Bok - 978 981 15 7031 5

Lecture Notes in Electrical Engineering 686

Pradeep Kumar Mallick

Preetisudha Meher
Alak Majumder
Santos Kumar Das Editors
Electronic
Systems and
Intelligent
Computing
Proceedings of ESIC 2020
Lecture Notes in Electrical Engineering

Volume 686
Series Editors
Leopoldo Angrisani, Department of Electrical and Information Technologies Engineering, University of Napoli
Federico II, Naples, Italy
Marco Arteaga, Departament de Control y Robótica, Universidad Nacional Autónoma de México, Coyoacán,
Mexico
Bijaya Ketan Panigrahi, Electrical Engineering, Indian Institute of Technology Delhi, New Delhi, Delhi, India
Samarjit Chakraborty, Fakultät für Elektrotechnik und Informationstechnik, TU München, Munich, Germany
Jiming Chen, Zhejiang University, Hangzhou, Zhejiang, China
Shanben Chen, Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai, China
Tan Kay Chen, Department of Electrical and Computer Engineering, National University of Singapore,
Singapore, Singapore
Rüdiger Dillmann, Humanoids and Intelligent Systems Laboratory, Karlsruhe Institute for Technology,
Karlsruhe, Germany
Haibin Duan, Beijing University of Aeronautics and Astronautics, Beijing, China
Gianluigi Ferrari, Università di Parma, Parma, Italy
Manuel Ferre, Centre for Automation and Robotics CAR (UPM-CSIC), Universidad Politécnica de Madrid,
Madrid, Spain
Sandra Hirche, Department of Electrical Engineering and Information Science, Technische Universität
München, Munich, Germany
Faryar Jabbari, Department of Mechanical and Aerospace Engineering, University of California, Irvine, CA,
USA
Limin Jia, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China
Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland
Alaa Khamis, German University in Egypt El Tagamoa El Khames, New Cairo City, Egypt
Torsten Kroeger, Stanford University, Stanford, CA, USA
Qilian Liang, Department of Electrical Engineering, University of Texas at Arlington, Arlington, TX, USA
Ferran Martín, Departament d’Enginyeria Electrònica, Universitat Autònoma de Barcelona, Bellaterra,
Barcelona, Spain
Tan Cher Ming, College of Engineering, Nanyang Technological University, Singapore, Singapore
Wolfgang Minker, Institute of Information Technology, University of Ulm, Ulm, Germany
Pradeep Misra, Department of Electrical Engineering, Wright State University, Dayton, OH, USA
Sebastian Möller, Quality and Usability Laboratory, TU Berlin, Berlin, Germany
Subhas Mukhopadhyay, School of Engineering & Advanced Technology, Massey University,
Palmerston North, Manawatu-Wanganui, New Zealand
Cun-Zheng Ning, Electrical Engineering, Arizona State University, Tempe, AZ, USA
Toyoaki Nishida, Graduate School of Informatics, Kyoto University, Kyoto, Japan
Federica Pascucci, Dipartimento di Ingegneria, Università degli Studi “Roma Tre”, Rome, Italy
Yong Qin, State Key Laboratory of Rail Traffic Control and Safety, Beijing Jiaotong University, Beijing, China
Gan Woon Seng, School of Electrical & Electronic Engineering, Nanyang Technological University,
Singapore, Singapore
Joachim Speidel, Institute of Telecommunications, Universität Stuttgart, Stuttgart, Germany
Germano Veiga, Campus da FEUP, INESC Porto, Porto, Portugal
Haitao Wu, Academy of Opto-electronics, Chinese Academy of Sciences, Beijing, China
Junjie James Zhang, Charlotte, NC, USA
The book series Lecture Notes in Electrical Engineering (LNEE) publishes the
latest developments in Electrical Engineering—quickly, informally and in high
quality. While original research reported in proceedings and monographs has
traditionally formed the core of LNEE, we also encourage authors to submit books
devoted to supporting student education and professional training in the various
fields and applications areas of electrical engineering. The series cover classical and
emerging topics concerning:
• Communication Engineering, Information Theory and Networks
• Electronics Engineering and Microelectronics
• Signal, Image and Speech Processing
• Wireless and Mobile Communication
• Circuits and Systems
• Energy Systems, Power Electronics and Electrical Machines
• Electro-optical Engineering
• Instrumentation Engineering
• Avionics Engineering
• Control Systems
• Internet-of-Things and Cybersecurity
• Biomedical Devices, MEMS and NEMS
For general information about this book series, comments or suggestions, please
contact [email protected].
To submit a proposal or request further information, please contact the
Publishing Editor in your country:
China
Jasmine Dou, Associate Editor ([email protected])
India, Japan, Rest of Asia
Swati Meherishi, Executive Editor ([email protected])
Southeast Asia, Australia, New Zealand
Ramesh Nath Premnath, Editor ([email protected])
USA, Canada:
Michael Luby, Senior Editor ([email protected])
All other Countries:
Leontina Di Cecco, Senior Editor ([email protected])
** Indexing: The books of this series are submitted to ISI Proceedings,
EI-Compendex, SCOPUS, MetaPress, Web of Science and Springerlink **
More information about this series at http://www.springer.com/series/7818

Pradeep Kumar Mallick Preetisudha Meher
• •
Alak Majumder Santos Kumar Das

•
Editors
Electronic Systems
and Intelligent Computing
Proceedings of ESIC 2020
123
Editors
Pradeep Kumar Mallick Preetisudha Meher
School of Computer Engineering Department of Electronics and
Kalinga Institute of Industrial Technology Communication Engineering
(KIIT) Deemed to be University National Institute of Technology Arunachal
Bhubaneswar, Odisha, India Pradesh
Yupia, Arunachal Pradesh, India
Alak Majumder
Department of Electronics and Santos Kumar Das
Communication Engineering Department of Electronics and
National Institute of Technology Communication Engineering
Arunachal Pradesh National Institute of Technology Rourkela
Yupia, Arunachal Pradesh, India Rourkela, Odisha, India
ISSN 1876-1100 ISSN 1876-1119 (electronic)

Lecture Notes in Electrical Engineering
ISBN 978-981-15-7030-8 ISBN 978-981-15-7031-5 (eBook)
https://doi.org/10.1007/978-981-15-7031-5
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore
Pte Ltd. 2020
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Programme Committee
Patron
Prof. Pinakeswar Mahanta, Director, NIT, Arunachal Pradesh
Conference Chair
Dr. Preetisudha Meher, NIT, Arunachal Pradesh
Conference Co-chair
Dr. Alak Majumder, NIT, Arunachal Pradesh
Special Session Chair
Dr. Sanjeev Kumar Metya, NIT, Arunachal Pradesh
Women in Engineering Chair
Dr. Yang Saring, NIT, Arunachal Pradesh
Technical/Reviewer Committee Chair
Dr. Subhodeep Mukhopadhyay, NIT, Arunachal Pradesh
Finance/Registration Chair
Dr. Yaka Bulo, NIT, Arunachal Pradesh
Local Arrangement Chair
Dr. Abir Jyoti Mondal, NIT, Arunachal Pradesh
Industry Liaison Chair
Dr. Sahadeb Roy, NIT, Arunachal Pradesh
General Chairs
Prof. Bidyut K. Bhattacharyya (Fellow IEEE), NIT Agartala
Prof. K. K. Mahapatra, Dean (Acad), NIT Rourkela
v
vi Programme Committee
Publication Chairs
Dr. Pradeep Kumar Mallick, Kalinga Institute of Industrial Technology (KIIT)
Deemed to be University
Dr. Santos Kumar Das, NIT Rourkela
Advisory Committee
Dr. Bhabani Sankar Swain, KU, South Korea
Dr. Benson Edwin Raj, Fujairah Women’s College Fujairah, UAE
Dr. Mohd. Hussain, Islamic University, Madina, Saudi Arabia
Dr. Frede Blaabjerg, Aalborg University, Denmark
Dr. Yu-Min Wang, National Chi Nan University, Taiwan
Dr. Gabriele Grandi, University of Bologna, Italy
Dr. Steve S. H. Ling, University of Technology, Sydney, Australia
Dr. Hak-Keung Lam, King’s College London, UK
Dr. Frank H. F. Leung, Hong Kong Polytechnic University, Hong Kong
Dr. Yiguang Liu, Sichuan University, China
Dr. Abdul Rahaman, Debre Tabor University, Ethiopia
Prof. Sanjeevikumar Padmanaban, Aalborg University
Dr. Bandla Srinivasa Rao, Debre Tabor University, Ethiopia
Prof. Pierluigi Siano, University of Salerno, Italy
Dr. Hussain Shareef, UKM, Malaysia
Dr. Anil Kavala, UKM, Malaysia
Dr. Hussain Shareef, Senior Engineer, Samsung Electronics, Seoul, South Korea
Prof. Michael Pecht, University of Maryland, USA
Prof. Josep M. Guerrero, Aalborg University, Denmark
Dr. Akshay Kumar Rathore, Concordia University, Montreal, Canada
Dr. Anup Kumar Panda, NIT Rourkela
Dr. Kishore Sarawadekar, IIT-BHU, Varanasi, India
Dr. Sachidananda Dehury, F.M. University, Odisha
Dr. Saroj Kumar Meher, ISI, Kolkata, India
Dr. Inderpreet Kaur, Chandigarh University
Dr. Debashish Jena, NITK, India
Dr. N. P. Padhy, IIT Roorkee
Dr. Sashikala Mishra, IIIT, Pune
Dr. Subhendu Pani, OEC, Odisha
Dr. Mihir Narayan Mohanty, ITER, SOA University
Dr. Sabyasachi Sen Gupta, Professor, IIT Kharagpur
Dr. P. Sateesh Kumar, Assistant Professor, IIT Roorkee
Dr. Swagatam Das, ISI, Kolkata
Dr. Sourav Mallick, NIT Sikkim
Dr. Ashok Kumar Pradhan, IIT Kharagpur
Dr. Bhim Singh, IIT Delhi
Dr. C. Sashidhar, CE, JNTUA
Dr. M. Sashi, NIT, Warangal
Programme Committee vii
Prof. Rajshree Srivastava, DIT University, Dehradun, India

Dr. Kolla Bhanu Prakash, K.L. University, Andhra Pradesh
Dr. Bhabani Shankar Prasad Mishra, KIIT University, Odisha
Preface
It is our pleasure to introduce the proceedings of the First International Conference

on Electronic Systems and Intelligent Computing (ESIC 2020) in relation to
complex issues of electronic circuit and system with the analysis of computational
techniques. The conference aims to create a forum for further discussion on inte-
grated information in the significant field incorporating a series of issues.
The relevance of the conference theme, to a wide variety of disciplines, is
reflected in the diverse range of papers submitted. The link between electronics and
computing techniques strengthens the area of research to be presented and provides
the precise information required for assessment. The international delegates and
keynote speakers like Prof. Indranil Hatai, NUS, Singapore, Prof. Kamalakanta
Mahapatra, NIT Rourkela, Prof. Goutam Sutradhar, Director, NIT, Manipur, Prof.
Roy P. Paily, IIT Guwahati, Prof. Sarat Kumar Patra, Director, IIIT Vadodara, Prof.
Sisir Kumar Nayak, IIT Guwahati, were highly impressed and have shown a high
level of international interest in the subject.
The level of interest in the subject matter of the conference was maintained by
submission of 225 papers at the conference through EasyChair. Every submitted
paper went through a precise review process. Each paper received at least three
reviews; where issues remained, additional reviews were commissioned. Finally,
140 papers were selected by 60 reviewers for presentation in four different tracks
like microwave and devices, communication and networking, signal and image
processing, and computations, mathematics, and control. Finally, 110 papers were
registered for the conference.
Finally, we would like to record our appreciation to the organizing committee
members for their work in securing a substantial input of papers to make the
conference successful. We are also indebted to those who served as reviewers,
faculties of our department, our students, and Prof. Pinakeswar Mahanta, Director
ix
x Preface
of NIT, Arunachal Pradesh, without their support the conference could not have
been the success that it was. We also acknowledge the authors themselves, without
their expert input there would have been no conference.
Bhubaneswar, India Pradeep Kumar Mallick

Yupia, India Preetisudha Meher
Yupia, India Alak Majumder
Rourkela, India Santos Kumar Das
About the Conference
Aim/Objective
The essential goal of the International Conference on Electronic Systems and
Intelligent Computing (ESIC 2020) is to give a stage to both electronic hardware
and software to communicate under one umbrella for further improvement of shrewd
electronic frameworks. Effective and secure information detecting, stockpiling, and
preparing play a significant part in the current data age. The cutting-edge savvy
electronic frameworks take into account the requirements of effective detecting,
stockpiling, and figuring. Simultaneously, effective calculations and programming
utilized for quicker examination and recovery of required data are winding up pro-
gressively. Storing and handling of the gigantic measure of organized and unstruc-
tured information are getting progressively hectic. Simultaneously, post-CMOS
technologies, Internet of things (IoT), and cyber-physical system (CPS) have been
moving with synchronous advancement of hardware and programming and lie over
ordinary customer gadgets. The exhibition and productivity of the present just as the
future ages of figuring and data handling frameworks are generally reliant upon
advances in both design and programming.
Contributions are sought in the following areas (but not limited to):
Track Details
Electronics, Devices, Circuits, and Systems
• Device modelling, post-CMOS devices
• Nano-technology, MEMS, and NEMS
• VLSI and microelectronic circuit design
• Hardware security
• Embedded systems and robotics
• System on chip (SoC)/network on chip design
• Field-programmable gate array (FPGA) design and applications
xi
xii About the Conference
• Electronic instrumentations
• Electronic power converters and inverters
• Electric vehicle technologies, etc.
Communications Engineering and Vehicular Technology
• Wireline communication
• Computer networking
• Telecommunication systems
• Wireless ad hoc and sensor networks
• Network/information security
• Security protocol
• Cognitive radio
• Cooperative communications
• Radio resource management and optimization
• Vehicular communication systems
• Information theory and coding systems, etc.
Computing and IT
• Software engineering
• Cloud computing, fog computing, and big data
• Business/enterprise architectures
• Mobile computing
• Distributed systems
• Information systems
• Knowledge discovery and data mining
• Artificial intelligence
• Decision support systems
• Visualization and computer graphic
• Approximate computing
• Image processing, information retrieval
• Natural language processing
• Machine learning, etc.
Pervasive Computing and Internet of Thing
• Internet of things, everything, and nano-things
• Location-dependent/personalized application
• System architectures and platforms
• Embedded computing with applications
• Low-power, energy-aware pervasive computing
• Smart devices
• Spaces and objects
• Middleware services and agent technologies
• Positioning and tracking technologies
• Intelligent devices and environments
About the Conference xiii
• User interfaces and interaction models

• Security and privacy issues for pervasive systems
• Ubiquitous computing, etc.
Control Systems
• Control theory and applications
• Robotics and autonomous systems
• Intelligent control
• Optimal control
• Robust control
• Adaptive control
• Linear and nonlinear control systems
• Complex adaptive systems
• Industrial automation
• Control systems technology, etc.
Energy and Power Systems
• Low-carbon power (CCS, nuclear, wind, solar, hydro, etc.)
• Low-carbon transport (fuel cells and hydrogen)
• Energy storage and network balancing
• Grid integration and infrastructure
• Distributed generation
• Energy efficiency and energy management
• Policy, targets, and resource efficiency
• Integrating finance, economies, and energy
• Computing, data, and ‘intelligent’ design
• Micro-grid operation
• System stability, dynamics, and control
• Power system protection and relaying
• Restructuring of power system, electricity, etc.
Engineering Education and Sustainable Development
• EESD and circular economy
• Global issues in EESD
• Overcoming barriers for EESD
• EESD and energy
• Stakeholders in EESD
• EESD and lifelong learning
• EESD and disciplinary approaches
Economics and Business Engineering
• Computational economics and finance
• FinTech
• Industrial and systems engineering
• e-commerce and e-business
xiv About the Conference
• Business information systems

• Technopreneurship, technology, and innovation
• Supply change management technology, etc.
Contents
A Reliable and Secure Wireless Network for VoIP Applications . . . . . 1

Vinod Kumar and O. P. Roy
Interdisciplinary Approaches Incorporating Computational
Intelligence in Modern Pharmacognosy to Address
Biological Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Tathagata Adhikary and Piyali Basak
Broadband Annular Ring Patch Antenna . . . . . . . . . . . . . . . . . . . . . . . 21
Thingbaijam Rajkumari Chanu, Somokanta Hijam, and Maxon Okramcha
Cooperative Spectrum Sensing with Improved Absolute Value
Cumulation Detection Based on Additive Laplacian Noise
in Cognitive Radio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Khushboo Sinha and Y. N. Trivedi
Solar Power Plant Site Selection: A Systematic Literature Review
on MCDM Techniques Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Rajkumari Malemnganbi and Benjamin A. Shimray
A FPGA-Based PUF Integrated Blockchain to Overcome
the Challenges of Internet of Everything (IoE) . . . . . . . . . . . . . . . . . . . 49
Lukram Dhanachandra Singh and Preetisudha Meher
Biometrics-Based Pig Identification: From Invention
to Commercialisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Subhranil Mustafi, Pritam Ghosh, Sanket Dan, Kaushik Mukherjee,
Kunal Roy, and Satyendra Nath Mandal
Adoption of IoT in Vehicular Traffic Control: An Overview . . . . . . . . 77
Joyeeta Goswami and Ashim Saha
xv
xvi Contents
Optimal Placement of FACTS Devices in the Power System

Using ANN to Increase the System Loadability . . . . . . . . . . . . . . . . . . 89
Mohammad Khalid Saifullah, Md. Monirul Kabir, Nasim Mia,
and Tanvir Ahmed
Impact of Noise Levels on SVM-GMM Based Speaker
Recognition System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
Renu Singh, Utpal Bhattacharjee, Arvind Kumar Singh,
and Madhusudhan Mishra
Digital Payment System and the Millennial in a Smart City:
An Antecedent to Technopreneurship . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Anil Kumar Singh and Tagiya Mudang
Efficient Fusion Based Multi-modal Biometric Authentication System
Using Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
R. Sindhuja and S. Srinivasan
An MOS-C Multifunction Filter Employing DXCCTA
for High-Frequency Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Karam Bharat Singh, Manoj Joshi, and Ashish Ranjan
Smart Car Parking Technique for Metropolitan City . . . . . . . . . . . . . . 143
Abu Shufian, Md. Asif Afroj, Md. Hasibuzzaman, Mehedi Al Miraz,
Avijit Sarker, Md. Jahidul Islam Rashed, and Md. Aman Miah
Implementation of Perturbation-Based MPPT Technique
Using Model-Based Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Sachin Angadi, Udaykumar R. Yaragatti, Yellasiri Suresh, and A. B. Raju
QRS Complex Detection Algorithm for Wearable Devices . . . . . . . . . . 167
Lalita Gupta
ORS: The Optimal Routing Solution for Smart City Users . . . . . . . . . 177
K. Krishna Rani Samal, Korra Sathya Babu, and Santos Kumar Das
An IoT-Based Pollution Monitoring System Using Data
Analytics Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Harshit Srivastava, Shashidhar Mishra, Santos Kumar Das,
and Santanu Sarkar
Modeling the Hierarchical Fuzzy System for Suitable
Crop Recommendation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
R. Aarthi and D. Sivakumar
Real-Time Emergency Vehicle Response System for Smart
City Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Lima Priyadarsini, Prashant Deshmukh, and Santos Kumar Das
Contents xvii
Modified Aggressive Packet Combining Scheme with Repetition

Code for Throughput Enhancement in High Error Rate Channel . . . . 221
Mancharla Ravi and Yaka Bulo
Review on Silent Speech Recognition Using EMG Sensors
and Voices After Laryngectomy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Pilla Devi Bhavani, V. Preethi, and Durgesh Nandhan
CNN-Based Individual Ghungroo Breed Identification
Using Face-Based Image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241
Kaushik Mukherjee, Sanket Dan, Kunal Roy, Subhojit Roy,
Subhranil Mustafi, Pritam Ghosh, Satyendra Nath Mandal,
Dilip Kumar Hajra, Santanu Banik, and Syamal Naskar
Development of Image-Based Disease Scale of Phoma Blight
of Potato Using k-Means Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
Kunal Roy, Sanket Dan, Kaushik Mukherjee, Subhranil Mustafi,
Pritam Ghosh, Satyendra Nath Mandal, Subrata Dutta,
and Ashis Chakraborty
Power Quality Improvement of Cascaded Savonius Rotor Based
Hydrokinetic Power System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Oying Doso and Sarsing Gao
Design and Simulation of a Sub-6 GHz Low Loss Band Pass Filter
Using Double Split Inductor for 5G Radio WLAN Applications . . . . . 275
Venkata Raghunadh Machavaram and Bheema Rao Nistala
Two Novel Configurations of Electronically Tunable Quadrature
Sinewave Oscillator Using CDBA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Shekhar Suman Borah, Ankur Singh, and Mourina Ghosh
Analog/RF and DC Performance Enhancement of a Pocket-Doped
Junction-Less TFET for Low Power Application . . . . . . . . . . . . . . . . . 295
Sneha Bharti, Suruchi Sharma, Abhishek Verma, Manisha Bharti,
and Baljit Kaur
A Study on Stability Analysis of QT Interval Dynamics of ECG
Using ARMAX Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
A. Jyothsana and J. Sivaraman
High Frequency Resonant Inverter with Shunt Active Power Filter
for Harmonic Compensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
Rohan Sinha and Rahul Raman
Modelling of Fiber Bragg Grating Sensor for Acoustic Pressure
Wave Detection of Partial Discharge in Power Transformer . . . . . . . . 329
Sorokhaibam Nilakanta Meitei, Kunal Borah, and Saibal Chatterjee
xviii Contents
Design and Investigation of Tunnel Junction Engineered

Dopingless-TFET with Improved DC and RF/Analog
Parameters Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Abhishek Verma, Suruchi Sharma, Sneha Bharti, Manisha Bharti,
and Baljit Kaur
The Role and Significance of Atrial ECG Components in Standard
and Modified Lead Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Shaik Karimulla and J. Sivaraman
Data Rate Enhancement Scheme and Mathematical Modelling
for Data Communication Through Blood . . . . . . . . . . . . . . . . . . . . . . . 357
Kauser Husainee and Tanmay Deshmukh
A Novel Design of Fast and Low Power Pull-Up and Pull-Down
Voltage Level Shifter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
Vickey Kumar, Suruchi Sharma, and Baljit Kaur
Modeling of Dual-Gate Carbon Nanotube Based Ion Sensitive Field
Effect Transistor (DG-CNTISFET) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
Hiranya Ranjan Thakur, Gaurav Keshwani, and Jiten Chandra Dutta
A Novel Approach to Calculate TF-IDF for Assamese Language . . . . . 387
Hsuvas Borkakoty, Chandana Dev, and Amrita Ganguly
A Low-Power Third-Order Passive Continuous-Time Sigma-Delta
Modulator Using FinFET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
Sarangam Kunamalla and Bheema Rao Nistala
Mitigation of Power Fluctuations in UPFC-Based Network . . . . . . . . . 407
Rajen Pudur, Pallavi Kumari, and Mrinal Kanti Rajak
An Efficient Approach for Barcode Encoding/Decoding
Using Pattern Substitution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419
S. Ieswaria, R. Bhavani, and R. Priya
Fabrication and Electrochemical Modeling of CNT-Based BioFET
for Cholesterol Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
Gaurav Keshwani, Kabyashree Hazarika, Hiranya Ranjan Thakur,
and Jiten Chandra Dutta
Design of LCL Filter for SPWM Inverter Based on Switching
Frequency and THD Criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
T. Saha, A. Mitra, and B. Halder
Level Control of Hopper Tank Process Using Model-Based
Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
V. Murugananthan, M. Valluvan, and G. Sakthivel
Contents xix
Hybrid L + C Edge Detection Algorithm for Harnessing Large

Image Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
Gurpartap Singh, Sunil Agrawal, and B. S. Sohi
Resolution Selective 2–6-Bit Flash ADC in 45 nm Technology . . . . . . . 475
Sarfraz Hussain, Rajesh Kumar, and Gaurav Trivedi
Measurement of Heartbeats for Well-Being Assessment Using
Photoplethysmographic Signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
Piyush Kumar, Arpan Sharma, Mayanmi Zimik, Shivam Parashar,
Renu Singh, Joyatri Bora, Aswini K. Patra, and Madhusudhan Mishra
An Iterative Node-Pair Time Synchronization (INTS) for Wireless
Sensor Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495
Sakuru K. L. V. Sai Prakash and N. Bheema Rao
Image Captioning Methodologies Using Deep Learning: A Review . . . 507
Virendra Kumar Meghwal, Namita Mittal, and Girdhari Singh
Performance of Speaker Recognition System Using Kernel Functions
Approach for Different Noise Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
Renu Singh, Arvind Kumar Singh, and Utpal Bhattacharjee
A Fuzzy Multicriteria Methodology for Selection Among Solar PV
Adoption Barriers in India . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
Anbesh Jamwal, Akshay Patidar, Rajeev Agrawal, and Monica Sharma
Suspension Modeling and Its Importance in Automotive Industries
for Implementation of Modern Vehicular Technologies . . . . . . . . . . . . 533
Sujan Neroula and Santanu Sharma
PIC Controller Based SVPWM Technique for Induction Motor
Variable Speed Drive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 541
Pabitra Kumar Guchhait, Suvadip Roy, and Abhik Banerjee
A Neural Network Based Strategic Placement and Task Assignment
for a Multi-agent System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555
Mukund Subhash Ghole and Anjan Kumar Ray
Review and Analysis of Charge-Pump Phase-Locked Loop . . . . . . . . . 565
Mriganka Gogoi and P. K. Dutta
Image Fusion: Challenges, Performance Metrics and Future
Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
S. B. G. Tilak Babu, I. Chintesh, V. Satyanarayana, and Durgesh Nandan
Factors Influencing Financing for Entrepreneurs . . . . . . . . . . . . . . . . . 585
Raman Kumar, Rajasve Kaushik, Shivani, Rishi Kant Kumar,
and Prashant Kumar
xx Contents
Simulation and Analysis of Hybrid PEMFC and Solar PV Cell Model

for Commercial Load Profile System . . . . . . . . . . . . . . . . . . . . . . . . . . 595
Archana Thoudam, Pabitra Kumar Guchhait, and Abhik Banerjee
How Does Perceived Desirability and Perceived Feasibility Effects
the Entrepreneurial Intention . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
Rajasve Kaushik, Raman Kumar, Manishankar Datta, Rishi Kant Kumar,
and Prashant Kumar
A New Topology for Photo Voltaic Energy Conversion System . . . . . . 617
Sudhir Maurya and Abhik Banerjee
A User-Centric Task Allocation Approach
in Mobile Crowdsensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 627
Moirangthem Goldie Meitei and Ningrinla Marchang
Fault Classification Based Approximate Testing of Digital
VLSI Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
Sisir Kumar Jena
Image Processing Using Case-Based Reasoning: A Survey . . . . . . . . . . 653
Rahul Barman, Saroj Kr. Biswas, Sunita Sarkar, Biswajit Purkayastha,
and Badal Soni
Automatic Speech Recognition of Galo . . . . . . . . . . . . . . . . . . . . . . . . . 663
Karter Nyodu and Samudra Vijaya
Early Detection of Diabetic Retinopathy Using Machine Learning
Techniques: A Survey on Recent Trends and Techniques . . . . . . . . . . 673
Dolly Das, Saroj Kr. Biswas, Sivaji Bandyopadhyay, and Sunita Sarkar
GUI-Based Secure Architecture Design for Distributed
Community Micro-grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685
Praveen Tiwari, Maj Sunil Kumar Panwar, Bikram Paul,
Pinakeswar Mahanta, and Gaurav Trivedi
A Hybrid Protocol for Stop and Wait ARQ in Markov
Two States Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 697
Sanjit Ningthoujam and Swarnendu K. Chakraborty
Comparative Study on a U-Slot Miniaturized CPW-Fed Multi-band
Antenna Applicable for 5G Communication . . . . . . . . . . . . . . . . . . . . . 707
Kaushal Mukherjee, Amitesh Das, and Sahadev Roy
An Experimental Investigation of FPGA-Based LMS Algorithm
for Adaptive Noise Cancellation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
Abinash Patnaik, Rajesh Kumar Patjoshi, and Rakhee Panigrahi
Contents xxi
ICI Mitigation with Fixed-Number (FN)- and Fixed-Region

(FRgn)-Based BS Coordination and Cooperation in a Multi-cell
Cellular Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
Janmoni Borah, Kshetrimayum Linthoinganbi Devi, and Joyatri Bora
InGaAs-Based Square-Shaped Nanopillar Array for High
Photodetector Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743
Smriti Baruah, Santanu Maity, and Joyatri Bora
Identification of Approximable Program Components
Using Edge Profiling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753
Priya Arundhati and Santosh Kumar Pani
Bidirectional LSTM with Attention Mechanism for Automatic
Bangla News Categorization in Terms of News Captions . . . . . . . . . . . 763
Md Shopon
Sentence Relation Classification Using Deep Learning Experiments . . . 775
B. Haripriya and Kishorjit Nongmeikapam
Spectral Analysis of Consonants in Arunachali Native
Language—Adi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 783
Sajal Sasmal and Yang Saring
Double Gate Tunnel FET Versus Double Gate MOSFET:
Electrical Properties Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 791
Menka Yadav
Wireless Access Support for Distribution Management System
(DMS) to Microgrid Communication in Power Network . . . . . . . . . . . 805
N. Himabindu, Rajashekar P. Mandi, Santoshkumar Hampannavar,
and Swapna Manasani
Implementation of a Charge Controller for Hybrid
Electric Vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
Gautam and Brajagopal Datta
Series-Stacked Non-parallel Multipath Differential Inductor
for C Band Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 823
Sunil Kumar Tumma and Bheema Rao Nistala
Implementation of DCGAN to Generate Gamocha
Design Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 831
Adarsh Pradhan, Arunav Buragohain, Uddipta Pathak, Sahil Ansari,
and Manali Baruah
Performance Evaluation of On-Grid Rooftop Solar PV System
in GMCH, Guwahati, Assam . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 839
Israfil Hussain, Utpal Chandra Baro, Rupa Chhetri, Vanita Agrawal,
and Abdul Latif
xxii Contents
Forced Alignment Method for Detection of Bodo

Word Boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 853
Raja Jwbthasa Daimari and Aniruddha Deka
Fault-Tolerant Cascaded Asymmetrical Multilevel Inverter
for the Solar-Based Electric Vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . 863
Hillol Phukan, Tamiru Debela, and Jiwanjot Singh
New H-Bridge for Symmetrical and Asymmetrical MLI
with Reduced Number of Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 875
Tamiru Debela, Hillol Phukan, and Jiwanjot Singh
Different EDGE Detection Techniques: A Review . . . . . . . . . . . . . . . . 885
Chinmoy Ghosh, Suman Majumder, Sangram Ray, Shrayasi Datta,
and Satyendra Nath Mandal
Design and Analysis of an Improvised Fully Differential Amplifier . . . 899
Swagata Devi, Koushik Guha, Naushad Manzoor Laskar, Sourav Nath,
and Krishna Lal Baishnab
Hyperspectral Band Selection Based on Variants of Rough
Set Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 909
Barnali Barman and Swarnajyoti Patra
Performance Analysis of Interference Alignment-Based Precoding . . . . 919
Amlan Deep Borah, Md. Anwar Hussain, and Joyatri Bora
Structural and Optoelectronic Properties of Spin-Coated
CH3NH3PbCl3 Thin Film Using Non-halide Source of Lead . . . . . . . . . 929
Paramita Sarkar, S. K. Tripathy, and K. L. Baishnab
Robust Gunshot Features and Its Classification Using Support
Vector Machine for Wildlife Protection . . . . . . . . . . . . . . . . . . . . . . . . 939
Vivek Singh, Kailash Chandra Ray, and Somanath Tripathy
Effect of pH Variations on Zinc Oxide (ZnO)
Nanowire Dispersions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 949
Vimal Kumar Singh Yadav and Roy Paily
Intensity Modulated U-Shaped Bent Tapered Optical Fiber
Concentration Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 957
Putha Kishore, D. Dinakar, Manchineellu Padmavathi, and L. Obulapathi
Performance Enhancement for Scattering Effect in Perovskite
Solar Cell with Distinct Cathode Materials . . . . . . . . . . . . . . . . . . . . . . 965
Sagar Bhattarai, Arvind Sharma, and T. D. Das
Need of Techno-Preneurship and Innovations for Community-Level
Cancer Screening in Arunachal Pradesh . . . . . . . . . . . . . . . . . . . . . . . 973
M. S. Momocha and K. Chaitan
Contents xxiii
Phrase-Based Machine Translation of Digaru-English . . . . . . . . . . . . . 983

Rushanti Kri and Koj Sambyo
EMT Analysis of Heavy-Duty EVs in Charging Station . . . . . . . . . . . . 993
Swati Shukla and Gaurav Trivedi
Analysis of the Transient Stability of a Multi-machine System
with SVC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1003
Srishail K. Bilgundi, M. Shivu, H. Pradeepa, and M. V. Likith Kumar
Projection-Based Throughput Computation of Three-Dimensional
Geographical Routing Protocol for Multi-hop Wireless Sensor
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1017
Afruza Begum and Md. Anwar Hussain
Digital Image Processing Based Automatic Fabric Defect Detection
Techniques: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1029
Vaishnavi Raut and Indrasen Singh
A Balanced Expert System to Manage Parkinson’s Disease
by Identifying Major Risk Factors: B-TESM-PD . . . . . . . . . . . . . . . . . 1039
Arpita Nath Boruah, Saroj Kr Biswas, Sivaji Bandyopadhyay,
and Sunita Sarkar
Bioinformatics Advancements for Detecting Epidemic Disease
Using Machine Learning Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 1051
Bikash Baruah and Manash Pratim Dutta
A Strategic Community Control-Based Power Flow Between
Grid-Integrated PV Houses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1061
Jatin M. Soni, Deep V. Patel, Rajendra V. Patel, and Hardik P. Modha
Classification of Buildings and Vehicles in Google Map Satellite
Images Using Random Forest Classifier . . . . . . . . . . . . . . . . . . . . . . . . 1073
Ritik Gautam, Saravanan Chandran, and Jose Hormese
Addressing the False Positives in Pedestrian Detection . . . . . . . . . . . . . 1083
N. J. Karthika and Saravanan Chandran
Hybrid CR Network: An Approach Based on Interweave-
and Underlay-Type CR Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1093
Varun Kumar, Mangal Singh, Sarat Kumar Patra, and Poonam Singh
Intensive Study on Compact Integrated Optic Couplers
Using Grating Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1105
Bidyut Deka, Aradhana Dutta, and Partha Pratim Sahu
Effect of AlN Spacer Layer on the Proposed
MIS-AlGaN/GaN HEMT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1115
Ravi Ranjan, Nitesh Kashyap, and Ashish Raman
xxiv Contents
Enhancement of Isolation in MIMO Antenna System . . . . . . . . . . . . . . 1123

Neelima Koppala, Sree Lochan Kanthi Kumar Kuppachi,
Rohith Sai Kumar Gajula, Bhanu Ravindranath Kasturi,
and Rupa Sree Jangam
Stochastic Frontier Model Approach of Technical Efficiency
for Paddy Farming in Odisha . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1135
Priyabrata Bhoi, Deepak Kumar Swain, Subhadra Mishra,
Debahuti Mishra, Gour Hari Santra, and Samarjeet Borah
Improving Security Concerns of Mobile Cloud Computing
Environment with Quantum Cryptographic Techniques . . . . . . . . . . . 1141
Sudhanshu Maurya, Kuntal Mukherjee, and R. Badlishah Ahmad
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1151

About the Editors
Dr. Pradeep Kumar Mallick is currently working as Associate Professor in the

School of Computer Engineering, Kalinga Institute of Industrial technology (KIIT)
Deemed to be University, Odisha, India. He has completed Post-Doctoral Fellow
(PDF) from Kongju National University South Korea, PhD from Siksha Ó’
Anusandhan University, M. Tech. (CSE) from Biju Patnaik University of
Technology (BPUT), and MCA from Fakir Mohan University Balasore, India.
Besides academics, he is also involved various administrative activities, Member of
Board of Studies, Member of Doctoral Research Evaluation Committee, Admission
Committee etc. His area of research includes Algorithm Design and Analysis, and
Data Mining, Image Processing, Soft Computing, and Machine Learning. He has
published 10 books and more than 70 research papers in National and international
journals and conference proceedings.
Dr. Preetisudha Meher is currently working as Assistant Professor in National

Institute of Technology, Arunachal Pradesh, India. She has completed her PhD
from National Institute of Technology, Rourkela, India in the year 2015. Her PhD
research work is on “Design and analysis of improved domino logic with noise
tolerance and high performance”. She has done M. Tech and B. Tech from BPUT
Odisha. She has guided almost 10 BTECH and 8 MTECH projects in various
domains. She has published 12 journals, 32 conference papers, 5 book chapters and
2 books during her industry and academic work. At present, she is working in the
field of Low Power VLSI, Bio Informatics and IOT. She has been awarded as
Outstanding Woman in Engineering Award conferred by Centre for Advance
Research and Design at International Women Meet-2018 on 1st March 2018 held at
Chennai, India. She has been awarded Best Paper in the year 2018 at International
Conference on Information System Design and Intelligent Applications, Mauritius
and also in the year 2019 at International conference on Intelligent System and
Control Communication at Bangkok. She has organized a Workshop on “Recent
Trends in VLSI Design” at NIT Arunachal Pradesh from 8th to 12th January
2018 and also International conference on Intelligent System and Control
Communication at Bangkok on 27-28 June 2019 as convener.
xxv
xxvi About the Editors
Dr. Alak Majumder has completed his PhD degree from NIT, Arunachal Pradesh,
India and is working there as an Assistant Professor in the Department of ECE. He
earned his M-Tech in Microelectronics & VLSI and BE in Electronics &
Telecommunication Engineering from NIT Agartala and TIT Agartala. He also
served as Assistant Professor in the Department of ECE at ICFAI University,
Agartala. He has 75+ technical articles, 6 filed Indian Patents and 1 US Patent to his
credit. His current research interests include Analog-Digital VLSI, Clock Gating
and Distribution and Wireline Communication Circuit. He is a Member of the
IEEE, IAENG and IACSIT.
Prof. Santos Kumar Das is a R&D professional with more than 17 (1 years after
BTECH, 7 years after MTECH (Foreign Industry), 5 years during PhD and 4 year
after PhD) years of valuable experience on various Software developments and
research activities both in industries and academics. At present, he is working at
National Institute of Technology as an Assistant professor. He worked at SDSU,
California, US as an adjunct faculty for one month on Software Defined Network.
He is academically astute and technically brilliant with a Master Degree special-
izing in Electrical Communication Engineering from Indian Institute of Science
(IISc) Bangalore, India and PhD Degree from National Institute of Technology,
Rourkela, India. His PhD research work was on “Quality of service estimation for
an Optical Virtual Private Network over WDM/DWDM network”. His current
research work is on software defined network (SDN), IoT, sensor networking,
optical networking, Smart City product development, and embedded system. He
has guided almost 40 BTECH and 30 MTECH projects in various domains. He has
published 12 journals, 25 conference papers, 8 patents (filed), and 4 book chapters
during his industry and academic work. At present, he has submitted one PhD
student and guiding 12 PhD ongoing students in optical networking, IoT and
embedded domain. He is also managing three high value funded projects (440
lakh), where 30 to 40 students are working.
A Reliable and Secure Wireless Network
for VoIP Applications
Vinod Kumar and O. P. Roy
Abstract In a wireless communication system, the Quality of Service (QoS)

becomes an active area of research as the demand for real-time applications like
voice or video communication increases day by day. In this paper, a new hybrid
INTER-SR & INTRA-SR based reliable wireless network is designed for Voice over
Internet Protocol (VoIP) applications. It is mainly consisting of one optimization
algorithm named as Genetic Algorithm (GA) in conjunction with two classification
approaches such as Artificial Neural Network (ANN) and Support Vector Machine
(SVM). The proposed algorithm selects a reliable and more reliable route based on
the nodes properties such as energy consumption and delay. In the end, three distinct
QoS parameters such as routing overhead, delay and jitter are analyzed for two
communication scenarios such as with attack and without attack in MANET simu-
lator. The results demonstrate that the proposed algorithm performs with satisfying
results compared to individual routing and optimization approaches.
Keywords Wireless network · Voice over internet protocol · Reliability ·

INTER-SR · INTRA-SR · GA · SVM · ANN
1 Introduction
The present day’s society observes a revolutionary change in communication tech-

nologies to make life comfortable. VoIP is the technology to make communication
possible over IP networks [1]. Using VoIP is very advantageous such as cost per call
[2]. Further, VoIP technique suffers from many disadvantages, which makes people
still remain on phone calls [3]. The major drawback faced by the VoIP calls is the
packet loss and the delay that occurs during the congestion problem over IP networks.
Another problem is the wastage of bandwidth [4]. A list of codecs along with their
V. Kumar (B) · O. P. Roy

Department of EE, NERIST, Nirjuli, Arunachal Pradesh 791109, India
e-mail: [email protected]
O. P. Roy
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer 1
Nature Singapore Pte Ltd. 2020
P. K. Mallick et al. (eds.), Electronic Systems and Intelligent Computing, Lecture Notes
in Electrical Engineering 686, https://doi.org/10.1007/978-981-15-7031-5_1
2 V. Kumar and O. P. Roy
Table 1 Codec in general

Codec Format size Codec Format size
VoIP
G. 729 10 G. 729 A 10
LPC 14 G-723.1 30
G. 729 E 10 G-729 D 10
format size is listed in Table 1. The payload produced by VoIP is small as the VOIP
is sensitive to delay; the larger is the size of the payload, more time is required by
the VoIP to generate the payload, and hence increase the delay. VoIP systems still
exist on data networks that means there may be a lack of security and types of attacks
related to any data network [5, 6]. In the VoIP, the voice message is translated into
IP packets. The data travels from different access points and chances of data loss
increases within the route while transmitting the data from its source to destination
[7]. From an existing survey, it has been observed that Denial of Service is one of
the most common attacks that affect VoIP data. Therefore, it is necessary to design
a reliable network.
2 Related Work
A number of protocols are coming into existence with improved security features,
but still not meet the desired requirement. Gupta and Shmatikov [8] have presented a
reliable network of VoIP protocol. The researchers started a replay attack that affects
voice and break the security of the transport layer [8]. VoIP evaluation has also been
performed by Audah et al. [9] using distinct codec schemes in NS-2 simulator The
author has examined the performance in terms of Quality of Service (QoS) [9]. Also,
Ad hoc On-demand Distance Vector (AODV), Dynamic Source Routing (DSR),
Temporally-Ordered Routing Algorithm (TORA) have been analyzed in OPNET
simulator [10]. It has been concluded that the data VoIP packet transmission using
the TORA protocol has been transmitted with better quality against the large data
traffic [10]. Furthermore, for VoIP data transmission carried out using AODV and
DSR routing mechanism has been conducted by Jasani [11]. The results indicate
that data transmission using AODV routing is more suitable compared to DSR while
VoIP packet is transmitted in the MANET network [11]. Sanchez-Iborra and Cano
[12] have investigated the quality of MANET routing protocols in transmitting video
data using OLSR and AODV [12].
3 Proposed Algorithm
The objective is to make VoIP data reliable and secure. A number of researchers have
contributed in this direction but not attained desired security. To increase data security
A Reliable and Secure Wireless Network … 3
INTRA-SR Domain 2 Domain 3

INTER-SR
5 6
4
S 3 7
1 Domain 1 Domain 4
D
Fig. 1 Simulation area of (1000 × 1000 m2 )
of VoIP application, SVM and ANN are used for decreasing routing overhead and
packet loss. The route has been selected based on the concept of INTER-SR (Inter-
State Routing) and INTRA-SR (Intra State Routing). INTER-SR is a type of protocol
in which the data is transmitted from one domain to another domain, whereas, in
INTRA-SR, the data is transmitted within the domain.
3.1 Network Design
A network of 1000 × 1000 m2 is created with N number of nodes as shown in Fig. 1.

Nine nodes are deployed denoted by (1–7), S and D. Nodes 1 to 4 and S comes under
domain 1 and the route formation occurred due to INTRA-SR process. To reach data
from source to destination, we also need INTER-SR technique in combination with
INTRA-SR technique.
3.2 Security and Prevention
The problem of routing in VoIP-based communication model is solved by integrating

INTRA-SR and INTER-SR approach. The problem of security and prevention is not
covered by this approach, explained in the result and analysis.
To resolve this problem, we designed a hybrid mechanism using SVM and ANN
techniques. SVM is a binary classifier for checking two conditions for the malicious
node or genuine node. To validate the detection probability of SVM for the malicious
or normal node, ANN and SVM are used as a classifier. In route formation, the data
is transmitted to the nearby node as shown in Fig. 2, node 2 pass VoIP packet to node
3 by using the concept of distance. But, the routing mechanism of VoIP does not
check whether the node is dropping the message or not. To degrade the VoIP packet
drop of the proposed wireless network, an optimization scheme has been used with
novel fitness function. The fitness function helps to select the node as in Table 2.
INTRA-SR Domain 3
INTER-SR 5
6
Domain 1 4
2 Domain 2
Domain 4
S 3
7
1 D
Fig. 2 Reliable network using GA with (SVM & ANN) approach
Table 2 Node’s properties

Nodes Energy consumption (mJ) Delay (ms)
2 10 4
3 12 3
4 12 2
In domain 1, during route formation, VoIP message packet is forwarded by node

1 to node 3. But, node 3 is not an accurate node as it drops data packet. It is necessary
to pass data to a genuine node, which has been performed here, using GA. The GA
checks energy consumed by each nearby node as in our example, node 3 and node 4
consumes same energy, therefore, in such condition, one must go for checking other
parameters that is a delay in this case. Node 4 forward data with less delay (∼ =2 s)
and hence node 1 transmits data to node 4.
Malicious Detection of VoIP communication using GA with SVM

and ANN
Required Input: T Training Data as nodes properties. C Target/Category
in terms of communicating and non-communicating nodes. N Number of
Neurons, MN Malicious Nodes,
Obtained Output: Start
1 To optimized the T, Genetic Algorithm (GA) is used
2 Set up basic parameters of GA: Population Size (P) – Based on the number
of properties, CO – Crossover Operators, MO – Mutation Operators
OT – Optimized Training Data Fitness Function:

1(T r ue); i f Fs < Ft = T hr eshold Pr oper ties
F( f ) =
0(False); Other wise
where Fs : Current Node Properties and Ft : Threshold properties for all

nodes based on energy consumption, delay and position
3 Calculate Length of T for R

4 Set Optimized Training Data OT = [] For i in range of R
5 Fs = T (i) = Selected N ode Pr oper ties . // Current Data from N
6 Ft = T hr eshold Pr oper ties . // Average of All Data
7 F( f ) = Fit Fun(Fs , Ft ). Nvar = Number of variables
8 BestProp = OT = GA (F(f), T, Nvar, Set up of GA)
9 End – For SVM training data initialization OT - Nodes optimized
10 For I = 1 All Nodes If Node Property (I) == Real
11 Define Cat category of training data Cat (1) = Node Properties (I)
12 Else Cat (2) = Node Properties (I) End - If
13 End – For Train_Struct=SVMTRAIN (T, Cat, Kernel function)
14 OT= Train_Structure.SupportVector //finding ANN training data
15 Initializing ANN basic parameters – Number of Epoch (E) // ANN
Iterations – Number of Neurons (N)
16 F i = 1 OT If T belongs to the properties of communicating nodes
17 Group (1) = Training data properties according to renodes
18 Else if T belongs to properties of non-communicating nodes
19 Group (2) = Training data properties for non-real nodes
20 Else Group (3) = Training data Extra properties End – If
21 End – For Initialize ANN for data
22 VoIP-Net = Newff (T, Gr oup, N )
23 VoIP = Train (VoIP, Training data, Group) Testing:
24 Current Node = Nodes Properties in Cloud-Net
25 Authentication = simulate (Cloud-Net, Current Node)
26 If Authentication = True Genuine node not consider as a malicious
27 Else MN = Malicious Node
28 End Return: MN a list of Malicious Nodes
29 End
Above algorithm is designed using hybridization of SVM, ANN and GA. Algo-
rithm first uses SVM using radial basis function (RBF) as a kernel and find the
most helpful properties for ANN which are called support vectors. ANN is used to
train and classify nodes authorization which helps in the improvement of network
performance.
4 Results and Analysis
In this research, the performance of the communication system for VoIP message
transmission is considered with two cases, one communication without attack and
another communication with attack. The effect of occurring malicious node inside
the network when making the voice call and data packet transmitted between the
source and destination node through different domain nodes have been discussed. The
performance has been analyzed by constructing network with multiple nodes (5, 10,
15, 20, 25 and 30). The quality of transmitting VoIP message transmission has been
observed on the basis of delay, packet loss and throughput. The network simulation
environment consists of 30 numbers of nodes that were positioned randomly using
random waypoint as a mobility model. The simulation parameters are mentioned in
Table 3. The results such as routing overhead, average delay and average jitter have
been computed in two scenarios that is communication with and without attack. To
obtain accurate results, the network is simulated at least ten times and their average
values are noted in the form of Tables 4, 5 and 6.
Figure 3 demonstrates the routing overhead for a reliable network that is used
for VoIP application. A very close variation has been observed for three different
approaches such as (1) with INTER-SR & INTRA-SR, (2) with GA and (3) for
Proposed Algorithm. For route construction, INTER-SR with INTRA-SR has been
used, which enables the user to communicate with the remote user that is located in
different domains. The route optimization is performed with GA as an optimization
technique. GA helps to find the properties of each node based on energy consump-
tion, delay and node’s coordinate and the data transmission is done based on node’s
properties rather than the concept of the closest node within the route. Later on, the
problem of identifying the best node has been resolved using SVM in hybridization
with the ANN algorithm. The graph is shown in Fig. 3 indicates that the proposed
Table 3 Simulation
Considered parameters Variation of parameters
parameters
Area 1000 × 1000 m2
Nodes 5, 10, 15, 20, 25 and 30
Number of malicious nodes 10
Simulation time (s) 400 s
Routing protocol INTER-SR and INTRA SR
Mobility model Random waypoint
Table 4 Routing overhead analysis

Number of Communication without attack Communication with attack
nodes INTER-SR With GA Proposed INTER-SR With GA Proposed
& algorithm & algorithm
INTRA-SR INTRA-SR
5 2540 2000 1800 4521 3254 1726
10 3002 2154 1985 5264 3642 1875
15 4251 3869 3125 5728 4782 3745
20 5126 3985 3547 6357 4759 3675
25 6482 5264 4597 8581 5025 4871
30 7356 6542 5942 9264 6235 5297
Table 5 Delay analysis

nodes INTER-SR & GA Proposed INTER-SR & GA Proposed
INTRA-SR algorithm INTRA-SR algorithm
5 1 0.6 0.5 5 4 2
10 6 5 4 8 8 4
15 8.2 8 2 14 10 5
20 15 9 2 16 12 6
25 25 11 2.5 22 19 8
30 30 18 10 40 22 10
Table 6 Jitter analysis (ms)

nodes With With GA Proposed With With GA Proposed
INTER-SR algorithm INTER-SR algorithm
& &
INTRA-SR INTRA-SR
5 1.9 1 0.3 3.5 3 2.5
10 1.8 1.1 0.35 7 5.9 3
15 6.5 3.2 1.8 16 14 4.6
20 11.5 4 2.5 20 18 8.7
25 8 4.2 3 21 20 12
30 15 6.5 4.5 28 22 24
8000
Routing Overhead
7000
6000 with INTER-SR
5000 &INTRA-SR
4000
3000 with GA
2000
1000
0 Proposed
5 10 15 20 25 30 Algorithm
Number of Nodes
Fig. 3 Routing overhead without Attack
algorithm transmits data with small routing overhead compared to individual routing
and routing with optimization approach.
Figure 4 represents the routing overhead observed with the variation of the number
of nodes in the presence of an attack. The graph indicates that with the increase in
10000
9000 with INTER-
Routing Overhead
8000 SR &INTRA-
7000 SR
6000
5000
4000 with GA
3000
2000
1000
0 Proposed
5 10 15 20 25 30 Algorithm
Number of Nodes
Fig. 4 Routing overhead with Attack
the communicating nodes within the network, the routing overhead increased and
the maximum routing overhead measured for 30 numbers of nodes with INTER-SR
and INTRA-SR, with GA and for the proposed algorithm are 9264, 6235, and 5297,
respectively.
To show the effectiveness of the proposed algorithm, we evaluated average delay
and jitter by generating VoIP messages in a large amount. The attained results related
to average delay obtained without and with attack are shown in Figs. 5 and 6. The
difference of delay values measured with INTER-SR and INTRA-SR, with GA and
for combination of all that is proposed algorithm has been presented in Fig. 5. From
the above graph, it has been observed that the average delay values recorded for with
GA are almost twice and for the proposed algorithm almost thrice compared to the
simple routing techniques performed by INTER-SR and INTRA-SR combination.
The jitter observed without attack and with the attack for VoIP application with
three different algorithms is shown in Figs. 7 and 8, respectively. It is clear from
Figs. 7 and 8 that when the classification algorithm in addition to the optimization
approach is used, the occurrence of jitter has been reduced in a great extent.
35
Average Dealy (ms)
30
with INTER-SR
25
20 &INTRA-SR
15 with GA
10
5
0 Proposed
5 10 15 20 25 30 Algorithm
Number of Nodes
Fig. 5 Average delay without attack for VoIP application

50
with INTER-SR
Average Delay (ms)

40
&INTRA-SR
30
20 with GA
10
0
Proposed
5 10 15 20 25 30
Algorithm
Number of Nodes
Fig. 6 Average delay with attack for VoIP application
20
15 with INTER-SR
Jitter (ms)
&INTRA-SR
10
with GA
5
0 Proposed Algorithm
5 10 15 20 25 30
Number of Nodes
Fig. 7 Average jitter without attack for VoIP application
30
25
with INTER-SR
Jitter (ms)
20
&INTRA-SR
15
10 with GA
5
0
5 10 15 20 25 30 Proposed
Number of Nodes Algorithm
Fig. 8 Average jitter with attack for VoIP application
5 Conclusion
In the paper, we have presented a reliable wireless network for VoIP-based applica-
tions. For routing, the concept of INTER-SR and INTRA-SR has been used to resolve
the problem of handover as the data has to be transmitted from one domain to another
domain. Two different scenarios of data communications such as in the presence
of with and without attack have been presented for three approaches as discussed
in the results and analysis section. The performance of the designed network has
been examined based on routing overhead, average delay and average jitter for VoIP
application. The results demonstrated that better results compared to existing routing
protocol has been obtained that guarantee the desired delay and jitter required for
data transmission in VoIP applications in a wireless network.
References
1. Mohd Ali A, Dhimish M, Mather P (2019) WLAN protocol and network architecture selection
for real-time applications (2019)
2. Chakraborty T, Misra IS, Prasad R (2019) VoIP over wireless LANs—Prospects and challenges.
In: VoIP technology: applications and challenges, Springer, Cham, pp 71–93
3. Alsaadi MS, Alotaibi ND (2019) Routing protocols design and performance evaluation in
wireless mesh networks. Int J Technol Diffus (IJTD) 10(1):60–74
4. Chakraborty T, Misra IS, Prasad R (2019) Technique for Improving VoIP performance over
wireless LANs. In: VoIP technology: applications and challenges, Springer, Cham, pp 95–121
5. Abualhaj MM, Al-Tahrawi MM, Al-Khatib SN (2019) A new method to improve Voice over IP
(VoIP) bandwidth utilization over Internet Telephony Transport Protocol (ITTP). In: Proceed-
ings of the 2019 8th international conference on software and information engineering.ACM,
pp 192–195
6. Estepa R, Estepa A, Madinabeitia G, Davis M (2018) Improving the energy efficiency of
VoIP applications in IEEE 802.11 networks through control of the packetization rate. In: XIII
Jornadas de Ingeniería telemática (JITEL 2017). Libro de actas, pp 23–29
7. Hsieh WB, Leu JS (2018) Implementing a reliable VoIP communication over SIP-based
networks. Wirel Netw 24(8):2915–2926
8. Gupta P, Shmatikov V (2006) Key confirmation and adaptive corruptions in the protocol secu-
rity logic. In: Proceedings of the Joint Workshop on Foundations of Computer Security and
Automated Reasoning for Security Protocol Analysis (FCS-ARSPA’06), pp 113–142
9. Audah L, Kamal AAM, Abdullah J, Hamzah SA, Razak MAA (2015) Performance evaluation
of voice over IP using multiple audio codec schemes. Asian Research Publishing Network
(ARPN). J Eng Appl Sci 10:8912–8919
10. Ravikanti S, Preeti G (2015) Evaluating the performance of reactive unicast routing protocols
with OPNET simulator in MANETS under VOIP. Int J Innov Res Sci Eng Technol 4(7) (2015)
11. Jasani H (2012) Evaluations of AODV and DSR for QoS requirements. In: Proceedings of the
1st annual conference on research in information technology ACM, pp 1–6 (2012)
12. Sanchez-Iborra R, Cano MD (2014) An approach to a cross layer-based QoE improvement for
MANET routing protocols. Netw Protocols Algor 6(3):18–34
Interdisciplinary Approaches
Incorporating Computational
Intelligence in Modern Pharmacognosy
to Address Biological Problems
Tathagata Adhikary and Piyali Basak
Abstract Computer simulations to address research problems have touched almost

all the sectors of science. Ethnopharmacological data demands linkage with different
biomedical databases for both prospective and retrospective studies. Computational
intelligence bridges the gap between diverse resources already available (from the
molecular level to macroscopic characteristics of drugs) marking its significance
in computer-aided drug design and gives an insight into the mechanism of action
of bioactive compounds. Computational methods are utilized to study the process
of extraction, isolation, structure prediction, metabolomics, biosynthesis, dereplica-
tion, phytochemical library construction, and assessment of their bioactivity. Further,
in silico models aim to minimize the cost and time associated with conventional
screening techniques. Electronic structure determination, docking, geometry opti-
mization are few methods to name that are popular in biological sciences. Softwares
like Gaussian 94, MOPAC, GAMES, Spartan, and Sybyl showed their dominant
role in deciphering the chemistry underlying a process. The complexity of collecting
vast biological data and extracting the information of interest for subsequent knowl-
edge provision and deduction of conclusions is mollified to a great extent using
bioinformatics. This review discusses the hurdles, current progress, and potential of
computational intelligence in drug discoveries/interactions.
Keywords Bioinformatics · Molecular simulation · Phytochemicals · Bioactivity ·

Statistics
1 Introduction
Pharmacognosy aims to develop databases on natural product-based drugs that

primarily hold information about sources, geographical distributions, identification,
T. Adhikary · P. Basak (B)

School of Bioscience and Engineering, Jadavpur University, Kolkata, West Bengal, India
T. Adhikary
12 T. Adhikary and P. Basak
extraction/isolation, bioactivity, and toxicity profiling of drugs. Traditional recipes

of medicine by apothecaries are still practiced ethnically but their scopes can be
extended globally. The conventional techniques of drug discovery can be made less
labor-intensive with minimal investment in terms of capital and time by the use of
high-throughput technologies. Approximately one-third of the present literature on
drugs involves plants as a source [1]. The immense data generated since centuries in
the field of pharmacognosy remain scattered and thus limit the essence of strategic
analysis of interlinked phenomena. Combining ancient manuscripts of ethnobotany
with modern researches and subsequent implementation of data mining techniques
forms a basic platform for logical representation of data.
Standing in the era of increasing pharmaceutical demands, computational intel-
ligence bridges the gap between diverse resources already available (from the
molecular level to macroscopic characteristics of drugs) marking its significance in
computer-aided drug design. Multifold perspectives leading to an exponential growth
of data demand for procedures to efficiently manage, integrate, visualize and analyze
data. Combining the knowledge of genomics, proteomics, and metabolomics using
bioinformatics advances in predicting pathways and accurate scientific reasoning of
any process [2]. Further, in silico models aim to minimize the cost and time associated
with conventional screening techniques.
Computer-assisted modeling of a physical entity or process has its own challenges
also. The vastness of databases, accuracy of the information on pharmacokinetics,
significance of assumptions, and their limitations influence the precision of predic-
tion, i.e., success of such models. Keeping in mind the pros and cons, this review
attempts to discuss the hurdles, current progress, and potential of computational
intelligence in drug discoveries/interactions.
2 The Diversity of Databases
Initial efforts in the digitization of existing literature like the Biodiversity Heritage
Library and the China-US Million Book Digital Library Project highlighted the
challenges faced in gathering resources of interest from them. Automated character
recognition and string-matching algorithms among many algorithms can be exploited
to overcome such hurdles without semantic knowledge. The Encyclopedia of Life
(EOL) and the Barcode of Life (BOL) provides a framework for taxonomic advance-
ments that eliminate the linguistic challenge in existing literature [3]. An attempt to
extract information from a semi-structured database of traditional Chinese medicines
used ontology by an executable knowledge extraction language [4]. In addition,
integration of heterogeneous data was achieved using Semantic Web technologies
and semantic-based interoperable e-Science for identifying plant species in Chinese
medicine having antidepressant activities and development of Traditional Chinese
Medicine (TCM) system, respectively [5, 6].
Among simple file directories, object-oriented databases, and other types of
databases, the relational database has gained popularity due to its efficient way of
Interdisciplinary Approaches Incorporating Computational … 13
storing and retrieving huge data via indexes, normalization, referential integrity, trig-
gers, and transactions. MySQL and PostgreSQL are noteworthy relational database
software that is free and quite popular in bioinformatics where data is assessed using
a structured query language (SQL) [7]. Considering one or multiple parameters (like
genomic data, the composition of bioactive compounds, ethnobotanical knowledge,
medicinal uses, etc.), several databases are developed for integrating and indexing
information. The International Ethnobotany Database (ebDB) is a non-commercial,
multilingual, and primary repository for ethnobotanical data that supports translation
and data entry by researchers. A large amount of experimental data in the field of
natural products (biochemical and biological activity from in vitro, in situ, in vivo, and
in-human experiments) is stored and analyzed by a relational database NAPRALERT
[8]. The database developed by The United States Department of Agriculture (USDA)
and maintained by Dr. Park provides information about secondary metabolites of
selected medicinal plants used in inflammation, heart disease, obesity, hypertension,
kidney disease, and diabetes whereas Dr. Duke’s Phytochemical and Ethnobotanical
Database at the USDA provides insights into a taxonomic classification of medicinal
plants, their active metabolites, and pharmacological behavior. Protein Data Bank
(PDB) being an important database in structural biology (for example, structural
genomics) stores 3D structural data of proteins and nucleic acids. SCOP and CATH
classify protein structures, while PDBsum provides a graphical overview of PDB
entries. UniPort is a database of protein sequences derived mainly from genome
sequencing projects, while KEEG specifies interactions and biosynthetic pathways
of secondary metabolite. Herb Ingredient’s Target (HIT) connects PDB, UniPort,
Therapeutic Target Database (TTD), Pfam, KEEG, and other databases to analyze
biological targets of herbal medicines listed in TCM.
Interestingly, the database CMKb is designed for aboriginal medicinal plants of
Australia relating their taxonomy, phytochemistry, biogeography, biological activ-
ities, and images. Raintree is also a geographic region-based database specific to
the plants of the Amazon rainforest [9]. Genome and transcriptome related data
of medicinal plants from sequencing projects and plant Expressed Sequence Tags
(ESTs) is analyzed by linking it with functional properties of plants and maintained
by Plant Genomes Central (PGC) and EGENES database. Annotation of EST data
with minimal redundancy is done in databases like KEGG genes, SwissProt, etc.
to aid in the exploration of molecular markers [10]. A recent database, Medicinal
Plants Genomics Resource (MPGR), collects transcriptomic and metabolomic data
from diverse taxonomic plant species to key out putative genes and pathways involved
in metabolite production. As a case in point, effector genes in the biosynthesis of
ginsenoside in Panax Ginseng have been identified by analyzing EST data [11].
Further, the transcriptome data can be used to design a co-expressed gene network
leading to gene discovery [12].
Instead of conventionally inspecting physical characteristics for bioscreening, the
authenticity of a plant species is now done by DNA barcoding using DNA based
markers due to its definitive results. For animals, the chosen sequence (i.e., DNA
barcode) is from the mitochondrial gene cytochrome c oxidase subunit 1 (CO1)
whereas matK-based barcode for angiosperms (representing a majority of land plants)
and ribulose-bisphosphate carboxylase gene (rbcL) were chosen as the plant barcode.
Barcode of Life Database (BOLD) exhibits a platform for barcode-based classifica-
tion. The Medicinal Materials DNA Barcode Database (MMDBD) combines data
available on plants with their DNA barcodes making scopes for computational tools to
search and sequence comparisons [9]. A database “MedHerb” is developed to manage
information on genes, proteins, primers (used in Polymerase Chain Reaction (PCR))
and statistics related to some selected plant species with a scope of its expansion with
time (a more feature-based database incorporating expressed sequence tags, active
constituents, etc.) [13].
These diverse databases call for integration among multiple databases to link
traditional information with experimental data of current researches in the field
of medicinal plants. Currently, “omic” technologies consider genomics, transcrip-
tomics, proteomics, and metabolomics related to a biological sample to understand
complex systems underlying the whole process in a non-targeted and non-biased
manner.
3 Approaches of Computational Intelligence (CI)
Generally, herbal remedies take into account a combination of medicinal plants

which makes the systemization of these remedies difficult due to limited knowledge
about the synergistic effect of active ingredients. Moreover, other influencing factors
like availability, effective dose, price, etc. should be considered while documenting
efficacious treatments. Chemical fingerprinting and characterization of chemical
constituents can play a vital role in quality control to achieve an optimum combi-
nation of active metabolites for a specific biological activity. Lately, computational
methods are utilized to study the process of extraction, isolation, structure predic-
tion, metabolomics, biosynthesis, dereplication, phytochemical library construction,
and assessment of their bioactivity (Fig. 1). Multivariate data analysis is addressed
by reducing the dimensions of data and portraying the relationship between our
observations and calculated parameters (i.e., molecular descriptors) [14].
Fig. 1 The chain of interconnection between different sectors of pharmacognosy using bioinfor-
matics
3.1 CI in Extraction, Isolation, and Dereplication
The foremost goal in pharmacognosy is to extract the phytochemicals from plant

matrixes, later investigating their bioactivity and mechanism of action. The tradi-
tional way of trialing for optimization is tedious and often leads to erroneous
results. Computational and mathematical models are applied to correctly value the
parameters determining the quality of extraction. This ensures efficient optimiza-
tion of process parameters to get maximum yield of desired bioactive compounds.
Response Surface Methodology (RSM) is a popular model used to design experi-
ments for obtaining an optimum response curve governed by multiple independent
variables. Central Composite Design (CCD) or Box-Wilson design, Box–Behnken
Design (BBD), Doehlert design, three-level full factorial, fractional factorial, and
D-optimal are the commonly used RSM designs to identify and evaluate effects of
factors involved in a process [15]. In a research work of extracting polysaccharides
(ABP), CCD followed by multiple regression analysis was performed. A thorough
analysis of the obtained 3D response surface plot and contour plot predicted the
optimized extraction parameters (i.e., the independent variables namely extraction
temperature, the ratio of water to raw material, number of extractions, and extrac-
tion time) [16]. Another study implemented ultrasonic extraction using red grape
skins and combined BBD with high-performance liquid chromatography (HPLC)
for maximizing the yield of flavonoids [17].
Recent trends are exploiting machine learning and probabilistic neural networks to
classify and predict the occurrence of secondary metabolites leading to commercial
drug designing [18]. Virtual screening of databases can also be used to mine thera-
peutic agents but its potential is restricted due to certain factors like large phytochem-
ical data sets, variations in conformation of receptor molecules due rotatable bonds
and complexities in calculating binding affinities [19]. Virtual High Throughput
Screening (vHTS) can single out a probable drug from a large library of compounds
based on the binding affinity of target receptors to accelerate drug discovery [20].
3.2 CI in Structure Prediction and Bioactivity Assessment
Chemometrics relates multiple parameters of a chemical system and includes various

tools for basic statistics, factorial design, multivariate regression, signal processing,
hierarchical cluster analysis, calibration curve fitting (using Principal Component
Analysis (PCA), Partial least square regression (PLS), soft independent modeling
of class analogy (SIMCA), etc.), factor analysis, pattern recognition and neural
networks [16]. PCA offers dimensionality reduction of a data set to three subsets
of variables namely principal components, scores, and loadings revealing variations
among samples. Apart from the dominant usage of PCA in metabolic fingerprinting,
computational models successfully discriminated cannabis varieties by performing
PCA on the quantitative chemical data generated from gas chromatography flame
ionization of each variety [21]. These approaches proved to be cost-effective in

discovering medicinal plants and their associated phytochemical patterns. This has
led to relate phytochemistry and medicinal properties of herbs with phylogeny
[22]. In a work involving phylogenetic analyses of 32 taxa of Amaryllidaceae tribe
Galantheae, 6 taxa of other Eurasian genera of Amaryllidaceae and Phaedranassa
dubia as an outgroup, plants to treat Alzheimer’s disease were identified and reported
a correlation between AChE inhibitory activity and alkaloid profiles of the plants [23].
Density functional theory (DFT) is a computational quantum mechanical
modeling method greatly used for “ab-initio” calculations of molecular proper-
ties (like vibrational frequency, atomization energy), ionization potential, orbital
energy and reaction paths [14]. The dominant role of bond dissociation energy of
OH followed by hydrogen atom transfer (HAT) mechanism in antioxidant activity
of a molecule was highlighted in a study to establish the relationship between struc-
tural electronic properties and antioxidant potentials of chosen isoflavonoids and
isoflavonoid glycosides using DFT [24]. The use of DFT is also seen in identifying
hepatoprotective phytochemicals of Cucumis trigonus Roxb. and Cucumis sativus
Linn. by linking physicochemical properties with chemical reactivity and biological
activity [15, 25].
Toxicogenomics relates genomics with toxicology to investigate the molecular
mechanisms underlying toxicity and designs molecular expression patterns to predict
it. Statistical learning methods like Probabilistic Neural Network (PNN), k-Nearest
Neighbor (kNN), Support Vector Machine (SVM) and Decision Tree (DT) are
employed to evaluate the genotoxicity of chemical compounds [26]. Docking predicts
the orientation of a molecule (a candidate ligand) bounded to a receptor molecule
forming a stable complex. Quantitative structure-activity relationship (QSAR) and
docking studies were investigated on 75 compounds having a plant origin for
computer-aided drug design against the dengue virus. QSAR studies reported 11
inhibitory molecules against NS5 protein and 8 inhibitory molecules against NS3
protein whereas docking studies revealed 13 inhibitory molecules against NS5
and three inhibitory molecules against NS3, out of which best-fit molecules were
predicted [27]. Molecular docking using software ArgusLab was done to predict the
anticancer activity of polytriterpenes namely triterpenes boswellic acid and ursolic
acid [14]. In a study, three selected acetylcholinesterase-targeted flavonoids were
addressed to Lipinski filter, ADME/Tox screening, molecular docking, and QSAR
to predict the anti-Alzheimer’s activity of the chosen pyranoflavonoids and the result
was compared with in vitro analysis involving Ellman’s method of bioactivity staining
[28]. Apart from network-based analyses, the use of machine learning and fuzzy logic
in Quantitative Composition–Activity Relationship (QCAR) is an initiative in this
direction and seems to be more convenient than QSAR as, unlike QSAR it doesn’t
depend on the structural information of participating molecules.
In a study relating chemical composition and decreasing cholesterol effect of Qi-
Xue-Bing-Zhi-Fang, experimental data validated that techniques like multiple linear
regression, artificial neural networks (ANNs), and support vector regression showed
a difference in the predictive accuracy of the models [29]. Experimental data on
agrochemicals were analyzed by ANN using Kohonen based Self Organizing Map
(SOM) to study the accuracy of the model in predicting the agrochemical activity
of various plant growth regulators [30]. A research conducted on the occurrences
and oxidation patterns of 4700 flavonoids considered them as chemical markers
for Asteraceae and taken as the input of SOM to establish phylogenetic relation-
ships among the subfamilies and tribes of Asteraceae (i.e., for plant classification)
[21]. A similar study combined Random Forest (RF) with SOM to classify 8411
phytochemicals from 240 herbs of TCM among 10 classes namely aliphatics, alka-
loids, simple phenolics, lignans, quinones, polyphenols (flavonoids and tannins), and
mono-, sesqui-, di-, and triterpenes (including sterols) [22].
The proteins separated are identified by using mass spectrometry (MS) nowadays
for its high-throughput nature in large-scale protein identification than other protein
identification practices like Edman degradation microsequencing. However, mass
spectrometric data are complicated and identification of a protein from it mandates
computational analyses of data. Computer-Aided Structure Elucidation (CASE) is an
emerging technique that involves thorough data analysis of (Nuclear Magnetic Reso-
nance) NMR, Mass spectroscopy (MS), and Infrared Radiation (IR) to extract all the
possible structural information available from spectral data. Molecular docking simu-
lations (MDSs) prognosticate time-dependent behavior of a system at the molecular
level studying the structure, dynamics, and thermodynamics of biological molecules
and their complexes from X-ray crystallography and NMR data [14].
The avenues of bioinformatics in solving research problems and examples
discussed here reflect the necessity of computer-assisted simulations prior to the
design of experiments. Its scope in the field of pharmacognosy is still explored
encouraging the implementation of interdisciplinary approaches for a clear repre-
sentation of scattered, unstructured data into a comprehensible form.
4 Summary
The immense data generated since centuries in the field of pharmacognosy remain
scattered and thus limit the essence of strategic analysis of interlinked phenomena.
Combining ancient manuscripts of ethnobotany with modern researches and subse-
quent implementation of data mining techniques forms a basic platform for logical
representation of data. Multifold perspectives leading to an exponential growth of
data demand for procedures to efficiently manage, integrate, visualize, and analyze
data. Computational intelligence bridges the gap between diverse resources already
available (from the molecular level to macroscopic characteristics of drugs) marking
its significance in computer-aided drug design. Interdisciplinary strategies proved to
be promising in solving scientific questions and are still at its infancy allowing scope
for thorough researches in the recent future.
References
1. Strohl WR (2000) The role of natural products in a modern drug discovery program. Drug
Discov Today 5(2):39–41
2. Saito K, Matsuda F (2010) Metabolomics for functional genomics, systems biology, and
biotechnology. Annu Rev Plant Biol 61:463–489
3. Sarkar, IN (2009) Biodiversity informatics: the emergence of a field. S1
4. Cao C, Wang H, Sui Y (2004) Knowledge modeling and acquisition of traditional Chinese
herbal drugs and formulae from text. Artif Intell Med 32(1):3–13
5. Samwald M et al (2010) Integrating findings of traditional medicine with modern pharmaceu-
tical research: the potential role of linked open data. Chin Med 5(1):43
6. Chen H et al (2007) Towards semantic e-science for traditional Chinese medicine. BMC Bioinf
8(3):S6
7. Rhee SY, Dickerson J, Xu D (2006) Bioinformatics and its applications in plant biology. Annu
Rev Plant Biol 57:335–360
8. Loub WD et al (1985) NAPRALERT: computer handling of natural product research data. J
Chem Inf Comput Sci 25(2):99–103
9. Sharma V, Neil Sarkar IN (2012) Bioinformatics opportunities for identification and study of
medicinal plants. Brief Bioinf 14(2):238–250
10. Zeng S et al (2010) Development of a EST dataset and characterization of EST-SSRs in a
traditional Chinese medicinal plant, Epimedium sagittatum (Sieb. Et Zucc.) Maxim. BMC
Genom 11(1):94
11. Chen S et al (2011) 454 EST analysis detects genes putatively involved in ginsenoside
biosynthesis in Panax ginseng. Plant Cell Rep 30(9):1593
12. Mochida K et al (2011) Global landscape of a co-expressed gene network in barley and its
application to gene discovery in Triticeae crops. Plant Cell Physiol 52(5):785–803
13. Rajoka MI et al (2014) Medherb: an interactive bioinformatics database and analysis resource
for medicinally important herbs. Curr Bioinform 9:23–27
14. Mukherjee PK (ed) (2015) Evidence-based validation of herbal medicine. Elsevier
15. Sarker SD, Nahar L (2018) An introduction to computational phytochemistry. Computational
Phytochemistry. Elsevier, pp 1–41
16. Liu J et al (2009) Optimization of polysaccharides (ABP) extraction from the fruiting bodies
of Agaricus blazei Murill using response surface methodology (RSM). Carbohydr Polym
78(4):704–709
17. Tomaz I et al (2016) Multi-response optimisation of ultrasound-assisted extraction for recovery
of flavonoids from red grape skins using response surface methodology. Phytochem Anal
27(1):13–22
18. Schwabe T et al (2005) Neural networks for secondary metabolites prediction in Artemisia
genus (Asteraceae). Internet Electron J Mol Des 4:9–16
19. Shoichet BK (2004) Virtual screening of chemical libraries. Nature 432(7019):862
20. Subramaniam S, Mehrotra M, Gupta D (2008) Virtual high throughput screening (vHTS)—A
perspective. Bioinformation 3(1):14
21. Fischedick JT et al (2010) Metabolic fingerprinting of Cannabis sativa L., cannabinoids and
terpenoids for chemotaxonomic and drug standardization purposes. Phytochemistry 71(17–
18):2058–2073
22. Paton AJ et al (2004) Phylogeny and evolution of basils and allies (Ocimeae, Labiatae) based
on three plastid DNA regions. Mol Phylogene Evol 31(1):277–299
23. Larsen MM et al (2010) Using a phylogenetic approach to selection of target plants in drug
discovery of acetylcholinesterase inhibiting alkaloids in Amaryllidaceae tribe Galantheae.
Biochem Syst Ecol 38(5):1026–1034
24. Do Minh T, Nguyen Van T (2019) Isoflavones and isoflavone glycosides: structural-electronic
properties and antioxidant relations—A case of DFT study. J Chem
25. Gopalakrishnan SB, Kalaiarasi T, Subramanian R (2014) Comparative DFT study of phyto-
chemical constituents of the fruits of Cucumis trigonus Roxb. and Cucumis sativus Linn. J
Comput Methods Phys
26. Li H et al (2005) Prediction of genotoxicity of chemical compounds by statistical learning
methods. Chem Res Toxicol 18(6):1071–1080
27. Mahendran R (2016) In silico QSAR and molecular docking studies of selected medicinal plant
compounds against ns5 & ns3 protein of dengue virus: a comparative approach. Int J Pharma
Bio Sci, Sci
28. Das S et al (2017) Prediction of Anti-Alzheimer’s activity of flavonoids targeting acetyl-
cholinesterase in silico. Phytochem Anal 28(4):324–331
29. Wang Y, Wang X, Cheng Y (2006) A computational approach to botanical drug design by
modeling quantitative composition–activity relationship. Chem Biol Drug Des 68(3):166–172
30. Bushkov NA et al (2016) Computational insight into the chemical space of plant growth
regulators. Phytochemistry 122:254–264
Broadband Annular Ring Patch Antenna
Thingbaijam Rajkumari Chanu, Somokanta Hijam, and Maxon Okramcha
Abstract This paper investigates an annular ring microstrip patch antenna with
improved bandwidth from 2.61 to 38.15% at the range of 5.95–8.43 GHz by using
pin-sorted using the same feeding point in each design. The obtained frequency range
lies within the C-band.
Keywords Annular ring patch antenna · Bandwidth · Gain · Antenna efficiency ·

Impedance
1 Introduction
Microstrip Patch Antenna (MPA) is a coaxial probe-fed antenna that offers excel-
lent segregation between radiating elements and the feed network, thus yielding a
worthy front to back ratio [1]. The major advantage of this MPA is its lightweight,
thin outlining, simplicity to fabricate, accordant to mounting surfaces, low cost and
capable of being assimilated in active devices [1–3]. Due to these beneficial aspects
the MPAs have versatile applications such as in space technology, satellite broad-
casting, tracking systems, missiles, GPS monitoring systems, remote sensing and
also in aircrafts [2, 3]. A small size circular MPA compared to a rectangular one
conveys similar radiation pattern characteristics, thus indicating design compactness
[2]. But, this circular MPA has a major drawback due to its narrow bandwidth in the
range of 1–5%, which can be overcome by cutting the slots (the mechanism used to
load the antenna) in various shapes. Al-Zoubi et al. [4] observed a 12.8% increase in
bandwidth for circular MPAs when coupled with annular ring structures while band-
width enhancement of 34% detected by Chang and Lien [5] using a stacked structured
annular ring patch antenna. Another alternative, for improvement of bandwidth, is
pin-shorting and is probably the modest technique to expand bandwidth without
implementing any structural change in the patch antenna.
T. R. Chanu (B) · S. Hijam · M. Okramcha

Manipur Technical University, Imphal, India
22 T. R. Chanu et al.
In this study, the properties of circular MPA and Annular Ring structure MPA are
presented and compared. The designs have been simulated using the electromagnetic
simulator, Zealand IE3D software.
2 Antenna Design
The conventional circular MPA is considered the reference antenna to compare the
results that obtained from the proposed annular ring structure MPA. The geometry
of the conventional circular MPA is shown in Fig. 1. The patch has the dimension
of radius = 20 mm and is printed on FR4 of dielectric constant, εr = 4.4 and the
thickness of the substrate, h = 1.6 mm. A coaxial probe is used to connect the
microstrip patch at coordinates and it is made fixed for both the conventional and the
proposed MPA. The coordinate of the feeding point is (8, 0).
The geometry of the proposed to extend the bandwidth probe-fed patch antenna
with embedding slots and pi-shorted is shown in Fig. 2. Impedance bandwidth of
38.15% can be obtained from the modified geometry. Its main advantage of this
structure is that it produces wider bandwidth than the conventional circular patch
with a single and simple topology. The feeding point and the shorted-pin coordinates
are (−9.6, −8.5) and (9.6, 8.2), respectively.
Fig. 1 Conventional CPA

Broadband Annular Ring Patch Antenna 23
Fig. 2 Pin-shorted annular

ring MPA
3 Simulated Results
3.1 Return Loss (S11 ) and Impedance Bandwidth
The temp the Return Loss (S 11 ) shown in Fig. 3 of the conventional CPA is −21.46 dB
Fig. 3 Return loss (S 11 ) for

conventional CPA
Fig. 4 Return loss (S 11 ) for

pin-shorted annular ring
MPA
at Resonating frequency at 2.03 GHz and the bandwidth obtained is 2.61%. For Pin-
Shorted Annular Ring MPA the Return Loss (S 11 ) is −29.24 dB at 6.5 GHz and the
bandwidth obtained is 38.15% which is shown in Fig. 4.
The BW of a patch antenna is calculated from the equation
BW = ( f h − f l ) ∗ 100/ f r
where
f h Higher Frequency at −10 dB Return Loss
f l Lower Frequency at −10bB Return Loss
f r Resonating Frequency of the antenna at the lowest Return Loss.
3.2 Radiation Pattern
The microstrip antenna radiates normal to its patch surface. So, the elevation pattern
for ϕ = 0 and ϕ = 90 degrees are important for the measurement. The simulated
E-plane and H-plane pattern, 2D pattern view the conventional CPA and the proposed
pin-shorted annular ring MPA are illustrated in Figs. 5 and 6, respectively.
3.3 Other Parameters
Positioning The Comparative study of a conventional CPA and Annular Ring MPA
with and without pin-short after the simulation is given in Table 1.
Broadband Annular Ring Patch Antenna 25
Fig. 5 Radiation pattern for

conventional CPA
Fig. 6 Radiation pattern for

pin-shorted annular ring
MPA
Table 1 Comparative results

S. No. Parameter Conventional CPA Annular ring without Annular ring with
pin-short pin-short
1. Gain (dBi) −1.57 6.12 3.1
2. Antenna efficiency 18.44 39.74 38.15
(%)
3. Impedance () 55.3 + 8.5i 49.66–0.59i 47.35 + 2.05i
4 Conclusion
The Pin-Shorted Annular Patch MPA is observed to have acquired the highest band-
width and also the bandwidth has increased by 14.6 times than the conventional
CPA. Considerable enhancement in gain from −1.57 to 3.1 dBi is detected, while
the impedance of both nearly matched. Due to the insertion of a circular slot and pin-
shorting technique, most of the parameters improved including a drastic enrichment
observed in the Return Loss (S 11 ). These factors led to a change in current density
on the circular patch and the path was also increased, hence the resonating frequency
of the antenna changes from 2.03 to 6.5 GHz.
Thus, on modifying the conventional CPA the proposed circular annular ring patch
antenna provided superior bandwidth of 38.15%. Also, the antenna was detected to
remain stable over the entire bandwidth and finally, the most beneficial part is this
antenna can be used for C-band applications.
References
1. Surmeli K, Turetkin B. U-slot stacked patched antenna using high and low dielectric constant
material combinations in S-Band. Antenna Test and Research Center (ATAM)
2. Garg R, Bhartia P, Bahl I, Ittipiboon A. Microstrip Antenna Design Handbook. Artech House
Publication, Boston London
3. Balanis CA. Antenna theory analysis and design, 3rd edn. Wiley
4. Al-Zoubi A, Yang F, Kishik A (2009) A broadband center-fed circular patch ring antenna with
a monopole like radiation pattern. IEEE Trans Antennas Propog 57(3)
5. Chang D-C, Lien H-C (2004) The study of wideband single fed cp ring antennas. In: 4th Internal
conference on microwave and millimeter wave technology proceedings
Cooperative Spectrum Sensing
with Improved Absolute Value
Cumulation Detection Based on Additive
Laplacian Noise in Cognitive Radio
Khushboo Sinha and Y. N. Trivedi
Abstract In this paper, the performance of cooperative spectrum sensing (CSS)

with absolute value cumulation detection (AVCD) based improved energy detector
(IED) in additive Laplacian noise environment is proposed. Each cognitive radio
(CR) uses improved AVCD (i-AVCD) as a test statistic. Decision from each CR
is then forwarded to the fusion center (FC) where the final decision on the pres-
ence or absence of primary user (PU) is made. Performance of CSS in the form
of receiver operating characteristic (ROC) and the total error probability (Pe ) is
discussed. Besides this, the optimum value of p, optimum number of CRs (n) out of
K CRs, and optimum number of K to achieve a specified target total error probability
(sum of false alarm probability and missed detection probability is also obtained.
Results are presented here using Monte Carlo simulations and it is concluded that by
carefully choosing the optimum value of p, performance can be improved in case of
CSS with OR rule while the same is not necessarily applicable in case of CSS with
AND and majority rule (n/K ) rule.
Keywords Spectrum sensing · Cooperative spectrum sensing · Cognitive radio ·

Receiver operating characteristic · Laplacian noise
1 Introduction
Cognitive radio is one of the most promising techniques for the efficient utilization
of the available radio frequency spectrum. According to the FCC report [1], a large
portion of radio frequency (RF) spectrum remains under-utilized. Motivated by it,
the concept of cognitive radio (CR) was first proposed in [2] as a possible solution
to the under-utilization of the limited RF spectrum. CR exploits the under-utilized
frequency bands by identifying spectrum holes. Spectrum sensing is one of the most
K. Sinha · Y. N. Trivedi (B)

Department of ECE, Nirma University, Ahmedabad, Gujarat, India
K. Sinha
28 K. Sinha and Y. N. Trivedi
important functions of CR system where unlicensed user or secondary user (SU)

at the CR terminal continuously monitors the spectrum to detect the presence or
absence of licensed user or primary user [2]. Most spectrum sensing algorithms such
as energy detection (ED) and improved energy detection (IED) [3], the eigenvalue
based spectrum sensing [4], the goodness-of-fit based spectrum sensing [5] assumes
Gaussian noise (AWGN) into consideration. In real scenarios, the noise may be non-
Gaussian. Possible impairments which result in non-Gaussian background noise
include artificial impulsive noise, co-channel interference from other PUs, emission
from microwave ovens etc. [6]. Further, the previously mentioned detection schemes
degrade heavily in non-Gaussian noise.
It has been shown in [6, 7] that Gaussian mixture model, Middleton class A noise
model, and Laplace distribution (Laplacian noise) are used for approximately charac-
terizing the statistics of multiple access interference (MAI) in TH-UWB communica-
tion system. It has been shown that Laplacian noise accurately model MAI than other
noise models [7]. Motivated by it, Laplacian noise has been addressed by advanced
test-statistics such as suprathreshold stochastic resonance (SSR) [8], polarity coinci-
dence array (PCA) [9], non-linear function Kernel function [10], soft limiting PCA
(SL-PCA) [11], AVCD [12], i-AVCD [13] among others. AVCD is based on fixed
fractional lower order statistics (FLOS) while i-AVCD is based on a flexible FLOS
strategy with exponent p ranging from 0 to 2 [13].
In this paper, i-AVCD based on the CSS scheme is proposed as a test statistic to
improve the detection performance in additive Laplacian noise. Performance of CSS
is shown for AND, OR, and majority (n/K ) rule. Optimum value of p for all three
CSS schemes (AND, OR, majority rule) is derived using simulations. Optimum
number of CRs out of a specified number of CRs is also expressed analytically.
Performance of CSS schemes for AVCD, i-AVCD, SL-PCA, SSR is compared and
it is shown here that by choosing optimum p for CSS based i-AVCD test-statistics,
performance can be improved greatly in case of CSS-OR rule but the same is not
necessarily true in case of AND (CSS-AND) rule and majority (CSS-n/K) rule.
2 System Model
Let y = [y1 , y2 , . . . , y N ] be a vector of N observations of a PU received at the

cognitive terminal where N ≥ 1. It is assumed that all the received observations are
real, independent, and identically distributed (i.i.d.). Each received observations yi
can be represented as
H0 : yi = wi
√
H1 : yi = γ hxi + wi , (1)
where h is the channel coefficient and it is assumed to be constant (h = 1), i = 1, 2,

…, N, γ denotes the average signal-to-noise ratio (SNR), ni is Laplacian noise which
Cooperative Spectrum Sensing with Improved … 29
follows Laplace distribution with mean 0 and variance 2b2 i.e., wi ~ L(0, b), b is the
scale parameter of Laplace distribution, x i is BPSK PU signal, i.e., xi ∈ {−1, 1} p is
the exponent (power) of received samples at the cognitive terminal (SU).
2.1 Decision Statistics
i-AVCD based decision rule can be expressed as

N
Z= |yi | p . (2)
i=1
Decision statistics of i-AVCD can be expressed as
H0 : Z < λ, PU absent,
H1 : Z ≥ λ, PU present,
where γ is the detection threshold of AVCD. It is should be noted that p = 1 denotes

AVCD. Here, CSS scheme is applied where multiple CRs are present in a cognitive
radio system. Each CR makes independent binary decisions in the form of 0 or
1. Then each CR forwards their independent binary decisions (hard decisions) to
the fusion center (FC). FC then combines the received decisions to determine the
presence or absence of PU as shown below:

K
S= si , (3)
i=1
where K is the total number of CRs present, S is the sum of independent binary
decisions from each CR. si is the hard decision made by each CR based on i-AVCD
test-statistics. The hypotheses H 0 (null hypothesis) and H 1 (alternative hypothesis)
at the FC using the CSS scheme can be expressed as
H0 : S < n, PU absent,
H1 : S ≥ n, PU present
where n < K denotes majority rule, n = 1 denotes OR rule and n = K denotes

AND rule.
3 Detection Probability and False Alarm Probability
Probability density function (pdf) of received samples yi at each CR under hypotheses

H0 and H1 can be expressed as
1 − |x|
f yi |H0 (x) = e b (4)
2b
1 − |x−√γ | |x+ γ |
√
f yi |H1 (x) = e b + e− b . (5)

4b
Using (4) f |yi ||H0 (x) can be expressed as [13]
1 −|b−1 x |
f |yi ||H0 (x) = e ,
b
= 0, x < 0. (6)

Mean E |yi | p |H0 and variance D |yi | p |H0 can be expressed as [13]

E |yi | p |H0 = b p Γ ( p + 1) (7)

D |yi | p |H0 = b2 p Γ (2 p + 1) − b2 p Γ 2 ( p + 1). (8)
False alarm probability (PF ) at each CR can be expressed as

⎛ ⎞
λ − N × E |yi | p |H0
PF = Q ⎝
⎠
. (9)
N × D |yi | p |H0
Similarly, detection probability (PD ) at each CR can be expressed as

⎛ ⎞
λ − N × E |yi | p |H1
PD = Q ⎝
⎠
. (10)
N × D |yi | p |H1
Similarly, expressions of E[|yi | p |H1 ] and D[|yi | p |H1 ] can be derived [13]
similarly as done in case of PF .
Finally, at FC, detection probability (Q D ) and false alarm probability (Q F ) can
be expressed as [14]
K

K
QF = PFl (1 − PF ) K −l (11)
l
l=n
K

K
QD = PDl (1 − PD ) K −l (12)
l
l=n
Total error probability (Pe ) at the FC can be expressed as
Pe = Q m + Q F , (13)
where Q m = 1 − Q D is the missed detection probability at the FC.
3.1 Optimum Values of p, N, and K
Optimum values of p( p ∗ ) at each CR can be obtained by partially differentiating PD

w.r.t. p and equating it to zero.
p ∗ = arg max(PD ) (14)

p
Similarly, optimum values of n can be obtained as

∗ K
n = arg min(Pe ) = min K , (15)
n 1+α

PF
ln
where α = and . denotes ceiling function.
1−Pm

Pm
ln 1−PF
Thus, for specified values of K, and Pm , n ∗ can be easily found out.
4 Simulation Results
Figure 1 shows PD versus SNR (in dB) comparison of i-AVCD with ED and AVCD
using CSS-n/K (majority) rule at SNR of −4 dB, N = 30 and PF = 0.1. It can
be seen clearly that at low values of SNR, i-AVCD with n/K rule underperform the
test-statistics without n/K rule. However, as SNR increases, n/K rule improves the
performance of i-AVCD. Figure 2 shows PD versus SNR comparison of i-AVCD
using CSS-AND rule. It is clear from figure that AND rule of CSS doesn’t perform
better than that without AND rule in case of the test-statistics ED, AVCD and i-
AVCD. Figure 3 shows PD versus SNR comparison of i-AVCD using CSS-OR rule.
It is clear from figure that OR rule of CSS improves the performance of ED, AVCD
and i-AVCD significantly. Figure 4 shows optimum p of i-AVCD with CSS-OR rule
for different PF . It can be clearly seen from the figure that optimum p( p ∗ ) increases
with decrease in PF . Values p ∗ at different values of PF and at γ = −4 dB are given
Fig. 1 Performance comparison of i-AVCD, AVCD and ED with CSS (n/K rule) for n = 2, K =
3 at γ = −4 dB, N = 30
Fig. 2 Performance comparison of i-AVCD, AVCD and ED with CSS (AND rule) for n = 3, K =
3 at γ = −4 dB, N = 30
in Table 1. Figure 5 shows the optimum values of K comparison to achieve a certain

target error probability for different PF . From the simulation result, it is found that to
achieve total error probability of 0.001739, optimum values of CR are 30 out of total
60 taken CRs. Similarly, to achieve an error target of 0.02311, only 15 CRs will be
sufficient out of 60 CRs. Figure 6 shows performance comparison of i-AVCD with
SL-PCA, SSR, ED, AVCD using CSS-OR rule at PF = 0.1 and N = 30. i-AVCD
at p ∗ = 0.101 shows better performance than all the mentioned test-statistics at
optimum p. However, as p increases ( p = 1.5), performance of i-AVCD decreases.
Fig. 3 Performance comparison of i-AVCD, AVCD and ED with CSS (OR rule) for n = 1, K = 3
at γ = −4 dB, N = 30
Fig. 4 PD versus p at γ = −4 dB and N = 30 at different values of PF showing optimum p
Table 1 Optimum values of p of i-AVCD with CSS at N = 30 and γ = −4 dB

Optimum p(p*)
PF i-AVCD i-AVCD with CSS-OR i-AVCD with CSS-AND i-AVCD with CSS-n/K
rule rule rule
0.1 0.101 0.101 0.101 0.101
0.05 0.151 0.151 0.101 0.201
0.01 0.251 0.251 0.101 0.301
*denotes optimum value
Fig. 5 Pe versus K at γ = −4 dB and N = 30 with i-AVCD at p = 1 for different detection

threshold (λ)
Fig. 6 Performance comparison of i-AVCD with other test-statistics using CSS scheme based on
OR rule at N = 30 and PF = 0.1.
Figure 7 shows receiver operating characteristic (ROC) plot of i-AVCD, AVCD, and
ED at N = 50 and γ = −4 dB at different values of p. ROC improves as p decreases.
Further, both simulations and analytical results closely match for the three compared
test-statistics in the figure.
Fig. 7 ROC comparison of i-AVCD with other test-statistics using CSS scheme based on OR rule
at γ = −4 dB, N = 50
5 Conclusion
Cooperative spectrum sensing with i-AVCD is discussed using different CSS rules:
AND, OR, and majority (n/K) rule. It is concluded that optimum p increases with a
decrease in PF . Further, it is also concluded that OR rule shows better performance
over a wide SNR range as compared to AND rule and majority rule. n/K rule also
works well but shows improved performance over only a limited SNR range. i-AVCD
is compared with ED, AVCD, SSR, SL-PCA and it is concluded that by carefully
selecting optimum p, detection performance increases over a wide SNR range, as
compared with other test-statistics in case of CSS-OR rule.
References
1. Federal Communications Communication (2004) Notice of proposed rule making. Unlicensed

operation in the TV broadcast band. ET Docket No. 04–186 (FCC 04 113)
2. Mitola J, Maguire GQ (1999) Cognitive radio: making software radio more personal. IEEE
Pers. Commun. 6(4):13–18
3. Chen Y (2010) Improved energy detector for random signals in Gaussian noise. IEEE Trans
Wirel Commun 9(2):558–563
4. Liu C, Wang J, Liu X, Liang Y (2019) Maximum eigen value-based goodness-of-fit detection
for spectrum sensing in cognitive radio. IEEE Trans Veh Technol 68(8):7747–7760
5. Trivedi YN, Patel DK (2015) Goodness-of-fit-based non-parametric sensing under Middleton
noise for cognitive radio. Electron Lett 51(5):419–421
6. Shao M, Nikias CL (1993) Signal processing with fractional lower order moments: stable
processes and their applications. Proc IEEE 81(7):986–1010
7. Hu B, Beaulieu NC (2008) On characterizing multiple access interference in TH-UWB systems
with impulsive noise models. In: IEEE radio and wireless symposium, pp 879–882
8. Li Q, Li Z (2014) A novel sequential spectrum sensing method in cognitive radio using

suprathreshold stochastic resonance. IEEE Trans Veh Technol 63(4):1717–1725
9. Wimalajeewa T, Varshney PK (2011) Polarity-coincidence-array based spectrum sensing for
multiple antenna cognitive radios in the presence of non-Gaussain noise. IEEE Trans Wirel
Commun 10(7):2362–2371
10. Margoosian A, Member S, Abouei J, Member S, Konstantinos N (2015) An Accurate kernelized
energy detection in gaussian and non-Gaussian/impulsive noises. IEEE Trans Signal Process
63(21):5621–5636
11. Karimzadeh M, Rabiei AM, Olfat A (2017) Soft-limited polarity coincidence array spectrum
sensing in the presence of non-Gaussain noise. IEEE Trans Veh Technol 66(2):1418–1427
12. Ye Y, Li G, Zhou F, Zhang H (2018) Performance of spectrum sensing based on absolute value
cumulation in Laplacian noise. In: IEEE vehicular technology conference, pp 1–5
13. Ye Y, Li Y, Zhou F (2019) Improved energy detection with Laplacian noise in cognitive radio.
IEEE Syst J 13(1):18–29
14. Zhang W, Mallik RK, Lataief KB (2009) Optimization of cooperative spectrum sensing with
energy detection in cognitive radio networks. IEEE Trans Wirel Commun 8(12):5761–5766
Solar Power Plant Site Selection:
A Systematic Literature Review
on MCDM Techniques Used
Rajkumari Malemnganbi and Benjamin A. Shimray
Abstract Site Selection is a crucial step in installing Solar Power Plant (SPP) as it is
determined by a set of quantitative and qualitative factors, which are vague in nature.
In this review, various suggestions for site location of Photovoltaic Power System
(PVPS) are studied. The solar power plants are mainly installed in remote regions
where solar radiation is high. But these regions are far from the generation site and will
face problems in transmission and distribution. Thus, various criteria/factors have
to be considered in the site selection. Factors like solar radiation, location, climate,
orography, environment, public acceptance have to be taken into account for optimal
site selection. Every form of electricity generation has its strengths and weakness. The
main objective in the site selection process is finding the optimum site satisfying the
desired conditions given by the selection criteria. This review suggests how to define
and classify particular criteria considered for Solar Photo Voltaic (PV) farm siting.
The best criteria among all the other evaluations are obtained from the multi-criteria
evaluation which gives scope for various applications. The aim of this paper is to
define how the ideal locations for solar PV are selected using various Multi-Criteria
Decision Making (MCDM) techniques. A large scale PV-project should generate at
least 5 MW power. In site suitability, India ranks third next to China and Spain.
Keywords Solar Power Plant (SPP) · Photovoltaic Power System (PVPS) ·

Multi-criteria Decision-Making (MCDM) · Geographical Information System
(GIS) · Site selection · Analytical Hierarchy Process (AHP)
R. Malemnganbi · B. A. Shimray (B)

Department of Electrical Engineering, National Institute of Technology Manipur,
Imphal, India
R. Malemnganbi
38 R. Malemnganbi and B. A. Shimray
1 Introduction
In developing countries like India, the increasing need for electricity has become a
major issue. At present many countries regarded fossil as the main source of energy.
Because of its massive use, fossil fuels like natural gas, coal, and petroleum are
getting depleted. Also, the burning of fossil fuels causes “Green House Effect” or
“Global warming”. Thus, this issue and the rising demand for electricity leads to
global effort. To meet the continuous requirement of electricity, we need to focus on
other alternative renewable sources of RES. In many developed countries, we see
that the use of renewable energy power supplies has decreased the global warming
issue. Solar energy is one of the alternatives to a Non-Renewable source of energy
that generates electric power from solar energy. Solar energy is found abundantly and
is pollution-free. Recently solar energy plays have been playing an important role in
the generation of electricity. In spite of all the advantages, as solar energy is uncertain
the output characteristics of PV are non-linear and thus become complicated when
the entire PV cell does not get uniform radiation and leading to more than one peaks.
India is a developing country and the need for energy has been increasing day by
day. Since our country is developing, the demand will rise by approximately 10% on
an installed base of 200 GW. We need to find other sources of electricity generation
and renewable sources of energy which can fill this gap. Renewable sources of energy
such as wind, solar, hydro, biomass, etc. play an important role to overcome the
shortfall of energy. We need not set up solar PV in large utility-scale. Installing solar
PV on the rooftop is amenable. By considering the solar resource data in the country,
we can install solar PV on our commercial and residential rooftops for electricity
production on a distributed basis. The energy generated from solar PV serves as the
power source for homes and offices. It can also be used for electrification in off-
grid sectors for rural India. As soon as this kind of project is set up, one can start
distributing to other countries. There are entrepreneurs working on such models.
India is considered one of the best suitable sites as renewable sources like sunlight;
wind, hydro, and biomass are found in abundance. People have also become aware of
the advantages of renewable energy. India is one of the countries that has the world’s
largest scheme for renewable energy. The challenge for installing solar PVPP is the
probable site selection for the annual production of electricity with financial viability.
The solar energy that requires massive areas of utilization, serves as one of the
best sources of energy with low-density power supply among all renewable sources.
Moreover, PV technology has the potential to deploy vast areas of land where the
scope of global radiation is very high per year. Though solar power development has
emerged as one of the leading renewable options for meeting demands in developing
countries like India and China, one of the hindrances is its variation and unevenness
of solar radiation which differs from site to site. Hence installing a PV project requires
an in-depth analysis of the sites where the plant is to be located. Thus, a question
arises, how one can select the optimal site for utility-size Solar PV? An extensive
analysis is a foremost step towards an economical and well-performed solar project.
Solar Power Plant Site Selection: A Systematic … 39
Therefore keeping a keen knowledge of the probable sites is a tactical breakthrough

for the prediction of the annual power plant output and economic feasibility.
Geographical Information system (GIS) is software used to visualize, analyze,
create, manage edit, and store the geographic data. Integrating GIS with various
decision support systems will benefit to determine the ideal location for the solar
power plant. Subsequently refining the performance of the solar PV project plays a
vibrant role in maximizing the output power generated which will help in reducing
the cost. Solar irradiation is considered one of the most important criteria while
temperature plays a vital role in the PV performance. Locating the plant nearby
ample consumer is one key factor that should be taken into account. From a location
point of view, India ranks third next to China and Spain.
Shimray et al. [1] proposed an MLP neural network trained by the Genetic Algo-
rithm for site selection of large Hydro Power Plant. In this paper, the author selected
19 criteria for analysis as a case study for site locating in Manipur, Arunachal Pradesh,
Sikkim, and Himachal Pradesh. A paper on MLP-GA integrated with fuzzy is also
presented for the four sites mentioned above. Criteria such as hostility, water quality,
air quality, environmental conditions are considered [2].
2 Techniques Used in Solar Power Plant Site Selection
Though it is well-known that considering various factors in the decision criteria can
enhance site selection, using the MCDM technique can ease site selection for an
optimal power Plant. The various methods used may vary in the decision maker’s
goal and the data required and their respective characteristics. Table 1 gives detailed
studies about different techniques used in site selection of solar PV plants. Various
techniques used by different researchers are also described in the following section.
Hong Fang proposed TOPSIS based on variable precision rough set theory and
prospect theory for selecting a suitable solar power Plant site. In this paper, a case
study is also portrayed for four sites. Here four experts provide the crisp importance
of the criteria chosen such as solar radiation, location, orography, policy support,
environmental influence, etc. The weights of the criteria and the ranking order of the
four sites are obtained by applying TOPSIS [33].
AHP is a decision an estimation method used for pairwise comparisons of spec-
ified criteria. Saaty [34] developed this method to measure vague factors relating
to uncertainties. The comparative scale judgment helps in prioritizing an element
over the other by taking the criteria into account. Thus, to make this choice, and to
create prioritizing the following steps are carried out. The first step is to define the
objectives and problems. Second, the decision hierarchy is created from the top, till
the lowest level. The next step is constructing a pairwise comparison matrix. Now
for every component, the priorities are used for every comparison done. Lastly, the
weighted values for each comparison level are added and the overall priority is taken.
The process of adding and weighing is done until the last level. Various application
of AHP is studied. There are various ways to apply AHP. Kohara and Sugiyama
Table 1 Methods for site Selection

S. No. Methods applied Types of RES Plant location Grid References
1 AHP PV South, On [3]
Morocco
2 AHP-fuzzy PV India On [4]
TOPSIS
3 Binary PSO PV Jaen, Spain On [5]
4 Grey cumulative PV Northwest, On [6]
China
5 WLC PV Southwest, On [7]
USA
6 Simulation PV India On [8]
scenarios
7 AHP PV Ismailia, On [9]
Egypt
8 GIS-join PV-wind-biomass United States On [10]
9 GIS-overlay PV-hydro-bioenergy Cumbria, On [11]
United States
10 AHP Solar-wind China On [12]
11 GIS-NEAR Solar-wind-hydro diesel Nigeria On [13]
12 GIS-solar Solar farms Brazil None [14]
irradiation
13 GIS-multiply PV-wind-biomass Appalachian On [15]
Mountains,
United States
14 WLC PV EU-28 On [16]
15 GIS—Solar Solar-wind-biomass-hydro Salta, None [17]
irradiation Argentina
16 GIS-intersect PV Rajasthan, On [18]
India
17 AHP Solar farms Konya, On [19]
Turkey
18 GIS—Solar PV Andalusia, On [20]
irradiation Spain
19 Axiomatic design PV Turkey On [21]
20 GIS—Solar Solar-wind-biomass Kujawsko, On [22]
irradiation Poland
21 FAHP PV Iran On [23]
22 WLC PV, CSP-wind Afghanistan On [24]
]23 AHP PV Ismailia, On [25]
Egypt
24 GIS-overlay PV Tibet, China On [26]
(continued)
Table 1 (continued)
S. No. Methods applied Types of RES Plant location Grid References
25 FUZZY ANP and PV Taiwan Off [27]
VIKOR
26 AHP Solar-wind-geothermal Indonesia None [28]
27 WLC PV Southwest, On [29]
USA
28 NASA—Solar PV India None [30]
irradiation
29 GIS-join Solar farms United Arab On [31]
Emirates
30 AHP-fuzzy-WLC Solar farms Isfahan, Iran On [32]
(35) propose an AHP tool to establish countermeasure practices in the incident of a

tragedy. Moreover, AHP can be improved by using optimization methods to specify
an objective. For example, Sumabat et al. [36] presented the usage of AHP integrated
with multiple response surface analysis to design a tool with reference to customer
product property and preferences.
GIS-AHP application is one of the commonly used methods for incorporating
GIS with other decision methods. AHP is considered a flexible and robust method
that can solve complex decision problems [37].
Hassan Z. et al. used GIS-AHP application for weighing and computing the land
suitability index (LSI) and estimate the probable site in Saudi Arabia. GIS integrated
with MCDM techniques is one of the most used techniques that deal with rich topog-
raphy information data and is one of the best methods to introduce the best sites for
PVPP. Tahri et al. applied GIS integrated with AHP in evaluating solar farm sites. Jun
et al. evaluated seven locations of solar-wind integrated power stations using elimi-
nation and choice transmission reality (ELECTRE-II) and concluded that the result
had more improved exactness to the related research findings. Chang established a
goal programming model for selecting suitable sites for various types of RE facili-
ties. Ingudam et al. proposed an optimization process which provides electrification
to rural areas of a particular state [38, 39]. Maleki et al. develop a framework that
integrates GIS and artificial bee swarm optimization in determining the most prob-
able sites for PVPP. Though the above-mentioned methods help ease the investors
in selecting the most probable optimal locations in implementing the power plant
projects, they are not able to deal with the uncertainties and vague environment as
they use crisp values in signifying the unclear judgment of DM.
To manipulate the subjectivity of judgment and the vagueness, Fuzzy theory has
been in use recently. Fuzzy theory permits the DM to integrate imperfect and uniden-
tifiable information in the decision model. Lee et al. provided multi-criteria decision-
making approach that incorporates interpretive structural modeling, Fuzzy Analyt-
ical Network Process and VIKOR for selecting the PVPP locations. Zoghi et al.
optimized PV site selection using fuzzy Logic, MCDM process, and weighted linear
combination. Sanchez-Lazano proposed a fuzzy TOPSIS method in evaluating the

site for PVPP to measure DM imprecision to enhance the subjective evaluations in
selecting the Power plant site. In this context theory of Fuzzy set is appropriate as
well as effective. However, in fuzzy, membership function and data distribution are
predetermined, due to which it fails to determine the spontaneous changes of various
decision-makers’ suggestions. Rough Set Theory is also a method that uses rough
intervals to define the uncertainties. Rough set theory is more objective and flexible
than the Fuzzy set theory as prior information is not required. Because of its various
advantages, rough set theory is used in various applications for solving vagueness
like sustainable supplier selection, product service recommendation, etc.
3 Site Selection Criteria
While developing a utility-scale solar power plant, various factors or criteria have to
be taken care of in selecting the site location. Probable Site Selection of Photovoltaic
Power Plant (PVPP) is a complex MCDM process, as the required site has to be
climatically and geographically acceptable. It must also have the highest generation
potentials. Thus, decision-makers must have enough information to assess sites under
different criteria, such as sunshine duration, sunshine radiation, soil, and transporta-
tion. The decision criteria reviewed here are deprived based on the existing literature
and accessibility to the geo-referenced database. The factors considered are mostly
related to climate and location. Solar irradiation is considered one of the most impor-
tant decision criteria. Public acceptance is also one of the important factors that have
to be considered. Various criteria used are shown in Table 2.
A large amount of solar energy will help in the generation of more electricity
from the existing sources. Figure 1 shows the most used criteria that have been
used in the selected power plant. Installing PVPP near the substation and power
lines will enhance the accessibility to the grid. New transmission line need not be
established, thereby reducing the overall cost and the power loss in the distribution
and transmission line.
Air temperature is an important factor that determines the performance of the PV
system. The air surrounding the PV system will cool the PV arrays and inverters’
connectivity thereby improving the efficiency of the PV system. In order to avoid not
in my backyard (NIMBY) opposition and its negative impact on the environment of
urban areas, the most effective location for installing a Solar Power Plant is far away
from cities for the development of Renewable energy. Some site selection criteria
are reviewed under Table 2
From Fig. 1, it is known that solar radiation is one of the most important or
used criteria by various DM followed by distance to power lines, main roads, and
temperature.
Table 2 Criteria for site selection

Solar radiation It is the energy emitted by the sun. The more energy it is
emitted, higher is the potential for generation of power from
the PVPP
Sunshine hour It measures the duration of sunshine in a particular site in a
given time period. Longer the duration higher will be the
power generation
Temperature The Plant Site has to be of suitable temperature for the panel.
Higher the temperature, lower the efficiency
Wind speed Wind can cause damage to the panel from dust or matters. On
the contrary, the distribution of heat from the panel makes the
panel’s performance higher
Rainfall Sites with high rainfall can deteriorate the efficiency of the
solar panels
Distance to faults The site near the fault line has a higher risk of earthquake
thus resulting in damage of the plant
Distance to roads Proximity to roadways can reduce the operational cost
Distance to urban and rural areas Sites near rural and urban areas cause environmental
problems and have a negative impact on the future
development of the residential areas
Slope Higher slope results in high investment and operational cost
Elevation Higher areas have more potential than lowlands for solar
energy. On the Contrary transportation and construction of
the plant is difficult with the increase in altitude
Land availability A PVPP requires a wide range of areas
Solar panel types Monocrystalline cell is the most used cell because it gives the
highest efficiency
Public acceptance A site where the attitude of the public is positive is considered
Policy Relevant law has to be considered to construct a solar power
plant in that particular area
4 Unsuitable Sites and Restriction Factors
This stage is for eliminating the non-feasible sites that have barriers to the develop-
ment and installation of the PVPP. Table 3 reviews the unsuitable or restricted factors
for the solar power plant installation.
Figure 2 represents the most common restriction applied in the study of the most
suitable site of PVPP. From the figure and the table above, it can be predicted that
protected land, airport military zones, urban land, cultural heritage, higher slope
areas, are the most omitted areas. Areas with high slope are excluded for their low
financial viability. Railway network is also excluded because of traffic issues, more
buildings in addition to a high-density population (Table 4).
35
30
30
25
No of arƟcles
20 18
15 14
15 12
10
10 8
6 5
5 3
0
Sub criteria
Solar irradiaƟon Distance from substaƟon Slope Distance from main roads
Land use Temperature Protected land Coastal zones
Cummunity interest sites ElevaƟon
Fig. 1 Various selection factors for PVPP
Table 3 Restriction factors

S. No. Limitations References
1 The land aspect [5, 6]
2 Dams [28, 31]
3 Elevation [23, 32]
4 Roads and railroad network [20, 22, 32]
5 Airport and military zones [32]
6 Watercourses and streams [20, 22, 23]
7 Soft soil [14, 15]
8 Natural disaster areas [30, 37]
9 Cultural heritage [3, 9]
10 Visual impact [24, 35]
11 Mountains [26, 29]
12 Water infrastructure [17, 21]
13 Coastal zones [22, 23]
14 Sand areas and sand dunes [3, 33]
15 Community interest sites [8, 30]
16 Protected land [13, 17]
17 Biological significant areas [23, 28]
18 Urban lands [29, 31]
19 Paleontological sites [17, 21]
20 Cultivated and high landscape [17, 18, 29, 32]
25
21
20 19
17
No of arƟcles
15 15
15 14
10 8
7 7
0
Criteria
Protected land CulƟvated and high landscape area
Watercourses and stream Urband land
Higher slope areas Roads and railroad network
Wildlife areas Airport and military zones
Cultural heritage
Fig. 2 Restriction factors
Table 4 Criteria for the selection of solar power plants

Factors Sub-factors References
Location Distance from urban areas [4, 7, 10, 14, 16, 20, 30]
Distance from main roads
Distance from substations
Population density
Distance from rivers
Climatic condition Solar radiation [4, 8, 9, 11, 19, 21]
Temperature
Humidity
Clearness index
Annual dusty days
Environmental factors Land use [7, 8, 10, 12, 17, 19, 32]
Agrological capacity
Emission level
Geographical feature Slope [5, 8, 11, 12, 19, 23, 29]
Orientation
Elevation
Plot area
Economic condition Energy saving [3, 4, 7, 9, 11, 26, 29]
Land cost
Transmission losses
Cost of energy
Construction cost
Energy generation
Initial investment
Demand for electricity
5 Conclusions
This paper reviews an in-depth assessment of various studies for a utility-scale PV

power plant. It can be concluded that GIS-based MCDM tool continues to enhance
the output for a proficient Decision Support system. It eases the Decision Makers
and the developers to decide the probable and efficient sites to have a substantial
technical performance with the least cost and less impact on the environment. Solar
irradiation is considered to be the most important decision criteria. Land slope and
Nearness to power lines are also considered as one of the most important factors while
urban and protected land followed by watercourses and streams are considered the
most restricted factors. In the last few years, the grid PV system has also overtaken
the off-grid installation globally. The on-grid PV system utilization has gained favor
where massive areas are available and solar radiation amount is high. Furthermore,
in the context, about 85% of the surveys are concerned with on-grid PV. China is
considered to be the world’s largest solar energy producer with an installed capacity
of 174 GW while Spain stands second with an installed capacity of 7011 MW. In the
location point of view, India ranks third next to China and Spain.
6 Future Work
This review does not include textbook and papers that are not published and is limited
to only English journals. It can be improved by taking real projects into account.
References
1. Shimray BA, Manglem K, Mehta RK (2017) Ranking of sites for installation of hydro power
plant using MLP neural network trained with GA: a MADM approach. Comput Intell Neuro
Sci 3:1–8
2. Shimray BA, Manglem K, Mehta RK (2017) A new MLP-GA-Fuzzy decision support system
for hydro power plant selection. Res Art Comput Eng Comput Sci 43(12): 6823–6835
3. Tahri M, Hakdaoui M, Maanan M (2015) The evaluation of solar farm locations applying
geographic information system and multi-criteria decision-making methods: case study in
southern Morocco. Renew Sustain Energy Rev 51:1354–1363–1362
4. Sindhu S, Nehra V, Luthra S (2017) Investigation of feasibility study of solar farms deployment
using hybrid AHP-TOPSIS analysis: case study of India. Renew Sustain Energy Rev 73:496–
511
5. Gómez M, López A, Jurado F (2010) Optimal placement and sizing from standpoint of the
investor of photovoltaics grid-connected systems using binary particle Swarm optimization,
Appl Energy 87:1911–1918
6. Liu J, Xu F, Lin S (2017) Site selection of photovoltaic power plants in a value chain based on
grey cumulative prospect theory for sustainability: a case study in Northwest China. J Clean
Prod 148:386–397
7. Brewer J, Ames DP, Solan D, Lee R, Carlisle J (2015) Using GIS analytics and social preference
data to evaluate utility-scale solar power site suitability. Renew Energy 81:825–836
8. Jain A, Mehta R, Mittal SK (2011) Modeling impact of solar radiation on site selection for
solar PV power plants in India. Int J Green Energy 8:486–498
9. Effat HA et al (2013) Selection of potential sites for solar energy farms in Ismailia Governorate,
Egypt using SRTM and multi-criteria analysis. Int J Adv Remote Sens GIS 2:205–220 (Cloud
Public)
10. Niblick B, Landis AE (2016) Assessing renewable energy potential on United States marginal
and contaminated sites. Renew Sustain Energy Rev 60:489–497
11. Gormally AM, Whyatt JD, Timmis RJ, Pooley CG (2012) A regional-scale assessment of local
renewable energy resources in Cumbria, UK. Energy Policy 50:283–293
12. Yunna W, Geng S (2014) Multi-criteria decision making on selection of solar-wind hybrid
power station location: a case of China. Energy Convers Manag 81:527–533
13. Mentis D, Welsch M, Fuso Nerini F, Broad O, Howells M, Bazilian M, Rogner H (2015) A
GIS-based approach for electrification planning—A case study on Nigeria. Energy Sustain Dev
29:142–150
14. Martins FR, Pereira EB, Abreu SL (2007) Satellite-derived solar resource maps for Brazil
under SWERA project. Sol Energy 81:517–528
15. Arnette AN, Zobel CW (2011) Spatial analysis of renewable energy potential in the greater
southern Appalachian mountains. Renew Energy 36:2785–2798
16. Perpiña Castillo C, Batista e Silva F, Lavalle C (2016) An assessment of the regional potential
for solar power generation in EU-28. Energy Policy 88:86–99 (2016)
17. Belmonte S, Núñez V, Viramonte JG, Franco J (2009) Potential renewable energy resources of
the Lerma Valley, Salta, Argentina for its strategic territorial planning. Renew Sustain Energy
Rev 13:1475–1484
18. Khan G, Rathi S (2014) Optimal site selection for solar PV power plant in an Indian state using
geographical information system (GIS). Int J Emerg Eng Res Technol 2:260–266
19. Uyan M (2013) GIS-based solar farms site selection using analytic hierarchy process (AHP)
in Karapinar region, Konya/Turkey. Renew Sustain Energy Rev 28:11–17
20. Carrión JA, Estrella AE, Aznar Dols F, Ridao AR (2008) The electricity production capacity of
photovoltaic power plants and the selection of solar energy sites in Andalusia (Spain), Renew
Energy 33:545–552
21. Boran FE, Menlik T, Boran K (2010) Multi-criteria axiomatic design approach to evaluate sites
for grid-connected photovoltaic power plants: a case study in Turkey. Energy Source B Econ
Plan Policy 5:290–300
22. Sliz-Szkliniarz B (2013) Assessment of the renewable energy-mix and land use trade-off at
a regional level: a case study for the Kujawskoe Pomorskie Voivodship. Land Use Policy
35:257–270
23. Noorollahi E, Fadai D, Akbarpour Shirazi M, Ghodsipour S (2010) Land suitability analysis
for solar farms exploitation using GIS and fuzzy analytic hierarchy process (FAHP)—A case
study of Iran. Energies 9:643
24. Anwarzai MA, Nagasaka K (2017) Utility-scale implementable potential of wind and solar
energies for Afghanistan using GIS multi-criteria decision analysis. Renew Sustain Energy
Rev 71:150–160
25. Effat HA (2013) Selection of potential sites for solar energy farms in Ismailia Governorate,
Egypt using SRTM and multi-criteria analysis. Int J Adv Remote Sens GIS 2: 205–220 (Cloud
Public)
26. Wang S, Zhang L, Fu D, Lu X, Wu T, Tong Q (2016) Selecting photovoltaic generation sites
in Tibet using remote sensing and geographic analysis. Sol Energy 13385–13393
27. Lee A, Kang H-Y, Liou Y-J (2017) A hybrid multiple-criteria decision-making approach for
photovoltaic solar plant location selection. Sustainability 9:184
28. Rumbayan M, Nagasaka K (2012) Prioritization decision for renewable energy development
using analytic hierarchy process and geographic information system. In: 2012 International
conference on advanced mechatronic systems (ICAMechS). IEEE, Tokyo, Japan, pp 36–41
29. Perpiña Castillo C, Batista e Silva F, Lavalle C (2016) An assessment of the regional potential
for solar power generation in EU-28. Energy Policy 88:86–99
30. Chakraborty S, Sadhu PK, Pal N (2014) New location selection criterions for solar PV. Int J
Renew Energy Res 4
31. Gherboudj I, Ghedira H (2016) Assessment of solar energy potential over the United Arab
Emirates using remote sensing and weather forecast data. Renew Sustain Energy Rev 55:1210–
1224
32. Zoghi M, Houshang Ehsani A, Sadat M, Javad Amiri M, Karimi S (2015) Optimization solar
site selection by fuzzy logic model and weighted linear combination method in an arid and
semi-arid region: a case study Isfahan-Iran. Renew Sustain Energy Rev
33. Fang H, Li J, Song W (2018) Sustainable site selection for photovoltaic power plant: an
integrated approach based on prospect theory. Energy Convers Manage 174:755–768
34. Saaty TL (2008) Decision making with the analytical hierarchy process. Int J Serv Sci 1:83–97
35. Kohara K, Sugiyama T (2016) Simulating Tsunami evacuation with multi-agents and
determining a countermeasure with AHP. Int J Anal Hier Process 8
36. Sumabat AKR, Manalac AJ, Nguyen NT, Kalaw ME, Tan RR, Promentilla MAB (2015) Opti-
mizing geopolymer-based material for industrial application with analytical heirarchy process
and multi-response surface analysis. Chem Energy Trans 45
37. Hassan Z, Garni A, Arabia A (2017) Solar PV power plant site selection using GIS-AHP based
approach with application in Saudi Arabia
38. Meitei IC, Singh TB, Denish K, Meitei HH, Singh NA (2019) Optimum design of photo-
voltaic system for a medical institute using HOMER. In: International conference on intelligent
computing and smart communication, pp 1337–1346
39. Meitei IC, Irungbam AK, Shimray BA (2019) Performance evaluation of hybrid renewable
energy system for supplying electricity to an institute and a hospital using HOMER. In:
International conference on intelligent computing and smart communication, pp 1317–1326
A FPGA-Based PUF Integrated
Blockchain to Overcome the Challenges
of Internet of Everything (IoE)
Lukram Dhanachandra Singh and Preetisudha Meher
Abstract Today almost everything is upgrading into a smarter one, connecting them
to the internet, which makes their data accessible to the network. Securing these data
is the most important thing in the present day situation. Many researches have been
proposing various solutions, PUF is one of them which is considered as hardware
security primitives and on another side, blockchain technology is an advancing tech-
nology which is vigorously taken up by many applications such as financial applica-
tions for secure transactions, smart applications to secure the data by decentralizing
them using some cryptography algorithm. But it is a bitter truth that the security
issues do not end with them. It is high time for us to search for an alternate solution
and to design a new better system that can bring these security issues to an end. So
we proposed an FPGA based PUF to be implemented with Blockchain technology
which can provide a better secure and reliable platform, which will be a decentralized
platform consuming ultra-low-power and having efficient computational resources.
Keywords Blockchain technology · FPGA · Programmable unclonable function ·

Internet of things · Internet of everything
1 Introduction
The introduction of the Internet in commercial applications helps in the develop-

ment of electronic commerce which encourages the financial trades among various
entities [1]. A central entity takes all the responsibility to provide security both in
communication and financial transaction among the different entities. And if any
fraud or failure occurs, the central entity will be questioned and will remain respon-
sible. In such centralized systems, at a single point of failure, it has a chance to cause
L. D. Singh (B)
Electrical Engineering Department, National Institute of Technology, Manipur, India
P. Meher
Department of Electronics and Communication Engineering, National Institute of Technology,
Tadepalligudem, Arunachal Pradesh, India
50 L. D. Singh and P. Meher
a terrible system failure. Thus, a central entity has challenges of being trusted and
issues of security and privacy. The central entity also has slow processing during the
transaction process [2].
To solve the issues of the central entity discussed above, it needs to be decentralized
and for the same, blockchain technology was introduced which practices a decentral-
ized ledger, in which all the members in the network uphold a synchronized copy of
the whole or partial ledger. A cryptocurrency called Bitcoin was introduced in 2008
which was a blockchain-based cryptocurrency [3]. Since then numerous researches
have been taken on to practice blockchain technologies in various applications of
finance services, Internet of Things (IoT), Smart Healthcare, Smart Governmental
schemes, etc. Every transaction achieved among multiple entities are deposited in the
distributed ledger and after all positive transaction, the ledger is synchronized among
all the nodes in the network. Now not only the industry is using cryptocurrency, but
also the healthcare, military, and many more are implementing this blockchain tech-
nology, as it excludes the need of a central entity, and also the problems discussed
above.
The Internet-of-Things (IoT) is known to be the pillar for various smart applica-
tion areas, comprising smart healthcare, smart cities and smart transport. The IoT
is a network formed with the interconnection of devices, which communicate and
exchange data among each other through the internet for smart resolution. But lately,
a new concept of interconnecting such various IoTs as one component of its archi-
tecture and environment has been presented [4], called the Internet of Everything
(IoE).
The IoE is consists of four components People, Data, Things and Processes. The
Things in the IoE represent the devices that are associated with each other and the
Internet exchanging data and decision making. People in the IoE come to be nodes
inside the network. Countless devices belonging to the People are endlessly linked
to the Internet and partake in performing communications among devices and other
People. The remaining component is the Data which takes the role of information.
The collected data is used for making intellectual choices in various phases of our
daily life. Once the data is collected, transferring the correct data to the correct place
at the accurate time is the Process in this IoE case. With every component of IoE,
data collection, processing and security can aspect numerous potential threats. As
an answer to all such problems, a FPGA based PUF integrated with Blockchain is
considered and proposed.
2 Problem Definition
There is a common smash at the moment about the Internet of things (IoT) and
its impact on almost everything we face in our day to day life from travelling to
shopping and health. The IoT is a huge network that connects people and the things,
all of which collect and share data about the way they are used. The collected data
amount on the Internet of Everything rises every day.
A FPGA-Based PUF Integrated Blockchain to Overcome … 51
The performance of devices used for these types of applications are low, which
consumes less power, unable to deliver sufficient computational power for the archi-
tecture. To come up with all these concerns, researches are going on around the
world which works on the development of solutions in which the computational
requirements are unloaded to the edge of the network [5]. In such situations, the inte-
gration of edge datacenters assisted to remove the restrictions of resource-limited
and low-performance devices [6].
2.1 Why Blockchain?
Privacy and security are two additional features of IoT architecture that require
the highest consideration through the design level. Numbers of crypto graphical
algorithms were suggested which can be used to reinforce the IoT security. But,
IoT needs a central entity where all the data are stored and various devices or people
communicated through; which is crucial if a cryptographic algorithm is used to secure
the IoT architecture. Blockchain technology can be used to eliminate the necessity of
a central entity from IoT architectures as the blockchain uses a decentralized public
ledger for arranging the data and completing the transactions. As a copy of the ledger
is shared with every node connected to the network, it helps to maintain consistency
and security [1].
Blockchain technology is one among the emerging technologies worldwide which
has the prospective to present it as a solution to several issues in multiple situations
and everyday uses, such as the IoT.
In the blockchain, blocks are formed combining several transactions and nodes
residing in the network produce blocks. All the blocks and their transactions should
be authenticated by the identified node of the network as a portion of the consensus
algorithm. Next to block confirmation, they are cryptographically linked with other
blocks in the chain.
2.2 Why PUF Need to be Integrated into Blockchain?
Along with the advantages of blockchain, it also has some challenges which are essen-
tial to be solved prior to integration with any environment. More or less challenges
require high computational resources, scalability, consumes more power, privacy and
security. As an example, if we consider an IoT atmosphere, almost all these chal-
lenges are at the point of congestion for the integration of blockchain which comes
up so fast and together at a time, which makes it hard to handle them quickly.
A Physically Unclonable Functions (PUF) is random functions which cause
intrinsic properties of hardware to give unique response for a set of inputs (or chal-
lenges). A PUF is in charge of creating a unique key as an identity for the IoT device.
A PUF can produce a sequence of unique keys that can only be generated from the
same PUF module. The set of keys produced by a PUF module can neither be dupli-
cated nor produced form any other module. The PUF keys are not to be store in the
IoT devices memory. When it required the keys, the PUF module will generate it and
is forwarded to the module used for hash, which increases the security of the IoT
device based on the architecture of PUF, an additional key can be created by varying
the input. The output of the PUF key can be altered during the progress and several
security threats can be dodged.
3 Type of Blockchain and Its Challenges
Blockchain technology is of several types. It is elaborated in Fig. 1 [2]. Blockchain

technology uses the idea of a distributed ledger in which the replica of the whole
ledger or a portion of it were shared with every node in the network. Central entity
is substituted by a consensus algorithm in case of a blockchain network [2]. Every
member in the network approve on the consensus algorithm which is like rules
essential to authenticate the transactions. For a block of transactions to be authorized
and added to the blockchain, the miner in the network runs the consensus algorithm
and confirm the transactions.
3.1 Blockchain Technology Challenges
The blockchain practices cryptographic hashes to uphold consistency and security.

When a block is added to the blockchain, it cannot be changed or removed. If anyone
tries to modify the data in the blocks linked with the blockchain, the entire ledger will
be destroyed signifying an inconsistency. There are several challenges for blockchain
[7]. Several challenges of blockchain technology that demands energy or computa-
tion power or prevent its application have found, such as, shortage of scalability,
consumption of high energy, high latency, lack of privacy and fake block generation.
Fig. 1 Different types of blockchain [2]

Different transactions in the blockchain collectively form the blocks and when
the block is formed at the nodes in the network, the process of mining begins which
check and confirm the transactions in blocks and the blocks it selves. The process of
mining needs higher computational power and dedicated hardware, consuming an
enormous power. The scalability is also poor due to the needs of dedicated hardware.
The latency also surges due to the rise in amount of data and nodes in the network.
The time taken to authorize the transactions also surges with increases in the number
of transactions which increases the issues too. There are also some issues where fake
blocks can be made to attack the blockchain.
4 Review on Existing Consensus Algorithm
Consensus algorithms comprises various processes to create, check and confirm

blocks. There are lots of key consensus algorithms, as shown in Fig. 2. We categorize
them into three groups: (1) Validation based, (2) Voting based and (3) Authentication
based.
Bitcoin practices Proof-of-Work (PoW) in which for consistent and secure trans-
action, it maintain miners inside the blockchain network and archive the same
distributed ledger. In PoW algorithm, the participating nodes contest compared
with each other to encrypt the new block into the prevailing chain [3]. New blocks
are mandatory to be under a target value. It consumes a lot of resources; though
provide security by determining the malicious nodes from inserting blocks into the
blockchain. It verifies the new blocks depending on total computing power.
Fig. 2 Various consensus algorithms used in the blockchain

Etherium and Peercoin use Proof-of-Stake (PoS) which utilizes coinage (time
for which the user held a particular currency) concept to determine the target value
[8]. As in PoW, new blocks are mandatory to be under a target value but here it
is determined individually. Higher the coinage is, it is easy for miner to find the
solution. For every new solution provide by miner, a coinstake block that comprises
a coin is referred back to the miner itself, which gives chance to miners [9].
The concept of PoW focused on rewarding the miners for validating the block,
but it also increases the chances for them to win for on-line peers, but off-line peers
cannot claim their reward as they are unable to sign their signature onto the blocks
to claim the reward. Proof of Activity (PoA) was established to reduce the effect of
these matters, whereby refining the security of the network [9].
Voting-based consensus algorithms consist of Ripple [10], Proof-of-Vote [11] and
Proof-of-Trust [12].
Proof of Authentication (PoAh) implements an old-fashioned PoW consensus
method for simple block confirmation [13]. The trusted nodes in the network act
as miners which validate the blocks, trailed by authenticated nodes which drive the
network peers to add block into the chain. This algorithm comprises two steps:
1. Validate the transactions in a block and respective sources, and
2. After positive validation, the corresponding trusted node’s trust value is increased
by one unit (Table 1).
Based on blocks from miners, the distributed ledger is shared to the network peers.
Individual transactions in a block are confirmed by the trusted nodes in the networks.
If it happens to be false identifications, after a certain number a trusted node will
turn into normal node. The trust value of a respective trusted node is reduced by a
unit for each false transaction.
However, for resource-limited IoT applications, any such consensus algorithm is
not yet developed or tested. PoW and PoS are strong consensus algorithms and will
not run in simple, battery-powered IoT devices.
Table 1 Comparison of different existing consensus algorithm

Consensus algorithm Year Types of blockchain Mining based on Prone to attack
used
PoW [3] 2008 Permission less Computation Bribe attack, Sibil
power attack
PoS [8] 2012 Permission less Validation DoS, Sibil
Ripple [10] 2014 Permissioned Voting DoS, Sibil
Proof of vote [11] 2017 Consortium Voting –
Proof of trust [12] 2018 Permission based Probability and DDoS
voting
PoAh [13] 2019 Permission based Authentication –
PoP [14] 2019 Permission based Authentication –
Another consensus algorithm that is integrated with PUF termed as Proof of

PUF enabled Authentication was proposed recently, which was permission-based
blockchain and its mining was based on Authentication [14].
5 Physical Unclonable Functions (PUF) as Hardware

Security Primitive
PUF was specially introduced for hardware security [15] as a random function which
led to manufacturing variations in chip levels from the devices on the Integrated
Circuit (IC). It introduces variations during the IC fabrication process which are
random, inescapable, and uncontrollable and automatic. Due to these Nano-electronic
manufacturing dissimilarities, none of the devices on a wafer are the same. The
input challenge and its resultant response obtained from PUF circuit are termed as
Challenge Response Pair (CRP). For evaluation of PUFs, three main figures of merit
are measured as Uniqueness, Reliability, and Randomness, due to which PUF is
considered as a hardware security primitives.
PUF can be categorized as Silicon PUFs and Non-Silicon PUFs based on fabri-
cation and it is also classified as Strong, Controlled and Weak PUFs based on
security-based. Many researches have been going on for efficient PUF designs
according to their targets. Some of them are for Intellectual Property protection,
anti-counterfeiting, test and debug security, device authentication, and key genera-
tion. And some are for implementing machine learning algorithm in it to provide
more security and some researchers also seek to merge it in FPGA for low power
consumption, to improve processing speed and also for size reduction [1]. Many
review papers are also published on different architectures of PUF, like MUX based
PUF, SRAM PUF, Butterfly PUF, Arbiter PUF, Ring Oscillator PUF and hybrid PUF
which is a combination of two or more PUFs.
6 Contribution of the Paper
6.1 PUF Integrated Blockchain–PUFchain
The main target is to integrate the consensus algorithm for blockchain in IoE network
which has a short form factor and less power for better security and data management.
Figure 3 displays the structure of nodes in PUFchain. The network consist is client
nodes that gather environmental data from devices and upload them to the network
and trusted nodes that are accountable for mining and validation of devices collecting
the data.
A high level of security can be attained with the integration of blockchain to the
network. A hybrid oscillator arbiter PUF was used to implement a PUFchain [14]. A
Fig. 3 Structure of nodes in PUFchain
Hashing module and PUF module is employed above the IoT device that improves
the computation of cryptographic hashes. Addition of PUF module for unique key
generation and cryptographic hash computation lessens the computational needs of
the IoT devices. By selecting FPGA based PUF designs, FPGA based hashing module
and optimization, we can significantly reduce the power consumption; improve secu-
rity, scalability and latency of the blockchain. Figure 4 presents the architecture of
the PUF-integrated blockchain.
The PUF and module used for hash consist of a cryptographic processor build up
on FPGA and the FPGA based PUF module. The cryptographic processor collects
the data from the IoT device and the PUF module supply the unique identification
key. A sequence of transactions was commenced, authenticated and inserted to the
blockchain which authorizes the consensus algorithm. Figure 5 shows the block
structure of PUF integrated blockchain.
Fig. 4 Block structure of PUF integrated blockchain

Fig. 5 Architecture of PUF-chain
6.2 Consensus Algorithm for PUF Integrated Blockchain
There are three phases in designing a consensus algorithm. First one is the device
enrollment phase (E phase) in which it checks the IoT Device at the node if it can be
selected or enrolled for PUF chain consensus algorithm. This can be achieved with
PUF module. All the IoT devices in the PUF chain network should comprises PUF
module to produce a unique identification key which is used later for authentication.
So, in this E phase, a set of challenge inputs will be selected for PUF module and
these challenges should satisfy a set of requirements to be considered as inputs
to PUF [14]. And thus the corresponding CRPs are stored in a secured database
which can be accessed later by trusted nodes. As soon as the IoT device is positively
registered into the network, it is qualified for the initiation of transactions which will
be authenticated by trusted nodes.
Second phase is the initiation of transaction (T phase) which will be generated
after collection of data from the sensor device. The transaction data will consist of
the collected sensor data and the device’s MAC address. The MAC address works as
identification for the remaining nodes in the network including trusted nodes. Just
after the transaction is processed, hardware accelerator will receive it. It consists of
the PUF which is proficient to compute the cryptographic hash. A challenge input
(Ci) is chosen for the PUF which is one among the challenges saved in the secure
database which can be accessed by the trusted nodes. The challenge is passed to the
PUF module and it will collect the response. Then the data linked with the generated
response are computed for hashing by the hardware accelerator. And it is directed
reverse to the IoT device which further broadcast it to the network.
And the last phase is Authentication phase (A phase), which authenticates the
block which is not broadcast to the network before adding to the blockchain. Just
after trusted nodes received the block, the data (Dn which consists of the data collected
by the node along with its MAC address) and the hash (H n ) are recovered which help
of identification of the device. The trusted node obtained the PUF keys from the
secure database with the help of the MAC address and it is sent along with the data to
the hardware accelerator to perform the hashing part. The function used for hashing
by the devices throughout the network should be the same. The resultant hash is
checked for matching with the Hn received from the node. If both the hashes found
to be a match, the device is authenticated and the block is added to the blockchain
and it is broadcasted to the network. If they are found not to be matched with each
other, the process carries on with all the PUF keys that are saved in the E phase. And
if it cannot authenticate the device, the block will be released and not broadcast.
6.3 Evaluation and Analysis of Result
We can evaluate/analyses the result for the proposed PUF-chain using both software
and hardware platform. We can develop a hardware accelerator containing PUF
module and hashing module on an FPGA and we can use Raspberry Pi module
as nodes as done in [14]. The consensus algorithm can be written in the Scyther
simulator which is an automatic verification tool of security protocols. In this, we
need to write the code in Security Protocol Description Language (.spdl), simulation
can be done in Scyther v1.1.3 with a system running with Ubuntu 18.04.3 or higher
version OS. The result can be checked as shown in Fig. 6 [14]. We can also perform
analysis the transaction time and power consumption for the PUFChain. Table 2 gives
the comparison of two existing consensus algorithms, i.e. PoAh [13] and PUFchain
[14].
7 Conclusion and Future Directions
In this paper, we elaborate the steps in various phases of consensus algorithm

which are required to authenticate the nodes in an FPGA-based PUF integrated
blockchain network, providing better security, scalability, and reliability to and
enhancing computational power of IoTs/IoE. We also promoted the ways to anal-
yses the proposed PUF integrated Blockchain module. Some results of the existing
consensus algorithm are also shown. We can implement SHA-256 on FPGA using
blockchain concept as mentioned in [1] for hashing module we can provide far better
security. And if we can implement the proposed consensus algorithm with machine
learning then its security will be very strong that it takes years to have a breakdown
of the barrier in the system.
Fig. 6 PUFchain security verification using Scyther

Table 2 Comparison of
Parameter PoAh based PUFchain
PoAh based blockchain and
blockchain
PUFchain
Blockchain type Permission based Permission based
Mining Authentication Authentication
based based
Security primitive Hashing Hashing and added
PUF key
Over head Device ID Device ID
Hardware needed IoT device capable IoT device
of performing
hashing
Time taken to add received the block (ms)
BlockPi 843 120.3
Clear Pi 85 46.5
(Raspberry Pi 3)
Clear Pi 162.4 120.3
(Raspberry Pi 1)
Time taken for a 950 198
complete
transaction
Power consumption range
BlockPi 3.1 W|3.6 W 4.3 W|6.6 W
Clear Pi 2.1 W|2.5 W 3.3 W|5.6 W
(Raspberry Pi 3)
Clear Pi 1.5 W|1.8 W 2.7 W|5 W
(Raspberry Pi 1)
References
1. Singh LD, Meher P (2019) Advancement of blockchain security solution with FPGA. Int J
Adv Sci Technol 28(3):276–283 (Elsevier)
2. Puthal D, Malik N, Mohanty SP, Kougianos E, Das G (2018) Everything you wanted to
know about the blockchain: its promise, components, processes, and problems. IEEE Consum
Electron Mag 7(4):6–14
3. Nakamoto S (2009) Bitcoin: a peer-to-peer electronic cash system. Cryptography Mailing List
03
4. Miraz MH, Ali M, Excell PS, Picking R (2015) A review on Internet of Things (IoT), Internet of
Everything (IoE) and Internet of Nano Things (IoNT). In: Proceedings of internet technologies
and applications (ITA), Sept 2015, pp 219–224
5. Stanciu A (2017) Blockchain based distributed control system for edge computing. In: Proceed-
ings of 21st international conference on control systems and computer science (CSCS), May
2017, pp 667–671
6. Wright K, Martinez M, Chadha U, Krishnamachari B (2018) SmartEdge: a smart contract
for edge computing. In: Proceedings of IEEE international conference on Internet of Things
(iThings) and IEEE green computing and communications (GreenCom) and IEEE cyber,
physical and social computing (CPSCom) and IEEE smart data (SmartData), July 2018, pp
1685–1690
7. Zheng Z, Xie S, Dai H, Chen X, Wang H (2017) An overview of blockchain technology:
architecture, consensus, and future trends. In: Proceedings of IEEE international congress on
big data (BigData Congress), June 2017, pp 557–564
8. King S, Nadal S (2012) PPCoin: peer-to-peer crypto-currency with proof-of-stake. https://dec
red.org/research/king2012.pdf. Last accessed on 17 May 2019
9. Tschorsch F, Scheuermann B (2016) Bitcoin and beyond: a technical survey on decentralized
digital currencies. IEEE Commun Surv Tutor 18(3):2084–2123
10. Schwartz D, Youngs N, Britto A (2014) The ripple protocol consensus algorithm. In: Ripple
Labs Inc white paper, vol 5
11. Li K, Li H, Hou H, Li K, Chen Y (2017) Proof of vote: a high-performance consensus protocol
based on vote mechanism & consortium blockchain. In: Proceedings of the IEEE 19th inter-
national conference on high performance computing and communications; IEEE 15th interna-
tional conference on smart city; IEEE 3rd international conference on data science and systems
(HPCC/SmartCity/DSS), pp 466–473
12. Zou J, Ye B, Qu L, Wang Y, Orgun MA, Li L (2018) A proof-of-trust consensus protocol for
enhancing accountability in crowdsourcing services. IEEE Trans Serv Comput 1–14
13. Puthal D, Mohanty SP, Nanda P, Kougianos E, Das G (2019) Proof-of-authentication for
scalable blockchain in resource-constrained distributed systems. In: Proceedings of IEEE
international conference on consumer electronics (ICCE), Jan 2019, pp 1–5
14. Mohanty SP, Yanambaka VP, Kougianos E, Puthal D (2019) PUFchain: Hardware-assisted
blockchain for sustainable simultaneous device and data security in the Internet of Everything
(IoE), Sept 2019. https://arxiv.org/1909.06496.pdf
15. Pappu R, Recht B, Taylor J, Gershenfeld N (2002) Physical one-way functions. Sci J
297(5589):2026–2030
Biometrics-Based Pig Identification:
From Invention to Commercialisation
Subhranil Mustafi, Pritam Ghosh, Sanket Dan, Kaushik Mukherjee,

Kunal Roy, and Satyendra Nath Mandal
Abstract The study aims to focus on the entrepreneurship possibilities in the live-
stock sector (particularly Pig) in India. The existing individual pig identification
system is tamperable, stressful, hurtful in nature resulting in low reliability among
the insurance companies in providing the insurance claim as demanded by the farmers
post animal’s death. Thus, a new method of identification has been developed, which
is tamperproof, reliable, unique, and non-invasive. Due to a large number of Pigs
in India (approx. 9.06 million), there is an entrepreneurship possibility to commer-
cialize the identification technology in the market. The market has been studied and
was concluded that charging Rs. 100 per biometric tag, revenue of about INR 90.6
Crores can be generated pan India. Finally, after analyzing the Lifetime Value and
Cost of Customer Acquisition, the break-even point was found near year two and
was predicted a huge profit on and after the second year.
Keywords Entrepreneurship · Animal-Aadhaar · Innovation · Start-up · LTV ·

COCA
S. Mustafi (B) · P. Ghosh · S. Dan · K. Mukherjee · K. Roy · S. N. Mandal

Department of Information Technology, Kalyani Government Engineering College, Kalyani,
Nadia, West Bengal 741235, India
P. Ghosh
S. Dan
K. Mukherjee
K. Roy
S. N. Mandal
64 S. Mustafi et al.
1 Introduction
The Total Early-Stage Entrepreneurship has been 10.6 with a rank of 31 amongst
the top 65 countries, according to the Global Entrepreneurship Monitor report in
India 2017–2018. Invention when mixes with commercialization, leads to innovation.
Entrepreneurs can also be proclaimed as innovators. They always tend to look for
something better and more profitable ways [1]. The process of social, political, and
environmental change expressed in the past 15 years of developing policy landscape
has focussed on the difficulties in persisting with the traditional ‘productivist’ model
and created an imperative for entrepreneurial opportunity-seeking behaviour [2].
“For de Lauwere et al. [3] it’s possible to categorize between groups of farmers:
• Economic entrepreneurs: people who can create significant economic change;
• Socially responsible entrepreneurs: for those who can recognize that success of
the farm may be able to balance socio-environmental role;
• Traditional farmers: those farmers who grow by believing in ‘guaranteed’
successful activities;
• New growers: those farmers who tend to follow diversification but the similarity
in areas of activity;
• Doubting entrepreneurs: those farmers who become disinclined to opt for and
accept change.” [4]
The authors focus on the invention of the unique trait of the pig, from where a
biometric tag can be generated for each animal which would help in identifying the
individual uniquely without hurting it. After the invention of the unique trait of the
species, the authors moved on to commercialize it and help the livestock (particularly
pig) sector widely. Different traits were tested to be inferred to a unique result.
However, the auricular venation pattern was proved the most feasible amongst them
[5].
In this paper, the Primary Market Research has been conducted and market
segments were identified. The pig breeding farms, the processors and the animal
husbandries were the systematic market segments discovered, respectively. Among
the first market segment, ‘DLG Farms [P] Ltd.’ was chosen to be the persona. The
authors have chosen Kalyani, Nadia, West Bengal, India to be the demographics for
the experiment of the invented product where pigs soon after weaning of about 2–
3 months were taken throughout its lifetime till death, to validate the product. Both,
the top-down and bottom-up approach of Total Addressable Market were calculated
where the top-down approach proved to be more effective. However, if the economic
buyer and the champion buyer agree upon the product and its technology, then the
product can easily be handed over to the end-user, who has both the ability to pay and
willingness to buy the product. The generation of the unique biometric tag is itself
a core competency in this domain as now work of non-stressful, non-manipulative
biometric tag for animals has not yet been generated till date, except the physical
tagging system by the government.
Biometrics-Based Pig Identification … 65
Eventually, we found the uprising Life Time Value (LTV) and down steeping
Cost of Customer Acquisition (COCA) coinciding at a breakeven point near the
second year. According to Gerard McElwee [4], farmers should start entrepreneurship
with strategic planning. And there should be a proper identification system for the
customers to rely upon and buy the product which is non-tamperable at any point
in time, and the authors are here to move into it of publishing a product based on
such reliable, biometric identification system. This type of effort has not been made
earlier in the pig sector. This is the reason for making this paper.
The paper is divided as follows: In Sect. 2, Systematic overview of the identifi-
cation process is shown followed by technology description in Sect. 3. Information
about the motivation for selecting pig is discussed. The fieldwork and motivation are
pursued are discussed in Sect. 4 followed by Total Addressable Market in Sect. 5.
The Value Proposition, which is described as the difference between the present
state without the product and the future state after acquiring the product is shown in
Sect. 6, followed by the decision-making journey in Sect. 7 and Core Competency
in Sect. 8. The technical methodology is discussed in Sect. 9 and plot of LTV versus
COCA in Sect. 10. The technology has been developed and validated under project
“ImageIDGP”, funded by ITRA, Govt. of India.
2 Systematic Overview for Individual Identification
The overview of the unique pig individual identification is depicted in Fig. 1 where
Fig. 1 Systematic overview for identification

after the trait acquired, a biometric template is generated which is stored in the
database. The template generator is matched with the template of the new picture
and is checked for equality.
3 Technology
Several traits had been examined but the auricular venation pattern of the pigs was
found to be most feasible amongst them for individual identification. A threshold
has thus been set to distinguish between the same or different pigs, based on the
developed algorithms.
3.1 Trait Selection
There are two main auricular veins in pig ears and a number of branches and sub-
branches arise from them. The number of branches, positions, and distances among
them is unique for each pig. The identification system has been developed based
on auricular venation pattern and it was found that branching up to secondary level
remains fixed for the entire lifetime of the pig (Fig. 2).
Fig. 2 Selecting ear vein pattern as the unique trait

3.2 Method of Identification
The ear vein image is taken as input and then processed through the phases of
segmentation and normalization to get the principal branching as depicted in the
picture and hence calculating the Euclidean distance to identify the threshold limit
for the same and different pigs. Also, most interestingly, the ear vein pattern is
completely different between the left and right ear of the same pig (Fig. 3).
4 Motivation and Field Work
This idea based on the technology push solves the problem where the traditional
methods of identifying individual pigs are tamperable and such kind of manipulation
creates a huge economic loss for the farmers as parent breeds can be manipulated
during breeding. Again, after the death of any animal, the farmers are either underpaid
or are refused to pay by the insurance company due to such unreliable identification
Fig. 3 Method of identification

systems, hence resulting in self-immolation for the farmers. The ability to solve the
problems by helping the people and enhance the economy of the country through the
livestock sector motivates the authors to pursue this opportunity, which may also be
viable simultaneously.
Primary Market Research has been conducted and was found out that the breeding
farms, organized farmers and animal husbandries can be identified as market
segments wanting a solution against the traditional method of stressful physical
tagging. Again, further research selected the breeding farms of West Bengal as the
beachhead market for immediate validation and commercialization of the product.
The demographics were chosen in Kalyani, Nadia, West Bengal and the Psycho-
graphics was the requirement for the replacement of the old traditional methods and
hence an increase in the profit and economy of the farm widely. The proxy product
was the pig restraining tool to minimize the utilization of the manpower in restraining
the animal during taking the ear vein images (Fig. 4).
Fig. 4 Motivation and problem statement

5 Total Addressable Market
The concept here generally revolves around the estimation of the size of the universe
(here considered as the total market), the product, and the technology would be
able to capture without the presence of any of its competitors. The revenue that can
be generated from such a 100% market share of the product helps in limiting the
usage of the product and its usefulness in the market. Of this, 30% market which
can be serviced at the present time denotes the service addressable market. Of the
service addressable market, target market (mostly 10% market share) is taken into
consideration for the most likely buyers at the very first instant. That target market
provides feedback to other customers regarding the product and the rest of the revenue
depends on that. This target market acts as advocates in the customer decision making
journey to the decision-making process for the rest of the market in purchasing the
product in nearby or distant future (Fig. 5).
Fig. 5 Distinction between the markets

5.1 The Total Addressable Market in the Pig Sector as Per

Latest Census
Total estimated user base, based on End User Profile

Number of people globally in your characteristic:
largest demographic or
psychographic characteristic Assumption(s): 4-5 million people (Preferably in
similar to end user = 12-15 million India)
people in pigs sector globally Sources(s): Market Survey (ICAR NRC on Pig)
Based on End User Profile (one particular

characteristic):
1st segmentaƟon based on
Percentage of total estimated user base =
end user profile = 5-6
million 40-45%
Breeders globally
Based on End User Profile (second particular

characteristic):
2nd segmentaƟon
based on end user Percentage of total estimated user base =
profile =3-4 million
FaƩeners & 50-55%
Processors globally
Based on third particular End User Profile

characteristic:
3rd
segmentaƟon Percentage of total estimated user base =
based on end
user profile = 0.12-0.15%
Husbandry
Based on End User Profile characteristic in the

End users targeted geography and particular characteristics:
in
beachhead Percentage of total estimated user base
market =
11.6
million 70%
pigs(WB)
This TAM forms a major portion in anticipation the value proposition. It is the
100% market share of the product and 30% market share of the product forms service
addressable market [6].
Fig. 6 ‘As-is’ state
6 Value Proposition
According to Frow and Payne [7], a value proposition is an offering to the customers
through a series of benefits that the customer is going to receive after the usage
experience. It is the change from the “as-is” state to the “possible” state on acquiring
the product [8]. Now, here the customer wants (1) Unique identification of parent
breeds of pigs to enhance the productivity and hence the economy of the breeding
farm. (2) Creation of database against Unique ID for morphological and vaccination
measures and (3) Proper claim of money for insurance post-death. Also, the customer
fears the sign of the belief for the new tag, whether the product would be able to
identify the animal uniquely, as per the statement. The drivers are the willingness and
ability to pay for the unique individual identification tags which would help, making
them economically stronger (Figs. 6 and 7).
7 Decision-Making Journey
The decision-making journey traverses from the Champion Buyer, the Gate Keeper,
the Decider, the Economic Buyer, the Supplier, the End User, and finally the Influ-
encer. The Gate Keeper helps in determining the type of information from various
sources that are necessary for the product purchase. Then comes the decider, who
helps in choosing the required supplier necessary for the purchase. The initiators
are the ones who help in picking the problem and finding a solution to the problem.
The champion buyer is the one who is willing to buy the product but is not sure
about his/her ability to buy. The economic buyer has the ability to buy the product
but depends on the willingness to buy the product from the champion buyer. When
Fig. 7 ‘Possible’ state
willingness and ability combine, he/she becomes the end-user. Later on, getting the
proper value, he/she advocates or influences others, wanting a product in the same
domain, thereby completing the decision-making process (Fig. 8).
Fig. 8 The consumer decision-making journey

8 Analyzing Core Competency
The Core-Competency is itself the technology developed revolving around the unique
character traits of Pigs & the technology used in developing the algorithm to produce
the exact correct result for any number of samples. The algorithms in staking with the
character traits are completely different from humans, and hence those human-based
algorithms cannot be put to use absolutely.
Man et al. [9] [4] categorized entrepreneurial competences in six key areas of
related competences. The key clusters are
• Opportunity recognition skills
• Relationship building
• Conceptual thinking and problem-solving skills
• Organizing
• Strategic competences.
Potential Competencies The potential competencies in this sector would be:

1. Creating a low-cost package plan for customers having more than 3000 animals
2. Providing a special service of handling and maintaining a database of the pedigree
of large farms
3. Training the customers for any kind of first-aid technical issues with the first-hand
experience.
Since, there is no such revolutionary work done in this sector of the livestock,
hence the product serves as a self-competent product and has covered nearly all the
loopholes of the area, thereby validated in the aforesaid demographics of the market.
Once commercialized, this would create an effective measure of verification during
the distribution of animals by the government for breeding and taking return the same
given individual after the breeding cycle. Also, since the technology revolves around
taking images of the ear vein, the method is feasible than any other ‘fingerprints’ of
the pig.
9 Plot of LTV [10] Versus COCA [11]
The lifetime value represents the total amount of money a customer is expected to
spend on the entrepreneurship possibility. It is based on the products or technology
they get and the value proposition satisfied to them after purchasing the product. The
Cost of Customer Acquisition is the amount of money spent on acquiring customers
through any way of advertisement or sales. As per our research, the LTV and COCA
intersect near the second year (break-even point) (Table 1; Fig. 9).
Table 1 General overview and scaling of LTV and COCA

Year 1 Year 2 Year 3 Year 4 Year 5
LTV 2680 7040 9600 24,000 26,000
COCA 6280 5620 2810 4149 2490
COCA VS. LTV

30,000
25,000
20,000
15,000
10,000
5,000
-
YEAR 1 YEAR 2 YEAR 3 YEAR 4 YEAR 5
LTV COCA
Fig. 9 Plot of LTV versus COCA
10 Conclusion and Future Work
The best part of management is to motivate other people more than anything else.
One can truly pursue such entrepreneurship opportunity in such a wide pig sector
in India of 9.06 million pigs using such biometric-based individual identification
technology, charging only Rs. 100 per biometric tag, which may consist of a unique
code consisting of a database of morphological and vaccination information from
birth till death, to guide to safe pork eventually. The point of intersection between
LTV and COCA as discussed in Sect. 9 clearly shows that from year two, a margin of
profit could be visualized which would grow in the later years. And most interestingly,
no foolproof system has yet been developed which would be able to store information
of the pig throughout the lifetime. This would be a flabbergasting opportunity to gain
profit in such a sector which is of daily need to commons. Later the system may be
extended globally to support the global pig population thereby generating a large
revenue at the end of the day.
Acknowledgements The authors would like to thank ITRA (Digital India Corporation, formerly
Medialab Asia), MeitY, Government of India, for funding, Padup Ventures Pvt. Ltd., and Dr.
Amitabha Bandyopadhyay, Senior Consultant, ITRA Ag&Food for helping to carry out this research
work.
References
1. Khan D (2012) Entrepreneurship in farming

2. Morris W, Henley A, Dowell Da (2017) Farm diversification, entrepreneurship and technology
adoption: analysis of upland farmers in Wales. J Rural Stud 53:132–143
3. Lauwere C, & Enting I, Vermeulen, P, Verhaar K (2002) Modern agricultural entrepreneurship.
International farm management association, 13th Congress, Wageningen, The Netherlands,
July 7–12
4. McElwee G (2006) The enterprising farmer: a review of entrepreneurship in agriculture. J R
Agric Soc England
5. Dan S, Mukherjee K (2020) Individual pig recognition based on ear images. In: Accepted at
COMSYS 2020
6. Mauboussin MJ, Callahan D (2015) Total addressable market methods to estimate a company’s
potential sales. Credit Sussie
7. Frow P, Payne A (2011) A stakeholder perspective of the value proposition concept. Eur J Mark
45(1/2):223–240
8. Anderson J, Narus J, Van Rossum W (2006) Customer value propositions in business markets.
Harvard Bus Rev 84(3):91–99
9. Man T, Lau T, Chan KF (2002) The competitiveness of small and medium enterprises: A
conceptualization with focus on entrepreneurial competencies. J Bus Ventur. 17. 123-142.
10. Hoekstra JC, Huizingh EKRE (1999) The lifetime value concept in customer-based marketing.
J Mark-Focused Manage 3:257–274. https://doi.org/10.1023/A:1009842805871
11. Ang L, Buttle F (2010) Managing for successful customer acquisition: an exploration. J Mark
Manage 22:295–317
Adoption of IoT in Vehicular Traffic
Control: An Overview
Joyeeta Goswami and Ashim Saha
Abstract The number of consumer vehicles is ever-increasing, and traffic manage-

ment is becoming a big challenge all over the world. In hilly or forest-covered areas,
the natural calamities are also an addition to this traffic problem. This gives rise to
economic difficulties as well as the environmental cataclysm of society. IoT (Internet
of Things) presently became a promising technology to solve various serious prob-
lems by establishing the link between the environmental parameters with cloud intel-
ligence. As the name, IoT or Internet of Things suggest it connects Hardware Devices
like a small sensor to cloud server through the Internet or Network. It’s like the
cryptic wand of the technical world. In this paper, we have presented a comprehen-
sive survey on IoT and different existing traffic management systems based on IoT
and the complexities related to them. We have also tried to approach a power-efficient
IoT based smart traffic management system which also incorporates the detection of
different unforeseen natural events like landslides, flood, etc. for all kinds of roads
in rural and urban areas, and will reduce the chances of road accidents.
Keywords IoT · Traffic control system · VANET · DSRC · LoRaWAN
1 Introduction
IoT, the modern span of technology combines all the available devices to commu-
nicate under one network using the internet. There are several reports and statistics
published that indicate that IoT is one of the hottest technologies of the twenty-first
century. The Term IoT is termed by Kevin Ashton, who is the co-founder of Auto-ID
Centre at MIT. in 1999. As per Business Insider’s Report [1], in the next five years,
a whopping $6 trillion will be invested in IoT based products and will give an RoI of
$13 Trillion by 2025. By 2020, M2M communication which is one of the fundamental
J. Goswami (B) · A. Saha (B)

National Institute of Technology, Agartala, India
A. Saha
78 J. Goswami and A. Saha
Fig. 1 IoV revenues in $Billion
for IoT communication will grow to 12.2 Billion from currently 4.9Billion globally
[2]. According to another report [3] published by Forbes IoT Market will reach up to
267billion USD in 2020. And domains like discrete manufacturing, Logistics, and
transportation will drive more than 50% of total IoT expenses by the companies.
Another report by McKinsey, suggests the total IoT market size in 2015 was up to
$900 M, growing to $3.7 B in 2020 attaining a 32.6% CAGR.
Along with this, IoT has a significant role to play on Vehicular Ad Hoc Network
or VANET. The use of Wireless Sensor Network (WSN) in VANET is widely popular
nowadays. It is estimated that in 2017 the Market Value of the Internet of Vehicle
(IoV) was $66,075 million and it will increase to $208,107 million by 2024 with a
CAGR of 18% [4]. According to a report by IDC [5] more than 70% Light-duty Vehi-
cles will be shipped with IoT and Internet-enabled with them. There are many appli-
cations of IoT including Agriculture, Automotive, Supply chain, Industrial Manu-
facturing, etc. In this paper, we will discuss the basic tech architecture of IoT along
with a detailed survey on the traffic control system across the globe. At last, we will
also discuss some areas researchers can work in this field. IoT based services are
also generating new business stream and are expected to generate around $152 B in
2020 (Fig. 1).
2 Communication Protocols
As an enormous and continuously advancing network, IoT has its range of protocols
also transcend. From those protocols, some are not well known but some are very
usual and used extensively in every field. These protocols establish communication
between devices to exchange information with each other in a secured manner. There
Adoption of IoT in Vehicular Traffic Control … 79
Table 1 IoT Communication mediums comparison

Technology Data rate Range Power use Application
GPRS/HSDPA [39, Up to Long range High Cellular/Remote sensing
40] 10 Mbps
LTE [39, 40] Up to Long range High Cellular/Remote sensing
20 Mbps
Wi-Fi [40, 41] Up to 300 ft High Home automation/Remote device
54 Mbps control
BLE (Bluetooth Up to 300 ft Low Home automation/Wearables
Low Energy) [39] 3 Mbps smartphone control/automotive
Xbee [39, 42] Up to Up to Low Home Automation/Mesh
250 Kbps 2 Kms LoS Network/Manufacturing/Traffic
control systems
LoRA [6, 40] Up to Up to 4 Moderate Remote sensing/Smart
50 Kbps Kms LoS city/Industrial
Sigfox [39–41] Up to Long Moderate Remote sensing/Smart city
1 Kbps Range
are multiple communication channels available for both Near Field and Far Field
Application. For devices that are in near proximity of each other, Bluetooth, Zigbee,
and NFC are dominantly used for communication. And for far-field communication,
technologies like Wi-Fi, Cellular. Recently Narrowband-IoT-based communication
channels are also becoming very popular due to their low bandwidth requirements.
Some of the technologies are LoRAWAN, Sigfox, Zwave, etc. Researches show that
[6] for Smart Cities, using LoRA the scalability of the network can be significantly
improved to get around 95% of the packet success rate for serving more than 15,000
nodes. The most popular communication technologies are tabulated here (Table 1).
3 Vehicular Ad Hoc Network (VANET) and Internet

of Vehicle (IoV)
Vehicle-to-Vehicle or V2V communication is the rapidly growing technology that

is possible by IoT as a foundation. In this type of communication, vehicles will
communicate with each other by sharing their real-time information in the around
360-degree range [7]. By virtue of this V2V communication a new traffic Manage-
ment technology has been established named ITS, i.e., Intelligent Transport System.
The role of IoT technology here is to gather the information from various devices
and analyze them to get the most preferable data and broadcast them to the unit or
device where it needed. There are many technologies that are evolved or used in
V2V Communication. One of such is VANET, which stands for Vehicular Ad hoc
Network. VANET is the ad hoc network that supports Inter-Vehicular and Vehicle
to Roadside Device Communication. It accumulates the real-time data from passing
by vehicles, analyze them, and produce some information message or alert message
to the point where needed. VANET uses DTMAC protocol which is a distributed
MAC protocol based on the TDMA scheme to establish communication between the
moving vehicles by using the same frequency range without any interference [8]. It
collects the real-time information of moving vehicles by using some combinations
like Zigbee and GPS or DSRC and GPS, etc. In recent days and also for the future,
the Dedicated Short Range Communication or DSRC protocol has taken a signifi-
cant part in V2V communication. It is the wireless messaging technology of moving
vehicles and operates on 5.9 MHz frequency. DSRC helps the drivers of moving
vehicles to get the information about the surroundings (whether there is any vehicle
or pedestrian is present or not) to control the vehicle according to the requirement,
also gives the message about the coming vehicles to the pedestrians [9]. DSRC not
only communicates between the vehicles but also sends the vehicle’s data to nearby
pedestrian and to VANET. VANET and DSRC commonly work together to control
different causes of accidents such as Blindspot, Sudden lane change, Hard braking,
Intersection movement, etc. For example, if any accident happens then DSRC will
transmit the data to nearby devices and VANET collects this information and broad-
cast the message to the upcoming traffics so that they can control the vehicles and
can avoid the congestion. VANET typically uses IEEE 802.11p, IEEE1609.1, 2, 3
and 4 communication standards [10].
IoV focuses on Intelligent integration of Vehicular communication with Environ-
mental, drivers, and other parameters [8]. IoV integrates technologies like Telem-
atics, Swarm Robotics, Machine Learning, Deep Learning, Artificial Intelligence,
etc. to provide a credible, manageable and highly operational network which consist,
multiple stakeholders, users, vehicle and decision-makers globally. A typical IoV
implementation has two separate communication modes: Peer to Peer Model and
Server-Client Model [11]. The communication between the Vehicles, i.e. V2V or
Vehicle to Roadside Device or Control Unit communication works on Peer to Peer
Model and Roadside Devices sends data to Cloud or Data Centre on Server-Client
Model. Typically the Vehicle to Control Unit Communication is performed over
IEEE 802.11 or IEEE 802.16 protocol [8]. The layer stack of IoV is as follows [11].
User Interaction Layer: This layer collects the data from Vehicles and it’s
environment.
Coordination Layer: Receives data from different sources and process them for
analyzing.
Processing and Analysis Layer: Based on cloud, this layer performs all analyses
and storage of the gathered data.
Application Layer: This layer offers intelligent services to end-users such as
congestion management, alert, streaming, etc.
Business Layer: This layer is used for making business decisions and models from
the IoV infrastructure (Fig. 2).
Fig. 2 VANET and IoV
4 Existing IoT Applications in Traffic Control
As day by day, the traffic problem is increasing, people from different corners of the
world also developing advanced solutions with the help of the latest technologies.
Some are trying to make it with wireless sensor networks [12, 13], some with RFID
[14], some with the neural network [15], and so on. But the well-liked in recent times
is the IoT based solutions. The Roadmap of IoT Based system is like this (Fig. 3).
Due to its mystic features, researchers are using this technology frequently. At
the very beginning, people started with the M2M communication process, consisting
of three parts: M2M devices, the server part, and a web-based user interface [16].
In this approach, the M2M server does the work to link the M2M devices with
the authentication server or AAA (authentication, authorization, accounting), where
authentication server will provide the number of vehicles in a particular area. The user
interface (INVENTOR) gives all the information about the operation of the working
unit to the authorized operator such as faulty units, errors, etc. [16]. After the basic
Fig. 3 IoV roadmap

structure, some improvisations had started. Cameras, snow sensors, fog monitors are
placed in the place of normal sensors in the perception layers, in the network or server
part cloud computing, data accuracy and validation has included, and the Service-
Oriented Architecture (SOA) part was improvised with intelligent traffic control
systems [17]. With the development people start to think of the separate parts of the
traffic system using IoT. Some researchers thought about the crowd indexing system
by using IoT structure to reduce the congestion on the roads [18]. Some started to
think about the speed limit of vehicles to reduce accidents due to rash driving [19]. In
[20] the system consisting of two parts the electronic system and the Software system.
The electronic system detects and gathers the traffic data by utilizing the inductive
loop detection method and transfers them to the base stations through Wi-Fi. And
the second part consists of Green Light Phase Time (GLPT) Calculation Algorithm
in cloud server which analyses the green light phase time and acknowledges back
to the base station [20]. With the advancement of technology, people are analyzing
the causes of traffic-related problems more deeply and try to solve them with more
accuracy. Some are analyzing the accidents happen by sudden traffic slowdown due
to other accidents, peak traffic time, work-in-progress, etc. [21]. Here it found that
the sudden slowdown of traffic is mainly affected by inadequate visibility of drivers
as a result of fog, tunnels, or any other reasons. In this case, Google maps are quite
helpful but not for all kinds of roads. Here, the main focus is on the cloud server. By
using the 4G network the real-time data from the sensors mounted in vehicles are
transferred to an OpenGTS server which will load the information in a SQL database
and gives a real-time OpenStreetMap traffic scenario [21]. In addition, those data
are also put in a MongoDB distributed database by GeoJSON parsing, which will
perform the marking of other vehicles on the same road in close proximity and to
notify them about the sudden slowdown of traffic [21]. Some designs are working
on the priority of traffic lanes using Wi-fi protocols [22], some are modifying the
sensors to IR sensors [23]. As a whole of those modifications, the ultimate is the
ITMS or Intelligent Traffic Management System to build a smart city. In this ITS
or ITMS system the sensors (Ultrasonic, IR, Inductive loop detector) on the roads
or in the moving vehicles will get the real-time information about the traffic density
and broadcast them through IoT protocols (like wi-fi [22, 23], LoRaWAN [24, 25]
and other above-mentioned protocols) to the nearest local units, other vehicles or
pedestrian in the working range and also to the nearest communities, it means all the
units in one system will get all the required information about other [26–29]. After
collecting that information the cloud server analyzes them and works according to
the given algorithm [29–32]. With the advancement, these smart systems have some
challenges also like security, interoperability, connectivity, energy consumption, etc.
[33]. But those problems are also been analyzed and solved by the researchers.
Fig. 4 Localization using RFID and vehicle to roadside communication
5 Way Forward
There are numerous researches are performed in this field as discussed in the previous
section. Many were also implemented nationally or internationally. Though in India
due to several reasons, not many actual implementations are done.
5.1 Location Issues
In [34] many issues of Internet of Vehicle such as localization, security and Radio
Propagation, etc. The GPS module has a location accuracy of 50 m in IoV systems.
Also, the problem of signal lost is pretty dominant in dense areas. Using High-
Frequency RFID along with GPS data can be helpful to mitigate localization issues.
Some researchers are already conducted to use BLE beacon as an indoor navigation
tool [35]. The security and privacy concerns are also important issues, that need to be
taken care of. Researches may be conducted to implement a single global identifier
and use the encryption algorithm for the same (Fig. 4).
5.2 Network Issues
As discussed in the last section, most of the system relies on the internet for commu-
nication between the sensor and cloud server. Based on the data sent from the sensors,
the cloud server analyses the data and generates a proper flag based on the logic set
by the administrator. This system heavily relies on internet connectivity but due to
geographical extremities, it is very possible that in remote hilly areas the internet
may not be available. So due to poor network connectivity, the cloud may not get
the sensor information and may not be able to generate an alarm. There are multiple
ways to solve it. One approach could be using satellite communication instead of
mobile networks as used by Ramesh [13]. But using VSAT increases the total cost
of deployment. Also, the energy required for powering the VSAT satellites is also
significant. To reduce the bandwidth requirement of the cellular network technolo-
gies like Narrowband IoT (NB-IoT) may be used. NB-IoT is an LP-WAN technology
which uses an unused 200 kHz band and has a lower bit rate compared to LTE-M1.
NB-Iot is supported by existing mobile networks and can support as long as 10 years
of battery life [36].
5.3 Data Analytics
Data Analysis is the big challenge for the Internet of Vehicles as the Volume of Data
generated is huge and mostly there is no trusted central authority. The automotive
business and traffic controller authorities need to understand the type and nature of
the data they required and integrate the solution to provide a rich contextual user
interface in the IoV network. The trigger for any alarm should be contextual data-
driven. That means the Analytical Model should be able to trigger the alarm based
on the present data intent. Based on the data generated, the researchers may device
predictive models which will help the drivers and other decision-makers to take
informed decision about any situation.
6 Societal Impacts of IoV
With the Advancement of IoV along with Artificial Intelligence and Machine
Learning, companies are coming up with autonomous vehicles, that will surely
benefit the society by helping persons with special ability to commute easily and
make driving a safer experience for all [37]. Around 90% of cases of car accidents
are due to human error only [38]. IoV along with Autonomous technologies can help
to save millions of people every year. The environment can also be less polluted by
them as the usage of fossil fuel will reduce substantially.
7 Summary/Conclusion
In this paper, we have briefly discussed the various communication mediums used in
IoT Implementation with their comparative study. We have discussed the overview
of Vehicular Ad Hoc Network or VANET along with IoV. Then we reviewed some of
the implementation and researches performed on the application of IoT in the traffic
Management system. Lastly, we have proposed multiple research options available

for this domain, where researchers can work on the practical implementation of
Traffic Management.
Acknowledgements This research work has been carried out in National Institute of
Technology, Agartala, India sponsored by Department of Science & Technology(DST),
Govt. of India under Interdisciplinary Cyber Physical System (ICPS) Program (Ref No:
DST/ICPS/Cluster/IoT/2018/General Dated: 26.02.2019).
References
1. Here’s how the Internet of Things will explode by 2020. https://www.businessinsider.

com/iot-ecosystem-internet-of-things-forecasts-and-business-opportunities-2016-4-28?IR=T.
Accessed 28 Apr 2016
2. Cisco visual networking index predicts near-tripling of IP traffic by 2020. https://newsroom.
cisco.com/press-release-content?type=press-release&articleId=1771211. Accessed on 07 June
07
3. Columbus L. Internet Of Things market to reach \$267B by 2020. https://www.forbes.com/
sites/louiscolumbus/2017/01/29/internet-of-things-market-to-reach-267b-by-2020/\#403928
ad609b. Accessed on 29 Jan 292017
4. Internet of vehicles (IoV) market analysis and forecast 2019–2024 by: Ford Motor, Texas
Instruments, Audi, Intel, SAP, NXP Semiconductors, Apple. https://www.marketwatch.
com/press-release/internet-of-vehicles-iov-market-analysis-and-forecast-2019-2024-by-ford-
motor-texas-instruments-audi-intel-sap-nxp-semiconductors-apple-2019-10-16
5. Worldwide connected vehicle shipments forecast to reach 76 million units by 2023, according
to IDC. https://www.idc.com/getdoc.jsp?containerId=prUS45092819
6. Magrin D, Centenaro M, Vangelista L (2017) Performance evaluation of LoRa networks in a
smart city scenario. In: 2017 IEEE international conference on communications (ICC). IEEE,
pp 1–7
7. Preview the future of transportation-connected vehicles!. https://www.its.dot.gov/factsheets/
pdf/Animation_preview.pdf
8. Fangchun Y, Shangguang W, Jinglin L, Zhihan L, Qibo S (2014) An Overview of Internet of
Vehicles. In: China Communications. http://sguangwang.com/PDF/An%20Overview%20of%
20Internet%20of%20Vehicles.pdf
9. Ashtaiwi A. Vehicle ad hoc networks. https://youtu.be/i2nGSUx9r_s
10. Borcoci E. From vehicular ad-hoc networks to Internet of Vehicles. https://www.iaria.org/con
ferences2017/filesICN17/EugenBrcoci_NexComm2017_Tutorial.pdf
11. Gasmi R, Aliouat M (2019) Vehicular ad hoc networks versus Internet of Vehicles—A
comparative view. In: 2019 International Conference on Networking and Advanced Systems
(ICNAS)
12. Elmrini A, Amrani AG (2018) Wireless sensors network for traffic surveillance and manage-
ment in smart cities. In: IEEE 5th international congress on information science and technology
(CiSt)
13. Ramesh MV. Real-Time wireless sensor network for landslide detection. In: 2009 Third inter-
national conference on sensor technologies and applications. https://doi.org/10.1109/sensor
comm.2009.67
14. Guo L, Fang W, Wang G, Zheng L (2010) Intelligent traffic management system base on
WSN and RFID. In: International conference on computer and communication technologies
in agriculture engineering
15. Gilmore J, Abe N (2010) A neural networks system for traffic congestion forecasting., In:
International conference on neural networks (IJCNN-93-Nagoya, Japan), 1993
16. Foschini L, Taleb T Corradi A, Bottazzi D. M2M-based metropolitan platform for IMS-enabled
road traffic management in IoT. IEEE Commun Mag 49(11):50–57
17. Yu X, Sun F, Cheng X (2012) Intelligent urban traffic management system based on cloud
computing and Internet of Things. In: International conference on computer science and service
system. https://doi.org/10.1109/CSSS.2012.539
18. Zhang SJ, Chen SY (2014) Design and implementation of real-time broadcast system of
urban public transport crowding index based on the Internet of Things. Adv Mater Res
1049–1050:1753–1758
19. Goswami J, Ghosh S, Katiyar S, Majumder A (2015) Development of a prototype to detect
speed limit violation for better traffic management. In: Eighth international conference on
contemporary computing (IC3). https://doi.org/10.1109/ic3.2015.7346723
20. Chong HF, Ng DWK (2016) Development of IoT device for traffic management system. In:
IEEE Student Conference on Research and Development (SCOReD)
21. Celesti A, Galletta A, Carnevale L, Fazio M, Ekuakille AL, Villari M (2017) An IoT cloud
system for traffic monitoring and vehicular accidents prevention based on mobile sensor data
processing. In: IEEE 1558-1748
22. Nagmode VS, Prof. Dr. Rajbhoj SM (2017) An IoT platform for vehicle traffic monitoring
system and controlling system based on priority. In: 3rd International conference on computing
and communication, control and automation (ICCUBEA)
23. Rani PJ, Kumar LK, Naresh M, Vignesh KS (2017) Dynamic traffic management system
using infrared (IR) and Internet of Things (IoT). In: Third International conference on science
technology engineering and management (ICONSTEM)
24. Kadir EA, Efendi A, Rosa SL Application of LoRa WAN sensor and IoT for environmental
monitoring in Riau Province Indonesia, In: 5th International conference on electrical engi-
neering, computer science and informatics (EECSI). https://doi.org/10.1109/eecsi.2018.875
2830, 2018
25. Khattak HA, Farman H, Jan B, Din IU (2019) Toward integrating vehicular clouds with IoT
for smart city services. IEEE Netw 33(2):65–71. https://doi.org/10.1109/mnet.2019.1800236
26. Sadhukhan P, Gazi F, An IoT based intelligent traffic congestion control system for road
crossings. In: International conference on communication, computing and Internet of Things
(IC3IoT). https://doi.org/10.1109/ic3iot.2018.8668131, 2018
27. Masum AKM, Chy KAM, Rahman I, Uddin MN, Azam KI (2018) An Internet of Things (IoT)
based smart traffic management system: a context of Bangladesh. In: International conference
on innovations in science, engineering and technology (ICISET)
28. Nafia HJ, Jacob SM, Nair N, Paul JJ (2018) Density based smart traffic system with real time
data analysis using IoT. In: International conference on current trends towards converging
technologies (ICCTCT)
29. Luo XG, Zhang HB, Zhang ZL, Yu Y, Li K (2019) A new framework of intelligent public
transportation system based on the Internet of Things, vol 7. https://doi.org/10.1109/access.
2019.2913288. Accessed on 17 April 2019
30. Mehmood Y, Ahmad F, Yaqoob I, Adnane A, Imran M, Guizani S (2017) Internet-of-Things-
based smart cities: recent advances and challenges. IEEE Commun Mag 55(9):16–24. https://
doi.org/10.1109/mcom.2017.1600514
31. Putra AS, Warnars HLHS. Intelligent traffic monitoring system (ITMS) for smart city based
on IoT monitoring. In: The 1st 2018 INAPR international conference, 7 Sept 2018, Jakarta,
Indonesia
32. Anitha, Babu KNRM (2018) IOT for ITS: an IOT based dynamic traffic signal control. In:
International conference on inventive research in computing applications (ICIRCA)
33. Du R, Santi P, Xiao M, Vasilakos AV, Fischione C (2018) The sensable city: a survey on the
deployment and management for smart city monitoring. IEEE Commun Surv Tutor 1–1. https://
doi.org/10.1109/comst.2018.2881008
34. Kaiwartya O, Abdullah AH, Cao Y, Altameem A, Prasad M, Lin CT, Liu X (2016) Internet
of vehicles: Motivation, layered architecture, network model, challenges, and future aspects.
IEEE Access 4:5356–5373
35. Singh A, Shreshthi Y, Waghchoure N, Wakchaure A (2018) Indoor navigation system
using bluetooth low energy beacons. In: 2018 Fourth international conference on computing
communication control and automation (ICCUBEA). IEEE, pp 1–5
36. https://www.gsma.com/iot/narrow-band-internet-of-things-nb-iot/
37. Connected autonomous vehicles—How does their social impact generate innovation? https://
www.lexology.com/library/detail.aspx?g=df22bfc0-90e1-46f5-9da3-8ba891e6e607
38. The social implications of driverless cars. https://www.bbvaopenmind.com/en/technology/art
ificial-intelligence/the-social-implications-of-driverless-cars/
39. 11 Internet of Things (IoT) protocols you need to know about. https://www.rs-online.com/
designspark/eleven-internet-of-things-iot-protocols-you-need-to-know-about. Accessed on 20
Apr 2015
40. Hasan M. Top 15 standard IoT protocols that you must know about. https://www.ubuntupit.
com/top-15-standard-iot-protocols-that-you-must-know-about/
41. Sakovich N (2018) Internet of Things (IoT) protocols and connectivity options:
an overview. https://www.sam-solutions.com/blog/internet-of-things-iot-protocols-and-con
nectivity-options-an-overview/. Accessed on 22 Aug 2018
42. Rouse M. Zigbee. https://internetofthingsagenda.techtarget.com/defin-ition/ZigBee. Accessed
on June 2017
Optimal Placement of FACTS Devices
in the Power System Using ANN
to Increase the System Loadability
Mohammad Khalid Saifullah, Md. Monirul Kabir, Nasim Mia,

and Tanvir Ahmed
Abstract Continuous monitoring of online voltage stability as well as increasing

loadability of the existing electrical power system transmission lines are the most
important issues for energy management systems in modern days. In this regard,
the task of monitoring online voltage stability for various loading conditions is very
much complex and time-consuming. This paper demonstrates an efficient voltage
stability monitoring system using the computation of line voltage stability indices
involving an artificial neural network (ANN) for reducing the complexity of an
online voltage monitoring system that can provide warning information to the oper-
ator before voltage collapses. Comparative studies of system voltage stability and
loadability at load point after placing the FACTS devices in the weakest lines of the
power system are presented in this paper. In order to evaluate our proposed system
how does it work; a series of experiments were conducted involving IEEE 14 bus
power system network. The simulation result confirms that the proposed system can
ensure power system operation security as well as reduce electrical power system
installation cost by increasing system loadability.
Keywords Line voltage stability indices · Loadability · Voltage stability · FACTS

devices · ANN
M. K. Saifullah · N. Mia · T. Ahmed

Department of EEE, Mymensingh Engineering College, Mymensingh, Bangladesh
N. Mia
T. Ahmed
Md. M. Kabir (B)
Department of EEE, Dhaka University of Engineering & Technology, Gazipur, Bangladesh
90 M. K. Saifullah et al.
1 Introduction
Nowadays, appraisement and controlling of voltage stability of power systems have

attained considerable attention although they are not new issues. The entire research
works in this field can be divided into several parts among which estimating the
distance towards voltage collapse is very important for the operators. For example,
the demand for electrical energy is increasing day after day but there is a shortage
in installed capacities. As a result, modern power systems are sometimes being
operated very closed to the stability margins [1]. Therefore, maintaining voltage
stability and managing demand has become major concerns for energy management
systems (EMSs) of modern time [2].
As an important tool of online voltage stability appraisement, an artificial neural
network (ANN) has earned worldwide attention to the scholars and researchers in
recent years. The voltage stability assessment problems are non-linear in nature,
therefore, the ANNs are considered as a better option over operable analytical
methods of voltage stability monitoring [3, 4].
However, in this paper, an efficient voltage stability monitoring system has been
presented called as VSMS that is formulated by four ways: (1) the computation of
line voltage stability indices (LVSI) of a power system network, (2) generating a
dataset of the particular network based on LVSI method with varying the loads, (3)
performing the ANN training process on the intelligent dataset as well as achieving
testing ANN for certain voltage collapsed load, (4) finding the ranking of different
buses of the particular network on basis of the testing ANN, (5) finally, placing
the FACTS devices in the weakest lines of the power system. The proposed VSMS
reduces the complexity of an online voltage monitoring system that can provide
warning information to the operator before voltage collapses. Comparative studies
of improvement of system voltage stability and loadability at load point after placing
the FACTS devices in the weakest lines of power system are presented in this paper.
The remainder of this paper is organized as follows. The proposed VSMS is briefly
discussed in Sect. 2. Experimental studies are mentioned in Sect. 3. Experimental
results are reported in Sect. 4. Short conclusions with few remarks are given in Sect. 5.
2 Proposed Voltage Stability Monitoring System (VSMS)
In this paper, an efficient voltage stability monitoring system (VSMS) has been
demonstrated that is mainly formulated by computing the LVSI methods. Generally,
LVSI methods are two types, such as the Jacobian matrix based LVSI method and
system variables based LVSI method. The latter one was selected for the proposed
VSMS because of its less complexity and less computational time. Two methods of
LVSI (1) Line stability index (L mn ) shortly known as L-index method and (2) Fast
voltage stability index shortly known as FVSI have been used in this work. However,
the major steps of VSMS are further explained as follows:
Optimal Placement of FACTS Devices in the Power System … 91
Step-1 Select the suitable two methods of LVSI from various methods and selection
of the optimum FACT device.
Step-2 Calculate the LVSI value of all lines of IEEE-14 bus power system using
verified data to generate the training data set for ANN.
Step-4 Complete the data prediction process by trained ANN and ranking of the
lines according to stability index to find out the weakest lines.
Step-5 Place the selected FACT device in the weakest line in power system.
Step-6 Compare the improvement of load point voltage stability and loadability of
the power system with and without FACT devices.
It is now clear that the idea behind VSMS is straightforward, i.e., generating the
training data and performing the ANN training and ranking the buses of power system
network. The following section gives more details about the different components
of our proposed system.
2.1 Selection of Stability Index and FACTS Device
As mentioned earlier, L-index method was chosen for training data generation of
ANN, whereas, FVSI method was used for the verification of generated data because
of their simplicity and popularity. A power system consists of two machine 100
and 5000 MVA, 50 Hz, 640 km transmission line, three buses with power system
stabilizers (PSS) had been simulated in MATLAB/SIMULINK with various FACTS
devices (UPFC, TCSC, SVC, STATCOM and SSSC) to find out the effective one.
From simulation results and other research results, it has been noticed that UPFC is
the quickest and most effective FACTS devices. Therefore, for the proposed system
UPFC was selected to install in electrical power system to improve voltage stability.
2.2 Formulation of LVSI
Various types of LVSI methods have been proposed by the authors mentioned in
[5–7]. In Fig. 1 a single line diagram of a simple single bus electrical power system
has been shown that has been used in formulation of LVSIs.
V S, δ S VR, δ R
R+jX
G
P load
Q load
Fig. 1 Simple single line diagram of a single bus electrical power system
Here, line impedance, resistance, reactance and line impedance angle are denoted
by Z, R, X and θ, respectively. V R is voltages in the receiving end, V S is denoted
for voltage in the sending end, δ R is phase angle at the receiving end and δ S is
phase angle at the sending end. Again, receiving end reactive power is denoted by
QR and PR is receiving end active power. The L-index method is presented in the
Eq. 2.1 mentioned in [8, 9] that was proposed on the basis of reactive power flow
in an electrical power systems transmission line. For the stability condition of line
voltages, the values of the L-index or L mn for the lines need to be equal or less than
one. Here, the sending bus voltage phase angle is δ m , the receiving bus voltage phase
angle is δ n , therefore δ = δ m - δ n and θ = tan−1 (X/R). Fast voltage stability index
(FVSI) method was provided by Musirin et al. [10]. The equation for FVSI method
is presented in Eq. 2.2. Here the parameters are the same as discussed before. For
stability, the value of FVSI should be equal or less than one.
4X Q R
L mn = (2.1)
[Vs sin(θ − δ)]2
4Z 2 Q R
FV SI = (2.2)
Vs2 X
3 Experimental Studies
This section presents VSMS’s performance on IEEE-14 bus power system network.
In this section, the ANN structure, its training dataset structure, training performance,
ranking of weakest line according to predicted output of ANN, and installation of
FACTS devices process are briefly discussed. The result and its discussion are given
in the next Sect. 4.
3.1 Experimental Setup
In this study, a feed-forward ANN was used including one input layer, one hidden
layer and one output layer as shown in Fig. 2. In this ANN, the input variables are
the variation of reactive loading conditions on the load buses and output or, target
variables are the LVSI values of the interconnection lines or, power transmission
lines between the buses of IEEE 14 bus power system. Here, the size of input layer,
hidden layer and output layer were considered as 14, 10 and 20, respectively. In data
generation, 10% of uniform increment the load variation was considered from 0%
to 500% of input parameters. Furthermore, LVSI values of 20 interconnected lines
in IEEE 14 bus power system were calculated for load variation of 0–500% as the
1 Output 1
Input 1
2
Input 2 Output 2
7
Input 11 Output 17
10
I nput 14 Output 20
Input 14 Neurons Hidden 10 Neurons Output 20 Neurons

Fig. 2 ANN structure for IEEE 14 bus power system
output parameters. The L-index method and FVSI method were used for generating
as well as verifying the training data, respectively.
3.2 ANN Training Performance
In this context, the total 51 patterns of training data for ANN were divided into three
parts (1) training patterns, (2) validation patterns and (3) testing patterns. The ANN
used the training data sets to train itself and used the validation data sets to update
the weight (relations between the layers and neurons) and finally used test data sets
to test itself. The permissible minimum error for the ANN was installed 10−9 . It is
seen in Fig. 3 that the minimum error during validation was found at epoch 577.
Therefore, the validation performance of the ANN network is satisfactory.
To verify the performances of the trained ANN two prototype data set of two
different loading condition was given to the ANN. The graphical comparison of
actual and predicted values of L-indexes (L mn ) for prototype Sample 1 and 2 input
shown in Fig. 4. In the figure, it was seen that the actual and predicted data sets were
all most same value except one small difference in line 19 for Sample 1.
For measuring accuracy criteria of prediction performances of the trained ANN for
the two different loading conditions given to ANN in prototype Sample 1 and 2, the
Mean Absolute Percentage Error (MAPE) of predicted output data were calculated
according to the method given in Eq. 3.1. The less value of MAPE indicates the more
accuracy of prediction performance of the trained ANN.
Best Validation Performance is 2.6504e-09 at epoch 577
Train
-2 Validation
10
Test
Mean Squared Error (mse)
Best
-4
10
-6
10
-8
10
0 100 200 300 400 500 600
634 Epochs
Fig. 3 Validation performance of the ANN network for IEEE 14 bus power system
0.45
0.4
0.35
Value of Lmn
0.3
0.25
0.2
0.15
0.1
0.05
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Line No.
Actual 1 Actual 2 Predicted 1 Predicted 2
Fig. 4 Actual and predicted L mn value for prototype input Sample 1 and Sample 2
N
1 An − Pn
MAPE = × 100 (3.1)
N n=1 An
Here, An represents the actual value of line voltage stability index and Pn represents
the predicted value of line voltage stability index and N represents the total sample
numbers available in the output samples. In Table 1 MAPE calculation results have
Table 1 MAPE calculation

MRPE in (%)
result of predicted outputs for
prototype input Sample 1 and Sample 1 Sample 2
Sample 2 −0.27 −0.099640
Table 2 Ranking of the

Rank Line No. From bus To bus
weakest lines of IEEE 14
electrical power system 1 9 4 9
2 3 2 3
3 19 12 13
4 20 13 14
5 10 5 6
been mentioned. From Table 1 is seen that the error was very low, therefore the
performance of trained ANN was satisfactory.
3.3 Ranking of Weakest Lines
Analyzing the simulation result of LVSI (value of L mn ) prediction by ANN of the

IEEE 14 electrical power system, the 5 weakest transmissions lines have been iden-
tified those have the most probability of voltage collapse at rated baseload condition
of buses in power system. The ranking of weakest lines has been shown in Table 2.
As discussed before for voltage stability of transmission lines the value of L mn should
not gather than one. Therefore, the smaller value of L mn signifies the better voltage
stability of the lines in power system. Similarly, the line having bigger value of L mn
is weaker than the line having smaller value of L mn . According to the values of L mn ,
the weakest lines of the power system had been ranked that is mentioned in Table 2.
From the predicted result it had been observed that the highest output value of L mn
was in line no. 9 which is the interconnection between bus no. 4 and bus no. 9,
whereas, the second-highest value is in line no. 3 (interconnection between bus no.
2 and bus no. 3) and successively the other 3 weakest lines are 19, 20, and 10. These
lines having the most probability of voltage collapse. Therefore, the selected FACTS
devices UPFC should be installed in these lines. If the loading condition of the buses
of power systems is changed, then the rank of the weakest line can be changed also.
3.4 Installation of FACTS Devices
Now, the FACTS device UPFC having capacitive reactance and inductive reactance
X c = 0.05 p.u. and X l = 0.05 p.u., respectively was installed in the weakest power
Table 3 Voltage stability

Line No. Stability without Stability with Stability
improvement data of the
FACTS FACTS Improvement in
weakest lines after installing
%
the FACTS devices
9 0.36927 0.213183 42.27
3 0.14634 0.1120 23.47
19 0.10306 0.069994 32.08
20 0.08644 0.056124 35.05
10 0.075608 0.048052 37.07
Reactive load in MVAR
200 170.48
156.08216
150 130.518
77.8674 82.86799 89.08844 99.195
100 57.844
44.954 56.3
50
0
Line 9 (4_9) Line 3 (2_3) Line 19 Line 20 Line 10 (5_6)
(12_13) (13_14)
Without FACTS With FACTS
Fig. 5 Lodability improvement of the weakest lines after installing the FACTS devices
transmissions lines one after one those has been ranked in Table 2 earlier [11]. After
installing the FACTS devices on the lines, the line voltage stability and lodability
were improved as shown in Table 3 and Fig. 5.
4 Results and Discussions
The line voltage stability and reactive loadability of the electrical power transmis-
sions lines have been significantly enhanced after installing FACTS devices. Voltage
stability improvement of the weakest lines of IEEE 14 bus power system at baseload
conditions showed in Table 3. The average improvement of the lines voltage stability
is more than 33% for IEEE 14 bus power system. If the X c and X l value is changed
then the amount of voltage stability improvement will change. In Fig. 5 the compar-
ison between loadability of the weakest lines before FACTS devices installation and
after FACTS devices installation in the weakest lines has been sowed for base loading
conditions of IEEE 14 power system.
The average loadability improvement that had been calculated for each line was
more than 37 MVAR (48%). The maximum line voltage stability improvement for
FACTS devices installation was found more than 42% in line no. 9 that is interconnec-
tion line of bus no. 4 and bus no. 9. But the maximum line loadability improvement
for the FACTS devices installation was found more than 50 MVAR in line no. 10
that is interconnection between bus no. 5 and bus no. 6. The minimum line voltage
stability improvement for FACTS devices installation was found more than 23% in
the line no. 3 that is interconnection line of bus no. 2 and bus no. 3. But the minimum
line loadability improvement for the FACTS devices installation was found more
than 25 MVAR in line no. 19 that is interconnection between bus no. 12 and bus
no. 13. The UPFC FACTS devices having X c = 0.05 p.u. and X l = 0.05 p.u. were
installed in the lines as mentioned before. If the value of X c and X l is changed then the
amount of loadability improvement will change. The improvement of lines lodability
after FACTS devices installation in the weakest lines in IEEE 14 bus power system
was calculated using LVSI method. Again, the amount of stability improvement and
line loadability increment also depends on various parameters of the power system
such as line impedance, line resistance, line reactance, and many more.
5 Conclusion
In this paper, an efficient voltage stability monitoring system, VSMS has been demon-
strated. Incorporation of the system variables based LVSI method and ANN training,
the proposed VSMS can easily find out the weakest electrical power transmissions
lines that have the maximum probability of voltage collapse in the power system.
After that, placing the suitable FACTS devices to the weakest lines improves the
overall voltage stability and loadability of the power system. In the experimental
studies, the proposed VSMS used standard IEEE 14 electrical power system, in where
the experimental results for various loading conditions confirmed that a significant
amount of improvement of system voltage stability and loadability was obtained after
installing the UPFC FACTS devices in the weakest power transmissions lines (see
Table 3 and Fig. 5). Furthermore, the proposed VSMS can minimize the complexity
of LVSI calculation using ANN. Thus, an operator of a certain power system network
can monitor and take necessary steps within the very shortest time. In Fig. 5, it was
found that the lodability of the existing lines was increased by installing FACTS
devices. Thus, new electrical loads can be added without installing a new power
system network. Finally, it can be said that the proposed system may ensure secured
and reliable service to the consumer as well as having economic significance by
reducing the electrical energy cost.
References
1. Goh HH, Chua QS, Lee SW, Kok BC, Goh KC, Teo KTK (2015) Comparative study of line
voltage stability indices for voltage collapse forecasting in power transmission system. Eng
Technol Int J Civil Environ Eng 9:2–7
2. Ashraf SM, Gupta A, Choudhary DK, Chakrabarti S (2017) Voltage stability monitoring of
power systems using reduced network and artificial neural network. Electr Power Energy Syst
87:43–51
3. Sobajic DJ (1993) Neural network computing for the electric power industry. Lawrence
Erlbaum Associates, Publishers, Mahwah, NJ, USA
4. Jayasankara V, Kamaraj N, Vanaja N (2010) Estimation of voltage stability index for power
system employing artificial neural network technique and TCSC placement. Neurocomputing
73:3005–3011
5. Ratra S, Tiwari R, Niazi KR (2018) Voltage stability assessment in power systems using line
voltage stability index. Comput Electr Eng 1–13
6. Mohamed O, Jasmon AG, Yusoff S (1989) A static voltage collapse indicator using line stability
factors. J Ind Technol 7:73–85
7. Goh HH, Chua QS, Lee SW, Kok BC, Goh KC, Teo KTK (2015) Evaluation for voltage
stability indices in power system using artificial neural network. In: International conference on
sustainable design, engineering and construction, procedia engineering, vol 118, pp 1127–1136
8. Moghavvemi O, Omar M, Omar F (1998) Technique for contingency monitoring and voltage
collapse prediction. In: IEEE proceedings—generation on transmission and distribution, vol
145, pp 634–640
9. Moghavvemi M, Omar FM (1998) Technique for contingency monitoring and voltage collapse
prediction. IEEE proceedings—Generation, transmission & distribution, pp 634–640
10. Musirin O, Rahman I, Rahman TA (2002) Novel fast voltage stability index (FVSI) for voltage
stability analysis in power transmission system. In: IEEE student conference on research and
development (2002)
11. Gerbex S, Cherkaoui R, Germond AJ (2001) Optimal location of multi-type FACTS devices
in a power system by means of genetic algorithms. IEEE Trans Power Syst 16:3
Impact of Noise Levels on SVM-GMM
Based Speaker Recognition System
Renu Singh, Utpal Bhattacharjee, Arvind Kumar Singh,

and Madhusudhan Mishra
Abstract Speaker recognition is one of the most popular voice biometric techniques
used for security reasons in many areas like the banking system, online-shopping,
database access, etc. The Speaker recognition system (SRS) works very well in noise-
free environments as compared to noisy environments. In this study, the impact of
noise levels has been studied by the application of support vector machine-Gaussian
mixture model (SVM-GMM) based speaker recognition system. It has been observed
that the recognition accuracy improves beyond SNR equals 10 dB with the use of
the proposed hybrid system.
Keywords Speaker recognition system · Support vector machine-Gaussian

mixture model · Mel-frequency cepstral coefficients · Signal-to-noise ratio
1 Introduction
Speaker recognition is the procedure of recognizing the voice of a speaker by utilizing

explicit information encompassed in voice samples [1, 2]. The methodology can be
used to verify the claimed identity of people for accessing the systems and enabling
R. Singh (B) · A. K. Singh

Department of Electrical Engineering, North Eastern Regional Institute of Science and
Technology, Nirjuli, Arunachal Pradesh 791109, India
A. K. Singh
U. Bhattacharjee
Department of Computer Science, Rajiv Gandhi University, Papum Pare, Arunachal Pradesh
791112, India
M. Mishra
Department of Electronics and Communication Engineering, North Eastern Regional Institute of
Science and Technology, Nirjuli, Arunachal Pradesh 791109, India
100 R. Singh et al.
access control over various services by recognizing the voice. The various applica-
tions of the speaker recognition system are voice dialing, phone banking, phone shop-
ping, database access, information and reservation facilities, voice mail, data security
control, and remote access of PCs. The other significant use of this technology of
speaker recognition is its utilization for criminological purposes [3].
Speaker recognition can be characterized into a number of classes, for example,
open and closed set, speaker identification and speaker verification, text-dependent,
and text-independent [4]. The recognition of a speaker using a machine comprises
the extraction of features, modeling of speaker, and its testing [5]. The estimation
of the specific features of the speaker from the speech signal is done using feature
extraction. Thereafter, the speaker model is trained with the use of these features.
Finally, in the stage of testing, recognition of the speaker is performed by matching
with the pattern. The objective of the present work is to investigate the effectiveness
of the speaker recognition system accuracy in different noise levels.
2 Speech Parameterization
The inspiration behind the speech parameterization for speaker recognition is to find
the phonetic-discriminative attributes of voice signals while neglecting the untrust-
worthy parts. The best parametric illustration of acoustic waves can be obtained
to create a superior performance of a recognition system. The proficiency of this
stage influences the activity of the consequent phases. The importance of cepstral
features particularly the mel-cepstral for speaker recognition has been studied by
many researches [6, 7]. In spite of several limitations, in practical applications using
the cepstral features due to high sensitivity in the channel and background noises [8].
Mel-frequency cepstral coefficients are the widely used feature both in speech and
speaker recognition systems. Figure 1 represents the block diagram of Mel-frequency
cepstral coefficients feature extraction process [9, 10].
Pre- Frame
Input Speech emphasize Blocking & |DFT|
Windowing
MFCCs DCT Logarithm Filterbank
Fig. 1 MFCC feature extraction process

Impact of Noise Levels on SVM-GMM … 101
3 Database and Experimental Setup
In this work, the NIST 2003 SRE dataset of clean speech has been used. The voice
samples are mixed with various kinds of noises ranging from 0 to 20 dB with an
interval of 5 dB signal-to-noise ratio. The AURORA databases are used in preparing
the sound noises. The different types of additive noises such as airport, car, exhi-
bition, restaurant, street, and train are added to the clean database of NIST 2003.
The following parameters viz. hamming window of length 20 ms, the frame rate
of 1600 Hz with 0.97 pre-emphasis factors are considered for analyzing the speech
signals. Apart from this for the computation of MFCC features 22 channel filter bank
and cepstral features of 14-dimension with cepstral coefficients ranging from c1 to
c13 are used in this analysis. In order to achieve the final feature dimension of 39, the
derivatives of the first- and second-order are added with the base features.
4 Hybrid Speaker Modeling
In this work support vector machine cum Gaussian mixture model has been used. The
structure of SVM and GMM-UBM framework was performed in three phases, i.e.,
training the data using Support Vector Machine (SVM) classifier, building a Universal
Background Model (UBM) as well as the speaker models. A gender-dependent 1024
component GMM-UBM has been offline trained for the used dataset considering
expectation-maximization algorithm with 100 iterations.
5 Results and Discussion
The performance of a speaker recognition system is degraded in the presence of noisy

conditions due to the acoustic mismatch between trained and testing conditions. The
influence of noise levels at different noisy environments on SVM-GMM hybrid
speaker recognition system is calculated using equal error rate (EER), detection cost
function (DCF), and recognition accuracy. The estimation of the EER value of an SRS
system depends on the computation of log-likelihood ratio (LLR) scores. Generally,
a large number of testing samples are required for the evaluation of false reject rate
(FRR) and false alarm rate (FAR), so that the EER can be determined. The DCF
is determined as a weighted sum of the two error probabilities and the recognition
accuracy as [11]:

DCF = Cmiss Pmiss Ptarget + Cfalse alarm Pfalse alarm 1 − Ptarget
Recognition accuracy(%) = 100 − EER
102 R. Singh et al.
The percentage EER values of SVM-GMM and GMM based speaker recognition
systems in clean conditions are shown in Table 1. Table 2 represents the individual
and average EER value of SVM-GMM based speaker recognition system at 0, 5, 10,
15, and 20 dB signal-to-noise ratios in various noisy environments viz. airport, car,
exhibition, restaurant, street, street, and train noises, respectively. From the results, it
has been observed that the impact of noise levels is varying for noisy environments.
From the table, it is seen that at 0 dB noise level, the percentage equal error rate (%
EER) is observed to be lowest for the street which equals 19.88, whereas the highest
% EER has been obtained for exhibition noise and equals 26.80. It can also be seen
that the percentage of EER is maximum at 5 dB for all the noises considered, and
it decreases with an increase in SNR in the range of 10–20 dB. The values of DCF
calculated falls in the range of 0.18–0.45.
The recognition accuracy percentage of SVM-GMM based speaker recognition
systems at different noise levels and at different noisy environments are calculated
and presented in Table 3. The recognition accuracy increases with an increase in
SNR in the range of 5–20 dB considered in the study. The comparison of recognition
accuracy of SVM-GMM based speaker systems of noisy environments with clean
environments is shown in Table 4. The recognition accuracy percentage is seen to
be closer for airport and restaurant when compared to a clean environment at 20 dB
SNR.
Detection tradeoff function (DET) plots are attained considering the probabilities
between “miss” and “false alarm” on a normal-deviate scale for all trials. Figures 2,
3, 4, 5, 6, and 7 show the DET plots between clean vs. noise at various noise levels.
Figure 2 reveals that the pattern of DET plot is similar for all noise levels. Figure 3
Table 1 EER value of GMM and SVM-GMM based speaker recognition system in clean condition
Baseline system Condition EER (%) DCF Recognition accuracy (%)
GMM [12] Clean 7.02 0.05 92.98
SVM-GMM Clean 4.12 0.04 95.87
Table 2 EER (%) value of SVM-GMM based speaker recognition system at different noise levels
and different noisy environments
Noise types 0 dB 5 dB 10 dB 15 dB 20 dB
EER DCF EER DCF EER DCF EER DCF EER DCF
Airport 20.62 0.24 22.68 0.27 17.52 0.18 14.43 0.15 9.27 0.09
Car 19.59 0.23 23.71 0.29 23.71 0.29 18.55 0.19 11.34 0.11
Exhibition 26.80 0.31 22.68 0.27 21.64 0.23 17.52 0.45 10.30 0.40
Restaurant 19.59 0.23 25.77 0.33 22.68 0.23 17.52 0.18 9.27 0.40
Street 19.58 0.22 23.71 0.29 19.58 0.29 19.58 0.29 11.34 0.40
Train 25.77 0.33 25.77 0.33 25.77 0.33 20.61 0.24 12.37 0.12
Average 21.99 0.26 24.05 0.30 21.82 0.27 18.04 0.25 10.65 0.25
Table 3 Recognition accuracy (%) of SVM-GMM based speaker recognition system at different
noise levels and environments
Airport 79.38 77.32 82.48 85.57 90.73
Car 80.41 76.29 76.29 81.45 88.66
Exhibition 73.20 77.32 77.23 82.48 89.70
Restaurant 80.41 74.23 74.23 82.48 90.73
Street 80.42 76.29 76.29 80.48 88.66
Train 74.23 74.23 74.23 79.39 87.63
Average 78.00 75.94 76.79 81.97 89.35
Table 4 Recognition accuracy (%) difference of SVM-GMM based speaker recognition system
between clean versus noisy environments at different noise levels
Airport 16.49 18.55 13.39 10.30 5.14
Car 15.46 19.58 19.58 14.42 7.21
Exhibition 22.67 18.55 18.55 13.39 6.17
Restaurant 15.46 21.64 21.64 13.39 5.14
Street 15.45 19.58 19.58 15.39 7.21
Train 21.64 21.64 21.64 16.48 8.24
Average 17.86 19.92 19.06 13.89 6.52
Fig. 2 DET plot clean

versus airport noise
104 R. Singh et al.
Fig. 3 DET plot of clean

versus car noise

versus street noise
represents the DET plot of clean versus car noise. In the above figure no separate
visible curve line for 5 and 10 dB is noticed.

versus train noise

versus exhibition noise
6 Conclusion
In the present work, an experiment has been performed using NIST 2003 and
AURORA database for the impact of noise levels for SVM-GMM based speaker
recognition system for various noisy environments. The recognition accuracy of
SVM-GMM baseline system is seen to be increased to that of GMM based system
106 R. Singh et al.

versus restaurant noise
for clean environment. From the results, it is also seen that recognition accuracy
increases with an increase in SNR in the range of 5 dB to 20 dB for various noisy
environments.
References
1. Furui S (1991) Speaker-dependent-feature extraction, recognition and processing techniques.

Speech Commun 10(6):505–520
2. Vuppala1 AK, Rao KS (2013) Speaker identification under background noise using features
extracted from steady vowel regions. Int J Adapt Control Signal Process 27:781–792
3. Kunzel H (1994) Current approaches to forensic speaker recognition. In: ESCA workshop on
automatic speaker recognition, identification and verification, pp 135–141
4. Sturim DE, Campbell WM, Reynolds DA (2007) Classification methods for speaker recogni-
tion. In: LNAI vol 4343, pp 278–297
5. Shridhar M, Mohan Krishnan N (1982) Text-independent speaker recognition: a review and
some new results. Speech Commun 1:257–267
6. Ghate PM, Sundar A, Kambale A (2013) Automatic speaker recognition system. In:
Proceedings of ICAdC, AISC 174, pp 1037–1044
7. Furui S (1981) Cepstral analysis technique for automatic speaker verification. IEEE Trans
Acoust Speech Signal Process 29(2):254–272
8. Wu Z, Cao Z (2005) Improved MFCC-based feature for robust speaker identification. Tsinghua
Sci Technol 10(2):158–161
9. Gokay D, Zekeriya T, Lütfü S, Ulus C (2016) A review on feature extraction for speaker
recognition under degraded conditions. IETE Tech Rev 1–12
10. Wang JC, Wang JF, Weng YS (2002) Chip design of MFCC extraction for speech recognition.
Integr VLSI J 32:111–131
11. Rao KS, Sarkar S (2014) Robust speaker recognition in noisy environments. In: Springer briefs
in speech technology http://doi.org/10.1007/978-3-319-071-5_3
12. Singh R, Bhattacharjee U, Singh AK (2019) Performance evaluation of normalization
techniques in adverse conditions. Procedia Comput Sci 171:1581–1590
Digital Payment System
and the Millennial in a Smart City:
An Antecedent to Technopreneurship
Anil Kumar Singh and Tagiya Mudang
Abstract Developing select cities into “smart cities” is an ambitious mission of

the current Indian government led by the Prime Minister, Shri Narendra Modi. It
aims at improving the lifestyle, the standard of living, the quality of life of Indians,
and elevating the possibilities of technopreneurship in the country. This research
attempts to highlight the potential dimensions leading to the success and possible
impediments ahead of this program, thereby estimating the level of success of the
Digital India campaign. This research at the assessment of the attitude of young
citizens towards digital payment systems. The study digs in to identify the factors
leading to the popularity and limitations of the digital payment system amongst the
millennials. The study is conducted in the proposed smart city of Itanagar, Arunachal
Pradesh, India. Primary data is collected through a self-administered questionnaire
built on a five-point Likert scale, and the secondary information is collected from
various government bulletins and published reports. A theoretical model of factors
determining the attitude towards digital payments is tested through confirmatory
factor analysis (CFA). The goodness of fit indices achieved is within the cut-off limits.
It is found that security concerns play the most significant role in determining the
attitude of the millennials in the Itanagar region. The findings of this research update
the existing literature on the acceptance and implementation of the digital medium
of financial transactions in northeast India. The study outlines relevant factors that
can determine the chances of success of the Digital India campaign in Itanagar.
Keywords IBM AMOS · Confirmatory factor analysis (CFA) · Digital payment

system · Millennials · Smart city · Technopreneurship
A. K. Singh
Assistant Professor, School of Business, Assam Kaziranga University, Jorhat, Assam, India
T. Mudang (B)
Research Scholar, Department of Humanities and Social Sciences, North Eastern Regional
Institute of Science and Technology, Nirjuli, India
110 A. K. Singh and T. Mudang
1 Introduction
The digitalization of data and seamless sharing of information has transformed the
way we conduct our business today. This fact is noteworthy in the Indian context
because unlike the developed countries in the west, information technology (IT)
revolution is a relatively new phenomenon here. It surfaced only in the early 1990s
during the liberalization, privatization, and globalization (LPG) era, and eventually
became a nationwide buzz after the launch of the “Digital India” campaign by the
Prime Minister Narendra Modi led Indian government in the year 2015. While some
of the technological changes that have happened in India have so far succeeded, others
have not [1]. One such significant technological switch is the digital payment system
that gained fresh impetus in India after the controversial decision of the government
on November 8, 2016, to demonetize high denomination Rs. 500 and Rs. 1000 bills
accounted for nearly 86% of the total cash in the economy [2]. The nation witnessed
a remarkable shift towards embracing alternate digital payment options of mobile
wallets such as Paytm, Mobikwik, m-Pesa, Airtel Money, and much more post-
demonetization [3]. By the year 2016, India had already become the world’s second-
largest smartphone consumer market [4], but 68% of the Indian population continued
to prefer cash to digital payments [1]. The transition from the traditional cash-based
business model to a less-cash model is still a major challenge. An impediment to
this initiative, however, would be the acceptability of digital mediums of payments,
and other such related issues. The attitude of people towards digital systems is of
considerable significance here. An individual’s perception of digital payment systems
can affect the program to a great extent. As mobile phones are particularly popular
amongst the generation Y, it no longer should be seen merely as a music player,
a navigation device, or a camera device. It is now time to disrupt online payment
systems through the immense possibilities of mobile devices [5]. But the matter
that bothers IT enthusiasts worldwide is the fact that though mobile technology got
warm and prompt reception worldwide, electronic payment services have witnessed
a slower adoption rate [6]. A list of benefits of the electronic/digital payment system
is listed in Table 1.
Table 1 Benefits of digital

S. No. Benefits Supporting literature
payment system
1 Security [7]
2 Reliability [8]
3 Scalability [9]
4 Anonymity [10]
5 Acceptability [11]
6 Privacy [12]
7 Efficiency [13]
8 Convenience [12]
Digital Payment System and the Millennial in a Smart … 111
This study aims to assess the perspective of the millennials towards digital
payment systems and predict the likelihood of success technopreneurship through
the Digital India campaign.
2 Identification of Constructs
To realize the main objectives behind making cities smart, it is essential that people’s
attitude towards digital lifestyle remains positive. To assess the view of millennials
towards digital payment systems, their technology readiness (TR) and adoption readi-
ness (AR) need to be assessed [1]. While TR is people’s tendency to welcome and
use new technology [14], AR is the willingness of people to take up a digital platform
for convenience and utility [15]. Apart from other measurements, security and trust
are the two original dimensions that determine TR [9, 16].
Performance expectancy (benefits), ease of use (convenience), social influence,
and facilitating conditions measure the AR [1]. Collectively AR and TR determine
the attitude of users towards the digital payment system. Based on this theory, a
conceptual research model to assess the attitude of millennials towards the digital
payment system is proposed (Fig. 1).
3 Research Methodology
This study is empirical, using data from both primary as well as secondary sources.
For secondary information, various authentic reports, high-end research papers were
referred. The primary data was collected using a self-administered questionnaire. The
questionnaire consisted of 15 close-ended items designed to measure the constructs
Security
Technology
Readiness
Trust Attitude
Towards
Digital Pay-
ment System
Convenience
Adoption
Readiness
Benefits
Fig. 1 Default relationship model

Table 2 Summary of factors

S. No. Parameters Constructs Literature support
measuring TR and AR
1 Technology Security [9, 16–20]
readiness Trust
2 Adoption Convenience [1, 15, 21, 22]
readiness Benefits
adapted from literature viz. security, trust, convenience, and benefits (refer to Table 2).
The questions were developed on a five-point Likert Scale ranging from 1 = strongly
disagree to 5 = strongly agree. The questionnaire was then checked for its face
validity and content validity by sending it to the subject matter experts. Data analysis
was carried out on the IBM SPSS v21 software platform.
3.1 Data Screening
Monotones are the responses collected that have no variance and hence should be
dropped before further analysis [23]. After transferring the dataset to an excel sheet,
the check for variances was done. All the response sets were found to have some
deviation, and therefore, no monotones were identified.
Missing values are incomplete responses. The possibility of missing values was
eliminated while preparing the questionnaire itself. All the items were marked as
compulsory, thereby preventing incomplete form submission.
Common method bias (CMB) may lead to questionable inferences drawn out of the
statistical analysis, and thereby posing a threat of wrong conclusions [24]. Harman’s
single-factor test, the statistical measure to test CMB, was also conducted on the IBM
SPSS software platform. It was found that the percentage of variance accounted for
the first component is 43.046%, which is less than 50%. This means that the instru-
ment used for this study is statistically free from any significant common method
bias.
3.2 Scale Reliability and Sampling Adequacy
Internal consistency, also known as scale reliability, is a valuable measurement prop-

erty of questionnaires that intends to measure a single underlying construct by using
multiple items [25], and the Cronbach’s alpha (α) is the most common measure of
scale reliability [26]. Literature suggests a minimum alpha (α) cut-off value at 0.5
[27, 28], and a more strict recommended value is 0.7 [29]. The observed Cronbach’s
alpha (α) value for the instrument used in this study obtained from SPSS is 0.906,
signifying excellent scale reliability. The KMO test for sampling adequacy value is
found to be 0.899.
3.3 Construct Validity
Construct validity is tested by a convergent validity test and discriminant validity test.
The convergent validity test is meant to determine the extent to which measures of the
same construct are correlated. One can look at the standardized factor loading values
in a CFA model and loadings of 0.7, and higher is considered excellent. Alternatively,
one can calculate the average variance extracted (AVE) to assess convergent validity.
The authors found out AVEs of the constructs using the following formula:
m
Li
AVE = i=1
n
where n is the number of items.
The discriminant validity test was not done in this study because it is done only in
the case of first-order models, and the research model identified from the literature
in this study is a third-order model. An AVE value of 0.5 and above is considered
acceptable. A summary of validity and reliability test results is given in Table 3.
4 Analysis and Findings
From the literature, it was established that technopreneurship to succeed in India,

people need to embrace technology-based business models, and technology readiness
and adoption readiness are the key criteria to assess the situation. A conceptual
model (Fig. 1) was developed and tested for acceptability through confirmatory
factor analysis (CFA). The model was recreated on the IBM AMOS v21 software
platform (Fig. 2). The four measurement dimensions viz. security, trust, convenience
and benefits were loaded with four independent variables respectively (except on
“benefits”, where only three measuring variables were loaded). The variables and
their respective factor loadings are given in Table 4. The CFA tests were run on
Table 3 Validity and reliability test results

Measures Items Factor loadings KMO AVE Cronbach’s α
Security 4 0.72–0.93 0.899 0.69 0.906
Trust 4 0.69–0.79 0.58
Convenience 4 0.54–0.69 0.40
Benefits 3 0.61–0.74 0.47
Fig. 2 IBM AMOS model
AMOS and the results are compiled in Table 5. After repeated improvements made
to the model, the authors could obtain acceptable values of the crucial model fit
indices.
4.1 Model Fit Indices
The proposed model has yielded encouraging fit indices. The absolute fit indices of
the normative χ 2 (CMIN/DF) = 2.140, RMR = 0.023, GFI = 0.915, RMSEA =
0.076, and the SRMR = 0.046, all satisfying the threshold levels comfortably. At the
same time, the incremental fit index CFI = 0.943 also satisfies the benchmark.
Conventionally, it is required to report one fit measure belonging to the absolute
fit measures and one from the incremental fit measures, along with the normative
χ 2 value. The results obtained from this study humbly satisfy that criterion. It may
now be determined with confidence that the proposed model to assess the attitude
of millennials towards the digital payment system is empirically validated, and the
relationship between the different endogenous and exogenous variables involved is
genuinely established.
Table 4 Constructs with their respective items’ path loadings in the AMOS model
Constructs Items Standardized path loadings
Security 1. Safety of my data while using a digital payment 0.91
system is vital for me (Code: V1)
2. The risk of data theft while performing digital 0.65
payments in low for me (Code: V2)
3. I stop using digital options of payment whenever 0.84
I learn about security breaches elsewhere (Code:
V3)
4. I make digital payments only if guaranteed by 0.73
web assurance seals (e.g., Verisign, Tech Process,
etc.) (Code: V4)
Trust 1. I trust the ability of digital payment systems to 0.77
protect my privacy (Code: V5)
2. I have full faith in reputed payment gateways like 0.79
PayPal, BillDesk, Verisign, etc. (Code: V6)
3. I prefer to use an internet security (antivirus) 0.79
solution for data protection because I cannot trust
the digital payment system’s ability (Code: V7)
4. I feel the risk associated with the digital payment 0.70
system is low (Code: V8)
Convenience 1. Digital payment is much convenient and 0.51
user-friendly as compared to the paper-based
conventional payment system (Code: V9)
2. The availability of a mobile-based digital 0.58
payment system has enhanced the level of
convenience (Code: V10)
3. I find using a digital payment system complicated 0.68
(Code: V11)
4. It took me little time to learn using the digital 0.68
payment system (Code: V12)
Benefits 1. I save my time and money by using a digital 0.75
payment interface (Code: V13)
2. The digital payment system has reduced my 0.68
reliance on cash-based transactions (Code: V14)
3. The digital payment system is quicker than the 0.62
traditional payment system (Code: V15)
5 Practical Implications and Conclusion
This study started with an idea to assess the preparedness of an ordinary city to be
transformed into a smart city and sustain technopreneurship. A city can be recognized
as a truly smart city if the application of IT is found in almost all the spheres of
life, including business and entrepreneurship. Attitude determines the chances of
adoption and application of any novel idea and technology. In this study, the authors
Table 5 CFA model fit

Fit measure Threshold Observed Background
indices
values values literature
CMIN/DF ≤5 2.140 [30, 31]
RMR <0.05 0.023 [32]
GFI >0.90 0.915 [32, 33]
AGFI >0.90 0.873 [32, 33]
NFI >0.90 0.900 [33]
RFI >0.90 0.872 [33]
TLI >0.95 0.928 [32]
CFI >0.90 0.943 [33, 34]
RMSEA <0.08 0.076 [32, 33, 35]
SRMR <0.05 0.046 [36, 37]
focused exclusively on respondent attitudes towards digital payment systems. The

authors chose to use AR and TR parameters only because of strong arguments in
favor of them. By applying the inputs from the literature, a conceptual model was
designed and later tested for confirmation using CFA on IBM AMOS. The results
were highly encouraging, signifying that a proper model fit was achieved. Sound
model fit indices indicated that the relational assumptions made in the theoretical
model were correct. AR and TR indeed have a significant impact on the attitude
of millennials towards the digital payment systems. TR values were slightly better
than the AR values, meaning that the generation—Y in the city is technologically
better prepared to embrace digital payment systems, but they need to be a little more
convinced about its benefits in realtime. The government faces challenges while
introducing technology into everyday business affairs. But it needs to spend more
resources in improving the conviction in people about the endless benefits and utility
of IT. Although the people give high importance to online security, the overall attitude
of the millennials is in favor of the government’s move to upgrade their city into a
smart city. The Digital India campaign has excellent chances of success in this part
of the nation if it can ensure enough and timely infrastructure. Technopreneurship is
a solution to many issues in the Itanagar region, but its chances of success are shady
now because of a lack of clarity in the minds of young people here about how IT
can be best used for business and entrepreneurship. But at the same time, they are
welcoming of technology applications in business, as evident from the AR and TR
scores.
References
1. Sinha M, Majra H, Hutchins J, Saxena R (2019) Mobile payments in India: the privacy factor.
Int J Bank Mark 37:192–209
2. Wade S (2017) How India is surviving post-demonetization. https://www.forbes.com/sites/wad
eshepard/2017/07/29/how-india-is-surviving-post-demonetization/#3f6841ee1164. Accessed
30 May 2019
3. Nielsen Report (2016) Demonetisation: the Nielsen view (part II). https://www.nielsen.com/
in/en/insights/reports/2016/demonetisation-the-nielsen-view-part-two.html. Accessed 01 June
2019
4. Iyengar R (2017) India poised for smartphone revolution. https://money.cnn.com/2017/09/26/
technology/india-mobile-congress-market-numbers/index.html. Accessed 30 May 2019
5. Kazan E, Damsgaard J (2014) An investigation of digital payment platform designs: a
comparative study of four european solutions. In: European conference on information systems
6. Guo J, Bouwman H (2016) An analytical framework for an m-payment ecosystem: a merchants’
perspective. Telecommun Policy 40:147–167. https://doi.org/10.1016/j.telpol.2015.09.008
7. Chou Y, Lee C, Chung J (2004) Understanding m-commerce payment systems through the
analytic hierarchy process. J Bus Res 57:1423–1430
8. Neuman BC, Medvinsky G (1995) NetCheque, NetCash, and the characteristics of Internet
payment services. J Electron Publ 1
9. Tsiakis T, Sthephanides G (2005) The concept of security and trust in electronic payments.
Comput Secur 24:10–15
10. Stroborn K, Heitmann A, Leibold K, Frank G (2004) Internet payments in Germany: a
classificatory framework and empirical evidence. J Bus Res 57:1431–1437
11. Linck K, Pousttchi K, Wiedemann DG (2006) Security issues in mobile payment from the
customer viewpoint. In: 14th European conference on information systems (ECIS 2006).
Munich Personal RePEc Archive, Göteborg, pp 1–11
12. Kousaridas A, Parissis G, Apostolopoulos T (2008) An open financial services architecture
based on the use of intelligent mobile devices. Electron Commer Res Appl 7:232–246
13. Cotteleer MJ, Cotteleer CA, Prochnow A (2007) Cutting checks: challenges and choices in
B2B e-payments. Commun ACM 50:56–61
14. Colby CL, Parasuraman A (2001) Techno-ready marketing: how and why customers adopt
technology. Simon and Schuster
15. Thakur R, Srivastava M (2014) Adoption readiness, personal innovativeness, perceived risk
and usage intention across customer groups for mobile payment services in India. Internet Res
24:369–392
16. Kim C, Tao W, Shin N, Kim K-S (2010) An empirical study of customers’ perceptions of
security and trust in e-payment systems. Electron Commer Res Appl 9:84–95
17. Lim B, Lee H, Kurnia S (2006) Why did an electronic payment system fail? A case study
from the system provider’s perspective. In: Collaborative electronic commerce technology and
research conference, Adelaide
18. Peha JM, Khamitov IM (2004) PayCash: a secure efficient Internet payment system. Electron
Commer Res Appl 3:381–388
19. Yousafzai SY, Pallister JG, Foxall GR (2003) A proposed model of e-trust for electronic
banking. Technovation 23:847–860
20. Zhou T (2011) An empirical examination of initial trust in mobile banking. Internet Res 21:527–
540
21. Eastin MS (2002) Diffusion of e-commerce: an analysis of the adoption of four e-commerce
activities. Telemat Inform 19:251–267
22. Chakravorti S (2003) Theory of credit card networks: a survey of the literature. Rev Netw Econ
2:50–68
23. Yeh K-JJ (2009) Reconceptualizing technology use and information system success: developing
and testing a theoretically integrated model
24. Podsakoff PM, MacKenzie SB, Podsakoff NP (2012) Sources of method biases in social science
research and recommendations on how to control it. Annu Rev Psychol 63:539–569. https://
doi.org/10.1146/annurev-psych-120710-100452
25. Terwee C, Bot S, de Boer M, Van der Windt DA, Knol D, Dekker J, Bouter L, de Vet H (2007)
Quality criteria were proposed for measurement properties of health status questionnaires. J
Clin Epidemiol 60:34–42. https://doi.org/10.1016/j.jclinepi.2006.03.012
26. Field AP (2005) Discovering statistics using SPSS. Sage, London
27. Zaichkowsky JL (1985) Measuring the involvement construct. J Consum Res 12:341–352
28. George D, Mallery P (2010) SPSS for windows step by step: a simple guide and reference.
Allyn & Bacon
29. Nunnally JC (1978) Psychometric theory. McGraw-Hill, New York
30. Wheaton B, Muthen B, Alwin DF, Summers GF (1977) Assessing reliability and stability in
panel models. Sociol Methodol 8:84–136
31. Tabachnick BG, Fidell LS (2007) Multivariate analysis of variance and covariance. In: Using
multivariate statistics, vol 3, pp 402–407
32. Hooper D, Coughlan J, Mullen MR (2008) Structural equation modelling: guidelines for deter-
mining model fit. Electron J Bus Res Methods 6:53–60. https://doi.org/10.1037/1082-989X.
12.1.58
33. Hair JF, Black WC, Babin BJ, Anderson RE, Tatham R (2010) Multivariate data analysis.
Pearson Prentice Hall, Upper Saddle River
34. Gefen D, Straub DW, Boudreau MC (2000) Structural equation modeling and regression:
guidelines for research practice. Commun Assoc Inf Syst 6:1–30
35. MacCallum RC, Browne MW, Sugawara HM (1996) Power analysis and determination of
sample size for covariance structure modeling. Psychol Methods 1:130
36. Byrne BM (2013) Structural equation modeling with LISREL, PRELIS, and SIMPLIS: basic
concepts, applications, and programming. Psychology Press
37. Diamantopoulos A, Siguaw JA (2000) Introducing LISREL: a guide for the uninitiated. Sage
Efficient Fusion Based Multi-modal
Biometric Authentication System Using
Machine Learning
R. Sindhuja and S. Srinivasan
Abstract In the recent past, a multi-modal biometric authentication system plays

a vital role in individual biometric authentication for private and public sections to
assure security. The conventional multimodal biometric system performance is poor
in various illumination conditions of biometric images during feature extraction.
Fusion-based Multimodal biometric improves the accuracy and efficiency of security
level, nonuniversality, variation of inter and intraclass, etc. In this paper, fusion-based
multi-modal based biometric modality verification is proposed through incorporating
modalities iris, fingerprint, and face based on Machine Learning (ML) algorithm. The
proposed system is based on the individual scores estimated through every biometric
modal and then normalized to receive score of fusion. Various statistical features
can be estimated through every biometric modality modal based machine learning
approach that can be applied for classification using ML. The biometric images are
undergone for preprocessing to remove noise occurrence and to improve the quality of
images even for various illumination conditions to achieve uniform illumination and
contrast. The proposed system is used to achieve an extensively multimodal system
with considering the improved wide range of quality of images. The segmentation
algorithm is applied to segment the required region for statistical feature estimation.
The fusion is applied using extracted features of biometric using pixel-based fusion.
The proposed fusion technique capture takes the information of high frequency such
as edges and slant textures of images for fusion. The estimated statistical features
could be feed to ML to identify the given image is obtainable in a secured database
or not. The various experimentation results are proof that the proposed methodology
is improved efficiency and accuracy.
Keywords Multimodal biometric authentication · Machine learning · Region of

interest (ROI) · Image enhancement · Image fusion
R. Sindhuja · S. Srinivasan (B)

E&I, Annamalai University, Chidambaram, India
R. Sindhuja
120 R. Sindhuja and S. Srinivasan
1 Introduction
Multi-modal biometric system (MBS) organized in necessary applications could be

unimodal like they utilize a mono source of data for recognition such as finger-
print, voice, face, etc. Few of the disadvantages exposed by the unimodal biometric
system may be overcome by improving several supplies of data for implementing
uniqueness. The fingerprint and face authentication could make a better integration
for a multimodal biometric system and utilized in the work for improving authenti-
cation by capture equipment, digital image processing unit, and personal computer.
The software unit incorporates database and modules of implementation [1]. The
MBS integrates data through various biometric modalities to enhance the efficiency
of every entity multi-modal modality biometric environment and create network
healthy to attacks of spoof. A secured MBS uses an integration of Convolutional
Neural Network (CNN) and the supervised learning method of Q-Gaussian Support
Vector Machine (QG-MSVM) is applied for the various status of fusion. Two recog-
nition methods with two various fusion level algorithms are used for secured MBS
with feature fusion level and decision level of the proposed system. The CNN algo-
rithm is used for feature extraction of biometric images. These two layers of CNN are
used for improving accuracy in that every layer is applied as descriptors of features.
The inner fusion method is used to generate multimodal biometric templates [2].
The MBS may procedure in various multi-modal biometric signs through individual
identification procedure detecting people. MBS gives a higher state of recognition
than a traditional biometric system that applied only a single data of biometric infor-
mation like fingerprint or face or some other information. The fingerprint and iris
are used to score matching techniques [3]. In the Information Technology (IT), the
high security is needed for every individual to ensure biometric model based scheme
of authentication. The multi-factor recognition using multimodal biometric method-
ology is used to overcome the disadvantages related to single biometric system.
The feature extraction from speech data is Mel Frequency Cepstrum Coefficient
(MFCC) statistical feature extracted from the speech signal. In the iris image, the
normalization is applied and the Daubechies wavelet scheme is applied to estimate
the features for comparing between various individual biometric to reduce computa-
tional difficulties [4]. Fingerprint recognition is used for the forensic department to
detect crime scenes. The entire crime scene fingerprint recognition method based on
deep learning with CNN is useful for efficient recognition. The various directional
images are captured at the crime site for further processing. The minutia extraction is
used as features for machine learning approach for the precision image too difficult
physical and compound chemical execution methods and stored as database. The
proper enhancement method should be applied to improve the quality of fingerprint
image for accurate recognition using ML [5]. The Gaussian and Poison noises should
be eliminated form the face database. The principle component analysis (PCA) algo-
rithm is applied to calculate eigenfaces and eigenvectors to estimate features of face
[6]. The computer vision and communication of an automatic control system is used
to provide accurate and efficient recognition of face using face detection and face
Efficient Fusion Based Multi-modal Biometric Authentication … 121
authentication procedures. The face detection is done to identity face through an

image that may have few attributes through the given image. Face detection is not
only used for the biometric system and also used for expression recognition and
tracking of face and pose. Face detection is a difficult task due to face is not rigid
and it changes in various parameters such as shape, size, and color illumination [7].
The automatic identification based on speech is also one of the important biometric
recognition systems in different languages. The Digital Signal Processing (DSP)
unit is used to restore and enhance the speech signal to eliminate harmonics from
a noisy speech signal. Features are applied to calculate various features from the
filtered speech signal. The database is created for features to process recognition
on a trained database. The testing signal is given to the DSP unit and proceeds all
algorithms finally calculate features and those features are given to a testing unit of
machine learning to give proper authentication if the features are already collected
and stored in the database of recognition system [8].
2 Related Works
Classification and quality improvement methodology thus multimodal modality

biometric verification network. Quality of image plays a vital role in the multimodal
modality biometric verification network. An authentication technique used exclu-
sively on only frequently capable is impossible convene structure concert needs
feature in low-quality image. The SVM classification for fingerprint and face is used
for outperforming considerably by applying image quality enhancement. Choras
[9] has proposed a biometric authentication method for iris retinal images based
on features and template matching. The multimodal biometric system is combining
both iris and retinal images using Gabor transformation. The Gabor transform is
used to spatial domain image into the frequency domain for calculating low and
high-frequency coefficients for fusion. Toh et al. [10] have investigated various
problems pertaining to the adaptive multimodal authentication system. The prob-
lems incorporate new user login, sensor decompose, and minute model information
size. The less sample size issue is solved by applying a feature scaling space learning
method. Viswanathan et al. [11] have presented a multimodal biometric verification
system using the invariant moment fusion method authentication process. Without
decreasing, preservation limitation novel methodology is implemented for biometric
authentication system. The proposed system uses fingerprint and faces biometrics
standard instant data could be estimated. The invariants are applied to fusion for
creating mono identification data by applying the variance of coefficients. The mono
data is recognized by estimating the variance, calculated based on the threshold data
that is considered. Sarier et al. have described the initial security preservation multi-
modal biometric recognition system protocol resistant to attacks of hill climbing.
The biometric system is encrypted and stored for additional execution of the spatial
domain of cryptography. The proposed system is working based on various cryptog-
raphy parameters functioning on crypto biometric templates created either from a
single attribute such as fingerprint multimodal biometrics. Ma et al. have proposed

that malicious attack highly damages the security and preservation of information
of biometric systems making the authentication of biometric information is making
highly demanding important. The watermarking based multi-stage recognition work
is address to solve security problems. While information selection, modality is water-
marked with fingerprint modality of information credibility token and recognition
resource. The first stage authentication process is executed, at the second stage, the
face patter is further secured by applying the encryption method. The discrete wavelet
transformation is applied to transfer time-domain information by frequency domain
data to extract the low and high-frequency coefficients for further watermarking.
The autocorrelation methodology is applied for watermarking and cross-correlation
method is applied for recover of fingerprint image. Jagadiswary et al. have investi-
gated patterns recognition method for multimodal biometric systems. The scanning
equipment are used to capture images, to store unique characteristics of individuals.
The methodology behind this can be utilized among various range of applications
such as verification, health industry, passport verification, etc. such various field
of application may be beneficial. Lokesh et al. have proposed speech recognition
in different environments such as robotics, healthcare, unmanned vehicle, etc. The
different speech recognition systems are implemented to solve different problems in
application of real world. The novel speech signal recognition is implemented by the
author by utilizing enhancement and feature extraction methodologies. The statis-
tical feature estimation are applied as features extraction from the speech signal. The
features considered as databases of various speech signals. The database is created
for template matching for testing signals. The speech signal biometric authentication
system provides a more range of efficiency and accuracy.
3 Modalities of Biometric Model
Various biometric modalities have been used that are widely utilized in multimodal
modality biometric verification scheme. Speech, face, signature, fingerprint, ear, etc.
are examples of biometric modalities. The biometric modalities are preprocessed
properly to improve the quality and then it is segmented to extract useful information
from biometric modality to extract the features and fused. The physical modalities
stay static with instance while the behavioral modalities modify besides with instance.
The multimodal modality biometric verification scheme utilizes various modali-
ties for authentication and verification. The simplest methods are the singles that
create utilize of a single biometric modality. The biometric authentication systems
creating utilize of more than two biometric modalities are multi-modal methods.
Every biometric system has its characteristics, advantages, and disadvantages.
4 Proposed Methodology
Multi-Modal Biometric Authentication and Verification System (MBAVS) can be

proposed by integrating various multi-modal modalities biometric like fingerprint,
iris, and face and fusion methodology. Various algorithms are applied to estimate
statistical data about various modalities and template creation for authentication and
verification using Machine Learning (ML). The fusion process of face, fingerprint,
and iris in this paper are done using the status of several features is shown in Fig. 1,
flow diagram of fusion methodology. The proposed system significantly consists of
five different parts;
1. Feature-based Fingerprint enhancement and recognition
2. Retinal Iris recognition
3. PCA based Face recognition
4. Multi-modal fusion
5. Machine Learning verification.
5 Fusion Using Score Level (SL)
The score level (SL) fusion is used for fusing various biometric features for classifi-
cation and recognition using machine learning. Various matching scores that could
be estimated by various classifiers through various biometric standards could be
applied fusion for matching at this status. Fusion at the level of matching could be
involved in two different ways. First stage is classification issues and as data incorpo-
ration issue. In the classification method, feature extraction is estimated by applying
a score matching method that is giving output using separate matches. In the next
state, feature vectors are classified as acceptance level and rejection level, that is,
genuine user and imposter classifications. In the data incorporation method, separate
matching scores are applied for fusion to create decision making.
6 Fingerprint Recognition
Fingerprint recognition primarily consists of image restoration, image enhancement,

image segmentation, feature extraction, and recognition using machine learning. We
are performing fingerprint recognition using various techniques such as fingerprint
minutiae extraction, orientation field, fingerprint thinning, etc.
7 Fingerprint Orientation Image
The orientation image of the fingerprint shows the exact ridge orientation in the
fingerprint image in all areas. Here are a variety of approaches used to determine
the fingerprint image orientation. The ridge orientation varies very slowly across
the given fingerprint apart from in the area of singularities such as core and delta.
Finally, there will be a full resolution of the orientation image. A single orientation
that relates to the highly credible or main orientation of the block is processed instead
of any non-overlapping block of size W /W of the image. Relatively, the horizontal
and vertical direction gradients AGx(x, y) and BGy(x, y) are calculated based on
simple gradient detection methods such as the Sobel edge detection technique.
The fingerprint minutiae block direction θ is given by θ = a{1/2} ∗
abs{tan −1Gyy/Gx x}
Gx y = u ∈ W v ∈ W 2 AGx(au, bv) × BGy(au, bv) (1)
Gx x = u ∈ W v ∈ W AG2x(au, bv) × BG2y(au, bv) (2)
8 Iris Normalization
The next step is to rectify the segmented image measurements to allow assessments
after successfully segmenting the iris region from a captured image. There are various
reasons for variations between the eye’s images. Some of them are due to dilation
of the pupil, movement of the eyes, tilt of the head and rotation of the eye within
the eyeball, and distance variations of the face. The most affected difference is due
to variations in light intensity and illumination causes dilation of the pupil leading
to stretching of the iris. To remove these inconsistencies, the segmented image is
transformed. The standardization approach would produce iris regions with the same
constant proportions so that two images of the same iris will have the same character-
istics under different conditions. The next stage is called “segmentation,” which is the
separation of the portion of the iris from the eye image. It is a technique for isolating
and excluding objects, as well as for identifying the iris’ inner and outer boundaries.
Standardization is the third stage. The product of the process of normalization is iris
regions with the same constant dimensions so that at the same spatial location two
iris images of the same individual will have characteristics under different condi-
tions. The fourth stage is to remove the iris and pupil features from the segmented
image by transforming the hair wavelet and the vector function (transforming wavelet
output) by using Principle Component Analysis (PCA). Wavelets could be used to
decompose information about the iris region into different components that appear
at different resolution levels. In this work, hair wavelet transformation was used to
extract the iris field characteristics. Decomposing wavelet images are each step results
in multi-resolution from accurate image to approximate images. The multilevel 2-D

wavelet’s decomposition output has four sub-images: LH, HL, HH represent detailed
images for horizontal, vertical, and diagonal orientation, and LL corresponds to an
approximation image as shown.
9 Iris Segmentation Using Modified Hough Transform
1. The image that helps process the valid region (ROI) is implied in the size map.
It reduces the total number of pixels to handle, reducing processing time and
memory use.
2. Symmetric division of image data allows searching the correct region effectively,
thus reducing the complexity level and reducing the time CHT has to locate
triplets.
The algorithm means that for each grid size P, the ROI only takes region from the
starting point SP to the endpoint EP. Therefore, the reduction of the number of pixels
to be processed to M/PN/P reduces the time complexity of modified algorithms by
order O ((M/P * N/P)).
The proposed methodology pseudo-code can be written as Start
1. Write the N M dimensions image matrix f(x, y).
2. If P is the size of the row, then Sp = (floor(P2)P) and Ep = (floor(P2) + 1)P.
3. Repeat steps 4 to 5 for I = Sp M to Ep[M].
4. For j = Sp, exit N to Ep, exit stage 5.
5. Find triplet (ho, ko, ro) and submit CHT.
10 Face Recognition Using Improved Principle Component

Analysis (IPCA)
The techniques of Principle Component Analysis (PCA) and Linear Discriminant

Analysis (LDA) are among the most popular techniques of extraction feature facial
recognition. This paper involves two face recognition systems, one based on PCA
followed by a feed-forward neural network (FFNN) called PCA-NN, and the other
based on LDA followed by an LDA-NN FFNN. All systems consist of two phases, the
PCA or LDA pre-processing phase, and the neural network classification phase. The
systems proposed display improvement in recognition rates over standard Euclidean
Distance-based LDA and PCA face recognition systems. Let a face image be a
two-dimensional N = N array. An image can also be viewed as a vector of the N2
dimension. A series of images map a set of points in this huge space. In this enormous
image space, face pictures, similar within an overall form, will not be distributed
randomly and can, therefore, be represented by a low-dimensional subspace. The
key idea of the PCA is to find the vectors that best account for the distribution of
Fig. 1 Feed forward neural

network architecture
face images in the whole image space. These vectors define the face image subspace
we call “face space.” By applying the Fisher’s linear discriminating criterion, the
Fisher approach overcomes the drawbacks of the system’s own-face. This criterion
seeks to maximize the ratio of the determinant of the in-class scatter matrix of the
projected samples to the determinant of the projected samples. Fisher discriminants
blend images of the same class and separate images from different classes. Images
are projected from N2-dimensional space (where N2 is the pixel number in the
image) to C − 1 (where C is the number of classes in the image). Consider, for
instance, two sets of points in two-dimensional space projected on a single line. As
with proper space projection, the training images are projected into a subspace. The
test images are projected and represented using a measure of similarity in the same
subspace that varies is the subspace estimate. The neurons in FFNN are arranged as
layers. The neurons receive input from the preceding layer in a layer and feed their
output into the next layer. In this type of network, neuron connections in the same
or previous layers are not allowed. Figure 1 shows the proposed face classification
system’s architecture.
Figure 2 demonstrates the identification of fingerprints using the minutiae extraction

feature. Figure 2b indicates this point’s result.
The fingerprint image is transformed into a skeleton of one pixel after thinning.
Thinned image is the one that allows minutiae to be identified. The image is basically
divided into 3 × 3 windows like Fig. 2c.
Figure 3 demonstrates the identification of iris by transforming modified Hough.
Iris is the sclera-pupil area. The inner and outer edges can be loosely seen as rings.
The method of position is to identify the centers and radii of the internal and external
circles. In the suggested algorithm, we first get the center point and radius of the
a) b) c) d)
Fig. 2 a Input fingerprint image, b enhanced image, c minutiae feature extraction, d ridge mapping
a) b) c) d)
Fig. 3 a Input iris image, b Canny edge detection, c enhanced image, d modified Hough transform
to detect iris circle
pupil. Then we get the center and radius of the iris using the pupil’s center and radius
information, the search scale. Iris images have some features in the distribution of
gray values in different regions. In general, the gray value of the pupil is lower than
the iris value, and the gray iris value is lower than the sclera value so that the pupil
has the lowest gray value in the gray histogram. Based on the above element, we
can differentiate between pupil and iris and sclera by taking binarization. The gray
histogram of the iris image has clear multi-peak features. The highest valley behind
the lowest peak gray value of the histogram is the threshold of segmentation of pupil-
to-other sections. We use an iterative approach to find the first local minimum value
on the left of the histogram’s first peak value is the threshold. Choosing the initial
seeds is a crucial step in the activity of the field. Generally speaking, when we get an
image through the sensor, the yaw angle is not high and the pupil area is close to the
center of the image. Therefore, it is needless to scan all the pixels in the image, and
we can only search for far fewer points. This means that time is small in complexity.
If all 0 (i.e., black) are the pixel values within the 5 * 5 representation of the point,
the point is considered to be within the pupil area. First, we look in the center of the
image, if the point meets the above requirements, it’s called in the region of the pupil
and the search is over otherwise, we’re looking for the white pixels. Table 1 displays
the feature extraction for fingerprint, iris, and face. The SL fusion method is used to
fuse all features based on decision-making theory (Fig. 4).
Tables 2, 3, and 4 are displayed for False Acceptance Ratio (FAR) and False
Rejection Ratio (FRR) for fingerprint, iris, and face. The thresholding value increases,
FRR is increased and FAR is decreased and vice versa. Suppose, FRR and FAR are
the same then it is known as Equal Error Rate (ERR). The ERR is decreasing the
system accuracy and efficiency is improved as shown in Table 5.
Table 1 Fusion using score level (SL)

Features Fingerprint Iris Face Score level
Autocorrelation 63.9782 61.3839 64.9283 62.3899
Amplitude 2.3399 6.38933 2.2389 4.3899
Energy 1.2389 1.4888 1.2389 1.2793
Entropy 0.3783 0.3899 0.3878 0.3989
Homogeneity 7.4899 7.3899 7.4784 7.3989
Sum of squares 10.7989 10.2383 10.2398 10.2389
Sum of average 5.3787 5.9899 5.3989 5.3899
Sum variance 4.2899 4.2389 5.3289 4.5899
Normalization factor 0.2323 0.3544 0.3788 0.2389
Inverse normalization factor 1.2389 1.2899 1.2389 1.9899
a) b) c) d) e)
Fig. 4 a Input face image, b preprocessing image, c face detection, d authentication using best
matching
Table 2 Parameter
FAR FRR Threshold
estimation for fingerprint
biometric authentication 0.010000 98 0
system 0.011633 94 4
0.013265 90 8
0.016531 82 16
0.019796 74 24
0.024694 62 36
0.027959 54 44
0.031224 46 52
0.034490 38 60
0.037755 30 68
0.041837 20 78
0.043469 16 82
0.045918 10 88
0.048367 4 94
0.050000 0 98
Table 3 Parameter
FAR FRR Threshold
estimation for iris biometric
authentication system 0.010000 98 0
0.011633 94 4
0.013265 90 8
0.014898 86 12
0.015714 84 14
0.017347 80 18
0.018163 78 20
0.019796 74 24
0.021429 70 28
0.023061 66 32
0.024694 62 36
0.026327 58 40
0.027959 54 44
0.029592 50 48
0.031224 46 52
Table 4 Parameter estimation for face biometric authentication system

FAR FRR Threshold
0.010816 96 2
0.012449 92 6
0.013265 90 8
0.014898 86 12
0.017347 80 18
0.018163 78 20
0.020612 72 26
0.022245 68 30
0.023061 66 32
0.024694 62 36
0.027143 56 42
0.028776 52 46
0.029592 50 48
0.031224 46 52
0.032041 44 54
Table 5 Classifiers comparison

Types of classifier Accuracy Sensitivity Specificity ERR
KNN 86.3893 83.3993 82.2398 10.3893
BPNN 91.2398 91.2398 90.3939 5.3999
ML (proposed) 98.3989 98.2389 97.2398 0.3989
12 Conclusion
A Multi-Modal Biometric Authentication and Verification System (MBAVS) can

be proposed using three biometric modalities including fingerprint, iris, and face
based on fusion of matching score state is proposed in this paper. The proposed
system provides an accurate and efficient authentication and verification than the
unimodal biometric system. Face recognition system is achieved by applying PCA
algorithm with FFNN based classification method (PCA-FFNN). The fingerprint
image is applied to preprocessing to enhance the excellence of the image and feature
extraction could be done for further recognition. The iris recognition is achieved by
applying anisotropic diffusion filter preprocessing, segmentation of iris using Hough
transform. The face image recognition is achieved using PCA-LDA hybrid algorithm
for accurate recognition. The Score Level (SL) decision making fusion technique is
applied to fuse all features of various modalities. The machine learning algorithm is
used to give accurate and efficient authentication and verification.
References
1. Rahal SM, Aboalsamah HA, Muteb KN (2016) Multimodal biometric authentication system—
MBAS. In: 2016 2nd international conference on information & communication technologies.
IEEE. https://doi.org/10.1109/ictta.2006.1684515
2. Hammad M, Liu Y, Wang K (2018) Multimodal biometric authentication systems using convo-
lution neural network based on different level fusion of ECG and fingerprint. Electronic ISSN:
2169-3536. IEEE. https://doi.org/10.1109/access.2018.2886573
3. Parkavi R, Chandeesh Babu KR, Ajeeth Kumar J (2017) Multimodal biometrics for user authen-
tication. In: 2017 11th international conference on intelligent systems and control (ISCO).
IEEE. https://doi.org/10.1109/isco.2017.7856044
4. Tayal A, Balasubramaniam R, Kumar A, Bahattacharjee A, Saggi M (2009) A multimodal
biometric authentication system using decision theory, iris and speech recognition. In: 2009
2nd international workshop on nonlinear dynamics and synchronization. IEEE. https://doi.org/
10.1109/inds.2009.5227988
5. Pavithra R, Suresh KV (2019) Fingerprint image identification for crime detection. In: 2019
international conference on communication and signal processing (ICCSP). IEEE. https://doi.
org/10.1109/iccsp.2019.8698014
6. Aiman U, Vishwakarma VP (2017) Face recognition using modified deep learning neural
network. In: 8th ICCCNT 2017, 3–5 July 2017. IIT Delhi, Delhi
7. Meena D, Sharan R (2016) An approach to face detection and recognition. In: 2016 international
conference on recent advances and innovations in engineering (ICRAIE). https://doi.org/10.
1109/icraie.2016.7939462
8. Dixit A, Vidwans A, Sharma P (2016) Improved MFCC and LPC algorithm for bundelkhandi
isolated digit speech recognition. In: 2016 international conference on electrical, electronics,
and optimization techniques (ICEEOT). ISBN 978-1-4673-9939-5/16/$31.00 ©2016. IEEE.
https://doi.org/10.1109/iceeot.2016.7755413
9. Choras RS (2010) Multimodal biometric personal authentication integrating iris and retina
images, pp 121–131. Springer
10. Toh K-A, Tran Q-L, Yau W-Y (2004) Some issues pertaining to adaptive multimodal biometric
authentication, pp 617–628. Springer
11. Viswanathan P, Venkata Krishna P, Hariharan S (2010) Multimodal biometric invariant moment
fusion authentication system, pp 136–143. Springer
12. Mansour A, Sadik M, Sabir E, Jebbar M (2017) AMBAS: an autonomous multimodal
biometric authentication system. In: 2017 13th international wireless communications and
mobile computing conference (IWCMC). IEEE. https://doi.org/10.1109/iwcmc.2017.7986607
13. Xu Y, Chuang T-C, Lai S-H (2017) Deep neural networks for accurate iris recognition. In:
2017 4th IAPR Asian conference on pattern recognition (ACPR). 2327-0985/17 $31.00 ©
2017. IEEE. https://doi.org/10.1109/acpr.2017.152
14. Chanukya PSVVN, Thivakaran TK (2019) Multimodal biometric cryptosystem for human
authentication using fingerprint and ear. Multimed Tools Appl. https://doi.org/10.1007/s11
042-019-08123-w
15. Elmir Y, Maadeed S, Amira A, Hassa|ne A (2012) A multi-modal face and signature biometric
authentication system using a max-of-scores based fusion. Springer
16. Huang H, Li J, Ma Z, Feng H (2007) Incorporating image quality in multimodal biometric
verification, pp 933–942. Springer
An MOS-C Multifunction Filter
Employing DXCCTA
for High-Frequency Operation
Karam Bharat Singh, Manoj Joshi, and Ashish Ranjan
Abstract This article introduces a biquad multifunction filter using an advanced

current-mode building block termed as “Dual X Current Conveyor Transconduc-
tance Amplifier (DXCCTA)” which is a combination of dual X second generation
Current Conveyor (DXCCII) followed by Operational Transconductance Amplifier
(OTA). Single Input Multiple Output (SIMO) characteristics of the multifunction
filter produce the filter responses and come with a single DXCCTA block with few
grounded capacitors along with MOS-based resistors. All high-frequency responses
of biquad multifunction filters are verified using CMOS model of DXCCTA as well
as experimental verification using off the shelf ICs AD844 as “Current Feedback
Operational Amplifier” and CA3080 as “Operational Transconductance Amplifier”.
Keywords DXCCTA · Multifunction filter · CFOA/ICAD844 · OTA/CA3080 ·

Frequency response
1 Introduction
An advance active block dominance in the current scenario is now considered as an

important active device in microelectronics engineering which gives a suitable plat-
form for the generation of signal processing circuits viz. active filters [1–13], oscilla-
tors [14–16], Schmitt trigger [17, 18], chaotic circuits [19, 20], active inductor design
[21] and many more. Initially, voltage-mode op-amp has the major contribution for
almost every electronic circuits but after the emergence of current-mode circuits
[22] became a new trend for the design of active mode circuitry with low power
dissipation, greater linearity, inbuilt tunability, wide bandwidth, higher frequency of
K. B. Singh (B) · M. Joshi · A. Ranjan

National Institute of Technology Manipur, Imphal, India
M. Joshi
A. Ranjan
134 K. B. Singh et al.
operation and few others [23]. Some very popular advance active blocks are CCII
[1, 4], CCCII [10], OTRA [24], FTFN [7, 13], CDTA [8], VDCC [9], DVCC [3,
12], CCTA [5], DDCCTA [25], DVCCTA [26], DXCCTA [27], FTFNTA [28] and
few other. A wide array of second-order filters is available and can be treated as
the fundamental filter circuitry in the analog domain. Moreover, literature is also
enriched with different input–output combination filter as with SIMO [1–3], MISO
[4–9] and MIMO [10–12]. In the case of multifunction filter, SIMO operation plays
a significant role due to single input source which causes attractive design procedure
without the selection of various input sources as in the case of MIMO and MISO.
An intensive study of literature [1–28] have the following characteristics:
1. Excess number of active block elements [1, 2, 5, 6, 11]
2. More input signal to realize output filter [7–9]
3. Absence of electronic tenability [1, 24]
4. Low cut off frequency response [1, 4, 8, 24]
5. Component mismatch and circuit topology to get different responses [1, 5–9]
6. Necessary to use external passive components for measurement of different filter
responses in current-mode topology [27].
This research paper brings a SIMO filter topology for the multifunction filter
responses viz. Low pass (LP), Band Pass (BP) and High Pass (HP) filter by utilizing
active MOS resistors and grounded capacitors with single active block DXCCTA.
The proposed design is suitable for a very high-frequency operation up to 200 MHz.
The examination of the frequency test of the multifunction filter is well executed
through PSPICE simulation. An experimental test of the filter is also examined by
using the ICAD844 and CA3080 to construct the DXCCTA for filter operation.
2 Circuit Description
A DXCCTA active block comprises a combination of DXCCII and an OTA. The

schematic of DXCCTA and its internal MOS-based design is shown in Fig. 1, where
the port characteristics can be mathematically defined as:
IY = 0, I X ± = I Z ± , VX ± = ±VY , I O± = ±gm VZ− (1)
where gm corresponds transconductance of DXCCTA and expressed as:

W
gm = μn C O X IB (2)
L
The term μn, C OX and W /L are the conventional parameters of a MOSFET and
I B be the input bias current.
A simple filter of the proposed SIMO filter is shown in Fig. 2. Where Z 1 , Z 2 and
An MOS-C Multifunction Filter Employing DXCCTA … 135
IB
Y Z+
IZ+
IY Z-
DXCCTA IZ-
O+
IO+
X+ X- O-
IO-
IX+
IX-
(a)
(b)
Fig. 1 DXCCTA a Schematic symbol, b internal CMOS circuit
Z 3 represent the impedance term. Z 1 and Z 3 be the parallel combination of the active
MOS resistor and capacitor and Z 2 be an active MOS resistor.
⎫
1 RMOS1 ⎪
Z 1 = RMOS1
sC = ⎪
⎪
⎪
1 1 + sC 1 MOS1 ⎪
R ⎬
Z 2 = RMOS2 (3)
⎪
⎪
1 RMOS3 ⎪
⎪
Z 3 = RMOS3 = ⎪
⎭
sC 1 + sC R
3 3 MOS3
The MOS resistor exhibits a numerical resistance value in the saturation region
as:
1
RMOS = W (4)
2μn C O X L
(VGS − VTh )
Fig. 2 Proposed MOS-C

current-mode multifunction
filter
O+
Y DXCCTA
IIN IHP
Z+
Z-
Z1 M3 C1 IBP
X+ X- O-
M1
C3
ILP
Z3
M2
Z2
A simple routine analysis gives the transfer functions for LP, BP and HP as:
gm
ILP RMOS1 RMOS2 RMOS3 C1 C3
TLP (s) = = 2 (5)
IIN s + gm
RMOS2 C3
s + RMOS1 RMOS2gmRMOS3 C1 C3
gm
IBP RMOS2 C3
s
TBP (s) = = 2 (6)
IIN s + gm
RMOS2 C3
s + gm
RMOS1 RMOS2 RMOS3 C1 C3
IHP s2
THP (s) = = 2 (7)
IIN s + gm
RMOS2 C3
s + gm
The above equations of the filter function give the pole frequency (ω0 ) and quality
factor (Q0 ) as:

gm
ω0 = (8)

RMOS2 C3
Q0 = (9)
gm RMOS1 RMOS3 C1
Here, the performance of (ω0 ) and (Q0 ) can be electronically tunable with a bias
current that tune gm and bias voltage for MOS resistors.
3 Non-Ideal Study
To study the non-ideal analysis, we have examined the transfer function using non-
ideal characteristic of DXCCTA. In non-ideal characteristic condition, DXCCTA
port relation is express as [27]
IY = 0, I X ± = α I Z ± , VX = ±βVY , I O± = ±γ gm VZ − ; (10)
where (α, β and γ ) be the current transfer gain, non-ideal voltage-transfer gains and
transconductance inaccuracies present in the DXCCTA, respectively. By using non-
ideal parameters (α, β and γ ), the current-mode LP, BP and HP transfer function of
the multifunction filters are observed as:
αβγ gm
ILP RMOS1 RMOS2 RMOS3 C1 C3
TLP (s)|α,β,γ = = (11)
IIN α,β,γ s2 + γ gm
RMOS2 C3
αβγ gm
s + RMOS1 RMOS2 RMOS3 C1 C3
αγ gm
IBP RMOS2 C3
s
TBP (s)|α,β,γ = = (12)
RMOS2 C3
s + αβγ gm

IHP αs 2
THP (s)|α,β,γ = = (13)
RMOS2 C3
s + αβγ gm
In this case, (ω0 ) and (Q0 ) becomes

αβγ gm
ω0 |α,β,γ = (14)

αβγ RMOS2 C3
Q 0 |α,β,γ = (15)
γ gm RMOS1 RMOS3 C1
The influence of non-ideal parameter is reflected in (ω0 ) and (Q0 ). Moreover, the
active and passive sensitivity of (ω0 ) and (Q0 ) are observed as:
1 ω0 1 ⎫
ω0 ⎪
Sα,β,γ ,gm = , S RMOS1 ,RMOS2 ,RMOS3 ,C1 ,C3 = − ⎬
2 2 (16)
1 1 ⎪
Q0
Sα,β,R MOS2 ,C 3
= , SgQm0,RMOS1 ,RMOS3 ,C1 = − , SγQ 0 = 0⎭
2 2
The sensitivity analysis (16) shows a low value that corresponds to good
performances for filter design.
To confirm the theoretical analysis, both simulation and experimental tests are
performed for a multifunction filter. The first section of validation is done through
PSPICE simulation in which DXCCTA is integrated with 0.18 µm CMOS TSMC
parameters with supply voltage ±1.25 V and bias voltage 0.43 V.
Multifunction filter frequency response for LP, BP and HP is observed for a very
high frequency in the range of 100 MHz. The active resistor have 1 K reactance
value and traditional capacitance with 1 pF exhibits ω0 of 177 MHz with a response
in Fig. 3. The filter frequency response of BP filter is below 0 dB which can be
improved by varying the bias current I B values which gives freedom for independent
gain control behaviour. The gain variation with different I B values is well observed
in Fig. 4 for the BP filter. Also, an experimental test for DXCCTA blocks is realized
with CFOA (ICAD844) and OTA (CA3080) as shown in Fig. 5.
A time-domain verification for filter is performed for ω0 100 MHz by selecting
the components values for the LP filter as RMOS1 = RMOS2 = RMOS3 = 2 K and
C1 = C2 = 1 pF. As we have designed for 100 MHz cut off frequency, the filter
output can pass the input signal which is less than 100 MHz frequency in LP filter.
Figure 6 shows the experimental result of a LP transient response for 50 MHz input
supply. For LP filter the input and output waveform will be in phase as shown in
Fig. 6.
Fig. 3 Proposed CM-MF frequency response

Fig. 4 Gain variation in BP with different bias current I O
Fig. 5 DXCCTA using ICAD844 and CA3080 for experimental test. a Circuit diagram, b physical
experiment set up
Fig. 6 Experimental LP
transient response for
50 MHz input
5 Conclusion
This article brings a high-frequency second-order filter using MOS resistors and
capacitors. The behaviour of the proposed design is simulated with the PSPICE simu-
lation by utilizing TSMC 0.18 µm technology and experimentally performed using
off the shelf ICs. The influence of non-ideal behaviour is also observed. The results
follow a close agreement with the theoretical prediction. Some useful characteristics
of the proposed filter are as follows:
1. Use of single DXCCTA.
2. With less passive components.
3. High-frequency response in 100 MHz.
4. Provides electronic tunability by using bias current.
5. No need to disturb the input signal for different output responses.
6. Suitable for LP, BP and HP filters.
Acknowledgements This work is supported by the Third Phase of Technical Education Quality
Improvement Program (TEQIP-III) under the Collaborative Research Scheme (CRS ID: 1-
5728068886), National Project Implementation Unit (NPIU), a unit of Ministry of Human Resource
Development (MHRD), Government of India for implementation of World Bank assisted projects
in technical education.
References
1. Cicekoglu O (1999) Multifunction filters using three current conveyors. Microelectron J 30:15–
18
2. Chen HP (2014) Voltage-mode multifunction biquadratic filter with one input and six outputs
using two ICCIIs. Sci World J (2014)
3. Abaci A, Yuce E (2016) Second-order voltage-mode universal filters using two DVCCs, two
grounded capacitors and four resistors. J Circ Syst Comput 25(12)
4. Ozcan S, Kuntman H, Cicekoglu O (2000) A novel multi-input single-output filter with reduced
number of passive elements using single current conveyor. In: Proceedings of the 43rd IEEE
midwest symposium on circuits and systems, vol 3, pp 1030–1032
5. Tomar RS, Singh SV, Chauhan DS (2013) Current processing current tunable universal biquad
filter employing two CCTAs and two grounded capacitors. Circ Syst 4:443–450
6. Ranjan A, Paul SK (2011) Voltage mode universal biquad using CCCII. Active Passive Electron
Compon
7. Ranjan A, Perumalla S, Kumar R, John V, Yumnam S (2019) Second order universal filter
using four terminal floating nullor (FTFN). J Circ Syst Comput 28
8. Shah NA, Quadri M, Iqbal SZ (2007) CDTA based universal transadmittance filter. Analog
Integr Circ Signal Process 52:65–69
9. Uttaphut P (2018) Single VDCC-based electronically tunable voltage-mode second order
universal filter. Przeglad Elektrotechniczny 1(4):24–27
10. Tangsrirat W (2007) Current-tunable current-mode multifunction filter based on dual-output
current-controlled conveyors. AEU-Int J Electron Commun 61:528–533
11. Abaci A, Yuce E (2017) A new DVCC+ based second-order current-mode universal filter
consisting of only grounded capacitors. J Circ Sys Comput 26
12. Horng JW, Jhao ZY (2013) Voltage-mode universal biquadratic filter using single DVCC.
Hindawi Publishing Corporation, ISRN Electronics, London
13. Abuelmaatti AT (2000) Universal current-mode filter using single four-terminal floating nullor.
Microelectron J 31(2):123–127
14. Joshi M, Bhatt V, Ranjan A, Benjwal P (2017) Realization of Colpitts oscillator using second
generation current controlled current conveyor. In: The second international conference on
research in intelligent and computing in engineering, pp 49–52
15. Kumari A, Ranjan A (2018) Realization of extra-X current controlled current conveyor based
resistorless quadrature sinusoidal oscillator. Int J Appl Eng Res 13:80–84
16. Ghosh M, Bhattacharya S, Ranjan A, Paul SK (2011) CMOS sinusoidal oscillator using current
conveyor. In: Proceedings of international colloquiums on computer electronics, electrical,
mechanical and civil, Kerala, India
17. Ranjan A, Pamu H, Tarunkumar H (2018) A novel Schmitt trigger and its application using a
single FTFN. Analog Integr Circ Signal Process 1–13
18. Pamu H, Ranjan A, Tarunkumar H (2018) Realization of schmitt trigger using IC AD844 based
differential voltage current conveyor transconductance amplifier. In: Proceedings of interna-
tional conference on innovations in electrical, electronics, power, smart grids and advanced
computing technologies, 9–10 Mar 2018
19. Joshi M, Ranjan A (2019) An autonomous chaotic and hyperchaotic oscillator using OTRA.
Analog Integr Circ Sig Process 101:401–413
20. Joshi M, Ranjan A (2018) Realization of multi scroll 2D chaotic oscillator using DVCC. In:
Lecture notes in electrical engineering, vol 553, pp 1093–1101
21. Kumar N, Vista J, Ranjan A (2019) A tuneable active inductor employing DXCCTA: grounded
and floating operation. Microelectron J 90:1–11
22. Sedra A, Smith KC (1970) A second-generation current conveyor and its applications. IEEE
Trans Circ Theory 17:132–134
23. Toumazou C, Lidgey FJ, Haigh DG (1993) Analogue IC design: the current-mode approach.
In: IEEE circuit and system series, vol 2. Peter Peregrinus Ltd., London, UK, pp 1–666
24. Bhatt V, Benjwal P, Joshi M (2017) OTRA based second order universal filter and its optimiza-
tion like Butterworth, Chebyshev and Bessel. In: Second International conference on research
in intelligent and computing in engineering, vol 10, pp 143–150
25. Pandey N, Paul SK (2011) Differential difference current conveyor transconductance amplifier:
a new analog building block for signal processing. J Electr Comput Eng
26. Pandey N, Paul SK (2011) VM and CM universal filters based on single DVCCTA. Active
Passive Electron Compon
27. Chaturvedi B, Kumar A (2016) DXCCTA: a new active element. In: 1st IEEE international
conference on power electronics. intelligent control and energy systems
28. Tarunkumar H, Singh YS, Ranjan A (2019) An active inductor employing a new four terminal
floating nullor transconductance amplifier (FTFNTA). Int J Electron
Karam Bharat Singh was born in 1994. He received his Bach-

elor of Technology in 2017 from the Department of Electronics
and Communication Engineering, North Eastern Regional Insti-
tute of Science and Technology (NERIST), Nirjuli, India. His
research interests include Current-Mode Analog circuit and
filter design, Mixed-signal IC design, low power VLSI design,
etc. At present, He is pursuing Master of Technology in VLSI
and Embedded System from the Department of Electronics and
Communication Engineering, National Institute of Technology
Manipur, Imphal, India.
Manoj Joshi was born in 1989. He received his B.Tech. degree

in 2011 from the Department of Electronics and Instrumenta-
tion Engineering, DIT University, Dehradun, India, and M.Tech.
degree in 2014 from BTKIT, Dwarahat (Uttarakhand Technical
University, Dehradun). His research interests include current-
mode analog circuits and study of Chaos Theory, Chaotic and
Hyperchaotic oscillator, Fractional-Order Devices, Mixed-signal
IC design, filter designs. At present, He is pursuing a Ph.D. in
VLSI in Electronics and Communication Engineering Depart-
ment, National Institute of Technology Manipur, Imphal, India.
Ashish Ranjan was born in 1986. He received his Doctor of

Philosophy (Ph.D.) degree in 2013 from the Department of
Electronics Engineering, Indian Institute of Technology (Indian
School of Mines), Dhanbad, India. His research interests
include current-mode analog circuits, the study of fractional-
order circuit, chaotic oscillators, filter designs, etc. At present,
he is Assistant Professor in Electronics and Communication
Engineering Department, National Institute of Technology
Manipur, Imphal, India.
Smart Car Parking Technique
for Metropolitan City
Abu Shufian, Md. Asif Afroj, Md. Hasibuzzaman, Mehedi Al Miraz,

Avijit Sarker, Md. Jahidul Islam Rashed, and Md. Aman Miah
Abstract The objective of this proposed model is to build a smart car parking
system that will make the parking process easier and more effective for the users of
the parking area. Due to the expansion of using car traffic problems mostly occur
all over the world and to reduce the problem smart parking system can play and
sustainable and innovative role. In the traditional parking system, the parking area’s
maximum works are dependent on manpower in the past. But now, it will reduce
on a high scale and make the whole process automated by the bless of the smart
system. A fire alarm process in the parking area makes this system more reliable
and safer. Also, the smart car parking system provides a proper security system and
proper parking area detection which are useful and makes the parking process more
comfortable for the users.
Keywords Smart parking · Arduino · RFID sensor (tag, reader) · IR sensor ·

Flame sensor · Bill calculation and payment
1 Introduction
Science and technology mean innovation and it spreads its innovative manner also
in the transportation sector. It changes all the transportation infrastructure and now
in today’s world a huge number of vehicles produce all over the world. Increasing of
the rapid number of vehicles also produce transportation problem, which is normally
known as traffic problem all over the world. So, to get a release from that problem
A. Shufian (B) · Md. Hasibuzzaman · M. Al Miraz · A. Sarker · Md. J. I. Rashed · Md. A. Miah
Department of Electrical and Electronic Engineering, American International
University-Bangladesh, Dhaka 1229, Bangladesh
Md. Hasibuzzaman
Md. A. Afroj
Department of Computer Science and Engineering, United International University, Dhaka 1212,
Bangladesh
144 A. Shufian et al.
smart parking system is the most effective and coolest way to decrease the rate of
traffic jams as most of the traffic problems normally occurred because of the parking
car here and there. In this paper, all the features are combined which are gathered
from the previous thoughts. The display will show the proper direction in which way
the car has to move and it also shows the parking slot which is free and which is
occupied [1]. In the parking area, an automatic counter system is provided that can
measure the time and give a calculation of various charges for different kinds of cars
for their parking time. In the smart parking process in and out gate can easily control.
When the car comes to the gate sensor or RFID can use and when the authorized cars
are detected the gate is automatically open and the same process is to open the out
gate. Sometimes, it shows that the driver or the user parks the car and goes outside;
then, there is no need to park on the load. A sensor can easily do the work. When
there is a person, it detects the person and on the load like light, and when there is no
one, it automatically offs the load. A fire alarm process can introduce in the parking
area. If any fire accident occurs there, the sensor can easily detect this and the alarm
will ring. This process makes the parking area more secure. Security is an important
issue for car parking. Normally in the parking area, an authorized person can park.
So, to make the parking process more secure sensor or RFID for an authorized person
can introduce. So, not only the parking cars are safe and get proper security but also,
unauthorized cars are easily detected and cannot park. In the parking area, various
types of cars are parked and to depend on this the parking area can be divided into a
different zone. If a different zone can introduce, it will easy to know which types of
cars with their numbers are parked in a different parking zone.
2 Literature Review with In-Depth Investigation
The smart car parking system is a process in which the whole parking system is
developed with merging of various features that make the parking process more
innovative and comfortable and secure for the users [2]. The smart car parking system
which is mainly implemented in Japan, USA, and Europe is developed with various
research centers and technical institutions. This developing process was not started
in a single day. It was a long-time process which was developed day by day. In 1951
automatic parking system was opened in Washington D.C. They have implemented
some systems and they are still in use bowser, pigeon-hole systems, and roto park
systems. In 1961 auto stacker was introduced in London. Advanced systems were
implemented in South America and Asia in the 1970s. Japan has crossed all the
records to implement more than 100,000 parking spaces which is a year from the
1990s which are fully automated [3]. In Hoboken, New Jersey robotic parking garage
was implemented in 2002. In the recent 2015, few students of Jazan University
have developed a computerized vehicle parking management and monitoring model
which utilizes programmed number plate acknowledgment cameras to monitor the
parking facilities of the University [4]. In 2018, a prototype car parking monitoring
system using a camera network was developed with the help of visual detection and
Smart Car Parking Technique for Metropolitan City 145
communication of parking information [5]. In the same year, some researcher has
developed a new localization approach that improved rear-view camera usage during
automatic parking. They proposed a vehicle localization process using an interacting
multiple model Kalman filter [6].
The smart parking system is considered the most beneficial, updated and innova-
tive process that makes the parking too much easier. It reduces the time and effort to
find out the parking slot. Also, it is helpful to reduce traffic jams. As in a smart parking
system, users can easily find out the parking place so it also places an important role
to decrease the emission of pollution. It also brings a revenue earning process for the
owner of the parking area. This system is fully automated that is why it reduces the
management cost and real-time data of users are stored in the system that makes the
smart car parking process more secure and beneficial for the users [7].
There are various methods of smart parking systems such as parking information
and guidance system, transit-based information system, smart payment system, an
automated parking system. Parking guidance and information system is a system in
which drivers can get proper information and direction sign or message sign to know
about the parking place [8]. Necessary data or messages of real-time information on
parking and departure times are provided in a transit-based information system [9].
The smart payment system is introduced to overcome the limitation of the accus-
tomed payment process by reconstruction of the payment method via parking meter.
It reduces the time limit of payment, maintenance, and staffing requirement. In the
e-parking process, it shows the availability of parking space and the user can easily
reserve place [10]. The main thought of this process is to ensure the vacant parking
facility for the users via SMS and the user can pay the bill by the smart payment
system. In the automatic parking system, the whole parking process is fully auto-
mated. Parking areas gate and other components are automatically worked through
RFID sensors and make parking easier to give proper direction in display [11, 12].
2.1 Aim of Smart Car Parking
This proposed system is fully automated by a microprocessor. Also, it provides proper

security protection with proper parking area detection which is useful and makes
the parking process more comfortable for the users. The smart parking system is a
suitable and efficient car parking system where cars are parked in a systematic and
proper way. In this system, parking areas decrease and also make parking process
easier and reduces parking time. A proper security system to detect authorized and
unauthorized cars is the added advantage of the smart car parking system. There are
some extra features like parking slot occupancy detection with a display, automatic
billing process for parking of public cars, automatic load control and fire alarm
system can add to make the parking system more effective. It can be introduced in
public places such as educational institutions, offices, and shopping malls and so
on with proper authorization. As a result, the parking process becomes more secure
and it will also contribute a great part to reduce the traffic problem because in this
system the owner cannot park a car here and there. The smart car parking system is
an adjustment of mechanical, electrical and control devices which can assure parking
a large number of cars in a little space in a systematic and automatic way with proper
direction.
3 Methodology and Designing of Smart Car Parking
The whole process was divided into various blocks and merged them to get the desired
output. Two types of the process were used to fulfill that project. One process was
the wireless connection and another was wired connection. In wireless process RFID
tag and RFID reader were used and in wired connection was between Arduino, LCD,
Flame sensor and other equipment. From the RFID tag information will send to the
RFID reader and from RFID reader, Arduino will get the information. For the valid
information, LCD will show information that had given to it from Arduino and the
gate will be opened. At the same time, Arduino would get information from the IR
Sensor about the vacancy of place and show the information in the LCD display. If
any fire accident occurs in the parking place Arduino will get the information from
the Flame sensor and the buzzer will ring. When the car will go out from the parking
place again the RFID tag will have to punch and Arduino will get information from
the RFID reader. Then parking time and money that have to pay to the users will
show in the display. When the user will pay the money, Arduino gets the information
and the gate will be opened (Figs. 1 and 2).
In smart parking system, the parking space is divided into a different slot and in
every slot, IR sensor is used and the main purpose of IR sensor main is to identify
if there any vacancy or not in parking space and an LCD display proper information
can be seen that would give from the IR sensor. Suppose the parking area is divided
into 6 slots and in every slot, the car was parked. So, LCD will show there is no
Motor IR Sensor Buzzer
RFID Arduino
RFID tag reader Flame Sensor
LED
Power Supply
Fig. 1 Block diagram of the smart car parking system

Fig. 2 Different slot in the

parking area
vacancy and parking is filled so extra no car can enter the parking area. Suppose
there is a vacancy in 5 and 6 slots. So, LCD will show the vacancy of 5 and 6 slots
and also give a proper direction of parking space thus a user can easily park their car
(Fig. 3).
First, the LED display was connected to the Arduino and controlled it by Arduino.
The input power was given to Arduino and RFID tag was used to control the gate.
Here, a fire sensor in Arduino input and a buzzer was used in Arduino output from
which a signal was found if there would any fire accident occur in the parking area.
There was an IR sensor that was used to detect the occupancy of a parking slot and
in LED display the user would easily see is there any free space or not in every slot.
Bill calculation process was also set up in Arduino from which the user would see
the amount of money in LED that would be paid to the owner of the parking area for
parking the car. In block diagram whole process was given with proper connection
and description thus how the system was going to be worked could easily understand
for people Block diagram and the flowchart of the smart car parking system introduce
the whole thought of this project in a shortage way. Step by step a process of this
project and the working system is shown in the block diagram properly.
Start
NO Display
Vacancy Parking Full
YES
Bill Calculation
Check RFID and Payment
NO Display
Authorized
Unauthorized
YES
Display Gate Open
Thank You Check Slot
Parking
Go Outside
Stop
Fig. 3 Flowchart of the smart car parking system
4 Simulation and Structural Result of Smart Car Parking

System
The simulation was the best way to find out the result in real-life circuits. For result
analysis, Proteus 8 professional simulator was used. Figure 4 shows the schematic
diagram of the proposed system where LCD display was attached to Arduino Uno.
The RFID reader and Flame sensor were simulated by Proteus 8 and replaced in
different combination pin number of Arduino through the wire. Red, yellow different
color LED was used to show the variety of signaling different sectors.
Gate open section was shown in Fig. 4a. When a user would use the ID card the
Arduino would justify the card pin number and give a result that the user was allowed
Fig. 4 a Car detect and bar open, b free slots in LCD display, c fire signal in LCD display, d car
detect and bar open, e free slots in LCD display
or not. If the card was valid, the signal from Arduino would make the servo motor
turn into the ninety-degree angle and the bar would open and the allowed users would
go inside to the parking area and after some delay, the bar was automatically closed.
The car parking free slots were shown in Fig. 4b. When valid user ID was punched
then yellow LED was turned on and the gate was opened and user would go to the
parking area.
In Fig. 4c, simulation was used for fire alarm purpose. The main purpose of the
Flame sensor is to detect the fire accident in the parking area was connected to the
Arduino and if any fire occurrence happened in parking area the occurrence was
detected by Flame sensor and from the Arduino the result would go to buzzer and
buzzer would make sound, Red LED would turn On and also in LCD display the fire
signal was showed to make aware of the fire accident to its users.
In Fig. 4d, the free and booked slot for parking system was shown. In the display,
there were six slots. In parking area, IR sensor was used to sense the vehicles and
the result was justified in Arduino and an LCD display user would see the result that
which slot was free and which slot was booked. In Fig. 4e, simulation was shown
about the outgoing process of users from the parking area. When the user would want
to go outside, the ID card was punched by the users and from the bill calculation
process which was already installed in Arduino showed the amount of money in LCD
display after calculating the parking time and money for per hour and user had to
pay the money for parking the car and after receiving the money the bar was opened
and user would go outside. After some delay, the bar was closed automatically.
In simulation part, the whole result was shown from which a total idea of the
system which was merged with different parts such as gate control, free and book
slots detection, bill calculation and fire system can easily understand. In the hardware
side, all the features were merged and a proper result of smart car parking system
was shown.
The first stage of the hardware system was shown in Fig. 5a. Here, a card reader,
servo motor, yellow LED and buzzer was connected to the Arduino according to
the Arduino code pin number through the wire and it was the initial stage where no
RFID was used to see the valid or invalid user ID and the gate was at its initial stage
that means bar was closed. Figure 5b showed the validation process of the user ID
card. When the valid ID was punched to the reader the information would go to the
Fig. 5 a First stage of the smart car parking system, b valid RFID check, c invalid user ID check,
d gate open, e free slots in LCD display, f free and booked slots in LCD display, g booked slots in
LCD display, h car detection and bar open in LCD display, i estimated bill and remaining balance
calculation, j invalid RFID check, k, l full hardware structure of the smart car parking system
Arduino and after checking the validity of the user ID the signal would go to the
yellow LED and LED would turn ON. Figure 5c was about the invalid ID users. If any
invalid id was punched in the RFID reader, the reader would detect the invalidity of
the ID and the bar would not open and the red LED was also turned ON. After getting
the valid ID signal from Arduino the gate would move into the ninety-degree angle
and the car was allowed to go inside the parking area and in Fig. 5d the process was
shown. When users have punched the ID, the validity was checked by the IR reader.
After checking the validity user ID, the bar will be opened and it would be shown
on display in Fig. 5d. When the valid ID was punched by the user, at the entrance
user would see the LCD display where free or book slot as shown in Fig. 5e was
usually shown in the free slot. That means every parking slot was empty and in any
parking slot, the car would be parked. In Fig. 5f where some free and booked slots
were shown. When the free slots were booked by the car, the IR sensor would detect
it and the signal was sent to the Arduino and after justifying the signal in display,
free and booked slot were shown. In Fig. 5g the booked slots are discussed. When
all the slots were booked by the cars that were shown in the LCD display.
When the user would want to go outside from the parking area user had to punch
the ID again and a particular amount of money was deducted from the user ID account
and after deduction, the bar would open and the user would go outside. This part was
shown in Fig. 5i. Figure 5j was about invalid ID users. If any invalid id was punched
in the RFID reader, the reader would detect the invalidity of the ID and the bar would
not open and the red LED was also turned ON.
4.1 Calculation
There is also a payment issue that users have to pay for using the parking space. The
owner of the parking space may fix a value for hourly parking suppose it will be $2
for one-hour parking. So, if the user parks the car for 2 h the user has to pay.
Per hour parking cost × 2 h = $2 × 2 = $4.
So, when the user will want to go out from the parking area first user has to pay
$4 and the gate will automatically open thus a user can go outside.
An important part of a project has measured the cost. To start any project first
rudiment was to select the equipment that would use in the project and calculate the
cost of this equipment. That’s why calculation was done and the total cost of the
project and the price of all components were given in Tables 1 and 2.
Table 1 Cost estimation

Components Quantity Price (US dollar)
Arduino Mega 2560 01 10.6
LCD display 16 × 2 01 1.95
RFID reader 01 4.26
RFID tag 01 0.36
Flame sensor 01 9.20
Relay 03 2.56
Servo motor 01 2.20
Volume pot resistor 01 0.15
LED (pin type) 08 0.20
Buzzer 01 0.18
Miscellaneous 5.48
Total cost 37.15$
Table 2 Compare with traditional models (list of car parking system and comparison table of car
parking system)
Present all the method Our proposed car Traditional car parking Smart car parking
of car parking parking method system system
Transit-based Smart car parking Dependent on Whole process
information system, system manpower [4] automated
automated parking, Cannot detect Detect authorized and
E-parking, parking authorized and unauthorized cars
guidance and unauthorized cars [9]
information system,
smart payment Manually check free Proper parking area
system, parking lot, space for parking [10, detects automatically
parking garages, 13]
carports, automated Cost and time Automatically
parking system, calculation process are calculate and give
semi-automated by hand [2] result instantly
parking system [3] Non-smart and manual Makes parking process
process [5] easier and reduces time
Some features like by Extra feature like
hand gate controlling, automatic billing
manual load control [8, process, automatic load
13] control, and fire alarm
system
Security system is weak Provides a proper
[6] security system
5 Limitations and Future Enhancements
In this research work, the main focus was concerned about the development of the
smart parking system. But there were some limitations also. There only one gate
was used to gate in and out from the parking area. A problem was created for that
when two users wanted to go and in at a time. In future improvements were carried
out with this proposed model. Only six slots were shown in this model and all the
components were accurately working with user demand. Sensors which were used to
sense the car and in the display, free or booked slots were shown. For a large parking
area, this model could be spread to fulfill the user demand in the future. Some of the
techniques in which the circuit could be modified were such as, making the system
with IoT and Wi-Fi based. Addition of more IR sensors for more car parking slots.
Advanced parking slot booking system by SMS from mobile.
6 Conclusion
This designed framework is coordinated by Arduino, RFID tag, RFID reader, LCD
display, servo motor, LED, buzzer, IR sensor, Flame sensor which make the parking
system more effective and useful for the users. This design of the smart parking
system will give proper direction, more security and provide an easy way of parking
to its users and also it will play an important role to reduce the traffic problem of a
country. On the other hand, this was an Arduino based project which would use in
large areas like corporate offices, institutions, shopping mall or residential area to
make the parking system easier. It would user-friendly and also the most effective.
It would make the parking area more secure and could give a fire alarm to detect the
fire with the flame sensor. For this purpose, the gate would automatically turn on and
off with the proper validation of the user’s ID card. So, it was not easy to make any
violence in the parking area because it could be easily detected. So, a whole security
system was created in the parking area through the project. The LCD display was
used in the parking area to give the proper instruction about the parking slots that
would make the parking process easier for the users because users would not have to
search the parking slot in the large parking place. And the owner of the parking area
would also earn money from the parking area easily. So, it was overall a friendly and
effective prototype model for the car users to park the car easily and safely.
References
1. Lu R, Lin X et al (2009) SPARK—a new VANET-based smart parking scheme for large parking
lots. In: IEEE conference on computer communications. https://doi.org/10.1109/infcom.2009.
5062057
2. Yan G, Yang W et al (2011) Smart parking: a secure and intelligent parking system. IEEE Intell
Transp Syst Mag 3(1):18–30
3. Pojani D, Stead D (2015) Sustainable urban transport in the developing world: beyond
megacities. Sustainability 7(6):7784–7805
4. Aalsalem M et al (2015) An automated vehicle parking monitoring and management system
using ANPR cameras. In: Conference on advanced communication technology. https://doi.org/
10.1109/icact.2015.7224887
5. Hammoudi K et al (2018) Towards a model of car parking assistance system using camera
networks: slot analysis and communication management. In: IEEE international conference on
smart city. https://doi.org/10.1109/hpcc/smartcity/dss.2018.00210
6. Lee S et al (2018) Localization for automatic parking system using interacting multiple model
Kalman filter with rear-view camera. In: International conference on control, automation and
systems. Available: INSPEC accession number: 18345640
7. Roßnagel A (2019) Protection of fundamental rights in the smart car. In: Communication,
security and data protection in the networked vehicle, p 84. https://doi.org/10.1007/978-3-658-
26945-6
8. Geng Y et al (2013) New “smart parking” system based on resource allocation and reservations.
IEEE Trans Intell Transp Syst 14(3):1129–1139
9. Rodier C, Shaheen S (2010) Transit-based smart parking: an evaluation of the San Francisco
Bay area field test. Transp Res Part C Emerg Technol 18(2):225–233
10. Wang H, He W (2011) A reservation-based smart parking system. In: IEEE conference on
computer communications workshops. https://doi.org/10.1109/infcomw.2011.5928901
11. Wei L et al (2012) Design & implementation of smart parking management system based
on RFID & internet. In: International conference on control engineering & communication
technology. https://doi.org/10.1109/iccect.2012.12
12. Khan S et al (2019) A smart intruder alert system based on microprocessor and motion detection.
In: International conference on robotics, electrical and signal processing techniques (ICREST).
https://doi.org/10.1109/icrest.2019.8644109
13. Barno M et al (2019) An android subjugated compact self-driving semi-autonomous vehicle.
In: IEEE region 10 symposium (TENSYMP). https://doi.org/10.1109/tensymp46218.2019.897
1224
Implementation of Perturbation-Based
MPPT Technique Using Model-Based
Design
Sachin Angadi , Udaykumar R. Yaragatti , Yellasiri Suresh ,

and A. B. Raju
Abstract Maximum Power Point Tracking (MPPT) algorithms are imperative in

solar Photovoltaic (PV) systems. This paper presents a simple approach for real-time
implementation of perturbation-based MPPT algorithms employing model-based
design (MBD) and Rapid Control Prototyping (RCP). To validate the proposed
approach, classic Perturb and Observe (P&O) algorithm is tested. Altair’s Embed
software and low-cost Texas Instruments (TI) DSP controller is used for implemen-
tation. All the fundamental blocks required for the implementation of perturbation-
based MPPT algorithms are discussed in detail. Hardware-in-Loop (HIL) simulation
using Altair Embed for selection of appropriate sampling time and perturbation step
size is demonstrated. The readers will also be presented with details of fundamental
fixed point blocks of Altair Embed software for the implementation of other MPPT
algorithms using low-cost controllers.
Keywords MPPT · Model-based design · Perturb and observe · Rapid control

prototyping · Solar PV system
1 Introduction
The need for energy is continuously increasing owing to the lifestyle of the millen-
nium and growing population. Exploring an alternative source of energy is the urgent
demand of the hour for a sustainable future. The solar PV source is seen as one of the
potential alternatives to meet the gap between energy demand and generation. India
is bestowed with solar radiation capable of generating 5000 trillion kWh/year [1].
Several papers have discussed various MPPT algorithms with the aid of simu-
lation results; however, developing code for a digital computer is still a challenge
for many core engineers. MBD, a new approach for embedded development, is fast
S. Angadi (B) · A. B. Raju

KLE Technological University, Hubballi, India
U. R. Yaragatti · Y. Suresh
National Institute of Technology Karnataka, Surathkal, Surathkal, India
156 S. Angadi et al.
emerging as alternative for code development. MBD is a model-centric approach to

the development of controls for a dynamic system. It facilitates system-level design,
RCP, automatic code generation and accelerated testing and verification [2]. Given
preliminary simulation studies and design specifications, RCP and HIL testing can
be performed on the system under test. If the simulation studies reveal substantial
details about the system, then the model can be directly used for production code
generation [3].
The rapid growth in MBD and rapid prototyping technology over the last few
decades has motivated core designers to adapt MBD to develop code for their algo-
rithms. Also, new MBD platforms like dSPACE, OPAL-RT are fast emerging to
promote MBD and RCP to reduce the time for product development. MBD requires
a hardware platform and dedicated software for performing necessary functions. The
work of selected research articles using MBD for power electronics and MPPT is
consolidated in Table 1. MATLAB-Simulink with embedded coder toolbox is the
most commonly used host PC software with a wide variety of supporting hardware
platforms like OPAL-RT, FPGA, DSP and microcontroller [4–10]. Few researchers
have also explored LabVIEW for the implementation of MPPT algorithms using
different compatible hardware platforms [11–13]. It is to be noted that backend
software compatible with the chosen hardware platform is required to compile c-
code generated by the embedded coder. In addition to automatic code generation,
both MATLAB-Simulink and LabVIEW facilitate HIL simulation. However, only
LabVIEW supports interactive HIL simulation. The development using LabVIEW
is not viable for commercial application as the code generation is not available for
low-cost microcontrollers.
Applications like MPPT requires real-time tuning of parameters like sampling
time and perturbation step size for optimal performance. This paper presents a simple
approach for RCP of perturbation-based MPPT algorithms. The proposed method
uses a Altair Embed software, TI TMS320F28027 controller and Code Composer
Table 1 Summary of software and hardware platform used for embedded controller development
employing MBD
References Host PC software Hardware platform Backend software Application
[4] MATLAB-Simulink FPGA HDL Inverter
[5] MATLAB-Simulink TMS320F28335 CC studio Power electronics
[6] MATLAB-Simulink Spartan 3E (FPGA) Xilinx P&O (MPPT)
[7] MATLAB-Simulink STM32F4 board ST-LINK MPPT
[8] MATLAB-Simulink OPAL-RT OP4500 RT-Lab Solar PV system
[9] MATLAB-Simulink STM32F429 board ST-LINK MPPT
[10] MATLAB-Simulink Arduino MEGA – Fuzzy MPPT
[11] LabVIEW MyRIO – Fuzzy MPPT
[12] LabVIEW cRIO-9075 kit – P&O (MPPT)
[13] LabVIEW NIcDAQ-9178 – MPPT
Implementation of Perturbation-Based MPPT Technique Using Model-Based … 157
Studio for the execution of MBD, automatic ‘C’-code generation and interactive
HIL simulation for controller tuning and development. Blocks like PWM, ADC,
unit delay, sample and hold, and digital filter required for the implementation of
classic perturbation algorithms are discussed in detail. Further, the interactive HIL
simulation feature of Altair Embed is explored to demonstrate live tuning of sample
time and perturbation step size for opt mal performance of MPPT algorithm.
2 Model-Based Design Framework
In this section, the system considered for the test using MBD is discussed. Also, the
framework of the MBD using Altair Embed for ‘C’-code generation, interactive HIL
simulation and data acquisition and monitoring is presented in detail.
2.1 System Description
The details of system implementation are shown in Fig. 1. The system consists of the
solar PV source, DC-DC boost converter feeding resistive load, a TI TMS320F28027
controller and Altair Embed software. The communication link is established
between the host computer and the controller using JTAG link at the rate of 100 Hz
PV Source Boost Converter Load

L
Ipv
+ +
Vpv Cpv Vdc
Cdc R
S
− −
Altair Ts Ipv
MPPT +
Embed
ΔD Algorithm Vpv
− S
JTAG
DSP TM320F28027
Interactive gains
D, Vpv and Ipv ADC inputs
Model Based Design Framework Data acquisition
Fig. 1 Block diagram representation of the system under test

to facilitate RCP. System variables namely PV voltage (V pv ) and current (I pv ) are

sensed and instantaneous power is calculated. The MPPT algorithm is processed to
generate appropriate duty ratio (D) to ensure maximum power extraction from solar
PV source.
2.2 MBD Using Altair Embed
The framework of MBD for controller design using Altair Embed is shown in Fig. 2.
The MBD software and the backend software (Code Composer Studio) are installed
on Host PC. A communication link is established between the target hardware
(TMS320F28027) and Altair Embed using the JTAG communication protocol. The
controller design and specifications are finalized before deployment using prelimi-
nary simulation studies. The MBD approach for controller design using Altair Embed
involves following steps:
1. The designed controller is drawn as model using fixed point blocks in Altair
Embed and is checked for logical, syntax error or any overflows.
2. The Altair Embed tool generates the c-code compatible with chosen target
hardware.
3. The generated c-code is compiled using Code Composer Studio to generate .out
file.
4. The .out file can either be used for interactive HIL simulation or can be burnt
into flash of the target hardware for final deployment.
Automatic c-code generation, Compilation,

Linking and downloading
sT-Embed on Host PC TMS320F28027
Sensing analog signals, PWM generation &

interactive simulation
sT-Embed Target Interface sT-Embed
block downloads and monitors generated
code running on Target .out file &
RTOS PLANT
(TMS320F28027)
Target Hardware PV Source fed

(TMS320F28027) Boost Converter
PWM (Duty ratio)
Interactive data ADC (Vpv & Ipv )
Exchange (Ts & dD)
sT-Embed blocks for:
@ 100Hz
• Interactive gains
(Sampling time and
perturbation size inputs)
• Sinks for recording
DSP response
Fig. 2 Model based design framework for rapid prototyping with interacting HIL simulation
3 Rapid Control Prototyping
The classic P&O algorithm is presented in this section along with its Simulink and
fixed point Altair Embed implementation. Also, the interactive HIL simulation using
Altair Embed and TMS320F28027 is discussed in this section.
3.1 Perturbation-Based MPPT Techniques
Figure 3 shows the implantation details of P&O algorithm for Boost DC-DC
converter. Direct duty ratio implementation is chosen for improved stability [14].
The Simulink implementation of the algorithm reveals the requirement of sample
and hold, multiplier, unit delay, sign detector and saturation blocks for implemen-
tation of the algorithm. All the aforementioned blocks are available in fixed point
block-set of Embed tool for implementation of the algorithm. The implementation
of the model using fixed point blocks facilitates code generation for fixed point
DSP/microcontrollers and reduced the memory utilization of the controller. The fixed
Start
Measure V[K],I[K]
Algorithm
Compute P[K],V[K-1],I[K-1],P[K-1]
dP = P[K]-P[KI1], dV = V[K]-V[K-1]
dP>0
dV>0 dV>0
D[k] = D[K-1] - dD D[k] = D[K-1] - dD D[k] = D[K-1] - dD D[k] = D[K-1] - dD
End
I I[K] P[K] dP Sign(dP) Simulink implementation of the algorithm

1 +
-
D
+ -
2 + 2
V V[K] - 0.04
dV Sign(dV) dD
Fig. 3 Direct duty ratio based P&O algorithm and its Simulink implementation
[email protected] %Duty Cycle A(1.16)

%Duty Cycle B(1.16) F28027-EPWM1A/EPWM1B
UD
z+1 V[K]
F28027-ADCRESULT0 c 0.000314 c S&H c
z-0.99937
P[K]
z+1 *
F28027-ADCRESULT1 c 0.000314 c S&H c 4.16
z-0.99937
ADC V[K] + 4.16

1/Z - [email protected] Vsign
Read
P[K] + 4.16
1/Z - [email protected] Psign
Ts UD Unit delay block used in 1/Z block

UD
c Convert block used for datatype conversion
Vsign
* [email protected] *
Psign 4.16 4.16 c %Duty Cycle A(1.16)
ΔD %Duty Cycle B(1.16) F28027-EPWM2A/EPWM2B
Fig. 4 Model based implementation of P&O algorithm using Altair Embed for TMS320F28027
point model developed using Altair Embed tool for TMS320F28027 is as shown in
Fig. 4.
3.2 Hardware-in-Loop Simulation for RCP
Altair Embed is a user-friendly tool for developing embedded systems. It automat-

ically generates a ANCI ‘C’-code compatible with the low-cost controller [15]. In
addition to the code generation, interactive HIL simulation facilitates live parameter
tuning and real-time data monitoring without data acquisition system. The software
supports wide variety of low-cost controllers [16]. TI TMS320F28027 controller is
chosen for study in this paper.
Figure 5 shows two files namely source file and the debug file. The source file
consists of the detailed block-level implementation of the algorithm shown in Fig. 4.
The source file is designed to facilitate interactive HIL simulation with real-time live
Fig. 5 Interactive HIL simulation setup in Altair Embed using TMS320F28027

data monitoring. The T s and D are chosen as two interactive inputs while D, V pv
and I pv acquired in real-time for data monitoring. The ANCI ‘C’-code is generated
using source file and compiled at the backend using Code Composer Studio software
to generate an executable .out file. A separate debug file is created consisting of target
interface block, interactive gains and display units for real-time data monitoring. The
debug file allows execution of the .out in real-time and displays the selected data for
monitoring in real-time at the rate of 100 Hz. The value of interactive gains can
be changed using the slider and the impact of the gains on the system under test is
monitored in the display units connected to the output of the debug file.
The details of the laboratory prototype developed for conceptual verification are
discussed in this section. Also, the experimental results of the interactive HIL simula-
tion studies are presented in detail to highlight the significance of the MBD approach
and RCP in the implementation of the MPPT algorithms. Further, the pointers are
given to the readers for the implementation of the other perturbation-based MPPT
techniques.
The laboratory prototype developed for conceptual validation is shown in Fig. 6. The
setup consists of a 1.5 kW Magna solar emulator, boost converter feeding a resistive
LEM LV-25p Altair’s

voltage transducer sTEmbed
TMS320F28027 LM7417 buffer circuit
High frequency
inductors
Magna Solar Emulator
LEM LA-25p
current transducer
Resistive
load
DC-DC Boost
converter
Fig. 6 Snapshot of laboratory prototype developed for conceptual validation

load, hall effect-based current and voltage sensors, TMS320F28027 controller and
host PC with Altair Embed for HIL simulation. The PV current and voltage are sensed
using LEM sensors and inbuilt ADC module of TMS320F28027 for processing of
MPPT algorithm. The PV current, voltage and duty ratio are acquired using the
Embed tool employing HIL simulation.
4.2 Results
Figure 7 shows the results obtained during the appropriate selection of T s . Initially,
the value of T s = 1 s larger than system time constant is selected to obtain stable
three-step output for the duty ratio around the maximum power point. Larger T s
increases the system response time, hence the value of T s is reduced in steps of 0.1.
It is observed that, stable output for the algorithm persists only up to value of T s
Ts = Sampling time in s
Ts =1
0.6 Ts =0.9 Ts =0.8
D
0.4 Ts =0.7 Ts =0.6 Ts =0.5 Ts =0.4

Ts =0.3 Ts =0.2
Ts =0.1
0
0 20 40 60 80 100 120 140
(a)
100
Vpv (V)
50
0
0 20 40 60 80 100 120 140
(b)
5
Ipv (A)
0
0 20 40 60 80 100 120 140
(c)
200
Ppv (W)
100
0
0 20 40 60 80 100 120 140
(d)
time(s)
Fig. 7 Experimental results of a duty ratio (D), b PV voltage (V pv ), c PV current (I pv ) and d PV

power (Ppv ) for varying sampling time (1–0.1 s)
D 0.6 Δ D = Perturbation step s ize
0.4 ΔD = 0.08 ΔD = 0.07 ΔD = 0.06

ΔD = 0.1 ΔD = 0.09 ΔD = 0.05 ΔD = 0.04 ΔD = 0.03
0 10 20 30 40 50 60 70 80 90 100
(a)
100
Vpv (V)
50
00 10 20 30 40 50 60 70 80 90 100
(b)
4
Ipv (A)
2
0 10 20 30 40 50 60 70 80 90 100
(c)
Ppv (W)
200
100
0
0 10 20 30 40 50 60 70 80 90 100
(d)
time(s)

power (Ppv ) for varying perturbation step (0.1–0.03)
= 0.5. The D, V pv , I pv start to exhibit larger perturbations leading to instability for

value of T s < 0.5 s.
The value of T s is chosen to be 0.5 s for the system under test. Another parameter,
which mainly decides the efficiency of the P&O algorithm, is the D. Figure 8
captures the response of the system for different values of D. Initially, the value
of D is chosen as 0.1. Further, the value of D is reduced in steps of 0.01 and
the response is captured in Fig. 8. For larger values of D, the perturbations in V pv
and I pv are larger leading to reduced power utilization. It is observed that the system
exhibits a stable three-step waveform for the value of D = 0.04, beyond which the
system tends to move towards instability, exhibiting random perturbations.
The values of two significant parameters of the P&O algorithm namely T s and D
are chosen with the aid of interactive HIL simulation. For these optimal parameter
values (T s = 0.5 s and D = 0.04), the response of the system for the varying load is
captured in Fig. 9. The system exhibits satisfactory performance for three different
values of the load hence proving the validity of the controller and chosen parameter
values.
0.6 R = 78 Ω
R = 58 Ω
D
0.4 R = 116 Ω
R = Load resistance R = 78Ω

0.2
0 10 20 30 40 50 60 70
(a)
100
Vpv (V)
50
0
0 10 20 30 40 50 60 70
(b)
3.5
Ipv (A)
3
2.5
0 10 20 30 40 50 60 70
(c)
Ppv (W)
200
100
0
0 10 20 30 40 50 60 70
(d)
(s)
time

power (Ppv ) for varying load resistor
4.3 Scope of the Study
In this paper, detailed implementation of the classic P&O algorithm for low-cost
controller is presented in detail using Altair Embed. Block-level requirement and
its fixed point implementation is conferred to the readers. The study can be further
extended to other perturbation-based MPPT algorithms with minimal modifications
[17–21]. Also, the impact of various parameters influencing the performance of the
algorithm can be easily analysed using the concept of interactive HIL simulation.
5 Conclusions
This paper has presented a simple approach for real-time implementation of the
P&O algorithm for low-cost controller TMS320F2027 using the concept of MBD
and RCP. The usage of the Altair Embed tool for MBD, RCP and HIL simula-
tion with interactive gains have been presented in detail. Results of the laboratory
prototype have been presented with the live tuning of sampling time and perturba-
tion step size, demonstrating the significance of interactive gains for RCP. Sufficient
details of fundamental fixed point blocks have been discussed for implementation of
other perturbation-based MPPT algorithms with minimal modifications to the model
discussed in this paper.
Acknowledgements Authors would like to thank KLE Technological University, Hubli—580031,

Karnataka (INDIA) for funding this research project under Capacity Building Project (CBP-2019)
scheme.
References
1. MNRE (2020) Solar energy systems. https://mnre.gov.in/solarsystems. Accessed 04 Jan 2020

2. Pulle DW, Darnell P, Veltman A (2015) Applied control of electrical drives. Springer
3. Roger A (2015) Managing model-based design. The MathWorks, Inc
4. Siwakoti YP, Town GE (2013) Design of FPGA-controlled power electronics and drives using
MATLAB-Simulink. In: 2013 IEEE ECCE Asia Downunder. IEEE, pp 571–577
5. Dos Santos B, Araujo RE, Varajao D, Pinto C (2013) Rapid prototyping framework for realtime
control of power electronic converters using Simulink. In: IECON 2013-39th annual conference
of the IEEE industrial electronics society. IEEE, pp 2303–2308
6. Selvamuthukumaran R, Gupta R (2014) Rapid prototyping of power electronics converters for
photovoltaic system application using Xilinx system generator. IET Power Electron 7(9):2269–
2278
7. Motahhir S, El Ghzizal A, Sebti S, Derouich A (2017) MIL and SIL and PIL tests for MPPT
algorithm. Cogent Eng 4(1):1378475
8. Jiang Z, Konstantinou G, Zhong Z, Acuna P (2017) Real-time digital simulation based labo-
ratory test-bench development for research and education on solar PV systems. In: 2017
Australasian universities power engineering conference (AUPEC). IEEE, pp 1–6
9. Cheddadi Y, Errahimi F, Es-sbai N (2018) Design and verification of photovoltaic MPPT
algorithm as an automotive-based embedded software. Sol Energy 171:414–425
10. Fannakh M, Elhafyani ML, Zouggar S (2018) Hardware implementation of the fuzzy logic
MPPT in an Arduino card using a Simulink support package for PV application. IET Renew
Power Gener 13(3):510–518
11. Bayrak G, Ghaderi D (2019) An improved step-up converter with a developed real-time fuzzy
based MPPT controller for PV-based residential applications. Int Trans Electr Energy Syst
12. Roy S (2017) Hardware based real time implementation of perturb and observe algorithm in
LabVIEW via cRIO-9075 interface. In: 2017 14th IEEE India council international conference
(INDICON). IEEE, pp 1–6
13. Jeyaprabha SB, Selvakumar AI (2017) Model-based MPPT for shaded and mismatched
modules of photovoltaic farm. IEEE Trans Sustain Energy 8(4):1763–1771
14. Elgendy MA, Zahawi B, Atkinson DJ (2011) Assessment of perturb and observe MPPT algo-
rithm implementation techniques for PV pumping applications. IEEE Trans Sustain Energy
3(1):21–33
15. Altair (2020) Altair Embed: visual environment for embedded systems. https://solidthinking.
com/product/embed/. Accessed 03 Jan 2020
16. Altair (2020) Target hardware support for Altair EmbedTM: TI: F28x, MSP430, ARM Cortex
M3 Arduino: Atmel ST Micro: STM32F3. https://solidthinking.com/white-paper.aspx?id=
8931. Accessed 03 Jan 2020
17. Alik R, Jusoh A (2017) Modified perturb and observe (P&O) with checking algorithm under
various solar irradiation. Sol Energy 148:128–139
18. Elgendy MA, Zahawi B, Atkinson DJ (2012) Assessment of the incremental conductance
maximum power point tracking algorithm. IEEE Trans Sustain Energy 4(1):108–117
19. Killi M, Samanta S (2015) Modified perturb and observe MPPT algorithm for drift avoidance
in photovoltaic systems. IEEE Trans Ind Electron 62(9):5549–5559
20. Killi M, Samanta S (2015) An adaptive voltage-sensor-based MPPT for photovoltaic systems
with SEPIC converter including steady-state and drift analysis. IEEE Trans Ind Electron
62(12):7609–7619
21. Killi M, Samanta S (2018) Voltage-sensor-based MPPT for stand-alone PV systems through
voltage reference control. IEEE J Emerg Sel Top Power Electron 7(2):1399–1407
QRS Complex Detection Algorithm
for Wearable Devices
Lalita Gupta
Abstract Electrocardiography (ECG) is one of the best ways to obtain health infor-
mation and monitor heart activity. QRS complex which is present in ECG is captured
from the wearable devices consisting of sensors. Determination of QRS complex is
crucial to understand the proper condition and functioning of the heart. Various QRS
complex detection techniques are presented so far based on multiple transforms and
designed filters. This paper presents an algorithm based on filtering and averaging
methods which is able to deliver sensitivity and positive prediction of 99.84% and
99.52%, respectively. The error prediction (DER) is found to be 0.65%. This method
is efficient in time when tested using MIT/BIH arrhythmia database.
Keywords Biometrics · ECG · QRS complex · Wearable sensors
1 Introduction
The research in the field of electrocardiogram (ECG) signal applications for wearable
devices has increased tremendously. Tiny electrical changes are observed on the
surface of the skin which arises due to the electrical activity of the heart. These
changes are recorded over a period of time by placing electrodes on the skin. These
signals are collected with the help of wearable sensors, which are gaining popularity
in biomedical science as bio-signal monitoring systems [1, 2]. Biometrics refers to
the physiological and biological characteristics which are unique for each individual
which can be used for their identification. ECG is the display of these recorded signals
that helps in determining the present biological condition of heart. Analyzing the ECG
signal is based on detecting the QRS complex (Fig. 1). The detection of QRS complex
has been creating wide interest in the research field [3]. Computational efficiency and
accuracy of the technique have been a major concern for the researchers over the past
few years. As the ECG signal is collected from the patient, it gets contaminated by the
presence of various noises which are basically other electrical signals arising due to
L. Gupta (B)
Department of Electronics and Communications Engineering, Maulana Azad National Institute of
Technology, Bhopal, Madhya Pradesh, India
168 L. Gupta
Fig. 1 Components of ECG

signal
different activities or surroundings [4]. These noises can be classified predominantly

as the electromyographic interference (EMG), baseline wander noise, and the power
line interference.
The QRS complex significantly contains energy of frequency range of 3–40 Hz
and even more for the Q, R, and S waves having sharp morphologies. The significant
energy components present in the P wave and T wave start from 0.5 Hz and not
more than 10 Hz. Power line interference is nothing but the harmonics of the 50 Hz
power supply. However, baseline wandering and motion artifacts have low frequen-
cies which range between 0.5 and 7 Hz. The ECG signal components get overlapped
in this frequency range. Several algorithms have been presented which are based
on Hilbert transform, wavelet transform [5], derivative filters [6], neural networks,
digital filtering, and heuristic approaches which are based on nonlinear transforma-
tion [7–11]. Accuracy of these algorithms varies due to presence of all these noises.
Besides accuracy and efficiency, computational complexity is a major concern for
designing any algorithm. The most popular and primitive algorithm is given by Pan
and Tompkins [12]. Similar method is proposed by Afonso et al. [13] and Hamilton
et al. [14]. However, these algorithms were too complex to provide good battery effi-
ciency in wearable devices. A novel dual slope QRS detection algorithm is proposed
by Wang et al. which utilizes the fact that the largest slope in the ECG signal is
found in the QRS complex (R peak) [15]. This method provides less complexity
and high accuracy. However, the same can be achieved by the conventional method
of successive filtering and processing the signal. This paper presents an algorithm
which is able to reduce time complexity as well as give better battery performance
with accurate results when applied for the MIT-BIH arrhythmia database.
QRS Complex Detection Algorithm for Wearable Devices 169
2 Methodology
ECG is a quasi-periodic signal which is composed of heartbeats that has three main
components: the P wave, the QRS complex, and the T wave. P wave is associated with
the right and left atrial depolarization. The QRS complex has a series of three deflec-
tions that depicts the right and left ventricular depolarization, where conventionally
the first negative deflection in the complex is Q wave. R wave is the following positive
deflection, and a negative deflection after the R wave is called the S wave. The T wave
denotes the repolarization (recovery) of ventricles, and its position depends on the
heart rate of the person. This sequence of waves typically lasts for about 0.6–1 s. The
morphology of the signal plays an important role, as the steepness, shape, and height
of the signal are used as criteria to detect QRS complex according to the algorithm
proposed by Wang et al. in [15]. The proposed algorithm is described briefly by the
flowchart in Fig. 2 and is discussed as follows.
Fig. 2 Flowchart of the

proposed algorithm
170 L. Gupta
2.1 ECG Signal Extraction
An ECG signal energy lies between 5 and 12 Hz. Designing a band pass filter of
given range increases complexity. So, the signal is passed through a band pass filter
which consists of cascaded low pass and high pass filters. The lower cutoff frequency
of the filter is kept 5 Hz, whereas higher cutoff frequency is 11 Hz with similar gains
of both high pass and low pass filters. The transfer function of high pass and low
pass filter is given by (1) and (2), respectively.

−16 1 − z −32
H1 (z) = z − (1)
1 − z −1
2
1 − z −6
H2 (z) = 2 (2)
1 − z −1
2.2 Derivative Filter
The band pass filtered signal is then passed through the derivative filter to get the
slope information of the QRS complex. The maximum slope in the ECG signal is
found in the QRS complex only since the P and T waves do not posses such high
peaks. The transfer function of the derivative filter used here is given by the following
equation.
1 −2
H3 (z) = −z − 2z −1 + 2z + z 2 (3)
8
It is a five point derivative which provides linear frequency response from DC to
30 Hz, which is favorable for the ECG signal.
2.3 Squaring Function
Squaring operation makes all the data points of the signal positive and gives a higher
value for the R peaks since the slopes of P and T waves are small. The output of the
derivative filter is nonlinearly amplified which emphasizes the higher frequencies.
The squaring operation is performed by the following equation.
xs (nT ) = [xd (nT )]2 (4)

2.4 Averaging and Threshold Detection
To detect peaks, a threshold value is required which is calculated on the basis of the
signal received after averaging of the squared signal. The squared signal undergoes
moving window integration to obtain the average value. An additional block adder
adds this signal with the original signal. The threshold value is determined as the
mean of the integrated signal after averaging and is applied on the output signal from
the adder. The peaks exceeding the threshold value are considered as R peaks and
are used in the calculation of accuracy, positive prediction sensitivity, and detection
error.
The proposed algorithm is implemented in MATLAB using the MIT-BIH arrhythmia

database. The database contains 48 half hour recordings of different patients for total
24 h of ECG data. These recordings have sampling rate of 360 Hz and have 11 bit
resolution over 10 mV [16, 17].
The calculation of FP and FN helps to evaluate the false detection rate. FP is
false detection of peaks, and FN is the count of those peaks whose detection has
failed. TP stands for true positive which refers to the correct detection of peaks. All
these parameters are used to calculate sensitivity (Se), positive prediction (+P), and
detection error rate (DER) using the following equations (Fig. 3 and Tables 1 and 2).
TP
Se (%) = (5)
TP + FN
TP
+P(% ) = (6)
TP + FP
FP + FN
DER = (7)
Total QRS
4 Conclusion
The algorithm presented in this paper gives faster results and better performance in
terms of accuracy and sensitivity. The algorithm is able to extract the ECG signal
from its accompanied noise signals. Hence, the processing of the ECG signal is
done precisely, and truthful results are obtained. The application of ECG signal
morphology can be extended for biometric application using feature extraction. The
172 L. Gupta
Fig. 3 Steps involved in QRS detection of tape 231 of MIT-BIH database

Table 1 Evaluation of implementation results of proposed algorithm

Tape No. Total beats FP FN Time (sec) Se (%) +P (%) DER (%)
100 2273 3 3 2.03 99.87 99.87 0.26
101 1865 6 1 2.19 99.95 99.68 0.38
102 2187 0 3 2.43 99.86 100 0.14
103 2084 1 2 2.37 99.90 99.95 0.14
104 2229 59 2 2.67 99.91 97.42 2.74
105 2572 29 9 2.62 99.65 98.88 1.48
106 2027 7 2 2.19 99.90 99.66 0.44
107 2137 9 0 2.3 100 99.58 0.42
108 1774 123 33 2.12 98.14 93.40 8.79
109 2532 12 0 2.21 100 99.53 0.47
111 2124 3 0 2.51 100 99.86 0.14
112 2539 3 1 2.31 99.96 99.88 0.16
113 1795 1 2 2.26 99.89 99.94 0.17
114 1879 37 17 2.54 99.10 98.05 2.87
115 1953 1 1 2.17 99.95 99.95 0.10
116 2412 31 34 2.36 98.59 98.71 2.69
117 1535 5 1 2.37 99.93 99.68 0.39
118 2278 2 0 2.47 100 99.91 0.09
119 1987 3 2 2.15 99.90 99.85 0.25
121 1863 3 0 2.13 100 99.84 0.16
122 2476 1 0 2.15 100 99.96 0.04
123 1518 4 4 2.32 99.74 99.74 0.53
124 1619 3 0 2.41 100 99.82 0.19
200 2601 10 2 2.36 99.92 99.62 0.46
201 1963 6 2 2.51 99.90 99.69 0.41
202 2136 1 0 2.51 100 99.95 0.05
203 2980 86 2 2.63 99.93 97.19 2.95
205 2656 2 7 2.43 99.74 99.92 0.34
207 1862 0 0 2.65 100 100 0.00
208 2955 35 9 2.68 99.70 98.83 1.49
209 3004 7 3 2.37 99.90 99.77 0.33
210 2650 4 0 2.12 100 99.85 0.15
212 2748 2 0 2.33 100 99.93 0.07
213 3251 1 2 2.36 99.94 99.97 0.09
214 2265 1 0 2.47 100 99.96 0.04
215 3363 2 0 2.51 100 99.94 0.06
(continued)
174 L. Gupta
Table 1 (continued)
Tape No. Total beats FP FN Time (sec) Se (%) +P (%) DER (%)
217 2209 4 2 2.66 99.91 99.82 0.27
219 2154 2 3 2.64 99.86 99.91 0.23
220 2048 1 1 2.64 99.95 99.95 0.10
221 2427 2 0 2.35 100 99.92 0.08
222 2483 2 0 2.17 100 99.92 0.08
223 2605 2 4 2.13 99.85 99.92 0.23
228 2053 2 0 2.35 100 99.90 0.10
230 2256 2 0 2.14 100 99.91 0.09
231 1571 2 2 2.34 99.87 99.87 0.25
232 1780 2 0 2.33 100 99.89 0.11
233 3079 0 2 2.14 99.94 100 0.06
234 2273 0 3 2.11 99.89 100 0.11
Total 109,510 524 161 2.36 99.84 99.52 0.65
Table 2 Comparison of different methods for QRS detection

Methods FP FN Se (%) +P (%) DER References
BPF/search back 248 529 99.61 99.51 0.96 [12]
Filter banks 406 374 99.6 99.45 0.71 [13]
Dual slope 405 199 99.82 99.63 0.55 [15]
Multiscale morphology 213 204 99.81 99.80 0.39 [7]
Proposed method 524 161 99.84 99.52 0.65 –
implementation of a modified version of this algorithm along with hardware interface

can enhance the automatic detection of cardiac diseases.
References
1. Wahabi S, Pouryayevali S, Hari S, Hatzinakos D (November, 2014) On evaluating ECG

biometric systems: Session-dependence and body posture. IEEE Trans Inf Forensic Secur
9(11)
2. Rachim VP, Chung WY (December, 2016) Wearable noncontact armband for mobile ECG
monitoring system. IEEE Trans Biomed Circuits Syst 10(6)
3. Kohler BU, Hennig C, Orglmeister R (January–February, 2002) The principles of software
QRS detection. IEEE Eng Med Biol Mag 21(1):42–57
4. Tomasini M, Benatti S, Milosevic B, Farella E, Benini L (May, 2016) Power line interference
removal for high-quality continuous biosignal monitoring with low-power wearable devices.
IEEE Sensors J 16(10)
5. Zhang F, Lian Y (2009) Wavelet and Hilbert transforms based QRS complexes detection
algorithm for wearable ECG devices in wireless body sensor networks. IEEE
6. Arzeno NM, Poon CS, Deng ZD (2006) Quantitative analysis of QRS detection algorithms
based on the first derivative of the ECG. In: IEEE Proceedings of 28th EMBS Annual
International Conference in New York City, USA (August 30–September 3)
7. Zhang F, Lian Y (August, 2009) QRS detection based on multiscale mathematical morphology
for wearable ECG devices in body area network. IEEE Trans Biomed Circuits Syst 3(4)
8. Kohama T, Nakamura S, Hoshiso H (1999) Preprocessing MAMD and RSEF for R–R interval
detection of ECG. IEEE TENCON
9. Phukpattaranont P (2015) QRS detection algorithm based on the quadratic filter. Elsevier Expert
Syst Appl 42:4867–4877
10. Rufas DC, Carrabina J (2015) Simple real-time QRS detector with the MaMeMi filter. Elsevier
Biomed Signal Process Control 21:137–145
11. Zhang F, Wei Y, Lian Y (June, 2008) Efficient QRS detection in wearable ECG devices for
body sensor network. In: IEEE Proceedings of 5th international workshop on wearable and
implantable body sensor networks, in conjunction with the 5th International summer school
and symposium on medical devices and biosensors HKSAR
12. Pan J, Tompkins WJ (March, 1985) A real-time QRS detection algorithm. IEEE Trans Biomed
Eng BME-32(3):230–236
13. Afonso VX, Tompkins WJ, Nguyen TQ, Luo S (1999) ECG beat detection using filter banks.
IEEE Trans Biomed Eng 46:192–202
14. Hamilton PS, Tompkins WJ (December, 1986) Quantitative investigation of QRS detection
rules, using the MIT/BIH arrhythmia database. IEEE Trans Biomed Eng BME-33(12)
15. Wang Y, Deepu CJ, Lian Y (2011) A computationally efficient QRS detection algorithm
for wearable ECG sensors. In: IEEE 33rd Annual International Conference EMBS Boston,
Massachusetts USA (August 30–September 3)
16. Arefin MR, Tavakolian K, Rezai RF (2015) QRS complex detection in ECG signal for wearable
devices. IEEE
17. Arefin MR, Rezai RF (2014) Computationally efficient QRS detection analysis based on dual-
slope method. IEEE
ORS: The Optimal Routing Solution
for Smart City Users
K. Krishna Rani Samal, Korra Sathya Babu, and Santos Kumar Das
Abstract Road transport-related air pollution is one of the critical issues for most
of the developing countries as the emission of air pollution during transportation
causes severe physical and psychological health problems. Hence, the identification
of pollution-free route is the most critical aspect in the case of air pollution manage-
ment for smart city users. This research paper proposes an optimal routing solution
algorithm (ORS) approach that predicts the pollutant value for the road network and
includes a weight assignment function to measure the safety level of each route to
build a developed transportation system for smart city users. The proposed method
experimented with both geostatistical and non-geostatistical algorithm to predict the
SPM pollutant level in each route. Each approach is evaluated to compute the weight
function to determine the optimized routing solutions. The streamlined routing solu-
tion can be accessible by the developed Web applications to access the healthier path,
which may not provide the shortest path every time but suggests a more robust route
for smart city users.
Keywords Air pollution · Spatial analysis · Geostatistics · Transportation
1 Introduction
Road transportation is the oldest means of communication in India. In developing

countries like India, it covers all the areas and facilitates to provide door to door
services. Due to modernization, several ways have been developed to travel across
K. K. R. Samal (B) · K. S. Babu

Department of Computer Science and Engineering, National Institute of Technology Rourkela,
Rourkela, Odisha 769008, India
K. S. Babu
S. K. Das
Department of Electronics and Communication Engineering, National Institute of Technology
Rourkela, Rourkela, Odisha 769008, India
178 K. K. R. Samal et al.
roads. This modernization phenomenon has also brought several negative impacts
on environments like the increasing level of air pollution due to overcrowded routes.
These busy streets bring high risk to the general public, like adverse effects on the
environment and human life. So, air pollution has become one of the most compli-
cated environmental issues due to its negative impact on human health. The trans-
portation system has become a severe issue in India, more specifically due to road
congestion and the increasing level of pollution generated due to traffic congestion.
Due to the high air pollution exposure, it is challenging to execute plans to develop
a smart transportation system for smart city users as it causes severe health issues
at nearby locations of toxic spots [1]. Therefore, there is a need to minimize traffic
congestion that increases due to the high level of air pollution at highways. Hence, it
is also essential to get notified of the air pollution level in routes, to prevent diseases
and plan daily activities accordingly. Thus, transportation has become a basic chal-
lenging problem in everyday life of citizens in a smart city. This issue can be solved
by spatial analysis of pollution levels at each location. As it is complicated to make
the construction of roads for smart city users due to the limitation of resources and
land use, it is better to enhance transportation quality by minimizing air pollution
and its negative impacts.
Following the introduction, the rest is organized as follows. Section 2 represented
state of the art, Sects. 3 and 4 include the proposed system model and results and
discussion. Section 5 presents the conclusion and future work part of this research
paper.
2 State of the Art for Routing Solutions
Zahmatkesh et al. [2] used air quality index as a standard to build urban travel route
planning. The optimum route-finding based on the Dijkstra shortest path algorithm
and AQI value. The author included a combination of both the parameters as a weight
function to find the optimum path. Sharker et al. [3] also present a similar type of
approach which differentiates the shortest route and the least air pollution exposure
route in terms of distance, time, and air quality index. Determine the source of pollu-
tion level is also a crucial part of pollution management. Therefore, Briggs et al.
[4] studied the causes of air pollution and proposed a regression-based approach for
mapping traffic generated air pollution using spatial analysis techniques. Li et al. [5]
also studied the source of increased levels of air pollution and its health impact during
the different modes of transportation. The author suggested walking communication
instead of motorcycles to minimize the pollution level. Vamshi et al. [6] attempted
to reduce the pollution level by reducing its source of origin. He developed a frame-
work based on Internet of things which distributes the traffic uniformly by providing
diversion suggestion during peak traffic duration, and therefore, it reduces the traffic
generated air pollution. Few types of research have been conducted to compute the
real-time pollution value on each route. Ramos et al. [7] proposed a sensor network
to identify the pollution level of each journey and provides all functionality that can
ORS: The Optimal Routing Solution for Smart City Users 179
give the user all the facilities to trace the least polluted route among all possible
ways.
Though all the existing work is made to alert citizens and focused on people’s
security, these applications are not user-friendly to access the analyzed information
and also not able to compute global variations of air pollution to predict pollution
level accurately. So, based on this study, the proposed work proposed a method to
predict the pollutant value of each route more appropriately to find an optimal routing
solution (ORS) for smart city users. The proposed system model is explained in the
next section.
3 Proposed System Model
The proposed system model is a technology-based framework that consists of

multiple components. Figure 1 represents the proposed framework consists of the
geospatial data layer, data integration layer, geospatial analysis stage, and client
application layer. Geospatial data layer includes data of air pollution monitoring
stations, IoT-based devices. Data integration (DI) layer builds the data fusion model,
which integrates study area coastline, natural, highway, administrative vector data,
and IoT sensor device data. The geospatial analysis stage takes data fusion data as
input and implements a geostatistical model to generate spatial interpolation maps.
This stage provides map layer services that publish interpolation maps to the cloud
server, where the navigation map will create a routing service based on the map
service layer after the client application layer HTTP request. The client application
Fig. 1 Proposed system model

layer supports Web and desktop as a client to start request (Req)-response (Res)
service with a cloud server to visualize the routing solution for a better decision.
Based on the system model, few things have been presented in the following
subsection, such as study area, proposed spatial interpolation model, cross-validation,
weight assignment, publish the map as a Web service.
3.1 Study Area
This study is carried out at Orissa (India). It extends between the latitudes (17.780 N
and 22.730 N) and between longitudes (81.37 E and 87.53 E), which covers an area
of 155,707 km2 . This state is in the northeastern part of India. A historical dataset
of several pollution monitoring locations, collected from Odisha central pollution
control board [8], is used to identify the spatial distribution of pollution levels across
several geolocations of Odisha. Road networks also included in the base map during
the spatial analysis of pollution over routes. During the study, it found that the SPM
level is very high at Orissa as compared to other pollutants. Hence, SPM considered
as a source of interest to identify the safety level at each route.
3.2 Spatial Interpolation Model
The Empirical Bayesian Kriging (EBK) model is used as a spatial interpolation

model in the geospatial analysis stage. It is a geostatistical algorithm that automates
the process of developing a kriging model. It follows the sub-setting and simulation
procedure to adjust the parameter while developing a valid model automatically. It
accounts for the error generated by the semivariogram model, which is possible by
estimating and implementing many semivariogram models instead of one model.
During the first iteration, it predicts a semivariogram model from the existing dataset
and simulates a new value at each input location. Those simulated data are used to
build a new semivariogram model and uses Bayes’ rule to assign a weight to that
model. The repetition of this process will compute standard prediction error at each
unmeasured locations and generates a semivariogram spectrum. The distribution of
such a semivariogram represented in Fig. 2, where blue crosses signify the empirical
semivariances and solid red lines represent the median of the distribution.
EBK model utilizes three types of semivariogram model which are expressed as
shown in Eqs. (1)–(3) [9]:
Power γ (h) = Nugget + b|h|a (1)
Linear γ (h) = Nugget + b(|h|) (2)

Fig. 2 Distribution of the semivariogram model shaded by density
Thin plate spline γ (h) = Nugget + b|h|2 ∗ ln(|h|) (3)
where h, b represent distance and slope, respectively. After the successful estimation
of the semivariogram at each iteration, EBK generated an interpolation map with the
help of some semivariogram model, as shown in Fig. 3, here color scale indicates
the SPM value over the layer.
3.3 Cross-Validation
The leave-one-out cross validation (LOOC) technique is adopted to compare the

performance of the proposed model with the inverse distance weighting (IDW) model
in ArcGIS [10]. The achievement of those models is estimated based upon their
standard prediction error, and therefore, root means square error (RMSE) and mean
error (ME) evaluated to measure the performance. Table 1 shows the RMSE and
MSE value for both the interpolation technique and found that EBK outperforms the
IDW method for non-stationary data and provides better performance for generating
the continuous surface.
3.4 Weight Assignment
Due to the non-uniform structure of air pollution monitoring stations, it is challenging

to identify the spatial correlation among them and hence to determine the pollution
level at each road segment for a given source and destination. This work followed
Fig. 3 SPM layer generated using EBK
Table 1 Comparison of error

Interpolation/error metrics Inverse distance EBK
metrics
weighting
Root mean square error 19.4994 16.3527
(RMSE)
Mean error (ME) 0.8302 −2.1326
the weighting scheme and compared the weight of each route, to estimate the least
polluted way among all possible paths. The weight computation method depends
upon the interpolation value and travel distance to cover a segment. Therefore, EBK
and IDW algorithms can be implemented for experimental purposes. Due to the high
performance of EBK, it is considered to get SPM prediction value. During EBK
implementations, the k-nearest neighbor technique is adopted to reflect the impact
of interpolation points on the interpolated points. Here, four number of minimum
neighbors and eight number of maximum neighbors are considered during EBK SPM
spatial distribution. Algorithm 1 shows the spatial interpolation technique performed
during the work. Once we get the continuous level of SPM value, it is possible to
assign a weight to each route to quickly identify the optimal solution from source to
destination, which is represented in Algorithm 2.
3.5 Optimal Routing Algorithm
3.6 Publish the Map as a Web Service
The interpolation output is projected to a WGS84 coordinate system to support the

features on the Web. After generating the EBK interpolation polygon map, it can be
published to ArcGIS Online as a Web service to design a Web application. The highly
polluted area can be treated as barriers to provide the safest routing services to drivers
and pedestrians. Here, we have developed a Web map application, which computes
the polluted area and considers it as barrier points for different transportation modes.
The navigation map is selected to find out the area of interest and intended navigation
routes within that area. Direction service enables the user to display the direction
while traveling toward the destination with an optimized travel time solution. High
level interpolated polygon is treated as barriers points during different types of travel
modes to obtain an optimal solution. Stops, barriers are the routing parameters that are
intended to modify the waypoints for getting an optimized routing solution. Travel
time is another most important feature, which is considered to get an optimized
solution. The default transportation mode provides optimized travel time for small
automobiles such as cars. This streamlined solution for a car driver represented in
Figs. 4 and 5, with and without restrictions points.
The driver requires 1 h 45 min to travel from Khorda to Cuttack, Maa Mangala
hotel, as shown in Fig. 4 without consideration of polluted area as a barrier and hence
navigates to the highly polluted area, but in Fig. 5, the driver requires 2 h 15 min to
Fig. 4 Most SPM polluted route

Fig. 5 Least SPM polluted route
cover that distance as he is considered the polluted area as barriers and followed the
longest but healthiest path for a safe journey.
The shortest distance may not be the best solution for smart city users every time.
Some routes may cause severe health impacts while traveling. Hence, it is essential
to identify the key features to determine the optimal path for smart city citizens.
The presented proposed work based on geospatial science and cloud-based mapping
service enables the user to generate network datasets and integrate spatial character-
istics into the database to obtain such key routing parameters. This research was an
attempt to find out the pollution exposure level at each route in Orissa. Therefore, it
also used the EBK algorithm, which is the best suitable interpolation method for non-
stationary data collected from unstructured sensors location than any other spatial
interpolation technique to determine spatial correlation and spatial distribution of
pollution level among those monitoring stations. Identification of the pollution level
at each route is an efficient method for those patients who suffer from chronic diseases
due to the high exposure of pollution. The final results of the work, which displays
as the least polluted route in the designed Web application, can be used by patients
and any other smart city users to improve their health conditions. The proposed work
has some limitations, which can be enhanced in our future work. The first and most
critical issue is the number of active sensors available in Orissa, whose value can
improve the performance of routing results. This work is limited to spatial analysis,
which can be enhanced by spatio-temporal analysis with a multivariate pollution
dataset.
Acknowledgements This research and product development work is financially supported by the
IMPRINT (Grant No.-7794/2016), an initiative of Ministry of Human Resource Development,
Government of India and Ministry of Housing and Urban Affairs, Government of India. The authors
would like to thank once again.
References
1. Zhang J, Wang F-Y, Wang K, Lin W-H, Xu X, Chen C (2011) Data-driven intelligent
transportation systems: a survey. IEEE Trans Intell Transp Syst 12(4):1624–1639
2. Zahmatkesh H, Saber M, Malekpour M (2015) A new method for urban travel route planning
based on air pollution sensor data. Curr World Environ 10:699–704
3. Sharker MH, Karimi HA (2014) Computing least air pollution exposure routes. Int J Geogr Inf
Sci 28(2):362
4. Briggs DJ, Collins S, Elliott P, Fischer P, Kingham S, Lebret E, Pryl K, Van Reeuwijk H,
Smallbone K, Van Der Veen A (1997) Mapping urban air pollution using GIS: a regression-
based approach. Int J 225 Geogr Inf Sci 11(7):699–718
5. Li HC, Chiueh PT, Liu SP, Huang YY (2017) Assessment of different route choice on
commuters exposure to air pollution in Taipei. Taiwan, Environ Sci Pollut Res 24(3):3163–3171
6. Vamshi B, Prasad RV (2018) Dynamic route planning framework for minimal 230 air pollution
exposure in urban road transportation systems. In: 2018 IEEE 4th world forum on the internet
of things (WF-IoT), IEEE, pp 540–545
7. Ramos F, Trilles S, Muñoz A, Huerta J (2018) Promoting pollution-free routes in smart cities
using air quality sensor networks. Sensors 18(8):2507
8. Central Pollution Control Board (6 January 2018) Air pollution. http://cpcb.nic.in/air-pollut
ion/
9. Gunarathna M, Kumari M, Nirmanee K (2016) Evaluation of interpolation methods for mapping
ph of groundwater. Int J Latest Technol Eng Manag Appl Sci 3:1–5
10. Esri, about ArcGIS (June, 2019). http://www.esri.com/software/arcgis
An IoT-Based Pollution Monitoring
System Using Data Analytics Approach
Harshit Srivastava, Shashidhar Mishra, Santos Kumar Das,

and Santanu Sarkar
Abstract Air pollution occurs when harmful gases such as CO and NH3 concen-
tration levels increase above the threshold level specified by the World Health Orga-
nization (WHO). Among this, one of the very important parameters is particulate
matter. These are tiny particulate that they reach directly to the lungs and cause
breathing problems. The standard level of range for pollution is already given by
the central governing body of India, i.e., Central Pollution Control Board (CPCB)
in terms of the air quality index (AQI). In this paper, a system has been developed
for detecting the air pollution index with the help of Raspberry Pi based on IoT
technology which sends an emergency notification (EN) if there are any chances
that the air pollution may raise above the given threshold in the future is developed
which measures physical parameters like temperature, humidity, dew point, wind
speed and pollutants parameters like suspended particulate matter (SPM) and carbon
monoxide (CO) are monitored, and the effect of these parameters in pollution level
is being predicted for pollution monitoring. The main objective of this is to apply the
machine learning algorithm for the prediction and analysis of gas sensors concentra-
tion levels, the effect of physical environmental parameters so that we can analyze
the future concentration levels (AQI) level of the gaseous pollutant, and based on
this, an emergency notification (EN) is sent to the public as well as the concerning
authorities. A system is developed for monitoring and alerting in real time. We are
discussing the different methods used in machine learning algorithm, i.e., support
vector machine (SVM) and random decision forests (RDF) to predict the multivariate
time series for forecasting and to use these predicted values to send an emergency
notification (EN).
H. Srivastava (B) · S. Mishra · S. K. Das · S. Sarkar

National Institute of Technology Rourkela, Rourkela, Odisha 769008, India
S. Mishra
S. K. Das
S. Sarkar
188 H. Srivastava et al.
Keywords Internet of things (IoT) · Raspberry Pi 3 · Support vector machine

(SVM) · Random decision forests (RDF) · Central Pollution Control Board
(CPCB) · Emergency notification (EN)
1 Introduction
Over the century, there has been a significant rise in the industries to fulfill the needs
of the evergrowing population. As a result of large numbers of industries, it has caused
a lot of major problems for the environment. As we know few of the impacts of air
quality on the animals, human and environment, the governing body of the world, i.e.,
World Human Organization (WHO) has created guidelines to curb the health-related
effects of air pollution on public by setting up the guidelines for the concentration
of harmful air pollutants likes particulate matter, carbon monoxide, wind speed,
temperature, humidity, etc. The first and the foremost effects that can be seen are
the environmental pollution due to which there is a degradation of the atmosphere
(breathable air) along with climate change with stratospheric ozone depletion and
reduction in biodiversity, hydrological imbalance, freshwater, etc. Mobile sources
and stationary releases chemical pollutants due to which suspended particulate matter
(SPM), carbon monoxide (CO) and other toxic gases.
There are three important parts of integral design architecture in a system.
1.1 Perception Layer
This network layer primarily includes a stationary field sensor that works on a front-
end acquisition device. As it can be realized by establishing a reliable as well as
stable monitoring system, which includes site selection, sensor deployment over an
area, and meteorological sensor deployment.
1.2 Network Layer
The importance of this is conveying the data related to the environment after assim-
ilating all the sensor nodes which are deployed along an area for monitoring after
this transmits the data received by the microcontroller to the data center in real time.
An IoT-Based Pollution Monitoring System Using … 189
1.3 Application Layer
One of the primary focuses of this is to analyze and then process the pollutants.
Then, calculate the air quality after which it should predict the trend which is to be
followed. From the working prospective, the layer means pollution forecasts and air
quality evaluation and then generating a notification to the concerning authorities.
2 Literature Survey
Hu et al. [1] presented an idea of how to trade off among the monitoring data quality
and the cost of the notification message using the adaptive approach by taking the
variance of the data received from the sensing nodes and adjusting the sensing rates.
Yi et al. [2] where they have tackled the main problem with a pollution monitoring
system is that, the presence of the limited amount of data available and non-scalability
of the system available this been removed in our work by using advance sensing
techniques like Wireless Sensor Networks (WSN). Hu et al. [3] presented an idea of
the Haze Est-a model, which links with fixed-station data and data from the sensor to
measure air quality, and the finding shows that decision tree and estimation accuracy
of support vector regression (SVR) are equivalent. Rybarczyk et al. [4] presented
an idea for the statistical correlation of particulate matter and wind after which it
proposes a simple machine learning model to predict the level of particulate matter.
Zhang et al. [5] presented an idea where they have used the advantage of a support
vector machine of the classification and algorithm to overcome practical problems
with which an algorithm proves the effectiveness of the algorithm. Saha et al. [6]
presented an idea for a region implementing an IoT-stationed method to monitor the
noise intensity and air quality index, a simple machine learning model to predict the
level of particulate matter using Raspberry Pi Wi-Fi-enabled module. Zhang et al. [7]
presented an idea of predicting the future outcome helps in reducing the harmful, so
in this paper, they have used prediction techniques to avoid the effects of air pollution
on the environment. Zhao et al. [8] presented a monitoring system for air pollution
using sensors that can detect particulate matter, temperature and humidity, and when
it crosses a safe range of the above parameters, then an alert signal is emitted. Yu et al.
[9] presented an idea of the random forest in which due to its recursive partitioning
to get many trees and then aggregate the total output to reduce the error. A Lyons
et al. [10, 11] presented an idea that an alert notification will help in reducing the
effects of air pollution. As they have used an alert notification to the people for
high levels of air pollution so as to reduce the damage caused on the health and can
be used to avoid such instances. Ayele et al. [12] presented a system that focuses
on monitoring of air pollutants having an IoT-enabled system using recurrent neural
network specifically Long Short-Term Memory (LSTM) which is a machine learning
algorithm that shows a quick convergence and reduction in the training cycles that
too with good accuracy.
3 Methodology
3.1 Architecture Model
The architecture model that we are using for our system will describe how the whole
system is going to work. As shown in Fig. 1, the model will be used to take the
environmental parameters reading, such as the particulate matter, temperature and
few gas parameters. Now, these data will be taken up by our microcontroller and
then will be sent to the cloud server through a modem. The model also has a display
device that can be used to see real-time data. It also helps to detect the exact location
of the device. The data sent in the server can be seen in the GUI interface that is a
web-page that can be accessed from anywhere across the world.
Sensors Nodes: The different kind of parameters which are analyzed is taken by
the different kind of the sensors, i.e., particulate matter sensing PM2.5 and PM10
(HPMA115S0), temperature sensor and humidity sensor (SHT10) and CO (carbon
monoxide). The reason why these parameters have been considered is because we
want to predict if there is an alarming situation in the near future so that necessary
steps can be taken by the concern authorities.
Control and Processing Unit: It contains a microcontroller which is used to control
the sensors to take values in real time. The microcontroller that we are using is
Raspberry Pi 3 Model B. It features Broadcom BCM2837 SOC, which has an ARM
Quad Cortex processor A53 with 1.2 GHz clock speed, and in addition to this, we
are using an ARPI600 chip is ADC which converts analog to digital values.
Cloud Database and Cloud Server: The server will be used to store, host and
manipulate the database along with a push notification that will be released when
the future values increase a threshold which is being defined by the Central Pollu-
tion Control Board (CPCB) and to send an emergency notification to the concerned
authorities. The server will store the pollution data received from the sensor node
Fig. 1 System architecture model

through wireless links, and this data can be used by the front-end application which
will then be used for the analytical purpose.
Android Emergency Notification: The data that is stored in the cloud database
is stored from the sensor node through wireless links, and this data is used by the
front-end application to analyze in which prediction algorithm is applied to trigger
an Android emergency notification (AEN) with the help of an application program
interface (API), which will be used to take proper precautions and to avoid the harmful
effects of air pollutions.
3.2 Flowchart Model
The primary goal of the model is for monitoring the air quality index over an area and
predicting the value to eliminate the problem of air pollution. As shown in Fig. 2, the
data is taken from a different kind of the sensor and then given to the Raspberry Pi
board and further, the AQI value is being calculated, and finally, the data is uploaded
to the cloud where the prediction models will be used to forecast the future values,
and if the values cross a certain threshold given by an IoT-based pollution Alert
will be shown using Data Analytics Approach the CPCB (Central Pollution Control
Board) in there website so that the concerned authorities can take the necessary steps.
3.3 Prediction Model
The prediction model can be explained by Fig. 3 whose working for each block is
given below.
Data Preprocessing: First of all, remove all the irrelevant features and missing values
from the database. Remove those values that are considered outside of the range to
make sure that the credibility of the pollution dataset is not reduced.
Input Feature Engineering: After taking the sensor values from the ground station
data splitting of the dataset into the hour, weekday/weekend and season by dividing
it into different rows and columns so that as per the need, prediction model can be
applied to get the desired output.
Target Preparation: In target preparation, we take an average of the data point
received over an hour as the data we are getting is coming after 6 s, so there will not
be a significant parameter change over such a short time, and then, the normalization
is done on this data to avoid over-fitting.
Model Training: Here, we use Random Forest and SVM for classification and
predicting the values with the help of Regression models based on the above methods
Fig. 2 Flowchart model diagram
now based on the output decision that will be taken to change the Hyper-parameters
to improve the accuracy of the model.
3.4 Regression Models
Support Vector Regression (SVR): It is one of the much blooming supervised

methods for the regression, support vector tries to give a mapping function which is
nonlinear for plotting the given training data points, i.e., D: (x1, y1), (x2, y2), …, (xi,
yi) to a higher dimension. Over this higher dimension, a separating decision boundary
hyperplane is defined which distinguishes all the dataset over the maximum margin
function.
Fig. 3 Flowchart of prediction model
Random Forest Regression (RFR): It is an ensemble learning algorithm that is based

on decision tree learning algorithm and bootstrap aggregating, which can be used
for regression, classification and other tasks. The vital and important concept is to
improve the prediction accuracy where subsamples play an important role. It is done
by fitting a large number of decision trees on random subsets to the features available
as a result of which it avoids over-fitting. As in this case, bootstrap aggregating is used
(or named bagging) to continuously train decision or regression trees, with random
feature subsets and sample subsets over which the algorithm is applied. Finally, after
this, it predicts all the samples which have been taken into consideration by using the
averaging technique and implying it to all the prediction from trees that have been
trained.
4 Results
4.1 Firebase Server Database
As shown in Fig. 4, we have obtained some of the real-time parameters like temper-
ature, humidity, dew point and particulate matter (PM2.5 and PM10) along with CO
gas sensor values stored in the Firebase database after this prediction model is used.
Fig. 4 Real-time data packets in database
Fig. 5 Classification model of SVM
4.2 Support Vector Machine (SVM)
The primary function of this support is to identify a hyperplane in an N-dimensional

hyperspace, which classifies the data points to distinguish the data points in between
two classes, as it is visible in Fig. 5. Our main goal is to find a hyperplane that has
the maximum margin as it is clear from the plot as the distance between data points
of both the classes should increase significantly over the testing, so the accuracy is
high. As shown in Fig. 6, the accuracy in our case is around 90%.
4.3 Random Forest Regression (RFR)
This technique uses a large number of decision trees in which these trees break a
class of prediction, and then, the class over which maximum votes are considered
Fig. 6 The accuracy of classification model of SVM
Fig. 7 Classification model of random forest regression
is considered as it is visible in our model for prediction in Fig. 7. The basic prin-
ciple behind why the random forest model gives high accuracy is that when several
uncorrelated trees will outperform any individual models. This can be explained as a
wonderful effect of an individual tree that protects each other from their errors. There
may be some trees that are wrong, and many others may be right, so overall, these
trees can move in the right direction. As shown in Fig. 8, it is visible the accuracy
that we are getting is around 99% which is indeed very good. Figure 9 shows the
testing values and the predicted values of classification model of random forest.
4.4 Android Emergency Notification (AEN)
An AEN is send to the public based on the prediction algorithm that it is not safe
for them move out for tomorrow based on the past data that has been received from
the sensing node, this notification is sent based on the data that is received by the
Fig. 8 Accuracy of classification model of random forest regression
Fig. 9 Testing values on the

left side and predicted values
on the right side of
classification model of
random forest
Firebase, this data is being taken up with the help of the API key over which the
data is being accessed, and then based on the threshold levels, a notification is sent
as shown in Fig. 10.
Fig. 10 Real-time images of emergency notification on left side and emergency pop-up on the right
side
5 Conclusion
We have proposed a model which uses Raspberry Pi microcontroller along with

the different kind of the sensors and then sending the data to the Firebase server
over which prediction algorithm is used to predict the future values, and then based
on the values, an emergency notification is sent. In an Internet of things (IoT)-
based environment to measure air quality, this gives us a macro description of our
proposed hardware and software module along with an explanation of the algorithm
for calculation of values of different gas sensors, temperature sensors, humidity
sensors, PPM data and predicts the value with the help of machine learning. Prediction
model shows that RFR outperforms SVM which can be seen from Fig. 9 where the
accuracy for this type of dataset is very high as it has to be the result because random
forest model uses a lot of individual decision trees due to which it reduces the chances
of predicting the wrong values overall due to considering an average result from a
large number of decision trees of air pollution.
6 Scope for Future
Future work is to predict the AQI value for analyzing the air quality using machine
learning with different techniques like LSTM, KNN and MLP and comparing the
accuracy of prediction. The real-time data is to be uploaded on a cloud server using

a LAMP server model, and routing algorithm is to be implemented for finding the
safest path based on environmental parameters from the similar nodes deployed
across different location.
References
1. Hu SC, Wang YC, Huang CY, Tseng YC (November, 2011) Measuring air quality in city areas
by vehicular wireless sensor networks. Elsevier J Syst Softw 84(11):2005–2012
2. Yi WY, Lo KM, Mak T, Leung KS, Leung Y, Meng ML (December, 2015) A survey of wireless
sensor network based air pollution monitoring systems 15(12):31392–31427
3. Hu K, Rahman A, Bhrugubanda H, Sivaraman V (1 June, 2017) Haze Est: machine learning
based metropolitan air pollution estimation from fixed and mobile sensors. IEEE Sensors J
17(11):3517–3525
4. Rybarczyk Y, Zalakeviciute R (2016) Machine learning approach to forecasting urban pollution.
IEEE Ecuad Tech Chap Meet (ETCM), pp 1–6
5. Zhang Y, Liu C, Wang L, Yang A (November, 2012) Support vector machine classification
algorithm and its application. In: Springer information computing and applications (ICICA
2012) communications in computer and information science, vol 308, pp 179–186
6. Saha AK, Sircar S, Chatterjee P, Dutta S, Mitra A, Chatterjee A, Chattopadhyay SP, Saha HN
(2018) A raspberry Pi controlled cloud based air and sound pollution monitoring system with
temperature and humidity sensing. In: 2018 IEEE 8th annual computing and communication
workshop and conference (CCWC). Las Vegas, NV, pp 607–611
7. Zhang J, Ding W (January, 2017) Prediction of air pollutants concentration based on an extreme
learning machine: the case of Hong Kong. Int J Environ Res Public Health 14(2):114
8. Zhao X, Zuo S, Ghannam R, Abbasi QH, Heidari H (2018) Design and implementation of
portable sensory system for air pollution monitoring. In: IEEE Asia Pacific conference on
postgraduate research in microelectronics and electronics (prime Asia), Chengdu, pp 47–50
9. Yu R, Yang Y, Yang L, Han G, Move OA (January, 2016) RAQ-a random forest approach
for predicting air quality in urban sensing systems. Sensors (Basel). Int J Environ Res Public
Health 16(1):PMCID: PMC4732119
10. Lyons RA, Rodgers SE, Thomas S, et al (May, 2016) Effects of an air pollution personal alert
system on health service usage in a high-risk general population: a quasi-experimental study
using linked data. J Epidemiol Community Health 70(12):1184–1190
11. Chen H, Li Q, Kaufman JS, Wang J, Copes R, Su Y, Benmarhnia T (January, 2018) Effect
of air quality alerts on human health: a regression discontinuity analysis in Toronto, Canada.
Elsevier Lancet Planet Health 2(1):19–26
12. Ayele TW, Mehta R (2018) Air pollution monitoring and prediction using IoT. In: 2018
second international conference on inventive communication and computational technologies
(ICICCT), Coimbatore, pp 1741–1745
13. Yang Y, Zheng Z, Bian K, Song L, Han Z (2018) Real-time profiling of fine-grained air quality
index distribution using UAV sensing. IEEE Internet Things J 5(1):186–198
14. Shaban KB (2016) Urban air pollution monitoring system with forecasting models. IEEE
Sensors J 16(8)
15. Martınez-Espana R, Bueno-Crespo A, Timon I, Soto J, Munoz A, Cecilia JM (2018) Air-
pollution prediction in smart cities through machine learning methods: a case of study in
Murcia Spain. J Univers Comput Sci 24(3):261–276
16. Xiao Jun C, Xian Peng L, Peng X (2015) IoT-based air pollution monitoring and forecasting
system. In: International Conference on Computer and Computational Sciences (ICCCS),
Noida, pp 257–260
Modeling the Hierarchical Fuzzy System
for Suitable Crop Recommendation
R. Aarthi and D. Sivakumar
Abstract To obtain successive sustainable agriculture, a farmer needs good prac-

tical and expert’s knowledge regarding the modern farming technologies. As per the
personal experience and traditional knowledge, the framers recommend a suitable
crop for cultivation and this prediction may go wrong several times. Therefore, fuzzy
expert system designed for crop recommendation systems for archiving the efficient
crop recommendation. In the traditional single-layer fuzzy system, the rule is expo-
nentially increased when the system variables are increased and a larger rule base will
affect the system performance and transparency. Therefore, the multi-layer system
is developed using the fuzzy hierarchical approach. The hierarchical fuzzy model is
applied in a Mamdani fuzzy inference system for a suitable crop recommendation
system. The crop recommendation system has 12 input variables, and it was decom-
posed into six fuzzy subsystems and arranged as per the priority. The final results
show that a hierarchical crop recommendation system provides better performance
than traditional fuzzy crop recommendation.
Keywords Fuzzy inference system · Crop recommendation · Hierarchical fuzzy

system · Rule explosion · Soil properties
1 Introduction
Traditionally, the crop selection process is based on a farmer’s own experience, but
in most cases, the prediction went wrong due to lack of knowledge. Because the
successful crop selection process is not only based on single criteria but also requires
several agriculture factors. In general crop, recommendation involves several phys-
ical, chemical, and environmental factors. Crop selection is indirectly related to
production and profit; this can be archived using the decision-making tool [1]. Devel-
oping the expert system is the best choice for crop recommendation because most of
the farmers do not get farming advice and expert knowledge collected from a different
R. Aarthi (B) · D. Sivakumar

Department of Electronics and Instrumentation Engineering, Annamalai University,
Chidambaram, Tamil Nadu 608002, India
200 R. Aarthi and D. Sivakumar
field, scientist, and organization. The fuzzy-based expert system is widely used in
the agriculture sector [2]. In the agriculture sector, most of the applications follow
the nonlinear system, and the fuzzy system is the best choice for nonlinear appli-
cations. Therefore, fuzzy-based expert system is developed for the crop selection
process. This type of expert system helps the farmers for the crop selection process.
The major issue the fuzzy system faced is the rule explosion. If the number of input
variables increased means, fuzzy rules also exponentially increased and affect the
system performance. The fuzzy tree or hierarchical approach is a promising method
to handle larger fuzzy variables spilled out into several subsystems; this will reduce
the complexity and improve the system performances [3]. The fuzzy hierarchical
system stared to evaluate the risk state in the driving assistant system [4]. In recent
years, smart city development project is a major research field in India. However,
maintenance and monitoring will cause several risks. To handle larger risk manage-
ment in smart cities, a fuzzy hierarchical approach is used [5]. The author utilizes
the hierarchical approach in the medical field that is designing a fuzzy system for
diagnosing the arthritis diseases [6]. Roel Bosma et al. develop the hierarchical
fuzzy layer in farming activity to improve the decision-making technique [7]. The
author develops a novel hierarchical approach that is a cohesive hierarchical fuzzy
system to monitor the water supply pipeline [8]. This study focused on the predic-
tion of daily gas consumption in the Czech Republic based on climate conditions.
The larger fuzzy system of the gas consumption model is decomposed into several
subsystems and connected in hierarchical order [9]. Greenhouse climate control is a
nonlinear approach and widely handled with the fuzzy system. The climate control
system is partitioned into several small FIS based on the hierarchical approach [10].
The proposed work is to develop the crop recommendation system with 12 different
agriculture parameters. This nonlinear approach is to deal with fuzzy systems. To
accomplish precise and less complex systems designing, the hierarchical fuzzy model
is employed for crop recommendation systems.
2 Material and Methodology
2.1 Hierarchical Fuzzy System
Rule generation is a fundamental part of a fuzzy system, but a larger number of

input variables will cause the rule explosion. The rule explosion is a major limitation
in the fuzzy system because it directly affects the system transparency and preci-
sion. In the standard fuzzy system, the rules are constructed in mn manner, where
n is an input variable and m is a membership function. Therefore, n increases will
automatically overload the fuzzy rules and directly affect the system performance.
The standard fuzzy system with a large rule base requires large memory and longer
computation time. Fuzzy hierarchical or tree model is the only way to reduce the rule
Modeling the Hierarchical Fuzzy … 201
(a) incremental (b) aggregated (C) cascade
Fig. 1 Types of fuzzy hierarchical structure
explosion-related problem and improve the fuzzy system transparency and compu-
tation efficiency because a larger system splits into several small fuzzy systems [11].
Several approaches are present to decompose the larger fuzzy system into a small
subsystem; they are incremental, aggregated, and cascade methods. The hierarchical
structure design is based on this topology. The term hierarchical fuzzy system also
named a fuzzy tree because the system is arranged in a tree-like structure. Figure 1
illustrates hierarchical fuzzy system structure.
In the incremental structure, the input variable is incorporated into multiple stages
to define the output. The input fuzzy system is arranged in the order of contribution;
the most important fuzzy input is arranged in the lower level and least at the top
level. In an aggregated structure, the fuzzy system is formed in a group manner
where a lower-level fuzzy structure is combined with a higher-level system [12].
In this structure, the subsystem is formed in the decision-making structure. The
cascade structure also named as a combined structure because it groups together,
both incremental and aggregated structure in the system. This fuzzy tree structure is
suitable for both correlation and uncorrelated input.
2.2 Methodology
Rule explosion is a major issue in the standard fuzzy system because it creates the
complexity of system designing. The crop recommendation requires the physical
and chemical properties of soil. In the proposed paper, crop recommendation model
developed using a fuzzy inference system. The block diagram of the standard fuzzy
crop recommendation system is illustrated in Fig. 2.
Fig. 2 Block diagram of standard fuzzy system
The system designed 12 input parameters to archive the effective result in crop
recommendation. The input parameter is sand, silt, clay, nitrogen, phosphorus, potas-
sium, soil color, soil pH, soil electrical conductivity, rainfall, climate zone, and water
resources. This standard fuzzy model has two layers that are the initial layer or input
layer and the final layer. As per the expert knowledge, each input variable has 3–4
membership functions. The total number of rules 499 for the crop recommendation
system. To improve the crop recommendation system performance and reduce the
complexity of rule formation, the fuzzy hierarchical system is proposed over the stan-
dard fuzzy system. The fuzzy hierarchical layout for crop recommendation system
is illustrated in Fig. 3.
In the hierarchal fuzzy model, the input variable is arranged in a lower-dimensional
order instead of using higher order in the standard fuzzy system. This hierarchical
model of the output variable of the previous fuzzy system is given as the input variable
to the next fuzzy system. And this work splits into three layers; they are input layer,
intermediate layer, and final layer. The input layer consists of 12 variables, and
intermediate layer is separated as five fuzzy inference systems; they are named as
soil texture, soil type, soil nutrient, soil fertility, and environmental factor. The inputs
of FIS 1 or texture class are sand silt and clay, and the outputs of the FIS 1 are given
as input to FIS 2. For FIS 3 input variables are NPK and soil pH and soil EC is
Fig. 3 Hierarchical fuzzy crop recommendation system
given as input parameter FIS 4. Climate zone, rainfall, and water resource are input
parameters for FIS 5. Finally, the output parameters of FIS 2, FIS 4, and FIS 5 are
given as input variables of FIS 6 that is a crop recommendation system. This fuzzy
hierarchical crop recommendation system drastically reduces the number of fuzzy
rules compared with the traditional fuzzy system.
3 Implementation and Discussion
The standard fuzzy model and hierarchical fuzzy crop recommendation system
were developed in the fuzzy inference toolbox in MATLAB version 2018a. Table 1
represents the basic configuration of FIS designing for crop recommendation.
The input and output membership for FIS 1 are shown in Fig. 4, and its corre-
sponding rule view is demonstrated in Fig. 5. In FIS 1, sand silt and clay are input
variables, and each input variable has three membership functions. Texture classes are
output variables and 12 membership functions. Totally, 12 fuzzy rules are developed.
Table 1 Basic FIS

Configuration Method
configuration
Membership function Triangular
Input fuzzification MIN and MAX
Implication MIN
Aggregation MAX
Defuzzification Centroid
Final hierarchical crop recommendation FIS 6 rule viewer in Fig. 7. Output of

fuzzy inference system FIS 2, FIS 4, and FIS 5 is applied as the input to FIS 6.
The input and output variables soil type, soil fertility, environmental factor, and crop
recommendation are illustrated in Fig. 6.
In this paper, five subsystems are designed for crop recommendation. They are
soil texture, soil type, soil nutrient, soil fertility, environmental factor, and crop
recommendation. The hierarchical fuzzy system rule table show in Table 2.
The rule table represents possible number rules in hierarchical system to numbers
of rules which drastically reduced when compared with standard fuzzy model.
Traditional fuzzy model total number of rules: 439.
Hierarchical fuzzy model total number rules: 152.
Fig. 4 Input and output membership function of FIS 1

Fig. 4 (continued)
Fig. 5 Rule viewer of FIS 1

Fig. 6 Input and output membership function of crop recommendation

Fig. 6 (continued)
Fig. 7 Rule viewer of crop recommendation system

Table 2 Rule formation in hierarchical fuzzy crop recommendation system

Model/fuzzy tree No. of FIS unit Input variable Total no. of No of fuzzy rules
membership function
Soil texture FIS 1 3 3 12
Soil type FIS 2 2 3 15
Primary nutrient FIS 3 3 2 38
Soil fertility FIS 4 3 3 35
Environmental FIS 5 3 3 12
factors
Crop FIS 6 3 20 40
recommendation
Total number of rules 152
4 Conclusion
The hierarchical fuzzy-based suitable crop recommendation system was developed.

This hierarchical model partitions the larger fuzzy system into several small subsys-
tems. Therefore, the number involved fuzzy rules also reduced compared to the
traditional fuzzy system for crop recommendation. The crop recommendation system
complexity is reduced, and the system performance is improved in the hierarchical
fuzzy system. The simulation results show that hierarchical fuzzy yields better
performance than the tradition fuzzy system in crop recommendation application.
References
1. Deepa N, Ganesan K (2017) Decision-making tool for crop selection for agriculture develop-
ment. Neural Comput Appl
2. Papageorgiou IE, Markinos A, Gemptos T (2009) Application of fuzzy cognitive maps for
cotton yield management in precision farming. Expert Syst Appl 36:12399–12413
3. Chen Y, Abraham A (2010) Hierarchical fuzzy systems, tree-struc. Based Hybrid Com Intell
129–147
4. Dkhil MB, Wali A, Alimi AM (2018) A hierarchical fuzzy system for an advanced driving
assistance system. Comput Vision Pattern Recogn Artif Intell
5. Ullah I, Fayaz M, Kim DH (January, 2019) Risk assessment based on hierarchical fuzzy
inference and prediction using Kalman filter for underground facilities in smart cities. Int J
Innov Technol Explor Eng (IJITEE) 8(3C)
6. Lim CK, Yew KM, Ng KH, Abdullah BJJ (2002) A proposed hierarchical fuzzy inference
system for the diagnosis of arthritic diseases. Australas Phy Eng Sci Med 25(3)
7. Bosma R, Kaymak U, van den Berg J, Udo H, Verreth J (2011) Using fuzzy logic modelling
to simulate farmers’ decision-making on diversification and integration in the Mekong delta,
Vietnam. Soft Comput 15:295–310
8. Fayaz M, Ahmad S, Hang L, Kim D (2019) Water supply pipeline risk index assessment based
on cohesive hierarchical fuzzy inference system. Processes 7:182
9. Radek Sindel a.r, Hierarchical Fuzzy Systems, 16th Triennial World Congress, Prague, Czech
Republic, Elsiver IFAC Publication, 2005 IFAC
10. Paulo Salgado J (2005) Boaventura Cunha, greenhouse climate hierarchical fuzzy modelling.
Control Eng Pract 13:613–628
11. Lee M-L, Chung H-Y, Yu FM (2003) Modeling of hierarchical fuzzy systems. Fuzzy Sets Syst
138:343–361
12. Yager RR (February 1998) On the construction of hierarchical fuzzy systems models. IEEE
Trans Syst Man Cybern Part C: Appl Rev 28(1)
Real-Time Emergency Vehicle Response
System for Smart City Applications
Lima Priyadarsini, Prashant Deshmukh, and Santos Kumar Das
Abstract The city applications and services are targeting how to frame the upcoming
Internet of things (IoT)-based applications and services for the perspective of the
smart city. The acceptance, arrangement and execution of cutting-edge wireless
services and applications have to be designed in the manner to allow facing the
difficulties of smart cities. There is intelligence, analytical and conduct automation
that permit the influencing better policy and smart deployment. Nowadays, various
works are going on to detect health disease, accident, crime and fire for achieving
the goal of smart cities. There are many individuals in the world to be alive if speedy
notice would have given to emergency services. This detection system provides only
alert to the respective authorities or their relatives but not able to implement real-
time lifesaver services. The objective regarding this paper is to provide online cloud
services to various emergency detection devices so that with the help of these devices,
life can also be saved. The smartphone has become an important instrument in our
life; with the help of a global position system (GPS) or cell phone location in mobile,
human can be tracked. This work mainly deals with emergency vehicle networks for
an ambulance, fire brigade and police vehicles like online cab services. In this work,
a mobile application is proposed which is immediately triggered after detection of
emergencies and provides real-time vehicle services for the nearby hospital, police
control room and fire brigade.
Keywords Cloud services · Emergency vehicle response network · Mobile

application · Google map
L. Priyadarsini (B) · P. Deshmukh · S. K. Das

Rourkela, Rourkela, Odisha 769008, India
P. Deshmukh
S. K. Das
212 L. Priyadarsini et al.
1 Introduction
Emergency vehicle armada dispatching and the board assume a significant job in
occurrence reaction times and the nature of administration offered to end clients. The
viability of the management procedure is important to limit the effect of occurrences
on human lives and properties. Exact and opportune choices are along these lines
required.
The key segment for crisis trade reaction scheme is the choice of stream ways
for vehicles and materials, by the reason for well-being and brief time, moving,
however, countless particular and vehicles influenced site could be expected under
the circumstances is one of the instinctive models to quantify the plausibility of
the arrangement [1]. Simultaneously, the city’s unique street organization has just
come to fruition; however, the development pace of vehicles is a lot quicker than
that of new streets [2]. The connection between organic market reversals prompts
the standardization and expanding the seriousness of urban clog.
The underdevelopment of data assortment innovation and the bother of data trans-
mission and dissemination will all lead that coordinators of crisis reaction cannot
take correct countermeasures; voyagers cannot acquire convenient, right and valuable
data to choose the ideal way for departure; salvage powers cannot convey alleviation
supplies to the scene in time [3–5].
To beat these issues, we propose a Real-Time Emergency Vehicle Response
System, which gives ongoing administrations to penniless people, accommodating
for diminishing the demise because of mishap cause, deferring of rescue vehicle
administrations and police administrations.
2 Related Work
One commonly applied to use safe for vehicular ways are forms that recognize crisis
channels. Such a request can reduce mishap risks through crisis effect hallucinations
and help additional important time. They diagram a far-reaching structure of such a
crisis channel warning structure that utilizes channel communication, yet addition-
ally involves roadside foundation alike traffic lights. Within this structure, different
vehicles do not only warned about moving toward crisis vehicles; they likewise get
nitty-gritty course data. Given the data, the favorable and proper acknowledgment
of various operators is understandable. A representation of this structure has been
proposed in a rush-hour gridlock condition including crisis vehicles and traffic lights.
To recognize prerequisites and assess the framework, this additionally led a point-
by-point investigation of recordings from crisis reaction trips and a specialist review
among individuals from a neighborhood crisis reaction association [6].
Urban crises are difficult to maintain a strategic distance from. Traffic crisis reac-
tion after an episode assumes a significant job in decreasing misfortunes and is a
key connection in urban crisis the executives. The framework, for the most part,
Real-Time Emergency Vehicle Response System … 213
incorporates sub-frameworks, for example, staff clearing information assortment

framework, vehicle activity information assortment framework, salvage material
conveyance information assortment framework, faculty settlement place information
assortment framework, traffic pike wise distinguishing proof framework and so on.
It likewise devises the working projects for ordering the executives, faculty clearing
and catastrophe transfer if there should arise an occurrence of crisis, and improves the
urban crisis bolster the board framework. Accompanied by the Internet of everything
and the details extract innovation, the trade crisis reaction framework is auspicious
as well as precisely check streaming data for faculty and motor vehicles, rapidly with
helpfully immigrate staff in addition to motor vehicle, adequately complete review
salvage tasks, successfully upgrade salvage proficiency with enhancing the degree
for the city administration [7–9].
Emergency circumstances require exact and opportune choices to lessen delay
and the extra effect of episodes on human lives and common properties. Nonethe-
less, basic leadership in urban emergency circumstances is a difficult assignment
because of the number of factors impacting the procedure and the unusual differ-
ence in some of them. This procedure trademark necessitates that leaders screen
and change their choices for all time. In this manner, data accessibility and trade
are basic to improve the basic leadership process. Despite refined enhancement
commitments in unique emergency vehicle dispatching issue, not many of them
have thought about their combination in intelligent transportation systems (ITS). In
this unique circumstance, dynamic occasions and non-intermittent blockage could
be better tended to. In this paper, we consider the dynamic crisis vehicles dispatching
issues utilizing real-time data (DEVDP-RT). We propose a responsive methodology
for vehicle dispatching dependent on a setting mindful and reconfigurable design.
Our methodology is planned to help apportion emergency vehicles to urban crises
and modify their courses as indicated by up and coming changes and unforeseeable
occasions. Recreation results feature the advantages of the proposed dispatching
model in improving the crisis reaction time and lessening delay [10].
Another framework identified with the traffic observing framework has been
suggested by creators. During venture, the group planned cutting-edge trade checking
framework as long as enhancing street trade stream accompanied by the point of gath-
ering present and coming times essential requirements for street move. The frame-
work expands this effectiveness of checking street trade conditions at giving perma-
nent information on weather forecasting parameters of various areas. The article
did not consider safety concerns by this cutting-edge trade observing framework
[11, 12].
Intelligent transportation systems are the whole common significant use of every
Internet of Everything empowered bright towns. It is dependent on a green wave
framework that has been exhibited, which empowers the trade sign framework to
use some signed field during a crisis channel which moves toward those trade signs.
As a result of using these signed fields, a crisis channel gains every single green
sign into the path. Creators exhibited this being a developing flow over a crisis
channel. During the responsibility, this trade sign framework can identify any taken
channel that goes through each developing sign. This significant burden from that
growing surge means that if this synchronization from this movement sign remains
upset, this package produces an enormous road turned parking lot (over-immersion).
To diminish the effect like non-having appropriate integration, RFID locate trade
controller framework has been prescribed [13–15].
3 Proposed System
Google Mobile Services (GMS) is a collection of Google applications and applica-

tion program interfaces that help support functionality across devices. The proposed
system using the cloud, database and Google services for the emergency vehicle
response platform. Smartphone and Internet are the common prominent things used
by the people and become very important in this era. Figure 1 shows the system
architecture of Real-Time Emergency Vehicle Response (RTEVR). In this left side,
all the sensor devices for emergency detection are already available in the market. All
the registered users having these devices with the Android phone have an emergency
calling application installed, when the sensor sends a request to our cloud system,
which is the middle of Fig. 1 shown, and detects emergency. The computing system
of the cloud server classifies the emergency response and generates a broadcasting
message for the corresponding vehicle response network, e.g., health-type request
nearest available ambulance driver will be called, similarly for fire and police. The
GMS provides the path to the driver to the victim using a real-time Google map and
the nearest hospital. Our proposed system stores all the records in the database, e.g.,
driver’s detail, user’s detail, journey time, hospital database, etc.
Transportation of emergency supply is dynamic. The system built up a numerical
model to depict the issue and afterward built up a database of the emergency orga-
nization on the base of the model. The database is the most significant piece of the
product that shows dynamic transportation. Therefore, the paper portrays the plan of
the database insistently. The database is the base of the emergency response system.
Fig. 1 System architecture of RTEVR

It very well may be utilized to store, recover and review the information shown
in Fig. 2, so the emergency system can obtain information from it precisely. The
information, which is associated with the paper, is examined through cluster hubs
and associating sides, the database structure of the essential system. The proposed
system uses MongoDB and Firebase database. MongoDB is the NoSQL database;
therefore, the system can handle a huge amount of data, and it is easy to interface
with Google services, an Android platform using the Atlas cloud or Parse server, and
this system uses Parse server for communicating with Android to MongoDB. The
system uses Firebase as a secondary data for providing an interface to sensor devices
in the proper format because the microcontroller cannot directly communicate with
MongoDB; therefore, firebase is a middleware between our system and emergency
detection devices.
In Fig. 2, user and emergency vehicle driver’s data is shown. The user database has
data field user name, device ID and device type. The system uses the Mac address of
sensor devices as device ID for classification of device type. The emergency vehicle
database consists of all the vehicle driver details, e.g., application ID, mobile number,
name and type in corresponding networks. Among the emergency organization, the
most significant of all components are the logins and connecting sides. Figure 3
depicts the data processing unit of RTEVR, for data to be transmitted from one
Fig. 2 Database design of RTEVR

Fig. 3 Data processing of

RTEVR
platform to another need of protocols and computation unit; therefore in the proposed
system, Amazon cloud is used as a computing platform and cloud services are using
for communication between database and user interface, i.e., Android system or
maybe Web interface.
Given the scientific model, the database of crisis organization is planned effectively.
The product that is on the base of the scientific model above has been created. As
indicated by the trademark and property of each component in the crisis arrange-
ment, we made a comparing element relationship model and tried information with
MongoDB server database and at last, built up the database of a crisis organization.
The achievement of the database demonstrated the exactness and sensibility.
Figure 4 shows the entire registered user’s database; it consists of a user-type field,
which indicates user as driver or victim and also his gender.
Similarly, Fig. 5 shows emergency-type vehicles’ database field which classifies
the request type for ambulance, police and fire brigade and also respected linked
driver details, e.g., e-mail, contact number.
Figure 6 indicates device ID database field link with a user profile for the respected
emergency alert device for accident, crime, fire or healthcare. The Emergency Service
Dashboard is displayed in Fig. 7, and all the requests are severed by response system
using the query dashboard, i.e., all the sensor devices are linked and responded
through the emergency dashboard.
Fig. 4 User database field

Fig. 5 Emergency type vehicles database field
Fig. 6 Device ID database field link with user
Fig. 7 Emergency service dashboard

5 Conclusion
During the evaluation, we developed the incident ID and mind-blowing salvage struc-
ture, which produces emergency alert and send it to the closest emergency commu-
nicator and will comparably send an SMS to emergency connection containing area
addresses of the mishap. By consistent region following for both hurt person and
communicator, the structure will develop the determination step of an event stun-
ning misfortune by giving crisis assistant during the time. The framework will
additionally assist in many crises, for example, when firing, crimes and different
prosperity-associated crises. Urgency communicator directions are gifted branch
location monstrous misfortune’s region toward a Google chart coherently.
References
1. Guo D, Gao C, Ni W, Hu X (June, 2016) Max-flow rate priority algorithm for evacuation route
planning. In: Proceedings of IEEE 1st international conference on data science in cyberspace,
pp 275–283
2. Zhou Z, Lin S, Du W, Liang H (2019) Integration of regional demand management and signals
control for urban traffic networks. IEEE Access 7:20235–20248
3. Zhou J, Beard C, Qian Y (July/August, 2018) Distributed device caching for information
updating in emergency communications. In: Proceedings of 27th international conference on
computer communication and network, pp 1–9
4. Oliveira A, Vazao T, Galao P, Varzea M (October, 2018) Performance evaluation of a hier-
archical low rate emergency system. In: 2018 16th international conference on intelligent
transportation systems telecommunications (ITST), pp 1–7
5. Yang B, Wang J, Ma W, Mao X (2018) Research on crowdsourcing emergency information
extraction of based on events’ frame. In: Proceedings of 10th International Conference on
Computer Electronics and Engineering
6. Buchenscheit A, Schaub F, Kargl F, Weber M (2009) A VANET-based emergency vehicle
warning system. In: IEEE Vehicular Networking Conference (VNC), Tokyo, pp 1–8
7. Sumi L, Ranga V (2016) Sensor enabled internet of things for smart cities. In: Parallel,
distributed and grid computing (PDGC), 2016 fourth international conference on, IEEE, pp
295–300
8. Liu Z, Wang C (2019) Design of traffic emergency response system based on internet of things
and data mining in emergencies. IEEE Access 7:113950–113962
9. Rani LPJ, Kumar MK, Naresh KS, Vignesh S (2017) Dynamic traffic management system
using infrared (IR) and internet of things (IoT). In: 2017 third international conference on
science technology engineering and management (ICONSTEM), pp 353–357
10. Yedder BH, Benyahia I (2017) Reactive emergency vehicles dispatching based real-time infor-
mation dissemination. In: 2017 8th IEEE annual information technology, electronics and mobile
communication conference (IEMCON), Vancouver, BC, pp 471–477
11. Costea I, Nemtanu F, Dumitrescu C, Banu C, Banu G (2014) Monitoring system with appli-
cations in road transport, in design and technology in electronic packaging (SIITME). IEEE
20th international symposium for IEEE, conference proceedings, pp 145–148
12. Aljawarneh S (2011) A web engineering security methodology for e-learning systems. Netw
Secur 2011(3):12–15
13. Nemtanu F, Moise I, Banica C (8–12 May, 2013) Hybrid simulation in intelligent transportation
systems. In: 36th international spring seminar on electronics technology (ISSE 2013), pp 106–
107
14. Andreev S, Galinina O, Pyattaev A, Gerasimenko M, Tirronen T, Torsner J, Sachs J, Dohler M,

Koucheryavy Y (2015) Understanding the IoT connectivity landscape: a contemporary M2M
radio technology roadmap. IEEE Commun Mag 53(9):32–40
15. Mittal AK, Bhandari D (2013) A novel approach to implement green wave system and detec-
tion of stolen vehicles. In: IEEE Conference Proceedings in Advance Computing Conference
(IACC), IEEE 3rd International, pp 1055–1059
Modified Aggressive Packet Combining
Scheme with Repetition Code
for Throughput Enhancement in High
Error Rate Channel
Mancharla Ravi and Yaka Bulo
Abstract Aggressive packet combining (APC) scheme is a well-established scheme

which reduces the number of retransmissions for receiving correct packet in error
prone wireless link. It is a well-known fact that the retransmissions consume extra
bandwidth and battery energy of the portable computer. Therefore, APC is an effi-
cient technique which reduces number of retransmissions for a packet, thus reducing
latency of the wireless system. In APC, three copies of a packet are transmitted, and
receiver does bit-wise majority decision to get the combined copy. Least reliable
bits identification is done in the combined packet if majority logic fails. Search the
correct bit pattern for least reliable bits so obtained. APC provides higher reliability
but at expense of throughput of the system. Therefore, major issue in APC is the
lower throughput which at maximum is 33.33% ( 13 × 100). In this paper, modified
APC with repetition code is proposed which addresses lower throughput issue of
conventional APC. Simulation results show significant performance improvement in
terms of throughput when packet combining scheme with repetition code is used in
APC scheme.
Keywords Bit-by-bit majority logic · Bit error rate · Conventional aggressive

packet combining scheme · Repetition code · Throughput
1 Introduction
The wireless networks suffer from high bit error rate 10−2 [1] to 10−3 [2]. In an auto-
matic repeat request (ARQ) scheme, a packet is retransmitted if it gets corrupted due
to transmission errors caused by the channel. Each retransmissions consume commu-
nication bandwidth and transmission energy. The bandwidth is a scarce resource
M. Ravi (B) · Y. Bulo

Arunachal Pradesh, Arunachal Pradesh, India
Y. Bulo
222 M. Ravi and Y. Bulo
because many portable computers may be sharing a limited free space spectrum. The
energy is also a scarce resource because the portable computer is usually powered
by battery.
Therefore, it is desirable to reduce the number of retransmissions in order to
utilize the bandwidth and battery energy efficiently. In many error control proto-
cols, the receiver discards the erroneous packet and requests the transmitter for a
retransmission. However, an erroneous packet may contain both erroneous bits and
correct bits, and hence, it may still contain useful information. The receiver may
be able to combine this information from multiple erroneous copies to recover the
correct packet. In order to achieve the desirable quality in high bit error rate wireless
channels, several modifications for applying basic ARQ are found in literature where
erroneous packets are combined and by applying logic, the error location is identified.
Once error location is identified, bit inversion is done to error so identified (Brute
method). The technique is variably applied in packet combining scheme [3], multiple
route packet combining scheme [4], and aggressive packet combining scheme [5].
In packet combining scheme [3, 6, 7], two copies of a packet are transmitted, and if
at least one of the received copies are error free, it is accepted as a correct transmis-
sion. However, if both copies are received erroneous, bit-wise modulo-2 sum of the
copies is computed to locate the errors in the combined copy. Thereafter, the packet is
corrected by brute force bit-by-bit inversion of bit located as erroneous and checking
for correctness using frame check sequence (FCS). However, PC scheme fails if there
is at least one-bit position where both copies have an error (termed as double error in
[3]), and the computational complexity increases when the number of errors in the
combined copy exceeds a predefined value N max [3]. To deal with the latency issue in
wireless network, Leung [5] proposed an efficient technique called aggressive packet
combining scheme. In aggressive packet combining scheme, transmitter sends three
copies of packet to the receiver. Receiver on receiving three copies of the packet will
apply bit-by-bit majority voting on three received copies to obtain a combined packet.
Receiver checks error on the combined packet, and if the packet is not corrected by
majority logic, receiver will identify least reliable bits and search the correct bit
pattern for the identified least reliable bits. If still correct copy is not obtained, it will
request for retransmission of the packet. APC gives better error correction capability
[8–14] with reduced latency, but suffers from low throughput. APC scheme requires
at least three copies of a packet to successfully retrieve the correct copy of the packet,
thus the best packet throughput of APC scheme is TAPC = 13 ×100% = 33.33% only.
In this article, we propose a modified technique of APC which increase the
throughput of the conventional APC using the idea of packet combining and three-
bit repetition code. We use the concept of packet combining scheme for incorrectly
received copies to identify the error locations and then use three-bit repetition code
when the retransmission call is received at the receiver side. We assume that the
round-trip delay is negligible, the feedback channel is error free, and CRC provides
perfect error detection capability.
Modified Aggressive Packet Combining Scheme … 223
2 Proposed Scheme
In proposed technique, instead of transmitting three copies of the original packet

as in conventional APC, two copies are sent initially. If the received two copies are
erroneous, then XORing of both the copies are done to locate the error positions as
that of the conventional packet combining scheme. Bit-by-bit inversion of the error
positions which is done in conventional PC is avoided because this step involves
computational complexity which is equal to C = 22nα −2 [3] where n is packet size in
bits and α is bit error rate. Thus, even for medium-sized packets and at moderate BER,
random fluctuations of the channel may make the bit inversion procedure extremely
complex. It also leads to battery energy consumption and delay of the system (as delay
is not acceptable in real-time traffic). Conventional APC is an efficient technique with
low latency, but it suffers from low throughput. Therefore, this proposal is an attempt
to increase the throughput of conventional APC. It is observed that current wireless
mess network protocols retransmit whole of the packet when a retransmission call is
received at the sender side [8]. These retransmissions end up sending bits that have
already received correctly, wasting network capacity. Retransmitting entire packets
work well over wired networks where bit-level corruption is rare and packet loss
implies that all the bits of the packet were lost. However, in case of wireless network,
all the bits in a packet do not share the same fate. Thus, it is wasteful to resend
the entire packet. Therefore, in the proposed technique, instead of retransmitting
entire packet when negative acknowledgement (NACK) is received, sender sends
only that bits which are erroneous by using basic three-bit repetition code. After
XORing, receiver sends NACK along with errored bit positions information to the
sender and keeping the errored copies at its buffer. At the sender side, instead of
retransmitting the whole packet, it sends only those bits which are detected to be
error by repeating each bit three times. Receiver applies bit-by-bit majority logic
in the third erroneous copy, and again error detection is done on the code obtained
after majority logic. Incorrect positions are rectified by comparing the bits obtained
after majority logic with those erroneous bit locations in any of the first two received
copies. Flow diagram of the proposed scheme assuming return channel to be error
free, the round-trip delay is negligible, and CRC used at the receiver provides perfect
error detection with example is shown in Fig. 1. In Fig. 1, P1_C1 is the copy 1 of
packet 1, and P1_C2 stands for copy 2 of packet 1. P1_Repetition_code stands for
retransmitted packet with three-bit repetition code.
In the proposed scheme, aggregate number of bits transmitted is lesser than the
conventional APC, so the throughput improvement is very high as compared to the
conventional APC. It gives satisfactory results when the packet is medium and large
sized.
Majority logic applied in the transmitted packet is successful if no two bits out
of three repetition code is error. Generally, codeword to have two bits flipped out
of three bits, the medium would have to be extremely noisy of the order of 10−3 to
10−2 [1, 2]. And practically, channel does not happen to be extremely noisy all the
Fig. 1 Flow diagram of the proposed APC scheme with the bit repetition code
time. To understand the proposed scheme clearly, let us take an example showing
proposed scheme.
An example showing proposed scheme for the packet b7 b6 b5 b4 b3 b2 b1 b0 =
11001100
Received copy, b7b6b5 —
b4 b3b2b1b0 = 110 −
1 1100
— —
Received copy, b7 b6 b4 b5 b3b2 b6 b0 = 1 −
0 0011 −
10
XOR = 01010010
The incorrect positions are identified as second, fourth, and seventh bit locations
from the left. Receiver now sends NACK along with this errored bit positions infor-
mation to the sender. Sender sends three-bit repetition code of this errored positions
as 111 000 000. In receiver side, if the received code is 101 100 000. Apply majority
logic, we get code as 100, and this code is sent for error detection. If found correct,
receiver checks correct packet using this code in one of the erroneous copies stored
in buffer of the receiver.
Error detection is done to check the packet.
3 Analysis and Simulation Results
In conventional APC scheme, three copies (i = 3) of a packet is transmitted at a

time to get reliable packet at the receiver and the actual bit error rate, £ (as then only
double or triple bits in error cannot be detected leading to actual bit error rate) for
APC is given by ζ [9]:

3 2 3 3
ζ = α (1 − α) + α (1)
2 3
Therefore, probability of erroneous transmission (PAPC ) or packet error rate (PER)

in conventional APC for packet size “n” is given by [9]:

PAPC = 1 − (1 − ζ )n (2)
where “α” is bit error rate and ζ is the actual bit error rate. Throughput of the
conventional APC is given by ηAPC [15]:
(1 − PAPC )
ηAPC =
(i + PAPC )
(1 − PAPC )
ηAPC = (3)
(3 + PAPC )
where “i” is the number of copies transmitted.

In the proposed scheme, initially two copies are transmitted. At the receiver side,
error detection is performed. If both the copies found erroneous, bit-wise modulo-2
sum is applied to find the error locations. Once error locations are identified, NACK
is sent along with error location information. Therefore, probability of erroneous
transmission PPC for the first two copies assuming no double error occurs is given
by [16]:
PPC = (1 − (1 − α)ni )
PPC = (1 − (1 − α)2n ) (4)
Since two copies are transmitted in the first go therefore, i = 2.

For the third copy with three-bit repetition code, if any of the retransmitted bits
cannot be recovered that is the case when two or all the three bits are erroneous.
Therefore, bit(s) which cannot be recovered is written as β:
β = α 3 + 3 C2 α 2 (1 − α) = α 3 + 3α 2 − 3α 3 = 3α 2 − 2α 3 (5)
Therefore, packet which is recoverable is given as φ:
φ = 1 − β = 1 − (3α 2 − 2α 3 ) = 1 − 3α 2 + 2α 3 (6)
Therefore, packet of k bits in error which are recoverable
= (1 − 3α 2 + 2α 3 )k (7)
Probability of erroneous transmission for the proposed scheme PNew is given by:
PNew = PPC × [(1 − 3α 2 + 2α 3 )k ] (8)
Therefore, throughput of the proposed scheme is given by ηNew :

(1 − PNew )
ηNew =
(i + PNew )
(1 − PNew )
ηNew = (9)
[(2 + 3k) + PNew ]
Performance evaluation is done in terms of packet error rate (PER) and packet
throughput, as determined by the computer simulation using MATLAB R2012b tool.
Bit error rate (α) in the range 10−5 to 10−3 (low error rate) and 10−3 to 10−2 (high error
rate) is taken for different values of bits in error (k) for comparison. Simulation results
show the effectiveness of the proposed scheme in terms of throughput improvement
for large-sized packet (n = 1000). But PER is almost same as that of the conventional
APC as observed from simulation results in Fig. 4.
Figures 2 and 3 show that the throughput of the proposed scheme increases with
change of bit error rate from low-to-high bit error rate, and this increment is more
for large-sized packet. Throughput of the conventional APC decreases with increase
in bit error rate. Not much improvement can be observed in terms of packet error
rate (P) in the proposed scheme as shown in Figs. 4 and 5. It becomes worst when
the packet size and bit error rate increase as chances of more number of bits getting
corrupted increase in wireless links.
90
80
70
60
Throughput,Vmod
APC Scheme
50
New scheme with Vmod1 for k1=2
Vmod2 for k2=30
40 Vmod3 for k3=300
30
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-3
x 10
Bit Error Rate,a
Fig. 2 Throughput plot of the proposed scheme for different values of k and APC scheme for n =
1000
70
APC Scheme
60 New scheme with Vmod1 for k1=2
Vmod2 for k2=30
Vmod3 for k3=300
50
Throughput,Vmod
40
30
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-3
x 10
Bit Error Rate,a
Fig. 3 Throughput plot of the proposed scheme for different values of k and APC scheme for n =
500
100
90
80
70
Packet Error Rate,P
60
50
40 APC Scheme, Papc

New scheme with Pnew1 for k1=2
30 Pnew2 for k2=30
Pnew3 for k3=100
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-3
x 10
Bit Error Rate,a
Fig. 4 Packet error rate plot for proposed scheme for different values of k and APC scheme for n
= 1000
40
35
APC Scheme, Papc
New scheme with Pnew1 for k1=2
30 Pnew2 for k2=30
Pnew3 for k3=100
Packet Error Rate,P
25
20
15
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-3
x 10
Bit Error Rate,a
Fig. 5 Packet error rate plot for proposed scheme for different values of k and APC scheme for n
= 156
4 Conclusion
The proposed scheme of bit repetition yields higher throughput as compared to APC
scheme. Increment in throughput is prominent when the packet size and bit error
rate increases as seen from Figs. 1 and 2. Not much improvement can be observed
in terms of packet error rate (P) in the proposed scheme as shown in Figs. 3 and 4.
It becomes worst when the packet size and bit error rate increase as chances of more
number of bits getting corrupted increase in wireless links. The proposed scheme
fails when double error occurs in the first two transmitted copies and two bits out of
three transmitted repeated code get corrupted. Generally, codeword to have two bits
flipped out of three bits, the medium would have to be extremely noisy of the order
of 10−3 to 10−2 [1, 2]. And practically, channel does not happen to be extremely
noisy all the time.
References
1. Liu H, Ma H, Zarki ME, Gupta S (1997) Error control schemes for networks: an overview.
Mob Netw Appl 2:167–182
2. Yuen O (1996) Design trade-offs in cellular/PCS systems. IEEE Comm Mag 34(9):146–152
3. Chakraborty SS, et al (1995) An ARQ scheme with packet combining. IEEE Comm Lett
2(7):200–202
4. Hirayama Y, Okada H, Yamazato T, Katayama M (2005) Time-dependent analysis of the
multiple-route packet combining scheme in wireless multihop network. Int J Wirel Inf Netw
42(1):35–44
5. Leung YW (2000) Aggressive packet combining for error control in wireless networks. Trans
Comm E83(2):380–385
6. Chakraborty SS, et al (1999) An exact analysis of an adaptive GBN scheme with sliding
observation interval mechanism. IEEE Comm Lett 3(5):151–153
7. Chakraborty SS, et al (1999) An adaptive ARQ scheme with packet combining for time varying
channels. IEEE Comm Lett 3(2):52–54
8. Jamieson K, Balakrishnan H (2007) PPR: partial packet recovery for wireless networks.
SIGCOMM’07, Kyoto, Japan
9. Bhunia CT (2012) Several modifications of aggressive packet combining scheme for wireless
network. In: Proceedings of the IEEE computer, information and telecommunication systems
(CITS), pp 1–5
10. Saring Y, Bulo Y, Bhunia CT (2015) New modifications of aggressive packet combining scheme
with improved performance. Int J Appl Eng Res 10(17):37610–37615
11. Khumukcham R, Goswami A, Saring Y (2015) Four new protocols for achieving better
correction capability of APC scheme. In: Proceeding IEEE international conference of
communication and signal processing, Melmaruvathur, Tamil Nadu, pp 741–745
12. Khumukcham R, Saring Y, Goswami A (2015) Combined APC-PC scheme for random and
time varying channels. In: Proceeding IEEE international conference of communication and
signal processing, Melmaruvathur, Tamil Nadu, pp 358–360
13. Bulo Y, Sharma P, Bhunia CT (2016) Improving error correction capability of aggressive packet
combining scheme by using half-byte packet reverse technique and even–odd selection method.
Int J Comput Appl 137(2):32–36
14. Saring Y, Singh SV (2019) Aggressive packet combining scheme with packet reversed and
packet shifted copies for improved performance. IETE J Res
15. Bhunia CT (2005) IT, network and internet. New Age International Publishers India, New
Delhi
16. Chakraborty SS, Liinaharja M, Ruttik K (2002) Selection diversity based ARQ schemes and
their enhancements using packet combining in a slow rayleigh fading channel. IEEE ICPWC
192–195
Review on Silent Speech Recognition
Using EMG Sensors and Voices After
Laryngectomy
Pilla Devi Bhavani, V. Preethi, and Durgesh Nandhan
Abstract In this paper, there are four ways of speech rehabilitation techniques after
laryngectomy. They are voice prosthesis, esophageal voice, electronic larynx, and
silent articulation. After the surgical removal of the larynx, the person cannot speak.
After laryngectomy, he/she cannot speak naturally. But using the four techniques
mentioned above is helpful to a person without larynx.
Keywords Larynx · Laryngectomy · Dysarthria · Amatsu tracheoesophageal

shunt · Electromyographic signals · Tracheophageal voice · Esophageal voice
1 Introduction
Speech is the most important tool for human beings to communicate with others.
The main part of the human body which helps a person to communicate verbally is
larynx. The larynx is also known as the voice box. In grown-up people, the larynx
is found in the front neck at the degree of the C3–C6 vertebrae. It interfaces the
second rate some portion of the pharynx (hypopharynx) with the trachea. The larynx
broadens vertically from the tip of the epiglottis to the sub-par outskirt of the cricoid
ligament. The voice box comprises of folds, which are known as vocal folds. The
vocal folds are also called vocal strings. Main functions of vocal cords are:
1. To shield the airway from chocking on the material within the throat.
2. To control the flow of air into the lungs.
3. The creation of sounds utilized for speech.
P. D. Bhavani (B) · V. Preethi

Department of Electronics and Communication Engineering, Aditya College of Engineering and
Technology, Surampalem, Andhra Pradesh, India
V. Preethi
D. Nandhan
Accendere Knowledge Management Services Pvt. Ltd, CL Educate Ltd, New Delhi, India
232 P. D. Bhavani et al.
If anyone of the above functions is not going properly, it results in the problem of
larynx. On earth, many people are suffering from throat disorders. Every year, thou-
sands of people who are suffering from throat trauma (disease) require the surgical
removal of larynx. The process of removing the larynx from the human body is
called laryngectomy. In this process, the surgeons remove the damaged voice box.
After laryngectomy, the patients require the alternative method of communication.
For that purpose, they require software and augmentative. These include artificial
voice sources after the loss of laryngeal function. But these kinds of artificial sources
may provide an unnatural voice to the patient. These sources may also require the
involvement of users’ hands which is complex for the user [1].
There is an innovation that enables individuals with discourse to scatter is auto-
matic speech recognition. In this innovation, the acoustic discourse is converted into
the grouping of tokens utilizing design characterization strategies. In this procedure,
there are numerical techniques for examining the connection between the vocal tract
and its acoustic yield. The connection between the state of the vocal tract and the
acoustic yield is spoken to with a scientific articulation that is y = f (x), where y is
the acoustic yield and x is the vocal tract shape. Here ‘y’ can be registered for any
estimation of ‘x.’ Furthermore, for rearranging ‘x,’ there is a method. Reversal by
PC arranging comprises of processing ‘y’ for some estimation of ‘x’ and arranging
the subsequent (y, x) sets into an advantageous request as per y, x for a given y is then
gotten by looking into y in the arranged information. Along these lines, the acoustic
yield is arranged (interpreted) into the succession [2].
2 Literature Review
In 1992, Jan Noyes and Clive Frankish estimated that about 15 million people in
the United States were suffering from throat disorder. The main application of auto-
matic speech recognition (ASR) is helping people with disabilities. In this paper,
it is discussed about the advantages of speech recognition. In the last decade, the
recognition performance has been increased due to the increase in the processing
power of the microcomputers. The processing power of microcomputers has been
increased due to the development of more sophisticated techniques. Now, the speech
recognizers are being considered in the industrial tasks and office and aerospace
environments. Mainly in three areas, the disabled users are using the speech recog-
nizers, and the areas are 1. communication, 2. environmental control, and 3. medical
applications [3].
In 1997, Fangxin Chen and A. Kostov developed the methods for improving the
performance of automatic speech recognition (ASR) of slurred speech (dysarthric
speech) by exploring the residual vocal ability of people with severe dysarthria. The
main application of this technology is to develop the communication and control
devices for the dysarthric people. To detect the words in the high-level dysarthric
speech, there is an algorithm. A discrete hidden Markov model (DHMM) was also
adopted for speech pattern recognition [4].
Review on Silent Speech Recognition … 233
In 1998, Mauro Becker presented a paper by explaining the importance of

Amatsu tracheoesophageal shunt for the patients after laryngectomy. Previously,
Radiotherapy and Hypopharynx lesions are employed for speech rehabilitation. By
comparing the radiotherapy and hypopharynx lesions with the Amatsu tracheoe-
sophageal shunt is much better. By using this Amatsu tracheoesophageal shunt 76%
of vocalization was achieved. Since March 1991, this operation was done in 2 women
and 33 men. And their ages are between 30 and 78 years. The main advantage of the
Amatsu tracheoesophageal shunt is its inexpensive nature [5].
In 2005, P. D. Polur and G. E. Miller were presented about the Markov model
which is helpful to increase the performance of the dysarthric speech recognition
system. Dysarthric speech is also known as slurred speech. Dysarthria occurs due to
neurological problems. The main aim is to allow the dysarthric patients to control
assistive technology. Therefore, this dysarthric speech recognition system acts as a
control tool. The fast Fourier transforms are used in dysarthric speech recognition
[6].
In 2006, Szuchen Jou, Lena Maierhein, Tanja Schultz, Alex Waibel were presented
about the classification of articulatory features based on electromyographic signals.
The articulatory features are manner, place, voicing, rounding, front-back, static.
The anticipatory behavior of electromyographic signals can be identified by using
the recorded audible speech and electromyographic signals. Every articulator has its
anticipatory behavior [7].
In 2009, James Heaton, Glen Colby was reported that five surface electromyo-
graphy sensors were enough for automatic speech recognition. The main aim is to
reduce the number of sensors because before 2009 authors reported that 11 surface
electromyographic sensors are required for automatic speech recognition and they
are fixed on the face and neck of the persons to recognize the speech. The sensors
are located on the muscles which are responsible for speech production [8].
In 2010, Harsha Vardhan Sharma proposed the speaker-adaptive recognizers for
the persons with spastic dysarthria. The spastic dysarthria is also known as upper
motor neuron dysarthria. It occurs due to the strokes, trauma. In this, there are two
modifications. They are (1) MAP (maximum a posterior) adaptation of speaker-
independent systems, (2) transition-interpolated (linear interpolation between fully
ergodic and left–right structures for both speakers adapted and speaker-dependent
systems). Based on the above two modifications, there are four results after a few
experiments [9].
In 2016, Tanmay Bhowmik reported that the two articulatory features have been
identified by using training data. After the identification of features, phonemes are
created by merging or combining the features. This experiment was conducted on
two persons of Bengal and England, respectively [4].
3 Methodology
There is a well-known way to recognize silent speech by using surface electromyo-

graphic sensors. In this, there are two subways in which silent speech can be recog-
nized. But the main component in both the ways is the surface electromyographic
sensors. The first one is (a). silent speech recognition using 11 surface electromyo-
graphic sensors, and the second one is (b). silent speech recognition using six surface
electromyographic sensors. The two ways are briefly described below.
3.1 Speech Recognition Using Eleven s-EMG Sensors
Firstly, eleven surface electromyographic sensors are used to recognize the silent
speech. The positions of the s-EMG are fixed. In this case, four sensors are fixed on
the face and the remaining seven sensors are fixed on the neck. The sensors which are
placed on the face are EMG-8, EMG-9, EMG-10, EMG-11. The sensors which are
placed on the neck are EMG-1, EMG-2, EMG-3, EMG-4, EMG-5, EMG-6, EMG-7.
The positions of EMG sensors are as Fig. 1.
This experiment was conducted on nine different speakers. In that nine members,
four were female and five were male. This experiment was conducted by using 65
words as given below in Table 1. When the experiment was conducted, there was a
problem with the order of the words. That problem was, every speaker pronounced
the same word for three times.
Fig. 1 Person with 11 s-electromyographic sensors

Table 1 65 words used for testing

No Fire That Right Forward Negative Maneuver
Go Five This Seven Latitude Hibernate Affirmative
One Four Zero Target Location Position Coordinates
Six Help Abort Three Measure Proceed Kilometers
Ten Days Block Assist Recover Transmit Rendezvous
Two Nine Eight Brings Reverse Thousand
Yes Pull Hours Cancel Seconds Longitude
Left Push Miles Meters Hundred Standby
Fast Slow Point Collect Monitor Shutdown
Feet Stop Reach Advance Continue Accelerate
Table 2 Recognition
Speaker Recognition approximated
approximated accuracy of
accuracy
speakers
Number Gender Mouthed (%) Vocalized (%)
1 Female 87 88
2 Female 83 93
3 Female 90 87
4 Male 78 91
5 Male 93 96
6 Male 91 98
7 Male 89 97
8 Female 89 85
9 Male 69 91
Mean value – 85.44 91.77
Recognition: As this experiment was conducted on nine members, their speech

modalities were calculated. The speech modalities are mouthed and vocalized. The
recognition rate was also calculated (Table 2).
Graph: After getting the recognition rate. The graph was plotted by taking the subset
size on the x-axis and the rate of recognition on the y-axis (Fig. 2).
3.2 Speech Recognition Using Six s-EMG Sensors
In this method, only six s-EMG sensors are used to recognize the silent speech.
Electromyography is a medical technique used to record the electrical activities of
the skeletal muscles. The positions of the six sensors are given below.
Fig. 2 Recognition rate
• EMG-1 => Under the chin

• EMG-2 => On the cheek (left side of the nose)
• EMG-3 => On the cheek (right side of the nose)
• EMG-4 => Bellow the EMG-3
• EMG-5 => Under the lower lip
• EMG-6 => On the neck.
This technique works very efficiently because of only six surface electromyo-
graphic sensors. This technique is affordable when compared with the technique of
11 sensors.
4 Speech Rehabilitation After Laryngectomy
There are four ways for the rehabilitation of speech after laryngectomy. The laryn-
gectomy is the surgical removal of the larynx from the human body. The larynx is
the most important tool for a person to communicate with others. The other name
for larynx is the voice box. The larynx is located above the trachea and in front of
the esophagus. With the removal of the larynx, a person cannot speak.
The four ways of speech rehabilitation are given below.
• Voice prosthesis
• Esophageal voice
• Electronic larynx
• Silent articulation.
Fig. 3 Air movement from

lungs to mouth
4.1 Voice Prosthesis
Voice prosthesis is also known as the speaking valve. This valve plays a major role
in helping the person to speak after laryngectomy. The valve is made up of silicon.
The valve is fixed inside the body of the patient at the position of the larynx. It is the
fixed back of the stoma and connects the trachea with the esophageal fistula (Fig. 3).
When the person wants to speak, then he/she has to push the valve inside. Then,
the airway will be formed between the windpipe and the esophagus. When the person
pushes the valve, then the air from the lungs passes through the valve and reaches the
esophagus which makes the person speak. But the voice of that person is not natural,
and that voice is known as tracheoesophageal voice (Fig. 4).
4.2 Esophageal Voice
The person with an esophageal voice has to push the air from his/her mouth down
to the esophagus. And he/she has to back up the air again. As the air passes, the
pharyngoesophageal segment vibrates, which results in the production of speech.
For this kind of speech, the person has to train very well. Because he/she has to
swallow the air from the mouth and again he/she has to release it out all the time
(when he/she want to speak). But this voice is not fluent and louder.
Fig. 4 Person with voice prosthesis
4.3 Electronic Larynx
The electronic larynx is also known as servox. It is most commonly used in the UK.
It is a mechanical vibrating device. It has no connection with the body. But the person
has to carry it along with him (Fig. 5).
When the person wants to speak, then he/she has to place that device at the side of
the neck. Then, the electronic signals pass through the tissues of the face and reach
the mouth. Then, a kind of electronic sound (speech) will be produced (Fig. 6).
Fig. 5 Electronic larynx

Fig. 6 Person with

electronic larynx
4.4 Silent Articulation
This is a kind of voice, in which a person may not speak using the speaking valve
and he/she may not have the esophageal voice and he/she may not be able to have
the electronic larynx. Then, that person will articulate silently. To understand what
that person is telling, others should watch the person’s lips very carefully.
5 Conclusion
This paper clearly shows the techniques of speech rehabilitation after laryngectomy.
This paper also shows the types of voices after laryngectomy. It also consists of
speech rehabilitation techniques using 11 surface electromyography sensors and
six surface electromyographic sensors. Finally, as the speech is the most important
thing for every human being, this paper clearly shows the way for the people with a
laryngectomy.
References
1. Colby G, et al (2009) Sensor subset selection for surface electromyograpy based speech recog-
nition. In: ICASSP, IEEE international conference on acoustics, speech and signal processing,
pp 473–476
2. Atal BS, Chang JJ, Mathews MV, Tukey JW (1978) Inversion of articulatory-to-acoustic trans-
formation in the vocal tract by a computer-sorting technique. J Acoust Soc Am 63(5):1535–1555.
https://doi.org/10.1121/1.381848
3. Noyes JM, Frankish CR (1992) Speech recognition technology for individuals with disabilities.
Augment Altern Commun 8(4):297–303. https://doi.org/10.1080/07434619212331276333
4. Bhowmik T, et al (2017) A comparative study on phonological feature detection from continuous

speech with respect to variable corpus size. In: 2016 IEEE Students’ Technol Symp Tech Sym,
pp 311–316. https://doi.org/10.1109/TechSym.2016.7872702
5. Polur PD, Miller GE (2005) Predictive and cepstral coefficients in dysarthric speech 13(4):558–
561
6. Vieira MBM et al (1999) Speech rehabilitation after laryngectomy with the Amatsu tracheoe-
sophageal shunt. Auris Nasus Larynx 26(1):69–77. https://doi.org/10.1016/S0385-8146(98)000
65-0
7. Jou SC, Maier-Hein L, Schultz T, Waibel A (2006) Articulatory feature classification using
surface electromyography. In: 2006 IEEE international conference on acoustics speech and
signal processing proceedings, Toulouse, France, pp 605–608
8. Colby G, Heaton JT, Gilmore LD, Sroka J, Deng Y, Cabrera J, Roy S, De Luca CJ, Meltzner GS
(2009) Sensor subset selection for surface electromyograpy based speech recognition. In: 2009
IEEE international conference on acoustics, speech and signal processing, Taipei, Taiwan
9. Meltzner GS et al (2017) Silent speech recognition as an alternative communication device for
persons with laryngectomy. IEEE/ACM Trans Audio Speech Lang Process 25(12):2386–2398.
https://doi.org/10.1109/TASLP.2017.2740000
CNN-Based Individual Ghungroo Breed
Identification Using Face-Based Image
Kaushik Mukherjee, Sanket Dan, Kunal Roy, Subhojit Roy,

Subhranil Mustafi, Pritam Ghosh, Satyendra Nath Mandal,
Dilip Kumar Hajra, Santanu Banik, and Syamal Naskar
Abstract Convolutional Neural Network (CNN) is widely used as a strong frame-

work for object classification. This framework is suitable for classification of group
of objects which carry similar features like animal breeds, plant varieties, etc. In this
paper, an attempt has been made to build a model of identifying individual Ghungroo
pig (one of the most popular registered pig breeds in India) based on its face image
using CNN. Individual Ghungroo pigs look very similar, and it is a very challenging
task even for humans to identify them from a group of Ghungroo pigs. Ten Ghungroo
pigs from ICAR-Indian Veterinary Research Institute (IVRI) Animal Farm Kalyani
(Block B, Kanchrapara, West Bengal 741235) have been captured using a DSLR
camera. The pigs have been captured in natural environment without any restriction.
The Inception-v3 model with transfer learning has been used to build the individual
pig identification model. The captured images have been separated with 90%:10%
into training images and test images sets, respectively. It has been observed that the
model has predicted individual pigs with 100% accuracy at 0.87 confidence level.
K. Mukherjee · S. Dan · K. Roy · S. Mustafi · P. Ghosh · S. N. Mandal (B)

K. Mukherjee
S. Roy
Gangarampur Government Polytechnic College, Gangarampur-Belbari Road, Kadighat,
Gangarampur, West Bengal 733124, India
D. K. Hajra
Department of Agronomy, Faculty of Agriculture, UBKV, Pundibari, Cooch Behar 736165, India
S. Banik
ICAR-National Research Centre On Pig, Rani, Guwahati, Assam 781131, India
S. Naskar
Indian Veterinary Research Institute (IVRI) Eastern Regional Station (ERS), 37, Belgachia Road,
Kolkata 700037, India
242 K. Mukherjee et al.
Keywords Deep learning · Convolutional neural network (CNN) · Object

classification · Ghungroo pigs · Accuracy · Confidence level
1 Introduction
It is very obvious for any farm to track their animal (individual as well as breed) in an
efficient manner; otherwise, it leads to some fraudulent activity. Traditional ways to
identify the livestock animal such as ear tag (electronic/non-electronic), tattoo, etc.,
is not healthy enough with respect to sustainability and those are also distressing
the animal. So there has always been a search for an optimized way to identify the
livestock animal which will longlasting as well as not to distress the animal too much.
Biometric models can play a vital role in this situation. There has been a tremendous
effort on the research of finding a better biometric model to identify the livestock
animal. Some of them are iris recognition, ear venation pattern, face recognition, etc.
Some successfully identify the animal with good accuracy.
In this paper, face-based images of ten Ghungroo pigs have been captured by
a DSLR camera without making any restriction. A total of 874 images have been
capture in the natural environment. The ten distinct folders have been created with
names Pig1, Pig2, …, Pig9 and Pig10 to store captured pictures. The captured images
have been stored into individual pig folders. The images kept in individual folders
have been separated as training images: test images with 90%:10% ratio. That means
if one individual pig folder contains 100 images; then, 90 images go to training images
and the rest ten images kept for testing purposes which were not part of the training
process. One model has been proposed based on the Inception-v3 model where ten
separate outputs give ten values from 0.0 to 1.0. The feature extraction of Inception-
v3 model kept intact, but the classification outputs have been changed into ten for
ten pigs. The height value in one output has been indicated by the corresponding
individual. The model has been trained in 200,000 iterations with learning rate 0.01.
Once the training process was completed, the test process was run with the test
images. This model has achieved the recognition accuracy of 100% with confidence
level of 0.87. The CNN is generally used for set of objects classification, but the
individual pig has been identified here. This type of effort has not been done earlier.
This is the reason for making this paper.
The paper is divided as follows: In Sect. 2, previous work of the classification
process is described followed by creation of image data set in Sect. 3. A brief descrip-
tion is discussed in the Sect. 4. Finally comes to architecture of proposed model which
is shown in Sect. 5 followed by result and discussion and conclusion and future work
in Sects. 6 and 7, respectively.
CNN-Based Individual Ghungroo Breed Identification … 243
2 Previous Work
Recognition of human through face has been a popular research area for at least
five decades [1]. Historically, geometric features extracted from face images have
been used for face recognition [2, 3] and now the recent trend is to apply deep
networks which has really improved the recognition performance in a high level.
The fundamental idea of applying deep networks has come from the human nervous
system (which is nothing but accumulation of infinite number of neurons). The
face recognition system is widely used for “non-intrusive access control and moni-
toring/surveillance purposes and as such represents a potentially useful research
area to apply to the problem of pig identification”. Although there has been a lot of
related work done on livestock animals, few such as automatically identify the pig
behaviours [4], biometric on livestock animal [5–10], breed identification on live-
stock animals [11] show very good result. Although there was little research on
pig identification where pig face was used as a biometric tool, eigenface technique
[12] was used and achieved 77% recognition accuracy on ten pigs, in this paper full
manually cropped face was used. Another research was on the pig face [13] where
augmented pig face was used as a data set for CNN and achieved 96.7%. In this
paper, original data has been used without any augmentations and achieved 99.73%
recognition accuracy.
3 Image Data set
Images were collected from ICAR-Indian Veterinary Research Institute (IVRI)

Animal Farm Kalyani (Block B, Kanchrapara, West Bengal 741235) using DSLR
camera (Nikon) with 18–55 mm lens (min focal length 18 mm and optical zoom
0.38X). The farm has been visited many times and collected face images of ten indi-
vidual Ghungroo pigs. A total of 60–70 face images have collected per individual
pig and collected 874 raw images in total. After that ten folders were created for ten
individual pigs and renamed these folders as Pig1, Pig2, up to Pig10. The images are
placed in respective folders, and all the ten folders are kept in a separate folder and
renamed as “Pig Data” (Fig. 1).
4 Convolutional Neural Network and Inception V3 Model
Convolutional neural network (CNN, or ConvNet) is a special kind of multi-layer

neural networks, designed to recognize visual patterns directly from pixel images
with minimal preprocessing. It has two parts—one is feature extraction, and other
is recognition. In this paper, Inception-v3 pretrained model developed by Google
Inc. has been used to develop the proposed model. The architecture of CNN with
Fig. 1 Pig face image
convolution, max pooling and fully connected layer [14] is depicted in Fig. 2. In
the convolution layer, kernel or filter K of specific dimension is applied over each
position on the input image or feature map x. The function provides the following
output (Eq. 1) [15]

xnk = xmk−1 × K mn
k
+ bnk (1)
m∈Mn
where bn represent bias, k is the layer number and M n is the feature map from n
channels. The size of the feature map can be estimated by using Eq. (2)
input dimension − kernel dimension

size of feature map = +1 (2)
stride
The obtained feature maps from each layer are carried forward to subsequent
convolution layer. This layer is responsible for feature extraction from an input image
based on a mathematical calculation. The mathematical calculation is basically an
element-wise multiplication with respect to filter and matrix. If the input image
matrix is of dimension [a × b × c] and filter matrix is of dimension [f1 × f2 × f3],
then the dimension of output matrix will be
Fig. 2 Convolutional neural network for pig identification
(a − f 1 + 1) ∗ (b − f 2 + 1) ∗ 1 (3)
Pooling layer (depicted in Fig. 3) is responsible for reduction in the dimension

of the feature matrix obtained from the convolution layer in order to reduce compu-
tational cost as well as to prevent overfitting. Among different types of pooling,
max pooling is popular and widely used for its ability to reduce the dimension
without losing important information. After creating feature vector, it is then passed
through a dense neural network for classification. This dense network is called fully
connected layer (Fig. 4). Actually in fully connected layer, the classification is deter-
mined by the softmax which is nothing but an activation function. The softmax layer
gives the output as 0 and 1. The following equation is responsible for the softmax
function (Eq. 4).
Fig. 3 Max pooling in CNN

Fig. 4 Fully connected (or

dense) layer in CNN
ez j
σ (z) j = K (4)
k=1 ez k
Above equation (Eq. 4) implies that the output activations are all positive, since the
exponential function is positive. By contrast, if the output layer was a sigmoid layer,
then certainly could not assume that the activations formed a probability distribution.
This cannot explicitly be proved, but it might be plausible that the activations from
a sigmoid layer will not be in general form of a probability distribution. And so
a sigmoid output layer does not have such a simple interpretation of the output
activations.
5 Proposed Model
In this paper, the prediction model has been proposed based on Inception-
v3 pretrained model in transfer learning mode. The feature extraction layers of
Inception-v3 have been kept intact, and the classification part has been changed
for the individual pig identification. Finally, the pretrained Inception-v3 has been
retrained with training pig image data set. The retrained model has been tested with
pig image data set as shown in Fig. 5. The images of all individual pigs have been
separated in certain distinct folders and renamed them to Pig1, Pig2, …, Pig10 and
put them in a folder called Pig Data. The subfolder names were written (as in the
order in Pig Data) in the “retrained labels.txt” file. After training, a “tmp” folder
was created automatically in the system disc drive where images were saved and
labelled. From the “tmp” folder, the “output graph.pb” and “output labels.txt” files
were copied and then pasted in the folder where “classify.py” file was located. This
process enabled the programme to classify test images in the newly trained model.
Fig. 5 Proposed architecture for individual pig identification based on Inception-v3
As there were ten subfolders, then there were ten output lines which gave the ten
distinct values, and the highest value implies the respective pig.
6 Result and Discussion
The model has ten outputs for identification of ten individual pigs. There were ten
distinct values for every result which has been produced from every test image. The
maximum number of those results implies the respective pig. That means if the test
result was 0.00004, 0.98863, 0.00013, 0.00009, 0.00026, 0.00431, 0.00059, 0.00402,
0.00089, 0.00104 (all are representing the confidence level) for Pig1, …, Pig10,
respectively, for a test image, then the result implies that the input test image was
Pig2. Table 1 consists of the overall result. The model has been tested with various
confidences levels from 0.87 to 0.99 for all test pig images as shown in Table 2,
Fig. 6 depicts the graphical representation of accuracy vs. different confidence level
with respect to pig, and Fig. 7 depicts the average accuracy vs confidence level. The
accuracy of proposed model has been compared with some accuracies of different
problem of object detection published in different research paper is shown in Table 3.
Table 1 Summary of
Ghungroo_pig Accuracy in Accuracy in test set
training set and test set results
individual training set (%) (%)
(with confidence level 0.87)
Pig1 100 100
Pig2 100 100
Pig3 100 100
Pig4 100 100
Pig5 100 100
Pig6 100 100
Pig7 100 100
Pig8 100 100
Pig9 100 100
Pig10 100 100
The images of individual pig have been captured in natural environment without any
restriction, and all captured images have been used in training and testing sets. There
was no threshold of accepting images based on pixel resolution, lighting conditions
and distance between camera and pig and also their movements. That is why the
accuracies are decreasing with increasing of confidence level. The model has given
lower accuracy in increasing the confidence level. The accuracy is given in published
paper without mentioning confidence level.
In this paper, the proposed model has given prediction accuracies from 100 to 76.67%
with confidence levels from 0.87 to 0.99. The result has been given based on limited
pig image data set (total 874 images). The pictures have been captured without
restraining the pigs and without any control. The model has given better result
compared to other results published in different research papers. The accuracies
have been obtained from all captured images without image accepting threshold like
pixel resolution, lighting conditions and distance between camera and pig and also
their movements. The model will be applied for prediction on more pigs, and it will
be tested on other pig breeds. The pig images will be collected from various ages
of the same pig, from different organized farms for same or different pig breeds in
future.
Table 2 Confidence level versus accuracy
Sl. No Pig Accuracy in various confidence levels (in %)
0.99 0.98 0.97 0.96 0.95 0.94 0.93 0.92 0.91 0.90 0.89 0.88 0.87
1. Pig1 83.33 83.33 83.33 83.33 83.33 83.33 83.33 83.33 83.33 100 100 100 100
2. Pig2 83.33 100 100 100 100 100 100 100 100 100 100 100 100
3. Pig3 50 66.67 66.67 66.67 83.33 83.33 100 100 100 100 100 100 100
4. Pig4 83.33 83.33 100 100 100 100 100 100 100 100 100 100 100
5. Pig5 100 100 100 100 100 100 100 100 100 100 100 100 100
6. Pig6 50 50 50 50 66.67 66.67 66.67 66.67 83.33 100 100 100 100
CNN-Based Individual Ghungroo Breed Identification …
7. Pig7 83.33 100 100 100 100 100 100 100 100 100 100 100 100
8. Pig8 83.33 83.33 83.33 83.33 100 100 100 100 100 100 100 100 100
9. Pig9 66.67 66.67 66.67 66.67 66.67 83.33 83.33 83.33 83.33 100 100 100 100
10. Pig10 83.33 83.33 83.33 83.33 83.33 83.33 83.33 83.33 83.33 83.33 83.33 83.33 100
Avg._accuracy 76.665 81.666 83.333 83.333 88.333 89.999 91.666 91.666 93.333 98.333 98.333 98.333 100
249
Fig. 6 Accuracy versus different confidence level with respect to pig
Fig. 7 Average accuracy

versus confidence level
Table 3 Accuracy level different object prediction

Author Crop Deep learning model Accuracy (%)
Liu et al. [15] Apple AlexNet precursors 97.62
with inception module
Ma et al. [16] Cucumber DCNN, based on 93.4
Lenet5
Aravind et al. [17] Grape AlexNet with SVM 99.23
Brahimi et al. [18] Tomato AlexNet 97.71
Lu et al. [19] Rice DCNN 95.48
Proposed algorithm Individual pig identification Proposed model (with 100
confidence level 0.87)
Acknowledgements The authors would like to thank Dr. Amitabha Bandyopadhyay, Senior
Consultant, ITRA Ag&Food for his valuable suggestions in this research work. The authors would
also like to thank Dr. Sourabh Kumar Das, Principal, Kalyani Government Engineering College,
Kalyani, Nadia, for his continuous support.
References
1. Bledsoe WW (1966) Man-machine facial recognition. Panoramic Research Inc., Palo Alto, CA
2. Turk M, Pentland A (1991) Eigenfaces for recognition. J Cognit Neurosci 3(1):71–86
3. Belhumeur PN, Hespanha JP, Kriegman DJ (1997) Eigenfaces versus fisher faces: recognition
using class specific linear projection. IEEE Trans Pattern Anal Mach Intell 19(7):711–720
4. P. Ahrendt, T. Gregersen, H. Karstoft (May, 2011) Development of a real-time computer vision
system for tracking loose-housed pigs. Comput Electron Agric 76(2):169–174
5. Gaber T, Tharwat A, Hassanien AE, Snasel V (March, 2016) Biometric cattle identification
approach based on Weber’s Local Descriptor and AdaBoost classifier. Comput Electron Agric
122(Supplement C):55–66
6. Kumar S, Singh SK, Singh AK (October, 2017) Muzzle point pattern based techniques for
individual cattle identification. IET Image Process 11:805–814
7. Tharwat A, Gaber T, Hassanien AE (November, 2014) Cattle identification based on Muzzle
images using Gabor features and SVM classifier, advanced machine learning technologies
and applications, communications in computer and information science, Springer, Cham, pp
236–247
8. Allen A, Golden B, Taylor M, Patterson D, Henriksen D, Skuce R (July, 2008) Evaluation
of retinal imaging technology for the biometric identification of bovine animals in Northern
Ireland. Livest Sci 116(1):42–52
9. Gonzales Barron U, Corkery G, Barry B, Butler F, McDonnell K, Ward S (March, 2008)
Assessment of retinal recognition technology as a biometric method for sheep identification.
Comput Electron Agric 60(2):156–166
10. Kumar S, Singh SK (December, 2016) Monitoring of pet animal in smart cities using animal
biometrics. Future Gener Comput Syst
11. Mandal SN, Ghosh P, Mukherjee K, Dan S, Mustafi S et al (2020) InceptGI: a ConvNet-Based
classification model for identifying goat breeds in India. J Inst Eng (India): Ser B
12. Wada N, Shinya M, Shiraishi M (2013) Pig face recognition using Eigenspace method. ITE
Trans Media Technol Appl 1(4):328–332
13. Hansen MF, Smith ML, Smith LN, Salter MG, Baxter EM, Farish M, Grieve B (2018) Towards
on-farm pig face recognition using convolutional neural networks. Comput Ind 98:145–152
14. Szegedy C (2015) Going deeper with convolutions. Proceedings of the IEEE conference on
computer vision and pattern recognition
15. Liu B, Zhang Y, He D, Li Y (2018) Identification of apple leaf diseases based on deep
convolutional neural networks. Symmetry 10(1):11
16. Ma J, Keming D, Zheng F, Zhang L, Gong Z, Sun Z (2018) A recognition method for cucumber
diseases using leaf symptom images based on deep convolutional neural network. Comput
Electron Agric 154:18–24
17. Aravind KR et al (2018) Grape crop disease classification using transfer learning approach.
International conference on ISMAC in Computational Vision and Bio-Engineering. Springer,
Cham
18. Brahimi M, Boukhalfa K, Moussaoui A (2017) Deep learning for tomato diseases: classification
and symptoms visualization. Appl Artif Intell 31(4):299–315
19. Yang L, Yi S, Zeng N, Liu Y, Zhang Y (2017) Identification of rice diseases using deep
convolutional neural networks. Neurocomputing 267:378–384
Development of Image-Based Disease
Scale of Phoma Blight of Potato Using
k-Means Clustering
Kunal Roy, Sanket Dan, Kaushik Mukherjee, Subhranil Mustafi,

Pritam Ghosh, Satyendra Nath Mandal, Subrata Dutta,
and Ashis Chakraborty
Abstract Disease identification of the plant at an early stage is a key to prevent

the major diseases by minimal application of chemical pesticides. In West Bengal,
phoma blight is now becoming an emerging dreaded disease of potato. Phoma blight
is associated with development of numerous spots on leaflets, thereby reducing green
photosynthetic area causing huge loss in potato tuber production. The disease rating
scale for severity analysis has not been developed for this disease till now. In this
paper, an image-based phoma blight disease rating scale has been developed using
k-means clustering. The image of phoma blight affecting potato leaflets has been
captured using a DSLR camera by placing white paper background of leaflets. The
percentage of affected areas has been calculated and an image-based phoma disease
scale has been developed. The number of affected leaflet images has been given to
several plant pathologists and they have assigned disease rating scores based on eye
estimations. The score has been assigned for each leaflet considering the maximum
number of same scores assigned by plant pathologists. The disease rating has also
been assigned based on the actual affected area within each leaflet image using k-
means clustering. The comparison has been performed between eye estimated scoring
and k-means-based scoring to verify the scale. Finally, it has been observed that the
new developed disease rating scale has given 87% accuracy in disease estimation
with plant pathologists.
Keywords Phoma blight · k-means clustering · Euclidean distance · Disease rating

scale
K. Roy · S. Dan · K. Mukherjee · S. Mustafi · P. Ghosh · S. N. Mandal (B)

K. Roy
S. Dutta · A. Chakraborty
Department of Plant Pathology, Bidhan Chandra Krishi Viswavidyalaya, Mohanpur, West Bengal
741252, India
254 K. Roy et al.
1 Introduction
Phoma blight is widely visible in many subcontinents which causes severe damages
in the potato production. It is a fungal disease. The symptom of the disease is found on
the leaflets. The early detection of the disease can be reduced the economic loss of the
farmer. Traditionally, phoma blight is detected through the naked eye observation by
the crop scout. The intensity of disease is measured by expert and plant pathologists.
They are assigned numbers, 0, 1, 3, 5, 7 and 9 based on eye estimated affected area
within leaflet, leaf and whole plant. k-means algorithm provides good clustering
result based on the Euclidean distance between several data points [1]. Disease of
the leaf spot can be identified using k-means clustering method as the color intensity
of the spot is different from the leaf color [2]. Prajapati et al. had identified the rice
plant disease using centroid feeding-based k-means clustering method and extract
the various features like shape, color and texture. Based on different features, they
have developed a support vector machine (SVM) for multiclass classification [3].
Manikrao et al. have identified disease of the cotton crop using image processing
technique [4]. In their proposed method, HSV-based model has used and classification
has done based on SVM method. Zhang et al. have used the k-nearest neighbor method
classifier to recognize the plant disease. [5]. Khirade et al. in their paper have used
the ANN method for classification of disease spot after considering H and B color
feature from HSV and LAB image [6].
In this paper, images of the unaffected and affected potato leaves due to phoma
blight have been captured with a DSLR camera placing white A4 paper below them.
The captured pictures are sent to seven plant pathologists for assigning disease scores
on each affected leaflet. One disease score has been assigned on affected leaflet after
considering same maximum scoring assigned by maximum plant pathologists. The
same affected leaflets have also been processed for calculating actual area affected
by phoma using k-means clustering. At first, all captured images are resized to make
them similar in size. The k-means clustering has been used to cluster green area
and affected portion of leaflets. The affected area and whole area of each leaflet
have been calculated and percentage of affected area has been measured. One new
disease rating scale 0, 1, 3, 5, 7 and 9 has been proposed and image-based scoring
has been assigned on same leaflets which have been sent to plant pathologists. One
comparison has been made between eye estimated scoring and k-means-based scoring
to verify the proposed scale. This type of effort has not been made in measurement of
intensity of phoma blight earlier. This is the reason for making this paper. The paper
is divided as follows: Introduction in Sect. 1. In Sect. 2, creation of Phoma Blight
Image Dataset, disease rating scale and k-means clustering are described. A brief
description proposed algorithm for image-based disease rating scale development is
discussed in Sect. 3. Finally comes to result and discussion which is shown in Sect. 4
followed by conclusion and future work in Sect. 5.
Development of Image-Based Disease Scale … 255
2 Phoma Blight Dataset, Disease Rating Scale and k-Means

Clustering
Phoma Blight Image Dataset

The unaffected and affected potato leaflets, leaves and plants have been captured by
DSLR camera at c block farm, Bidhan Chandra Krishividyalaya, Kalyani, Nadia,
West Bengal-741235. The 1000 potato leaflet has been captured during image
acquisition and stored them at phoma blight potato leaflets database.
Disease Rating Scale
The intensities of diseases are measured using disease rating scales. The healthy
leaves or plants are assigned scale “0” and dry leaves or plants or maximum portions
which are affected by disease are assigned scale “9.” The scores between 0 and 9 are
divided into either 1, 3, 5, 7 or 1, 2, 3, 4, 5, 6, 7, 8 scales. The scores are assigned
by plant pathologists based on their eye estimation. The percentages of affected area
along with their associated scores based on eye estimation are given in Fig. 1.
k-means Clustering
k-means is the simplest method to solve the clustering problem. Main idea of this
clustering is to find out k centroid for each cluster. Centroid should be placed in the
different location of the cluster space. Each pixel has a specified location. Assign
each pixel to the group which is closest to the centroid. This grouping has been done
based on the Euclidean distance metric. This algorithm aims at minimizing a squared
error function. The equation is given in Fig. 2.
Fig. 1 Percentages of
affected leaflets
256 K. Roy et al.
where is a chosen distance measure between a data point and the cluster
centre , is an indicator of the distance of the n data points from their respective cluster
centres.
Fig. 2 Objective function of k-means clustering
3 Proposed Algorithm for Image-Based Disease Rating

Scale Development
The algorithm named “IBPBRS” has been proposed to calculate actual phoma blight
affected area of leaflet images of potato. The algorithm will take phoma blight affected
leaflets and it will give percentage of affected area due to phoma blight. Finally, one
disease scoring will be given based on percentage of affected area. The algorithm
has been presented into flow chart in Fig. 3.
The detail descriptions of algorithm are as follows.
Image Acquisition
Images of both healthy and disease affected leaves have been captured using DSLR
camera in different daylight conditions. The dry potato leaves and soil color are
looked similar. To reduce this problem, the leaflets, leaves and plants have been
captured by placing a white A4 paper below them. Images with low contrast,
unwanted leaflets and noise were discarded. Image acquisition process is shown
in Fig. 4.
Pre-processing
Both healthy and phoma affected leaf images acquisition process. The images with
white background were resized into 1000 × 1280 pixel for making all pictures in
same size. The unwanted area of the background; outside of A4 paper have been
discarded for keeping the background as white (Fig. 5).
Affected Area Segmentation
In segmentation, the k-means clustering has been used for classification of the affected
portion due phoma blight and unaffected part. At first RGB leaflet images have been
converted into L * a * b color space. L * a * b color space consists of luminosity layer
L* which is lightness from black to white, chromaticity-layer “a*” specify where the
color falls along the red–green axis, and chromaticity-layer “b*” specify where the
color falls along the blue–yellow axis. All the color information has been stored in
a and b layer which has been used for measuring the Euclidean distance metric for
classification. In k-means clustering, each pixel has positioned a specified location
in the cluster space. In this experiment, three clusters have been used. Out of these
Fig. 3 “IBPBRS”—image-based phoma blight rating scale
Fig. 4 Image acquisition

258 K. Roy et al.
Fig. 5 Pre-processing
three clusters classification, two clusters have been used in this experiment whereas
one cluster which only contains affected area, another has full outline of the leaf
(shown in Fig. 6). The aim of this paper is to find out the total outline of leaf image
and affected portion of leaflet. Mean of the each cluster was calculated. The total
outline of the leaf images has the maximum mean value and affected portion of the
image cluster has the minimum mean (Fig. 7). We can easily isolated total leaf and
affected portion of the leaf.
Fig. 6 Image segmentation using k-means clustering

Fig. 7 Disease affected leaf after segmentation
Calculation of image-based scoring of Phoma

Total outlined leaf image and affected portion images are considered for scoring.
Total area each clustered image was measured, divided the affected leaf portion area
by the total leaf area and calculate the percentage (shown in Fig. 8).
Proposed Disease Rating Scale
One disease rating scale for phoma blight is proposed at Table 1. The scale is divided
into six grades 0, 1, 3, 5, 7 and 9 where score 0 indicates there is no disease and
9 indicates disease percentage greater than 55%. The proposed algorithm will be
calculated the percentages of affected area due to phoma blight and announce the
grade.
Fig. 8 Calculation of affected percentage
Table 1 Proposed disease rating scale

Score/Grade Percentage Disease intensity
0 0.0 No disease
1 0–10 Small disease area has been identified
3 10.1–25 Lesions area between 10.1 and 25% of total leaf
9 >55 Lesions area are greater than 55% of total leaf
260 K. Roy et al.
Fig. 9 Sample pictures from “phoma blight potato leaflets”
Phoma Blight affected Leaflets database

The leaflets have been captured keeping white A4 paper below them using DSLR
camera. The captured phoma blight affected leaflets are about 1000 pictures and they
are stored into “phoma blight potato leaflets” database. Some captured pictures with
their picture identification number from database are shown in Fig. 9.
In this paper, the captured potato phoma affected leaflets have been processed based
on proposed algorithm and scoring have been assigned on each leaflet using proposed
disease rating scores. Same pictures are sent to seven plant pathologists to assign
disease rating scores. The maximum number of scoring with the same values, the
assigned scoring of plant pathologists and scoring based on algorithm is shown in
Table 2. The matching comparison between plant pathologist scores and algorithm-
based scores of twelve pictures is shown in Table 3. The 87 pictures have been tested
and a percentage of matching scores of proposed algorithm with plant pathologist is
given in Table 4. It is observed that the overall matching score is 87%.
In this paper, proposed algorithm has been calculated percentage of affected area
from phoma blight affected leaflets and disease scores have been assigned based
on proposed disease rating scale. The algorithm has been given 87% accuracy with
Table 2 Disease rating of pathologists and proposed scale
Image ID Plant Plant Plant Plant Plant Plant Plant Maximum Scoring based
pathologist 1 pathologist 2 pathologist 3 pathologist 4 pathologist 5 pathologist 6 pathologist 7 scoring with on proposed
same value scale
130001 1 3 3 3 1 3 3 3 3
150001 3 3 3 3 3 3 3 3 3
170001 5 5 5 5 5 3 3 5 5
180001 5 7 7 7 7 5 7 7 7
Development of Image-Based Disease Scale …
190001 9 9 9 9 9 9 9 9 9
200001 7 7 5 5 7 7 7 7 5
210001 7 5 7 7 7 5 7 7 7
210002 9 9 9 9 9 9 9 9 9
230001 7 7 7 7 7 7 7 7 5
240001 3 3 3 3 5 3 3 3 3
270001 7 7 5 5 7 7 7 7 9
290001 5 5 5 5 5 7 5 5 5
261
262 K. Roy et al.
Table 3 Plant pathologist score versus proposed disease score

Image ID Plant pathologist score Proposed disease score Matching
130001 3 3 Yes
150001 3 3 Yes
170001 5 5 Yes
180001 7 7 Yes
190001 9 9 Yes
200001 7 5 No
210001 7 7 Yes
210002 9 9 Yes
230001 7 5 No
240001 3 3 Yes
270001 7 7 Yes
290001 5 5 Yes
Table 4 Matching percentage of different scoring between plant pathologist and scoring of image-
based proposed scale
Total 0 1 3 5 7 9
87 100% 77% 86% 83% 76% 100%
assigned scores of plant pathologists. It is observed that proposed cluster-based algo-

rithm has been given 100% accuracy on rating scale 0 and 9. The plant pathologists
are confused for scoring 3 and 5. The matching percentages almost same of these two
scoring. The database contained 1000 phoma blight affected potato leaflets pictures.
All pictures will be assigned score values and the disease rating scale may be updated
in future.
Acknowledgements The authors would like to thanks to Department of Higher Education,

Science and Technology and Bio-Technology, West Bengal for funding this research work (Memo
No. 33(Sanc)/ST/P/S&T/6G-41/2017 dated 12/06/2018). The authors would also like to thank
Dr. Amitabha Bandyopadhyay, Senior Consultant, ITRA Ag&Food for his continuous valuable
suggestions and Dr. Sourabh Kumar Das, principal, Kalyani Government Engineering College.
References
1. Sidqi HM, Kakbra JF (February, 2014) Image classification using K-mean algorithm. Int J Emerg
Trends Technol Comput Sci (IJETTCS) 3(1)
2. Singh MK, Chetia S (February, 2018) Diseased leaf detection using K-mean clustering and
texture features. Int J Recent Sci Res 24427–24429
3. Shah JP, Prajapati HT, Dabhi VK (2016) A survey on detection and classification of rice plant
diseases. In: International conference on current trends in advanced computing (ICCTAC), pp
1–8, IEEE
4. Manikrao ND, Vyavahare AJ (2015) Disease detection of cotton crop using image processing
technique: a survey. IJRASET 3(VI):ISSN: 2321–9653
5. Zhang SW, Shang YJ, Wang L (2015) Plant disease recognition based on plant leaf image. J
Animal Plant Sci 42–45
6. Khirade SD, Patil AB (2015) Plant disease detection using image processing. In: IEEE,
international conference on computing communication control and automation, Pune, pp
768–771
Power Quality Improvement of Cascaded
Savonius Rotor Based Hydrokinetic
Power System
Oying Doso and Sarsing Gao
Abstract Cascaded Savonius rotor based hydrokinetic power system is a novel

method that used a vertical axis wind rotor used to generate power from running water.
This article presents modelling and power quality study of Cascaded Savonius rotor
based hydrokinetic power system aimed at supplying power without affecting the
environment. The model is tested with load and with STATCOM with changing water
velocity in MATLAB environment and its performance is found to be acceptable.
Keywords Hydrokinetic power system · Savonius rotor · Power quality
1 Introduction
India is one of the major producers of hydropower in the global scenario but there
are many drawbacks related to traditional hydropower plants particularly large hydro
plants. Therefore, the Hydro plant presented in this paper is a new technique where
running river water is employed to push Savonius rotor that is usually developed for
wind energy conversion systems (WECS). The Savonius rotor is directly connected
to generators and working is the same as WECS except the driving force is running
river water. Here in this hydro plant, there is no damming and thus does not cause
a large area to be submerged. In addition to producing Hydro energy, this plant will
fulfil the regular energy requirement of the rural area. Savonius hydrokinetic rotors
can continuously yield power 24 h per day. Additional benefit is that water density
is 835 times higher than air, hence if a swept area is equal energy generated by
water is 835 times higher than wind [1]. Over 80% of the hydro capability of our
nation (India) is circulated among the Himalayan state and the main rivers of the
northeastern part of India continue flow throughout the whole year. So this scheme
is expected to work well especially in North East India. The main advantage of this
scheme is that it uses water as it comes so it is not harmful to nearby areas and is
environment friendly.
O. Doso (B) · S. Gao

266 O. Doso and S. Gao
2 Savonius Rotor
Savonius rotor is an important tool for energy conversion in wind and hydropower
systems. S. J. Savonius invented savonius rotor based on Flettner’s rotor model.
Savonius rotor is suitable in a condition where wind or water speed is low. Savonius
rotors have the advantage of redirecting approaching debris instead of infusing it and
thus, have reduced upkeep cost and do not require costly hydrofoil-shaped blades
[2] and have higher stating torque [3]. Drag force is the vital thrust in Savonius rotor
[4]. Savonius rotor produces higher torque even at low speeds and it is self-starting.
The performance of Savonius rotor at the beginning is caused by the drag force after
that rotation is caused by lift force. The aspect ratio is obtained by dividing the rotor
height by its diameter. When used as a hydrokinetic rotor, aspect ratio should be
between 1.5 and 2.0.
Output power of rotor is given by
P = 0.5CP AS ρV 3 (1)
Here P = power, C P = power coefficient, AS = swept area, ρ = air density, and

V = velocity.
Rotor power output is greatly influenced by swept area AS and is given by
AS = H × D (2)
Here, H = height and D = diameter.

Rotational speed (n) is given by
60 ∗ 2π
n= (rpm) (3)
ω
Tip speed ratio (λ) is
ωD
λ= (4)
2V
Here, ω = angular velocity, D = diameter, and V = velocity.
The torque coefficient (C T ) is given by
T T
CT = = 1 (5)
Tw 2
ρ As V 2
where T = rotor torque and T w = torque available in wind.

Power Quality Improvement of Cascaded Savonius Rotor … 267
3 Static Synchronous Compensator(STATCOM)
STATCOM is a shunt device that basically works on the control of the voltage. It
basically consists of a voltage source converter whose input is provided by a DC
voltage source (Capacitor). V 0 is the generated voltage by the STATCOM and V is
system voltage. By varying the output of V 0 reactive power can be controlled. If V0 is
greater than V current will flow from STATCOM to the system means the converter is
generating the reactive power and vice versa. Under steady-state condition, no active
power involve (if V is in phase with V 0 ). However, active power can also control
by a slight phase shift between the V and V 0 . If V 0 lags V in that case STATCOM
will absorb the active power and vice versa. In this way, STATCOM absorbs both
active and reactive power. The modelled STATCOM is a three-legged IGBT based
voltage source converter(VSC) with a DC link capacitor and filter. The STATCOM
performance is fully dependent on its control scheme. The capacitance-voltage vari-
ation is because of power consumption by the voltage source converter(VSC) and
filter. The change in DC capacitor voltage is examined, which indicates active power
unbalance and direct axis component of current I d . After that system peak voltage
is compared with a reference value which indicates reactive power needed by the
system and quadrature component of reference current, Iq [5]. The reference current
I d and I q are changed to three-phase with Inverse Park’s transformations (Fig. 1).
Inverse Park’s transformation (dq0-abc) given by
⎡ ⎤ ⎡ ⎤
cos(wt) cos wt − 2π
3
cos wt − 2π
3 Ia
⎢ 2π ⎥
Idq0 = T Iabc ⎣ sin(wt)
√
2π ⎣
sin wt√ − 3 sin wt√ − 3 ⎦ × Ib ⎦ (6)
2 2 2 Ic
2 2 2
Fig. 1 STATCOM iL
V
Rf if Non-linear
load
Lf
V0
Cdc
Static Compensator
(STATCOM)
Gating signals to the VSC is generated by comparing the reference source currents
with that of measured currents in an HCC. The current controller output determines
the switching pattern of the IGBTs of the VSC.
4 Proposed System and Simulation Result
The cascaded Savonius rotor based hydrokinetic power system is modelled and simu-
lated in MATLAB software. The simulation is implemented using MATLAB 2016Rb
(9.1.0.441655) in a discrete form with 5e-6 step size for 2.8 s. Figure 2 shows the
schematic diagram of cascaded Savonius hydro plant where only 3 Savonius gener-
ator sets are presented. The modelled Cascaded Savonius plant consist of 6 unit and
Each unit consist of 5 generators. Each unit can generate nearly 15 kW so from 6
unit nearly 90 kW can be generated. If the number of generation unit is increased it
can generate more power up to large megawatt. The power quality issues have been
analyzed with STATCOM and the result of the simulation is presented through the
figure (Fig. 3).
Figure 4 shows the voltage with harmonic spectra and the current waveform of
the system.
When a load of 60 kw +20 kVAR is loaded between 1.5 and 2.7 s voltage waveform
is reduced up to 420 V as shown in Fig. 5
And with the use of STATCOM from 1 to 2.7 s. the voltage waveform is improved
as given in Fig. 6.
The total harmonic distortion (THD) is used to quantify the presence of harmonics
in the system. THD of the system voltage with load and STATCOM is within IEEE
519-1992 [6].
TO GRID
G2 G3 G4
PLATFORM/BRIDGE
RIVER BANK
RIVER BANK
T3 T4
T2
W
O
FL
R
TE
A
W
TRASH RACKS
Fig. 2 Schematic diagram of Cascaded Savonius hydro plant

Va Ila
A
Vb Ilb B Linear and
Vc Non- Linear
ILc C
Load
G1 G2 G3
Rf
Lf
Filter
Cacaded Savonius Vdc

Hydro system
VSI
STATCOM
Source voltage Va Vb Vc
Reference
current Hysteresis Gate
Load current Ila Ilb ILc generation Current Pulses for
Controller VSI
dc link voltage Vdc
Fig. 3 Model of Cascaded Savonius hydro plant with load and STATCOM
5 Conclusion
The performance characteristics of the cascaded Savonius hydro system is found to

be satisfactory and the result is presented in Table 1. Under no-load condition, the
system voltage is very well maintained and when load is added there is some voltage
drop but with the addition of STATCOM, the voltage waveform is improved. This
Cascaded Savonius hydro system is feasible in powering rural locations as well urban
locations.
Fig. 4 Voltage waveform, current waveform and harmonic spectra of voltage of Cascaded Savonius
hydro system
Fig. 5 Voltage, current waveform and harmonic spectra of voltage of Cascaded Savonius hydro
system with load
Fig. 6 Voltage, current waveform and harmonic spectra of voltage of Cascaded Savonius hydro
system with load and STATCOM
Table 1 Performance result

S. No. Frequency Voltage (V) % THD of
of the three models
(Hz) voltage
Without load 50 560 0.13
and STATCOM
With load 50 420 2.68
With load and 50 525 0.96
STATCOM
References
1. Doso O, Gao S (2019) Application of Savonius rotor for hydrokinetic power generation. ASME
J Energy Resour Technol 142(1):014501 (2020). https://doi.org/10.1115/1.4044555
2. Zhou T, Rempfer D (2013) Numerical study of detailed flow field and performance of Savonius
wind turbines. Renew Energy 51(3):373–381
3. Pudur R, Gao S (2016) Savonius rotor based grid connected hydrokinetic power generation
scheme. Sustain Energy Grids Netw 5:148–155
4. Roy S, Saha UK (2013) Review of experimental investigations into the design, performance and
optimization of the Savonius rotor. J Power Energy 227(4):528–542 (2013). (Proceedings of the
institution of mechanical engineers, part A)
5. Mosobi RW, Chichi T, Gao S (2015) Power quality analysis of hybrid renewable energy system.
Cogent Eng 2(1):1005000
6. IEEE Std. 519-1992 (1993) IEEE recommended practices and requirements for harmonic control
in electric power systems. © Institute of Electrical and Electronics Engineers, Inc
Design and Simulation of a Sub-6 GHz
Low Loss Band Pass Filter Using Double
Split Inductor for 5G Radio WLAN
Applications
Venkata Raghunadh Machavaram and Bheema Rao Nistala
Abstract In this paper, a novel double split (2_split) inductor and a square spiral
inductor are developed to design a low loss bandpass filter (BPF). The 2_split inductor
has a quality factor of 23.858 at 5.1 GHz with a self-resonating frequency of 13.5 GHz.
It shows a 32% improvement in quality factor and 17% improvement in inductance
against the square spiral inductor. A planar square spiral capacitor is also designed,
which has a capacitance of 0.206 pF and Q value of 27.42981 at 5.1 GHz. The
double split inductor BPF had 1300 MHz bandwidth while square inductor BPF had
1900 MHz at 5.1 GHz. Both filters yielded a very good return loss of −25.644 dB
and −17.943 dB, respectively. They also yielded an excellent insertion loss of
−0.432 dB and −0.641 dB. The BPF designed using the 2_split inductor has 32%
less insertion loss and 7.7 dB (30%) less return loss, when compared to the BPF with
square spiral inductor. Both filters occupied lowest on-chip area of only 0.16 mm2 .
Simulation results prove that these BPFs possess excellent loss characteristics that
suit the sub-6 GHz 5G radio WLAN applications.
Keywords Double split inductor · Quality factor · HFSS · Return loss · Insertion
loss · On-chip area
1 Introduction
Implementation of low-cost and high-performance radio-frequency integrated

circuits (RFICs) that suit the emerging 5G ultra-speed smart wireless devices and
services is a big challenge for the RFIC designers [1]. Currently under development
are miniaturized 5G radio BPF designs on silicon substrate CMOS ICs based on
integrated passive device (IPD) technology. They enable small footprints, lowest
V. R. Machavaram (B) · B. R. Nistala

Warangal, Warangal, India
B. R. Nistala
276 V. R. Machavaram and B. R. Nistala
power, flexible bandwidths, and easy integration. The spiral inductor and capacitor
are most sought silicon passive components, which consume major areas of RFICs
(60–70%). Low insertion loss, high-quality factor (Q) along with minimal on-chip
area, is a challenging task for passive BPF designs [2]. Power handling, selectivity is
traded with center frequency and on-chip area [3]. Good performing low-cost minia-
ture passive devices are developed using IPD technology [4]. Compact passive LTCC
BPFs are developed at high frequencies but suffer from heat and size [5]. IPD BPF
and LNA circuits are now realized as co-located CMOS chips as a system in package
(SIP). Such RFIC circuits have the least chip area with superior RF performance
[6]. Therefore, passive 5G radio BPFs using Si IPD CMOS technology are being
researched heavily to meet stringent 5G spectral demands.
A microstrip BPF demonstrated smaller insertion loss of 0.8 dB at 3.6 GHz but
occupied large 99 mm2 area [7]. Tapered inductors showed enhanced Qmax from
22.8 to 26.7 with reverse excitation. But the Qmax frequency increased to 8 GHz
and resonant frequency to 14.6 GHz [8]. A 2.31 GHz BPF with inter-wined spiral
inductor and capacitor had a good return loss of 26.1 dB but occupied 0.63 mm2
area [9]. A quad BPF reported 22–33 dB return loss and 0.2–1.2 dB insertion loss
in 1.8–4.2 GHz. It had a very large area of 545 mm2 [10]. Another compact 3.3–
3.8 GHz band 5G nRN78 BPF using Si IPD had produced smaller insertion loss of
<1.8 dB but occupied a large area of 1.28 mm2 [11]. An LC silicon IPD BPF resulted
in 3–4 dB insertion loss but with 0.75 mm2 area [12]. A GaAS substrate IPD BPF
produced smaller 0.26 dB insertion loss and good return loss of 25.6 dB at 1.35 GHz
but had a large chip area of 1.25 mm2 [13].
We report here, the design and simulation of a spiral double split inductor and
a spiral capacitor connected in cascade structure to form an LC resonator BPF. We
found its S parameters in HFSS by concentrating the design for 5 GHz, as this sub
6 GHz radio band is being heavily explored for 5GnR N78 and N79 radio reception.
This paper comprises the Sects. 2, 3, and 4.
2 Design and Analysis of BPF
The inductance L, capacitance C, quality factor Q, and self-resonant frequency (SRF)

depend on device geometrical parameters and technology process parameters: spiral
diameter, type, and thickness of the substrate, conductor width, spacing, number of
turns, ring shapes, metal ring thicknesses, etc. These are traded off [14]. Selection
of thick substrate and thin oxide layers, decreasing permittivity of metal, not only
reduces component losses but also enhances the capacitance and inductor Q value
along with filter selectivity [15]. Analytical optimization techniques are performed by
varying the geometrical parameters and technology process parameters to accomplish
high performance [16].
A compact on-chip BPF is designed and simulated with a novel double split spiral
inductor and spiral capacitor connected in a series configuration, in a single layer
on the Si substrate. They are simulated in HFSS and the filter analysis is performed
Design and Simulation of a Sub-6 GHz Low Loss Band Pass … 277
with an equivalent lumped model. S parameters are employed as they are good at
high-frequency analysis of RF circuits.
2.1 On-chip Double Split Spiral Inductor
The inductor with more number of turns gives the best quality factor and inductance.
In the double split inductors, the conductor width is reduced and the number of
turns are increased with in the same specified area. Reduction in the conductor
width increases inductance value, so as to enhance the effective inductance. There
is also an increase in number of turns due to which again the effective inductance is
increased due to the mutual coupling between the tracks. Series stacked multi-turn
spiral inductor structures shown in Fig. 1 possesses high inductance and Q around
10.
L total = L 1 + L 2 + L 3 + M12 + M13 + M23 (1)
Peak Magnetic Energy − Peak Electric Energy

Q = 2π (2)
Energy Loss
The proposed double split inductor has a novel double split structure (conductor
path is split into two parallel tracks). The physical dimensions of the double split
inductor are: width of each track −4.5 µm, the spacing between the tracks −1 µm
and spacing between the turns −2 µm. The width of the conductor strip is 10 µm.
Fig. 1 Typical on-chip

passive spiral inductor
structure- planar view
Fig. 2 On chip spiral inductors. a Square inductor. b Double split inductor (2-split)
We selected copper conductor on a thick Si substrate to reduce its parasitic resistance

to enhance Q value. This double split inductor occupies an area of 180 × 180 µm
square on the chip footprint.
We also developed a square spiral inductor with similar geometrical parameters
for performance comparison. The physical dimensions of the double split inductor
are: width of each conductor track −5 µm and spacing between the turns −2 µm.
This square spiral inductor occupies an area of 180 × 180 µm square on the chip
footprint. The structures of both the inductors are shown in Fig. 2.
If mutual inductance increases then effective inductance also increases and can
make it suitable for high-frequency applications. We have utilized the fact that mutual
inductance depends on the space between the tracks. With the decrease in the space
between the tracks the mutual inductance increases, thereby enhancing overall induc-
tance. The same conductor width is chosen for all the 3 turns. Spacing between layers
is selected so as to minimize the negative mutual inductance. The Q and L values for
these inductors are found from the Y parameters (found from simulated S parameters)
employing Eq. (2) [3].
Im[Y11 ] −1
Q= and L = (2)
Re[Y11 ] 2π f {Im[Y11 }]
Inductance and Quality factor variation against frequency for the simulated
inductor structure are given in Figs. 3 and 4.
The value of inductance is 6.75 nH for the 2-Split double split inductor and 5.6 nH
for the square inductor at 5.1 GHz. The self-resonant frequency is 13.5 GHz. The
double split spiral inductor has an inductance of 1.21 times that of the square inductor.
Both inductors have the same chip area of 180 × 180 µm2 .
The quality factor is 16.177 for a square spiral inductor and 23.858 for 2-split
double split inductor. From simulation results, it is observed that the double split
inductor has a quality factor of 1.48 times that of square spiral inductor. The double
split inductor has an inductance of 1.21 times that of square spiral inductor.
Fig. 3 Variation of inductance (nH) with frequency for on-chip square inductor and double split
(2-split) inductor
Fig. 4 Variation of a quality factor with frequency for on-chip square inductor and double split
(2-split) inductor
2.2 On-Chip Spiral Capacitor
Proposed and simulated single layer square spiral capacitor is designed in HFSS. All
turns are placed in the same layer, with the same conductor width. The capacitance
value majorly depends upon material used, conductor dimensions in the annular
geometry of spiral. The geometry of the spiral capacitor is given in Fig. 5.
The dimensions of planar on-chip capacitor are conductor width −3 µm, thickness
−2 µm, the spacing between turns −2 µm. The occupied area of the on-chip capacitor
Fig. 5 On-chip square spiral capacitor

Fig. 6 Variation of the capacitance (pF) and the quality factor with frequency for on-chip spiral
capacitor
is 180 × 180 µm square. Capacitance and the quality factor Q for the capacitor are
extracted from the S- parameters’ as given in Eq. (3) [3].
−Imaginary[Y21 ] Imaginary [Y11 ]

C= and Q = (3)
ω Real [Y11 ]
Change in the values of the capacitance and the quality factor w.r.t. the frequency
for this proposed capacitor is shown in Fig. 6.
The simulated value of capacitance is 0.206 pF at 5.1 GHz. The capacitor HFSS
simulation showed higher Q value of 27.42981 at 5.1 GHz and 10.3 GHz resonant
frequency (SRF). Simulated response of this capacitor on 0.18 µm CMOS technology
showed a significant enhancement of the quality Q and capacitances. Therefore, this
capacitor is sure to satisfy the 5G radio lower band applications.
2.3 On-Chip Passive Band Pass Filter
We employed a series LC resonator model to simulate the BPF. This permits easier
analysis and implementation to prove the filter performance. The simulation of BPF
replaces the circuit elements L and C with the designed two split inductor and a spiral
capacitor. The equivalent circuit for LC filter configuration. The filter simulation and
optimization were carried out in HFSS tool. We found stable simulated values for the
inductance and capacitance during the entire passband. Given below is the depiction
of the simulated series LC bandpass filter using both the square and double split
inductors with the same capacitor as shown in Fig. 7.
Fig. 7 Planar views of the series LC BPF employing both the on-chip square inductor and double
split (2-split) inductor
2.4 Band Pass Filter Simulation
The first order series LC passive BPF is designed using a lumped LC model and
simulated in HFSS for its S parameters. The obtained simulation results and theo-
retical values matched well for the proposed BPF. The insertion and return loss
characteristics of the proposed BPF are shown in Fig. 8.
Both the double split and square spiral inductors and the spiral capacitor have their
maximum quality factor value of 23.858, 16.177, and 495, respectively. It is observed
that the proposed double split inductor showed an improvement of 32% in quality
factor and 17% in inductance value. HFSS simulation of the 5.1 GHz BPF with 2- split
inductor had shown 1400 MHz bandwidth from 4.5 to 5.9 GHz. The 5.1 GHz BPF
using square inductor had shown 1900 MHz bandwidth from 4.1 to 6.0 GHz. The
two filters yielded a very good return loss of −17.943 and −25.644 dB, respectively
as per Fig. 8. They also yielded an excellent insertion loss of −0.432 dB and −
0.641 dB, respectively at 5.1 GHz center frequency. Both the filters have lowest
occupied on-chip area of only 0.16 mm2 .
These simulation results of loss performance have clearly proven the two filters
would definitely satisfy the performance requirements of a pre-6 GHz 5G bandpass
filter. These simulation results demonstrate superior passband and stopband perfor-
mance for the designed bandpass filter that yields good quality output signals suitable
for a 5G radio (Table 1).
Fig. 8 Insertion and return loss variations of designed series LC passive BPF
Table 1 Summarized BPF parameters

5.1 GHz BPF Design specifications 2-split L BPF Square L BPF
simulation results simulation results
Center frequency f 0 —GHz 4–5 5.1 5.1
Bandwidth f —MHz 1500 1400 1900
Fractional bandwidth—% 20–25 27 37
Quality factor Q 5–10 3.65 2.68
Return loss S 11 dB −20 to −30 −25.644 −17.943
Insertion loss S 12 dB <−1.0 −0.432 −0.641
On-chip area sq.mm <1 0.16 0.16
3.1 Performance Comparison
The performance of our filter is compared with that of the corresponding filters
in references reported, so as to prove its effectiveness to suit the requirements of
5G WLAN communications. The minimum and maximum fractional bandwidths
reported in references [1–13] are 15 and 38%. Our filter had 27% fractional band-
width. Minimum and maximum insertion loss reported in references [1–13] are −
1.3 and −5.6 dB. Our simulated BPF yielded an insertion loss of −0.432 dB for
double split inductor BPF. The minimum and maximum return loss reported in refer-
ences [1–13] are −25 and −35 dB. Our BPF simulations produced a return loss
of −25.644 dB for double split inductor BPF and −17.943 dB for square inductor
BPF. The minimum and maximum on-chip filter occupied areas reported in refer-
ences [1–13] are 0.852 and 4 mm2 . Our simulated BPF occupies only an area of
0.16 mm2 . These comparisons have proved that this proposed BPF employing double
split inductor and a square spiral capacitor will surely meet the stringent spectral
requirements of a sub-6 GHz 5G radio WLAN applications.
4 Summary and Conclusion
In this paper, a miniature IPD BPF is modeled and developed for the pre-6 GHz 5G
radio access applications. A double split inductor has been proposed and compared
with the square spiral inductor. The proposed inductor resulted in an improvement of
32% in quality factor and 17% in inductance value, compared to that of a square spiral
inductor. So, the proposed inductor is useful for high-frequency applications. The
filter is successfully simulated and its analysis is done in HFSS. The BPF designed
using the 2_split inductor has 32% less insertion loss and 7.7 dB (30%) less return
loss, when compared to the BPF with a square spiral inductor. Also, it had smaller
bandwidth of 1400 MHz (26% less) against 1900 MHz bandwidth possessed by
the BPF with a square spiral inductor. Due to several such key enhancements, this
5.1 GHz BPF will surely make the realization of RFICs for the pre-6 GHz 5G radio
wireless access WLAN applications.
References
1. Lee CY, Chen TS, Deng JDS, Kao CH (2005) A simple systematic spiral inductor design
with perfected Q improvement for CMOS RFIC application. IEEE Trans Microw Theory Tech
53(2):523–528
2. Lin CC, Huang CZ, Chen CY (2011) Compact and highly selective millimeter-wave mean-
dered bandpass filter using 0.18-µm CMOS technology. In: Proceedings of Asia-Pacific
microwave conference 2011, pp 49–52. ISBN: tel:978-0-85825-974-4">978-0-85825-974-4
© 2011 Engineers Australia
3. Nagesh Deevi BVNSM, Bheema Rao N (2016) Miniature on-chip band pass filter for RF
applications. Microsyst Technol 23(3):633–638. https://doi.org/10.1007/s00542-016-3052-7
4. Xinhai B et al (2014) Fabrication and measurement of BPF using IPD technology. In: 2014
15th international conference on electronic packaging technology, pp 36–38. ISBN: 978-1-
4799-4707-2/14/© 2014 IEEE
5. Lin YC, Huang HH, Horng TS (2014) Synthesis of LTCC multi-band bandpass filter using
reflection and transmission zeros. In: 2014 IEEE MTT-S international microwave symposium
(IMS2014). ISBN: 978-1-4799-3869-8/14/ ® 2014 IEEE
6. Wang S, Cho K-F (2016) CMOS, IPD switchable bandpass circuit for 28, 39 GHz fifth-
generation applications. IET Microwaves Antennas Propag 10(14):1461–1466. https://doi.org/
10.1049/iet-map.2015.0806
7. Al-Yasir YI et al (2019) Design, simulation and implementation of very compact dual-band
microstrip bandpass filter for 4G and 5G applications. In: 2019 16th international conference
on synthesis, modeling, analysis and simulation methods and applications to circuit design
(SMACD), pp 41–44, ISBN: 978-1-7281-1201-5/19/ © 2019 IEEE
8. Vanukuru VN, Chakravorty A (2014) High-Q characteristics of variable width inductors with
reverse excitation. IEEE Trans Electron Devices 61(9):3350–3354
9. Chuluunbaatar Z, Adhikari KK, Wang C, Kim NY (2014) Micro-fabricated bandpass filter
using intertwined spiral inductor and interdigital capacitor. Electron Lett 50(18):1296–1297
10. Chen YW, Tai TC, Wu HW, Su YK, Wang YH (2017) Design of compact multilayered quad-
band bandpass filter, pp 1815–1818. ISBN: 978-1-5090-6360-4/17 ©2017 IEEE
11. Shin KR, Eilert K (2018) Compact low cost 5G NR n78 band pass filter with silicon IPD
technology. ISBN: 978-1-5386-1267-5/18 © 2018 IEEE
12. Mao C, Zhu Y, Li ZA (2018) Design of LC bandpass filters based on silicon based IPD
technology. In: 19th international conference on electronic packaging technology, pp 238–240.
ISBN: 978-1-5386-6386-8/18/ © 2018 IEEE
13. Quan CH, Wang ZJ, Lee JC, Kim ES, Kim NY (2019) A highly selective and compact bandpass
filter with circular spiral inductor and an embedded capacitor structure using an integrated
passive device technology on a GaAs substrate. Electronics 8:73. (MDPI, pp. 1–8). https://doi.
org/10.3390/electronics-8010073
14. Wang C, Kim N-Y (2012) Analytical optimization of high performance and high-yield spiral
inductor in integrated passive device technology. Microelectron J 43:176–181
15. Büyüktas K, Koller K, Müller KH, Geiselbrechtinger A (2010) A new process for on-chip
inductors with high Q factor performance. Int J Microwave Sci Technol
16. Nagesh Deevi BVNSM, Bheema Rao N (2016) Analysis of miniature on-chip 3-D inductor
for RF circuits. ARPN J Eng Appl Sci 11(5). ISSN: 1819-6608
Two Novel Configurations
of Electronically Tunable Quadrature
Sinewave Oscillator Using CDBA
Shekhar Suman Borah, Ankur Singh, and Mourina Ghosh
Abstract This paper presents two novel voltage mode quadrature sinewave oscil-
lator configurations employing Current Differencing Buffered Amplifier (CDBA) as
active elements and a few passive components. Tuning of frequency and condition of
oscillation can be performed independently. The proposed CDBA-based quadrature
sinewave oscillator circuit offers two voltage outputs with a 90° phase difference from
high output impedance terminals and the performance of the proposed circuit is also
valued by well-known Monte-Carlo simulations. The sensitivity analysis also carried
out. The simulated and hardware verification is performed using PSPICE 0.5 µm
CMOS process parameters and Current Feedback Operational Amplifier (CFOA)
based IC AD844AN, respectively, to produce periodical sinewave oscillations.
Keywords Analog-integrated circuit design · Current differencing buffered

amplifier (CDBA) · Quadrature sinewave oscillators
1 Introduction
The popularity attained by the Quadrature Sinewave Oscillators in various disciplines

of electronic circuits such as signal processing and measurements, control systems,
and to produces single sideband (SSB) as well as quadrature mixtures in the field
of communication, etc. Based on the operational amplifier (Op-Amp) as an active
element, many oscillator circuits have existed in the previously reported papers. But,
the disadvantages of these circuits are not able to perform at higher frequencies
S. S. Borah (B) · A. Singh · M. Ghosh

Department of Electronics and Communication Engineering, Indian Institute of Information
Technology Guwahati, Guwahati, India
A. Singh
M. Ghosh
286 S. S. Borah et al.
because of limited gain-bandwidth product and slew rate [1–3]. This restricted gain-
bandwidth product (GBWP) affects the oscillation frequency and its condition. To
overcome these shortcomings, different oscillator circuits have been reported in the
literature [4–7]. But, in those, the requirement of passive elements or active elements
is more to generate oscillations. CDBA, a unique active ABB has attained much
attention from the analog circuit designers [8]. It has the capability of working in
both the current mode (CM) and voltage mode (VM) operations has a broad dynamic
range and gives considerably wider bandwidth compared to other Analog Building
Blocks (ABBs) [9, 10]. These advantageous characteristics of CDBA encouraged a
far-reaching use of this ABB for the implementation of converters, oscillators, and
filters as well as in the area of measurement and control system both adequately and
efficiently.
In this work, two new configurations of CDBA-based sinewave oscillators are
presented. The proposed oscillator circuits require two CDBAs and six passive
components to produce the oscillations. Also, the simulations were performed
using PSPICE 0.5 µm CMOS technology. The analysis of sensitivity determines
that the proposed circuits have low-sensitivity. The performance of the proposed
circuit is evaluated through Monte-Carlo simulations. The circuit is also verified
experimentally using commercially available IC AD844AN.
2 Proposed Circuit Description
The electrical notation of CDBA [1] is shown in Fig. 1 and the relationship between
current and voltage is represented by the following equation as:
V p = Vn = 0, Iz = I p − In , Vw = Vz (1)
where I p and I n are the input currents, V p and V n are the input voltages at the
terminals p and n, as well as z and w, are the current and the voltage output terminals,
respectively. Terminal-w follows the voltage of terminal z. Current Differencing
Buffered Amplifier (CDBA) is a current-controlled voltage source device and can be
realized with CMOS technology. CDBA CMOS structure and 0.5 µm parameters are
taken from ref. [11]. The analysis can be performed as a set of a non-inverting and
an inverting current mode (CM) and a non-inverting voltage mode (VM) unity-gain
cells. The two proposed quadrature sinewave oscillator circuits are shown in Fig. 2.
Fig. 1 Electrical notation of

CDBA
Two Novel Configurations of Electronically Tunable … 287
Fig. 2 Proposed sinewave

oscillator circuits.
a Configuration 1.
b Configuration 2
The straightforward circuit analysis of the proposed configuration 1 as shown in

Fig. 2a gives the characteristics equation as
s 2 Ca1 Ca2 Ca3 Ra1 Ra2 + s(Ca3 Ca2 Ra2 + Ca1 Ca2 Ra1 − Ca1 Ca3 Ra3 ) + Ca2 = 0 (2)
Equation (2) implies the oscillation condition (OC) and oscillation frequency (OF)
as
OC : Ca3 Ca2 Ra2 + Ca1 Ca2 Ra1 = Ca3 Ca1 Ra3 (3)

1 1
OF : fo = (4)
2π Ca1 Ca3 Ra2 Ra1
Similarly, as shown in Fig. 2b, for the proposed configuration 2, the characteristics
equation is,
s 2 Ca1 Ca3 Ra1 Ra2 Ra3 + s(Ca3 Ra3 Ra2 − Ca2 Ra1 Ra2 ) + Ra1 = 0 (5)
Therefore, from Eq. (5), oscillation condition (OC) and oscillation frequency (OF)
is found to be
OC : Ca3 Ra2 Ra3 = Ca2 Ra1 Ra2 (6)

1 1
OF : fo = (7)
2π Ca1 Ca3 Ra2 Ra3
An essential performance measure of an integrated circuit is the study of sensi-

tivity. Considering Eqs. (2) and (5), the sensitivities of OF (ω0 ) to the variation of
passive components such as resistances and capacitances are obtained as
1
SCωa1o = SCωa3o = S Rωa1o = S Rωa2o = S Rωa3o = − (8)
2
From Eq. (8), it can be seen that the proposed configurations exhibit low-sensitivity
performance and it is acknowledged as a good characteristic.
3 Non-ideality Analysis
For a complete review of the proposed circuit, it is pretty much necessary to analyze
the non-idealities of CDBA [5]. Therefore, considering non-idealities, the voltage-
current relationship given in (1) can be modified as
V p = Vn = 0, Iz = Y p I p − Yn In , Vw = λVz (9)

where Y p = 1 − β p and β p (β p 1) signifies the current- tracking error from p to
z terminal, Yn = 1 − βn and βn (|βn | 1) is the current-tracking error from terminal
n to terminal z, λ = 1 − βv and βv (|βv | 1) is the voltage-tracking error from
terminal z to terminal w. So, the revised transfer functions of the stated configuration
1 after examining the factors due to the non-idealities of CDBA presented in Eq. (2)
can be described as
s 2 Ca1 Ca2 Ca3 Ra1 Ra2 + s(Ca3 Ca2 Ra2 + Ca1 Ca2 Ra1
− Yn1 Yn2 λ1 λ2 Ca1 Ca3 Ra3 ) + Ca2 = 0 (10)
Therefore, modified
OC : Ca3 Ca2 Ra2 + Ca1 Ca2 Ra1 = Yn1 Yn2 λ1 λ2 Ca1 Ca3 Ra3 (11)

1 1
OF : fo = (12)
2π Ca1 Ca3 Ra2 Ra1
Similarly, the modified transfer functions of the proposed configuration 2 given

in Eq. (5) can be described as
s 2 Yn1 Yn2 λ1 λ2 Ca1 Ca3 Ra1 Ra2 Ra3

+ s(Yn1 Yn2 λ1 λ2 Ca3 Ra3 Ra2 − Ca2 Ra1 Ra2 ) + Ra1 = 0 (13)
Therefore,
OC : Yn1 Yn2 λ1 λ2 Ca3 Ra3 Ra2 = Ca2 Ra1 Ra2 (14)

1 1
OF : fo = (15)
2π Yn1 Yn2 λ1 λ2 Ca1 Ca3 Ra2 Ra3
Also, while considering non-idealities the sensitivities of oscillator frequency ω0

to the variation of passive components are
1
SCωa1o = SCωa3o = S Rωa1o = S Rωa2o = S Rωa3o = − , SYωn1o ,Yn2, Yλ1 ,Yλ2 = Sλω1o,λ2 = 0 (16)
2
The above equation implies that the proposed sinewave oscillator configurations
exhibit a low-sensitivity performance and the absolute measures ω0 with respect to
Y pi , Y ni and λi are within unity.
4 Simulated and Experimental Results
To validate the theoretical interpretation, the proposed circuits in Fig. 2a, b is simu-
lated with PSPICE 0.5 µm CMOS parameters. The supply voltage V DD = V SS =
±1.2 V and the bias currents, I B1 = 20 µA, I B2 = 20 µA and I B3 = 25 µA. In Fig. 3,
the simulated transient response of the proposed oscillators during the initial state is
shown. Simulated output waveforms (V 01 and V 02 ) for configuration 1 and config-
uration 2 is displayed in Fig. 4. The simulated frequency of oscillation for both the
circuits is 91.3 kHz and 16.4 kHz which closely agrees to the theoretical frequency
of 91.88 kHz and 16.776 kHz respectively. The passive components values for both
the circuits are Ra1 = Ra2 = 1 k, Ra3 = 30 k, C a1 = C a2 = 1 nF, and C a3 = 3 nF.
The Fourier response of the output voltage, which verifies that the oscillations
present at the desired frequency, are depicted in Fig. 5. The plot for voltage V 01
versus V 02 for both the configurations is shown in Fig. 6. The Monte-Carlo analysis,
by considering 100 samples for 5% tolerance which provides the robustness in passive
components is shown in Fig. 7.
The practical workability of this work is also verified using a macro model of
commercially available IC of CFOA (AD844AN) [1] with supply voltage ± 5 V and
the realization of the proposed oscillator configurations using AD844AN is shown
Fig. 3 The simulated transient response of the oscillators during initial state
Fig. 4 Simulated quadrature output waveform of a configuration 1 b configuration 2
Fig. 5 Simulated frequency spectrum of a configuration 1 b configuration 2

Fig. 6 Plot of V 01 versus V 02 for a configuration 1 b configuration 2
Fig. 7 Monte-Carlo analysis results for a configuration 1 b Configuration 2
in Fig. 8. The hardware arrangements for the proposed configuration 2 is shown in

Fig. 9.
The passive elements values are, Ra1 = Ra2 = 1 k, Ra3 = 30 k, C a1 = C a2 =
1 nF and C a3 = 3 nF and the oscillation frequency for both the circuit is 89.4 kHz
and 15.2 kHz, respectively, which is close to the simulated results. The quadrature
oscillations output waveform and the steady-state responses of the experimental
circuits are shown in Fig. 10a, b, respectively.
In Table 1, a comparative analysis of the proposed work with the earlier reported
work is shown. From this, it can be observed that, although the proposed work
uses two ABBs which is similar to other reported works, but it uses less amount
of transistors which can decrease power dissipation. Moreover, the use of passive
components reported in ref [10] is more than the proposed topology. Therefore,
these can be considered as useful characteristics of this proposed sinewave oscillator
topology employing CDBA.
Fig. 8 Proposed oscillator circuits using AD844 a configuration 1 b configuration 2
Fig. 9 Hardware
arrangements for the
proposed configurations
5 Conclusion
This paper presents two CDBA-based quadrature sinewave oscillator configura-

tions including low passive sensitivities. The proposed topology is simpler than
the usual waveform generators which provide independent control of the frequency
as well as the condition of oscillations. Here, CDBAs operate as active components
and requires only three resistors and capacitors each as passive components. The
Fig. 10 Experimental results (using AD844): a quadrature output waveform b steady-state response
Table 1 Comparative study

References No. of active No. of passive Transistor No. of transistors
elements components
(CDBA) R C
[8] 2 3 2 CFOA(BJT) 76
[9] 2 4 2 CFA(BJT) 76
[10] 2 4 4 CFOA(BJT) 76
[12] 2 3 3 CFOA(BJT) 76
Proposed 2 3 3 CMOS 40
outcomes presented by the stated topologies are congruent with simulated and corre-
sponding experimental results. Simulations were performed using PSPICE 0.5um
CMOS technology and hardware experimentation was fulfilled using commercially
available IC AD844AN. The grounded resistor is considered to be an additional
feature from the integration point of view. Finally, the well-known Monte-Carlo
analysis is performed to check the performance of the proposed work. Hence, the
circuits can provide new possibilities to analog designers.
References
1. Acar C, Ozoguz S (1999) A new versatile building block: current differencing buffered amplifier
suitable for analog signal-processing filters. Microelectron J 30:157–160
2. Pal D, Srinivasulu A, Pal BB, Demosthenous A, Das BN (2009) Current conveyor-based
square/triangular waveform generators with improved linearity. IEEE Trans Instrum Measur
58:2174–2180
3. Senani R, Bhaskar DR, Singh AK (2014) Current conveyors: variants, applications and
hardware implementations. Springer, New York
4. Sharma RK, Arora TS, Senani R (2016) The realization of canonic single-resistance-controlled
oscillators using third generation current conveyors. IET Circ Devices Syst 11:10–20
5. Sotner R, Jerabek J, Prokop R, Kledrowetz V (2016) Simple CMOS voltage differencing current
conveyor-based electronically tunable quadrature oscillator. Electron Lett 52:1016–1018
6. Keskin AÜ (2005) Voltage-mode notch filters using single CDBA. Frequenz 59(9–10):225–228
7. Gupta SS, Sharma RK, Bhaskar DR, Senani R (2010) Sinusoidal oscillators with explicit
current output employing current-feedback Op-Amps. Int J Circ Theor Appl 38:131–147
8. Prasertsom D, Tangsrirat W (2010) CDBA-based sinusoidal quadrature oscillator with
current controlled amplitude. In: 10th IEEE international symposium on communications and
information technologies, pp 187–191. IEEE, Tokyo, Japan (2010)
9. Tangsrirat W, Pukkalanun T, Surakampontorn W (2008) CDBA-based universal Biquad filter
and quadrature oscillator. Act Passive Electron Compon 2008:1–6
10. Keskin AU, Aydin C, Hancioglu E, Acar C (2006) Quadrature oscillator using current
differencing buffered amplifiers (CDBA). Frequenz 60(3–4):57–59
11. Keskin AÜ, Hancioglu E (2005) Current mode multifunction filter using two CDBAs. AEU
Int J Electron Commun 59(8):495–498
12. Tangsrirat W, Pisitchalermpong S (2007) CDBA-based quadrature sinusoidal oscillator.
Frequenz 61(3–4):102–104
Analog/RF and DC Performance
Enhancement of a Pocket-Doped
Junction-Less TFET for Low Power
Application
Sneha Bharti, Suruchi Sharma, Abhishek Verma, Manisha Bharti,

and Baljit Kaur
Abstract The Junction-less tunnel field-effect transistor (JLTFET) is a captivating

device due to its excellent electrical properties and less variability in comparison
to MOSFET at the nanometer regime. In this regard, we investigate a silicon-
based pocket doped JLTFET in which an InAs pocket is inserted across the source-
channel junction to enhance tunneling probability. In this respect, we have consid-
ered analog/RF and DC Figure of merit analysis for the conventional and pocket
doped JLTFET (PD- JLTFET) in terms of an electric field, transfer characteris-
tics, transconductance, parasitic capacitances, cut-off frequency, gain-bandwidth
product, and maximum oscillation frequency. Additionally, we have examined the
effect of spacer length variation across the junction between source and channel. The
ATLAS device simulator is used for the simulations of the conventional JLTFET and
PD-JLTFET. The proposed PD-JLTFET has shown a higher I ON /I OFF ratio (~1013 )
and improved sub-threshold swing (~9.08 mV/decade). The notable characteristics
demonstrated by PD-JLTFET make it an optimum device for low power and high
switching application.
Keywords High-k dielectric material · InAs · Junction-less TFET · Sub-threshold

swing
S. Bharti (B) · S. Sharma · A. Verma · M. Bharti · B. Kaur

Delhi, Delhi, India
S. Sharma
A. Verma
M. Bharti
B. Kaur
296 S. Bharti et al.
1 Introduction
Development in the field of semiconductors has led to an increase in the peremptory

call for submicron devices for high switching speed and DC/analog applications [1,
2]. However, in deep submicron technology (due to reduction in device dimension)
power dissipation and scalability have become the most critical factor due to an
increase in transistor density on a chip, increasing leakage power and need of high
speed of operation for the proper functioning of the device. For this reason, MOSFET
is mostly used for the design of integrated circuits as it provides a simple fabrication
process and is available at a cheaper cost to fulfill consumer demands. According to
Moore’s law, transistors density on a chip increases due to continued scaling of the
device.
However, conventional MOSFET suffers from short channel effect, lowering of
barrier potential caused by the drain voltage, low ON-OFF current ratio, and also it
is not an energy-efficient device [3, 4]. Along with this, increment in leakage current
occurs from one technology node to another, results in more static and dynamic power
dissipation and also due to tunneling of charge carrier occurs over the barrier [5],
conventional MOSFET provides minimum sub-threshold swing (>60 mV/decade),
which doesn’t provide high I ON /I OFF (ratio of ON-OFF current) required for applica-
tion in low power devices. The reduction in power consumption is possible by scaling
down the supply voltage but the reduction in supply voltage is also not possible after
a particular technology node in the case of MOSFET because it reduces the compe-
tency of the device in carrying current [6]. So as to conquer these limitations, TFET
is used over MOSFET.
The operation of TFET is based on the quantum mechanical tunneling(band to
band) due to which it is immune from the short- channel effect and barrier lowering
due to drain voltage [7, 8]. It also provides low OFF-state current and low sub-
threshold swing [9–11]. The low sub-threshold swing of TFET enables it to reduce
the supply voltage and threshold voltage to a value of less than 0.5 V [6, 12–14].
TFET, also known as “Green Transistors,” can be operated at lower voltage source,
so it is termed as a highly energy-efficient device [15].
While TFET has all these benefits, it suffers from low ON-state current that occurs
due to insufficient quantum tunneling of charge carrier caused by large bandgap mate-
rial present at the interface between source and channel [16]. Another disadvantage
of TFET is its ambipolar behavior (current conduction for positive as well as negative
voltage at the gate terminal) [17, 18]. Apart from all these, as the technology node
reaches below 100 nm, fabrication complexity increases for such types of devices
[19].
Thus, in order to overcome the limitations of fabrication complexity and low ON-
state current with regards to TFET, a new transistor named JLTFET is used. It makes
use of the concept of both JLFET and TFET [20]. It works on the principle of gate
work function engineering, also termed as charge plasma concept [21]. As the name,
JLTFET itself suggests that it does not have any metallurgical junction across drain-
channel and source-channel interface, because of their equal doping concentration
Analog/RF and DC Performance Enhancement of a Pocket-Doped … 297
throughout the region from drain-to-source. JLTFET does not have any junction, so
it will be easier to fabricate [20] and has better performance than TFET.
Considering these advantages of JLTFET, we proposed a device pocket doped
JLTFET (PD-JLTFET), which is composed of two gates, i.e., polarity-gate (P-gate)
and control-gate (C-gate) in order to control the device operation. The C-gate is
placed near the channel in this system, P-gate is placed near the source side, and
its work functions are varied to inhibit ambipolar current. A pocket of InAs, which
has a very low bandgap as compared to Si element, is inserted at the tunneling
junction of the interface between source and channel to improve the current when
the device is on. The proposed PD-JLTFET structure performance is analyzed and
compared with conventional JLTFET [20]. The system suggested has better simula-
tion results in terms of I ON /I OFF , I ON , I OFF , sub-threshold swing (SS), electric field,
and transconductance.
PD-JLTFET design can be constructed using the steps specified by Daley et al.
[22], as something of a single sheet piled up of several sheets of hetero components
with a vertical structural layout.
The rest parts of this manuscript are structured as mentioned further: Sect. 2
elaborates device structure and simulation variables. Section 3 describes the results
and discussion part, and at last, the conclusion is discussed in Sect. 4.
2 Device Structure and Simulation Variables
Figure 1a depicts the cross-sectional view of conventional JLTFET used in reference

[20] and Fig. 1b shows the cross-sectional view of proposed PD-JLTFET with N+
doped structure. The device is shown in Fig. 1b is different from the structure shown
in Fig. 1a by insertion of InAs as a pocket at the source-channel interface to shorten
the bandgap at tunneling junction of the interface between source and channel, so that
proper tunneling of charge carrier can occur at the junction. To obtain the enhanced
device performance, InAs is used because it has low bandgap nearly equal to 0.36 eV
(much less than Si having bandgap 1.1 eV) and also a high melting point. As InAs
is a compound type semiconductor, so energy will be radiated in the form of light
energy. The device parameters used in the simulation results are n-type heavily doped
Si-channel, with a doping concentration of 1 × 1019 cm−3 . In order to separate both
the gates, i.e., C-gate and P-gate a spacer of 5 nm is used in the device structure. Few
device simulation parameters and material constants are given in Table 1.
All the simulation results are carried out using 2D device simulator Silvaco Atlas
5.19.20.R [23]. Fermi-Dirac statistics are included in determining the required carrier
concentration (intrinsic) in Shockley-Read-Hall (SRH) recombination expression
[24]. In TFET current conduction, occur due to band to band tunneling. Therefore,
the nonlocal BTBT (BBT.NONLOCAL) model is applied to determine the tunneling
of charge carriers along the lateral direction for a particular device [25]. This nonlocal
BTBT model is applied to simulate JLTFET [26]. Bandgap narrowing (BGN) model
is taken into consideration due to excessive dopant mixing along the channel region
Fig. 1 Cross-sectional views

of a conventional JLTFET
b proposed PD-JLTFET
Table 1 Devices structure parameters

Parameters Symbol JLTFET PD-JLTFET
Silicon film thickness TSi 5 nm 5 nm
Source/channel/drain doping concentration ND 1 ×1019 1 ×1019
cm−3 cm−3
Channel length LG 20 nm 20 nm
Drain/source length L Ext 20 nm 20 nm
Control-gate work function WC -gate 4.54 eV 4.3 eV
Polarity-gate work function W P -gate 5.93 eV 5.93 eV
Oxide thickness Tox 2 nm 2 nm
Electrode thickness Tetd 5 nm 5 nm
Gate-source spacer L GS 5 nm 5–10 nm
in the device structure [23]. Because of the material having a large value of dielectric
constant present at the gate electrode gate leakage model is exempted from the
consideration. Shockley read hall (SRH) recombination model fits well due to high
impurity atom present throughout the channel [23]. Quantum confinement model
given by Hansch [23, 27] on BTBT is also taken into consideration, as the device
dimension is less than 10 nm. Trap assisted tunneling model (SCHENK.TUN) given
by Schenk is used to give the importance of trap assisted tunneling in the sub-
threshold region [23, 28]. In this paper, the used nonlocal BTBT model uses Wentzel–
Kramer–Brillouin method to calculate electron and hole tunneling probability [25].
The metalwork function for the proposed device is varied from 4.3 to 4.74 eV and
5.5 to 5.93 eV for C-gate and P-gate to create a layer just under the gates as P-
type as well as intrinsic. Proposed device structure gives best outcome in view of
sub-threshold swing and I ON /I OFF ratio at work function of 4.3 and 5.93 eV for
corresponding control-gate and polarity-gate. As the sub-threshold swing is very
less for the proposed device, hence it can be used for low power applications.
3.1 DC Analysis
This section describes the performance of proposed PD-JLTFET with the conven-
tional JLTFET. The comparative analysis of carrier concentration in OFF-state for
both the devices, i.e., PD-JLTFET and JLTFET is shown in Fig. 2a, from the concen-
tration contour, This could be noted that the system looks like the layout N+-I-P+
[20]. In the proposed PD-JLTFET device structure InAs, having band-gap equal to
0.36 eV is used as a pocket at the source-channel interface that means bandgap
will be very less as compared to Si when used at the interface of source-channel
region hence electron concentration will be somehow more at source side in case
InAs is used as pocket as have done in the case of the proposed structure. Figure 2b
shows the comparison of OFF-state energy band diagram for conventional JLTFET
with proposed PD-JLTFET. For both the device as There is a large tunneling barrier
between source and channel, so tunneling is likely an electron from valance band of
the source to conduction band of the channel will be negligible in this case. There-
fore, the current will flow here is only due to diode leakage. As the bandgap of InAs
is very less, hence some spike may occur at the junction between source and channel.
The device operation is controlled by the application of a voltage on C-gate. As the
device will be turned on, then the barrier existing between the channel and the source
region will become narrow, so that proper tunneling of majority carrier can occur
from source to channel. As Fig. 2c reflects the comparison of carrier concentration
in ON-state for both the devices. From the figure shown below, we can note that on
application of voltage on C-gate the concentration of electrons beneath the C-gate
escalates and behaves like an n-type region. This results in a reduction of barrier
Fig. 2 Comparison between JLTFET and PD-JLTFET a Carrier concentration on OFF-state

b Energy band diagram in OFF-sate c Carrier concentration in ON-state d Energy band diagram in
ON-state
between the source-channel junctions of the device due to the insertion of InAs
also concentration of electron in the source region for PD-JLTFET will be more as
compared to JLTFET. Figure 2d shows the comparative analysis of ON-state energy
band diagram for both the devices. As we know that for TFET, as the barrier between
source and channel tunneling region will decrease corresponding ON-current will
increase, which can also be seen from the given figure. As the InAs has much lower
bandgap as compared to Si, hence tunneling barrier in proposed PD- JLTFET will
be less as compare to conventional JLTFET. So proper tunneling of the electron will
occur in the case of PD- JLTFET compared to conventional JLTFET. For turning the
device on the C-gate voltage is varied from 0 to 1 V, while P-gate voltage is kept at
zero volts [20].
High-k dielectric material yields superior results in context to ON-current, SS,
and OFF-current [5]. Interface trap effect is taken into consideration because of the
direct exposure to material having high k-value with Si channel, as it leads to a
kind of defect at the semiconductor-dielectric interface. Here in the structure, TiO2
Fig. 3 Comparison of
transfer characteristics
between conventional
JLTFET and Proposed
PD-JLTFET
is considered as a dielectric material having a dielectric constant 80 [29]. Figure 3

depicts the transfer characteristics of existing JLTFET and intended PD-JLTFET. As
the lower subthreshold swing of TFET reduces the supply and threshold voltage of
a device to a value lower than 0.5 V [6, 12–14], hence as the device has very less
sub-threshold swing equal to 9.08 mV/decade so the proposed device can be used
at an input voltage less than 0.5 V, i.e., at 0.4 V. From the above characteristics it is
clear that the proposed structure has higher ON-current and lower OFF-current as
compared with conventional JLTFET.
Figure 4 shows the comparison graph of the electric field for JLTFET and PD-
JLTFET at ON-state. From the graph, we can say that the proposed device has a better
electric field in the source-channel tunneling junction due to the presence of narrow
bandgap material as compared to conventional JLTFET. Whereas electric field in the
other portion is the same.
Figure 5a displays the input to output characteristics of PD-JLTFET at various
supply voltages (V ds ). Nanometer technology is perfect to use for low power appli-
cations as we can conclude from the graph that the intended structure gives better
results for low supply voltage (0.4 V) than for high voltage (1.0 V). Figure 5b shows
the transfer characteristics of PD-JLTFET at different isolation thickness. As the
spacing between source-channel decreases, it leads to increase in tunneling proba-
bility of carriers, leads to escalates in ON-current and trails in subthreshold swing for
lesser isolation thickness at the interface between source and channel, which can be
seen from the graph that lesser isolation thickness of the provides more ON-current
and as the isolation thickness at the interface increases, ON-current will decrease.
Transconductance (gm ) describes System operation capability. Therefore, the
value of gm must be high enough to achieve high linearity performance and high-
frequency characteristics [1]. In this regard, Fig. 6 shows the comparison plot of gm
versus V gs for both the devices. From the graph, it can be concluded that transcon-
ductance of PD-JLTFET is higher as compared to the JLTFET due to lesser work
Electric field under ON-state
for JLTFET and PD- JLTFET
Fig. 5 Transfer characteristics for PD-JLTFET at a different supply voltage (V ds ) b different

isolation thickness
function value of C-gate and material having less band gap present at the tunneling
junction of the channel-source region, results in large current operating capability of
the device, leading to high transconductance.
transconductance (gm )
between JLTFET and
PD-JLTFET
3.2 Analog/RF Parameter Analysis
In the high-frequency region, parasitic capacitances play a significant role in device

performance. Figure 7a shows the graph between C gd versus V gs for both the devices.
From the graph, it can be observed that C gd increase along V gs due to significant
barrier at drain channel interface created due to C-gate work function. The C gd of
PD-JLTFET is more as compared to JLTFET due to the presence of InAs at tunneling
junction, which causes a relatively more parasitic effect on the devices. The parameter
fT and gain-bandwidth product (GBP) can be calculated using the formula as

f T = gm /2π Cgd + Cgs (1)
and GBP = gm /20πCgd (2)
Figure 7b, c shows the comparative analysis of cut-off frequency and Gain band-
width product graph. Cut-off frequency for JLTFET is 16 GHz and for PD-JLTFET
is 5.5 GHz. Corresponding gain-bandwidth product for JLTFET is 1.6 GHz and for
PD-JLTFET is 0.55 GHz. F max is the other constraints to determine the RF execution
of the device. F max (frequency) needed for power gain to be unity. It is defined as

Fmax = 2
f T /8πCgd Rds (3)
Figure 7d depicts the graph of maximum frequency from which it can be concluded
that F max for JLTFET is 2.49 GHz and for PD-JLTFET F max is 1.36 GHz due to
parasitic effect in case of PD-JLTFET.
Fig. 7 Comparison between JLTFET and PD-JLTFET a gate-to-drain capacitance (Cgd ) b cut-off
frequency (f T ) c GBP d maximum oscillation frequency (F max ) versus Vcgs
4 Conclusion
We have shown an effective technique to enhance the capability of the device in

context to sub-threshold swing (SS), OFF-state current, ON-state current. The device
performance of the intended PD-JLTFET is compared with the existing JLTFET.
The device performance for the various spacing length between control-gate and
polarity-gate is also analyzed. The SS of PD-JLTFET is 9.08 mV/decade, i.e., Very
less as compare to conventional JLTFET, which has SS is equal to 40.2 mV/decade.
The PD-JLTFET shows a significantly higher I ON /I OFF ratio, therefore improving its
switching capability and makes it best suited for the use of low power applications.
Acknowledgements This work is supported by the SMDP-C2SD under the reference letter no.
9(1)2014-MDD (NIT Delhi, INDIA).
References
1. Kilchytska V et al (2003) Influence of device engineering on the analog and RF performances

of SOI MOSFETs. IEEE Trans Electron Devices 50(3):577–588
2. Mohankumar N, Syamal B, Sarkar CK (2010) Influence of channel and gate engineering on the
analog and RF performance of DG MOSFETs. IEEE Trans Electron Devices 57(4):820–826
3. Mallik A, Chattopadhyay A (2012) The impact of fringing field on the device performance of
a p-Channel tunnel field-effect transistor with a high-κ gate dielectric. IEEE Trans Electron
Devices 59(2):277–282
4. Bangsaruntip S, Cohen GM, Majumdar A, Sleight JW (2010) Universality of short-channel
effects in undoped-body silicon nanowire MOSFETs. IEEE Electron Device Lett 31(9):903–
905
5. Boucart K, Ionescu AM (2007) Double-gate tunnel FET with high- κ gate dielectric. IEEE
Trans Electron Devices 54(7):1725–1733
6. Seabaugh AC, Zhang Q (2010) Low-voltage tunnel transistors for beyond CMOS logic. Proc
IEEE 98(12):2095–2110
7. Bhuwalka KK, Schulze J, Eisele I (2005) Scaling the vertical tunnel FET with tunnel bandgap
modulation and gate work function engineering. IEEE Trans Electron Devices 52(5):909–917
8. Nirschl T et al (2006) Scaling properties of the tunneling field effect transistor (TFET): Device
and circuit. Solid State Electron 50(1):44–51
9. Choi WY, Park BG, Lee JD, Liu TJK (2007) Tunneling field- effect transistors (TFETs) with
subthreshold swing (SS) less than 60 mV/dec. IEEE Electron Device Lett 28(8):743–745
10. Wang PF et al (2004) Complementary tunneling transistor for low power application. Solid
State Electron 48(12):2281–2286
11. Zhang Q, Zhao W, Seabaugh A (2006) Low-subthreshold-swing tunnel transistors. IEEE
Electron Device Lett 27(4):297–300
12. Kam H, King-Liu TJ, Alon E, Horowitz M (2008) Circuit-level requirements for MOSFET-
replacement devices. In: 2008 IEEE international electron devices meeting 2008. IEDM, vol
11, no 6, p 1 (2008)
13. Mookerjea S, Datta S (2008) Comparative study of Si, Ge and InAs based steep subthreshold
slope tunnel transistors for 0.25v supply voltage logic applications. In: 2008 device research
conference. Dig. DRC, vol 567, no 2, pp 47–48 (2008)
14. Collet J, Tharaud O, Chapoton A, Vuillaume D (2000) Low- voltage, 30 nm channel length,
organic transistors with a self- assembled monolayer as gate insulating films. Appl Phys Lett
76(14):1941–1943
15. Ionescu AM, Riel H (2011) Tunnel field-effect transistors as energy-efficient electronic
switches. Nature 479(7373):329–337
16. Kao KH, Verhulst AS, Vandenberghe WG, Soree B, Groeseneken G, De Meyer K (2012) Direct
and indirect band-to-band tunneling in germanium-based TFETs. IEEE Trans Electron Devices
59(2):292–301
17. Abdi DB, Kumar MJ (2016) Suppressing ambipolar conduction using dual material gate in
tunnel-FETs having heavily doped drain. Int J Electron Commun Eng 10(5):594–598
18. Hraziia A, Vladimirescu A Amara, Anghel C (2012) An analysis on the ambipolar current in
Si double-gate tunnel FETs. Solid State Electron 70:67–72
19. Tirkey S, Sharma D, Yadav DS, Yadav S (2017) Analysis of a novel metal implant junction
less tunnel FET for better DC and analog/RF electrostatic parameters. IEEE Trans Electron
Devices 64(9):3943–3950
20. Ghosh B, Akram MW (2013) Junctionless tunnel field effect transistor. IEEE Electron Device
Lett 34(5):584–586
21. Saurabh S, Kumar MJ (2011) Novel attributes of a dual material gate nanoscale tunnel field-
effect transistor. IEEE Trans Electron Devices 58(2):404–410
22. Daley DM, Tran HH, Woods WH, Zhang Z (2015) Tunnel field-effect transistors with a gate-
swing broken-gap heterostructure. U.S. Patent 9,087,717 B2, 21 Jul 2015. (Online). Available:
https://www.google.ch/patents/US9087717
23. Silvaco, Atlas user’s manual, Silvaco, Inc., 408: 567–1000 (2016)
24. Raushan MA, Alam N, Siddiqui MJ (2019) Electrostatically doped drain junction less transistor
for low-power applications. J Comput Electron 0123456789
25. Raushan MA, Alam N, Siddiqui MJ (2018) Dopingless tunnel field-effect transistor with over-
sized back gate: proposal and investigation. IEEE Trans Electron Devices 65(10):4701–4708
26. Aghandeh H, Ziabari SA (2017) Gate engineered heterostructure junctionless TFET with
Gaussian doping profile for ambipolar suppression and electrical performance improvement.
Superlattices Microst 1(111):103–114
27. Mouritsen OG, Ipsen JH, Zuckermann MJ (1989) Lateral density fluctuations in the chain-
melting phase transition of lipid monolayers. J Colloid Interface Sci 129(1):32–40
28. Schenk A (1992) A model for the field and temperature. Solid State Electron 35(11):1585–1596
29. Daineka D, Bulkin P, Girard G, Bour J (2004) High density plasma enhanced chemical vapor
deposition. Phys J 9:3–9
A Study on Stability Analysis of QT
Interval Dynamics of ECG Using
ARMAX Model
A. Jyothsana and J. Sivaraman
Abstract The stability analysis of Electrocardiogram (ECG) intervals helps in deter-

mining its response to various disturbances in the cardiac cycle. ARMAX (Autore-
gressive with Moving Average and Exogenous Input) modeling of QT Interval (QTI)
using RR Interval (RRI) as exogenous input shows the dependency of ventricular
repolarization on heart rate. This study predicts the QTI using the ARMAX model,
to evaluate the stability of QTI dynamics and then, to identify the factors responsible
for causing instabilities. One minute-long ECG signals were recorded from Normal
sinus Rhythm (NSR) subjects and the QTI of each signal are then modeled using the
ARMAX approach. The stability of QTI dynamics was determined using the pole-
zero plot. The proposed ARMAX model was capable of detecting unstable QTI in
NSR subject and the unstable QTI was mainly due to Heart Rate Variability (HRV)
within one-minute ECG.
Keywords ECG · Heart rate variability · QT interval · Stability analysis
1 Introduction
The action potential of cardiac muscle cells is obtained as an Electrocardiogram

(ECG) recording. The QRS complex of the ECG signal is obtained as a result of
depolarization of the ventricles and the T wave of ECG signal represents ventricular
repolarization. Thus, the QT Interval (QTI) which consists of both QRS complex
and T wave signifies the Action Potential Duration (APD) of the ventricle. The
instability of cellular APD can be seen in the QTI of the ECG signal. The unstable
QT Intervals (QTIs) increase the chances of arrhythmia in patients having different
A. Jyothsana · J. Sivaraman (B)

Bio-Signals and Medical Instrumentation Laboratory, Department of Biotechnology and Medical
Engineering, National Institute of Technology, Rourkela, Odisha 769008, India
A. Jyothsana
308 A. Jyothsana and J. Sivaraman
cardiovascular diseases like QT syndrome [1–3] and acute myocardial infarction [4–
6]. In some studies, the instabilities of the QTI indicate Ventricular Tachycardia (VT)
and Ventricular fibrillation (VF) in patients [7]. Any abnormalities in the ventricular
repolarization of the ECG signal can be the root cause of different types of ventricular
arrhythmia. APD restitution is the dependency of APD on its preceding Diastolic
Interval (DI). The restitution slope detects the abnormality in APD. For a larger APD
restitution slope (>1) even smaller variations in DI can cause an unstable APD [8].
The adaptability of the ventricular APD with the heart rate was studied by Halamek
et al. [9] and the role of Premature Activation (PA) and QTI instability to the onset
of VT was explored by Chen et al. [10]. The normal QT variability can change with
the presence of PA in the ECG signal. The presence of such unstable QTIs can lead
to arrhythmias such as sustained VT in cardiac patients.
Recently, Imam et al. [11] have conducted a study to analyze the stability of
ECG interval and came up with the factors responsible for causing instability in
QTI. From that study, they found that HRV is also responsible for causing unstable
segments because HRV is more in healthy hearts and is more asymmetry as compared
to an unhealthy heart. A methodology for determining the stability of PTa Interval
(PTaI) was developed by Sivaraman et al. [12]. The authors studied the variation of
atrial repolarization using ARX model and explored the role of PA and HRV in PTaI
stability. Sivaraman et al. [13] had also studied the morphology of depolarization and
repolarization wave of atria in healthy subjects using the signal-averaging method
and they observed a decrease in the duration of repolarization wave with the increase
in heart rate. They also concluded that the duration of the repolarization wave varies
with the age of subjects. As ventricular arrhythmia continues for a longer time, the
refractory period of ventricular tends to become shorter. This will have an impact on
ventricular repolarization phase. The changes in ventricular repolarization can cause
ventricular arrhythmia.
This study is performed to analyze stability in QTI dynamics in NSR subjects.
The QTI dynamics are predicted using the ARMAX model. The equation derived
using the ARMAX approach in the time domain is then converted to the Z-domain
for accessing the stability. Further, the factors responsible for causing changes in the
QTI variability is also investigated in this study.
2 Methods
2.1 Subjects
One minute-long ECG data of fifty NSR subjects were collected using the EDAN
PC ECG system. The subject group had twenty-five males and twenty-five females
of mean age 29.4 ± 5.3 years (range 20–40 years). The subjects were medically
examined to exclude any form of cardiovascular disease. Smokers and patients with
congestive heart failure, valvular disease, and other cardiopulmonary diseases that
A Study on Stability Analysis of QT Interval Dynamics … 309
may alter the ECG morphology were excluded from this study. ECG data with less
noise is selected so that the QRS and T waves are visible and no further denoising
process is required.
2.2 ECG Data Analysis
All the collected ECGs were imported to MATLAB(R 2018a) for further digital data
analysis. The time duration from the beginning of the Q wave to the end of T wave
was defined as the QTI. The QTI time series is extracted from the ECG signals and
the RRI time series is extracted to detect the presence of PA and Huikuri et al. [14]
proposed a method to find out the count of PA from the RRI time series.
PA was encountered when the RRI of a beat was smaller than the previous beat by
at least 100 ms. The proposed methodology for detecting the instability in the QTI
consists of two steps. First, the dependency of QTI on its previous QTIs and RRIs
are represented using the ARMAX model. Second, the instability is determined by
converting the derived ARMAX model equation into Z-domain.
2.3 ARMAX Modeling in QTI Dynamics
ARMAX approach is used to model the QTI in one-min ECG

A
B
C
ai QTI(K − i) = bi RRI(K − i) + ci e(K − i) (1)
i=0 i=1 i=0
where K denotes the beats in 1 min ECG Segment. QTI and RRI are the discrete
signals which represent APD and DI. QTI(K), QTI (k − i) and RRI(k − i) denotes
the duration of the QTI and RRI for beat k and k − i, respectively. e(K) represents the
discrete signal of error values which occurs between successive QTIs. ai , bi and ci are
the weight constants for QTI, RRI, and error terms, respectively, which contribute
to QTI. A
A B and C represent the number of model parameters. The term i=0 ai QTI(K −
B
i) represents the autoregressive part, the term i=1 bi RRI(K − i) represents the
C
exogenous input and the last term i=0 ci e(K − i) represents the moving average
part of the ARMAX model. We have used A = B = C in this study where the value
of B and A indicate the effect of heart rate and ventricular repolarization that are used
in the model for the prediction of QTI duration as mentioned by Imam et al. [11].
We have used RRI instead of TQ Interval (TQI) because there is a chance for the
TQI to get affected by the previous QTI hence; TQI is no longer an exogenous input.
The system identification toolbox in MATLAB is used to determine the parameters
of the ARMAX model.
a0 QTI(k) + a1 QTI(k − 1) + · · · a A QTI(k − A)

= b1 RRI(k − 1) + · · · + b B RRI(k − B) + C0 e(k)
+ C1 e(k − 1) + · · · + Cc e(k − C) (2)
Equation (2) shows the expanded form of Eq. (1)

For each iteration, the order of QTI, i.e., A was incremented by 1. Then the
minimum value of A at which instability is obtained for the first time is denoted as
Amin . That is, the number of poles required to determine the instability for the first
time is Amin . Similarly, the value of A at which the predefined prediction value of
QTI is obtained was denoted as Amax . Hence, the number of poles required to predict
the QTI accurately for the first time is Amax . The Mean Square Error (MSE) between
the calculated and measured QTI value should be less than 5 ms2 for better accuracy.
2.4 Stability Analysis of QT Intervals
To evaluate the stability of QTI dynamics, the ARMAX model equation in the time
domain is converted to Z-domain.
Equation (3) represents the Z-Transform of Eq. (2)
a0 QTI(Z ) + a1 Z −1 QTI(Z ) + · · · + a A Z −A QTI(Z )

= b1 Z −1 RRI(z) + · · · + b B Z −B RRI(Z )
+ C0 e(z) + C1 Z −1 e(Z ) + · · · + Cc Z −C e(Z ) (3)
Equation (4) is obtained by rearranging Eq. (3)

QTI(Z ) a0 + a1 Z −1 + · · · + a A Z −A

= RRI(Z ) b1 z −1 + · · · + b B Z −B + e(Z ) C0 + C1 Z −1 + · · · + CC Z −C (4)
The relationship between the output QTI and input RRI can be expressed using a
transfer function based model as in Eq. (5)
QTI(z) b1 z −1 + b2 z −2 + · · · + b B z −B
G(z) = = (5)
RRI(z) a0 + a1 z −1 + · · · + a A z −A
The solution of the numerator and denominator polynomials in Eq. (5) represents
zeros and poles, respectively. If the values of poles and zeros are equal then they get
canceled each other. This pole-zero plots can be used to determine the stability of
ECG intervals and if any of the poles is having a magnitude greater than 1 or when
it comes outside the unit circle then such a condition is considered to be unstable.
3 Results
ARMAX model was developed for each one-minute-long ECG of NSR subjects. The
accuracy of predicted QTI is evaluated by measuring the MSE between the measured
QTI and calculated QTI. The model order required to predict the measured QTI with
better accuracy in one of the NSR subjects is Amax = 12, whereas at Amin = 7 the
accuracy of prediction is less, as the MSE between the measured and predicted values
of QTI is not less than 5 ms2 . However, the model shows instability condition for the
first time at Amin = 7 in NSR subject. The measured and predicted value of QTI at
Amin = 7 and Amax = 12 for NSR subject is obtained in Fig. 1a, b. In the pole-zero
plot of NSR subject as the value of ‘A’ is increased from 7 to 12 it was observed that
the number of poles is coming outside the unit circle is increased but the position
of poles having a magnitude greater than unity obtained at Amin = 7 remains the
same as observed in Fig. 2a, b, respectively. The results are showing that the larger
number of model orders are essential for better prediction of QTI but the instability
can be found using the minimum number of model orders in the NSR subjects. The
variation of the MSE with the number of beats for NSR subject is obtained in Fig. 3a.
4 Discussion
Chen et al. [10] used an autoregressive model with exogenous input as RRI to access
the stability of the ventricular repolarization process and the model includes the ECG
artifacts which can cause instability in QTI. The proposed ARMAX model is capable
of separating the system dynamics from ECG artifacts. Imam et al. [11] found that
PA is not the main cause for instabilities in QTI dynamics and is in agreement with
the present finding of this study. He found that the HRV is more complex in healthy
subjects as compared to VT subjects and hence, the PA beat is not the only factor for
causing the unstable and stable conditions in QTI dynamics.
Using the ARMAX model, we have analyzed the effect of HRV on QTI in NSR
subjects. Since the QTI affects the impending TQI, the focus of this study is on
HRV based on analyzing RRIs as it covers both QTI and TQIs. Many studies have
concluded that QTI dynamics can get affected by several factors such as respiration,
age, temperature, gender, genetic profile, and autonomous nervous system [15, 16].
A study was conducted on the variations of PTaI dynamics with sinus rate, pacing,
and drugs in heart block patients by Debbas et al. [17]. In a study conducted by
Sivaraman et al. [13] on P and Ta wave morphology it was noted that with age there
is a prolongation in the P and Ta wave duration and P–Ta interval get shortened with
the increase in heart rate. Acharya et al. [18] have reported that HRV is more in a
healthy heart and it is also more asymmetric as compared to the diseased heart. In this
study, the impact of HRV has caused instability in the QTI dynamics of NSR subject
whereas the chances of PA are minimum. In addition to HRV, the other intrinsic
factors which are described in [15, 16] can also be held responsible for causing
Fig. 1 a Measured and predicted value of QTI for NSR subject at Amin = 7. b Measured and
predicted value of QTI for NSR subject at Amax = 12
Fig. 2 a Pole-zero plot for NSR subject at Amin = 7. b Pole-zero plot for NSR subject at Amax =
12
instabilities in NSR subject which are in agreement with the study done by Imam
et al.
Fig. 3 a Variation of mean square error with the order
5 Conclusion
In this study, for the stability analysis of ventricular repolarization, the ARMAX
model has been used in the determination of impending QTIs. The value of QTI
predicted using derived ARMAX model was in agreement with the measured QTI in
NSR subjects. In this study, we have found the instability condition in NSR subjects
and the presence of instability in NSR subjects was mainly attributed to the presence
of HRV. The result of this study indicates that the ECG signal predicted using the
ARMAX model can be used to find out the development of instability in QTI due to
HRV in healthy subjects. These instabilities are responsible for the development of
arrhythmia risk. From this preliminary study, it was found that the ARMAX model
was capable of predicting the QTI more accurately. Also, it determines the unstable
intervals at minimum order which helps in saving the computational time.
Acknowledgements The authors acknowledge the support from MHRD, Government of India,
for sponsoring the M.Tech. Programme of the first author. The present study was supported by
financial grants from the Science Engineering Research Board (SERB), Department of Science and
Technology, Government of India (EEQ/2019/000148).
References
1. Merri M, Moss AJ, Benhorin J, Locati EH, Alberti M, Badilini F (1992) Relation between
ventricular repolarization duration and cardiac cycle length during 24-hour Holter recordings:
findings in normal patients and patients with long QT syndrome. Circulation 85(5):1816–1821
2. Yamauchi S, Yamaki M, Watanabe T, Yuuki K, Kubota I, Tomoike H (2002) Restitution proper-
ties and occurrence of ventricular arrhythmia in LQT2 type of long QT syndrome. J Cardiovasc
Electrophysiol 13(9):910–914
3. Chinushi M, Restivo M, Caref E, El-Sherif N (1998) Electrophysiological basis of arrhythmo-
genicity of QT/T alternans in the long-QT syndrome: tridimensional analysis of the kinetics
of cardiac repolarization. Circ Res 83(6):614–628
4. Schwartz PJ, Wolf S (1978) QT interval prolongation as predictor of sudden death in patients
with myocardial infarction. Circulation 57(6):1074–1077
5. Bonnemeier H, Wiegand UK, Bode F, Hartmann F, Kurowski V, Katus HA, Richardt G (2003)
Impact of infarct-related artery flow on QT dynamicity in patients undergoing direct percuta-
neous coronary intervention for acute myocardial infarction. Circulation 108(24):2979–2986
6. Szydlo K, Trusz-Gluza M, Wita K, Filipecki A, Orszulak W, Urbanczyk D, Krauze J, Kolasa J,
Tabor Z (2008) QT/RR relationship in patients after remote anterior myocardial infarction with
left ventricular dysfunction and different types of ventricular arrhythmias. Ann Noninvasive
Electrocardiol 13(1):61–66
7. Couderc JP (2009) Measurement and regulation of cardiac Ventricular repolarization: from
the qt interval to repolarization morphology. Philos Trans R Soc A Math Phys Eng Sci
367(1892):1283–1299
8. Nolasco JB, Dahlen RW (1968) A graphic method for the study of alternation in cardiac action
potentials. J Appl Physiol 25(2):191–196
9. Halamek J, Jurak P, Lipoldova J, Leinveber P (2010) QT/RR coupling and gender differences.
Comput Cardiol 37(2010):365–368
10. Chen X, Trayanova NA (2012) A novel methodology for assessing the bounded-input
bounded-output instability in qt interval dynamics: application to clinical ecg with ventricular
tachycardia. IEEE Trans Biomed Eng 59(8):2111–2117
11. Imam MH, Karmakar CK, Khandoker AH, Palaniswami M (2013) Effect of premature acti-
vation in analyzing QT dynamics instability using QT-RR model for ventricular fibrillation
and healthy subjects. In: 2013 proceedings of annual international conference of the IEEE
engineering in medicine and biology society (EMBC), pp 2559–2562
12. Sivaraman J, Uma G, Langley P, Umapathy M, Venkatesan S, Palanikumar G (2016) A study
on stability analysis of atrial repolarization variability using ARX model in sinus rhythm and
atrial tachycardia ECGs. Comput Methods Programs Biomed 137:341–351
13. Sivaraman J, Uma G, Venkatesan S, Umapathy M, Kumar NK (2014) A study on atrial ta wave
morphology in healthy subjects: an approach using P wave signal-averaging method. J Med
Imaging Healthinfo 4(5):675–680
14. Huikuri HV, Valkama JO, Airaksinen KE, Seppanen T, Kessler KM, Takkunen JT, Myerburg RJ
(1993) Frequency domain measures of heart rate variability before the onset of nonsustained
and sustained ventricular tachycardia in patients with coronary artery disease. Circulation
87(4):1220–1228
15. Almeida R, Gouveia S, Rocha AP, Pueyo E, Martinez JP, Laguna P (2006) QT variability and
HRV interactions in ECG: quantification and reliability. IEEE Trans Biomed Eng 53(7):1317–
1329
16. Porta A, Tobaldini E, Gnecchi-Ruscone T, Montano N (2010) RT variability unrelated to heart
period and respiration progressively increases during graded head-up tilt. Am J Physiol Hear
Circ Physiol 298(5):1406–1415
17. Debbas NMG, Jackson SHD, De Jonghe D, Robert A, Camm AJ (1999) Human atrial repo-
larization: Effects of sinus rate, pacing and drugs on the surface electrocardiogram. J Am Coll
Cardiol 33(2):358–365
18. Acharya UR, Joseph KP, Kannathal N, Lim CM, Suri JS (2006) Heart rate variability: a review.
Med Biol Eng Comput 44(12):1031–1051
High Frequency Resonant Inverter
with Shunt Active Power Filter
for Harmonic Compensation
Rohan Sinha and Rahul Raman
Abstract This paper presents a high frequency resonant inverter which has been
used as a power supply unit for an induction heating unit. Generally, power electronic
devices being nonlinear loads introduce quite a high amount of harmonics, which
in turn deteriorates the power factor. Therefore, this paper proposes a power filter
stage has been devised for the mitigation of harmonics. As the main harmonic-related
problem is with the input current, a shunt active power filter has been used. A voltage
source inverter (VSI) has been used for the design of the filter, which has an additional
control circuitry. The main aim here is to produce the compensating currents, which
in turn governs the switching of the VSI. For this purpose, the instantaneous power
theory (PQ theory) has been used for generation of compensating currents; these
currents are then fed into a hysteresis controller (aversion of PWM) for generation
of the gating signals. All the circuits are simulated using the PSIM platform and the
results with and without the filter have been compared to obtain the desired results.
Keywords Resonant inverter · Shunt active power filter · VSI · Instantaneous

power theory · Hysteresis control · THD · FFT · PSIM · Induction heating
1 Introduction
As the use of electronic and power electronic devices like is increasing at a steady
rate, they pose some disorder in the utility supply because of their nonlinear nature.
On such device is an inverter, a device which converts DC signals into AC signals.
The inverter which will be discussed in this project is a high frequency resonant
inverter, [1] which will be used as a power supply unit for induction heating. Induc-
tion heating is one of the most popular heating technologies used in the present
R. Sinha (B)
Department of Electrical Engineering, Jorhat Engineering College, Jorhat, Assam, India
R. Raman
Department of Electrical Engineering, Indian Institute of Technology (ISM), Dhanbad, India
318 R. Sinha and R. Raman
era right from domestic heating to industrial heating. This is mainly because of
the advantages associated with it like compact size, high power density, efficiency,
easy controllability over power, etc. There are many configuration of the inverter
which can be used for induction heating purpose like Class E resonant inverters [1],
single ended resonant inverter [3], full bridge inverter [2–5], modified half bridge
inverter [7], etc. For induction heating to be effective, the inverter concerned should
switch at a very high frequency so as to be able to produce a high frequency AC
output signal. This high frequency oscillating voltage signals induce eddy current
in the workpiece causing the object to heat given by Joule’s heating law. But being
nonlinear loads and high frequency switching devices, they produce a high amount
of harmonics. These harmonics have a natural tendency to flow back into the utility
supply, thereby causing distortion of the input signal. Also, these harmonics are
responsible for creating undesirable conditions in the grid, special when sensitive
equipment is connected to the supply. Therefore, there is a need of circuits which
can eliminate these harmonics making the utility supply sinusoidal. These circuits are
referred to as filter circuits which are classified broadly as passive filters [5] and active
power filters [6–9]. Passive filters do play a vital role in harmonic mitigation, but they
come with some limitations such as large inductor or capacitor size, their tendency
to undergo resonance with the source and the poor dynamic response of these filters.
These are the reasons why passive filters fail to compensate for the harmonics under
certain circumstances. Active power factors are therefore a preferable option which
can overcome the limitations of passive filter and provide an effective harmonic
compensation.
In this paper, the filter which will be coupled along with the high frequency
resonant inverter is a shunt active power filter(SAPF). An SAPF [10] is a variant
of power filters which is used for current harmonic mitigation connected between
the nonlinear load and supply. This filter is generally preferred due to simplicity in
design, easy control circuit and low cost. It basically consists of a voltage source
inverter (VSI) for producing the compensating current and a control circuit which
produces the reference current for generation of the gating pulse for the VSI. The
instantaneous power theory has been employed for extraction of the current and a
hysteresis current controller is used for generating the gating pulses. The shunt active
power filter works by producing harmonics of equal magnitude but is 180° phase
shifted. These signals generated by the filter are injected at the PCC and cancel out
the harmonics making the utility supply near sinusoidal.
2 High Frequency Resonant Inverter
As it has been mentioned above, for the purpose of induction heating, the voltage
signal to the load should be of high frequency. Therefore, a high frequency resonant
inverter is employed for the purpose of generating this high frequency output [2, 5].
The inverter used here is a full bridge inverter (four IGBTs are used as switch) which
is PWM controlled. The resonant condition created by the load, where the load forms
High Frequency Resonant Inverter with Shunt … 319
a parallel resonating circuit. Only when the resonant frequency of the load matches
the switching frequency of the inverter, the inverter is said to be operating in the
resonant state. This condition has to be fulfilled for the source to be able to transfer
the maximum power to the load or the induction heater in this case.
2.1 Calculation of Switching Frequency and Load

Parameters
For the resonant condition to deliver the maximum power to the load without power
loss, the switching frequency of the inverter has to be precisely determined [2]. The
switching frequency f s should be near the resonance of the tank circuit. Here, the
formula of parallel resonance is used for calculation. For it to satisfy the resonant
condition, the susceptance part (imaginary) should be zero.

L eq − Ceq R 2
ω= (1)
L 2eq Ceq

1 L eq − Ce Req 2
fs = (2)
2π L 2 Ceq
Always the capacitor and inductor value should be selected first because variable
capacitors and inductors are expensive and easily available. On putting the values
of the different parameters Req = 1 , L eq = 57.1 µH, Ceq = 0.2 µF, the value of
f s = 47 kHz. Therefore, Eq. (2) gives the resonant frequency which is essential for
the inverter to resonate and to ensure ZVS.
3 Reference Current Generation
One of the most important part of the circuit is the control circuit based on which the
gating pulses of the SAPF is determined. The proper control of the filter is required for
effective compensation. To achieve the reference currents, the instantaneous power
theory (PQ) has been used [6, 8, 10]. The instantaneous power theory was proposed
by Akagi in the year 1983 and is a time domain analysis approach. This is a theory
which is based on the conversion of abc coordinate to the αβo coordinate. And then
define instantaneous power based on these coordinates. This control through current
generation method can be used for both three-phase three-wire or four-wire systems.
The main transformation consists of the Clark’s transform given as,
⎡ ⎤ ⎡ √1 √1 √1
⎤⎡ ⎤
Vo Vsa
⎣ Vα ⎦ = 2 ⎢ ⎥⎣
2 2 2
⎣ 1 1
√2
−√21 ⎦ Vsb ⎦ (3)
3
Vβ 0 2
3
− 23 Vsc
⎡ ⎤ ⎡ ⎤⎡ ⎤
Io √1 √1 √1 Isa
⎣ Iα ⎦ = 2 ⎢ 2 2 2
⎥⎣ ⎦
⎣ 1 1
√2
−√21 ⎦ Isb (4)
3
Iβ 0 2
3
− 23 Isc
After the above transformation is performed, the instantaneous powers have to be

now calculated. As in a three-phase system, the current in Io = 0 the power Po is
also equal to zero. So we have to now consider only the p and q part of the power.
Therefore, the power can be given as follows,

p Vα Vβ Iα
= (5)
q −Vβ Vα Iβ
As it is clear from the above equation, the instantaneous power consists of two
main parts the oscillatory and the average components. The SAPF should be able
to compensate for this oscillatory power. After the power calculation, an low pass
filter (LPF) is required to eliminate the spurious components and extract the funda-
mental components from the power p. The next step is to determine the compensating
currents given as,

Isα 1 Vα Vβ p
= (6)
Isβ Vα2 + Vβ2 Vβ −Vα q
Having calculated the reference currents which are given by Eq. (6), these currents
now need to be transferred back to the abc coordinate. For this purpose, the inverse
Clark’s transformation is used which gives back the compensating currents in abc
coordinates.
⎡ ∗ ⎤ ⎡ ⎤
Ica 1 √0

⎣ I∗ ⎦ = 2⎢ ⎣ −1 3 ⎥ Isα
⎦ (7)
cb
3 −1 2 √
2 Isβ
∗ − 3
Icc 2 2
One of the issues associated in controlling the power filter is to maintain the
DC link voltage. A separate controller has to be present in order to maintain a
constant DC link. For this purpose, a proportional integral controller has been used
which compares the voltage across the DC link capacitor and a reference voltage to
produce a signal. This signal is than compared with the “p” part of the calculated
power to obtain the required compensating power signal. So as to be able to mitigate
the harmonics effectively, the reference voltage is always kept higher than that of the
input utility supply. Therefore, the recommended reference voltage is always kept
three to four times higher than the input.
4 Hysteresis Current Control
Hysteresis control is one of the most well know methods for controlling inverters and
other switching devices [10]. It is actually an improvised version of PWM generation.
In this proposed method, a Schmitt trigger has been used to form the PWM signal. A
Schmitt trigger is a comparator with a positive feedback which forces the Op Amp
into the saturation region. Using which a hysteresis band is formulated having two
thresholds, the upper threshold and the lower threshold. Every time the error signals
tries to cross these thresholds, the controller produces a change in the external signal.
The error current deviate within the hysteresis band and forms the required pulsating
signal (Fig. 1).
The hysteresis bands upper and lower thresholds can be determined by the formula
given below. (A non-inverting type Schmitt trigger is used for this purpose)
−(+Vsat )R1
VLT = (8)
R2
−(−Vsat )R1
VUT = (9)
R2
R1
VH = [+Vsat − (−Vsat )] (10)
R2
Fig. 1 Hysteresis controller

[11]
Fig. 2 General layout of the method used
5 Block Diagram of Method Used
The above general layout represents the method which has been used for the compen-
sation of the harmonics produced by the high frequency resonant inverter. The control
is divided into two parts, the reference current generation unit-based PQ theory and
the hysteresis controller which is responsible for generating the switching pulses.
These pulses are then fed into a three-phase VSI acting as the SAPF. The compen-
sating currents are further injected at the point where the load is connected to the
supply. The resonant inverter is fed from a three-phase power supply via a diode
bridge rectifier which provides the pulsating DC supply for the inverter. For the
reference current generation, the source voltages and currents are utilised (Fig. 2).
6 Simulation Diagrams and Experimental Results
All the simulation has been performed in PSIM platform. The simulation diagrams
consist of high frequency resonant inverter without the filter and another one with
the shunt active power filter. After that, the input currents are compared and their
FFT analysis is done. The core motive of this work is to improve the signal quality
of the input signal which is fed to the induction heater, which further says that the
THD content of the signal has to be improved within some tolerable limit (Figs. 3,
4, 5, 6, 7 and 8).
Fig. 3 High frequency resonant inverter without SAPF
Fig. 4 Supply input current of inverter without SAPF
Fig. 5 FFT analysis of input current without filter

Fig. 6 Simulated circuit of the inverter with the SAPF
Fig. 7 Input current signal with SAPF
Fig. 8 FFT analysis of the input signal using SAPF

7 Calculation of Total Harmonic Distortion
As it is evident from the FFT analysis of the two signals, the one without the filter
is quite non-sinusoidal and contains a high amount of harmonic. The one where the
filter has been installed contains lower amount of harmonic and the utility power
supply is almost near sinusoidal. The THD can be calculated as given below [5],

∞ 2
n=2,3 In rms
THD = (11)
I1 rms
7.1 THD Without Filter

√
5.2082 + 2.86082 + 2.3912 + 0.73322 + 0.67272
THD =
14.5
= 45.25%
7.2 THD with Filter

√
22
THD =
20.6
= 9.7%
8 Conclusion
This work therefore deals with the compensation of harmonics to improve the overall
power factor and quality of the input current signal. Also, a controller has been
designed based on the of instantaneous power theory to control the active filter.
Then the experimental results of the simulation have been compared to determine
the effectiveness of the filter. Without the filter the THD content has been found out
to be at a 45.25% and with the SAPF the THD has been significantly brought down
to 9.7%. So, it can be conclude that the filter has been able to reduce the harmonics,
thereby improving the overall power factor.
References
1. Gorain C, Sadhu PK, Raman R (2018) Analysis of high-frequency class E resonant inverter
and its application in an induction heater. In: 2nd international conference on power, energy
and environment: towards smart technology, IEEE
2. Kumar A, Sadhu PK, Raman R, Singh J (2018) Design analysis of full-bridge parallel reso-
nant inverter for induction heating application using pulse density modulation technique. In:
International conference on power energy, environment and intelligent control, IEEE, 13–14
April 2018
3. Yeon JE, Cho KM, Kim HJ (2015) A 3.6 kW single-ended resonant inverter for induction
heating applications. IEEE, 8–10 Sept 2015
4. Sugimura H, Sumiyoshi S, Mun SP, Kwon SK, Nakaoka M (2008) A novel one stage ZVS-PWM
high frequency resonant inverter with boost single diode conducting mode passive ZCS-PFC
scheme, IEEE, 10–13 Nov 2008
5. Raman R, Sadhu P, Kumar A, Kumar Sadhu P (2018) Design and analysis of EMI and
RFI suppressor for induction heating equipment using vienna rectifier. In: 2nd international
conference on power, energy and environment: towards smart technology, IEEE, 1–2 June 2018
6. Fujita H, Akagi H (1991) A practical approach to harmonic compensation in power systems-
series connection of passive and active filters. IEEE Trans Indus Appl 27(6):1020–1025
7. Pal P, Sadhu PK, Pal N, Sanyal S (2015) An exclusive design of EMI-RFI suppressor for
modified half bridge inverter fitted induction heating equipment. Int J Mechatron Electr Comput
Technol (IJMEC) 5(15):2084–2100
8. Afonso JL, Pinto JG, Gonçalves H. Active power conditioners to mitigate power quality
problems in industrial facilities. In: Power quality issues. IntechOpen
9. Budhrani AH, Bhayani KJ, Pathak AR (2018) Design parameters of shunt active filter for
harmonics current mitigation. PDPU J Energy Manage 59–65
10. Shah A, Vaghela N (2014) Shunt active power filter for power quality improvement in
distribution systems. Int J Eng Dev Res 1(2):23–27
11. Kazmierkowski MP (2002) Pulse width modulation techniques for three-phase voltage source.
In: Krishnan R, Blaabjerg F (eds) Control in power electronics, 1st edn. Academic Press, 84
Theobolds Road, London WC1X 8RR, UK, pp 89–160
Rohan Sinha hails from Jorhat, Assam, India. He received his

B.Tech degree in Electronics and Communication Engineering
from DUIET, Dibrugarh University, Dibrugarh, Assam. He is
currently pursuing his M.Tech degree in Electrical Engineering
from Jorhat Engineering College, Assam, India.
Rahul Raman received the B.Tech degree in Electrical Engi-

neering from National Institute of Technology, Agartala, India,
and the M.Tech degree in Power System Engineering from
Indian Institute of Technology (ISM), Dhanbad. He is currently
a Doctoral Research Scholar with the Department of Electrical
Engineering, Indian Institute of Technology (ISM), Dhanbad,
and he is also working as an Assistant Professor in the Depart-
ment of Electrical Engineering, Jorhat Engineering College,
Assam, India. His area of research interest includes induc-
tion heating applications, power quality improvement, high
frequency resonant inverter, filter designing, etc. He has several
journal and conference publications at national and international
levels along with a patent published in the same domain.
Modelling of Fiber Bragg Grating Sensor
for Acoustic Pressure Wave Detection
of Partial Discharge in Power
Transformer
Sorokhaibam Nilakanta Meitei, Kunal Borah, and Saibal Chatterjee
Abstract Partial discharge (PD) detection in power transformer is a useful condition

monitoring process. Various sensors like piezoelectric, optical fiber, capacitive and
ultra-high frequency (UHF) sensors have been used for the monitoring of the power
transformer. FBG sensor can be installed inside the transformer tank and is able to
capture even low acoustic signal of PD due to its high sensitivity, small size and
immunity to EMI. Hence in this work, an FBG sensor has been modelled for PD
detection in transformer using COMSOL Multiphysics. Authors have studied the
sensing performance of the model FBG sensor with different coating radii.
Keywords Power transformer · PD · Fiber Bragg grating (FBG) · Finite element

method
1 Introduction
In the power grid, power transformers are costly and important equipment because
of its working performance is direct influences the security and stability of the whole
network [1]. Complete breakdowns of the transformer can happen without any signs,
following in dangerous insulating oil spills, widespread damage to nearby apparatus
and extreme interruption of service. Insulation breakdown is the main reason for
S. N. Meitei
K. Borah
Department of Physics, North Eastern Regional Institute of Science and Technology, Nirjuli,
Arunachal Pradesh 791109, India
S. Chatterjee (B)
Department of Electrical and Electronics Engineering, National Institute of Technology Mizoram,
Aizawl, Chaltlang, Mizoram 796012, India
330 S. N. Meitei et al.
the transformer failures, and frequently, PDs are included in such failure by deterio-
rating the dielectric characteristics of the insulating oil [2]. Therefore, continuously
checking the health condition of the insulation system is necessary to protect from
a complete breakdown and to avoid the huge cost. The most commonly adopted PD
detection technique for power transformer is namely, electrical, chemical, acoustic,
optical and opto-acoustic techniques [2, 3]. Among these PD detection methods of
power transformer, the opto-acoustic method is adopted in this work.
The PD acoustic wave can be detected by using a PZT sensor. Recently, Meitei
et al. [4] have modelled a piezoelectric sensor for PD detection and localization
in distribution transformer using COMSOL Multiphysics software. An externally
installed PZT sensor has the merits of a simple arrangement [5]. But it may deteri-
orate from surrounding noise like EMI and corona event, and it is also complicated
in discovering the precise location of PD source due to its various routes of acoustic
sound wave propagation in the transformer. To overcome these demerits of an exter-
nally installed piezoelectric sensor, an optical-based FBG sensor is applied for the
application of PD detection [6]. FBG sensors have been designed for their application
in different fields, like hydrogen gas detection in an oil-filled power transformer [7,
8], structural condition monitoring [9–11] and underwater acoustic detection [12].
Few researchers [13–15] have been experimentally investigated the PD detection in
an oil-filled power transformer using Fiber Bragg Grating (FBG) sensor. Massimo
Moccia et al. [16, 17] have been designed the FBG sensor using COMSOL Multi-
physics software for the application of an underwater acoustic sensor. In this work,
authors have designed an acoustic FBG sensor using FEM-based software COMSOL
Multiphysics for PD acoustic wave detection in power transformer.
2 Theoretical Background and Simulation Model
FBG is a type of distributed Bragg reflector constructed in a short segment of optical

fiber core [18]. When the broadband light is incident on the FBG, it reflects a particular
wavelength and transmits all others. The reflected wavelength is known as the Bragg
wavelength, and as follows [19].
λ B = 2n eff α (1)
where α represents the grating period.

n eff represents the effective refractive index.
The working principle of PD detection in a power transformer using an FBG
sensor is presented in Fig. 1. Generally, it includes a complicated interaction of
several physical phenomena.
When the PD acoustic wave distributing in an oil (i.e. transformer insulation
oil) strikes on the composite structure of FBG sensor, its geometry gets mechanical
deformation. The strain appears on the FBG due to mechanical deformation. In this
work, transformer oil where PD acoustic pressure waves distribute is defined as an
Modelling of Fiber Bragg Grating Sensor for Acoustic Pressure … 331
Fig. 1 Working principle of

PD detection using an FBG
sensor
acoustic medium. The model FBG sensor consists of cylindrical shape optical fiber
and coating is defined as the mechanical medium.
2.1 Power Transformer and FBG Sensor Model
The 3D model of 3-phase, 35 kV power transformer [20] is modelled using the CAD
tools of COMSOL Multiphysics and is presented in Fig. 2. The sizes of the power
transformer and its materials properties are given in Tables 1 and 2.
The 3D model of partial discharge acoustic FBG sensor is presented in Fig. 3.
It is a composite structure which are consists of two-cylinder, inner cylinder which
represents the optical fiber, and the another one is the outer cylinder which represents
Fig. 2 Power transformer

model
Table 1 Dimensions of 3D
Component Dimension (m)
model of power transformer
[20] Transformer tank (L × W × H) 2.2 × 1 × 1.5
Table 2 Properties of the

Components Material Density Speed of
material [20]
ρ0 kg/m3 sound (m/s)
Transformer Structural 7850 5100

tank and core steel
Winding Copper 8700 4760
Oil Insulating 890 1420
oil
Fig. 3 3D model of FBG

sensor
the coating of the fiber. The FBG model dimensions and its materials properties are
given in Tables 3 and 4.
The modelled FBG sensor is installed inside the oil-filled power transformer at
coordinate (0.2, 0.1, 0.7) to detect PD acoustic wave signal. Considering that the PD
acoustic plane wave is incident on the FBG sensor in the form of an acoustic plane
wave. At the boundary between the transformer oil and transducer emitting side, the
constant acoustic pressure plane wave is defined (Dirichlet boundary condition) with
the amplitude of 1 MPa [17]. The PD investigation is conducted within the frequency
range of 80–200 kHz [3].
Table 3 Dimensions of 3D
Components Dimensions
model of FBG sensor [17]
Radius of optical fiber (rf ) 62.5 µm
Coating radius (rc ) 1.25 mm
Height of the optical fiber and coating (H) 5 mm
Table 4 Material properties

Components Young’s Poisson’s ratio Density (ρ)
of the model FBG sensor
modulus (E) (v)
components [17]
Optical fiber 72 GPa 0.17 2200 kg/m3
Coating 78 MPa 0.3 1180 kg/m3
2.2 Sensitivity Calculation of Model FBG Sensor
The sensitivity can be calculated by [17]:

λ 1 n 2eff
S= = εz − p11 εx + p12 εz + ε y (2)
λ0 p 0 p0 2
where λ, λ0 , p0 and n eff represent the Bragg wavelength shift, the central wavelength
of the FBG, pressure amplitude, the effective refractive index (i.e. n eff = 1.465), p11
and p12 represent elasto-optic parameters (i.e. p11 = 0.121, p12 = 0.265) and εi (i
= x, y, z) are the coordinate at the FBG position where the strain components are
calculated.
The sensitivity gain as follows [17].

S
Sensitivity Gain = 20 log10
(3)
S bare
where Sbare = −2.76 × 10−6 MPa−1 [17].
To study the opto-acoustical response of the coated FBG sensor, the authors consider
a specific sensor structure. For this first simulation case, the radius of the cylindrical
shape coating is 1.25 mm, the radius of the optical Fiber Bragg grating is 62.5 µm,
and the height is 5 mm [17]. The simulation is conducted in the frequency range of
80–200 kHz with an interval of 5 kHz [3].
When the acoustic plane wave produced by PD strikes on the FBG sensor, its
geometry gets mechanical deformation due to its elastic properties, and the strain
distributes on the surface of the model FBG sensor. The strain at the Bragg position
defines the appear Bragg wavelength shift according to Eq. (2). Figure 4 shows the
deformed structure of the model FBG sensor.
The sensitivity gain of the model FBG sensor is calculated using Eq. (3). The
sensitivity gain curve of the model FBG sensor is presented in Fig. 5. The average
sensitivity gain of the model FBG sensor was −230.43 dB, and the maximum sensi-
tivity gain was −114.52 dB. From Fig. 5, obtained that the response of the model
FBG sensor is decreasing from the lower frequencies to higher frequencies.
To analyze the sensing performance of the model FBG sensor, we have conducted
the parametric study by changing the coating radius. In this study, dimensions and the
material properties of the FBG sensor are kept as presented in Tables 3 and 4 except
for the coating radius. The coating radius has been taken following a paper [17]
to check the validity of our model based on COMSOL Multiphysics software. The
(a) (b) (c) (d)
Fig. 4 Deformed structure and the z-strain distribution on the FBG sensor surface for various
frequencies
Fig. 5 Sensitivity gain curve
authors have investigated the sensing performance of the model FBG sensor by taking
three different coating radii (i.e. rc = 0.625 mm, rc = 1.25 mm and rc = 2.5 mm).
Figure 6 shows the sensitivity gain curve for different coating radii. It can be
seen that the, when the coating radius is rc = 0.625 mm, the average sensitivity gain
of the model FBG sensor was −236.80 dB and the maximum sensitivity gain was
−117.40 dB. When rc = 1.25 mm, the average sensitivity gain was −230.43 dB
and the maximum sensitivity gain was −114.52 dB. Similarly, when rc = 2.5 mm,
the average sensitivity gain was −227.53 dB and the maximum sensitivity gain was
−107.72 dB. From the analysis, we found that increasing the radius of coating,
enhancing the sensitivity of the model FBG sensor. Therefore, the model FBG
sensor with coating radius rc = 2.5 mm had a better sensing performance for partial
discharge acoustic pressure wave detection. To the best authors’ knowledge, there
Fig. 6 Sensitivity gain curve

for different coating radii
exists no work on modelling of FBG as a PD sensor using COMSOL Multiphysics

software. But, similar work on modelling of FBG sensor for underwater acoustic
detection application using COMSOL Multiphysics software can be found in [16,
17] and the sensitivity performance of their underwater acoustic FBG sensor model
is in good agreement with our model.
4 Conclusion
In this paper, modelled a coated FBG sensor for PD detection in power transformer
using FEM-based software COMSOL Multiphysics. A parametric analysis was
carried out by varying the coating radius of the FBG sensor to study its sensing perfor-
mance. It obtained that increasing the radius of coating, enhancing the sensitivity of
the model FBG sensor. Hence, the sensitivity response to the same acoustic wave
signals in the frequency range of 80–200 kHz of the model FBG sensor with a radius of
rc = 2.5 mm has a better performance than the other two models. The proposed FBG
sensor model shows the usefulness and effectiveness of the PD detection technique
in a power transformer.
References
1. Xue N, Yang J, Shen D, Xu P, Yang K, Zhuo Z, Zhang L, Zhang J (2019) The location of partial
discharge sources inside power transformers based on TDOA database with UHF sensors. IEEE
Access 7(7):146732–146744
2. Deng J, Xiao H, Huo W, Luo M, May R, Wang A, Liu Y (2001) Optical fiber sensor-based
detection of partial discharges in power transformers. Optics Laser Technol 33(5):305–311
3. Zhou HY, Ma GM, Zhang M, Zhang HC, Li CR (2019) A high sensitivity optical fiber interfer-
ometer sensor for acoustic emission detection of partial discharge in power transformer. IEEE
Sens J
4. Meitei SN, Borah K, Chatterjee S (2020) Modelling of acoustic wave propagation due to partial
discharge and its detection and localization in an oil-filled distribution transformer. Frequenz
74(1–2):73–81
5. Zheng Q, Ma G, Jiang J, Li C, Zhan H (2015) A comparative study on partial discharge
ultrasonic detection using fiber Bragg grating sensor and piezoelectric transducer. In: 2015
IEEE conference on electrical insulation and dielectric phenomena (CEIDP), pp 282–285.
IEEE (2015)
6. Yu B, Kim DW, Deng J, Xiao H, Wang A (2003) Fiber Fabry-Perot sensors for detection of
partial discharges in power transformers. Appl Optics 42(16):3241–3250
7. Samsudin MR, Shee YG, Adikan FR, Razak BB, Dahari M (2016) Fiber Bragg grat-
ings hydrogen sensor for monitoring the degradation of transformer oil. IEEE Sens J
16(9):2993–2999
8. Jiang J, Ma GM, Li CR, Song HT, Luo YT, Wang HB (2015) Highly sensitive dissolved
hydrogen sensor based on side-polished fiber Bragg grating. IEEE Photon Technol Lett
27(13):1453–1456
9. Ma G, Li C, Quan J, Jiang J, Cheng Y (2011) A fiber Bragg grating tension and tilt sensor applied
to icing monitoring on overhead transmission lines. IEEE Trans Power Deliv 26(4):2163–2170
10. Goossens S, De Pauw B, Geernaert T, Salmanpour MS, Khodaei ZS, Karachalios E, Saenz-
Castillo D, Thienpont H, Berghmans F (2019) Aerospace-grade surface mounted optical fibre
strain sensor for structural health monitoring on composite structures evaluated against in-flight
conditions. Smart Mater Struct 28(6):065008
11. Tan CH, Shee YG, Yap BK, Adikan FM (2016) Fiber Bragg grating based sensing system: Early
corrosion detection for structural health monitoring. Sens Actuators A Phys 1(246):123–128
12. Bai X, Hu M, Gang T, Tian Q (2019) An ultrasonic sensor composed of a fiber Bragg grating
with an air bubble for underwater object detection. Optics Laser Technol 112:467–472
13. Ghorat M, Gharehpetian GB, Latifi H, Hejazi MA, Bagheri M (2019) High-resolution
FBG-based fiber-optic sensor with temperature compensation for PD monitoring. Sensors
19(23):5285
14. Ma G-M, Zhou H-Y, Shi C, Li Y-B, Zhang Q, Li C-R, Zheng Qing (2018) Distributed
partial discharge detection in a power transformer based on phase-shifted FBG. IEEE Sens
J 18(7):2788–2795
15. Kanakambaran S, Sarathi R, Srinivasan B (2017) Identification and localization of partial
discharge in transformer insulation adopting cross recurrence plot analysis of acoustic signals
detected using fiber Bragg gratings. IEEE Trans Dielectr Electr Insul 24(3):1773–1780
16. Moccia Massimo, Consales Marco, Iadicicco Agostino, Pisco Marco, Cutolo Antonello, Galdi
Vincenzo, Cusano Andrea (2012) Resonant hydrophones based on coated fiber Bragg gratings.
J Lightwave Technol 30(15):2472–2481
17. Moccia M, Pisco M, Cutolo A, Galdi V, Bevilacqua P, Cusano A (2011) Opto-acoustic behavior
of coated fiber Bragg gratings. Opt Express 19(20):18842–18860
18. Ma G-M, Li C-R, Mu R-D, Jun J, Luo Y-T (2014) Fiber Bragg grating sensor for hydrogen
detection in power transformers. IEEE Trans Dielectr Electr Insul 21(1):380–385
19. Onn BI, Arasu PT, Al-Qazwini Y, Abas AF, Tamchek N, Noor AS (2012) Fiber Bragg grating
sensor for detecting ageing transformer oil. In: 2012 IEEE 3rd international conference on
photonics, pp 110–113. IEEE (2012)
20. Gao C, Yu L, Xu Y, Wang W, Wang S, Wang P (2018) Partial discharge localization
inside transformer windings via fiber-optic acoustic sensor array. IEEE Trans Power Deliv
34(4):1251–1260
Design and Investigation of Tunnel
Junction Engineered Dopingless-TFET
with Improved DC and RF/Analog
Parameters Analysis
Abhishek Verma, Suruchi Sharma, Sneha Bharti, Manisha Bharti,

and Baljit Kaur
Abstract Today’s emerging trends of low power applications dopingless tunnel

field-effect transistor (DLTFET) will replace MOSFET in the future. Hence, in this
paper, we described a new tunnel junction engineered charge plasma-based DLTFET.
A pocket of gallium arsenide (GaAs) is placed across the source and channel inter-
face of the silicon film, thus increasing the likelihood of tunneling. Although GaAs
bandgap is slightly more than silicon bandgap it has high electron mobility and low
tunneling mass, resulting in increased current junction drivability. Besides, DC and
analog/RF merit figures were evaluated in consideration of the carrier concentra-
tions, energy bandgap diagram, parasitic capacitance, electric field, transconduc-
tance, transfer characteristics, cutoff frequency. Using the ATLAS device simulator,
the simulation of conventional DLTFET and proposed device (TJE-DLTFET) was
carried out. The proposed system showed a higher ON-current (~100 μA/μm) and
better subthreshold swing (~10.25 mV/decade). TJE-DLTFET showed a worthy
choice for future high-frequency applications.
Keywords Non-local BTBT · GaAs · Dopingless-TFET · Subthreshold swing ·

Tunneling junction
1 Introduction
MOSFET continuous scaling down enables us to scale down supply voltage,

enhancing current driving capacity with low power supply. Also, achieving a low
OFF-state leakage current is useful. However, at the same moment due to its low
ON-state current and its elevated subthreshold swing (SS) decreased to 60 mV per
decade, the device is restricted to some extent. Along with this, it suffered from SCE,
DIBL, and RDF because of doped regions MOSFET [1–5]. To reduce this limita-
tion of MOSFET, further introduced conventional doped TFET with low OFF-state
current that eliminates the problem of SCE, DIBL and it also reduces the subthreshold
A. Verma (B) · S. Sharma · S. Bharti · M. Bharti · B. Kaur

Department of Electronics and Communication, National Institute of Technology Delhi, Delhi,
India
338 A. Verma et al.
swing but increases the ON-state current by little amount because of poor BTBT [9–
11]. For getting high ON-state current and to eliminate the RDF because of the
doping mechanism, dopingless-TFET has been introduced. For the manufacture of
dopingless-TFET, the annealing process and the ion implantation techniques do not
need high temperature, this makes the dopingless-TFET realized with a low thermal
budget [12–14]. The most delicate step in the fabrication of TFET is the formation of
the abrupt junction for effective tunneling. Additionally, abrupt junction formation
isn’t an easy thing because of the massive diffusion of doping atoms existing in the
drain and source section through the intrinsic doped TFET channel [6–8]. However,
these problems related to doped TFET was solved by dopingless-TFET but for getting
more tunneling efficiency and further increment in ON-state current, Tunnel Junction
Engineered Dopingless TFET device is being proposed device (TJE-DLTFET).
For further comparative study purposes, the paper is broken down into four parts.
Part 1.2 consist of device structure and simulation parameters, part 1.3 familiar of
results and discussion which is sectioned into two subparts that is 1.3.1 DC anal-
ysis of transfer characteristics, carrier concentration, transconductance (gm ), energy
band diagram, electric field, and 1.3.2 analog/RF study of the drain-gate capacitance
(C gd ), gain bandwidth product (GBP), cutoff frequency (fT ), and maximum oscilla-
tion frequency (f max ) are addressed. Finally, the investigation throughout this paper
will be concluded in the conclusion part 1.4.
2 Device Structure and Simulation Parameters
An intersectional overview of DLTFET and the suggested system (TJE-DLTFET) are

included in Fig. 1a, b combined. Both devices are simulated using TCAD Silvaco
tool [12, 17, 18, 22, 23]. The parameters considered for the device structures are
shown in Table 1 Here, the charge plasma principle helpful to establish source p
(a) (b)
Fig. 1 Intersectional view of a DLTFET, b TJE-DLTFET

Design and Investigation of Tunnel Junction Engineered … 339
Table 1 Device structure and

Parameters Values Unit
design parameters were seen
in the simulation Length of drain (L D ) 50 nm
Length of source (L S ) 50 nm
Length of channel (L CH ) 20 nm
Drain and gate space (L GD ) 5 nm
Source and gate space (L GS ) 2 nm
The thickness of Silicon film (t Si ) 10 nm
The thickness of physical oxide 2 nm
(t ox )/EOT
Gate oxide material HfO2 (ε = 21) unitless
The work function of drain (M2) 3.9 eV
The work function of source 5.93 eV
(M3)
The work function of gate (M1) 4.3 eV
The thickness of inserted GaAs 2 nm
material
+ region and drain n + region. For source p + region, platinum with work func-
tion as 5.93 eV is chosen to induce holes and for drain n + region, hafnium with
work function as 3.9 eV is chosen to induce electrons [15, 16]. In the proposed
device (TJE-DLTFET), a vertical strip of gallium arsenide (GaAs) compound semi-
conductor material is incorporated across the source and channel of the thin silicon
film tunneling junction. Although GaAs compound semiconductor material has a
larger energy band gap than silicon, its inherent advantage of high electron mobility
(8500 cm2 /Vs) and smaller effective mass helps further enhancement of the band-
band tunneling (BTBT) efficiency all over the source and channel tunneling junction
[17]. The proposed device (TJE-DLTFET) can easily be fabricated in the same fashion
as given to Daley et al. of IBM in a US patent, where a vertical structure is arranged
as one layer stack composed of hetero-material surfaces [24]. The bandgap nature
of GaAs components is direct, while the bandgap nature of silicon is indirect. The
momentum corresponding to a minimum conduction band and also the maximum
valence band is just the same indirect bandgap material. Thus, an electron can jump
from conduction to valence band and vice versa by directly emitting or absorbing
the photon.
However, the momentum corresponding to the conduction band minima and
also the valence band maxima are different for an indirect bandgap material such
as silicon. As a consequence, making a shift from one band to another needs a
shift in momentum, and the electrons also need to absorb some additional energy
from the crystal lattice vibrations to achieve the necessary momentum shift. This
additional barrier in creating the transition significantly reduces the probability
of the carriers for Band to Band tunneling (BTBT) [18]. The simulation models
used for the simulation of the device in TCAD Silvaco include Fermi–Dirac,
340 A. Verma et al.
Table 2 Material Parameters

Parameters Silicon (Si) Gallium arsenide
used in Simulation [10, 21]
(GaAs)
Bandgap, E g 1.1 eV 1.42
Electron affinity, χ 4.07 4.07
Static dielectric constant, 11.7 13.2
ε
Electron effective mass, 0.522 0.072
me *
Hole effective mass, mh * 0.44 0.55
Auger recombination, Shockley–Read–Hall (SRH) Recombination model, Lombardi

(CVT) model, Concentration Dependent Recombination model, and Field Depen-
dent mobility models are all included for additional precise and reliable simula-
tion. Calculation of tunneling probability and the numerical solution is done by the
Wentzel–Kramers–Brillouin (WKB) method [19–21] (Table 2).
3.1 DC Analysis
This part holds the analysis of the concentration of carrier and energy band diagram
under OFF-state and also under ON-state. The impact of source and channel Tunnel
Junction Engineering is used in TJE-DLTFET, it could be verified in Fig. 2a–d
wherever it is seen that bandgap narrowed the same as DLTFET but a little sharp
variation of the valence band of source and channel interface in TJE-DLTFET because
the source and channel tunnel junction area comprises low effective mass material
(GaAs). Under the OFF-state, between source and channel, the width of the tunneling
barrier is high enough as shown in Fig. 2a, so charged carriers may not tunnel to
channel conduction band from the source valence band in TJE-DLTFET and DLTFET
which can be verified in Fig. 2b. Under ON-state condition, this can be seen as the
positive bias applied through the gate electrode, the width of the barrier is reduced
can be depicted in Fig. 2c which is sufficient to allow the majority of charge carriers to
passageway easily to the channel conduction band from the valence band of the source
region, it leads to an increase in the bulk of carriers charges across the channel, which
Fig. 2d depict. The conduct of drain current (I DS ) concerning gate voltage (V GS ) for
both DLTFET and TJE-DLTFET which can be verified from Fig. 3, which illustrates
that the ON-state drain current (I DS ) of TJE-DLTFET showed better improvement
compared to DLTFET. This increment in drain current (I DS ) is due to the availability
of GaAs material across the source and channel tunnel junction. By keeping drain
voltage V DS = 0.5 V constant and gate voltage, V GS varied where ON-state drain
(a) (b)
(c) (d)
Fig. 2 a Carrier concentration in OFF-state, b energy band diagram in OFF-state, c carrier

concentration in ON-state, d energy band diagram in ON-state for both DLTFET and TJE-DLTFET
Fig. 3 Transfer
characteristics for both
DLTFET and TJE-DLTFET
current (I DS ) for TJE-DLTFET is recorded as ~10−4 A/μm and ~ 0−5 A/μm for that
of DLTFET at gate voltage V GS = 1.5 V.
The spacer width (L GS /L GD ) in both the gate/source electrode and gate/drain elec-
trode, respectively must be chosen correctly as this is the primary factor across either
junction for the tunneling of electrons. Reducing the gate/source spacer width (L GS )
increases tunneling of the majority of charge carriers at the source and channel inter-
face, optimizing drain current (I DS ) under ON-state. [12]. In the ON-state, one can
observe the electric field is maximum at the source/channel tunnel junction interface
since the barrier range is reduced across it as well as the lesser effective mass of
majority carriers. Consequently, there is an increment in tunneling efficiency; which
in turn, allows the electric field across the source and channel interface to be reached
at a peak, as shown in Fig. 4. Transconductance (gm ) is defined to be the rate at
342 A. Verma et al.
Fig. 4 The electric field

under ON-state for both and
TJE-DLTFET and DLTFET
Fig. 5 Transconductance for

both DLTEFT and
TJE-DLTFET
which gate voltage can be converted into drain current. TJE-DLTFET showed better
sensitivity compared to DLTFET for the conversion of gate voltage to drain current
because proposed TJE-DLTFET attains the maximum peak at lower gate voltage
(V GS ) which starts decreasing at greater gate voltage (V GS ) due to degradation of
mobility as depicted in Fig. 5.
3.2 Analog/RF Parameter Analysis
From the figure, it can be inferred that the proposed device has somehow the same
parasitic capacitance as that of the DLTFET. The improvement in transconductance
(gm ) has concluded the significant increment in the cutoff frequency (f T ) for the TJE-
DLTFET compared to DLTFET because gate-to-drain capacitance (C gd ) as shown in
Fig. 6. It shows a minimal increment in the Cgd value of TJE-DLTFET compared to
DLTFET, therefore, the cutoff frequency of proposed TJE-DLTFET mainly depen-
dent on the variation of transconductance (gm ) rather than the gate-to-drain capac-
itance (C gd ) as depicted in Fig. 7. The formula for the calculation of the cutoff
frequency is as:
gm
fT = (1)
2π Cgd + Cgs
However, Maximum oscillation frequency (f max ) is explained as the frequency

where unit value for power gain. It can be formulated as:
Fig. 6 Gate to drain

capacitance plot
Fig. 7 Cutoff frequency plot
Fig. 8 Maximum oscillation

frequency plot

fT
f max = (2)
8πCgd Rgd
Gain Bandwidth Product (GBP) variation concerning gate voltage (V GS ) for both
TJE-DLTFET and DLTFET is depicted in Fig. 9. It shows that the increment in gate
Fig. 9 Gain-bandwidth
product (GBP) plot for both
DLTFET and TJE-DLTFET
344 A. Verma et al.
Table 3 Performance
Parameters DLTFET TJE-DLTFET
comparison for both the
devices I ON (A/μm) ~10−5 ~10−4
I OFF (A/μm) ~10−19 ~10−19
I ON /I OFF 1014 1015
SSavg (mV/decade) 12.06 10.25
V T (V) 0.55 0.40
voltage (V GS ) result into initial increment in Gain Bandwidth Product (GBP) due to
increase in the transconductance (gm ), but it starts decreasing for a higher value of gate
voltage (V GS ) because of increment in gate-to-drain capacitance (C gd ). Therefore,
TJE-DLTFET shows a higher GBP compared with DLTFET. The maximum value of
GBP in TJE-DLTFET can approach to 29.17 GHz, which is higher as compared to
18.40 GHz in DLTFET. Thus, for high-frequency applications, TJE-DLTFET gives
better performance.
gm
GBP = (3)
20πCgd
As depicted in Fig. 7, the maximum value of fT for TJE-DLTFET is obtained at

26 GHz, which is higher than that for the DLTFET (of 17 GHz) obtained at the bias
of V ds = 0.5 V. Furthermore, Fig. 8 illustrates that the TJE-DLTFET exhibits a peak
value of maximum oscillation frequency f max = 8.01 GHz, and 2.68 GHz for the
DLTFET for the bias of V ds = 0.5 V (Table 3).
4 Conclusion
In this paper, DLTFET and TJE-DLTFET are explored by using the TCAD Silvaco
simulation tool. In this regard, TJE-DLTFET (proposed device) has shown that the
use of GaAs material through the source and channel tunnel junction, it facilitates
the dropping of the energy band at the source and channel interface that also allows
a strong tunneling efficiency of charge carriers due to high mobility and an effective
mass of majority charge carriers. Thus, the enhanced drain current (I ds ), further
showed the significant increment in the transconductance (gm ), which indicate the
better current driving capability of the TJE-DLTFET compared to DLTFET. The
result of TJE-DLTFET shows that the peak value of cutoff frequency (f T ) is 26 GHz,
maximum oscillation frequency (f max ) is 8.01 GHz, and GBP is 29.17 GHz at lower
gate voltage (V GS ), which offers the proposed TJE-DLTFET beneficial for high-
frequency application. The threshold voltage (V T ) of TJE-DLTFET is 0.4 V which is
lesser than compared to 0.55 V of DLTFET. Therefore, the proposed TJE-DLTFET
opens the door for embedded ultra-low power device applications.
Acknowledgements The actual work is favored by the SMDP-C2SD under the specific reference
letter No.9 (1)2014-MDD (NIT Delhi, Delhi, INDIA).
References
1. Choi WY, Park B-G, Lee JD, Liu T-JK (2007) Tunneling field-effect transistors (TFETs) with
the subthreshold swing (SS) less than 60 mV/dec. IEEE Electron Device Lett 28(8):743–745
2. Wang PF, Hilsenbeck K, Nirschl T, Oswald M, Stepper C, Weis M, Schmitt-Landsiedel D,
Hansch W (2004) Complementary tunneling transistor for low power application. Sold-State
Electron 48(12):2281–2286
3. Bhuwalka KK, Schulze J, Eisele I (2005) Scaling the vertical tunnel FET with tunnel bandgap
modulation and gate work function engineering. IEEE Trans Electron Devices 52(5):909–917
4. Verhulst AS, Vandenberghe WG, Maex K, De Gendt S, Heyns MM, Groeseneken G (2008)
Complementary silicon-based heterostructure tunnel-FETs with high tunnel rates. IEEE
5. Saurabh S, Kumar MJ (2009) Impact of strain on drain current and a threshold voltage of
nanoscale double-gate tunnel field effect transistor. Jpn J Appl Phys 48(6):064503
6. Chiang M-H, Lin J-N, Kim K, Chuang C-T (2007) Random dopant fluctuation in limited-width
FinFET technologies. IEEE Trans Electron Devices 54(8):2055–2060
7. Asenov A, Brown A, Davies J, Kaya S, Slavcheva G (2003) Simulation of intrinsic parameter
fluctuation in decananometer and nanometer-scale MOSFETs. IEEE Trans Electron Devices
50(9):1837–1852
8. Shin C, Sun X, King Liu T-J (2009) Study of random dopant-fluctuation (RDF) effects for the
tri-gate bulk MOSFET. IEEE Trans Electron Devices 56(7):1538–1542
9. Jhaveri R, Nagavarapu V, Woo JCS (2011) Effect of pocket doping and annealing schemes on
the source-pocket tunnel field-effect transistor. IEEE Trans Electron Devices 58(1):80–86
10. Royer CL, Mayer F (2009) Exhaustive experimental study of tunnel field effect transis-
tors (TFETs): from materials to architecture. In: 10th international conference on ultimate
integration of silicon, pp 53–56 (2009)
11. Leonelli D, Vandoren A, Rooyackers R, Gendt SD, Heyns MM, Groeseneken G (2010) Opti-
mization of tunnel FETs: impact of gate oxide thickness, implantation and annealing conditions.
In: Proceedings of ESSDERC, Sep 2010, pp 170–173
12. Jagadesh Kumar M, Janardhanan S (2013) Doping-less tunnel field effect transistor: design
and investigation. IEEE Trans Electron Devices 60(10):3285–3290
13. Damrongplasit N, Shin C, Kim SH, Vega RA, Liu TJK (2011) Study of random dopant fluctua-
tion effects in germanium-source tunnel FETs. IEEE Trans Electron Devices 58(10):3541–3548
14. Damrongplasit N, Kim SH, Liu TJK (2013) Study of random dopant fluctuation induced
variability in the raised-Ge-source TFET. IEEE Electron Device Lett 34(2):184–186
15. Krishnamohan T, Kim D, Nguyen CD, Jungemann C, Nishi Y, Saraswat KC (2006) High-
mobility low band-to-band-tunneling strained-germanium double-gate heterostructure FETs:
simulations. IEEE Trans Electron Devices 53(5):1000–1009
16. Yadav DS, Raad BR, Sharma D (2016) A novel gate and drain engineered charge plasma
tunnel field-effect transistor for low subthreshold swing and ambipolar nature. Superlattices
Microstruct 100:266–273
17. Shockley W, Read WT (1952) Statistics of the recombination of hole and electrons. Phys Rev
87:835–842
18. ATLAS device simulation software. Silvaco International, Santa Clara, CA (2012)
19. Saurabh S, Kumar MJ (2016) Fundamentals of tunnel field-effect transistors. CRC Press,
London
346 A. Verma et al.
20. Boucart K, Ionescu AM (2007) Double-Gate tunnel FET with high-k gate dielectric. IEEE
21. Raad BR, Tirkey S, Sharma D, Kondekar P (2017) A new design approach of dopingless tunnel
FET for enhancement of device characteristics. IEEE Trans Electron Devices 64(4):1830–1836
22. Raad BR, Sharma D, Kondekar P, Nigam K, Yadav DS (2016) Drain work function
engineered doping-Less charge plasma TFET for ambipolar suppression and RF perfor-
mance improvement: a proposal, design, and investigation. IEEE Trans Electron Devices
63(10):3950–3957
23. Li W et al (2015) Polarization-engineered III-nitride heterojunction tunnel field-effect
transistors. IEEE J Explore Solid-State Comput Devices Circ 1(1):28–34
24. Daley D, Tran H, Woods W, Zhang Z (2015) Tunnel fieldeffect transistors with a gate-swing
broken-gap heterostructure. 21 Jul 2015, US Patent 9,087,717. (Online). Available: https://
www.google.ch/patents/US9087717
The Role and Significance of Atrial ECG
Components in Standard and Modified
Lead Systems
Shaik Karimulla and J. Sivaraman
Abstract Comparative study between Standard limb lead system (SLL) and Modi-
fied limb lead system (MLL) was carried out in this paper. Electrocardiograms (ECG)
of Forty-five female subjects were recorded in sinus rhythm with SLL and MLL for
60 s with the help of EDAN SE-1010 PC ECG system. In MLL system, the ampli-
tudes of P wave increased and the amplitude of ventricular components reduced.
In PR interval atrial repolarization wave (Ta) wave was noticed and it is opposite to
direction of P wave. There were no changes observed in durations of ECG wave. The
MLL system with observable Ta wave improves the diagnosis of arrhythmias and
atrial related diseases. The developed MLL system provides significant information
related to atrial components. This study shows that MLL system is an alternate lead
system to provide detailed information about atrial components of ECG.
Keywords Standard 12-lead system · Modified limb lead system · Atrial

repolarization wave
1 Introduction
The positioning of Standard Limb Lead (SLL) was first introduced by Willem
Einthoven [1], later positioning of precordial leads was introduced by Wilson et al.
[2]. Limb leads are concerned with motion artifacts which affects the amplitudes of
ECG. Placement of limb leads is not accessible in exercise and stress electrocardio-
gram and this problem was solved by Mason et al. [3] by placing the limb electrodes
on torso. The placement of limb electrodes on torso affect the ECG amplitudes
in frontal plane and false positive ECG changes also observed. No changes were
S. Karimulla · J. Sivaraman (B)

Bio-Signals and Medical Instrumentation Laboratory, Department of Biotechnology and Medical
Engineering, National Institute of Technology Rourkela, Rourkela, Odisha 769008, India
S. Karimulla
348 S. Karimulla and J. Sivaraman
observed in transverse plane and precordial leads [4] and many other lead systems
were developed to record the atrial ECG components [5, 6].
Generally, Standard 12-lead system is used to record the electrical activity of heart
and the placement of electrodes gives the view of all electrical axis of anatomical
space of ventricles. It provides more information about ventricular components (i.e.)
ventricular depolarization and ventricular repolarization and less information about
atrial depolarization and repolarization. It is more important to study about atrial
components and it is difficult to observe the activity of atrium in case of atrial arrhyth-
mias. Study of entire cardiac cycle is essential for electro cardiologists to observe the
electrical activity of entire heart and focusing more on ventricular components. The
study related to ventricular disorders, arrhythmias and related diseases were studied
widely but the disorders related to atria are not focused till recently.
The atrial depolarization wave (P) observed in standard 12-lead system is having
less amplitude compared to ventricular depolarization wave (QRS complex) and
is having major role in study of clinical information related to atrial arrhythmias.
Furthermore, the standard 12-lead system is not able to provide information about
atrial repolarization wave (Ta) because it is having very low amplitude and it coincides
with the QRS complex as the activity of QRS complex starts before the ending of
atrial repolarization [7]. The Ta wave was first observed in AV block patients and the
observed Ta wave starts after end of P wave and it is opposite in direction to P wave.
Ta wave is located in PR interval and is having duration nearly three times to P wave
[8]. The Ta waves differ with different AV block patients. Standard 12-lead systems
are not able to record Ta wave in normal sinus rhythm subjects (NSR) and only
few studies describes the study of Ta wave in sinus rhythm subjects with different
lead techniques. The study related to P and Ta wave improves clinical diagnosis
related to atrial diseases. To observe the Ta wave in NSR subjects, Modified limb
lead system (MLL) is developed by changing the positions of limb lead electrodes of
SLL system. Recently Sivaraman et al., suggested MLL system describes the clinical
standards of MLL system and studied about dynamics of atrial repolarization wave
(Ta) in normal sinus rhythm (NSR) and atrial tachycardia subjects [9]. It is also noted
atrial repolarization affected with that sinus rhythm and heart rate variability. The
main goal of the present study is to explain the importance of MLL system in ECG
diagnosis and on understanding the atrial ECG components. The lead placement in
MLL system is placed in such a way where atrial repolarization wave is seen clearly
which is absent in SLL system. MLL system improves in analyzing and diagnosis of
atrial related diseases whereas SLL provides more information related to ventricular
diseases.
The Role and Significance of Atrial ECG Components … 349
Table 1 Subject details

Age details
Total subjects 45 (24 ± 4.57) (21.0, 25.0, 28.0)
All values are expressed in n (mean ± SD), (minimum, median
and maximum)
2 Methods
2.1 Subjects
The experimental study included 45 healthy female subjects. The mean age of the
subjects is 24 ± 4.57 years. There was a prior written consent taken from the volun-
teers before undergoing the study. None of the volunteers had any prior or ongoing
medical issues (Table 1).
2.2 Data Collection
In MLL system the position of limb electrodes is changed and the position of chest
electrodes remains same. The placement of electrodes as follows: Right arm (RA)
electrode is placed on the third right intercostal space near the Left side of midclav-
icular line, left arm (LA) electrode is placed on fifth right intercostal space near the
right side of midclavicular line, Right leg (RL) electrode is preferred for reference
electrode is placed is same position like SLL system, left leg electrode is placed on
fifth right intercostal space in midclavicular line [10].
The recording of the ECG signal was done with EDAN SE—1010 PC ECG
instrument using surface Ag–AgCl electrodes. The subjects were in supine position
and the duration of recording was 60 s for each subject with frequency response of
0.05–150 Hz. The recorded ECG signal was processed using EDAN software, and
noise filtering was done using 25 Hz filter. The recording of ECG consists two-step
process initially subjects were requested for to take 10 min rest after that ECG was
recorded with standard limb lead system and by changing the position of limb lead
electrodes ECG were recorded with modified limb lead system.
2.3 Statistical Analysis
The recorded data is mentioned as mean ± standard deviation. For correlation anal-
ysis between ECG values Pearson correlation coefficient is considered. The data
values are statistically significant with P < 0.05 for amplitudes and P > 0.05 for time
Table 2 Changes in amplitudes of ECG components in SLL and MLL ECG’s

ECG wave Standard limb lead ECG Modified limb lead ECG P value
(µV) (µV)
Mean SD Mean SD
P wave 141.47 22.24 84.84 25.41 <0.05
R wave 947 346.99 512.81 195.83 <0.05
T wave 266.55 99.10 134.39 58.30 <0.05
Table 3 No changes in temporal components of ECG with SLL and MLL ECG’s
ECG intervals Standard limb lead ECG Modified limb lead ECG P value
(ms) (ms)
Mean SD Mean SD
PR interval 141.47 22.24 142.15 21.33 >0.05
QRS interval 90.26 11.46 88.42 11.94 >0.05
QT interval 357.97 24.74 358.63 25.20 >0.05
durations of ECG components. All the statistical analysis was done using Origin
Pro—8 of Origin Lab Corporations, 2016 version (Tables 2 and 3).
3 Results
The box and whisker plot shown in Fig. 1 describes the analysis of P and R wave
in Standard Limb Lead system (SLL) and Modified Limb Lead system (MLL).
Figure 1a) shows the P wave amplitude in SLL and MLL system. The minimum P
a) b)
Fig. 1 Analysis of P and R wave in SLL and MLL. a Box plot of P wave amplitude in SLL and
MLL system. b Box plot of R wave amplitude in SLL and MLL system
Fig. 2 Box plot of P and Ta

wave amplitudes in MLL
wave amplitude value in SLL system is 40 µV, mean value is 70 µV and maximum
value in SLL system is 130 µV. The minimum, mean and maximum value in MLL
system for P wave amplitude is 30 µV, 77 µV, and 140 µV respectively. Figure 1b
shows the R wave amplitude in SLL and MLL system. The minimum R wave ampli-
tude in SLL and MLL system is 200 µV and 190 µV respectively. The mean R
wave amplitude in SLL and MLL systems is 1000 µV and 500 µV respectively. The
maximum R wave amplitude in SLL and MLL systems is 1500 µV and 1000 µV
respectively.
Figure 2 shows the box and whisker plot of P and Ta wave amplitude in MLL
system. The minimum, mean and maximum P wave amplitude in MLL system is
25 µV, 75 µV, and 149 µV respectively. The maximum Ta wave amplitude in MLL
system is −90 µV, mean Ta wave amplitude is −50 µV and minimum Ta wave
amplitude is −3 µV.
Figure 3 shows the T wave amplitude in SLL and MLL system. The minimum
T wave amplitude in SLL and MLL system is 100 µV and 50 µV respectively. The
mean R wave amplitude in SLL and MLL systems is 250 µV and 150 µV respectively.
The maximum R wave amplitude in SLL and MLL systems is 550 µV and 350 µV
respectively.
Fig. 3 Box plot of T wave

amplitude in SLL and MLL
Fig. 4 Box plot of P wave

duration in SLL and MLL
system
Figure 4 shows the P wave duration in SLL and MLL systems. The minimum,
mean and maximum value in SLL system for P wave duration is 88 ms, 97 ms and
112 ms respectively. The minimum P wave duration in MLL system is 85 ms, the
mean P wave duration in MLL system is 95 ms, and maximum P wave duration in
MLL system is 110 ms.
Figure 5 expresses the box and whisker plot of QRS duration in SLL and MLL
system. The minimum QRS duration in SLL and MLL system is 65 ms and 67 ms
respectively. The mean QRS duration in SLL and MLL systems is 90 ms and 85 ms
respectively. The maximum QRS duration in SLL and MLL systems is 115 ms and
108 ms respectively.
Figure 6a shows the box and whisker plot of PR interval duration in SLL and
MLL system. The minimum PR interval duration in SLL system is 90 ms, mean PR
interval duration is 140 ms, and maximum PR interval duration in SLL system is
180 ms. The minimum, mean, and maximum PR interval duration in MLL system
are 100 ms, 135 ms, and 177 ms respectively.
Figure 6b illustrates the box and whisker plot of QT interval duration. The
minimum QT interval duration in SLL and MLL systems is 290 ms and 320 ms
respectively. The mean QT interval duration in SLL and MLL systems is 355 ms and
Fig. 5 Box plot of QRS

interval duration in SLL and
MLL system
a) b)
Fig. 6 Analysis of PR and QT interval duration in SLL and MLL. a Box plot of PR interval duration
in SLL and MLL system. b Box plot of QT interval duration in SLL and MLL system
345 ms respectively. The maximum QT interval duration in SLL and MLL system
is 420 ms and 410 ms respectively.
Figure 7 expresses the box and whisker plot of Heart rate in SLL and MLL system.
The minimum heart rate in SLL and MLL system is 69 bpm and 65 bpm respectively.
The mean heart rate in SLL and MLL systems is 85 bpm and 82 bpm respectively. The
maximum heart rate in SLL and MLL systems is 100 bpm and 95 bpm respectively.
In general, heart rate plays major role in health of human body. It measures the
number of contractions or beats per minute. The speed of heart rate varies with
exercise, emotional responses etc. Resting heart rate denotes heart rate when body
is relaxed. In SLL system heart rate is calculated by using RR interval, i.e., Heart
Rate = 60/(RR interval), whereas heart rate in MLL system is calculated by using
PP interval and is given as Heart Rate = 60/(PP interval). Heart Rate measurement
in SLL and MLL system is same and there is no effect on heart rate with the type of
lead system.
Fig. 7 Box plot of Heart

rate (bpm) in SLL and MLL
system
4 Discussion
Petrenas et al. described that modified Lewis lead system producing atrial amplitudes
three times more than atrial amplitudes in Lewis lead system [10]. Kennedy et al.
described P leads were derived from standard 12-lead systems by using transforma-
tions. P lead increases the P wave Root mean square (RMS) signal strength and AV
RMS ratio in atrial activation. The increased signal strength and AV RMS results in
accurate P wave detection and cancellation of QRST in standard electrocardiogram
[11]. Sivaraman et al. presented a work in 2012 on two different lead systems (SLL
and MLL) results in minimizing the atrial activity of ventricular QRS complex. The
amplitudes of QRS complex in ECG reduced with MLL system [12]. Atrial repo-
larization wave was observed in PR segment having more duration than P wave and
following the P wave in opposite direction [13]. Sivaraman et al. described that the
amplitude of P wave increases in MLL system as compared with SLL system and
there is axis shift in P wave also observed. The amplitude of QRS complex is reduced
with MLL system. The P wave signal averaging method producing best explanation
of P and Ta waves. The R peak algorithm gives the position of peak in P and R wave
of all leads of MLL system. The MLL system changes the amplitudes of ECG waves
and STa and also produces axis deviation in QRS complex and T wave [14].
5 Conclusion
The standard 12-lead system is not able to provide significant information about
atrial components and the atrial repolarization wave (Ta) is not observed in ECG
trace because it coincides with QRS complex in normal conditions. Ta wave was first
observed in AV block patients. Recently developed MLL provides more information
related to atrial depolarization wave and atrial repolarization (Ta) wave. Ta wave
is clearly observed in MLL system which is opposite in direction to P wave and
it improves the diagnosis related to atrial arrhythmias and related diseases. The Ta
wave is useful in the study of dynamics of atrial activity change.
Acknowledgements The authors acknowledge the support from MHRD, Government of India, for
sponsoring the M.Tech. program of the first author. The present study was supported by financial
grants from Science Engineering Research Board (SERB), Department of Science and Technology,
Government of India (EEQ/2019/000148).
References
1. Einthoven W (1912) The different forms of the human electrocardiogram and their signification.
Lancet 179(4622):853–861
2. Wilson FN, Johnston FD, Macleod AG, Barker PS (1934) Electrocardiograms that represent
the potential variations of a single electrode. Am Heart J 9(4):447–458
3. Mason RE, Likar I (1966) A new system of multiple-lead exercise electrocardiography. Am
Heart J 71(2):196–205
4. Rautaharju PM, Prineas RJ, Crow RS, Seale D, Furberg C (1980) The effect of modified limb
electrode positions on electrocardiographic wave amplitudes. J Electrocardiol 13(2):109–113
5. Takuma K, Hori S, Sasaki J, Shinozawa Y, Yoshikawa T, Handa S, Horikawa M, Aikawa N
(1995) An alternative limb lead system for electrocardiographs in emergency patients. Am J
Emerg Med 13(5):514–517
6. Szekely P (1944) Chest leads for the demonstration of auricular activity. Br Heart J 6:238–246
7. Sivaraman J, John R (2017) Effects of sinus rhythm on atrial ECG components using modified
limb lead system. In: 4th international conference on signal processing, computing and control
(ISPCC), pp 527–530 (2017)
8. Sprague HB, White PD (1925) Clinical observations on the T wave of the auricle appearing in
the human electrocardiogram. J Clin Invest 1:389–402
9. Sivaraman J, Uma G, Venkatesan S, Umapathy M, Dhandapani VE (2015) Unmasking of atrial
repolarization waves using a simple modified limb lead system. Anatol J Cardiol 15(1):2–6
10. Petrenas A, Marozas V, Jarusevicius G, Sornmo L (2015) A modified Lewis ECG lead system
for ambulatory monitoring of atrial arrhythmias. J Electrocardiol 48(2):157–163
11. Kennedy A, Finlay DD, Guldenring D, Bond RR, McEneaney DJ, Peace A, McLaughlin J
(2015) Improved recording of atrial activity by modified bipolar leads derived from the 12-lead
electrocardiogram. J Electrocardiol 48(6):1017–1021
12. Sivaraman J, Uma G, Umapathy M (2012) A modified chest leads for minimization of
ventricular activity in electrocardiograms. In: 2012 international conference on biomedical
engineering (ICoBE) 2012, pp 79–82
13. Sivaraman J, Gandhi U, Sangareddi V, Mangalanathan U, Shanmugam RM (2015) Unmasking
of atrial repolarization waves using a simple modified limb lead system. Anatol J Cardiol
15(8):605–610
14. Sivaraman J, Uma G, Venkatesan S, Umapathy M, Kumar NK (2014) A study on atrial ta wave
morphology in healthy subjects: An approach using P wave signal-averaging method. J Med
Imaging Heal Inf 4(5):675–680
Data Rate Enhancement Scheme
and Mathematical Modelling for Data
Communication Through Blood
Kauser Husainee and Tanmay Deshmukh
Abstract The data communication for medical purposes such as nano-surgical

robots is being established by applying information theory on the kinetic, drift and
diffusive molecular flow of blood. The modelling of blood flow can be used further
for multiple scientific and medical research purposes. Modelling and simulation of
blood flow is analysed in COMSOL Multiphysics tool using laminar flow model.
Efficient scheme is developed for data rate enhancement in the field of molecular
communication. Propagation delay is estimated using simulation results for commu-
nication establishment. Various differential equations are solved by modelling the
blood pressure and velocity of blood.
Keywords Mathematical modelling · Molecular communication · COMSOL

Multiphysics tool · Blood flow · Data rate enhancement
1 Introduction
In order to establish data communication for medical applications through nano-

surgical robots placed inside the blood vessels of the human body using molecular
fluid dynamics theory, a part of the cardiovascular system has to be studied, math-
ematically modelled, graphically plotted and simulated. A cardiovascular system
implies the network of blood distribution in the human body. It consists of the heart,
the blood and the blood vessels. The term blood pressure can be described as the
pressure detected on the walls of the blood vessels when blood flows through them.
It depends on the rate of blood flow, pressure gradient and vessels’ size [1]. The types
of blood vessels in the human body are the arteries, the arterioles and the veins. The
arteries are relatively larger blood vessels which carry pure and oxygen-rich blood at
high pressure to the body from the heart. Veins are the low-pressure blood collecting
system which returns impure blood to the heart [2, 3]. The arterioles are the vessels
which get divided into smaller vessels (to supply oxygen and nutrients to all small
K. Husainee (B) · T. Deshmukh

Department of Electronics and Communication Engineering, Visvesvaraya National Institute of
Technology, Nagpur, India
358 K. Husainee and T. Deshmukh
Fig. 1 Concept diagram
tissues). They are the constitutional units for the connection of arterial and venous
systems with each other [1].
The approach is to inject a biodegradable solute into the blood stream from the
transmitter side, and the volume rate of the solvent along with the time period for
which it is transmitted is detected at the receiver side, through the sensors on the
nano-surgical robots, which determine the analog value of the data sent as shown in
Fig. 1. The data can be any symbol which can be exclusively designed for medical
science purposes and not necessarily the characters that we use in keyboards or
everyday life [4].
2 Data Rate Enhancement
The comparison of the conventional way of establishing molecular communication

with the proposed model which uses quantization theory is done on the basis of baud
rate and bit rate.
2.1 Conventional Model
In the convention model, general scheme followed as when the number of molecules
above certain threshold is detected in certain time frame t can be taken as Logic-1 and
the number of molecules below that threshold is taken as Logic-0. The time frame
has to be decided as per the capabilities of hardware. The time frame directly gives
us the data rate in the convention scheme. Total characters needed in the scheme
are first defined, and then based on the corresponding character length, the baud
rate is determined. Let the time frame which is the time required to receive 1 bit be
t = 10 ms .
Data Rate Enhancement Scheme and Mathematical Modelling … 359
bits
i.e., bit rate = 100
sec
If the maximum number of distinct input symbols available is 512, the number of
bits per symbol is given by
bits
log2 512 = 9 .
symbol
From this, time required to send 1 symbol, i.e. time required to send 9 bits, is
9 ∗ t = 90 ms.

1 symbols symbols
Baud rate = = 11.11 .
90 msec sec
Therefore,
bits
Bit rate = 100
sec
If the total number of molecules detected at the receiver is above the threshold
value, data is received as 1, otherwise 0. Let the 9 bits 100101011 represent one of
the symbols as shown in Fig. 2.
Fig. 2 Data transfer by conventional model for one symbol Tb—bit duration and Ts—symbol
duration
2.2 Proposed Model
In our proposed model, as shown in Fig. 3, the earlier decided time frame is quantized
in q1 levels. Also, volume rate is quantized in q2 . Here, again the quantization levels
q1 and q2 depend on the hardware capability. Quantization levels in time frame q1
depend on the sensor capabilities on the reception side. The fast response of the sensor
is desired for the higher quantization levels. Similarly, the quantization levels in the
volume rate depend on the injection capability of the device. High range volume rate
injection may enable us for increase in quantization levels. So,
q
2
the overall quantization levels = q1 ∗ i = 550 available symbols.

i=1
Thus, here the maximum number of distinct input symbols available is 550. Here,
we use 512 symbols out of these 550 symbols considering there may be high corre-
lation between some symbols. Now in the same time frame, a complete symbol, i.e.
characters, is transmitted instead of a bit.
Baud rate obtained for the above scheme is given by
Fig. 3 Quantization levels in the proposed model and the possible cases of the symbols

1 symbols symbols
Baud rate = = 100
10 msec sec
The number of bits per symbol is given by
bits
log2 512 = 9
symbol
So, the bit rate obtained is 900 bits per sec.

Thus, there is an enhancement in the bit rate by 9 times.
According to the possible cases in Fig. 3, the molecule reception pattern in time
frame can be correlated on the reception side with predefined symbol patterns to
determine the symbol received.
3 Mathematical Modelling
3.1 Assumptions for the Fluid Characteristics
• Blood is contemplated as a ‘Newtonian fluid’ which has the relation that the
viscous stress is directly proportional to the strain developed [5]. Such a system
is governed by the conservation of momentum as well as mass relations which
are represented by Navier–Stokes equation and continuity equation, respectively.
• The orientation of blood vessels in space and the effect of gravity have been
neglected.
• Blood vessels are assumed to be non-deformable components cylindrical in shape.
They do not change their shape as blood flows through them.
• Properties of blood before and after addition of oxygen from lungs are considered
to be as it is.
In order to do the mathematical modelling, the characteristics of a particular

section of the cardiovascular system has to be considered and studied [6]. In this
model, we have studied the upper part of inlet and outlets of the aorta. For simulation
purposes, these inlet and outlet blood vessels are placed inside the tissues which
are particularly called as the ‘cardiac muscles’, which help to keep the vessels non-
deformable. Hence, the blood fluid characteristics inside the vessels do not change
[7, 8]. The modelling results give out the distribution of velocity of fluid inside the
aortic section which is variable in space as well as time.
3.2 Navier–Stokes and Continuity Equation
Considering the assumptions as stated above, the blood is considered to be governed

by the equations as given below
Continuity equation (conservation of mass):
∇ · V = 0 (1)
Navier–Stokes equation (conservation of momentum)
D V
ρ = −∇ P + μ∇ 2 V (2)
Dt
where P, μ, ρ, V are the blood pressure, blood viscosity, blood density and velocity
vector, respectively.
4 Simulation and Results
The function of heart beat which incorporates the functions of systole and diastole
defined for the purpose of modelling is given in Table 1. which is graphically plotted
in Fig. 4. The simulation is performed in COMSOL Multiphysics tool based on the
heart beat function defined in Table 1 to get the result as shown in Fig. 5.
Table 1 Mathematical
Start End Function
equation of heart beat
function 0 0.5 (1 − α) ∗ sin(π t)
0.5 1.5 1 − α ∗ cos(2π (t − 0.5))
Fig. 4 Function of heart beat

Fig. 5 Velocity distribution

inside the aortic section
5 Modelling Based on Simulation
Using the simulation results obtained from Fig. 5, velocity distribution inside the
aortic section, the mathematical modelling for axial and radial velocity is proposed
as follows
For axial flow velocity component w(z) in cylindrical coordinate system,
z = (a − w)2 + (a − w)4 + (a − w)6 + (a − w)8 + · · · ∞ (3)
where ‘a’ is the maximum axial velocity at the centre of the blood vessel.

z = (a − w)2 1 + (a − w)2 + (a − w)4 + (a − w)6 + (a − w)8 + · · · ∞
z = (a − w)2 (1 + z)

z
w(z) = a − (4)
z+1
For radial flow velocity component f (r) in cylindrical coordinate system,
∞
m− f
= r2 + r4 + r6 + r8 + · · · ∞ = r 2n (5)
q n=1
where ‘m’ is the maximum radial velocity at the centre of the blood vessel and ‘q’
is a parameter which depends on the cross-sectional area of the blood vessel which
varies throughout the different parts of the body.
∞

f (r ) = m − q r 2n (6)
n=1
For a particular small section of the cardiovascular system of length ‘k’, average
velocity ‘s’ in the blood vessel can be calculated as,
∫k0 w(z)dz
s= (7)
∫k0 dz
∫(a)dz = az (8)
Now, using Eq. (4),

√
k k z
∫ w(z)dz = ∫ a − √ dz (9)
0 0 z+1
solving,

ln z
+1
z z+1
∫ dz = −
z+1 2

1 1 ln z
z+1
−1
− − + (10)
2 z
+ 1 2 z
− 1 2
z+1 z+1
Using Eqs. (8) and (10), we get

√ √ √ √
z ln z + 1 + z − ln z + 1 − z
∫ a− dz =
z+1 2

+ z(z + 1) + az + c
Thus, Eq. (9) becomes

√
k k z
∫ w(z)dz = ∫ a − √ dz
0 0 z+1
√ √
ln k(k + 1) + k − ln k(k + 1) − k
= − k(k + 1) + ak (11)
2
Substituting Eqs. (8) and (11) in Eq. (7), the average velocity ‘s’ in the vessel is
given by
√ √
ln k(k + 1) + k − ln k(k + 1) − k k+1
s= − +a (12)
2k k
So, the time delay between the transmission and reception ends required for
synchronization in order to establish communication successfully is given by T,
using Eq. (12)
k k
T = = √ √ √
s ln( k(k+1)+k )−ln( k(k+1)−k )
2k
− k(k+1)
k
+ ak
k
2
k
T = √ √ √ (13)
ln( k(k+1)+k )−ln( k(k+1)−k )
2
− k(k + 1) + ak
6 Conclusion
For the establishment of the communication, synchronization of data between trans-

mitter and receptor is essential. In the mathematical model which represents the
aorta in human cardiovascular system, the velocity distribution of axial and radial
components is obtained using the simulations performed in COMSOL Multiphysics
tool by solving Navier–Stokes equation and continuity equation using finite element
method. From the velocity distribution expression, the estimation of propagation
time from injection point to the reception sensor is derived. Using this propagation
time, synchronization and thus communication can be established successfully. The
proposed model for the communication gives the nine times enhancement in the data
rate.
References
1. Ehrlich A, Schroeder CL (2004) Medical terminology for health professions, 5th edn, Thomson
Delmar Learning, pp 131–132
2. Katila T, Magnin I, Clarysse P, Montagnat J, Nenonen J (2001) (eds) Functional imaging and
modeling of the heart (FIMH’OI). In: Lecture notes in computer science, vol 2230. Springer,
Berlin
3. Magnin I, Montagnat J, Nenonen J, Katila T (2003) (eds) Functional imaging and modeling of
the heart (FIMH’03). In: Lecture notes in computer science, vol 2674. Springer, Berlin
4. Taylor CA, Draney MT (2004) Experimental and computational methods in cardiovascular fluid
mechanics. Annu Rev Fluid Mech 36:197–231
5. Sankar DS, Hemalatha K (2007) A non-Newtonian fluid flow model for blood flow through a
catheterized artery—steady flow. Appl Math Modell 31(9):1847–1864
6. Jayaraman G, Tewari K (1995) Flow in catheterized curved artery. Med Biol Eng Comput
33(5):720–724
7. Sherwood L (2005) Fundamentals of physiology: a human perspective, 3rd edn. Thomson
Brooks/Cole, pp 276
8. Yang BH, Asada HH, Zhang Y (1999) Cuff less continuous monitoring of blood pressure using
hemodynamic model, the home automation and healthcare consortium” progress report no 2–3
9. Sultanov RA, Guster D (2009) Full dimensional compute simulations to study pulsatile blood
flow in vessels, aortic arch and bifurcated veins: investigation of blood viscosity and turbulent
effects. In: 31st annual international conference of the IEEE EMBS Minneapolis, Minnesota,
USA, 2–6 Sep 2009
A Novel Design of Fast and Low Power
Pull-Up and Pull-Down Voltage Level
Shifter
Vickey Kumar, Suruchi Sharma, and Baljit Kaur
Abstract This article presents a low power consumption and fast voltage level
shifter, which can increase the input voltage level lower or comparable to a threshold
voltage to nominal voltage level and high voltage level to low voltage level. The
proposed level shifter is based on a pull-up network which has regulated cross-
coupled structure. This helps to enhance the switching speed and reduction of
dynamic power. Furthermore, split input inverter and pMOS diode used to reduce
the static power consumption. For proposed level shifter simulation results have
performed in a 180 nm CMOS technology using cadence virtuoso tool. While
converting supply voltage from low to the high voltage level of 0.4 V and 1.8 V,
respectively at applied frequency 1 MHz, power consumption and propagation delays
are 139.6 nW and 19.92 ns, respectively. And during conversion high to the low supply
voltage are 1.8 V and 0.4 V, 120.2 nW and 15.31 ns, respectively.
Keywords Dual-supply · High speed · Low power · pMOS diodes · Up/down

level shifter
1 Introduction
Power dissipation has become the most critical concern nowadays in integrated
circuits (IC). Due to technology scaling, IC performance has increased but at a
result of high power dissipation. One of the efficient ways of power reduction is the
voltage scaling method [1–3]. By this dynamic power and static power consumption
reduced. But the speed of the circuit also reduced. So where we need high-speed,
we can’t reduce supply voltage. Conventionally, limited power is reduced by supply
voltage scaling well in the super threshold region, but in recent studies, it has been
found that supply input voltage can be reduced near below than threshold voltage [4–
13]. Different modules like digital, analog, passive circuits are fabricated on a single
chip. And different modules have different requirements. For some modules, power
V. Kumar · S. Sharma (B) · B. Kaur

Delhi, Delhi, India
368 V. Kumar et al.
consumption matters more as compared to the speed and vice versa. So two or more
supply voltages were used for the efficient performance of SoC [2, 3]. Therefore, we
need voltage level shifting between the two modules of SoC. For this purpose voltage
level shifter is used which can revise low voltage comparable to the threshold voltage
into high nominal voltage and vice a versa for the efficient performance of the SoC.
There might be the use of many level shifters (LS) in SoC. Therefore, power demand,
propagation delay and circuit area become the primary concern while designing the
level shifter (LS). Hence, in short, power-efficient and fast level shifter is proposed
which can increase low voltage level to high voltage level and vice versa, which is
controlled by the applied voltage.
Further, this paper proceeds as follows: in Sect. 2, conventional LS circuits
reviewed, Sect. 3 introduced the proposed LS and Sect. 4 shows the simulation
results. In the end, in Sect. 5, conclusions are presented in brief.
2 Literature Review
Figure 1a represents the conventional-level shifter, which uses a pull-up network

based on the current mirror structure. As it does not have regenerative interaction
between a pull-down part and pull-up part in both sides of branches of the level shifter.
Therefore, the operation speed of the circuit is very low. Moreover, because of the
flow of static current in one of the circuit branches, it consumes high standby power
according to the input state. Figure 1b shows level shifter having a pull-up network
based on a cross-coupled structure. This pull-up part increases the difference between
Q1 and Q2 . Due to this regeneration process, operation speed increases. When input
rises to high MN1 becomes on whereas MN2 turned off. So Q1 begins to discharge,
(a) (b)
Fig. 1 Two different types of conventional LS architecture based on a current mirror b cross-
coupled structure
A Novel Design of Fast and Low Power … 369
which tries to switch on the MP4 , so Q2 starts to reach V ddh , which turns off MP3 . It
results in a fast discharging of Q1. If the supply voltage range is of sub-threshold or
less than that, then the contention problem comes into existence between pMOS and
nMOS. To remove the contention problem size of MN1 and MN2 , we have to increase.
Therefore, the overall efficiency of the circuit is reduced. One of the approaches to
solve this problem is to reduce the drive strength of the pull-up circuit.
3 Proposed Circuit
Figure 2 represents a simplified proposed circuit schematic. This design uses a pull-
up network having a cross-coupled regulated structure. The problem of making more
size of NMOS in the pull-down network is resolved here by using MP5 and MP6 . By
using these two PMOS transistors strength of the pull-up network is reduced. This
charging and discharging of pull-up critical nodes becomes fast, which enhances the
speed of switching and decreases dynamic power. Figure 3 explains the working of
a simplified proposed level shifter in case of low to high voltage level conversion.
Initially, the voltage at a node Q1 is V H, which is slightly less than V ddh and voltage
at Q2 is low. Therefore, MP6 and MP3 are on, and MP4 and MP5 are off. When there
is the transition of input from low level to a high level then MN1 becomes switched
on and MN2 switched off so Q1 begins to discharge very fast because of the current
flowing through MP3 is very small, because of MP5 and Q1 discharge approximately
from V H to V ddh − V H at this time MP4 and MP5 start to turn on.
So consequently, Q2 begins to charge, and voltage at Q2 increases. This flow
causes MP3 and MP6 to start to turn off. This regenerative time process of pull-up
network increase voltage at Q2 very fast and caused MP6 to turn OFF before Q2
reaches V ddh and Q2 fixed at V H . At the last step, MP3 will be turned off because
of Q2, and MN1 would be able to discharge node Q1 . The left and right branches do
not contain any flow of static current due to switched OFF MP5 and MP6 . Moreover,
interval nodes voltage remains at V H always, which are lower than V ddh , so dynamic
Fig. 2 Simplified proposed

LS
370 V. Kumar et al.
Fig. 3 Operation of a simplified schematic of proposed LS [14]
power also reduced. This flow will also be similar for high to low transition of input
voltage exact reverse of it (Fig. 4).
For the complete schematic circuit of voltage level shifter, which we have proposed,
the transient waveforms are as depicted in Figs. 5 and 6, respectively. This is done
by keeping V ddl = 0.4 V and V ddh = 1.8 V. In the proposed LS circuit, two PMOS
diodes body connected to V ddh are used. So that pMOS and nMOS of inverter do
not switch on at the same time which results in small short circuit current and so
power dissipation of LS circuit reduced. The transistors sizes are reported in Table 1.
Figure 7 shows the effect of PVT variation on the overall power consumed and delay
of the proposed LS structure for various values of V ddl keeping supply frequency
1 MHz and V ddh value 1.8 V. A fast pMOS and fast nMOS give the least delay
contrarily slow pMOS and slow nMOS results the worst case. While Fig. 8 shows
variation values concerning V ddh at frequency 1 MHz and V ddl value 0.4 V. Figures 9
Fig. 4 Schematic of a complete proposed LS
Fig. 5 Proposed level shifter simulated waveforms in case of low input voltage level
and 10 represents power and delay characteristics of LS at the typical corner for up
and down voltage level conversion, respectively, for an applied frequency of 1 MHz.
Table 2 shows consumed power and delay, where Ps denotes the static power and
Pt total power, in case of up voltage level, having input voltage value 0.4 V, V ddl
= 0.4 V, and V ddh = 1.8 V at frequency 1 MHz at 27 °C typical corner. Moreover,
Table 3 shows performance for converting voltage level down, having input voltage
1.8 V.
372 V. Kumar et al.
Fig. 6 Proposed level shifter simulated waveforms in case of high input voltage level
Table 1 Size of transistors

Transistors W/L (in µm) Transistor W/L (in µm)
MN1 , MN2 , MN3 , MN4 , MP1 , MP2 , MP3 , MP4 , MP5 , 0.4/0.18 MP7 0.8/0.18
MP6 , MP8
(a) (b)
Fig. 7 Proposed LS PVT analysis a delay b power versus V ddl for V ddh = 1.8 V by keeping input
supply voltage of 0.4 V at the frequency of 1 MHz for low to high-level shifting time
5 Conclusion
A novel up/down LS is proposed, which is fast and consumes low power. The
proposed LS utilizes a pull-up network, which is based on the regulated cross-coupled
circuit. Furthermore, it can convert supply input from a low threshold level to a high
nominal voltages level, and it can also convert the supply voltage of high-level to
(a) (b)
Fig. 8 Proposed LS PVT analysis a delay b power versus V ddh for V ddl = 0.4 V by keeping input
supply voltage of 18 V at a frequency of 1 MHz for high to low-level shifting time
(a) (b)
Fig. 9 Proposed LS characteristics a Delay b power versus V ddl for V ddh = 1.8 V and frequency
of 1 MHz at the typical corner
low. Performed simulation results of the proposed up and down voltage level shifter
using 180 nm CMOS technology indicates high performance and supports a wide
range voltage level conversion. It can provide an up-conversion for the input signal
as low as 80 mV and the input signal of high voltage level to about 0.4 V. Further-
more, corner analysis of the proposed circuit is performed to validate the superior
performance over the conventional-level shifter.
374 V. Kumar et al.
(a) (b)
Fig. 10 Proposed LS characteristics a delay b power versus V ddh for V ddl = 0.4 V and frequency
of 1 MHz at a typical corner
Table 2 Comparison with previous work (V ddh = 1.8 V, F = 1 MHz)

References Min V ddl Delay (ns) Ps (nW) Pt (nW) PDP (nW ns)
Figure 5 0.28 19.92 0.36 139.6 2781
[13] 0.31 31.29 0.23 326.7 10,222
[15] 0.28 31.54 0.27 166.3 5245
[16] 0.29 29.85 0.06 139.3 4158
[17] 0.28 23.75 1 183.9 4368
[18] 0.3 29.59 0.78 631.4 18,683
[19] 0.3 28.83 0.23 218.3 6293
[20] 0.27 21.49 0.26 418.7 8998
[21] 0.29 25.23 1.76 137 3456
[22] 0.28 22.32 0.5 347 7745
[23] 0.38 41.12 0.13 228.1 9379
[24] 0.28 24.31 1.06 150 3646
Table 3 Performance
Delay (ns) Ps (nW) Pt (nW) PDP (nW ns)
parameters of the proposed
LS 15.31 0.17 120.2 1840.26
References
1. Wang A, Chandrakasan AP (2005) A 180-mV subthreshold FFT processor using a minimum

energy design methodology. IEEE J Solid-State Circuits 40(1):310–319
2. Usami K et al (1998) Automated low-power technique exploiting multiple supply voltages
applied to a media processor. IEEE J Solid-State Circ 33(3):463–472
3. Zhang D, Bhide A, Alvandpour A (2012) A 53 nW 9.1-ENOB 1-kS/s SAR ADC in 0.13-µm

CMOS for medical implant devices. IEEE J Solid-State Circ 47(7):1585–1593
4. Wang A, Chandrakasan A (2005) A 180-mV subthreshold FFT processor using a minimum
energy design methodology. IEEE J Solid State Circ 40(1):310–319
5. Raychowdhury A, Paul BC, Bhunia S, Roy K (2005) Computing with subthreshold leakage:
device/circuit/architecture co-design for ultralow-power subthreshold operation. IEEE Trans
Very Large Scale Integr (VLSI) Syst 13(11):1213–1224
6. Markovic D, Wang CC, Alarcon LP, Liu TT, Rabaey JM (2010) Ultralow-power design in the
near-threshold region. Proc IEEE 98(2):237–252
7. Seok M, Jeon D, Chakrabarti C, Blaauw D, Sylvester D (2011) A 0.27 V 30 MHz 17.7
nJ/transform 1024-pt complex FFT core with super pipelining. In: Proceedings IEEE inter-
national solid-state circuits conference digest of technical papers (ISSCC), 20–24 Feb 2011,
pp 342–344
8. Alioto M (2012) Ultra-low power VLSI circuit design demystified and explained: A tutorial.
IEEE Trans Circ Syst I Regul Pap 59(1):3–29
9. Jeon D, Seok M, Zhang Z, Blaauw D, Sylvester D (2012) Design methodology for voltage-
overscaled ultra-low-power systems. IEEE Trans Circ Syst II Exp Briefs 59(12):952–956
10. Klinefelter A, Zhang Y, Otis B, Calhoun BH (2012) A programmable 34 nW/channel sub-
threshold signal band power extractor on a body sensor node SoC. IEEE Trans Circ Syst II Exp
Briefs 59(12):941
11. Jeon D, Kim Y, Lee I, Zhang Z, Blaauw D, Sylvester D (2013) A 470 mV 2.7 mW feature
extraction-accelerator for micro-autonomous vehicle navigation in 28 nm CMOS. In: 2013
IEEE international solid-state circuits conference digest of technical papers (ISSCC), 17–21
Feb 2013, pp 166–167
12. Craig K, Shakhsheer Y, Arrabi S, Khanna S, Lach J, Calhoun BH (2014) A 32 b 90 nm processor
implementing panoptic DVS achieving energy efficient operation from sub-threshold to high
performance. IEEE J Solid-State Circ 49(2):545–552
13. Hosseini SR, Saberi M, Lotfi R (2014) A low-power subthreshold to above-threshold voltage
level shifter. IEEE Trans Circ Syst II Express Briefs 61(10):753–757
14. Kabirpour S, Jalali M (2019) A low-power and high-speed voltage level shifter based on a
regulated cross-coupled pull-up network. IEEE Trans Circ Syst II Express Briefs 66(6):909–913
15. Lanuzza M et al (2017) An ultralow-voltage energy-efficient level shifter. IEEE Trans Circ
Syst II Express Briefs 64(1):61–65
16. Maghsoudloo E, Rezaei M, Sawan M, Gosselin B (2017) A high-speed and ultra low-power
subthreshold signal level shifter. IEEE Trans Circ Syst I Regul Pap 64(5):1164–1172
17. Hosseini SR, Saberi M, Lotfi R (2017) A high-speed and power-efficient voltage level shifter for
dual-supply applications. IEEE Trans Very Large Scale Integr (VLSI) Syst 25(3):1154–1158
18. Luo SC, Huang CJ, Chu YH (2014) A wide-range level shifter using a modified Wilson current
mirror hybrid buffer. IEEE Trans Circ Syst I Regul Pap 61(6):1656–1665
19. Lanuzza M, Corsonello P, Perri S (2015) Fast and wide range voltage conversion in multisupply
voltage designs. IEEE Trans Very Large Scale Integr (VLSI) Syst 23(2):388–391
20. Huang CR, Chiou LY (2014) A limited-contention cross-coupled level shifter for energy-
efficient subthreshold-to-superthreshold voltage conversion. In: International SoC design
conference (ISOCC), 2014, pp 142–143
21. Lotfi R, Saberi M, Hosseini SR, Ahmadi-Mehr AR, Staszewski RB (2018) Energy-efficient
wide-range voltage level shifters reaching 4.2 fJ/transition. IEEE Solid-State Circ Lett 1(2):34–
37
22. Osaki Y, Hirose T, Kuroki N, Numa M (2012) A low-power level shifter with logic
error correction for extremely low-voltage digital CMOS LSIs. IEEE J Solid-State Circuits
47(7):1776–1783
23. Lanuzza M, Corsonello P, Perri S (2012) Low-power level shifter for multi-supply voltage
designs. IEEE Trans Circ Syst II Express Briefs 59(12):922–926
24. Wen L, Cheng X, Tian S, Wen H, Zeng X (2015) Subthreshold level shifter with self-controlled
current limiter by detecting output error. IEEE Trans Circ Syst II Express Briefs 63(4):346–350
Modeling of Dual-Gate Carbon
Nanotube Based Ion Sensitive Field
Effect Transistor (DG-CNTISFET)
Hiranya Ranjan Thakur, Gaurav Keshwani, and Jiten Chandra Dutta
Abstract In this paper, modeling of a dual-gate carbon nanotube based ion sensi-
tive field effect transistor (DG-CNTISFET) fabricated by electrochemical deposition
(ECD) technique is discussed. The surface potential variation of the device was deter-
mined by using Bousse’s model. The device showed high sensitivity because of large
coupling ratio between top and bottom gate capacitances. High-κ and low-κ dielec-
tric materials have been used as the top and bottom gate dielectrics, respectively, to
achieve large coupling ratio. The high carrier mobility of CNT also contributed in
achieving this high sensitivity of the device. The developed model was simulated in
MATLAB environment. The simulated results were compared with the experimental
results of the fabricated device and good fit was obtained.
Keywords CNT · High-κ · Modeling · DG-CNTISFET
1 Introduction
Among the various nanomaterials, field effect transistors based on carbon nanotube
(CNT) have attracted lot of attention in the field of biosensors [1–3]. Some of the
excellent properties of semiconducting CNT include ballistic transport, robustness,
good thermal, mechanical, large surface area and high chemical stability [4–7]. The
use of CNT in the channel can overcome the traditional ISFETs scaling limitations [8,
9]. The unique chemical bonding and surface stability of CNT impart high compat-
ibility with high-κ dielectric material. Thus, the capacitance of CNTFET increases
without using small film thickness. Consequently, the direct tunneling leakage current
reduces [4, 5]. The drain current and on–off current ratio increases because of
H. R. Thakur (B) · G. Keshwani · J. C. Dutta

Tezpur University, Tezpur, Assam 784028, India
G. Keshwani
J. C. Dutta
378 H. R. Thakur et al.
increased capacitance. The combination of high-κ material with biocompatible and

size compatible CNT leads to advanced miniature device [1, 4].
The single-gate CNTISFETs fabricated by using solution process showed near
Nernstian sensitivity at room temperature [10]. Some application such as measure-
ments of biological interactions or activities (e.g., DNA hybridization, protein func-
tion, cellular activity and DNA-protein interactions) requires highly sensitive device
[11, 12]. Introduction of a second gate in single-gate CNTISFET can enhance this
low sensitivity. The reference electrode and an electrolyte solution form the top gate.
A metal layer is used as the bottom gate [13]. The sensitivity of the device improves
due to the capacitive coupling. The capacitive coupling is enhanced with materials of
high dielectric constant (κ) as top gate and bottom gate with low dielectric constant
(κ).
Modeling of DG-CNTISFET helps in understanding the device physics and
explains experimental data. The parameters to achieve high sensitivity and selec-
tivity can be chosen properly [14]. The surface potential varies due to change in
solution pH. Bousse’s model was used to model this potential change. The drain
current changes as a result. The current transport model of DG-CNTISFET was used
to model the drain current change.
Thus, the modeling of DG-CNTISFET is discussed in this paper [15]. The
schematic of the fabricated device is shown in Fig. 1.
Chemical solution process was used to fabricate the device. An ITO coated glass
instead of (dimension ~5 mm × 2 mm) was used as substrate. The substrate was
cleaned by using a solution of water (H2 O), hydrogen peroxide (H2 O2 ) and ammo-
nium hydroxide (NH4 OH). A ZnO (dielectric constant ~1.5) layer (thickness ~10 nm)
was deposited on the substrate by electrochemical deposition(ECD) technique. This
layer prevents current flow from channel (CNT) to ITO. ECD was carried out by
using a three electrodes system. The three electrodes are reference electrode Ag/AgCl
(silver/silver chloride), counter electrode Pt (platinum) wire, and working electrode
ITO coated glass. 10 mg Zn(CH3 COO)2 (zinc acetate) was dissolved in distilled
water of 10 ml to prepare ZnO solution. Ammonium hydroxide (NH4 OH) of 2 ml
Fig. 1 DG-CNTISFET with

high-κ (HfO2 ) top and low-κ
(ZnO) bottom gate dielectric
[15]
Modeling of Dual-Gate Carbon Nanotube Based … 379
was then added to this solution. Gravimetric analysis technique was used to measure
the thickness of the layer (~10 nm) [9]. PEI-doped COOH-functionalized MWCNT
was deposited on this ZnO layer. PEI donates electrons to CNT as it is an amine
rich polymer. The solution was prepared as: disperse CNT (5 mg) in PEI/methanol
(5 ml). The n + S (source), D (drain), and channel regions were formed by this
layer. The contact S and D regions were deposited on the top of the PEI-doped
COOH-functionalized MWCNT layer with metal aluminum (Al) by PVD (filament
evaporation) method. Quasi-ohmic contact was formed by Fermi-level depinning
technique between CNT (work function ~4.8 eV) and Al (work function ~4.08 eV).
HfO2 (high-κ) layer of thickness ~10 nm as the sensing layer has been deposited after
metallization process on the CNT channel. HfO2 solution was prepared as: dissolve
10 mg HfCl4 (solid) in de-ionized water (5 ml). This layer was heated (~180 °C) to
form dry HfO2 layer. The thickness was measured by gravimetric analysis method.
2 Modeling
Modeling in general can be described as the representation of real-world objects

in terms of physical or mathematical equations. Modeling helps in understanding
the operation of a device in details and reduces time as well as development cost.
There are many models of ISFETs and BioFETs developed in SPICE. These are
based on MOSFET models with modifications. However, the SPICE simulations do
not include the electrochemistry of the device. The physical and chemical behavior
of ISFET is modeled using physico-chemical condition by Grattarola et al. They
have carried out the simulation in electronic circuit simulation program SPICE.
Electrochemical models have been developed for silicon-based BioFETs only [16].
But, there is no electrochemical model for DG-CNTISFET.
2.1 Site-Binding Model
The concentration of H+ ions plays an important role in ISFETs. These ions determine
the surface potential at the interface between the electrolyte and insulator depending
on their binding with the surface sites of the gate insulator. The surface charge change
can be well-explained by the site-binding model. Yates et al. [17] in 1973 introduced
the site-binding model that described oxide and aqueous electrolyte interface prop-
erties. Fung et al. [18] in 1986 generalized this model to characterize the properties
of ISFET with oxide insulators.
The relationship
+ between the +hydrogen
ion (H+ ions) concentration at the oxide
surface ( H s ) and the bulk ( H b ) is given by Eq. (1):

+ + −qψ0
H s = H b exp (1)
kT
where ψ0 is the surface potential between the electrolyte and insulator interface, T
denotes the absolute temperature, k denotes Boltzmann constant, and q denotes the
elementary charge.
Hence, a surface potential, ψ0 is developed as the H+ ions present in electrolyte
solution binds with the oxide surface sites. The potential developed at the interface
changes with solution pH.
2.2 Model of Interfacial Potential
The potential developed at the electrolyte and insulator interface changes with the pH
of the solution. Bousse et al. [19] developed a relationship based on the site-binding
theory that describes the potential change with solution pH. The relation is given by
Eq. (2) [21]:

kT β
ψ0 = 2.3 p H pzc − p H (2)
q β +1
where the value of pH for the oxide surface to be electrically neutral is pHpzc; β is a
dimensionless parameter; k denoted Boltzmann constant; T denotes absolute temper-
ature. The values of different parameters for determination of surface potential, ψ0
of HfO2 surface are listed in [20].
2.3 Current Transport Model of DG-CNTISFET
The drain current equation of DG-CNTISFET can be determined by modifying the

equations of MOSFET drain current. The conventional MOSFET like behavior of
DG-CNTISFET has been verified experimentally in [15]. The drain current through
the channel region in DG-CNTISFET depends on the potential both at top and bottom
gate.
Thus, the linear and saturation region drain current in DG-CNTISFET can be
written as:
W

IDS,lin,dual = μ VDS Cox,top VGS,top − VTH,top − 0.5VDS
L

+Cox,bottom VGS,bottom − VTH,bottom − 0.5VDS (3)
W
IDS,sat,dual = μ Cox,top (VGS,top − VTH,top )2
2L
+Cox,bottom (VGS,bottom − VTH,bottom )2 (4)
In the above equation, Cox,top is the top gate capacitance and Cox,bottom is the
bottom gate capacitance; VTH,top is the top gate threshold voltage, and VTH,bottom is
the bottom gate threshold voltage.
In DG-CNTISFET, total threshold voltage is due to the sum of the threshold
voltage of the top and bottom gates. The VTH,top is given by ISFET’s threshold voltage
equation without the term 2 f . The bottom gate VTH is represented by MOSFET’s
threshold voltage equation except the term 2 f .
CNT Q ox,top + Q ss,top + Q CNT

VTH,top = E ref − ψ0 + χsol − − (5)
q Cox,top
M − CNT Q ox,bottom + Q ss,bottom + Q CNT
VTH,bottom = − (6)
q Cox,bottom
where Q ox,top is the top gate and Q ox,bottom is the bottom gate insulator charge; Q ss,top
is the interface charge for insulator-CNT top gate and Q ss,bottom is the interface charge
for insulator-CNT bottom gate; M is the work function of bottom gate material;
E ref is the reference electrode potential; CNT is the CNT work function; q is the
elementary charge; χsol is the solvent surface dipole potential; ψ0 is the interfacial
potential between the electrolyte and insulator interface. The value of ψ0 depends
on the analyte ions activity.
The total charge inside the carbon nanotube is given by Eq. (7)
Q CNT = −qn CNT L CNT (7)
where n CNT is carbon nanotube carrier concentration; L CNT is carbon nanotube

length.
The oxide layer charge can be written as Eq. (8)
Q ox = qn ox (8)
where, n ox is the oxide layer carrier concentration.
2.4 pH Sensitivity of DG-CNTISFET
The device sensitivity due to the top gate can be determined using Eq. (2). However,
the sensitivity of DG-CNTISFET depends on both the gates (top and bottom). The
capacitive coupling ratio is high when the top gate capacitance is higher compared to
bottom gate. As a result, the sensitivity of the device will be more. Large capacitive
coupling ratio can be obtained with high-κ and low-κ as the top and bottom gate
material, respectively. The thickness of the top gate dielectric should be less than the
bottom gate dielectric material. The threshold voltage shift of DG-CNTISFET can
be obtained by Eq. (9)
Cox,top
VTH = − ψ0 (9)
Cox,bottom
This equation shows that the threshold voltage change due to changes in the surface
potential is increased by a factor of Cox,top /Cox,bottom . This enhanced sensitivity makes
it advantageous to detect small changes.
Equation (9) can also be written as:
Cox,top
Stotal = Stop (10)
Cox,bottom
where Stotal denotes sensitivity of the DG-CNTISFET and Stop is sensitivity of the
device due to the top gate only.
The change of the device surface potential with pH change has been determined
using Eq. (2). Figure 2 shows the variation of surface potential with pH change. The
surface potential decreases with the increase the of solution pH. Figure 3 shows the
threshold voltage change of the top gate with pH. The threshold voltage increases
with increase in pH value. This parameter acts as input for drain current variation of
the device.
The sensitivity of the DG-CNTISFET was found to be 907 mV/pH compared to
experimental value of 943 mV/pH using Eq. (10). The increased sensitivity compared
to single-gate device is due to the dual-gate structure. Figure 4a shows the drain
characteristics of the device in liquid medium. The experimental results have also
been embedded in the plot. The plot shows the experimental and modeled results are
in good agreement. Table 1 shows the values of different parameters that influenced
the result most.
Fig. 2 Surface potential 0.15

change with pH
0.1
Surface potential in V
0.05
-0.05
-0.1
5 6 7 8 9
pH
0.5
0.45
Threshold voltage in V
0.4
0.35
0.3
0.25
0.2
5 6 7 8 9
pH
Fig. 3 Threshold voltage change of top gate with pH
Fig. 4 a Drain characteristics of DG-CNTISFET in liquid b drain characteristics of DG-CNTISFET

outside liquid
Table 1 Different influential parameters used in simulation

S. No. Different parameter Values
1 Reference electrode potential (E ref ) 0.6 V
2 CNT work function (CNT ) 4.7 eV
3 Solvent surface dipole potential (χsol ) 0.2 V
4 Length of CNT (L CNT ) 5 mm
5 Number of charge carriers in CNT (n CNT ) 9 × 1015 cm−3
6 Number of charge carriers in the oxide (n ox ) 9 × 1012 cm−2
7 HfO2 oxide thickness (tox ) 10 nm
8 Charges in the surface and interface states (QSS ) −3.5 × 10−6 C/cm2
9 ITO work function (M ) 4.6 eV
10 Thickness of HfO2 (t ox, top ) 10 nm
11 Thickness of ZnO (t ox, bottom ) 10 nm
12 Dielectric constant for ZnO (κ ox, bottom ) 1.5
13 Dielectric constant for HfO2 (κ ox, top ) 25
The device drain characteristic outside liquid medium has also been investigated.
For this, a metal Al gate has been deposited by physical vapor deposition (PVD)
method on the top of the gate oxide layer. This is now similar to a MOSFET device.
The threshold voltage was found to be 0.2 V. The drain characteristics were deter-
mined using Eqs. (3) and (4). The characteristics are similar to MOSFET. A compar-
ison is shown in Fig. 4b between the modeled and experimental results. A good fit
was obtained. Thus, it can be concluded that the device performance depends on the
gate both in liquid and outside liquid.
4 Conclusion
A model was developed for DG-CNTISFET, which considers all the physical and
electrochemical parameters of the fabricated device. The device showed good sensi-
tivity with high-κ HfO2 (dielectric constant ~25) as top gate (thickness ~10 nm) and
low-κ ZnO dielectric material in the bottom gate. The enhanced sensitivity is due to
the high coupling ratio. The drain current variation of the device is mostly due to the
top gate of the device. The high mobility of CNT greatly contributes toward enhanced
sensitivity so that the device can detect even small surface potential change. A small
surface potential change causes a large change in drain current. A good fit was found
between the experimental and the modeled results. Thus, this electrochemical model
can be used for understanding DG-CNTISFET working.
Acknowledgements The authors thank the Department of Electronics and Information Tech-
nology, MCIT, Government of India for the support through Visvesvaraya PhD scheme and Tezpur
University for providing the laboratory facilities.
References
1. Gruner G (2006) Carbon nanotube transistors for biosensing applications. Anal Bioanal Chem
384(2):322–335
2. Harvey JD, Jena PV, Baker HA, Zerze GH, Williams RM, Galassi TV, Roxbury D, Mittal J,
Heller DA (2017) A carbon nanotube reporter of microRNA hybridization events in vivo. Nat
Biomed Eng 1(4):0041
3. Allen BL, Kichambare PD, Star A (2007) Carbon nanotube field-effect-transistor-based
biosensors. Adv Mater 19(11):1439–1451
4. Peng LM, Zhang Z, Wang S (2014) Carbon nanotube electronics: recent advances. Mater Today
17(9):433–442
5. Javey A, Kim H, Brink M, Wang Q, Ural A, Guo J, McIntyre P, McEuen P, Lundstrom M,
Dai H (2002) High-κ dielectrics for advanced carbon-nanotube transistors and logic gates. Nat
Mater 1(4):241
6. Javey A, Guo J, Farmer DB, Wang Q, Wang D, Gordon RG, Lundstrom M, Dai H (2004) Carbon
nanotube field-effect transistors with integrated ohmic contacts and high-κ gate dielectrics.
Nano Lett 4(3):447–450
7. Javey A, Tu R, Farmer DB, Guo J, Gordon RG, Dai H (2005) High performance n-type carbon
nanotube field-effect transistors with chemically doped contacts. Nano Lett 5(2):345–348
8. Barik MA, Dutta JC (2014) Fabrication and characterization of junctionless carbon nanotube
field effect transistor for cholesterol detection. Appl Phys Lett 105(5):1–6
9. Barik MA, Deka R, Dutta JC (2016) Carbon nanotube-based dual-gated junctionless field-effect
transistor for acetylcholine detection. IEEE Sens J 16(2):280–286
10. Dutta JC, Thakur HR (2017) Sensitivity determination of CNT based ISFETs for different
high-κ dielectric materials. IEEE Sens Lett 2(3):1–4
11. Spijkman M, Smits ECP, Cillessen JFM, Biscarini F, Blom PWM, De Leeuw DM (2011)
Beyond the Nernst-limit with dual-gate ZnO ion-sensitive field-effect transistors. Appl Phys
Lett 98(4):043502
12. Liu S, Guo X (2012) Carbon nanomaterials field-effect-transistor-based biosensors. NPG Asia
Mater 4(8):e23
13. Jang H-J, Bae T-E, Cho W-J (2012) Improved sensing performance of polycrystalline-silicon
based dual-gate ion-sensitive field-effect transistors using high-k stacking engineered sensing
membrane. Appl Phys Lett 100(25):253703
14. Dutta JC, Sharma PK (2018) Fabrication, characterization and electrochemical modeling of
CNT based enzyme field effect acetylcholine biosensor. IEEE Sens J 18(8):3090–3097
15. Dutta JC, Thakur HR, Keshwani G (2019) High performance dual-gate carbon nanotube ion-
sensitive field effect transistor with high-κ top gate and low-κ bottom gate dielectrics. IEEE
Sens J 19(14):5692–5699
16. Sharma PK, Thakur HR, Dutta JC (2017) Modeling and simulation of carbon nanotube-
based dual-gated enzyme field effect transistor for acetylcholine detection. J Comput Electron
16(3):584–592
17. Yates DE, Levine S, Healy TW (1974) Site-binding model of the electrical double layer at the
oxide/water interface. J Chem Soc Faraday Trans 1 Phys Chem Condens Phases 70:1807–1818
18. Fung D, Cheung PW, Ko WH (1986) A generalized theory of an electrolyte-insulator-
semiconductor field-effect transistor. IEEE Trans Electron Devices 33(1):8–18
19. Bousse L, De Rood NF, Bergveld P (1983) Operation of chemically sensitive field-effect sensors
as a function of the insulator-electrolyte interface. IEEE Trans Electron Devices 30(10):1263–
1270
20. Sharma PK, Thakur HR, Dutta JC (2016) Fabrication and characterization of a carbon
nanotube based junctionless ion sensitive field effect transistor (CNT-JLISFET). In: Inter-
national conference on computing, communication and automation (ICCCA), IEEE, Noida,
pp 1450–1453
21. Dutta JC (2012) Ion sensitive field effect transistor for applications in bio-electronic sensors:
a research review. In: 2nd national conference on computational intelligence and signal
processing (CISP), IEEE, Guwahati, pp 185–191
A Novel Approach to Calculate TF-IDF
for Assamese Language
Hsuvas Borkakoty, Chandana Dev, and Amrita Ganguly
Abstract TF-IDF is one of the earliest and most effective word vectorization tech-
niques that provide the basis for many natural language processing (NLP) tasks.
This paper provides a novel approach to create TF-IDF vectors for Assamese text.
A considerable number of experiments are carried out throughout the process, and a
significant result was found which is discussed here in detail.
Keywords TF-IDF · Assamese · Corpus · NLP
1 Introduction
Term frequency–inverse document frequency (TF-IDF) is one of the effective

weighing factors in many natural language processing tasks, such as information
retrieval and text mining. It calculates the probability of occurrence of the word in
a document, which can be used to find the originality of the word. The strength of
the value of TF-IDF vector of a word determines the relationship of the word in the
document, and the stronger the value, higher the relation is. The approach described
here tries to determine the TF-IDF value of documents in Assamese language, which
is a morphologically rich Indo-Aryan language. The approach calculates TF-IDF
based on the novel approach to find the frequency of the words in the document. The
experimentation is carried out with different sets of data, with sentences, paragraphs
and documents. The results produced to provide substantial information about the
H. Borkakoty (B)
Jorhat Institute of Science and Technology, Sotai, Jorhat-10, India
C. Dev · A. Ganguly
Department of Electrical Engineering and Industrial Engineering, Assam Engineering College,
Jalukbari, Guwahati-13, India
A. Ganguly
388 H. Borkakoty et al.
probability of occurrence of the word in the document(s). The following subsec-

tions cover different aspects of the approach. Section 2 covers a brief literature
review of the topic. Section 3 sheds light on the methodology that is used to find the
results. Section 4 provides information on the experimentation and results; finally,
the conclusion is given in Sect. 5.
2 Literature Review
Spärck Jones [1] first revealed the relationship between exhaustivity of document
description and the specificity of words in the document which first led to the idea
of TF-IDF. The inherent relation was mathematically represented by Robertson [2],
where he imprinted the IDF as log of total number of documents divided by the
number of documents where the word occurs. Since then, TF-IDF has become a
widely used word vectorization technique, used in many NLP tasks, such as clas-
sification, summarization, clustering and keyword extraction. Some of the uses are
reviewed throughout this section.
Chan et al. [3] offered a classification system for Chinese news segments, where
term frequencies are used to build the news words database, which in turn was
used in classifier. Similarly, Abu-Errub [4] proposed a classification method to clas-
sify Arabic text documents where the categorization was done by using TF-IDF
frequencies to find out the index terms, which were then used in the classifier. The
classification task for Indian languages is also been done using TF-IDF. Notable
work includes Dhar et al. [5] work on Bengali text classification where the features
were determined using TF-IDF with dimensionality reduction using 40% of term
frequencies of the word, which resulted in the highest of 97.78% accuracy in the
classification of sports documents. Islam et al. [6] used a support vector machine
[7] and a modified TF-IDF for feature selection which then used to classify a set of
Bengali text documents with 92.57% of accuracy.
Apart from document classification, TF-IDF is also used in other language
processing tasks for Indian languages. Notable works include Jayashree et al. [8]
where they used TF-IDF along with GSS [9] coefficients to calculate the scores of
the sentences in order to employ the extractive summarization technique. Sarkar [10]
developed a summariser which ranked sentences according to their cumulative TF-
IDF values which are then measured against a predefined threshold in order to find
the thematic terms of a document. As for document clustering, Hannan et al. [11]
used TF-IDF for vector space modelling of the words after the cleaning of corpuses.
Hanumanthappa et al. [12] devised a keyword extraction mechanism using TF-IDF
for three Dravidian languages, viz. Tamil, Telugu and Kannada, where TF-IDF was
used to evaluate the importance of the word in the document and to model the vector
space model for the vocabulary.
After reviewing all these literature, it has become evident that TF-IDF can serve
as a very important pre-processing component in many language processing tasks.
A Novel Approach to Calculate TF-IDF for Assamese Language 389
3 Methodology
Our proposed methodology is provided in accordance with the standard approach to

find the TF-IDF of the document(s). The approach is divided into a set of steps as
follows (Fig. 1).
3.1 Pre-processing
The pre-processing tasks that are carried out are tokenization and removal of stop
words and special characters of words from the text. Since there is no predefined
functions available to remove the stop words of Assamese language, hence a function
is defined, which takes the text file as input removes special characters, etc., tokenizes
the sentences using word tokenizer of NLTK [13]. After these two tasks, it uses a file
Fig. 1 Proposed
methodology for calculation
of TF-IDF
that contain the stopwords, obtained on request from the Indian Language Technology
Proliferation and Deployment Centre (TDIL-DC) to find and remove the stop words.
The output that is returned is in the form of a list of words with empty string for the
stop words.
3.2 Computation of Term Frequency (TF)
A function is designed that takes a dictionary containing the pre-processed file (output
of pre-processing function) and creates the term frequency of the words which is
returned in the form of a dictionary. Term frequency is calculated as follows.
word count
Term Frequency =
length(bag of Words)
3.3 Computation of Inverse Document Frequency (IDF)
The inverse document frequency (IDF) is calculated by converting the text files
(documents) which are first to dictionaries and passed to the function as list of
dictionaries. After that, the length of the list, that is, the total number of words in the
dictionaries, is found. After that, IDF values of the words are calculated as per the
formula mentioned below which is then stored in another dictionary against every
word.

Total number of documents
Inverse Document Frequency = log10
appearence of the word in the documents
3.4 Computation of TF-IDF
After calculating both TF and IDF for every word in the document, the calculation
of TF-IDF is done using the following formula and the result is stored in another
dictionary. The TF-IDF score is calculated by multiplying the TF and IDF score of
the word.
TFIDF = TF × IDF
4 Experiment and Results
The code, written in Python 3.7, contains the functions mentioned in the section
above. The functions are defined to carry out the tasks from pre-processing till finding
out the TF IDF, and inputs are given accordingly. The functions return the values
after the calculation of respective components. The experiments are designed in
accordance with the size of text, that is, starting with a set of sentences and continuing
with set of paragraphs, a single document and ending with two different documents.
All these data are accepted from the Assamese general text corpus obtained from
the Indian Language Technology Proliferation and Deployment Centre (TDIL-DC).
The sentences that are considered here are from an Assamese non-fiction named
“Asomiya Upanyashr Bhumika” which is also a part of the aforesaid corpora. The
results of the experiments are given as a set of screenshots in Figs. 2, 3, 4 and 5. It
has been observed that although the code works with efficiency for a set of inputs,
with the increase of dataset, the time consumption goes higher. However, the results
are quite satisfying and the program is scalable to accept n-number of documents to
find the TF-IDF of the words contained in those documents.
The results seem promising and can be given as input to different NLP tasks.
However, it is also noticed that with increase of documents, the TF-IDF values also
increase for a certain set of words, whereas for some words, it shows the opposite
result. This observation leads to the existence of certain frequent words and their use
in different senses.
Fig. 2 TF-IDF value for two sets of sentences
Fig. 3 TF-IDF value for words of two consecutive paragraphs. Taken from one file
Fig. 4 TF-IDF value for the words of one file
Fig. 5 TF-IDF value for the words of two different files
5 Conclusion
This paper provides a novel approach to calculate TF-IDF feature vector for Assamese
texts. The experimentation carried out on the code shows the word vectors for the
words of the given corpus files. This can find applications in many NLP tasks such as
document classification, named entity recognition and text summarization where the
vectors can be fed as input parameters to the system. Thus, the future work should
include the prospect of using the vectors to carry out the aforementioned NLP tasks.
References
1. Sparck JK (1972) A statistical interpretation of term specificity and its application in retrieval.
J Documentation 28(1):11–21
2. Robertson S (2004) Understanding inverse document frequency: on theoretical arguments for
IDF. J Documentation 60(5):503520
3. Chan TY, Chang YS (2017) Enhancing classification effectiveness of Chinese news based on
term frequency. In: 2017 IEEE 7th international symposium on cloud and service computing
(SC2). IEEE, pp 124–131
4. Abu-Errub A (2014) Arabic text classification algorithm using tf-idf and chi square measure-
ments. Int J Comput Appl 93(6)
5. Dhar A, Dash NS, Roy K (2018) Application of tf-idf feature for categorizing documents of
online bangla web text corpus. Intelligent engineering informatics. Springer, Singapore, pp
51–59
6. Islam MS, Jubayer FE, Ahmed SI (2017) A support vector machine mixed with TF-IDF algo-
rithm to categorize Bengali document. In: 2017 international conference on electrical, computer
and communication engineering (ECCE). IEEE, pp 191–196
7. Kecman V (2005) Support vector machine- an introduction. https://doi.org/10.1007/109846
97_1
8. Jayashree R, Srikanta KM, Sunny K (2011) Document summarization in Kannada using
keyword extraction. Proc AIAA 11:121–127
9. Galavotti L, Sebastiani F, Simi M (2000) Experiments on the use of feature selection and
negative evidence in automated text categorization. In: International conference on theory and
practice of digital libraries. Springer, Berlin, Heidelberg, pp 59–68
10. Sarkar K (2012) Bengali text summarization by sentence extraction. arXiv preprint arXiv:1201.
2240
11. Hannan A, Boro SR, Sarma JPSSK (2015) An approach to Bodo document clustering. Int J
Innovative Res Sci Eng Technol (ISSN: 2319-8753) 4(12), https://doi.org/10.15680/IJIRSET.
2015.0412069
12. Hanumanthappa M, Narayana Swamy M, Jyothi NM (2014) Automatic keyword extraction
from Dravidian language. Int J Innovative Sci Eng Technol 1(8):87–92
13. Loper E, Bird S (2002) NLTK: the natural language toolkit. arXiv preprint cs/0205028
A Low-Power Third-Order Passive
Continuous-Time Sigma-Delta
Modulator Using FinFET
Sarangam Kunamalla and Bheema Rao Nistala
Abstract FinFETs are known to be one of the promising devices for sub-50 nm
regime. It involves better channel control, reduced short channel effects and low
leakage current. A third-order continuous-time sigma-delta modulator presented in
this paper is constructed by adding a capacitor to the stable second-order low-pass
RC filter. It improves the quantization noise shaping and linearity of the modulator
without increasing the power dissipation. A low-power, low-offset SR latch-based
clocked comparator is used as quantizer. The designed modulator is implemented in
FinFET 16 nm process, achieves the result of SNDR of 64 dB, ENOB of 10.5 bits,
power dissipation of 48 µW and operating with 1 V supply voltage. It is suitable for
low-power ADC in biomedical applications with a signal bandwidth of 4 kHz.
Keywords FinFET · Continuous-time · Sigma-delta modulator · Low power ·

Analog to digital converter
1 Introduction
The development of embedded biomedical systems in recent years has increased the
demand for low-power ADC [1]. The bio-potential signals are weak, and they require
high resolution for accurate conversion. Sigma-delta modulators are commonly used
for low-power ADC because of its simple circuitry, oversampling and noise shaping
property [2]. There are two possible implementations of sigma-delta architectures.
CT Modulators are more power efficient than discrete-time modulators because no
need of anti-aliasing filter [3].
The basic architecture of a sigma-delta modulator is presented in Fig. 1. It
comprises of loop filter, comparator and a DAC in the feedback path. The loop filter
S. Kunamalla (B) · B. R. Nistala

B. R. Nistala
396 S. Kunamalla and B. R. Nistala
Fig. 1 Basic architecture of a sigma-delta modulator [4]
is mostly realized using active integrator. The comparator is a clocked comparator

which is a 1-bit quantizer, and DAC is an inverter. Operational amplifier is the main
building block in an integrator and consumes significant portion of the power in
an overall sigma-delta modulator. The order of the filter depends on the number of
active integrators. Each integrator requires one op-amp. As the order of the filter
increases, the number of op-amps increases which limits the low power dissipation
of the sigma-delta modulator.
The trend is towards smaller number of amplifiers than the order of the loop filter
and without any amplifier [5–8]. A hybrid active–passive loop filter is implemented
[9] with three active integrators and two passive integrators. This saves the power
but resolution is restricted to 10 bits. In ref [10], a second-order passive sigma-delta
ADC for low-power applications is implemented using cascade of two first-order RC
circuits. This paper achieves simple circuit and low power but two poles in a single-
loop cause stability problem. A 1 MHz continuous-time sigma-delta modulator using
genetic algorithm approach is implemented by adding zero to the second-order filter
to solve the problem of stability [11].
The passive filters are simple, no power consumption, linearity and operating at
low supply voltage [12]. These filters are limited to second order only because it
depends on the comparator gain. As the order of the filter increase, resulting a low
voltage swing at its input. One solution to this is adding a capacitor to the second order
to filter and pre-amplifier to the latched comparator. A third-order continuous-time
sigma-delta modulator is implemented in this paper using passive only this achieves
bandwidth using FinFET devices. FinFETs are promising devices in near future and
have less short channel effects and low leakage currents than planar MOSFETs [13].
The paper is structured in the following manner. In Sect. 2, first- and second-
order passive sigma-delta modulator architecture and its problems are described.
The proposed third-order passive SDM architecture and its implementation details
A Low-Power Third-Order Passive Continuous-Time Sigma-Delta … 397
are demonstrated in Sect. 3. In Sect. 4, the simulation results are discussed. Finally,
the paper concludes in Sect. 5.
2 First- and Second-Order Passive Sigma-Delta Modulator

Architecture
2.1 First-Order Passive Sigma-Delta Modulator
In this topology, the loop filter is realized with R1 , C 1 network as shown in Fig. 2. The
input and feedback resistors are same value to get unity gain. Here, the integration
function is performed by the capacitor C 1. The main disadvantage of this topology is
voltage across the capacitor is varying due to input and feedback resulting distortion
present in the modulator output.
Equation (1) indicates the output voltage as related to input voltage, integration
node voltage and quantization error voltage. The first-order STF, NTF, DTF is as
shown below Eqs. (2), (3), (4), respectively.
Vin S R1 C 1 −2
Vout = + · Vqe + · Vint (1)
1 + S R1 C 1 1 + S R1 C 1 1 + S R1 C 1
1
STF = (2)
1 + S R1 C 1
S R1 C 1
NTF = (3)
1 + S R1 C 1
Fig. 2 First-order passive sigma-delta modulator architecture [10]

−2
DTF = (4)
1 + S R1 C 1
2.2 Second-Order Passive Sigma-Delta Modulator
This topology is obtained by cascading of two RC sections as shown in Fig. 3.

It improves the linearity and reduces the distortion in the modulator output. The
drawback of this topology is cascading of RC sections attenuates the input signal at
the positive input of the comparator resulting the noise of the comparator increases.
The second drawback is two poles in a negative feedback loop cause instability.
node voltage and quantization error voltage. The second-order STF, NTF, DTF is as
shown below Eqs. (6), (7), (8), respectively.
(1 + SC2 R2 ) S(S R1 C1 R2 C2 + R1 C1 )
Vout = · Vin + · Vqe
1 + S R2 C 2 + S R1 C 1 1 + S R1 C 1 + S 2 R1 C 1 R2 C 2
1
+ · Vint (5)
1 + S R1 C1 + (1 + S R2 C2 )
(1 + SC2 R2 )
STF = (6)
1 + S R2 C 2 + S R1 C 1
S(S R1 C1 R2 C2 + R1 C1 )
NTF = (7)
1 + S R1 C 1 + S 2 R1 C 1 R2 C 2
Fig. 3 Second-order passive sigma-delta modulator architecture [10]

1
DTF = (8)
1 + S R1 C1 + (1 + S R2 C2 )
3 Proposed Third-Order - Modulator
The proposed third-order - modulator is demonstrated in Fig. 4. In this modified

architecture, the conventional second-order - modulator is stabilized by series
R3 and C 2 to create the left-hand zero. Now, in proposed topology, we add one more
capacitor C 3 parallel to series R3 and C 2 . By adding this capacitor, we are getting
third-order - modulator and less signal voltage at the positive node capacitor.
In order to get the input signal at the output, a 2-stage RC filter is added to the
modulator output. This topology saves a resistor when compared to cascading of
three RC sections.
node voltage and quantization error voltage. The proposed third-order STF, NTF,
DTF is as shown Eqs. (11), (12), (13), respectively.
(1 + SC1 R2 ) D
Vout = · Vin + · Vqe
D + (1 + SC1 R2 ) D + (1 + SC1 R2 )
(2R1 + SC3 R3 R1 )(1 + SC1 R2 )
+ · Vint (9)
D + (1 + SC1 R2 )

D = (2R1 + R3 + SC3 R3 R1 ) S(C1 + C2 ) + S 2 C1 C2 R2 + (1 + SC1 R2 ) (10)
Fig. 4 Proposed third-order - modulator

(1 + SC1 R2 )
STF = (11)
D + (1 + SC1 R2 )
D
NTF = (12)
D + (1 + SC1 R2 )
(2R1 + SC3 R3 R1 )(1 + SC1 R2 )
DTF = (13)
D + (1 + SC1 R2 )
3.1 Passive Filter Implementation
The dominant pole is coming through the resistance R3 and the capacitor C 3 . We are
choosing the values of R3 and C 3 by criteria of R3 C 3 > 5T where T is input signal
time period. The above condition is hold for proper integration. The values of R1 , R2 ,
C 1 , C 2 are selected based on resolution limit of the comparator and system stability.
The overall transfer function of the proposed SDM is displayed in Fig. 5. The
bandwidth of the proposed modulator is 10 kHz. The first pole, second pole and
third pole of this modulator are 3.8 kHz, 63 MHz and 350 MHz, respectively, and
zero is compensated by second pole. The STF and NTF are low-pass and high-pass
Fig. 5 Overall transfer functions of proposed modulator

Fig. 6 Circuit diagram of comparator
characteristics, respectively, having a slope of 50 dB/decade with values of R1 , R2 ,

R3 , C 1 , C 2 , C 3 are set to 10 M, 600 , 600 , 15pF, 2pF, 2pF, respectively.
The sigma-delta modulator’s critical portion is a comparator. The poor perfor-
mance of comparator can degrade the resolution of the modulator. Figure 6 indicates
the circuit diagram of the comparator. It comprises one pre-amplifier stage followed
by a positive feedback latch and S–R flip-flop. The pre-amplifier stage is simple
common source amplifier (M 1 and M 13 , M 4 and M 14 ) having a gain of gm13 /gm1 and
gm14 /gm4 to increase the resolution of the comparator. After the design of pre-amplifier
stage, a cross-coupled latch (M 2 –M 3 and M5 –M 12 ) together with an S–R latch built
with NAND gates (M 13 –M 16 and M 17 –M 20 ) to generate the accurate output signal.
The S–R latch reduces the meta-stability occurrence in case of large regeneration time
and output changes only raising edge of the clock. The inverters (M 21 –M 22 , M 23 –
M 24 ) increase the available output current. The current mirror (M 25 –M 27 ) provides
the bias current of 10 µA.
The designed comparator has resolution limit of 5 mV, and simulation results are
plotted in Fig. 7.
The proposed architecture has been designed in a 16 nm FinFET process with a

supply voltage of 1 V. Figure 8 shows the transient response of the proposed sigma-
delta modulator for a sine wave input of 4 kHz, 200 m V pp . Resulting a pulse width
modulated signal having more number of one’s corresponding to 1 V and more
number of zero’s corresponding to 0 V. Intermediate values will get 1, 0 transition.
The SDM modulator output spectrum with the sampling rate of 1 MHz is depicted
in Fig. 9. From this plot, we are getting a gain of −34.54 dB at 4 kHz sine wave
Fig. 7 Simulation results of the comparator
Fig. 8 Transient response of the proposed modulator
input signal and the average noise factor floor value is around −90 dB up to 10 kHz
bandwidth.
Figure 10 depicts the FFT plot of PWM output waveform. Regarding dynamic
performance of the modulator, a signal to noise and distortion ratio (SNDR) of
64 dB was achieved. This corresponds to the effective resolution of 10.5 bits and
power dissipation 48 µW.
Fig. 9 SDM modulator output spectrum (4 kHz, 500mVpp )
Fig. 10 FFT plot of the PWM output
The proposed SDM’s performance is outlined and distinguished with the previous
published passive sigma-delta modulator in Table 1. Based on the comparison Table 1,
the proposed modulator achieves lower FOM and is designed in a lower technology.
Table 1 Performance comparison with previous published papers

S. No. Parameter This work Ref [10] Ref [14] Ref [15]
1 Process 16 nm FinFET 500 nm 350 nm 500 nm
2 Signal BW 4 kHz 5 kHz 5 kHz 3 kHz
3 Sampling rate 1 MHz 10 MHz 1 MHz 1 MHz
4 Resolution (ENOB) 10.5 bits 8 bits 10 bits –
5 SNDR 64 dB 50 db – 57.6 dB
6 Power consumption 48 µW 100 µW 50 µW –
7 Supply voltage 1V 5V 3.3 V 2.5 V
5 Conclusion
A third-order sigma-delta modulator presented in this paper is designed by two

cascaded passive RC circuits, a capacitor and comparator. The comparator is low-
offset latched type implemented using FinFET devices. The stability of the modulator
is achieved by overall transfer function plot. The designed modulator has an SNDR
of 64 dB, effective resolution ENOB of 10.5 bits, power dissipation 48 µW, and it is
suitable for low area, low power dissipation and low-scale process ADC applications.
References
1. Benabes P, Kielbasa R (2002) Passive Sigma-Delta converters design. In: Proceedings of 19th
IEEE instrumentation and measurement technology conference, IMTC 2002, Anchorage, AK,
USA, 21–23, pp 469–474
2. Baker RJ (2009) CMOS mixed-signal circuit design, 2nd edn. Wiley, London, pp 59–60. ISBN:
978-0-470-29026-2
3. Schreier R, Temes GC (1997) Understanding delta-sigma data converters. IEEE press,
Piscataway
4. de la Rosa JM (2011) Sigma-delta modulators: tutorial overview, design guide, and state-of-
the-art survey. IEEE Trans Circ Syst I Reg Pap 58(1):1–21
5. Srinivasan V, Wang V, Satarzadeh P, Haroun B, Corsi M (2012) A 20 mW 61 Db SNDR (60 MHz
BW) 1b 3rd-order continuous-time delta-sigma modulator clocked at 6 GHz in 45 nm CMOS.
In: EEE international solid-state circuits conference (ISSCC), Digest of Technical Papers, pp
158–160
6. Zeller S, Muenker C, Weigel R, Ussmueller T (2014) A 0.039 mm inverter-based 1.82 mW
68.6 dB-SNDR 10 MHz-BW CT ADC in 65 nm CMOS using power- and area-efficient design
techniques. IEEE J Solid-State Circ 49(7):1548–1560
7. Matsukava K et al (2010) A fifth-order continuous-time delta-sigma modulator with single-
opamp resonator. IEEE J Solid-State Circ 45(4):697–706
8. Weng CH, Wei TA, Alpman E, Fu CT, Tseng YT, Lin TH (2014) An 8.5 MHz 67.2 dB SNDR
CTDSM with ELD compensation embedded twin-T SAB and circular TDC-based quantizer
in 90 nm CMOS. In: Proceedings IEEE international symposium on VLSI circuits, pp 1–2
9. Song T, Cao Z, Yan S (2008) A 2.7-mW 2-MHz continuous-time modulator with a hybrid
active–passive loop filter. IEEE J Solid-State Circ 43(2):330–341
10. Roy A, Baker RJ (2014) A passive 2nd-order sigma-delta modulator for low-power analog-
to-digital conversion. In: 2014 IEEE 57th international midwest symposium on circuits and
systems (MWSCAS), College Station, TX, pp 595–598
11. De Melo JL (2013) A low power 1-MHz continuous-time M using a passive loop filter designed
with a genetic algorithm tool. In: Proceedings EEE international symposium on circuits and
systems (ISCAS), pp 586–589
12. Sai T, Sugimoto Y (2009) Design of a 1-V operational passive sigma-delta modulator. In: 2009
European conference on circuit theory and design, Antalya, pp 751–754
13. Thakker RA, Srivastava M, Tailor KH, Baghini MS, Sharma DK, Rao VR, Patil MB (2011) A
novel architecture for improving slew rate in FinFET-based op-amps and OTAs. Microelectron
J 42(5):758–765
14. Guessab S, Benabes P, Kielbasa R (2004) Passive delta-sigma modulator for low-power
applications. In: MWSCAS ‘04, pp 295–298
15. Gundel A, Carr WN (2007) A micro power sigma-delta A/D converter in 0.35 m CMOS for
low frequency applications. LISAT 2007 pp 1–7
Mitigation of Power Fluctuations
in UPFC-Based Network
Rajen Pudur, Pallavi Kumari, and Mrinal Kanti Rajak
Abstract Unified power flow controller which is the latest family in FACT family
is used for real-time control and dynamic compensation of AC transmission system.
UPFC controls all the parameters that affect power flow (voltage, phase angle and
impedance) of transmission line by either selectively or simultaneously. But with
conventional control schemes, the power fluctuation cannot be completely mitigated,
and the time constant of damping is independent of real and reactive power feedback.
In this paper, an attempt is made to validate the advance control scheme available
for mitigation of the power fluctuation with MATLAB simulation. Complete model
is prepared and analysed ignoring the dynamics of shunt device; the power fluctu-
ation attenuation has been successfully achieved by simulation, and the results are
presented.
Keywords Unified power flow controller · Power feedback control ·

Cross-coupling control
1 Introduction
The power transmission lines are in continuous improving to serve its best function of
catering uninterrupted power to the consumers. These power systems are comprised
of three major parts, generation, transmission and distribution which are governed
by the demands of loads at different places, and their availability and controllability
largely depend on the latest technology breakthrough. The Electric Power Research
R. Pudur (B) · P. Kumari

National Institute of Technology Arunachal Pradesh, Yupia, Arunachal Pradesh, India
P. Kumari
M. K. Rajak
Department of Electrical Engineering, National Institute of Technology Arunachal Pradesh,
408 R. Pudur et al.
Institute (EPRI) has introduced a new concept which was called Flexible AC Trans-
mission System (FACTS) [1] in the late 1980, which gives a concept or power to
the existing power line supply more power to load up to its thermal capabilities. A
FACTS device gives us freedom to choose how to control and increase the efficiencies
of power line. Unified power flow controller (UPFC) is the youngest in the family
of FACTS, and it can give both controls of real power and reactive powers indepen-
dently in single go. Many researches have conducted to have smooth power flow
without fluctuations, but it generally seems power fluctuation in these lines. Phase
angle control and cross-coupling control are two known control methods which can
give smooth flow of power, but power fluctuation has seen in power line [2]; hence,
Hideaki Fujita [3] proposed an advance control scheme combing the two methods
available to damp the power fluctuation in transient state.
In this paper, an attempt is made to validate the result obtained by Hideaki Fujita,
with MATLAB simulation with the new value of additional feedback is presented
with validates the experimental result obtained by Hideaki Fujita.
2 UPFC Generalized Configuration
UPFC is basically a dual converter system which is required for realistic implementa-
tion of the concept so that real power easily flows in two ends though reactive power
can be changed in between the two converters. Using two voltage source converters,
the UPFC can be easily realized, with common DC link in between as shown in
Fig. 1; as shown in figure, the shunt converter is connected with grid by a shunt
transformer and the series converter is connected with grid by series transformer.
E c is the voltage to be injected in the line for controlling the real power. The two
Ec
Series
Transformer
Es Er
Shunt
Transformer Shunt Series
Converter Converter
AC DC
DC AC
Fig. 1 Simplified diagram of UPFC

Mitigation of Power Fluctuations in UPFC-Based Network 409
converters are connected with common DC line capacitor to keep the input voltage
at series converter constant.
The UPFC plays a vital role in optimizing the operation of the transmission line
system as it has multiple functions. The damping of the system will be boost by the
UPFC, and it is also used to improve the transient margins. The shunt converter will
control the bus voltage by absorbing or generating reaction power, and the series
converter will control the power flow by injecting a controllable series voltage via a
series transformer with the transmission line.
2.1 Operating Principle of UPFC
As we know that there are two numbers of the converter whichever are used in UPFC
as shown in Fig. 1; the function of these converters is given in detail in the following
1. Shunt converter: Shunt converter will provide or consume the real power which is
required by series converter at its common DC link which is to help the exchange
of real power which comes from the series voltage injection. Shunt converter can
also feedback the DC link voltage to AC line through shunt transformer, and shunt
converter can also be used to produce or consume reactive power compensation;
thus, an independent control can be achieved, and shunt converter is generally
functioned as STATCOM.
2. Series converter: The main function of UPFC is to inject AC voltage (E c ) whose
magnitude is controlled in the range 0 ≤ E c ≤ E c(Max) and phase angle ρ in (0 ≤ ρ
360) which is only possible using series converter. The voltage injection will be
done by series converter connected with transmission line through series trans-
former at power frequency. The voltage injected will be taken as a synchronous
AC voltage sources. There is an exchange of real and reactive power between
series converter and AC line due to transmission line current. At the output of
series transformer, there is an exchange of reactive power which is produced
within the converter. At AC terminal, there is an exchange of real power which is
converted into DC power which seems as positive or negative real power demand
at DC terminal. This series converter basically functions as SSSC
There is an independent generation of reactive power at AC terminal by each
converter and it is necessary to know that there is occurrence of closed—direct patch
for the negotiation of real power through the action of injection of series voltage
injection by shunt converter and series converter back to the line, and then, there
is corresponding generation of consumption of reactive power exchange locally by
series converter, and thus, it will not flow through the line. Therefore, shunt converter
will act at unity power factor or will be controlled to acquire reactive power exchange
with the line which is independent of the reactive power exchange by series converter.
410 R. Pudur et al.
2.2 Real and Reaction Power Control
Sending end voltage and phase angle:
δ
E s = Ee− jθ1 θ1 =
2
Receiving end voltage and phase angle:
δ
E r = Ee− jθ2 , θ2 =
2
The current phase is:
Es − Er
I = ,
Z
where Z = jX is the complex impedance of the line.
Receiving end complex AC power:
S = Er I ∗ (1)
P − j Q r = Er I ∗ (2)
Receiving end transmitted power P and reactive power −jQr:

∗
Es − Er
P − j Q r = Er (3)
jX
Substituting

δ δ
E s = E jδ/2 = E cos + j sin (4)
2 2

− jδ/2 δ δ
Er = E = E cos − j sin (5)
2 2
Since E s = E r = E
The final expression for real power transfer and reactive power transfer is:
E2
Po (δ) = sin δ
X
E2
Q 0r (δ) = − (1 − cos δ)
X
3 Proposed Set-up of UPFC
A proposed circuit of UPFC is shown in Fig. 2; it is observed that a UPFC is connected

between two sides of lines with voltage marked as E s and E r . Generally, the series
converter is operated as controllable voltage source whereas shunt converter as the
current source. By adjusting the bulk of real power in the line, the shunt converter is
utilized to control the DC link voltage and also control the reactive power. IGBT is
used to make series converter, and diodes are used to make shunt converter. A PWM
inverter is used to generate six gate pulses, and its AC terminal will be connected
via a matching transformer in series with the line, and a resistor and inductance are
used to show the impedance (Z) of the line.
It is shown that we have one resistor (R) and inductor (L) which is used to show
the impedance of transmission line. As the impedance of transmission line is much
larger than the resistance of the line, it is acceptable to discard that line resistance at
the overhead line.
Therefore, an expression for the line current phase will be given by,
Es − Er + Ec
I =
jω ◦ L
Single-phase current as shown in Fig. 3.

It is assumed here that sending end-side voltage, E s equal to the receiving end
side voltage, E r for the account of ease as shown in the figure. From Fig. 4a, active
power will be accomplished by adjusting the output voltage E c which will be led
Ec Z
Series
Es Transformer
Er
Shunt
Transformer Shunt Series
Converter Converter
AC DC
DC AC
Es
Control Circuit PWM
p q
Fig. 2 Proposed to set up of UPFC

412 R. Pudur et al.
Fig. 3 Single-phase Ec
equivalent circuit E
I R L
Es Er
a E b
E
Ec
Es Ec
δ
I
I Es
Fig. 4 Phasor diagram for a control of real power and b control of reactive power
by 90º as in relative to E s in result I will be in phase with E s . On the other hand,

by adjusting E c to be in phase with E s , current I will lags by 90º; then, it results in
control of reactive power flow shown if Fig. 4b. The above discussion is not required
because phasor vectors are used.
3.1 Phase Angle Control
In phase angle control, the control of real power will likely to be achieved by
controlling the amplitude of 90º leading or lagging output voltage.
Consider a d–q reference frame based on space vector in which q-axis voltage
E cq is utilized to control d-axis current id which is corresponding to real power.
∗
Thus, for the series device, the voltage reference vector E c∗ = E cd ∗
, E cq will be
given as,
∗
E cq = K p (i d∗ − i d )(K p = [V /A] is an real power feedback gain)
(i d∗ = real current reference frame)

So, E cq is used to control the phase angle between E m and E r which is given as

E cq
δ = tan−1
Es
3.2 Cross-Coupling Control
In this control which is comprised of both real power feedback loop and reactive
power feedback loop, through the control of both the magnitude of E m and phase
angle δ, this control scheme can be described. Hence, the cross-coupling control
empowers the UPFC to accomplish both control of real and reactive power, so on
d-q frame coordinates d-axis, voltage E cd is used to control q-axis current iq which
is related to reactive power. Thus, for the series device, the voltage reference vector
is given by
∗
E cd = −K q (i q∗ − i q )
∗
E cq = K p (i d∗ − i d )
where K q [V/A] reactive power feedback gain
4 Proposed Control Scheme
In this control scheme, for the series device, the voltage reference vector is given by
∗
E cd = K r (i d∗ − i d ) − K q (i q∗ − i q )
∗
E cq = K p (i d∗ − i d ) + K r (i q∗ − i q )
In this scheme, both the schemes cross-coupling and phase control scheme are
combined, with new power feedback gain K r which will perform like a damping
resistor across the current error.
Let us assume that the voltage of sending end will be the same as the voltage of
receiving end, i.e. E s = E r .
⎡ ⎤ ⎡ ⎤
E sa
2 cos θ 2π
⎣ E sb ⎦ = E s ⎣ cos θ − 3 ⎦ (6)
3
E sc cos θ + 2π 3
Here, θ = ω ◦ t, ω◦ = supply angular frequency
di a
E sa = Ria + L (7)
dt
414 R. Pudur et al.
di b
E sb = Rib + L (8)
dt
di c
E sc = Ri c + L (9)
dt
Now, applying d-q transformation in Eq. (6)
⎡ ⎤ ⎡ ⎤
E sa
2 cos θ 2π
⎣ E sb ⎦ = E s ⎣ cos θ − 3 ⎦
3
E sc cos θ + 2π 3
Apply to apply d-q transformation
di d
Ecd = Ri d + L − ω ◦ Li q
dt
di q
Ecq = Riq + L + ω ◦ Li d
dt
Then, instantaneous real and reactive power is given by
p = E sd · i q + E sq · i q = E s · i d
q = E sd · i q − E sq · i d = E s · i q
Taking the Laplace transform and assume initial condition i d (0) = i q (0) = 0,
then
R Id (s) − R Id (0+ ) + K r Id (s) − K r Id (0+ )

+ s L Id (s) − s L Id (0+ ) − ω0 L Iq (s)
− K q Iq (s) = K r Id∗ (s) − K q Iq∗ (s) As I d(0+ ) = 0
(R + K r + s L)I d(s) − (ω0 L + K q )Iq (s) = K r Id∗ (s) − K q Iq∗ (s)

(ω0 L + K p )Id (s) + (R + K r + s L)Iq (s) = K p Id∗ (s) + K r Iq∗ (s)
After solving the above equations.

(R + K r + s L)2 + (ω0L + K p )(ω0L + K q ) Id (s)
= {K r (R + K r + s L) + K p (ω0 L + K q )}Id∗ (s)
− {K q (R + K r + s L) − K r (ω0 L + K q )}Iq∗ (s)
2 2
s L + s(2R L + 2K r L) + (R 2 + K r2 + 2R K r

+ω02 L 2 + ω0 L K q + ω0 L K p + K p K q )Id (s)

= s K r L + (K r R + K r2 + ω0 L K p + K p K q ) Id∗ (s)
− {s K q L + (K q R + K q K r − ω0 L K r − K r K q )}Iq∗ (s)
Equation of the second-order system is s 2 + 2ς ωn s + ωn2
2(K r + R)
Here 2ς ωn =
L
⎡
⎤
(ω0 L + K p ) ω0 L + K q + (K r + R)2
Or, ς ⎣ ⎦ = (K r + R)
L L
Kr + R
Or, ς (Damping Factor) =
(10)
ω0 L + K p (ω0 L + K q ) + (K r + R)2
The equation for fluctuating frequency is given as,

ω = ωn 1 − ς 2

(K r + R)2
= ωn 1 −
L 2 ωn2

1 ω0 L + K p (ω0 L + K q ) + (K r + R)2
= L 2 − (K r + R)2
L L2

ω0 L + K p (ω0 L + K q )
ω= (11)
L
From Eq. (10), it can be observed that as larger the feedback gain K r , then there
will be ς close to unity. And from (11), there is an increase in K p and K q ; there
will be an increase in fluctuating frequency ω. The final control diagram is shown in
Fig. 5.
Fig. 5 Control scheme

416 R. Pudur et al.
5 Proposed Model of UPFC in MATLAB
A three-phase UPFC model is modelled in MATLAB software whose model is shown

in Fig. 6. In this model, two generating units are taken which is of same rating, i.e. of
440 V. UPFC has two transformers—one is shunt transformer and series transformer
where a series device is used to inject a voltage in series with the line. It consists of
two converters, i.e. AC– DC–AC converter which is least expansive. This converter
consists of the uncontrolled rectifier, and IGBT is used as voltage source inverter.
The DC voltage at series converter should remain constant. A PWM inverter is used
to give the gate pulse to the series inverter by which it will compare the carrier signal
with a reference voltage. Also, a three-phase RLC load is used in order to show its
effect in controlling real and reactive power.
5.1 Simulated Results
The model is run with the calculated value of additional power feedback gain one
with ignoring the value of gain and other including it; the result clearly distinguished
the active power and reactive power profile, and Figs. 7 and 8 show that the power
profile is successfully attenuated.
Fig. 6 Proposed UPFC model in MATLAB

Fig. 7 Simulated waveform without the proposed control scheme with K p = K q = 2.5 [V/A] K r
= 0 [V/A]
Fig. 8 Simulated waveform when the proposed control scheme is applied with K p = K q = 2.5
[V/A] K r = 4 [V/A]
418 R. Pudur et al.
6 Parameters Used
Parameters Ratings
Generating unit 440 V, 50 Hz
Rated power 10 KVA
Line reactance R = 0.04 , L = 1 mH
Shunt transformer ratio 1:1
Series transformer ratio 1:12
Controller parameters
Kp 2.5 V/A
Kq 2.5 V/A
Kr 4 V/A
RLC load P = 5 kW, QL = 800 VAR, Qc = 15,000 VAR
7 Conclusion
As here, a UPFC model is developed with a new control scheme where a new feedback
gain is proposed, K r . The proposed model is run without the implementation of the
value of new power feedback gain K r , and it has been observed that prominent
distortion or fluctuation in the real and reactive power waveform. The value of K r
was calculated and implemented in the same model. After proper inclusion of K r ,
it was observed that the unwanted fluctuation in real and reactive power has been
reduced tremendously. Hence, the proposed control system gives smooth control of
active and reactive power, and the same is tangibly advocated with all output graphs
and validates the findings of Hideaki Fujita.
References
1. Gyugyi L (1992) Unified power-flow control concept for flexible AC transmission systems.
In: IEE proceedings C (generation, transmission and distribution), vol 139, no 4. IET Digital
Library, Jul 1 1992, pp 323–331
2. Yu Q, Round SD, Norum LE, Undeland TM. Dynamic control of a unified power flow controller.
In: 27th annual IEEE power electronics specialists conference 1996, pp 508–514
3. Fujita H, Watanabe Y, Akagi H (1999) Control and analysis of a unified power flow controller.
IEEE Trans Power Electron 14(6):1021–1027
An Efficient Approach for Barcode
Encoding/Decoding Using Pattern
Substitution
S. Ieswaria, R. Bhavani, and R. Priya
Abstract The storage of data with reduced space and faster retrieval has always
been the main concern in every field. Barcodes ensure faster storage and retrieval of
data in a secured form. This paper enlightens the importance of barcodes in every
field of today’s world and a proposed new barcode generation method with a picture
embedded overcoming the traditional methods of storing information. Maintaining
details of patient, student, etc., have always been into confusion due to manual
maintenance. It would be a revolutionary change in maintaining patients’ detail using
barcodes in the form of a person’s picture avoiding the manual error.
Keywords Barcodes · Picture-embedding · Modulation · Demodulation ·

Encoding · Decoding
1 Introduction
Manual records maintenance has always been a hectic task; apart from its difficulty
in maintaining large amount of information and faster retrieval, it was also prone to
unacceptable human error. Medical records include a variety of documentation of the
patient’s history, clinical findings, diagnostic test results, preoperative care, operation
notes, postoperative care and daily notes of a patient’s progress and medications.
Student database with details is like admission, fees, marks, arrears, and courses.
Barcodes are of two types: (a) one-dimensional(1D) barcodes and (b) two-
dimensional(2D) barcodes (quick response (QR) codes and as an extension of QR
codes, an additional property of color). The one-dimensional barcodes though being
the traditionally used one for a long period, QR codes took over every field since it
S. Ieswaria (B) · R. Bhavani · R. Priya

Department of Computer Science and Engineering, Annamalai University, Chidambaram, India
R. Bhavani
R. Priya
420 S. Ieswaria et al.
is seen as an advancement of the former one with the capacity to store more data,
requires less space, enabling it to be read from all angles and also with the advan-
tage of including more languages than 1D barcodes such as numeric, alphanumeric,
kana(Japanese character), and kanji(Chinese character).
The mapping between the barcodes and the message stores into it is called
symbology. Barcodes though being classified broadly as 1D and 2D, various barcode
symbologies exist like Code39, Code128, Codabar, Postnet, PDF417, Datamatrix,
Maxicode, Aztec, and QR code.
Yang et al. [1] proposed an artcode with a dynamic color pallet by error diffusing
algorithm and also performed shuffling to embed encoded modules and bits into color.
Karthikeyan and Bender [2] introduced the embedding of molecular structures into
QR codes by converting them into SMILES and compressing by Lempel–Ziv–Welch
(LZW) and automatic chemical structure (ACS) for inventories in pharmaceutical
laboratories. Gaikwad et al. [3] discussed embedding the image into barcodes and
hiding information into them.
Divya et al. [4] proposed a technique to enhance security with the help of holog-
raphy to recognize items. van der Westhuizen and Chai [5] introduced the ability
to capture dynamic barcodes obtained from frames and produced the highest trans-
mission rate of 622 bytes/s and 1% error for 720p. They have suggested ways in
which the work could be extended like fault finding and amendment, live decoding,
efficient data ordering, and more colors. Abdelnour et al. [6] combined barcode and
radio frequency identification technology. Lee et al. [7] proposed a method to encode
voice into binary digits, embedding them into barcode and vice versa achieved the
lowest execution time of 18 ms for complex words.
This paper is organized by starting with an introduction to the problem and
barcodes in Sect. 1 and followed by brief research works previously done on barcodes
in Sect. 2 and proposed image embedded barcode model in Sect. 3. Section 4 provides
the experimental results, and Sect. 5 concludes the work.
2 Literature Review
Various researchers have made continuous approaches to overcome the time and
space complexity problem that is prevailing in the environment in storage and
retrieval of data. The concept of barcodes has gone various domain-specific
improvements to suit individual application requirements.
Chen et al. [8] proposed modulation method and demodulation method for Picode,
which is a remarkable approach to embed a message into QR codes resembling the
picture of the content that the barcode holds, achieving highest average successful
detection probability (SDP) of 0.9964 and normalized data capacity (NDC) of
0.959. Chandran et al. [9] enhanced the Picode proposed before by amending
the encoding/decoding algorithm with Bose, Chaudhuri, and Hocquenghem (BCH)
algorithm, thereby strengthening the security.
An Efficient Approach for Barcode Encoding/Decoding … 421
Umaria and Jethava [10] discussed increasing the storage capacity of QR code
by compressing and multiplexing, thereby producing it in 1 Kb and also providing a
hint to extend it to audio and video files. Liu et al. [11] proposed an encoding method
using pattern substitution for 2D barcodes. Chen et al. [8] proposed new modulation
and demodulation techniques for Picode (picture-embedded barcode) and Vicode
(video-embedded barcode).
Chu et al. [12] introduced a new encoding and decoding technique for barcodes
that are visually relatable for humans and providing efficiency in terms of reliability
and regularization with pattern assignment. Wasule and Metkar [13] proposed an
encoding method for generating 2D barcodes in color with a better data capacity of
grayscale QR codes to ‘n’ fold.
Xu et al. [14] proposed a barcode scanning method based on module scanning
probability estimation model for binary, grayscale, and color images. Dinesh and
Sharma [15] proposed a framework where three individual barcodes were displayed
in R, G, and B channels, thus producing a color barcode with increased storage
capacity.
Chen et al. [16] introduced a lightweighted encoding and decoding algorithm
for generating color barcodes and also brought forth the importance of barcodes in
M-commerce. Gabor Soros et al. [17] introduced a method that located 1D and
2D barcodes on smart phones. Garateguy et al. [18] optimized QR codes with
halftoning techniques with multi-scale–structure similarity index measure (MS-
SSIM) of 0.5590. Yu Zhang and Tong Lu proposed a model to detect color barcodes
with an efficiency of 88.9%.
3 Proposed Method
This section gives the procedure followed in encoding and decoding the picture-
embedded barcodes for binary images.
3.1 Encoding Barcode
The binary images that are used to encode are set for a uniform size of 126*126. The
binary image is divided into equal size blocks. On the other hand, the message data
undergoes channel encoding to avoid data loss/corruption. The encoded message
along with the image blocks is passed for the modulation technique being proposed.
Then finder pattern is applied to the modulated image in case of the binary image,
otherwise passed through an intermediary binary to grayscale conversion. The final
result giving the picture-embedded barcode. Figure 1 shows the block diagram of
the encoding procedure done in this work to produce the picture-embedded barcode
when provided with the picture and the message to be encoded.
Dividing
blocks
Applying
Modulation Finder
pattern
Channel
Encoding
Fig. 1 Block diagram of barcode encoding process
3.1.1 Image Data
The original images are taken in the context from standard image processing picture,
real-time personalities, and logo which are divided into blocks of 1*2 for modulation
procedure. The message to be encoded undergoes channel coding for error correc-
tion. The encoding algorithm used here is Reed–Solomon encoding which is being
traditionally used for channel coding [19]. The encoded data stream is then sent for
modulation with the image.
3.1.2 Proposed Modulation Procedure
Figure 2 depicts the procedural steps of the modulation process proposed in this paper.
The proposed modulation process acts as the procedure for combining the image and
message in a manner controlling visual distortion of the image and protecting the
message stored in the barcode. The modulation process undergoes through three
steps: pattern counter, pattern assignment, and pattern replacement. The human eye
is sensitive to sharp changes; taking account of this, the images are processed in its
binary form, and also, the block size is taken as per the intermediary pattern size. As
per the block size, 4 patterns are analyzed to be repeated throughout the image, which
is defined as X 1 , X 2 , X 3, and X 4 . In the pattern counter step, the number of times the
4 patterns are being repeated is stored as n(X i ) where, i = 1, … 4. The result of the
pattern counter gives the pattern enumeration of X 1 to X 4 . Based on this, the pattern
assignment step assigns the highest repeating pattern as O and the second-highest
repeating pattern as Z, in order to maintain the visual responsiveness and followed
Fig. 2 Proposed flow diagram of modulation process

Fig. 3 1*2 patterns defined

in the binary image
X1 X2
X3 X4
by which the pattern replacement takes place according to the following:

O, Mk == 1
CB
Z , Mk == 0
where
CB represents current block,
M k represents the kth element in the message data.
Figure 3 depicts the 1*2 block patterns X 1 , X 2 , X 3, and X 4 that are being repeated
throughout the image. The traditional pattern assignment is refined in order to main-
tain the visual similarity of the original image and the encoded barcode by not
completely encoding all the blocks in the image. The encoding method is done
by encoding every nth block according to the message bits. Thus, this modulation
method acts as the new symbology that maps the encoded data to the message encoded
into the barcode.
3.1.3 Barcode Generation
The modulated image is directly sent to the finder pattern generation consisting of
the original image as binary. The finder pattern applied for the barcode is as followed
in the data matrix, to leave the generated picture-embedded barcode undisturbed.
The finder pattern is in the form of ‘ ’-shaped along the left and down and broken
lines or dots along the right and up, acting as the alignment and timing pattern. Thus,
the barcode has been generated and MS-SSIM was evaluated.
3.1.4 Barcode Encoding Algorithm
Step 1: Input message x

Step 2: Generate RS code for X
Step 3: Input image I
Step 4: Divide image blocks with block size 1*2
Step 5: Define 4 patterns and obtain the value from pattern counter
Step 6: Assign, O=maximum repeated pattern
Z= second maximum repeated pattern
Step 7: For alternate blocks, replace current Block,
Where, Mk is the kth element in encoded message
Step 8: Generate finder pattern
Step 9: Calculate MS-SSIM
3.2 Decoding Barcode
The barcode generated undergoes the proposed demodulation procedure in accor-

dance with the modulation proposed as shown in Fig. 4. The encoded image is
extracted from the barcode that consists of potential regions; i.e., the finder pattern
and the timing pattern that was applied to the encoded image during barcode gener-
ation. The image is then divided into a moving window of size 1*2 with respect to
the modulation procedure.
The demodulation contains the same steps as in the modulation with pattern
counter, pattern assignment, and pattern replacement. Figure 5 shows the result
of pattern counter that computes the number of times each pattern defined being
repeated. But the pattern replacement step acts as a reversed process of barcode
encoding pattern replacement. The demodulation takes place by pattern counting
and assigning the highest and second-highest repeating patterns as O and Z. Here
the pattern replacement in order to obtain the encoded message is carried by,

1, CB == O
Enc
0, CB == Z
Fig. 4 Block diagram of barcode decoding process

Fig. 5 Result of pattern 6567

7000
counter for panda image Pattern Count
6000
Repeat count
5000
4000
3000
2000 1143
1000 113 115
0
X1 X2 X3 X4
Patterns
where
Enc represents the encoded message.
The decoding also takes place by considering the nth block of the encoded message
as it was taken during encoding. The encoded message is then decoded by the Reed–
Solomon decoding algorithm giving the message that was encoded into the picture-
embedded barcode.
3.2.1 Barcode Decoding Algorithm
Step 1: Input barcode image

Step 2: Remove finder pattern
Step 3: Input image I
Step 4: Divide image blocks of block size 1*2
Step 5: Define 4 patterns and obtain the value from pattern counter
Step 6: Assign, O=maximum repeated pattern
Z= second maximum repeated pattern
Step 7: For alternate blocks, replace encoded message rsmsg(i,j)
Where, CB is the current block in the barcode image
Step 8: Decode rsmsg and obtain the message
4 Experimental Results
The model was implemented using MATLAB r2016a on Intel corei5 processor-based
personal computer with 4 GB RAM. The images taken for encoding were binary with
an uniform size of 126*126 for pattern considered for modulation. The images are
divided into a minimum rectangular block size of 1*2 for which the patterns possible
are shown in the figure. The proposed modulation procedure is carried out with the
main aim to maintain the similarity to the original image as close as possible and at
the same time securing the data encoded into the barcode, which is measured using
MS-SSIM.
Figures 6a, 7, 8, and 9a show the original binary images sample taken and Figs. 6b,
7, 8, and 9b show the barcodes obtained for the original images, respectively, with
MS-SSIM. Figure 10 compares the resulting MS-SSIM produced by the previous
related research works, producing MS-SSIM of 0.46 for Halftone QR, 0.84 for QR
image, and 0.49 for Picode with the proposed work which produces a result of 0.99.
The proposed model was proved to produce higher MS-SSIM for binary images.
Fig. 6 a Original binary

image b barcode with
MS-SSIM = 0.9648

MS-SSIM = 0.9989

MS-SSIM = 0.9755

MS-SSIM = 0.9271
Fig. 10 Comparison results 0.99

of MS-SSIM of proposed 1 0.84
and previous works
0.8
0.49
MS-SSIM
0.6 0.46
0.4
0.2
0
Halftone QR QR image Picode Proposed
5 Conclusion
The binary image was thus embedded with the encoded message through the proposed
modulation procedure, and the resulting barcode was proved to produce high MS-
SSIM in reference to the original image. The message was successfully decoded
using the proposed demodulation procedure from the generated barcodes. Thus,
facilitating data retrieval of databases is like medical, student, product, employee in
the most efficient manner overcoming all traditional methods and manual errors.
References
1. Yang Z et al (2016) ARTcode : preserve art and code in any image. UBICOMP ’16, pp 904–915
2. Karthikeyan M, Bender A (2005) Encoding and decoding graphical chemical structures as
two-dimensional (PDF417) barcodes. J Chem Inf Model 45:572–580
3. Gaikwad AM (2015) Halftoning Technique. In: International conference on innovations in
information, embedded and communication systems ICIIECS’15, vol 2
4. Divya PS, Sheeja MK (2013) Security with holographic barcodes using computer gener-
ated holograms. In: 2013 international conference on control communication and computing.
Computers and security, pp 162–166
5. Van Der Westhuizen Y, Chai D (2018) 3D dynamic barcode. Electron Lett 54(17):1024–1026
6. Abdelnour A, Kaddour D, Tedjini S (2018) Transformation of barcode into RFID. IEEE Microw
Wirel Compon Lett 28(5):398–400
7. Lee JR, Ruan S, Lin CH (2016) VoiceCode: a 2D barcode system for digital audio encoding.
IEEE Glob Conf Consum Electron VoiceCode 5:2–3
8. Chen C, Huang W, Zhang L, Mow WH (2019) Robust and unobtrusive display-to-camera
communications via blue channel embedding. IEEE Trans Image Process 28(1):156–169
9. Chandran V, Sekhar A (2017) A secure and reliable channel error correction technique for
picode. In: 2017 IEEE international conference on electrical, instrumentation and communi-
cation engineering (ICEICE) 2017, pp 1–4
10. Umaria MM, Jethava GB (2015) Enhancing the data storage capacity in QR code using
compression algorithm and achieving security and further data storage capacity improve-
ment using multiplexing. In: 2015 international conference on computational intelligence and
communication networks, CICN 2015, pp 1094–1096
11. Liu X, Luan R, Huang C (201) A novel encoding method for visual two-dimensional barcode
using pattern substitution. In: 2010 international conference on machine vision and human-
machine interface, MVHI 2010, pp 662–665
12. Chu HK, Chang CS, Lee RR, Mitra NJ (2013) Halftone QR codes. ACM Trans Graph 32(6):1–8
13. Wasule S, Metkar S (2017) Improvement in two-dimensional barcode. Sādhanā 42(7):1025–
1035
14. Xu M, Li Q, Niu J (2015) ART-UP: a novel method for generating scanning-robust aesthetic
QR codes. J Latex Files 14(8):1–14
15. Dinesh K, Sharma G (2016) Per channel color barcodes for displays. IEEE, ICIP, pp 3992–3996
16. Chen B (2013) A lightweight color barcode algorithm and application in mobile E-commerce.
Int Conf Comput Sci Educ 8:805–809
17. Soros G (2014) GPU-accelerated joint 1D and 2D barcode localization on smartphones. In:
2014 IEEE international conference on acoustics, speech and signal processing, pp 5095–5099
18. Garateguy GJ, Arce GR, Lau DL, Villarreal OP (2014) QR images: optimized image embedding
in QR codes. IEEE Trans Image Process 1:1–11
19. Chen C, Huang W, Zhou B, Liu C, Mow WH (2016) PiCode: a new picture-embedding 2D
barcode. IEEE Trans Image Process 25(8):3444–3458
Fabrication and Electrochemical
Modeling of CNT-Based BioFET
for Cholesterol Detection
Gaurav Keshwani, Kabyashree Hazarika, Hiranya Ranjan Thakur,

and Jiten Chandra Dutta
Abstract In this work, a high sensitivity BioFET has been fabricated and modeled
for cholesterol detection in serum samples. The device consists of indium tin oxide
(ITO) coated glass as substrate, high κ-dielectric (ZrO2 ~ 25) as gate insulator, PEI-
doped MWCNT as transporting layer and K/PPy/CNT nanocomposite as sensing
membrane. Alkali metal potassium (K) and conducting polypyrrole have been used
to enhance the conductivity and dispersion property of the nanocomposite in order
to get better response time and detection limit. The layers have been deposited by
electrochemical deposition and spin coating technique. The morphology and charac-
teristics of the deposited layers were characterized using Fourier transform infrared
spectroscopy (FTIR), field emission scanning microscopy (FESM), scanning elec-
tron microscopy (SEM), X-ray diffraction (XRD), and cyclic voltammetry (CV).
The modeling has been done using the enzymatic reactions between cholesterol
substrate and cholesterol oxidase enzyme, site binding theory, pH dependency of
surface potential, and current transport model.
Keywords Cholesterol · BioFET · Cholesterol oxidase · CNT · Modeling
1 Introduction
Cholesterol is an important lipid present in human cells majorly produced in the

liver [1]. The level of cholesterol helps in the determination of diseases such as
arteriosclerosis, myocardial infarction, hypertension, and coronary heart diseases
G. Keshwani (B) · K. Hazarika · H. R. Thakur · J. C. Dutta

Tezpur University, Tezpur, Assam 784028, India
K. Hazarika
H. R. Thakur
J. C. Dutta
430 G. Keshwani et al.
[2]. At prescribed level, cholesterol is essential for proper metabolism of body but
high concentration of cholesterol in blood serum can be life-threatening. Enzymatic
methods offer better selectivity and specificity than analytical methods reported in
the literature [3]. Cholesterol oxidase (ChOx) has emerged as an important tool for
detection and measurement of cholesterol level in serums [4]. Few enzyme-based
biosensors for cholesterol detection have been reported in past few decades but due to
shortcomings such as short life span, high power consumption, narrow linear range,
and sophisticated fabrication process could not be commercialized [5]. Thus, there
is a need for a reliable, sensitive, and inexpensive biosensor for routine monitoring
of cholesterol [6].
FET-based biosensors have gained a lot of attention due to its high amplification
capability and possibility of ‘on chip fabrication.’ Traditional Si-based FET devices
have drawbacks such as low sensitivity, scaling limitation, less carrier mobility, and
incompatibility with high-κ dielectrics [7]. These drawbacks can be overcome to
a large extent with the use of nanomaterials such as carbon nanotubes (CNT) [8–
10]. With the use of nanomaterials, these devices can be used for the detection of
enzymatic reactions, proteins, and other micro-molecules [11]. Moreover, CNT has
high compatibility with high-κ dielectrics such as zirconia and hafnia which helps in
further miniaturization of the device. High-κ dielectrics reduce tunneling current to
a large extent even at nanometer thickness which makes it suitable as gate insulator
and biosensor applications.
2 Experimental
2.1 Apparatus Used
For electrochemical deposition and layer characterization, AUTOLAB

PGSTAT128N has been used. Spin coating has been done using SPIN NXG
P28 spin coater. REMI RM centrifuge and SONAPROS PR-1000MP probe soni-
cator were used for sonication and centrifugation of prepared solutions. The samples
were vacuum dried and heated using LABFREEZ vacuum oven.
2.2 Solution Preparation
Multi-walled carbon nanotubes (MWCNTs) with carboxyl group functionaliza-

tion, zirconium (IV) tetrachloride (ZrCl4 ), cholesterol oxidase (ChOx) enzyme,
polyethylenimine (PEI) with average molecular weight Mn ~5000 g/mol, and
polypyrrole (PPy) have been purchased from Sigma-Aldrich. All the other chemicals
and reagents used were of analytical grade (Fig. 1).
Fabrication and Electrochemical Modeling of CNT-Based … 431
Fig. 1 Solutions for deposition of (a) top gate and bottom gate insulating layer (b) transporting
layer (c) sensing layer
2.3 Fabrication
The high-κ dielectric cholesterol BioFET is fabricated on an ITO-coated glass

substrate having sheet resistance ~50 ohm/cm2 . The layers have been deposited
by solution process. High-κ zirconia layer and PEI-doped CNT channel have been
deposited by electrochemical deposition technique using three electrode systems. The
three electrode systems consist of Ag/AgCl reference electrode (RE), platinum wire
counter-electrode (CE), and the deposition surface as the working electrode (WE).
K/PPy/CNT sensing layer is deposited by spin coating technique. The various steps
involved in fabrication of the BioFET are shown in Fig. 2. Gravimetric analysis is used
to calculate the thickness of deposited layers. ChOx is covalently immobilized on
carboxyl group of functionalized CNT by using EDC and NHS as crosslinking acti-
vators [5]. The deposited layer is washed using PBS solution to remove unbounded
ChOx, dried for 12 h, and kept at temperature as low as 2–4 °C when not in use. The
process of covalent immobilization of cholesterol oxidase on functionalized CNT
is shown in Fig. 3. K/PPy/CNT layer facilitates the fast transfer of charge between
reaction site and the oxide layer.
Fig. 2 Steps for fabrication and characterization of cholesterol BioFET
Fig. 3 Covalent immobilization of ChOx on MWCNT
2.4 Results and Discussions
The oxidation and reduction potential of the ceramic and nanocomposite were found
out using cyclic voltammetry. Figure 4a, b shows the cyclic voltammetry plots for
zirconia (ZrO2 ) and PEI-doped CNT layer, respectively. The layers were deposited
using chronoamperometric process. Zirconia was deposited keeping working elec-
trode (WE) potential at −1 V (Fig. 4c) and PEI-doped CNT at WE potential at +
0.8 V (Fig. 4d). The deposition time was optimized by trial and error method.
Fig. 4 Autolab waveforms showing (a) cyclic voltammetry of zirconia layer (b) cyclic voltam-
metry PEI-doped CNT layer (c) chronoamperommetry process for deposition of zirconia layer (WE
potential = −1 V), (d) chronoamperommetry process for deposition of PEI-doped CNT layer (WE
potential = 0.8 V)
2.5 Transport Layer Characterization
Fourier transform infrared spectroscopy (FTIR) and scanning electron microscopy

(SEM) analysis have been used for chemical characterization of PEI-CNT transport
layer. The peaks at 1750 and 3450 cm−1 may be attributed to CO and N–H stretching
mode [12]. FESM results show uniform deposition but agglomeration can be seen at
few sites. The FTIR and FESM results are shown in Fig. 5a, b.
Fig. 5 (a) Fourier transform infrared spectroscopy and (b) field emission scanning microscopy of
PEI-doped CNT layer
Fig. 6 Characterization of zirconia layer showing (a) XRD results (b) SEM image [13]
2.6 Oxide Layer Characterization
Scanning electron microscopy (SEM) results of deposited zirconia layer give infor-
mation about the uniform deposition and fine particle size ranging ~40–90 nm. Peaks
at 2θ = 28°, 35°, and 50° show deposition of tetragonal phase and peaks at 2θ =
32° and 45° evidences monoclinic phase of zirconia [13]. The patterns are semi-
amorphous containing a mixture of t–ZrO2 and m–ZrO2 . The XRD and SEM images
are shown in Fig. 6a, b.
3 BioFET Modeling
BioFET works on the basic principle of pH variations at the oxide surface of ISFET.
The modeling has been done using the acid/base reactions, enzymatic reactions, and
the pH detection properties of ISFET [14]. The enzymatic reactions are shown in
Eqs. (1) and (2). The reactions between substrate and enzyme release proton, which
alters the surface potential and modulates drain current of the BioFET device.
Cholesterol + O2 ChOx Cholest - 4 - en - 3 - one + H2 O2 (1)

→
H2 O2 Electode O2 + 2H+ + 2e− (2)

→
Michaelis–Menten equation (Eq. 3) is used to model the enzymatic reaction which

is responsible for cholesterol oxidation.
[S]
a = aM (3)
[S] + K M
Michaelis constant K M is calculated using Lineweaver–burk plot and found out to

be 2.1 mM. [S] is the cholesterol concentration in PBS solution, and aM is the maximal
enzyme activity of ChOx enzyme. The pH value can be obtained by calculating the
hydrogen ion concentration released in the solution using Eqs. (2) and (4).

pH = − log H+ (4)
Substrate and product concentration change with time. Decrease in substrate

concentration increases the product concentration and the H+ ion concentration
resulting in decrease of pH in the solution. Variation of pH with substrate and product
concentration is shown in Fig. 7 (Table 1).
Fig. 7 Variation of pH with (a) substrate concentration, (b) product concentration

Table 1 Values taken for

S. No. Parameter Values for modeling
simulation of the device
1. Maximal enzyme activity aM 16.68 × 10−9
2. Michaelis–Menten constant 2.1 mM
3. Channel length 1 mm
4. Reference voltage 0.2 V
5. Acid dissociation constant 2.24 × 10−12
6. Substrate concentration 0–20 mM
4 Conclusion
In this work, CNT-based high sensitivity cholesterol BioFET has been fabricated
and modeled. The electrochemical modeling of the device is in good agreement with
the practical results obtained from electrical characterization of the device. Bonding
structure between nanocomposites has been proposed in the paper. The deposited
layers have been electrically and chemically characterized. The chemical characteri-
zation results of nanocomposite layer and oxide layer are in good agreement with the
proposed structure. The thickness of the deposited layers is calculated using gravi-
metric technique and found to be in nanometer range. The use of carbon nanotubes
in nanocomposites resulted in a better adhesion between the layers. FTIR, SEM,
FESM, and CV results obtained for ZrO2 oxide layer and PEI-MWCNT transporting
layer validates uniform deposition and proposed structure.
Acknowledgements The authors would like to thank the Deity, MCIT, Government of India, for
the support through Visvesvaraya PhD scheme and Tezpur University for providing laboratory
facilities.
References
1. Narwal V, Deswal R, Batra B, Kalra V, Hooda R, Sharma M, Rana JS (2019) Cholesterol

biosensors: a review. Steroids 143:6–17
2. Yıldırımoğlu F, Arslan F, Çete S, Yaşar A (2009) Preparation of a polypyrrole-
polyvinylsulphonate composite film biosensor for determination of cholesterol based on
entrapment of cholesterol oxidase. Sensors 9(8):64356445
3. Ansari AA, Kaushik A, Solanki PR, Malhotra BE (2009) Electrochemical cholesterol sensor
based on tin oxide-chitosan nanobiocomposite film. Electroanalysis Int J Devoted Fundam
Pract Aspects Electroanalysis 21(8):965–972
4. Vrielink A, Ghisla S (2009) Cholesterol oxidase: biochemistry and structural features. FEBS
J 276(23):6826–6843
5. Keshwani G, Thakur HR, Dutta JC (2019) Fabrication and electrical characterization of carbon
nanotube based enzyme field effect transistor for cholesterol detection. In: TENCON 2019–
2019 IEEE region 10 conference (TENCON). IEEE, pp 1821–1824
6. Ibupoto ZH, Elhag S, Nur O, Willander M (2014) Fabrication of sensitive potentiometric

cholesterol biosensor based on Co3O4 interconnected nanowires. Electroanalysis 26(9):1928–
1934
7. Barik MA, Deka R, Dutta JC (2015) Carbon nanotube-based dual-gated junctionless field-effect
transistor for acetylcholine detection. IEEE Sens J 16(2):280–286
8. Dutta JC, Thakur HR, Keshwani G (2019) High-performance dual-gate carbon nanotube ion-
Sens J 19(14):5692–5699
9. Salaets N, Ervin M (2009) Solution deposition methods for carbon nanotube field-effect tran-
sistors (No. ARL-TR-4854). Army research lab Adelphi MD sensors and electron devices
directorate
10. Dürkop T, Getty SA, Cobas E, Fuhrer MS (2004) Extraordinary mobility in semiconducting
carbon nanotubes. Nano Lett 4(1):35–39
11. Dutta JC, Sharma PK (2018) Fabrication, characterization and electrochemical modeling of
CNT based enzyme field effect acetylcholine biosensor. IEEE Sens J 18(8):3090–3097
12. Dutta JC, Thakur HR, Keshwani G (2019) High performance dual-gate carbon nanotube ion-
Sens J
13. Keshwani G, Thakur HR, Dutta JC Characterization of thin zirconia films deposited by ECD
on ITO coated glass for biosensing applications. In 6th international conference on signal
processing and integrated networks, Amity University, Noida, Uttar Pradesh
14. Sharma PK, Dutta JC, Barik MA, Sarma MK (2017) Numerical modeling of potassium-doped
polypyrrole/carbon nanotube graphene-based cholesterol enzyme field effect transistor. Int J
Numer Model Electron Netw Devices Fields 30(6):e2223
Design of LCL Filter for SPWM Inverter
Based on Switching Frequency and THD
Criterion
T. Saha, A. Mitra, and B. Halder
Abstract Over the time, power generation from renewable energy sources and its
grid integration is being increased rapidly. Subsequently, the quality of power has
been degraded because of injected harmonic contents to the grid from large number of
power electronic devices required for inverter applications. An optimized LCL filter
can improve the quality of power for single-phase inverter connected with the utility.
This perspective motivates the authors to design an LCL filter with optimum perfor-
mance under the constraint of least switching frequency variation, total harmonic
distortion (THD), size of the inductors, and power factor of the grid current. All the
results have been obtained through MATLAB/Simulink environment in justification
with the optimum satisfactory design.
Keywords SPWM inverter · LCL filter · Switching frequency · THD
1 Introduction
The upward global energy demand as well as saving the consumptions of fossil
fuels is manipulating to drive renewable energy sources to meet the energy demand.
Among them, the use of solar photovoltaic (PV) energy is most popular because
of its several advantages. The energy is available in the form of DC source and
requires converting through an inverter of suitable power rating as per demand by
the consumers. Also due to non-sinusoidal AC output from the inverter, proper design
of filter circuit is highly essential to make it sinusoidal to avoid unnecessary energy
loss of the domestic appliances. On the other hand, when the PV inverter system is
integrated with grid, the filter will provide a better power quality.
T. Saha · A. Mitra (B) · B. Halder

Department of Electrical Engineering, Narula Institute of Technology, Kolkata, India
T. Saha
B. Halder
440 T. Saha et al.
To lower the total harmonic distortion (THD) of the inverter current, LCL filter is
most popular nowadays in comparison with L filter and LC filter [1]. Proper selection
of the values of inductor and capacitor for an LCL filter is an important criterion
for the betterment of inverter output current and grid current to maintain improved
quality of power fed to the grid [2, 3]. The magnitude of current as well as active and
reactive power of the inverter can be controlled through a closed-loop controller in
comparison with a reference signal [4–6], where the THD of the current cannot be
controlled without LCL filter. In the presence of the passive components of the filter,
a small resistance is to be inserted in series with the filter capacitance for proper
damping of the filter. Moreover, due to power losses caused by the passive elements
of the filter, the introduction of active damping in the circuit is much preferable [7].
The THD can be controlled directly from the inverter through proper switching
topology at higher switching frequency. The sinusoidal pulse width modulation
(SPWM) and hysteresis band control technique can be a solution to lower the THD
[8, 9].
In addition to this, the size of the filter will be lowered, but the switches of the
inverter will experience high switching stress and high switching loss accordingly.
Though this can be reduced by applying soft-switching technique, the circuit overall
switching circuit will be more complicated. This proposed work makes a compromise
between the switching frequency of the inverter and the size of the LCL filter with
THD of the grid current.
2 Single-Phase SPWM Inverter
Among different switching topologies of inverters, sinusoidal pulse width modulation

(SPWM) methodology is widely used because of its several advantages and low
level of harmonic contents at its output. The output voltage from the inverter can be
controlled by varying modulation index which is defined as the ratio between the
peak of the modulating signal and carrier signal, and therefore,
v̂sin
ma = , (1)
v̂tri
where v̂sin and v̂tri are the peak value of modulating sine wave and the peak value of
carrier triangular wave, respectively. Since output voltage of the inverter is defined
by the peak value of the modulating signal, the fundamental voltage from the inverter
will accordingly be
m a Vdc
Vi1 = √ (2)
2
Figure 1 shows the development of the SPWM signal through the comparison of
a sinusoidal signal with a triangular signal. Figure 2 is the enlarged view of a section
Design of LCL Filter for SPWM Inverter Based … 441
Fig. 1 Development of SPWM signal
Fig. 2 Enlarged view of

sine-triangle comparison
of Fig. 2 to understand the relationship between the duty ratio and the modulation
index.
3 LCL Filter Modeling
In general, the output of the inverter is not a sinusoidal and contains several
harmonics. Figure 4 shows the circuit diagram of the LCL filter, which can be
governed by the following equations
442 T. Saha et al.
di i
Li = vi − vc − (Ri + Rf )i i + Rf i g (3)
dt
di g
Lg = vc − vg + Rf i i − Rf + Rg i g (4)
dt
dvc ii − ig
= (5)
dt Cf
where vi is the inverter-side AC voltage, vg is the grid-side AC voltage, vc is the

voltage across capacitor, Ri is the internal resistance of the inverter-side inductor
L i , Rg is the internal resistance of the inverter-side inductor L g , Cf is the filter
capacitance, and Rf is the damping resistance.
The state matrix of the filter can be expressed as
⎡ ⎤ ⎡ ⎤⎡⎤ ⎡ 1 ⎤
di i − Ri L+R f Rf
Li
− L1i ii 0

1 ⎥ vi
Li
⎦=⎢ 1 ⎥⎣ ⎢
dt i
Rf +Rg
⎣ di g
⎣
Rf
− Lg ⎦ + −
dt Lg Lg ⎦ i g ⎣ 0 Lg ⎦ v (6)
g
dvc
dt
1
Cf
− C1f 0 vc 0 0
which is expressed as ẋ = Ax + Bu. The transfer function of the filter circuit can
be obtained by the following equation.
ig
= C[s I − A]−1 B (7)
vi
where
⎡ ⎤ ⎡ ⎤
− Ri L+R f Rf
− L1i 0 1
⎢ i Li
Rf +Rg 1 ⎥ ⎢ 1 ⎥
Li
A=⎣ Rf
Lg
− Lg L g ⎦; B = ⎣ 0 − L g ⎦ and C = 0 1 0
1
Cf
− C1f 0 0 0
Neglecting the value of the internal resistance of the inductors, i.e., considering
Ri = Rg = 0, the transfer function of the system can be written as
ig Rf C f s + 1
= (8)
vi s L i L g C f s + Rf C f L i + L g s + L i + L g
2
The characteristic equation of Eq. (10) is written as

1 1 1 1 1
s + Rf
2
+ s+ + =0 (9)
Li Lg Cf L i Lg
Comparing the equation with s 2 + 2ξ ωn s + ωn2 = 0, the damping ratio can be

estimated as

1 1 1
ξ = Rf C f + (10)
2 Li Lg
4 Design of Filter
To design LCL filter, several characteristics of the filter have to be considered, such
as, current ripple, switching ripple attenuation, size of the filter, etc. The damping
factor of an LCL filter is an essential criteria, and thus, a passive damping must
be included by inserting a resistance in series with the capacitance as indicated in
Eq. (10).
The parameters required to design the filter circuit are the DC link voltage Vdc ,
fundamental voltage output of the inverter Vi1 , rated active power Pn , switching
frequency f s , resonant frequency f r , and grid frequency f g .
The base impedance and the base capacitance are defined by the following
equations.
Vi21
ZB = (11)
Pn
and
1
CB = (12)
ωg Z B
The filter capacitor is to be designed considering maximum variation of power

factor at the grid side which is 5%, resulting the value of the filter capacitance will
be Cf = 0.05CB . If required the value of the filter capacitance may be considered
higher that 5% of the base capacitance in order to compensate the inductive reactance
of the filter.
The ripple current of the inverter-side inductor i i can be estimated as
i i
Vdc − Vo = L i (13)
DTs
or,
D × (Vdc − Vo )
i i =
L i × fs
or,
D(1 − D)Vdc
i i = (14)
L i fs
444 T. Saha et al.
Fig. 3 LCL filter circuit

diagram
where Vo is the grid-side voltage referred to the DC side and is simply can be
considered as DVdc . Since the output voltage waveform of the inverter is the reflection
of the duty cycle varied sinusoidally as shown in Figs. 2 and 3, the ripple current can
be referred to the AC side can be written using Eq. (14) as
m a (1 − m a )Vdc
i i = (15)
L i fs
It is evident from Eq. (15), that the maximum ripple current will occur at D = 0.5
and the value can be estimated as
Vdc
i imax =
4L i f s
or,
Vdc
Li = (16)
4 f s i imax
Most of the cases, the grid-side inductor is designed considering 20% of the
inductor value connected at the inverter side. This will provide better attenuation
particular at the frequency above 50th harmonic and can be calculated as
L g = 0.2L i (17)
The natural frequency of oscillation can easily be estimated from Eq. (9) as follows

Li + Lg
ωn = (18)
L i L g Cf
From Eqs. (10) and (18), the damping resistance Rf can be calculated simply by
considering the damping ratio ξ = 0.5, which gives
1
Rf = (19)
ωr Cf
Also, the active power and the reactive power flow through the inductor due to
the fundamental component of the inverter current can be expressed as
Vi1 Vg
Pi = sin δ (20)
x
and,
Vi1
Qi = Vi1 − Vg cos δ (21)
x
where δ is the angle between the fundamental component of inverter output voltage
and the grid voltage and x is the series inductive reactance of the filter equal to
L i + L g.
5 System Specifications and Performance Analysis
The specifications of the proposed system are as follows:

DC bus voltage Vdc = 100 V
Fundamental voltage output from inverter Vi1 = 50 V
Active power output of the inverter Pn = 200 W
The modulation index from the output voltage from the inverter can be calculated
as
√
2Vi1
ma = = 0.707
Vdc
The fundamental output current from the inverter
Pn 200
I i1 = = = 4A
Vi1 50
Allowing 5% ripple current of the peak value of the fundamental current of the
inverter, i imax = 0.28 A. Switching frequency of the inverter is studied to find a
suitable design of inductors for LCL filter. Also, variation of THD and acceptable
power factor has been taken into consideration for this work. The performance of
the proposed LCL filter with respect to switching frequency, THD, and power factor
has been given in Table 1.
Although switching frequency is appearing in Table 1 as integer multiple of 1 kHz,
an incremental frequency of 125 Hz has been studied for the whole range. At lower
switching frequency with a fixed percentage of ripple current, the size of inductors is
446 T. Saha et al.
Table 1 Design criterion of LCL filter

Switching Inverter-side Grid-side Filter bandwidth THD Power factor
frequency inductance inductance (kHz) (%)
(kHz) (mH) (mH)
3 29.46 5.89 0.81 0.09 0.93
4 22.1 4.42 0.93 0.08 0.97
5 17.68 3.54 1.05 0.13 0.98
6 14.73 2.95 1.15 0.19 0.99
7 12.63 2.53 1.24 0.38 0.99
8 11.05 2.21 1.32 0.18 0.99
9 9.82 1.96 1.4 0.21 0.99
10 8.84 1.77 1.48 0.21 0.99
11 8.04 1.61 1.55 0.36 1.00
12 7.37 1.47 1.62 0.43 1.00
13 6.8 1.36 1.69 0.33 1.00
14 6.31 1.26 1.75 0.36 1.00
15 5.89 1.18 1.81 0.43 1.00
16 5.53 1.1 1.87 0.37 1.00
so high that it may result in sin δ > 1, which is infeasible design. In this point of view,
the minimum switching frequency has been considered to be 3 kHz, which causes a
power factor of 0.93 and with significantly large inverter-side inductor of 29.46 mH.
It is evident from Table 1, at the switching frequency of 4 kHz, the corresponding
THD level is found significantly low along with the acceptable level of power factor
of 0.97 and the required inverter-side inductor size of 22.1 mH. On the other hand,
at higher switching frequencies greater than 4 kHz, significant improvement of THD
level has not been found, although inductor sizing and the power factor remain
in the acceptable level. However, due to an increase in the switching frequency, the
switching losses of the inverter will increase accordingly and all the power electronic
switches of the inverter will experience a high switching stress and may damage if
proper snubber protection is not associated with the switches. Thus, the bolded part
of Table 1 is showing the optimum design parameters for the LCL filter the switching
frequency of 4 kHz.
To find out the value of the filter capacitor, base impedance and base capacitance
are required to be calculated using Eqs. (11) and (12). From the above data, the value
of base impedance will be Z B = 12.5 ohm and the value of base capacitance will be
CB = 255 µF. Since the value of the filter capacitance will be considered as 5% of
the base capacitance, Cf = 12.7 µF. Also from Eq. (19), the value of the damping
resistance will be Rf = 17.01 ohm.
If the switching frequency of the inverter switches is considered as f s = 4 kHz,
from Eq. (16), the value of the inverter-side inductor will be L i = 22.1 mH and from
Eq. (17), the value of the inductor connected at the grid side will be L g = 4.42 mH.
From Eq. (18), the resonant frequency of the system will be f n = 735 Hz, which
maintains the criteria 10 f g ≤ f n ≤ 21 f sw to avoid the resonant condition of the LCL
filter, where f g is the grid frequency.
From Eqs. (20) and (21), δ = 41.8◦ and the reactive power flow from the inverter
to the grid will be Q i = 76.36 VAr.
For a ripple current of 5%, the variation of inverter-side and grid-side inductors
with respect to the switching frequency is shown in Fig. 4. Also the variation of
frequency bandwidth of the filter, THD and power factor are shown in Fig. 5 accord-
ingly. Figures 6,7, and 8 represent THD of the inverter current, frequency response
of the LCL filter and THD of the grid current, respectively. Figure 9 shows the output
voltage of the inverter and the inverter output current, filter capacitor current and the
grid current are shown in Figs. 10, 11, and 12, respectively.
Fig. 4 Variation of
inductance with respect to
switching frequency
Fig. 5 Variation of
frequency bandwidth,
%THD and pf w.r.t.
switching frequency
448 T. Saha et al.
Fig. 6 THD of the inverter current
Fig. 7 Frequency response

of LCL filter
Fig. 8 THD of the grid current

Fig. 9 Inverter output

voltage
Fig. 10 Inverter output

current
6 Conclusion
The proposed work shows that the choice of switching frequency at 4 kHz results
in an inductor size of 22.1 mH at the inverter side and 4.42 mH at the grid side
corresponding to a reduction of THD in the grid current of 89%. The designed filter
also satisfies the choice of resonant frequency criteria. In this work, the minimization
of the switching frequency of the inverter has been emphasized, which is unique in
comparison with other research works in this area. In view of increase in single-
phase quality power generation from renewable sources, the proposed work has
high level of acceptability because of its cost-effectiveness and simplicity. In future,
this performance will be experimentally validated through the implementation of
single-phase grid-connected renewable power generation.
450 T. Saha et al.
Fig. 11 Grid current
Fig. 12 Filter capacitor

current
References
1. Hanju C, Vu TK (2010) Comparative analysis of low-pass output filter for single-phase grid-
connected photovoltaic inverter. In: 25th annual IEEE applied power electronics conference and
exposition (APEC), pp 1659–1665
2. Sedo J, Kascak S (2016) Control of single-phase grid connected inverter system. In: IEEE
ELEKTRO, pp 207–212
3. Cossoli P, Cáceres M, Vera L, Firman A, Busso A (2018) Proportional-resonant controller and
LCL filter design for single-phase grid-connected PV micro-inverters. In: IEEE PES transmission
and distribution conference and exhibition
4. Liu X, Ma P, He G, Wang D (2018) Study on control strategy of LCL type single-phase
grid-connected inverter. In: 3rd international conference on mechanical, control and computer
engineering (ICMCCE), pp 37–42
5. Tang H, Zhao R, Tang S, Zeng Z (2012) Linear quadratic optimal control of a single-phase
grid-connected inverter with an LCL filter. IEEE Int Symp Indus Electron, pp 372–376
6. Li X, Balog RS (2015) PLL-less robust active and reactive power controller for single phase
grid-connected inverter with LCL filter. In: IEEE applied power electronics conference and
exposition (APEC), pp 2154–2159
7. Cai P, Wu X, Yang Y, Yao W, Liu W, Blaabjerg F (2018) Design of digital filter-based
highly robust active damping for LCL-filtered grid-tied inverters. In: IEEE 4th southern power
electronics conference (SPEC)
8. Maghraby M, Ashour HA, Hossam-Eldin AA (2017) Analysis and implementation of discrete
SPWM signals for grid-connected inverter with LCL filter. In: 9th international middle east
power systems conference (MEPCON), pp 515–520
9. Sudhakar B, Satish Kumar GVS (2016) A unipolar fixed hysteresis band based sliding mode
control of single phase grid connected LCL filtered voltage source inverter. In: IEEE international
conference on power electronics, drives and energy systems (PEDES)
Level Control of Hopper Tank Process
Using Model-Based Controller
V. Murugananthan, M. Valluvan, and G. Sakthivel
Abstract In industries, many processes are nonlinear, in-order to understand and

analyze the process, linearization step is involved. The nonlinear process taken is
Hopper tank process. This paper highly concentrates on level control of Hopper
tank process. Conventional PID, internal model controller, modified internal model
controller are developed and tested in MATLAB simulation environment. Time
domain and error criterion analysis are done and better performances are highlighted.
Keywords Conventional PID · Internal model controller (IMC) · Modified internal

model controller (MIMC) · Internal model control based PID(IMC-PID) ·
First-order plus dead time (FOPD)
1 Introduction
Nonlinear process tanks like conical tank process, Spherical tank Process, and contin-
uously stirred tank process are highly in application of many process industries. The
above-mentioned processes are nonlinear because they never follow linearization
principle and superposition principle, in particular all the above-mentioned tank
processes area is not fixed, i.e., radius of the tank has been varied at each and every
level, thus area is also variable. Because of the above-mentioned reasons, those tanks
are nonlinear in nature.
Hopper tank process is nonlinear in nature, this is because it never follows
linearization and superposition principle, in its physical structure, it has cylindrical
tank, which is mounted on a conical tank bottom. This kind of tank processes
V. Murugananthan (B) · M. Valluvan · G. Sakthivel

Department of Electronics and Instrumentation Engineering, Faculty of Engineering and
Technology, Annamalai University, Annamalainagar, Chidambaram 608002, India
M. Valluvan
G. Sakthivel
454 V. Murugananthan et al.
are highly in use of petrochemical industries, pharmaceutical industries, and food

processing industries. Those industries use this tank setup because it can store large
amount of feed and at the same time, higher advantages of evacuation sludge material
[1]. While coming to offline analysis, FOPD transfer function is taken into account.
Many research papers are available for spherical tank, conical tank and cylindrical
tank process. But very limited articles are available for hopper tank process. Suresh
et al. [1] studied this tank system and developed FOPD models for different levels
and designed conventional PID using different tuning techniques, IMC-PID, then he
compared time domain performance indices and suggested better tuning technique.
Vinothkumar [2] applied Ziegler Nicholas and Cohen Coon tuning techniques for
the linearized model, which are developed by Kesaven [1]. Wang [3] numerical
model technique for hopper tank, then the results of pressure and flow dynamics
are compared. Wang [4] developed discrete model for hopper tank by using discrete
element method and finite element method and compared the results of pressure
and flow pattern. Dinesh Kumar and Meenakshi Priya [5] took two spherical tank
process and implemented gain schedule PID controller, Arasu et al. [6] took conical
tank process and implemented model-based nonlinear controller.
1.1 Work Flow
The main objective of this paper is to implement basic model-based linear controllers
IMC and MIMC, because in many papers, conventional controllers based on different
tuning techniques are implemented for hopper tank process while compared with
basic model-based linear controllers. The overall work of this paper is shown in
Fig. 1
2 Hopper Tank Modelling
Any process can be model by means of three methods; they are

i. Black box modelling
ii. Gray box modelling
iii. White box modelling.
White box modelling is applicable only if all process parameters are known.
If partial process parameters are known means gray box modelling is applicable.
Completely the process parameters are unknown means black box modelling can be
applied. Here, black box modelling technique is applied.
The geometrical diagram of Hopper tank process is shown in Fig. 2 and the tank
parameters [1] are
Overall height of the tank H is 100 cm.
Level Control of Hopper Tank Process Using Model-Based Controller 455
PROCESS IDENTIFICATION BLACK BOX MODELLING
ZIEGLER NICHOLAS
CONVENTIONAL PID
CONTROLLERS
TUNING
COHEN COON
IMC
MODEL BASED
CONTROLLERS
TUNING(λ)
MIMC
Fig. 1 Work flow diagram
Fig. 2 Hopper tank process

Variable height h.
Top radius of cone and radius of cylinder R Diameter of cylinder is 20 cm.
Variable radius of cone is r.
Fin inflow rate.
Fout outflow rate.
Valve co-efficient is β.
Volume v.
Area A.
Based on the parameters taken [1], the mass balance equation of the process is
dv/dt = Fin − Fout (1)
√
A dh/dt = Fin − β h (2)
The generalized FOPD transfer function model which relates input F in and out
variable height h is
ke−θs
G p (s) = h(s)/Fin (s) = (3)
τs + 1
By solving the mass balance Eqs. (1) and (2), then substitute the process
parameters, the FOPD model for level step change 015 cm is [1],
2.7e−0.15 s
G p (s) = h(s)/Fin (s) = (4)
(0.75 s + 1)
3 Controller Design
Three types of controllers are developed for obtained process model of setpoint
change 0–15 cm; they are conventional PID, IMC, MIMC, and IMC-PID.
3.1 PID Controller
The PID controller parameters are obtained by reaction curve method, i.e., by
getting open-loop response of the model, the tuning parameters are identified. In
PID controller, process model does not have any influence on controller structure.
The open-loop block diagram is shown in Fig. 3
The open-loop response FOPD is shown in Fig. 4,
Fig. 3 Open-loop structure SET POINT PROCESS OUTTPUT
PROCESS
Fig. 4 Open-loop response
From the response, the tuning gain K, dead time is θ and time constant τ is
obtained. Here, the time constant τ is 63% of rise time. The general tuning formulas
for Ziegler Nicholas and Cohen Coon are given. In Table 1, based on the parameter
obtained, the Ziegler Nicholas and Cohen Coon PID tuning parameters are given in
Table 2.
Table 1 Ziegler Nicholas

Tuning P P+I P+I +D
and Cohen Coon tuning
techniques
formulas
Ziegler τ/K θ – –
Nicholas 0.9τ/K θ 3.3θ –
1.2τ/K θ 2θ 0.5θ
Cohen Coon τ/K θ [1 + – –
θ/3]
τ -
K θ [0.9 + θ[30 +
θ/12τ ] 3θ/τ ]/9 +
20θ/τ
τ/Kθ[4/3 + θ[32 + 4θ/11 + 2θ/τ

θ/4τ 6θ/τ ]/13 +
8θ/τ
Table 2 Ziegler Nicholas

Tuning technique PID
and Cohen Coon tuning
values Ziegler Nicholas K P = 0.01278, K I = 0.000701, K D =
0.0568
Cohen Coon K P = 0.016232, K I = 0.00092, K D =
0.0472
Fig. 5 Closed-loop structure
The parameters from Table 2 are applied to the PID controller in Simulink model.
The closed-loop block diagram of the process is shown in Fig. 5,
In closed-loop, based on feedback process variable, the error is calculated, and
the PID controller gives manipulated variable to the process to make process variable
settle in the setpoint level given. The general PID controller structure is

G c (s) = K p e(t) + K p /Ti e(t) + K p ∗ Td de(t)dt (5)
Here, K p —proportional gain, T i —intehral time constant, T d —differential time

constant.
3.2 Internal Model Controller
Internal model control is a basic model-based control strategy, in which unlike PID
controller, the controller is purely influenced by process model, i.e., the control is
an inverse of the actual process model, so in IMC, the setpoint is tracked accurately,
i.e., there is no steady-state error theoretically; the proof is as follows,
yset(s)
G p (s) = (6)
u(s)
Then,
u(s)
G −1(s)
p = (7)
yset(s)
So, we can write
y(s) u(s) G p (s)

= (8)
yset(s) yset(s)
Equation (8) can be written as
y(s)
= G −1(s)
p ∗ G p (s) (9)
yset(s)
y(s)
=1 (10)
yset(s)
From the above result, it is clear that, the output exactly tracks setpoint, i.e., theo-
retically no steady-state error but practically steady-state error cannot be eliminated
but it can be minimized by using IMC. This is the major advantage of model-based
controller over PID controller. While inverting our process model, there is a mismatch
between numerator and denominator part of transfer function, i.e., numerator order is
greater than denominator order, and to compensate this, a low pass filter is multiplied
with the inverted process model. The controller model is as follows; from Eq. (4),
the process model is
2.7e−0.15 s
G p (s) = h(s)/Fin (s) =
(0.75 s + 1)
By inverting the above transfer function,
(0.75 s + 1)e−0.15 s
Ĝ p (s) = (system is unstable) (11)
2.70
To avoid transfer function mismatch, a low pass filter is included
1
LPF = (12)
λs + 1
G IMC (s) = Ĝ p (s) ∗ LPF (13)
It is clear that while we are inverting the process model, we get positive dead time
which means it make the system to give inverse response, so before inverting, pades
approximation [7] is applied for dead time. The general transfer function of pades
approximation for FOPD process is as follows
−0.5θ s + 1
e0.15 s = (14)
0.5θ s + 1
By solving Eq. (13), (14), we get IMC controller models as

Fig. 6 Internal model control structure
0.80625s + 1.075
G IMC (s) = (15)
2.4975λ s + 2.4975
λ is the filter co-efficient, generally its value lies between 0 and 2 [8], and for this
work, the filter co-efficient we consider as 0.6, 1.2, and 1.5. The IMC structure is
shown in Fig. 6.
3.3 Modified Internal Model Controller
The modified internal model controller block diagram is shown in Fig. 7. MIMC
controller mathematical part is same as that of IMC and the only difference is in the
structure, that is in IMC, actual process output and process model output is compared
and given as feedback but in MIMC, the process model output is directly added with
the controller output and given as input to the process.
Fig. 7 Modified internal model control structure

The responses of all control strategies are given in Figs. 8, 9, 10 and 11. The time
domain and error criterion of all control strategies are given in Table 3 and from
the process response, it is clear that conventional PID using Ziegler Nicholas and
Cohen Coon tuning technique, the time taken to settle the process in actual setpoint
is very large and also the process attains initial settlement prior to actual setpoint for
some time period and then starts rising and settles in actual setpoint. But in IMC and
MIMC, smooth settling is achieved as well as settling time taken is highly reduced.
The responses of IMC and MIMC for different õ are shown in Figs. 10 and 11,
From the response, it is clear that higher õ values increase the time taken by the
Fig. 8 Ziegler Nicholas

tuning response
Fig. 9 Cohen Coon tuning

response
Fig. 10 IMC response for λ

= 0.6, λ = 1.2, λ = 1.5
Fig. 11 MIMC response for

λ = 0.6, λ = 1.2, λ = 1.5
process to settle at setpoint, and time domain criterion values also proven the same.
From the response and time domain criterion, both IMC and MIMC are similar but
when come to error criterion, MIMC gives better results. Overall MIMC outperforms
IMC and conventional PID
5 Conclusion
The results of model-based controllers outperforms conventional PID controller;

this is because the controller has the influence of process model. In this work, for
model-based control structure, the FOPD model is taken as plant and process model
Table 3 Time domain and error criterion values

Controllers Ziegler Cohen IMC MIMC
Nicholas Coon λ = 0.6 λ = 1.2 λ = 1.5 λ = 0.6 λ = 1.2 λ = 1.5
PID PID
Time domain specifications
Rise time 3.5935e 2.8393e 3.8008 9.6265 12.4543 3.8008 9.6265 12.4543
in sec + 04 + 04
Settling 5.8695e 4.8168e 15.3969 25.9288 30.9372 15.3969 25.9288 30.9372
time in sec + 04 + 04
Peak time 84213 84208 53 54 54 53 54 54
in sec
Error criterion
ISE 5.62e + 4.348e 4906 4934 4949 78.1 133.9 162.7
04 + 04
IAE 7708 6031 376.8 378 378.5 7.95 15.49 19.6
approximated after pades approximation is taken as process mathematical model. In

the future, the simulation parameters will be applied to real-time process. The scope
of the work will be in model predictive controller mechanism and adaptive control
mechanism.
References
1. Kesavan SM, Padmesh TV, Shyan CW (2014) Controller tuning for non-linear hopper process
tank-A real time analysis design. J Eng Sci Technol (Special Issue):59–67 (2014). (EURECA
2013)
2. Vinothkumar C, Esakkiappan C (2019) Level control of nonlinear hopper tank process in phar-
maceutical industry using ziegler Nicholas and cohen coon tuning techniques. Int. J. Innovative
Technol Explor Eng (IJITEE) 8(10). ISSN: 2278–3075
3. Wang Y, Lu Y, Ooi J (2013) Numerical modelling of dynamic pressure and flow in hopper
discharge using the Arbitrary Lagrangian-Eulerian formulation. Eng Struct 56:1308–1320
4. Yin W, Ooib JY (2015) A study of granular flow in a conical hopper discharge using discrete
and continuum approach. Sci Direct Proc Eng 102:765–772
5. Dinesh Kumar D, Meenakshi Priya B (2012) Design and implementation of non linear system
using gain scheduled PI controller. Procedia Eng 38:3105–3112
6. Arasu SK, Panda A, Prakash J (2016) Experimental validation of a nonlinear model based control
scheme on the variable area tank process. IFAC-PapersOnLine 49(1):30–34
7. Bequette BW (2003) Process control: modeling, design, and simulation. Prentice Hall,
Englewood Cliffs
8. Jogwar SS (2018) NPTEL study Material, Department of Chemical Engineering, Indian Institute
of Technology Bombay
Hybrid L + C Edge Detection Algorithm
for Harnessing Large Image Datasets
Gurpartap Singh, Sunil Agrawal, and B. S. Sohi
Abstract The primary focus of this work is to improve the performance of pre-
existing techniques used to crop fragments of images from a larger image or a form.
The form mentioned here can consist of any number of rectangles having different
dimensions and in varying arrangements. The requirement of such forms arises with
the staggering growth in the application of artificial intelligence (AI). This growth has
made the collection of large image datasets an essential task, as we need a colossal
amount of data to train algorithms. These algorithms are in turn used in the trending
AI applications. This has made collection of datasets a prominent task as any in the
development of an AI. Therefore, in the field of handwriting recognition also we
need a large dataset. To create such large datasets, we need general formats that can
be applied anywhere around the globe and also do not require any special equipment.
In this work, a hybrid approach with the above said attributes have been proposed
to collect the Punjabi handwritten dataset. This proposed technique primarily uses
pre-existing operators designed for edge detection and locates rectangles within the
forms. Then, it matches the dimension and location of rectangles within the forms
with pre-assigned values. On applying the proposed technique, the algorithm was
able to extract segments with quite impressive accuracies and create large enough
datasets for future work. The main reason for achieving such high accuracy is that
the proposed technique exploits the robustness of both Canny and Laplace operators
for edge detection.
Keywords Datasets · Handwriting · Dilation · Contours · Canny operator ·

Laplace operator · Edge detection
G. Singh (B) · S. Agrawal

UIET, Panjab University, Chandigarh 160017, India
S. Agrawal
B. S. Sohi
Chandigarh University, Gharuan, Punjab 140413, India
466 G. Singh et al.
1 Introduction
Image processing has been a prevailing field of research for decades because of its
wide range of applications in almost every field. But recently, staggering growth
in computer vision and artificial intelligence(AI) has led to tremendous growth in
the field of image processing. These fields are entirely based on data and a signif-
icant amount of data is in the form of images. Hence, it is natural that image
processing will be one of the most prominent research fields in the future [1]. Image
processing includes several fundamental steps but this work is entirely focused on
image retrieval/fragmentation. As we all know that most of the available data is not in
a structured form neither, it is available in a format in which they can be used to train
AI or computer vision algorithms. Most of the algorithms that are being used today
are deep learning algorithms and they need very specific datasets so that they can
learn the general and in-depth representation of datasets. Only then, these algorithms
can be used in real-time applications. If the datasets used to train these algorithms is
somewhat compromised, then it will create issues in achieving the ultimate goal of a
generalized AI. This has made creating a dataset quite significant but yet a daunting
task as any in the development of an AI.
Recently, some synthetic methods [2, 3] have also been developed to create image
datasets that mimic the existing images. But the images obtained using synthetic
methods cannot accurately represent the data encountered in the real-time scenarios.
The main reason for above-mentioned issue is that synthetically generated datasets
are nothing but anomalies created by introducing variance in the existing images.
Even in recent experiments done by Jahanian et al. [4], it was observed that gener-
ative adversarial network’s (GAN’s) ability to introduce variance is limited which
means GANs can introduce only a certain degree of variation in dataset. This clearly
substantiate that the data generated using synthetic methods yet cannot represent
the real-time data; it can only generate samples within the scope of existing data or
extrapolate entirely outside the support of the training data.
Although the main feature of these synthetic techniques is that they learn the
distribution from the existing data and then accordingly introduce variance in order to
generate new data samples but even, then we cannot completely rely upon these tech-
niques for future applications. Hence, this work is dedicated to find an approach using
which data can be collected conveniently for automation in the future applications.
2 Literature Review
Since last few decades, a significant amount of work has been done in the field of
image segmentation. In [5], authors have done inspection of several autonomous
approaches which have been used for image segmentation and to what extent
Hybrid L + C Edge Detection Algorithm for Harnessing … 467
they have achieved success. There are several application where image segmen-
tation is required like preservation of historical handwritten text documents [6–
8], vehicle number plate extraction from images [9], segmentation of handwritten
words/alphabets for classification [10], etc. Based on applications, there are different
algorithms that have been used for segmentation and these algorithms uses different
techniques for segmentation [11]. In this work, the main aim is to crop rectangular
segments from larger images. Hence, the techniques mentioned in [11], namely edge
detection-based image segmentation techniques have great potential for yielding
excellent results in the scenario considered in this work. The two most commonly
used edge-based segmentation methods are Canny [12] and Laplace [13]. The main
difference between these two techniques is that Canny is first derivative order-based
edge detection whereas Laplace is a second order edge detection technique. Moreover
in canny filter, Otsu’s thresholding technique [14] is used in order to obtain better
edges. The main reason for using Otsu’s thresholding is that it generally searches for
the threshold that minimizes the intra-class variance which is defined as a weighted
sum of variances of the two classes. Hence, yielding a high accuracy in edge detec-
tion. Moreover, there are several variants of Laplace and Canny filters [15, 16] that
have been used for several applications, which suggests that these techniques are
quite versatile.
From literature survey, it is evident that these edge detection techniques are capable
of extracting fragments of images from a larger image irrespective of the format of
form being used for data collection. Hence, the work presented in this article is
entirely based on edge detection techniques and also the proposed methodology is
derived from these techniques only.
3 Methodology
Present work consists of a form printed on A4 page, but this method can be very easily
extended for pages of any size. Here, we have considered three formats in which we
have collected handwritten Alphabets/Vowels, digits, and words in Punjabi language.
The formats used are as shown in Fig. 1a–c.
From the above figures, it is clear that we can place rectangles in different arrange-
ments and in different sizes thus assuring wide range of applications. These forms
are then printed on A4 size paper and are filled by school students. Later, the filled
forms are scanned using a CanonMF246dn Digital Multifunction Laser Printer. A
scanned image of the filled forms is shown in Fig. 2.
As we can observe from Fig. 2, there are several imperfections and noise that
have been introduced while scanning hence, the algorithm must be able to identify
them ideally. Otherwise, it will be difficult to extract the fragments accurately. The
imperfections occurring at the time of scanning also include some rotation as well
which makes quite difficult to pinpoint the exact location of rectangular boxes. Hence,
the required methodology must be robust enough to locate the exact edges despite of
468 G. Singh et al.
Fig. 1 a A segment of format for alphabet and vowel dataset collection. b A segment of format for
collection digit dataset. c A segment of format for collection word dataset
Fig. 2 A scanned image of format for alphabet and vowel dataset collection
470 G. Singh et al.
Fig. 3 Algorithm for extracting handwritten fragments from scanned images
all the deformations. Taking into consideration all these factors, the final algorithm
used to extract images is shown in Fig. 3.
The algorithm shown in Fig. 3 comprises of features of both Laplacian and Canny
filters for edge detector. The reason for using both the filter at the same time is that they
both have excellent features but neither of them is individually capable of detecting
all the edges from a noisy and deformed images. This has been demonstrated in the
results mentioned in Table 2. That is the reason that we have combined both the edge
detectors so that we can get better image extraction, which can be confirmed from
the results mentioned in Table 3.
In this work, we have collected data from over 300 individuals. The details of forms
filled by individuals and the number of fragments in the scanned images that have to
be extracted are mentioned in Table 1.
After we have obtained the scanned images, the next step is to extract the frag-
ments. We have initially used Canny and Laplace separately to extract the frag-
ments. This also helps in analyzing how these algorithms individually and collectively
perform. The details of extracted fragments are mentioned in Table 2.
Table 1 Details of image fragments within scanned images

S. No. Type Number of scanned Number of boxes in each Rectangle count
IMAGES format
1. Alphabets 323 47 15,181
2. Digits 325 40 13,000
3. Words 325 20 6500
Table 2 Details of extracted fragments using Laplace and Canny operators individually
S. No. Type Fragments Accuracy Fragments Accuracy
extracted (Canny) extracted (Laplace)
(Canny) (Laplace)
1. Alphabets 13,813 90.9 13,719 90.3
2. Digits 11,641 89.5 11,609 89.3
3. Words 5442 83.7 4113 63.2
As we can observe in both the cases, that the number of extracted fragments are
quite less than the actual number. But there is another observation that the fragments
extracted by both the filters are not exactly same. There are some images that are
extracted only by either of the algorithms. Hence, all we need is to combine both the
algorithms in a manner that we can extract relatively higher number of fragments.
Therefore, in this work, both the filters have been used in unison in the algorithm
suggested in this work; hence, we are able to integrate the mathematical foundations
of both the filters in a single filter. After combining both the filters, we were able
to extract more number of fragments than we were able to extract with individual
filters. The details of extracted fragments from proposed algorithm are mentioned in
Table 3.
Another important fact regarding this method is that the amount of scanned images
will have no effect on the accuracy. So it will not matter if we are extracting frag-
ments from 100 or 1000 scanned images. A comparison of accuracy for all the
three algorithms is shown in Fig. 4 which indicates that the suggested algorithm has
outperformed the existing algorithms.
From Fig. 4, it can be visualized that our proposed method is able to extract the
fragments with higher accuracy as compared to Laplace and Canny individually.
Table 3 Details of extracted fragments using proposed algorithm

S. No. Type Extracted fragments using L + C Accuracy
(%)
1. Alphabets 14,839 97.7
2. Digits 12,635 97.1
3. Words 5843 89.8
472 G. Singh et al.
Fig. 4 Accuracy comparison of Laplace, Canny, and L + C algorithms for alphabets, digits, and
words
This improvement can be attributed to the fact that the proposed method is able to
combine the benefits of both the existing techniques.
5 Conclusion and Future Scope
From experimental results shown in this work, it can be observed that proposed
algorithm did reasonably well. In order to make this technique easily accessible,
hardware used in this work is quite ordinary. An improvement that can be made
to this proposed methodology is by the use of noise cancellation techniques which
would make edge detection facile. Also the quality of the scans can be improved by
using hardware of superior grade.
The main advantage of this work is convenience in collecting large image datasets.
Hence, the methodology proposed in this work will be of great help for researchers
working on collection and segmentation of large data. For a futuristic technique, we
can introduce adaptiveness as well, which will help in designing an edge detection
algorithm capable of maintaining robustness even at the most sophisticated scenarios.
Acknowledgements This Publication is an outcome of the R&D work undertaken project under the
Visvesvaraya Ph.D. scheme of Ministry of Electronics and Information Technology, Government
of India, being implemented by Digital India Corporation.
References
1. Shih F (2011) Image processing and mathematical morphology. CRC Press, Boca Raton
2. Martin Arjovsky SC, Bottou L (2007) Wasserstein generative adversarial networks. In:
International conference on machine learning, pp 214–223
3. Pu Y, Gan Z, Henao R, Yuan X, Li C, Stevens A, Carin L (2016) Variational autoencoder for
deep learning of images, labels and captions. NIPS 2352–2360
4. Jahanian A, Chai L, Isola P (2019) On the “steerability” of generative adversarial networks.
ArXiv abs/1907.07171
5. Zhang H, Fritts JE, Goldman SA (2008) Image segmentation evaluation: a survey of
unsupervised methods. Comput Vis Image Undestanding 110(2):260–280
6. Toledo JI, Carbonell M, Fornés A, Lladós J (2018) Information extraction from historical
handwritten document images with a context-aware neural model. Pattern Recogn (2018)
7. Sánchez JA, Bosch V, Romero V, Depuydt K, De Does J (2014) Handwritten text recogni-
tion for historical documents in the transcriptorium project. In: ACM international conference
proceeding series
8. Sastry PN, Krishnan R, Ram BV (2010) Classification and identification of Telugu handwritten
characters extracted from palm leaves using decision tree approach. J Applied Engn Sci 5(3):22–
32
9. Sharma C, Kaur A (2011) Indian vehicle license plate extraction and segmentation. Int J Comput
Sci Commun 2(2):593–5999
10. Louloudis G, Gatos B, Pratikakis I, Halatsis C (2009) Text line and word segmentation of
handwritten documents. Pattern Recogn 42(12):3169–3183
11. Zaitoun Nida M, Aqel Musbah J (2015) Survey on Image Segmentation Techniques. Procedia
Comput Sci 65:797–806
12. Canny J (1986) A computational approach to edge detection. IEEE Trans Pattern Anal Mach
Intell 6:679–698
13. Wang X (2007) Laplacian operator-based edge detectors. IEEE Trans Pattern Anal Mach Intell
29:886–890
14. Otsu N (1979) A threshold selection method from gray-level histogram. EEE Trans Syst Man
Cybern 9(1):62–66
15. Rong W, Li Z, Zhang W, Sun L (2014) An improved CANNY edge detection algorithm. In:
2014 IEEE international conference on mechatronics and automation, pp 577–582
16. Fan P, Zhou RG, Hu WW, Jing N (2018) Quantum image edge extraction based on Laplacian
operator and zero-cross method. Quantum Inf Process 18(1):27
Resolution Selective 2–6-Bit Flash ADC
in 45 nm Technology
Sarfraz Hussain, Rajesh Kumar, and Gaurav Trivedi
Abstract This paper presents a 2–6-bit resolution selective flash ADC (RSA). A
re-configurable or resolution selective flash ADC is designed for use with different
applications. In this design, resolution for a particular application can be chosen
without the need of replacing the ADC with another low or high resolution ADC.
Speed, resolution, and power are inter-dependent; therefore, to maintain a good
balance re-configurable ADCs are very much efficient. A resolution selective design
gives user a choice for selecting the resolution according to the application and thus
save the manufacturing cost. For full resolution of 6-bit the ADC consumes 1.57 mW
and for the lowest resolution of 2-bit it consumes 268.9 µW for 1 V supply voltage.
Keywords Mixed-signal circuits · Encoders · Flash ADC · Fat-tree encoder ·

Re-configurable ADCs
1 Introduction
Resolution is one of the most important factors in an ADC. Resolution is defined as

1 LSB which is the step size of an ADC. More the resolution, better the output. With
performance comes trade-off. Speed decreases and power consumption increases
with increase in resolution. For an n-bit Flash ADC, 2n − 1 number of comparators
are required. For a 2-bit Flash ADC 3 comparators are required whereas for 6-bit
flash ADC 63 comparators are required. High speed applications sometimes doesn’t
require high resolution therefore a low resolution ADC can give a good required
S. Hussain (B) · R. Kumar

Department of Electronics and Communication Engineering, NERIST, Nirjuli, Arunachal
Pradesh, India
R. Kumar
G. Trivedi
Department of Electrical and Electronics Engineering, Indian Institute of Technology Guwahati,
Guwahati, Assam, India
476 S. Hussain et al.
outcome. Similarly, in less speed and more accurate applications like CD read-write
operation high resolution is required. Earlier two different ADCs are required for
high speed and high accuracy applications but with resolution selective ADC a single
ADC is sufficient. By selecting the resolution in accordance with our requirements
we can get optimum results. 45 nm technology provides better speed performance
and a torsion between different technologies.
In [1], a resolution selective flash ADC was proposed for a maximum resolution
of 4-bit. It was designed for 90 nm technology. The proposed design has a higher
resolution and is designed for 45 nm technology node. In [2], a portable application
based 6-bit flash ADC is designed. Many ADCs have been and are being designed
in bulk. TIQ-based comparators have an advantage of being fast [3, 4]. Therefore,
a resolution selective TIQ based ADC can reduce the manufacturing cost as well as
increase speed.
Section 2 describes the idea of resolution selective flash ADC. An algorithm is
proposed for the selection of analog and digital part. Section 3 shows the simulated
results and comparison table. All the simulations are carried out in Cadence Virtuoso
IC 6.1.
2 Resolution Selective Flash ADC
In the proposed flash ADC, one can select the resolution from 2 to 6-bit as per the
specific application. The main idea of the resolution selective ADC is to reduce the
manufacturing cost by implementing the same ADC for multiple applications. A flash
ADC has two main blocks: comparator block (analog part) and encoder block (digital
part). The first step for designing such an ADC is to select the highest resolution.
Once highest resolution ADC is designed then the proposed algorithm is applied to
select the inner components for different resolutions. Algorithm 1 is used for analog
part election and algorithm 2 is used for digital part selection.
2.1 TIQ Based Comparator
Threshold Inverter Quantizer is an Inverter based comparator proposed in [5] which is

used in this paper. The algorithm used in [6] is used to select different transistor sizes
for different thresholds. Such a comparator doesn’t require resistor ladder circuitry
and the threshold acts as a reference node voltage. The comparator selective circuit
is shown in Fig. 1. The following equation is used to select the transistor sizes [4, 5]:

Vth = sqrt(k) VDD − |Vtp | + Vtn /1 + sqrt(k)
where k = µp Wp/µn Wn (1)
Resolution Selective 2–6-Bit … 477
Fig. 1 Comparator circuit
Table 1 Step size for

Resolution Step size (mV)
different resolutions for the
same input range 2 112
3 32
4 16
5 7.46
6 3.61
In [6], the transistor sizes are optimized by utilizing the property of INL and DNL.
Optimization is required because of a vast range of transistor values for a specific
threshold. For choosing a particular threshold or reference points for the comparator,
the full scale range is selected by choosing the highest and the lowest input swing.
V DD − |V tp | is the highest input and V tn is the lowest input. The input range for the
proposed design varies from 368 to 592 mV. With this range step size is chosen by
dividing it into 2n − 2 parts. Respective step size for different resolutions is tabulated
in Table 1.
2.2 Encoder
Encoder is the digital part of an ADC [7–9]. Modified Fat-tree encoder is used along-
with a bubble error corrector circuit for better results. The selection of output is done
Fig. 2 Proposed encoder circuit
Table 2 Selection of binary

Resolution MSB LSB
outputs based on resolution
2 B5 B4
3 B5 B3
4 B5 B2
5 B5 B1
6 B5 B0
through the proposed algorithm. Only the required components are chosen rest are
turned OFF. For an encoder only the output block selection is made for easy and fast
response. Selector block is added to the output of each of the successor blocks of the
binary encoder. The proposed encoder is shown in Fig. 2. For different resolutions
the MSBs and LSBs are tabulated in Table 2.
2.3 Proposed Algorithm for RSA
An algorithm is proposed for resolution selective flash ADC (RSA). It helps us

select only the required components for a specific resolution thus saving power and
Table 3 Algorithm for analog part selection

Algorithm 1 Algorithm for selecting the comparators for varying resolution
Input: Resolution in bits
Output: Selected Comparator Array
Initialisation :
1: Initialise 6-bit array
‘a[]’
Setting 6-bit array elements:
2: for i = 0 to 62 do
3: a[ i ] = i+1
4: end for
Selection of components from array ‘a’ based on the selected resolution :
5: Res = required resolution
6: n = 2**res
7: b= [[] for k in range(n-1)]
8: b[0] = a[0]
first element is always same
9: x = 6-res
10: for i = 1 to length of b array do
11: j = (i+2)*(2**x)-2
12: b[i] = a[j]
13: end for
14: return b
**Note: Total number of comparators in a block is 63 (i.e. 26 -1). Out of the 63
comparators, selection is being made.
manufacturing cost. Algorithm is tabulated in Tables 3 and 4. Separate algorithms are

followed: Algorithm 1 is meant for comparator block selection while the algorithm 2
is meant for selecting digital part.
3 Simulated Results
The parameters considered for the simulations are: (i) 1 V supply for 45 nm tech-
nology, (ii) minimum channel length is taken as 45 nm and (iii) ramp input is given.
In the proposed flash ADC, one can select the resolution from 2 to 6-bit as per the
specific application. For 2, 3 and 4-bit resolution different graphs are obtained to
show the selectivity results. The performance of the proposed resolution selective
flash ADC is tabulated in Table 5. The normalized power-delay product (PDP) for
2-bit resolution is 1.8 and for 6-bit is 47.1. The outputs of ADC for different resolu-
tions are shown in Figs. 3, 4 and 5. The results signify that it can be used for multiple
applications. Compared to a TIQ flash ADC the proposed ADC shows more or less
good results when 4-bit resolution is selected. Table 6 shows the comparison of both
the 4-bit ADCs. Though, it has a drawback of consuming more power relative to ADC
Table 4 Algorithm for digital part selection

Algorithm 2 Algorithm for selecting the encoder outputs for varying resolution
Input: Resolution in bits
Output: Selected Binary Output Array
Initialisation :
1: Initialise 6-bit output array
‘a[]’
Setting 6-bit array elements:
2: for i = 0 to 5 do
3: a[i] = i+1
4: end for
Selection of components from array ‘a’ based on the selected resolution :
5: Res = required resolution
6: b= [[] for k in range(0,res)]
7: x = 6 - res
8: for i = 0 to length of b array do
9: b[i] = a[i] + x
10: end for
11: return b
Table 5 Performance chart for resolution selective flash ADC

Resolution Delay (pico-sec) Power dissipation (µ-W) Normalized PDP Normalized EDP
2 70 268.9 1.8 1.31
3 200 591.6 11.8 23.66
4 250 941.8 23.5 58.86
5 270 933.1 25.2 68.02
6 300 1570 47.1 141.30
Fig. 3 2-bit flash ADC output

Table 6 Proposed 4-bit flash ADC performance comparison

ADC Delay (pico-sec) Power dissipation Normalized PDP Normalized EDP
(µ-W)
4-bit flash ADC 252.8 740.4 1.87 47.31
Proposed 4-bit 250 941.8 2.35 58.86
RSA
without the resolution selective blocks but it saves the manufacturing cost and can be
utilized for multiple applications. Thus, a resolution selective ADC can be used with
less cost of implementation for multiple applications. Table 7 exhibits the compar-
ison of the proposed ADC with the recent flash ADC designs. We can observe that
the chosen flash ADC design architecture is advantageous and the proposed design
enhances the implementation capability of the ADC.
Table 7 4-bit flash ADC performance comparison

ADC Proposed flash Proposed flash Wu et al. [10] Varghese and
ADC (RSA) ADC (RSA) Mahapatra [11]
Technology node 45 nm 45 nm 180 nm 90 nm
Resolution (in bits) 6 4 4 4
Max sampling rate 1.25 1.5 2 3
(in GHz)
Supply voltage (in 1 1 1.8 1.2
volts (V))
Power dissipation 1.57 0.94 42 2.42
(mW)
4 Summary
The paper depicts the methodology for selection of different components for different
resolutions. Two algorithms are proposed for the selection procedure. A 4-bit flash
ADC without the selection block is compared with flash ADC with selection block
having the same architectural components. The resultant of proposed ADC shows
more power consumption due to the additional blocks. When compared to other recent
4-bit ADCs, the chosen architecture and the proposed design show improvement in
the design. For the complete 6-bit resolution selective flash ADC (RSA), maximum
sampling rate is 1.25 GHz with a power dissipation of 1.57 mW, and for a 4-bit
RSA, it is 1.5 GHz with 0.94 mW. The objective of the design is to reduce the
manufacturing cost and to use the same ADC for multiple applications. By selecting
different resolutions it can be used for applications required.
Acknowledgements Authors would like to thank ECE department, NERIST, Arunachal Pradesh
and EICT Academy, IIT Guwahati, Assam for their resource and assistance. We would also like to
thank MeitY (under Govt. of India) for facilitating research grant under Visvesvaraya PhD Scheme.
References
1. Tiwari S, Kumar A (2018) Reconfigurable flash ADC using TIQ technique. In: 4th international
conference on computing sciences (ICCS). Jalandhar, pp 204–208
2. Mishra SN, Arif W, Mehedi J, Baishya S (2013) A power efficient 6-bit TIQ ADC design for
portable applications. In: IEEE third international conference on consumer electronics Berlin
(ICCE-Berlin). Berlin, pp 235–239
3. Hussain S, Kumar M (2016) Design of an efficient 8-bit flash ADC for optical communication
receivers. In: International conference on electrical, electronics, and optimization techniques
(ICEEOT). Chennai, India, pp 449–453
4. Abumurad A, Ozdemir A, Choi K (2019) 10-bit flash ADCs and beyond: an automated
framework for TIQ flash ADCs design. Circuits, Syst Signal Process 38(9):4314–4330
5. Lee D, Yoo J, Choi K (2002) Design method and automation of comparator generation for flash
A/D converters. In: IEEE international symposium on quality electronic design, pp 138–142
6. Ozdemir Alrizah M, Choi K (2019) Optimization of comparator selection algorithm for

TIQ flash ADC using dynamic programming approach. In: IEEE computer society annual
symposium on VLSI (ISVLSI). Miami, FL, USA, pp 495–500
7. Lee D, Yoo J, Choi K, Ghaznavi J (2002) Fat tree encoder design for ultra high speed flash
A/D converters. In: Proceedings of 2002 MWSCAS, vol 2. Oklahoma, USA, pp II-87-II-90
8. Sall E, Vesterbacka M (2008) Thermometer-to-binary decoders for flash analog-to-digital
converters. In: 18th European conference on circuit theory and design. Seville, Spain, pp
240–243
9. Varghese GT, Mahapatra K (2016) A low power reconfigurable encoder for flash ADCs.
Global colloquium in recent advancement and effectual researches in engineering, science
and technology (RAEREST 2016), Elsevier Procedia Technology, vol 25, pp 574–581
10. Wu L, Huang F, Gao Y, Wang Y, Cheng J (2009) A 42 mW 2 GS/s 4-bit flash ADC in 0.18 µm
CMOS. In: International conference on wireless communications and signal processing.
Nanjing, China, pp 1–5
11. Varghese GT, Mahapatra K (2016) A low power reconfigurable encoder for flash ADCs.
Procedia Technol 25:574–581
Measurement of Heartbeats
for Well-Being Assessment Using
Photoplethysmographic Signals
Piyush Kumar, Arpan Sharma, Mayanmi Zimik, Shivam Parashar,

Renu Singh, Joyatri Bora, Aswini K. Patra, and Madhusudhan Mishra
Abstract Heart rate carries vital information for the diagnosis of cardiac abnormal-
ities. By monitoring the heart rate, one can acquire crucial information regarding the
health condition of a person. Expansion and contraction of the heart bring up varia-
tions in the volume of the blood flowing through the entire human body. On observing
the blood flow via blood vessels, one can determine the rate by which heart pumps the
blood. The proposed work built on an embedded platform is aimed to determine the
heartbeat counts in order to detect cardiac abnormalities like arrhythmia in real-time
scenario. The principle used in this study is based on photoplethysmography (PPG)
which is a simple and non-invasive technique to achieve heartbeats information. It
calculates the heartbeat by analyzing the change in the volume of blood flowing
P. Kumar · A. Sharma · M. Zimik · S. Parashar · J. Bora · M. Mishra (B)

Science and Technology (NERIST), Nirjuli, Arunachal Pradesh 791109, India
P. Kumar
A. Sharma
M. Zimik
S. Parashar
J. Bora
R. Singh
Technology (NERIST), Nirjuli, Arunachal Pradesh 791109, India
A. K. Patra
Department of Computer Science and Engineering, North Eastern Regional Institute of Science
and Technology (NERIST), Nirjuli, Arunachal Pradesh 791109, India
486 P. Kumar et al.
through the blood vessels. Arteries near the skin of the fingers are used for detection
of the pulse. The density of blood in arteries changes with each beat of the heart. This
variation is strengthened by the amplifiers and is fed to the microcontroller (Arduino
Uno) for further analysis. An LCD display shows the heartbeat rate thereby paving
a way to assist medical professionals for reliable and accurate treatment of cardiac
abnormalities.
Keywords Heart rate · Arrhythmia · Arduino Uno · Photoplethysmography ·

Cardiovascular abnormalities
1 Introduction
Heart rate is defined as the rate by which the heart of an individual beats in a minute.
It is also known as pulse. It depends on various factors like age, size, emotions,
anxiety, body temperature, dehydration, etc. [1]. Generally, the heart rate is low for
a healthy person. It is because of the efficient functioning of the heart muscles. A
non-invasive technique like the ANFIS system is used to find out fetal heart rate
[2]. Some surveys reported that a small increase in blood pressure leads to a rise in
the risk of cardiovascular diseases (CVDs) [3]. The normal heartbeats for a human
by age are given in Table 1 [4]. Arrhythmia is a very common cardiac disorder in
peoples having heart-related issues. It is diagnosed when the heartbeat of a person
varies from the normal heartbeat. About 80% of the heart attack incidents are due to
arrhythmia, causing 12% of total deaths annually [5].
Arrhythmia may be of different types like bradycardia, tachycardia, premature
contraction, atrial fibrillation, and ventricular fibrillation. When the heart rate of an
adult goes below 60, then the person suffers from bradycardia. It does not show
any symptoms until the heartbeat reaches below 50. The heart rate more than the
normal range is referred as tachycardia. Even after so much technical advancement
in medical science, cardiac arrest largely affects public health and is the leading
cause of death globally [6–10]. A large number of studies have been performed to
make arrhythmia detection processes uncomplicated [11–13]. Due to a large number
of people who needs continuous monitoring, a self-operating diagnosis system is
required. People of any age group may become the victim of arrhythmia. Common
arrhythmia symptoms include premature rhythms, fatigue, weariness, giddiness, and
Table 1 Heart rates of healthy human for different age groups [4]
AG 40 45 50 55 60 65 70
AMHR 180 175 170 165 160 155 150
THR 90–153 88–149 85–145 83–140 80–136 78–132 75–128
Here, AG age group, AMHR average maximum heart rate in beats per minute, and THR target heart
range in beats per minute
Measurement of Heartbeats for Well-Being Assessment … 487
fainting. Morphological arrhythmia and rhythmic arrhythmia are the two major cate-
gories of arrhythmia. Morphological arrhythmia refers to only one irregular heartbeat,
while rhythmic arrhythmia refers consists of a series of irregular heartbeats.
2 Proposed Method
Heart diseases need a routine checkup and proper diagnosis by the physicians. By
doing this, doctors can treat patients well. Some diseases like arrhythmia and prehy-
pertension need continuous monitoring by the doctors. Prehypertension does not indi-
cate any symptoms but damages the atrial wall and the nervous system [3]. For this
type of diagnosis, a person has to stay in hospitals for a longer duration. This creates
many problems such as a shortage of doctors and beds in hospitals for all the patients.
Some other patients who need urgent treatment may not get admitted to hospitals.
In this case, hospital authority usually discharges patients who are admitted only for
medical examination. Due to technological advancement, health centers use wireless
communications to control traffic [14]. Heartbeat is also an important parameter to
know about the health condition of an individual in this context [15]. Electrocardio-
gram (ECG) recording-based heart monitoring system employed in hospitals usually
leads to discomfort to the patients due to wiring arrangements. It is not applicable to
wearable devices. PPG-based instruments are economical, easy to use and portable.
PPG uses light rays for diagnosis purpose, and ECG uses electrodes to transmit and
receive information. ECG-based bedside cardiac monitoring system mostly shows
incorrect values, 86% approximately [16]. The PPG uses optical sensors, so it is
used in wearable devices. Our proposed method is based on PPG which has mainly
a pulse sensor module, Arduino board, and a display. A typical system comprising
of these different units for heart rate measurement using PPG signals is illustrated in
Fig. 3. The main objective of the proposed scheme is to measure the heart rates from
PPG signals. Pulse sensor consists of light-emitting diode (LED) and light-dependent
resistor (LDR), which generates potential difference when a tip of a finger is placed
over the sensor for monitoring heartbeat. Arduino is an open-source microcontroller.
It is used to design various electronic equipment. It has hardware as well as soft-
ware. It consists of “Integrated Development Environment” (IDE), which is used to
write and upload programs from computer to Arduino board. Arduino does not need
separate hardware (programmer) to upload new code onto the board. Uno is the most
used board among all in the Arduino family. It is mainly used by the novices because
it is cheap and simple to learn (Fig. 2).
Arduino communicates with buttons, LEDs, motors, speakers, camera, GPS units,
and even with TVs and smartphones. It has an Atmega 328 which is 28 pin IC consists
of 14 digital I/O pins and 6 analog inputs and many more. Arduino Uno operates at a
frequency of 16 MHz, and dimension of the board is 68.6 × 53.44 mm. It has a flash
memory of 32 KB out of which 0.5 KB is used by the bootloader [17]. “Liquid crystal
display” (LCD) screen is an electronic display unit. It can be found in many circuits
wherever a display is required. A “16 × 2” LCD is very generic equipment and is
488 P. Kumar et al.
Fig. 1 Arduino Uno board
Fig. 2 Experimental setup

for heart rate measurement
Fig. 3 Block diagram of a

general system
commonly used in the countless number of devices. LCDs are more advantageous
than seven-segment display and other multi-segment LEDs. It is because LCDs are
cheap, easy to program, and have no limitations in showing alphabets and numerals
and so on. In a “16 × 2” LCD screen, there are two lines. Each line can show 16
characters. To control the flow of current in the circuit, a resistor is used [18] (Fig. 4).
Fig. 4 Interfacing diagram

[17]
Heartbeat sensor converts the change in the amount of blood flowing in the human
body into electrical pulses [19]. The sensor unit consists of an LED and LDR. A tip
of a finger is placed over the sensor for monitoring heartbeat. LED transmits light,
and some parts of it are absorbed by the blood cells inside the arteries within the
finger, and rest of the light rays are reflected back to the receiving unit, i.e., LDR.
The extent of light received by the LDR rest on the bulk of blood flowing over
the arteries. So, when the heart pumps the blood into the body, then the amount of
blood in blood vessels also changes. This change in the amount of blood leads to
the change in intensity of light captivated by blood cells and eventually, it leads to
variation in the magnitude of light rays acknowledged by the LDR. The resistance
of LDR changes according to the intensity of light received by it. This change in
resistance of LDR can be converted into a varying voltage by using op-amps [20].
This change is very small such that it needs to be amplified before feeding into the
microcontroller. An amplifier with large gain can amplify these small variations. The
amplified signal is passed through a low-pass filter to filter out the unwanted signals.
A computer is used to program the Arduino board and to select the age group of the
person to be monitored. The low-pass filter consists of two stage op-amps having
cutoff frequencies of 2.5 Hz each, which count up to 2.5 * 60 = 150 BPM.
2.1 Formulae for Counting Heartbeat

Pulse A = time B−time A,
490 P. Kumar et al.
Pulse B = pulse A/5,

Rate = 60, 000/pulse B
where Pulse A = five-pulse time, Pulse B = single pulse time, Time A = first pulse
counter value, and Time B = last pulse counter value.
Timer counter starts with the arrival of the first pulse. The first pulse counter
value is recorded. After this, counter waits for five pulses. Then, the counter value
for Time B is recorded. Time A is subtracted from Time B to get actual time taken
by five pulses. Then, this five-pulse time is divided by 5 to get a single-pulse time.
Finally, the pulse can be calculated by using the above rate expression, i.e., rate =
60,000/pulse. The pulse sensor module is supplied with 5 V and GND and the signal
out pin is applied to A3 pin in Arduino module as an analog input pin. The digital
output pins of Arduino pin 4, 5, 6, 7 is connected to dB4, dB5, dB6, dB7 of LCD as
a data transfer pins and RS and E of LCD is also connected to the pins of 2 and 3
of digital I/O pins in Arduino module. VCC, GND, and contrast adjustment are also
applied to the respective pins of LCD where potentiometer is adjusted to 10 K.
The heartbeat sensor counts the heartbeat of the patient in a minute and sends it to the
filters. The final heartbeat is stored in the Arduino board by using five-pulse formula.
Many countries are facing scarcity of doctors and hospitals due to the increase in
population. Many patients do not get medical facilities, so this study is aimed to help
reduce the crowd. It also decreases the cost of monitoring pulse, because it is a very
easy and simple process of doing it. It is also very user-friendly to operate; it does
not need any experts to do so. We have monitored the heartbeat of 32 persons of
four different age groups. We have also calculated the heartbeat of the athletes, and
we got the result as shown in Table 3. We have taken the heartbeat of three infants
(0–1 years), eight children of age between 1 and 4 years, and four individuals of age
between 4 to 9 years. We have also recorded pulse rate of 17 persons of age above
10 years old.
Finally, in our analysis of 32 different persons, all of them were found to be
normal. Arrhythmia was not detected in any of them. These recorded values were
cross-validated with the help of a fasttrack reflex smartwatch. The readings of the
established market product (smartwatch) were close to that of our proposed method.
This study was extended to measure the heat rate of four different persons at two
different conditions as depicted in Table 3. The values were close to the readings
obtained using smartwatch which validates the effectiveness of our proposed method.
Table 3 shows the heartbeat of three peoples in normal situations and after running
for 20 min. The heartbeat of a person increases while running because the body cells
demand more oxygen, so the heart starts to pump more blood in order to increase
the supply of oxygen and other nutrients to all the cells present in the body. In our
Fig. 5 Output displayed on the computer screen of the experimental setup
experiment, all three persons were above 18 years old. The heartbeat of a normal
adult person lies between 60 and 100 BPM. From Table 3, we can note that after
running when the person is in an idle position or taking rest, the heartbeat starts
decreasing and tends to attain normal heartbeat.
4 Conclusion
The significance of this device is that it can reduce the cost of monitoring heart
diseases like bradycardia and tachycardia. It also reduces the physician’s efforts and
helps them to know about the heartbeat of the patients. The device is mobile and
cost-efficient and only the knowledge of its working would be enough for a person
to even use it at their home without any clinician’s help. Thus, it could be also used
by the general public who has almost no prior knowledge of medical equipment and
terms associated with these diseases. As we are aware of the present-day scenario of
492 P. Kumar et al.
Table 2 Heartbeat recordings of 32 different persons of varying age group using proposed
experimental setup
S. no. Age group Beats per minute Status Arrhythmia
1. 0–1 121 Normal X
2. -do- 111 Normal X
4. 1–4 108 Normal X
6. -do- 92 Normal X
7. -do- 96 Normal X
8. -do- 97 Normal X
9. -do- 80 Normal X
12. 4–9 62 Normal X
16. Above 10 68 Normal X
Table 3 Experimentally recorded four sets of heartbeat (measured per minute) for three different
healthy persons at rest and unrest conditions
Heartbeats at rest condition Heartbeats after running
Person 1 Person 2 Person 3 Person 1 Person 2 Person 3
78 70 82 144 164 147
68 72 73 143 159 144
80 79 74 130 146 107
74 89 89 107 130 98
the less availability of doctors in many places, especially in the developing countries
like India, this type of mobile biomedical devices can be a helpful for the people as
well as for the clinicians at the local hospital where no physicians are present. This
device is an example of a collaboration of medical science with engineering.
References
1. Das S (2013) The development of a microcontroller based low cost heart rate counter for health
care systems. Int J En Trends Technol 4(2)
2. Marchon N, Naik G (2016) Detection of fetal heart rate using ANFIS displayed on a smartphone.
In: 2016 IEEE region 10 conference (TENCON), Singapore
3. Zhu Y, Chen Y, Qi L, Ma HT (2016) Spectral analysis of heart rate variability and its coherence
with pulse transit time variability in prehypertension. In: 2016 IEEE region 10 conference
(TENCON), Singapore
4. https://www.health.harvard.edu/heart-health/what-your-heart-rate-is-telling-you
5. Mehra R (2007) Global public health problem of sudden cardiac death. J Electrocardiol
40:S118–S122
6. Owis V, Abou-Zied AH, Youssef A-BM, Kadah YM (2002) Robust feature extraction from
ECG signals based on nonlinear dynamical modeling. In: Proceedings of the 23rd annual
international conference of the IEEE, engineering in medicine and biology society
7. Mishra M, Banerjee S, Thomas DC, Dutta S, Mukherjee A (2018) Detection of third heart
sound using variational mode decomposition. IEEE Trans Instrum Meas 67(7):1713–1721
8. Kumar P, Parashar S, Gautam AK, Bora J, Mishra M (2019) Techniques for lung sounds analysis
to assist medical professionals: a study. J Anal Comput 12(1):1–8
9. Mishra M, Pratiher S, Banerjee S, Mukherjee A (2018) Grading heart sounds through vari-
ational mode decomposition and higher order spectral features. In: 2018 IEEE international
instrumentation and measurement technology conference (I2MTC), Houston, TX, pp 1–5
10. Paul Nikita et al (2019) Advancements in signal processing techniques for detecting cardiac
abnormalities: a study. J Anal Comput 12(1):1–9
11. de Albuquerque VHC et al (2016) Robust automated cardiac arrhythmia detection in ECG beat
signals. Neural Comput Appl 1–15
12. Kora P, Krishna KSR (2016) ECG based heart arrhythmia detection using wavelet coherence
and bat algorithm. Sens Imaging 17(1):1–16
13. Daamouche A, Hamami L, Alajlan N, Melgani F (2012) A wavelet optimization approach for
ECG signal classification. Biomed Signal Process Control 7(4):342–349
14. Mamoon IA, Islam AKMM, Baharun S, Komaki S, Ahmed A (2016) A novel medical priority
aware transmission mechanism for cognitive radio based hospital. In: 2016 IEEE region 10
conference (TENCON), Singapore
494 P. Kumar et al.
15. Jha RS, Singh VP, Mittal VK (2016) Discriminant feature vectors for characterizing ailment
cough versus simulated cough. In: 2016 IEEE region 10 conference (TENCON), Singapore
16. Lawless ST (1994) Crying wolf: false alarms in a pediatric intensive care unit. Crit Care Med
22:981–985
17. Arduino Uno board specifications. https://www.edgefx.in/arduino-uno-board-tutorial-and-its-
applications/
18. Burke MJ, Whelan MV (1987) The accuracy and reliability of commercial heart rate monitor.
21(1):29–32
19. Fezari M, Bousbia-Salah M, Bedda, M (2008) Microcontroller based heart rate monitor. Inte
Arab J Inf Technol 5(4)
20. Adams TD, Yanowitz FG, Fisher AG, Ridges JD, Lovell K, Pryor TA, (1981 Noninvasive
evaluation of exercise training in college-age men. Circulation 64:958–965
An Iterative Node-Pair Time
Synchronization (INTS) for Wireless
Sensor Networks
Sakuru K. L. V. Sai Prakash and N. Bheema Rao
Abstract An essential function of the spread sensor network is time synchroniza-

tion. This paper presents a novel iterative node-pair time synchronization (INTS)
algorithm that assists the sensor nodes in collaborating with the anchor node in syn-
chronizing its clock. Even though the clocks on anchor and the sensor nodes are
presumed to set to the universal time, they invariably have a different offset. How-
ever, in the present formulation, we make a simplifying assumption that all the anchor
clocks and all the sensor clocks run at the same speed and taken as one. The varying
clock offsets could lead to inconsistencies in data gathering in a distributed environ-
ment. A synchronization mechanism for a global ordering of the packets generated
at different nodes is essential. Bounded but variable delays during the exchange of
message packets between the sensor nodes prevent the determination of the exact
clock offsets between the anchor transmitting node and the senor receiving nodes.
Our algorithm tries to handle this problem by limiting the uncertainty bounds to
minimum application-specific value by an iterative method. Finally, we present the
simulation and the analytical results, which are identical and encouraging.
Keywords Time synchronization · Anchor node · Sensor node · Iterative ·

Uncertainty
1 Introduction
Effective collaboration and collection of data by the small form factor sensor nodes
in the sensor field play an imperative role in sensor networks. Sensor nodes are
composed of sensors and actuators, the computational module (microcontroller and
S. K. L. V. Sai Prakash (B) · N. Bheema Rao

Department of Electronics & Communication Engineering,
National Institute of Technology, Warangal, Telangana State 506004, India
N. Bheema Rao
496 S. K. L. V. Sai Prakash and N. Bheema Rao
memory), a communication module, and a power supply. To make the right interpre-
tation of the data collected by the nodes, the time on the sensor nodes play a crucial
role. In a distributed environment, any two sensor nodes will not have the same clock
showing the same time, as different oscillators drive them. To fulfill the application
demand, time synchronization of the distributed sensor nodes is of utmost priority.
Time synchronization algorithms, contrarily, yield a beneficial solution to keep every
networked node to the coincident time scale concerning a common clock. Typical
cases to point up the need for time synchronization are:
1. Localization: The location of the randomly deployed distributed sensor nodes
heavily depend on time synchronization. As the data generated by the node should
have a tag from which location it belongs to, the node location plays a vital
role. The localization algorithms use time of arrival information to calculate the
position from different nodes; a firm agreement of time between the nodes highly
required.
2. Network access: Scheduling and accessing the network in time division also
demands a highly time-synchronized sensor network. Time-synchronized media
sharing sensor nodes coordinate for collisions avoidance and conserve energy, as
presented in Fig. 1.
3. The lifetime of sensor nodes: Duty cycling making the nodes to sleep and making
them wake up for reception and transmission of packets will extend the life of the
sensor node and also the network. For this operation, tight time synchronization
is required.
4. Target localization: When a set of sensor nodes are trying to localize the target,
time synchronization plays an essential role in calculating the position, direction,
and velocity, as presented in Fig. 2.
Because of the importance of time synchronization, as illustrated by the above
examples, we have developed a novel time synchronization algorithm that addresses
the problem due to different offsets of the individual clocks.
The remaining of the paper written as follows: Sect. 2, a brief survey of the related
work presented, and we introduce the system model, terminologies, and algorithm in
Sect. 3. An iterative node-pair time synchronization procedure and statistical analysis
are explained in Sect. 4. Discussion on the results is in Sect. 5, and finally concludes
with the future direction of our work in Sect. 6.
Fig. 1 TDMA frame

structure
An Iterative Node-Pair Time Synchronization (INTS) for Wireless Sensor Networks 497
Fig. 2 Target localization
2 Related Work
For the last few years, several time synchronization algorithms have been proposed
and evaluated for the performance of sensor networks. Most existing synchroniza-
tion schemes considered a time-stamped message exchange process for obtaining the
offset between two clocks in the distributed environment and used it as a correction
factor. The exchange process can, therefore, classified as uni-directional, bidirec-
tional, and receiver–receiver protocols [1, 2]. The network time protocol [3] and the
global positioning system are not suitable for sensor nodes as they are power hungry
and energy consuming. The predominant time synchronization algorithms, such as
RBS [4], TPSN [5], DTSP [6], and FTSP [7], are all based on timestamp exchanges
and tried to eliminate the uncertainty in exchange process such as: sent, access, prop-
agate, and receive delays. The authors in RBS protocol [4] uses the broadcast nature
of the wireless media to bring down the delays in the synchronization procedure.
The reference node transmits a sync time-stamped message and the same received
by every node in the communication range of the reference node. The nodes which
received the packet note the arrival time and exchange this information among all
the recipients and use it for finding the offset among themselves. In this method, the
delays which are common at the transmitting side are eliminated, and only receiver
side delay exists. The authors in the DTSP protocol [6] uses multiple transmitter
timestamps and receiver timestamps at each receiver node and perform recursive
least squares estimate to calculate associated time offsets and skews. The authors in
the TPSN protocol [5] find the root node and generated a spanning tree. Under the
assumption that all are symmetrical links in the spanning tree, a node-pair synchro-
nization is run long the branch of tree between parent node and child node, and round
trip measurements are performed. The communication overhead when compared to
RBS, TPSN is more, and the node failure, makes the whole process of tree formation
to repeat. The authors in the FTSP protocol [7] select the node having lowest ID as
the reference node, and use its time as the reference for synchronization. The refer-
ence node periodically broadcasts the time-stamped messages. The receiver nodes
use these timestamps to calculate the offset and skew by linear regression. As the
messages are periodically flooded, the node failure does not affect the performance
of the algorithm.
Localization and synchronization in sensor networks are well studied and have
many similarities, such as the message exchange process and use of mathematics to
estimate position and time. Due to this correlation, a small survey of the literature
on localization carried out.
In Refs. [8–10], iterative localization, range-free localization, and distributed
localization algorithms are discussed. The authors in the iterative localization [5]
uses a one-dimensional array of sensor nodes randomly placed and found the loca-
tion of the nodes, using a communication range as a disk. Based on the disk infor-
mation, an intersection method used to get the location information. Once the nodes
are localized, they help the neighbor nodes to localize. The authors in the range-free
localization [4] uses a two-dimensional array of sensor nodes randomly placed and
found the location of the nodes, using a communication range as a circular disk.
Based on the disk area intersection method, the location information is calculated.
Once the nodes are localized, they help the neighbor nodes to localize. The authors
in the distributed localization [11] uses a two-dimensional array of sensor nodes ran-
domly placed found the location of the nodes. The unlocalized nodes only iteratively
talk with localized nodes and use correlation technique to find their locations. A
square intersection model is considered.
3 Terminology, System Model, and Algorithm
The distributed clocks which are present on the sensor nodes are to be synchro-
nized utilizing message packet transmissions that undergo variable delays [4–7].
One transmission is sufficient if the application-dependant clock accuracy is more
than the delay.
Considering node-pair for the time synchronization procedure, assuming the trans-
mitter node referred to as anchor node represented by A and the clock on it by tA ,
and receiver node referred to as sensor node by R and the clock on it by tR are related
as:
tR (n) = m ∗ tA (n) + O + δ(n) (1)
where m defined as relative slope, O defined as relative offset, δ defined as the random
transmission delay of the message packet amid the anchor node and the receiver node,
and n is the nth packet transmissions. These delays hamper the synchronization
process and consist of many components, inclusive of the sender processing delay
caused by kernel processing, system calls, and context switches which are dependent
on node load processing; media accessing delay depends on the MAC procedure and
network load, media transmission delay by the radio frequency signal to transit
from anchor to receiver node, receiver receiving delay at the receiver node, the time
captured to receive the complete message and translate.
If the clocks on both the nodes are moving at the same frequency, i.e., m is 1, the
relative offset is given by:
O = tR (n) − tA (n) − δ(n) (2)
If the clocks on both the nodes have no offset, i.e., O is 0, the relative skew is
given by:
tR (n) − δ(n)
m= (3)
tA (n)
We assume δ(n) ∈ [ par , Par ] and 0 ≤ par ≤ Par < α, i.e., δ(n) are bounded ran-
dom variables and that par and Par are known. Then, after receiving n time-stamped
message packets, offset O should satisfy
O ∈ [tR (n) − tA (n) − par (n), tR (n) − tA (n) − Par (n)] for 1 ≤ n ≤ x (4)
where x is the max value where the iterative process stops when the application
specific uncertainty value is reached.
This means O ∈ [O(n), O(n)] where,
O(n) = max {tR (n) − tA (n)} − Par = O − (Par − max {δ(n)}) (5)
1≤n≤x 1≤n≤x
O(n) = min {tR (n) − tA (n)} − par = O − ( min {δ(n)} − par ) (6)
1≤n≤x 1≤n≤x

Let O(n) := O(n) − O(n) denote the width of uncertainty in our estimate of O
offset after n timestamps were received. The O corresponds the left lower bound,
and from here on, it will be represented as Le; O corresponds the right upper bound

of uncertainty interval and from hereon represented as Ri. Then, we have O(n) =
(Par − par ) − (max1≤n≤x δ(n) − min1≤n≤x δ(n)). Note that, we have not made any
assumptions on the random sequence δ(n) of delays experienced by the timestamp
packets except that they are bounded.
4 Mathematics
4.1 Estimation Procedure
Let anchor node A be transmitting clock messages, and the same is collected by the
sensor node R. The 1st timestamp message of the anchor node A is designated by
tA (1), and the 1st packet collection time by the sensor node R is designated tR (1).
The δ(1) is the random delay which is contributed by all the delays, experienced
by the 1st packet at the destination sensor node R from the anchor node A and is
Sensor Node clock Le(2) Ri(2) Le(n) Ri(n)
Re(1)=t_A(1)+O+P
Li(1)=t_A(1)+O+p
t_R(1) t_R(2) t_R(n)
t_A(2)+O+p t_A(2)+O+P t_A(n)+O+p t_A(n)+O+P

Le(1)+t_A(2)−t_A(1)Ri(1)+t_A(2)−t_A(1)
Ri(n−1)+t_A(n)−t_A(n−1
Le(n−1)+t_A(n)−t_A(n−1)
t_A(1) t_A(2) t_A(n)
Anchor Node clock
Fig. 3 Iterative time synchronization
confined by [ par , Par ]. Given single anchor node A and single sensor node R pair
under study, we discard the subscripts of the random delay bounds.
After first packet transmission, time tR (1),
tR (1) ∈ [tA (1) + O + p, tA (1) + O + P] (7)
Le(1) = max{tA (1) + O + p, 0} (8)
Ri(1) = min{tA (1) + O + P, ∞} (9)
under the assumption, the starting time of sensor R can be anywhere between [0, ∞]
(Fig. 3).
After second packet transmission, time tR (2),
tR (2) ∈ [tA (2) + O + p, tA (2) + O + P] (10)
and from advancing the first packet tR (2) lies in
tR (2) ∈ [tA (1) + O + p + tA (2) − tA (1), tA (1) + O + P + tA (2) − tA (1)] (11)
Now
Le(2) = max{tA (2) + O + p, tA (1) + O + p + tA (2) − tA (1)} (12)
Ri(2) = min{tA (2) + O + P, tA (1) + O + P + tA (2) − tA (1)} (13)

now substituting Eqs. 8 and 9
Le(2) = max{tA (2) + O + p, Le(1) + tA (2) − tA (1)} (14)
Ri(2) = min{tA (2) + O + P, Ri(1) + tA (2) − tA (1)} (15)
After nth packet transmission, time tR (n),
tR (n) ∈ [Le(n), Ri(n)] (16)
where
Le(n) = max{tA (n) + O + p, Le(n − 1) + tA (n) − tA (n − 1)} (17)
Ri(n) = min{tA (n) + O + P, Ri(n − 1) + tA (n) − tA (n − 1)} (18)
The important point here is to get the random variables and from Eqs. 17 and 18
Le(n) = tA (n) + O + p + δ(n) − min{δ(n), δ(n − 1), . . . , δ(1)} (19)
Ri(n) = tA (n) + O + P + δ(n) − max{δ(n), δ(n − 1), . . . , δ(1)} (20)
The uncertainty interval width Z (n) can therefore be written as
Z (n) = Ri(n) − Le(n) (21)
Z (n) = P − p − max{δ(n), δ(n − 1), . . . , δ(1)}

+ min{δ(n), δ(n − 1), . . . , δ(1)}
If the a-priori bounds are accurately selected, it is intuitivly clear that max
will go close to P and min will go close to p, thereby Z (n) converges to zero.
4.2 Statistical Analysis
Let
X (n) = max{δ(n), δ(n − 1), . . . , δ(1)} (22)

Y (n) = min{δ(n), δ(n − 1), . . . , δ(1)} (23)
Then
Z (n) = (P − p) − (X (n) − Y (n)) (24)
Let δ(1), δ(2), . . . , δ(n) be independent random variables with identically distributed
and has similar probability density function f (z) and distribution function F(z)
having span [ p, P]. Expectation of Z will be
E[Z ] = (P − p) − E[X − Y ] (25)
Being continuous, the joint pdf of FX Y (x, y) is
∂2
f X Y (x, y) = − Pr[X ≤ x, Y ≥ y] (26)
∂ x∂ x
All the n random delays are i.i.ds, and variables possess similar PDF f (z) with the
support [p, P].
Pr[X ≤ x, Y ≥ y] = [F(x) − F(y)]n (27)
Differentiating
∂2
[F(x) − F(y)]n = n(n − 1)(F(x) − F(y))n−2 f (x) f (y) (28)
∂ x∂ y
The joint pdf

n(n − 1)(F(x) − F(y))n−2 f (x) f (y) x ≥ y
f X Y (x, y) = (29)
0 x≤y
Expectation value of [X − Y ] is

E[X − Y ] = (x − y) f X Y (x, y)dxdy
x≥y
P x
= n(n − 1) (x − y)((F(x) − F(y))n−2 f (x) f (y)dxdy
x= p y= p
Assuming uniform density in [p, P], the distribution becomes
z−p
F(z) = z ∈ [ p, P] (30)
P−p
Expectation of [Y − Z ] is derived as
(n − 1)(P − p)
E[X − Y ] = (31)
n+1
Expectation of Z is shown by
2(P − p)
E[Z ] = (32)
n+1
The variance of Z is given by
Var[Z ] = E[Z 2 ] − [E[Z ]]2 (33)

2(P − p)2 (n − 1)
= (34)
(n + 1)2 (+2)
5 Results
The novel iterative node-pair time synchronization algorithm has analytically com-
puted, and the same is simulated in MATLAB software. The uncertainty a-priori
bounds considered as 0.3-time units as the lower bound and 0.9-time units as the
upper bound. The anchor node transmits time-stamped message packets periodically
with a regular interval of the 1-time unit. The process was monitored for the behavior
of the uncertainty width as the message transmission packets are incremented. The
i.i.d random delays assumed to take uniform density function with the above uncer-
tainty support. Both the analytical and simulation studies carried out considering
uniform distribution with the support of 0.3 and 0.9-time units and measurements
made over 100 packets, as shown in Fig. 4.
The same is drawn in the log–log scale for the bounds on uncertainty interval
with 0.3 and 0.9-time units and over 1000 transmission message packets. The graph
is shown in Fig. 5:
We present here an iterative node-pair time synchronization algorithm motivated by

the localization methods in the WSN literature. Even though it is a node-pair time
synchronization, it can also view it as a one to many single-hop synchronization
approaches. It reduces the voluminous transmission, thereby reducing the energy
at the forfeiture of the enhanced computation. From Fig 4, we can conclude that if
the data of the uncertainty width of the interval, if carefully selected, they would
much impact the convergence curve. The analytical and simulation results show
that approximately 20 anchors transmitted time-stamped message packets are in a
single-hop network for 10-time units to 1-time unit reduction in the uncertainty with
the node clock times. This profound yield improvements compared to RBS method
need 30 message packets for a pair of receivers, thereby reducing the uncertainty
Beacon Node Synchronization Simulation

0.7
Simulation
Analytical
0.6
0.5
Width of the Bounds
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50 60 70 80 90 100
No. of message packets
Fig. 4 Convergence characteristics
0 Iterative Time Synchronization in WSN

10
Width of the Bounds
−1
10
−2
10
−3
10
0 1 2 3
10 10 10 10
No. of message packets
Fig. 5 Log–log convergence characteristics

width from 11-time units to 1.6-time units. Our INTS protocol uses unidirectional
communication, while TSPN protocol employes bidirectional communication. Thus,
our protocol halves of the transmission messaging time and thus increasing energy
efficiency. The construction of a spanning tree process not necessary. In our method,
the messages are broadcasted periodically by the anchor node in predetermined time
periods, tree reformation either not required, in the event of failed node. The FTSP
authors claimed a minimum requirement of six message packets for uncertainty
reduction 10 to 1-time unit. Though these results seem to be better than ours, elec-
tion procedure in the FTSP method incurs much communication and computational
overhead. The clocks in FTSP only synchronized to the leader local time rather than
global time. In our INTS protocol, GPS assumed to be present at the anchor node,
and the time is universal. Every sensor node in the sensor network gradually synchro-
nized to the global time present with the anchor node. The results are encouraging
us to extend the work for whole network synchronization, and various strategies to
cover the entire network are to be studied.
References
1. Simic S, Shankar Sastry S (2002) Distributed localization in wireless ad hoc networks. Tech-
nical report No. UCB/ERL M02/26. EECS Department, University of California, Berkeley
2. Feng Z, Leonidas G (2005) Infrastructure establishment, chapter 4, wireless sensor networks:
an information processing approach. Elsevier
3. Sundararaman B, Buy U, Kshemkalyani AD (2005) Clock synchronization for wireless sensor
networks: a survey. Elsevier J Adhoc Netw 3(3):281–323. https://doi.org/10.1016/j.adhoc.
2005.01.002
4. Sivrikaya F, Yener B (2004) Time synchronization in sensor networks: a survey. IEEE Netw
18(4):45–50. https://doi.org/10.1109/MNET.2004.1316761
5. Mills DL (1994) Precision synchronization of computer network clocks. ACM SIGCOMM
Comput. Commun. Rev. https://doi.org/10.1145/185595.185651
6. Jeremy E, Lweis G, Deborah E (2002) Fine grained network time synchronization using ref-
erence broadcasts. ACM SIGOPS Oper Syst Rev. https://doi.org/10.1145/844128.844143
7. Ganeriwal S, Kumar R, Srivastava MB (2003) Timing-sync protocol for sensor networks. In:
SenSys’03: proceedings of the 1st international conference on embedded networked sensor
systems, Nov 2003, pp 138–149. https://doi.org/10.1145/958491.958508
8. Solis R, Borkar VS, Kumar PR (2007) A new distributed time synchronization protocol for
multihop wireless networks. In: 45th IEEE conference on decision and control. https://doi.org/
10.1109/CDC.2006.377675
9. Miklos M, Branislav K, Gyula S, Akos L (2004) Protocol the flooding time synchronization,
SenSys’04. Baltimore, Maryland, USA
10. Karnik A, Kumar A (2004) Iterative localization in wireless ad hoc sensor networks: one
dimensional case. In: IEEE SPCOM, international conference on signal processing and com-
munications, Dec 2004. https://doi.org/10.1109/SPCOM.2004.1458387
11. Stupp G, Sidi M (2004) The expected uncertainty of range free localization protocols in sensor
networks. In: International symposium on algorithms and experiments for sensor systems,
wireless networks and distributed robotics, ALGOSENSORS 2004. https://doi.org/10.1016/j.
tcs.2005.06.025
Image Captioning Methodologies Using
Deep Learning: A Review
Virendra Kumar Meghwal, Namita Mittal, and Girdhari Singh
Abstract Image captioning involves knowledge of both “Computer Vision” as well

as “Natural Language Processing” to generate a caption for digital images. Various
approaches have been used to generate captions for images. Among them, Deep
Learning based approaches have shown impressive results in image captioning. This
paper presents a concise review of recent image captioning methodologies based
on Deep Learning to automatically generate captions and discusses the datasets and
the evaluation metrics that are used to evaluate various methodologies in image
captioning.
Keywords Image caption · Visual attention · Semantic attention · Deep Learning ·

CNN · LSTM
1 Introduction
Computer Vision deals with the study of digital images and videos in order to under-
stand the content in them, so as to enable computers to understand images and videos
like humans do. While, Natural Language Processing deals with the understanding
of Natural Languages (languages used by humans), so as to enable the computer
to interact with humans. Image captioning combines both these fields of study to
enable the computer to generate captions for images. As both these tasks are difficult
for computers, thus image captioning is a much more difficult and complex task for
computers as it combines them both.
Image captioning using Deep Learning approaches automatically learns to extract
various features from the data and then these features are used to generate captions for
V. K. Meghwal · N. Mittal (B) · G. Singh

Malaviya National Institute of Technology, Jaipur, India
V. K. Meghwal
G. Singh
508 V. K. Meghwal et al.
Fig. 1 A general image captioning framework
the given image. Most of these approaches use encoder–decoder framework (Fig. 1).
In this framework, the input image is first fed into an encoder. The encoder, which is
generally a CNN, finds various features and encodes it into a vector called context
vector. This vector is passed as an input to the decoder, which decodes it and generates
an appropriate caption for the given image. Attention-based encoder–decoder models
instead of taking context vector as input, takes input from the last convolution layer
of the CNN Encoder, so as to preserve spatial details of various features (Fig. 1).
2 Datasets
There are a few open-source datasets that are available to be used for image
captioning. Three of them are listed below.
2.1 Mscoco [1]
This is the largest dataset available for image captioning. It contains around 330 K
images and five captions per image. Most researchers who are working on image
captioning prefer to work on this dataset as it is the largest dataset and has lesser
noisy data.
2.2 Flickr30k [2]
This is the second most popular dataset in the image captioning domain. It contains
around 31,783 images, with 158,915 crowd-sourced captions describing them.
2.3 Flickr8k [3]
It is a predecessor of the Flickr30k dataset. It contains eight thousand images with

five captions each.
Image Captioning Methodologies Using Deep Learning: A Review 509
3 Evaluation Metrics
3.1 Bleu [4]
BLEU stands for Bi-lingual Evaluation Understudy. This evaluation metrics is

popular in machine translation. As, the earlier no better metric was available for
measuring image captioning, so BLUE was used for this purpose. It does not take
grammatical correctness of sentences into account. It ranges from 0 to 1.
3.2 Rouge [5]
Recall-Oriented Understudy of Gisting Evaluation is a popular metric for summa-

rization evaluation. It ranges from 0 to 1.
3.3 Meteor [6]
It stands for Metric for Evaluation of Translation with Explicit ORdering. It

was developed for measuring machine translation. Compared to BLUE, it has a
significantly better correlation with human judgments.
3.4 CIDEr [7]
Consensus-based Image Description Evaluation was developed for evaluating image

captioning. It measures the similarity of a generated sentence against a set of ground-
truth sentences written by humans.
3.5 Spice [8]
Semantic Propositional Image Caption Evaluation was developed for evaluating

image captioning. It is based on semantic scene graphs and has shown that it correlates
much better with human judgments.
4 Methods
Various methodologies have been proposed for generating captions from images. A
few of the recent methodologies are discussed below.
Huang et al. [9] proposed a pipelined framework consisting of two cascaded
agents in the decoder. In this model, the image is encoded using a Faster-RCNN, this
encoded vector is then passed to the decoder for caption generation. The decoder
consists of a semantic adaptive agent and a caption generating agent. In the semantic
adaptive agent, this encoding is passed to an LSTM, the output of the LSTM is
fed into the visual attention module to generate a vector which is passed to the
semantic adaptive attention module along with the output of the LSTM to generate
the output vector of the semantic adaptive agent. This output vector is then passed
to the caption generating agent which consists of an LSTM followed by softmax, to
generate the caption. Peri et al. [10] used a CNN for encoding image and an LSTM
for encoding captions (during training) and then these features were projected into
a Common Vector Space which aligns similar or correlated concepts in images and
texts and dissimilar ones far apart. This CVS representation is then used by decoder
LSTM to generate semantically similar captions. Huang et al. [11] proposed an
attribute refinement network to model the correlation between attributes. For this,
they utilized joint training on images and captions which helped them to boost the
performance of their model. They used image features, object features and attribute
distribution in their model. The model employs a combination of an attention LSTM,
a visual attention module, a language LSTM along with attribute refinement and
semantic attention to generate captions. He et al. [12] proposed Visual-Semantic
Double Attention (VSDA) model with a new semantic attention model. This model
uses two CNNs to extract visual attention features and semantic attention features
which are then fed as an extra input to the LSTM for generating caption. These
extra inputs help the LSTM to recognize the relevant semantic features and reduce
the effect of irrelevant features. Shi et al. [13] proposed a cascade attention module
that enables high priority inputs to affect the attention of other inputs to emphasize
inequality among inputs. In this model, the encoder uses a Faster-RCNN to extract a
feature vector representing features of each region of interest. Another feature vector
representing features of the global context in the image is also extracted by encoder.
The decoder consists of a cascaded attention module and two LSTMs—Language
LSTM and Attention LSTM. In the decoder, the cascaded attention module attends to
these regions of interest and global context based on the hidden state of the attention
LSTM. The language LSTM builds a visual sentinel vector to represent semantic
context information which helps to determine whether the language LSTM requires
the visual features as input or not. Based on this language LSTM generates each
output word of the caption.
Zhang et al. [14] combined spatial and adaptive attention into a transformer. This
model uses a CNN to extract the image features into a context vector which is then fed
into the adaptive transformer in place of LSTM in the decoder to generate captions.
The transformer helps the model to determine when and where to use image features
Image Captioning Methodologies Using Deep Learning: A Review 511
and achieved better performance and boost in training speed. Li et al. [15] proposed
a visual-semantic LSTM model. They used Region Proposal Network to extract low-
level features and a CNN to extract high-level features. The low-level features are
passed to LSTM-V to localize the objects in the image. The output of LSTM-V and
the high-level features generated by CNN are passed to LSTM-S to generate the
corresponding words in the caption.
5 Comparison of Evaluation Metrics
See Tables 1 and 2.
Table 1 Comparison among few recent methodologies in image captioning (dataset: MSCOCO)
Approach name Methodology Results
Show, translate and CNN + common vector space B1 = 70.6, B2 = 53, B3 = 38.5,
tell [10] + RNN B4 = 27.9, M = 24.6
Visual-semantic double CNN (visual-semantic attention) B1 = 75.3, B2 = 59.1, B3 =
attention [12] + LSTM 45.1, B4 = 34.4, M = 26.5, R =
55.2, C = 106.3
Attribute refinement [11] CNN + visual Att + attribute B1 = 77, B2 = 61, B3 = 47.1,
refine + semantic Att + LSTM B4 = 36.3, M = 27.6, R = 56.6,
C = 115.1, S = 20.9
Cascade attention [13] R-CNN + cascade Att + (lang B1 = 79.4, B2 = 63.7, B3 =
LSTM + Att LSTM) 48.9, B4 = 36.9, M = 27.9, R =
57.7, C = 122.7
Two Cascaded Agents [9] CNN + visual Att + LSTM + B1 = 80.6, B4 = 38.2, M = 28.3,
semantic Att + LSTM R = 58.4, C = 126.4, S = 21.7
Table 2 Comparison among few recent methodologies in image captioning (dataset: Flickr30k)
Approach name Methodology Results
Show, translate and CNN + common vector space B1 = 61.1, B2 = 42.7, B3 =
tell [10] + RNN 29.3, B4 = 20.3, M = 19.3
Adaptive transformer [14] CNN + transformer B1 = 67, B2 = 49.6, B3 = 35.5,
B4 = 25.2, M = 20.4, C = 53
Visual-semantic double CNN (visual-semantic B1 = 68.1, B2 = 49.8, B3 =
attention [12] attention) + LSTM 35.7, B4 = 25.6, M = 20.8, R =
47.4, C = 53.2
Visual-semantic (Region proposal network + B1 = 75.5, B2 = 57.1, B3 =
LSTM [15] CNN) + LSTM 42.9, B4 = 31.7, M = 22.9, C =
71.5
6 Conclusion
A lot of effort has been put by researchers in image captioning, thus, resulting in a
vast improvement over the past few years. This paper reviewed some of these recent
methodologies used in the image captioning domain. Despite the vast improvement,
there is still scope for improvement in methodologies. Further research direction may
be an improvement in visual and semantic attention, evaluation metric, etc.
References
1. MSCOCO Homepage. http://cocodataset.org/. Last accessed 02 Oct 2019

2. Plummer BA, et al (2015) Flickr30k entities: Collecting region-to-phrase correspondences
for richer image-to-sentence models. In: Proceedings of the IEEE international conference on
computer vision
3. Hodosh M, Young P, Hockenmaier J (2013) Framing image description as a ranking task: data,
models and evaluation metrics. J Artif Intell Res 47:853–899
4. Papineni K, et al (2002) BLEU: a method for automatic evaluation of machine translation.
In: Proceedings of the 40th annual meeting on association for computational linguistics.
Association for Computational Linguistics
5. Lin CY (2004) Rouge: a package for automatic evaluation of summaries. Text summarization
branches out
6. Banerjee S, Lavie A (2005) METEOR: an automatic metric for MT evaluation with improved
correlation with human judgments. In: Proceedings of the ACL workshop on intrinsic and
extrinsic evaluation measures for machine translation and/or summarization
7. Vedantam R, Lawrence Zitnick C, Parikh D (2015) Cider: consensus based image description
evaluation. In: Proceedings of the IEEE conference on computer vision and pattern recognition
8. Anderson P, et al (2016) Spice: semantic propositional image caption evaluation. In: European
conference on computer vision. Springer, Cham
9. Huang L, Wang W, Wang G (2019) Image captioning with two cascaded agents. ICASSP 2019
IEEE international conference on acoustics speech and signal processing (ICASSP). IEEE
10. Peri D, Sah S, Ptucha R (2019) Show, translate and tell. arXiv preprint arXiv:1903.06275
11. Huang Y, et al (2019) Image captioning with attribute refinement. In: 2019 IEEE international
conference on image processing (ICIP). IEEE
12. He C, Hu H (2019) Image captioning with visual-semantic double attention. ACM Trans
Multimed Comput Commun Appl (TOMM) 15(1):26
13. Shi J, Li Y, Wang S (2019) Cascade attention: multiple feature based learning for image
captioning. In: 2019 IEEE international conference on image processing (ICIP). IEEE
14. Zhang W, et al (2019) Image caption generation with adaptive transformer. In: 2019 34rd youth
academic annual conference of Chinese association of automation (YAC). IEEE
15. Li N, Chen Z (2018) Image captioning with visual-semantic LSTM. IJCAI
Performance of Speaker Recognition
System Using Kernel Functions
Approach for Different Noise Levels
Renu Singh, Arvind Kumar Singh, and Utpal Bhattacharjee
Abstract Speaker recognition is one of the most popular voices biometric tech-
nique used for security reason in many areas like banking system, online-shopping,
database access etc. The recognition performance of speaker recognition system is
very satisfactory in noise-free environments, whereas the improved performance in
case of low level signal-to-noise ratio (SNR) is the need in the present days. Hence,
in this study speaker recognition performance has been evaluated at different signal-
to-noise ratios using SVM-based various kernel functions approach and principal
component analysis (PCA). The proposed scheme has been applied on NIST 2003
and AURORA dataset and found that the recognition accuracy and running time
improves at low level SNRs using kernel functions.
Keywords Speaker recognition system · SVM · Principle component analysis

(PCA) · Signal-to-noise ratio
1 Introduction
Speaker recognition is a method of recognizing a speaker’s voice by utilizing explicit

information comprised in speech waves [1, 2].This methodology can be used to
substantiate the characteristics claimed by the person who wants to access the systems
i.e. enabling the access control of different services with the help of voice. The various
applications of speaker recognition system are voice dialing, phone banking, shop-
ping over phone, database access services, reservation services, services on voice mail
R. Singh (B) · A. K. Singh

A. K. Singh
U. Bhattacharjee
Department of Computer Science, Rajiv Gandhi University, Itanagar, Arunachal Pradesh 791112,
India
514 R. Singh et al.
and controlling security aspects for secret data and remote access of PCs. Another
significant use of speaker recognition technology is its utilization for criminolog-
ical purposes [3]. Generally, the speaker recognition by a machine comprises feature
extraction, speaker modeling or training and testing [4]. The estimation of the speaker
specific features from the speech signal is done using feature extraction. Thereafter,
the speaker model is trained with the use of these features. Finally, in the stage of
testing, recognition of speaker is performed by matching with the pattern.
Support vector machine (SVM) is one of the highly used classification technique in
analyzing the performance of speaker recognition [5–7]. The support vector machine
provides a natural solution for binary class problems. It is one of the popular learning
techniques for solving regression and classification problems. SVM has been applied
in various areas viz. face, text, 3D object, speech, speaker recognition systems. SVM
uses different kernel functions such as linear, nonlinear, polynomial, radial basis,
sigmoid, Gaussian, cubic and quadratic etc. Astuti and Adiwijaya have carried out
SVM with principal component analysis (PCA) for the classification of microarray
data [8]. Bassma have used SVM for the improvement of vehicle localization in urban
canyons [9]. Zhao et al. [10] have used nonlinear kernel nuisance attribute projec-
tion for speaker verification using NIST SRE data. Kernel PCA feature extraction
along with support vector machine classification are effectively used for through-
wall human being detection [11]. Das et al. have used linear polynomial and radial
basis function kernels for the analysis of gene data set [12].
The focus of this work is to investigate the accuracy of the speaker recognition
system in different noise levels using different kernel functions viz. linear, quadratic,
cubic and Gaussian.
2 Speech Data Classification
The various steps in this analysis consist of pre-processing/feature extraction, classi-

fication and performance evaluation as shown in Fig. 1. In the feature extraction step,
principal component analysis (PCA) has been used for dimension reduction which
uses mapping function. The different processes which are used in this extraction are
data centering, covariance matrix calculation, Eigen value and Eigen vector calcu-
lation. In classification, SVM helps in modeling the complex data set and separate
different classes and maximize the margin. SVM uses a nonlinear mapping approach
for converting the original data to a high dimension and helps in linear separation
of data creating a hyperplane (decision boundary). The performance evaluation is
executed in the final step.
In this work, two different types of experiments are performed. Initially the exper-
iment is performed to analyze the effectiveness of number of components considered
in PCA. The various experiments are carried out on NIST 2003 SRE clean data set to
visualize the effect on the accuracy of various Kernel functions. The second exper-
iment in the proposed method is performed to visualize the recognition accuracies
as well as running time. The experiments are conducted on noisy data namely: 0, 5,
Performance of Speaker Recognition … 515
Input Evaluation
of
Speech
Performance
Data
Pre- Noise
Processing Levels
Classification
Fig. 1 Performance evaluation steps
10, 15 and 20 dB signal-to-noise ratio. The total numbers of 12 scenarios of experi-

ment were considered for each data set, as shown in Table 1. The general classifier
characteristics for Kernel function operation has been given in Table 2.
Table 1 Experiment
Scenario Dimension reduction SVM Kernels
scenarios [8]
1. None Linear
2. None Quadratic
3. None Cubic
4. None Fine Gaussian
5. None Medium Gaussian
6. None Coarse Gaussian
7. PCA Linear
8. PCA Quadratic
9. PCA Cubic
10. PCA Fine Gaussian
11. PCA Medium Gaussian
12. PCA Coarse Gaussian
Table 2 General classifier

Kernel function Linear, quadratic, cubic, Gaussian
characteristics
Kernel scale Automatic
Box constraint level 1
Multiclass method One-vs-One
Standardize data True
516 R. Singh et al.
3 Kernel Functions and Expressions
The various kernel functions used in the SVM classifier are linear, quadratic, cubic
and Gaussian functions. The expressions of these kernel functions are expressed as
follows [9];
Linear function:

k xi , x j = xiT x j + c (1)
Quadratic function:

xi − x j n
k xi , x j =1− (2)
xi − x j n + c
Cubic function:
n
k xi , x j = xiT x j + 1 (3)
Gaussian function:

xi − x j n
k xi , x j = exp − (4)
2σ 2
where xi and x j are input space vectors, n is order of the kernel and c is a constant.
4 Results and Analysis
The effect of kernel function approach for the clean speech data and noisy data
at different SNR has been studied in this paper. For the effective analysis, various
scenarios with and without PCA are considered for clean speech data which can be
seen in Table 3. It is clearly seen that accuracy percentage of scenario using PCA is
found to be improved except linear case, whereas there are no changes in accuracy
percentages for fine and coarse Gaussians. The comparison of accuracy for with and
Table 3 Accuracy (%) of SVM Kernel function in various scenarios

Case Linear Quadratic Cubic Fine Gaussian Medium Coarse
Gaussian Gaussian
Without PCA 95 85 85 85 85 100
With PCA 90 90 90 85 100 100
without PCA are done at number of components after PCA equals 1 and variance
kept after PCA equals 95%.
Here, total 9 numbers of experimental scenarios are considered for the calculation
of the accuracies of 6 SVM Kernel functions. The various scenarios for different
variance thresholds which are 36.5, 48.44, 66, 80.14, 90.34, 95.08, 97.52, 98.85
and 100% are considered as per Astuti and Adiwijaya [9]. To check the Kernel
functions accuracy the first scenario chooses a minimum number of components and
the corresponding value of the variance ≥36.5%, and so on. The results are calculated
for 6 numbers of SVMs. The accuracies for various numbers of components and
variance for different scenarios are presented in Table 4.
The experimental results for noisy speech data using various scenarios and SNRs
are shown in Table 5. The accuracies with PCA are found to be increased for linear,
Table 4 Accuracyfor PCA parameters for clean speech data

Components after PCA 1 2 4 8 16 25 34 39 39
Variance kept after PCA(%) 36.5 48.44 66 80.14 90.34 95.08 97.52 98.85 100
Accuracy (%)
Linear 75 80 85 90 75 75 95 75 75
Quadratic 75 85 90 85 80 90 90 85 90
Cubic 70 85 90 90 80 95 85 85 90
Fine Gaussian 65 80 85 75 75 75 85 75 75
Medium Gaussian 75 80 95 90 75 75 75 75 75
Coarse Gaussian 75 75 75 75 75 75 75 75 75
Average 72.5 80.8 86.6 84.2 76.66 80.83 84.16 78.33 80
Table 5 Accuracyfor proposed methods for noisy speech data

Scenario Accuracy (%)
0 dB 5 dB 10 dB 15 dB 20 dB
1. 60 70 75 85 95
2. 65 70 85 90 90
3. 60 70 85 90 90
4. 60 65 60 85 85
5. 65 65 65 85 85
6. 60 65 60 85 95
7. 70 75 75 85 90
8. 75 75 80 85 85
9. 70 85 80 90 85
10. 60 65 60 85 85
11. 65 65 60 85 95
12. 60 65 60 85 95
518 R. Singh et al.
Table 6 Training time for proposed methods

Scenario 0 dB 5 dB 10 dB 15 dB 20 dB
1. 0.5883 1.0826 1.0440 1.1933 1.1467
2. 0.5903 1.0258 1.0317 1.2308 1.0862
3. 0.6298 1.0316 1.0216 1.0771 1.0594
4. 0.7259 1.0421 1.1473 1.0756 1.0530
5. 0.5891 1.0299 1.0804 1.1434 1.0081
6. 0.6342 1.1789 1.0507 1.0156 1.1173
7. 0.5663 0.5293 0.5312 0.5538 0.6021
8. 0.5876 0.5774 0.6192 0.7662 0.6056
9. 0.5881 0.5767 0.5321 0.5983 0.5410
10. 0.5815 0.5412 0.6226 0.7907 0.6189
11. 0.5887 0.5477 0.5414 0.6390 0.5456
12. 0.6125 0.5477 0.5508 0.5913 0.5497
quadratic and cubic Kernels at 0 and 5 dB SNRs. However, it remains same for all
the Gaussian Kernels. It is also seen that the accuracies do not improve for higher
SNRs in the range of 10–20 dB. The running times for all the scenarios for different
SNRs are depicted in Table 6. It is seen that PCA greatly reduces the running time
for all the SNRs considered.
5 Conclusion
In the present work, experiments have been performed using NIST 2003 and
AURORA database for the analysis of Kernel functions at various scenarios for
different SNRs. The accuracy and running time of SVM using Kernel approach and
PCA are found to be improved for low level SNRs. The quadratic Kernel function
is found to be better at 0 dB, whereas Cubic Kernel function is found to be better at
5 dB SNR.
References
1. Furui S (1991) Speaker-dependent-feature extraction, recognition and processing techniques.

Speech Commun. 10(6):505–520
2. Vuppala AK, Rao KS (2013) Speaker identification under background noise using features
extracted from steady vowel regions. Int J Adapt Control Signal Process 27:781–792
3. Kunzel H (1994) Current approaches to forensic speaker recognition. In: ESCA workshop on
automatic speaker recognition, identification and verification, pp 135–141
4. Shridhar M, MohanKrishnan N (1982) Text-independent speaker recognition: a review and
some new results. Speech Commun 1:257–267
5. Wan V, Renals S (2005) Speaker verification using sequence discriminant support vector
machines. IEEE Transa Speech, Audio Process 13(2):203–210
6. Sturim DE, Campbell WM, Reynolds DA (2007) Classification methods for speaker recogni-
tion. LNAI 4343:278–297
7. Chakroun R, Zouari LB, Frikha M, Hamida AB (2015) A novel approach based on support
vector machines for automatic speaker identification. IEEE conference
8. Astuti W, Adiwijaya (2018) Support vector machine and principal component analysis for
microarray data classification. J Phys Conf Series 971:012003
9. Bassma G, Hassan EG, Tayeb S (2018) Support vector machines for improving localization in
Urban canyons. MATEC Web Conf 200:00004
10. Zhao X, Dong Y, Yang H, Zhao J, Lu L, Wang H (2008) Nonlinear kernel nuisance attribute
projection for speaker verification. In: IEEE international conference on acoustics, speech, and
signal processing, 4125–4128
11. Wang W, Zhang M, Wang D, Jiang Y (2017) Kernel PCA feature extraction and the SVM
classification algorithm for multiple-status, through-wall, human being detection. EURASIP J
Wireless Commun Netw 151:1–7
12. Das SR, Das K, Mishra D, Shaw K, Mishra S (2012) An empirical comparison study on kernel
based support vector machine for classification of gene expression data set. Procedia Eng
38:1340–1345
A Fuzzy Multicriteria Methodology
for Selection Among Solar PV Adoption
Barriers in India
Anbesh Jamwal, Akshay Patidar, Rajeev Agrawal, and Monica Sharma
Abstract At present selection of energy policies of any country affect its environ-
ment and economic development. There is a need for the selection of the most appro-
priate energy policy which can promote the economic development of the country. In
developing countries like India, there is a lack of energy resources and these sources
are exhausting day by day due to an increase in population. To preserve the un-
conventional sources of energy Govt. of India is now promoting renewable sources
of energy. In the last few years, studies were carried out which were focused on the
best energy policy selection and these studies are based on the hybrid multicriteria
decision-making techniques like Fuzzy-AHP. In the last few years, the fuzzy set
theory has been proved out as a powerful tool to treat the uncertainty when informa-
tion is vague or incomplete. In the present study fuzzy multicriteria decision making,
the based methodology is suggested to assess the barriers which resist the adoption
of Solar PV panels in the Himachal Pradesh region. The methodology is based on
the AHP (analytical hierarchy process) under the fuzziness which allows the ranking
of barriers from the industry experts and inputs from Govt. of Himachal Pradesh to
be crisp, linguistic expressions and fuzzy number. In the application of this study,
critical barriers are found which resist the implementation of Solar PV panels in
Himachal Pradesh.
Keywords Solar PV · Barriers · Fuzzy-AHP · Triangular fuzzy numbers · India
1 Introduction
Energy is an important factor in the present time for the economic as well as social
development of any nation [1]. At present, Electricity has become a promising energy
A. Jamwal · R. Agrawal · M. Sharma (B)

Department of Mechanical Engineering, Malaviya National Institute of Technology, J.L.N. Marg,
Jaipur, Rajasthan 302017, India
A. Patidar · M. Sharma
Department of Management Studies, Malaviya National Institute of Technology, J.L.N. Marg,
Jaipur, Rajasthan 302017, India
522 A. Jamwal et al.
source to fulfil human needs. In the last few decades, the demand for electricity has
grown due to the rapid advancement in technology [2]. Developing countries like
India, where economic development is depending on industries [3]. All the major
industries sectors such as the Information Technology sector and the Manufacturing
sector consumes a large quantity of energy. India is the third largest producer of elec-
tricity [4]. But still, there are many areas in India that are far away from the regular
supply of electricity [5]. The production level of Electricity in India is not enough to
meet the rising demand from Industries and other consumers [6]. At present, India is
focusing more on gas-based power, hydro energy and renewable sources of energy.
However, Govt. of India is also looking for the adoption of clean coal technology to
meet the rising demands of electricity. The generation of electricity from renewable
sources of energy using PV systems is the major need of the present generation [7]. At
a global level, Europe has the most developed PV market, where 3% of total electricity
is generated by PV systems [8]. The electricity consumption level of any country
indicates its economic prosperity. Himachal Pradesh is situated in the northern region
of India. In Himachal Pradesh, the growing population and rapid growth in the indus-
trial area, an increase in the standard of living in cities has increased the dependence
on imported energy. Himachal Pradesh energy consumption has risen in the last
10 years because of the combined demand for urbanization and industrialization.
Solar energy contributes to the power generation in the Himachal Pradesh, although
this region of India receives the maximum sunlight exposure which has potential for
the energy generation from Solar Photovoltaic systems. Govt. of Himachal Pradesh
has launched several schemes to promote solar energy in state but sadly genera-
tion of energy from these sources remained narrow and unpractical when compared
to conventional electricity generation because of a lot of Implementation barriers.
Traditionally, solar energy in India has been only used for limited applications [9].
But as our fossil fuels are consuming rapidly due to an increase in the population
of the world in the last few decades. Man is moving towards renewable sources of
energy to secure the future of upcoming generations [10, 11]. Photovoltaic systems
have become more popular and much competitive at the present time but the diffu-
sion of PV systems is still less than in the comparison to conventional energy sources
[12]. At present, the production of electricity is shifting from conventional sources
of energy to renewable sources of energy to save the environment from the effect
of fossil fuels. Another reason for shifting is to reserve fossil fuel sources for future
generations [13]. Generation of electricity from renewable sources like solar energy
a profuse means and use of PV systems is one means.
Diffusion research on a PV system is old since the 1980s. Few studies have
carried out on the economic perspective and analysis of policies [14]. Many studies
have been carried out in past not only based on the countries, for example, India
[15], Pakistan [16], China [17], Sri Lanka [18] and Nepal [19], but also the type
of applications like off grid systems and On-grid systems [20]. In past, the cost
and technology of PV systems were not assumed to commercially competitive with
conventional sources of energy but at present because of the maturity in technology
and lesser production costs made the solar PV systems very popular [21]. However,
after technology advancement and lesser costs still, the adoption of the solar PV
A Fuzzy Multicriteria Methodology … 523
system is limited as compared to conventional sources of energy. There are always

some barriers that affect the adoption of PV systems over conventional sources of
energy [22]. The study of these adoption barriers has been a common interest in
many studies which include renewable energy research. However, only a few studies
report about the analysis of these barriers with the use of MCDM techniques and
very fewer studies reported about the assessment of these barriers with the hybrid
techniques like Fuzzy-AHP [23]. To fill this gap, the aim of the present study is to
investigate the barriers in the adoption of Solar PV in Himachal Pradesh and the
assessment of these barriers with Fuzzy-AHP. To address this, data is collected from
industry experts, Govt. of Himachal Pradesh and academia related to this field. The
rest of the paper has 5 Sections. There are 13 major barriers are found out in the
study. These 13 barriers are divided into three criteria. The most influencing barrier
is found out with the ranking of barriers done by Fuzzy-AHP.
2 Methodology
In the present paper, a fuzzy multicriteria decision making methodology is recon-

structed to select the most influencing adoption barrier among all barriers. Fuzzy-
AHP method is used in this study to prioritize the weights of adoption barriers. In
the Analytical hierarchy process experts related to the research, the area has to give
a definite input number which varies from 1 to 9 scale for the pairwise comparison
which helps in finding the priority vector [24]. If the two factors in the study have
equal importance then it has a scale of 1; if a factor is weakly important or strongly
important than other than it has a scale of 3, 5, 7 and 9 on the basis of importance
over another. The scale used in this study is discussed in Table 1. Even sometimes
the scale of 2, 4, 6 and 8 also used when there is a slight difference between two
classifications. Reciprocals of 1, 1/2, 1/3, 1/4, …, 1/9 used for the reverse comparison
in the study. Sometimes, factor comparison involves certainty and subjectivity which
don’t give accurate results of the study. For example, In a study may be expert know
that one factor is important than another factor but he may not be able to give the
definite scale for the comparison because he is not sure about the degree on factor
A over the factor B. Experts provides a range of scale from 3 to 7 to compare the
two factors but due to lack of adequate information sometimes it is difficult to do
Table 1 Saaty rating scale

Scale Description
for study
1 Equal importance
3 Less weakly importance
5 Strong weakly importance
7 Strongly importance
9 Very strongly importance
Table 2 Adoption barriers in

Barrier No. Barrier name Barrier notation
solar PV systems in Himachal
Pradesh 1. Effect of temperature Barrier B1
2. Shadowing Barrier B2
3. Irradiance Barrier B3
4. High initial investment cost Barrier B4
5. Absorption and transmittance Barrier B5
6. Power efficiency Barrier B6
7. Cost of material used Barrier B7
8. Solar panel mounting Barrier B8
9. Cloud cover Barrier B9
10. Ineffective quality control Barrier B10
11. Competition with land used Barrier B11
12. Solar plant infrastructure Barrier B12
13. Humidity Barrier B13
Table 3 List of criteria of

Criteria No. Criteria name
barriers in the adoption of
solar PV systems C1 Economic perspective
C2 Technical perspective
C3 Atmospheric perspective
comparison. In such cases, the AHP method failed due to the existence of Fuzziness
or incomplete comparisons. [25] Fuzzy-AHP is an extension of the AHP method
which was introduced earlier of 1983. With the development in MCDM over the
years and application of these MCDMs in engineering, there is a lot of development
in these techniques [26]. In the present study modified Fuzzy-AHP is used to find
the most influencing barrier among all barriers which resist the adoption of Solar PV
systems in Himachal Pradesh. Table 2 shows the barriers in the adoption of a Solar
PV system. Table 3 shows the list of criteria.
2.1 Consistency Ratio Check for Pairwise Comparison
In the present study following steps are followed to measure the consistency ratio.
1. Calculation of relative weight and Eigen-vector and λmax for each of the matrix.
2. Computation of CI (consistency index) for each of the matrix by the following
equation:
Table 4 CR calculation for

Criteria 3RD root Priority vector
criteria’s
C1 1.0000 0.2583
C2 2.4662 0.6370
C3 0.4055 0.1047
Lambda max 3.0384
CI 0.0192
CR 0.0331
CI = (λmax − n)/(n − 1) (1)
3. The consistency ratio of each of the matrix is calculated by following mathemat-

ical equations.
CR = (CI/RI) (2)
2.2 Questionnaire Design and Experts of Solar PV Industries

Response Collection
The questionnaire survey for this study is carried out by preparing a questionnaire
and responses are collected from 28 experts related to the Solar PV field. After
the expert’s selection for this study, detailed information about this study objective
and reason for this study were sent to experts by mails. At last, 28 responses were
collected related to this study from experts. The experience of experts were more
than 10 years in this area. The nine-point scale was used to assign the relative weights
for the pairwise comparison of both the criteria and barriers under criteria. Further,
the consistency ratio calculation for criteria’s and barriers under different criteria’s
is shown in Tables 4, 5, 6, and 7.
2.2.1 Conversion of Crisp Data into Triangular Fuzzy Numbers
From the data collected from experts and consistency ratio checked from the pairwise
comparison, it is found that all the matrices are consistent. A TFN (triangular fuzzy
number) is represented by (a, b, c) and the member function is defined as:
⎧
⎪
⎨ b−a , a ≤ x ≤ b
x−a
μv(x) = μv(x) = c−b , b ≤ x ≤ c

c−x
(3)
⎪
⎩ 0, otherwise
Table 5 Consistency ratio

C1 Criteria 3RD root Priority vector
calculation for barriers under
criteria C1 B3 3.4087 0.5574
B5 0.0441 0.0441
B6 1.8481 0.3022
B8 0.5886 0.0962
Lambda max 4.202
CI 0.0674
CR 0.0748

C1 criteria 3RD root Priority vector
criteria C2 B3 4.2129 0.5574
B5 1.3161 0.2045
B6 0.6148 0.0955
B8 0.2934 0.0456
Lambda max 4.1901
CI 0.0634
CR 0.0704

C3 criteria 3RD root Priority vector
criteria C3 B4 1.6787 0.3020
B7 0.8027 0.1444
B10 1.5518 0.2792
B11 1.0845 0.1951
B12 0.4409 0.0793
Lambda max 6.7491
CI 0.4373
CR 0.3904
The crisp data collected from the responses of experts were converted to fuzzy
data by the use of TFN. The nine-point scale was used in the study for the pairwise
comparison [27]. TFN is used for the pairwise comparison representation. The scale
is given in Table 8.
The conversion of crisp data into the fuzzy numbers is done by a formula shown
in the equation below:

1 1
A = x − , x, x + (4)
2 2
Table 8 Conversion of crisp inputs to triangular fuzzy number

Crisp number Triangular fuzzy number
Alpha Beta Gamma
1 1.0000 1.0000 1.0000 1.0000
2 2.0000 1.5000 2.0000 2.5000
3 3.0000 2.5000 3.0000 3.5000
4 4.0000 3.5000 4.0000 4.5000
5 5.0000 4.5000 5.0000 5.5000
6 6.0000 5.5000 6.0000 6.5000
7 7.0000 6.5000 7.0000 7.5000
8 8.0000 7.5000 8.0000 8.5000
9 9.0000 9.0000 9.0000 9.0000
1/2 0.5000 0.4000 0.5000 0.6667
1/3 0.3333 0.2857 0.3333 0.4000
1/4 0.2500 0.2222 0.2500 0.2857
1/5 0.2000 0.1818 0.2000 0.2222
1/6 0.1667 0.1538 0.1667 0.1818
1/7 0.1429 0.1333 0.1429 0.1538
1/8 0.1250 0.1176 0.1250 0.1333
1/9 0.1111 0.1111 0.1111 0.1111
The alpha cut value of 0.5 is considered in the present case for the conversion of
crisp data into fuzzy numbers. The alpha cut values of zero are considered for both
upper bound and lower bound numbers.
Table 9 shows the fuzzy weight vector calculations for criteria; Table 10 shows the
weight vector calculations for C1 criteria. Table 11 shows the fuzzy weight vector
calculations for C2 criteria. Table 12 shows the fuzzy weight vector calculations for
C3 criteria.
Table 13 shows the barriers affecting the adoption of Solar PV in Himachal Pradesh
with their normalized global weights. Barrier B1 (effect of temperature) with an
Table 9 Fuzzy weight vector

Alpha Beta Gamma
calculations for criteria’s
C1 0.8939 1.0000 1.1187
C2 2.2407 2.4662 2.6801
C3 0.3731 0.4055 0.4463
Sum 3.5077 3.8717 4.2450
Fuzzy weight vector 0.2106 0.2583 0.3189
0.5278 0.6370 0.7640
0.0879 0.1047 0.1272
Table 10 Fuzzy weight

Alpha Beta Gamma
vector calculations for
barriers under C1 B3 3.1721 3.4087 3.6280
B5 0.2551 0.2697 0.2876
B6 1.7003 1.8481 2.0154
B8 0.5362 0.5886 0.6448
Sum 5.6636 6.1151 6.5758
0.0388 0.0441 0.0508
0.2586 0.3022 0.3559
0.0815 0.0962 0.1138

Alpha Beta Gamma
B2 1.1959 1.3161 1.4382
B9 0.5555 0.6148 0.6812
B13 0.2756 0.2934 0.3152
Sum 6.0551 6.4371 6.8242
0.1752 0.2045 0.2375
0.0814 0.0955 0.1125
0.0404 0.0456 0.0521

Alpha Beta Gamma
B7 0.7028 0.8027 0.9221
B10 1.3488 1.5518 1.7654
B11 0.9607 1.0845 1.2332
B12 0.4097 0.4409 0.4797
Sum 4.8870 5.5587 6.3002
0.1115 0.1444 0.1887
0.2141 0.2792 0.3613
0.1525 0.1951 0.2523
0.0650 0.0793 0.0982
Table 13 Normalized global

Normalized global weights
weights of all barriers
Barriers Weight ×100
B1 0.4141 41.4062
B2 0.1297 12.9750
B3 0.1443 14.4274
B4 0.0322 3.2186
B5 0.0115 1.1476
B6 0.0787 7.8701
B7 0.0155 1.5500
B8 0.0250 2.5039
B9 0.0608 6.0842
B10 0.0298 2.9788
B11 0.0209 2.0913
B12 0.0085 0.8454
B13 0.0290 2.9016
approximate weight of 41% has the highest power among all the barriers. This is
the most influencing barrier among all the barriers which resist the adoption of
solar PV in Himachal Pradesh. On the other hand barrier, B12 has approximately
0.84% weight which is the lowest among all the barriers. Barrier B3 and B2 are
also influencing barriers with a weight of 14 and 12% approximately. Simulation of
most influencing barrier with new parameters is discussed in below to overcome this
barrier. It is expected that this simulation with new parameters can help to improve
the performance of Solar PV which will help to overcome this barrier.
In this study Technical, economic and environment factors were categorized into
13 barriers. These barriers were then studied using the AHP-Fuzzy methodology to
find the most critical barrier, which was affecting the implementation of solar panels
in the Himachal Pradesh region. Temperature (B1) is the most critical barrier and
Irradiance (B3) is the second most critical factor. With any change in the value of
temperature the output power production of the solar panel gets affected such that
its implementation in higher latitudes becomes uncertain due to a high variation in
temperature and more cloud cover. Irradiance also affects the output power, as the
value of currents varies linearly with any change in the amount of energy absorbed
by the solar panel.
In the present study, 13 critical barriers were found from the data collected from Solar
PV industry experts. These 13 barriers were categorized into 3 criteria. Further, hybrid
Fuzzy-AHP methodology is adopted for the assessment of barriers that influence
the adoption of Solar PV in Himachal Pradesh. Ranking of barriers was got by
normalizing the global weights of barriers. It is found that Barrier B1 (Effect of
temperature) is the major barrier among all barriers which influence the adoption of
Solar PV in Himachal Pradesh. It is also found that B12 (Solar Plant Infrastructure)
barrier is a very less influencing barrier among all barriers. The effect of temperature
should be taken care of before the implementation of Solar PV in the city areas of
Himachal Pradesh. This study was carried out in the city areas of Himachal Pradesh.
Further, this study can be extended by the identification of barriers in both rural areas
and urban areas of Himachal Pradesh using hybrid techniques.
References
1. Jamwal A, Aggarwal A, Gupta S, Sharma P (2019) A study on the barriers to lean manufacturing
implementation for small-scale industries in Himachal region (India). Int J Intell Enterp 6(2–
4):393–407
2. Sahu BK (2015) A study on global solar PV energy developments and policies with special focus
on the top ten solar PV power producing countries. Renew Sustain Energy Rev 43:621–634
3. Jamwal A, Agrawal R, Gupta S, Dangayach GS, Sharma M, Sohag MAZ (2020) Modelling
of sustainable manufacturing barriers in pharmaceutical industries of Himachal Pradesh: an
ISM-Fuzzy approach. In: Proceedings of International Conference in Mechanical and Energy
Technology, pp 157–167. Springer, Singapore
4. Zahedi A (2006) Solar photovoltaic (PV) energy; latest developments in the building integrated
and hybrid PV systems. Renew Energy 31(5):711–718
5. Katiraei F, Aguero JR (2011) Solar PV integration challenges. IEEE Power Energ Mag 9(3):62–
71
6. Tyagi VV, Rahim NA, Rahim NA, Jeyraj A, Selvaraj L (2013) Progress in solar PV technology:
research and achievement. Renew Sustain Energy Rev 20:443–461
7. Nema P, Nema RK, Rangnekar S (2009) A current and future state of art development of hybrid
energy system using wind and PV-solar: a review. Renew Sustain Energy Rev 13(8):2096–2103
8. Denholm P, Margolis RM (2007) Evaluating the limits of solar photovoltaics (PV) in elec-
tric power systems utilizing energy storage and other enabling technologies. Energy Policy
35(9):4424–4433
9. Wang G, Ciobotaru M, Agelidis VG (2014) Power smoothing of large solar PV plant using
hybrid energy storage. IEEE Trans Sustain Energy 5(3):834–842
10. Nugent D, Sovacool BK (2014) Assessing the lifecycle greenhouse gas emissions from solar
PV and wind energy: a critical meta-survey. Energy Policy 65:229–244
11. Tsoutsos T, Frantzeskaki N, Gekas V (2005) Environmental impacts from the solar energy
technologies. Energy policy 33(3):289–296
12. Asif M, Muneer T (2007) Energy supply, its demand and security issues for developed and
emerging economies. Renew Sustain Energy Rev 11(7):1388–1413
13. Rehman S, Bader MA, Al-Moallem SA (2007) Cost of solar energy generated using PV panels.
Renew Sustain Energy Rev 11(8):1843–1857
14. Koppelaar RHEM (2017) Solar-PV energy payback and net energy: meta-assessment of study
quality, reproducibility, and results harmonization. Renew Sustain Energy Rev 72:1241–1255
15. Sharma NK, Tiwari PK, Sood YR (2012) Solar energy in India: strategies, policies, perspectives
and future potential. Renew Sustain Energy Rev 16(1):933–941
16. Bhutto AW, Bazmi AA, Zahedi G (2012) Greener energy: issues and challenges for Pakistan—
solar energy prospective. Renew Sustain Energy Rev 16(5):2762–2780
17. Liu LQ, Wang ZX, Zhang HQ, Xue YC (2010) Solar energy development in China—a review.
Renew Sustain Energy Rev 14(1):301–311
18. McEachern M, Hanson S (2008) Socio-geographic perception in the diffusion of innovation:
solar energy technology in Sri Lanka. Energy Policy 36(7):2578–2590
19. Bhandari R, Stadler I (2011) Electrification using solar photovoltaic systems in Nepal. Appl
Energy 88(2):458–465
20. Sohag MAZ, Kumari P, Agrawal R, Gupta S, Jamwal A (2020) Renewable energy in
Bangladesh: current status and future potentials. In: Proceedings of International Conference
in Mechanical and Energy Technology, pp 353–363. Springer, Singapore
21. Kabir E, Kumar P, Kumar S, Adelodun AA, Kim KH (2018) Solar energy: potential and future
prospects. Renew Sustain Energy Rev 82:894–900
22. Weldekidan H, Strezov V, Town G (2018) Review of solar energy for biofuel extraction. Renew
Sustain Energy Rev 88:184–192
23. Kahraman C, Kaya İ, Cebi S (2009) A comparative analysis for multiattribute selection among
renewable energy alternatives using fuzzy axiomatic design and fuzzy analytic hierarchy
process. Energy 34(10):1603–1616
24. Gautam N, Ojha MK, Swain P, Aggarwal A, Jamwal A (2019) Informal investigation of fourth-
party and third-party logistics service providers in terms of Indian context: an AHP approach.
In: Advances in industrial and production engineering, pp 405–413. Springer, Singapore
25. Tasri A, Susilawati A (2014) Selection among renewable energy alternatives based on a fuzzy
analytic hierarchy process in Indonesia. Sustain Energy Technol Assessments 7:34–44
26. Sharm P, Jamwal A, Aggarwal A, Bhardwaj S, Sood R (2018) Major challenges in adoption
of RFID for Indian SME’s. Int Res J Eng Technol 5:2247–2255
27. Singh PL, Sindhwani R, Dua NK, Jamwal A, Aggarwal A, Iqbal A, Gautam N (2019) Eval-
uation of common barriers to the combined lean-green-agile manufacturing system by two-
way assessment method. In: Advances in industrial and production engineering, pp 653–672.
Springer, Singapore
Suspension Modeling and Its Importance
in Automotive Industries
for Implementation of Modern Vehicular
Technologies
Sujan Neroula and Santanu Sharma
Abstract Losing traction by wheels while driving can deteriorate the vehicle effi-
ciency and dynamic stability, and might lead to catastrophic outcomes in some cases.
The traction provided by the tire is a function of the normal reaction force on a wheel
and the Tire Road Frictional Coefficient (TRFC). Knowledge of the normal reaction
force on the wheel can enable the vehicle to abide by the Acceleration Slip Regu-
lation (ASR) while driving and can help in optimizing the performance in braking
and traction control. A suspension being a linkage between the sprung and unsprung
components of a vehicle and plays an important role in-vehicle system and dynamics.
This paper presents a model of suspension taking into account different spring stiff-
ness and damping coefficients. The behavior of the suspension has been analyzed
for different obstacle heights, vehicle speed, and acceleration. The model has been
further assessed on-road using a lightweight Electric Vehicle (EV).
Keywords Suspension modeling · Electric vehicle · Traction control
1 Introduction
A vehicle suspension system works as a linkage between the vehicle wheels and the
chassis. The suspension system isolates the vehicle from the vibration and harshness
arising due to the road surface and also nullifies the vertical acceleration arising due
to road undulations [1]. A suspension must be rigid enough so it can withstand the
static load effectively and on the other hand, it must be soft enough so as to ensure
comfort and isolation from vibrations and harshness [2].
S. Neroula · S. Sharma (B)

Department of Electronics and Communication Engineering, Tezpur University, Tezpur, Assam
784028, India
S. Neroula
534 S. Neroula and S. Sharma
In [3, 4], a two-dimensional nonlinear model of MacPherson suspension is

reported by analysis of the suspension kinematics and the dynamics have been derived
using Euler-Lagrange’s approach [5].
An electromagnetic suspension model has been reported in [6], wherein the opti-
mization has been done by the introduction of constraints in current, voltage, and
airgap.
The suspension of the vehicle can be mounted to the chassis through a load sensor
that can sense the wheel load in real-time. The real-time data of the wheel load along
with the vehicle TRFC can enhance the vehicle ASR by optimizing the deliverable
torque and also can enhance the braking performance [7, 8].
The following sections present the vehicle suspension model using Newtonian
mechanics and Lagrange’s model, followed by the simulation and actual on-road
analysis of the suspension system.
2 Suspension Modeling
2.1 Suspension Modeling by Newtonian Mechanics
Vehicle suspension modeling is crucial for the inclusion of technologies like traction
control, rollover prevention, etc. Figure 1b illustrates the vehicle suspension model
where b is the damping coefficient of the damper and K S and K W are the spring
constants of suspension and wheel, respectively. The relation governing equations
for the spring and the damper can be expressed as [9].
F = K d x, F = b ẋ (1)
From the FBD illustrated in Fig. 1b, the equation for the suspension model can
be derived as
F1 = K S (y − x), F3 = K W (x − r ), F2 = b( ẏ − ẋ) (2)
Fig. 1 Figure showing a vehicle parked on an incline b free body diagram (FBD) of suspension
model using Newtonian mechanics c dynamic off-road suspension model using Lagrange’s model
Suspension Modeling and Its Importance … 535
Here, the damping force is proportional to the velocity of the piston. Further force
F 3 , i.e., the force acting between the wheel and road can be expressed as
MWheel aW = K s (y − x) + b( ẏ − ẋ) − K W (x − r ) (3)
MCar ac = −K s (y − x) − b( ẏ − ẋ) (4)
Here (x − r) is the relative displacement of the wheel and the road surface.
From the above equations, we can estimate the total force acting on the wheel and
the vehicle. Equations 3 and 4 describe the suspension model using Newtonian
mechanics wherein the mass of the tyre, suspension stiffness, and damping have
been considered.
2.2 Suspension Modeling by Lagrange’s Mechanics
Another very effective and lucid approach is an analysis using Lagrange’s equation.
When we neglect the mass of the tyre and its roll moment, a simple model can be
formed by taking into consideration; the suspension damping and stiffness along
with the vehicle pitch. The spring constant (front and rear) of the vehicle suspension
systems has been measured as K S = 35,000 Nm−1 , and damping coefficient has been
determined to be b = 2000 N-sm−1 . Lagrange’s model has been presented in Fig. 1c.
Here the index f and r refer to the front and rear of the vehicle, respectively.
The vehicle moment of inertia I, for the designed vehicle has been measured to be
144.5 kg m2 . The kinetic energy (KE) and the potential energy (PE) for the system
can be formulated as
1 2 1 2
KE = m ẋ + I θ̇ (5)
2 2
1 2 1
PE = K f y f − x − a1 θ + K r (yr − x − a2 θ )2 (6)
2 2
The energy (Q) dissipated in damper can be expressed by Rayleigh’s function as
1 2 1 2
Q damper = b f ẏ f − ẋ + a1 θ̇ + br ẏr − ẋ − a2 θ̇ (7)
2 2
Further implementing the Eqs. 8 and 9, along with Eqs. 5, 6, and 7; the equation
describing the motion for the suspension system can be obtained as (Eq. 10). (Here
P = KE − PE) [10].

d ∂P ∂P ∂ Q damper
− = (8)
dt ∂ ẋ ∂ ẋ ∂ ẋ

d ∂P ∂P ∂ Q damper
− = (9)
dt ∂ θ̇ ∂ θ̇ ∂ θ̇

m 0 ẍ(t) b f + br a2 br − a1 b f ẋ(t)
+
0 I θ̈(t) a2 br − a1 b f a22 br + a12 b f θ̇ (t)

K f + K r K r a2 − K f a1 x(t)
+
K r a2 − K f a1 a12 K f + a22 K r θ (t)

Kf Kr yf bf br ẏ f
= + (10)
−K f a1 K r a2 yr −b f a1 br a2 ẏr
The parameters that are considered for the simulation closely resemble the values
obtained for the designed vehicle and data obtained from the manufacturers of various
components that have been used in designing the vehicle. K f = K r = 35,000 Nm−1 ,
a1 = 0.9 m, a2 = 0.8 m, vehicle gross mass, m = 270 kg (including the driver), I
= 145 kg m2 , and wheelbase of the vehicle is 1.7 m. The height of the CG for the
vehicle designed has been measured to be 0.32 m.
MATLAB simulation is used to extract the solution for Lagrange’s equation
which describes the system’s natural frequency and oscillations. Simulation of
Eq. 10 with vehicle parameters yields the roots of the equation as; −3.6552 ±
16.4810i, −8.3448 ± 7.3886i.
From these roots, the first and second damped natural frequencies for the
suspension system are obtained as

ωd1 = ωn1 1 − ζ12 = 7.3386

ωd2 = ωn2 1 − ζ22 = 16.4810
−ζ1 ωn1 = −8.3448 rad s−1 , −ζ2 ωn2 = −3.6552 (11)
Negative sign that persists in the real part of the root signifies the decaying nature
of the system oscillations. Using Eq. 11, the damping coefficient and damped natural
frequency has been obtained to be
ζ1 = 0.75, ζ2 = 0.22
ωn1 = 11.15 rad s−1 , ωn2 = 16.88 rad s−1
The system transfer function for the two natural frequencies and the damping
coefficients has been obtained as
284.9
H1 (s) = and
+ 7.43s + 284.9
s2
124.2
H2 (s) = 2 (12)
s + 16.7s + 124.2
Fig. 2 a Step response for H 1 (s), and H 2 (s) b step response for the superposition of H 1 (s), and
H 2 (s)
The step response for these transfer functions H 1 (s) and H 2 (s) obtained has been
illustrated in Fig. 2a. The actual step response (superposition) for the system has
been illustrated in Fig. 2b.
3 Experimental Design
To illustrate these damping, we have considered an obstacle over which the vehicle
will be passing at a uniform speed. This obstacle acts as a unit step input for the
transfer function illustrated in Eq. 10. The variation in load at each wheel has been
logged in real-time. The comparison of the two results (simulation and on-road) has
been presented in the result section.
The vehicle drives over an obstacle of height 2.4 cm having width 500 cm at a
constant speed of 5 m/s. The real-time force acting on the wheel and suspension has
been plotted. This experiment is followed by a deceleration at a rate of 5 m/s2 and
the force on the front wheel has been recorded in real-time.
Figure 3a illustrates the experimental design which analyzes the responsiveness and
damping of the suspension setup. Figure 3b shows the actual vehicle suspension setup
equipped with a load sensor for the estimation of normal force in dynamic conditions.
The following part presents in brief the mathematics involved to estimate the holding
torque and the techniques employed.
Fig. 3 a The designed obstacle for evaluation of suspension b the actual designed vehicle with
1-load cell, two-wheel, three-suspension, four-drivetrain
Accurate data acquisition from the load cell associated with each wheel is impor-
tant and the results must comply with the simulation carried out in the previous
section (Fig. 2a, b). The geometry of the suspension and swing arm introduces a
deviation in the wheel load sensor output. Initially, the system was designed using
a Kalman filter. As this filter is defined only for the linear systems, it was unable
to filter out the error(s) introduced due to the suspension. The road grade estimated
using this technique can be expressed as
θ = f (Fi )i= FL, FR, RL, RR. (13)
To eliminate the error introduced, an NLO has been adopted thus establishing a
linear relation between the wheel load and sensor output. The actual sensor output,
state estimation using Kalman filter and NLO has been illustrated in Fig. 4a, b.
The use of NLO over the Kalman filter also reduces the computational speed and
increases design flexibility [11]. Using the NLO, the nonlinearities, and the errors in
estimated load have been reduced to a great extent. Figure 4c exhibits the results of
the experiment design discussed in Fig. 3a. Rolling of the wheels over an obstacle of
height 2.4 cm and width 50 cm; at speed of 5 ms−1 exhibits a considerable variation in
the estimated wheel load. The evaluated damping coefficient and natural frequency
from the real-time suspension response resemble the simulation results.
This test was followed by braking test wherein the vehicle was subjected to a
high deceleration force of 5 ms−2 , creating a negative pitch. A subsequent increase
in the front wheel load on the front wheels was registered is shown in Fig. 4d. Also,
a comparison between the simulation result (expected) and the actual load has been
presented.
With the knowledge of the wheel load, the gradient of a road surface can be
determined, and technologies like Hill Start Assist (HSA), Traction control, and
vehicle Eco-routing can be implemented effectively [12, 13].
Fig. 4 a Load sensor linearization to mitigate undesired effect caused by vehicle suspension geom-
etry b the estimated road grade using G sensor, the load cell output without linearization and with
load cell output with adaptive calibration factor c variation in load on front left and front right wheel
when passing over an obstacle of height 2.4 inches and 50 cm wide d variation in wheel load of
front left during deceleration of 5 ms−2
5 Conclusions
This paper presents an approach to suspension modeling in lightweight vehicles and

dynamic estimation of the real-time load on the wheel of a vehicle when vehicle
passes over an obstacle. The novel contribution of the paper is as follows:
1. Lagrange’s model provides a simple approach to suspension modeling and taking
into consideration the spring stiffness, damping coefficient and vehicle param-
eters, the natural frequency and damping constant for a suspension can be
derived.
2. From the on-road test results, it can be affirmed that the damping constant and
natural frequency of the suspension is in close resemblance to the results obtained
in simulation.
Future developments of this research will include the implementation of traction
control and optimization of braking in lightweight electric vehicles based on the load
on each wheel.
Acknowledgements We are immensely grateful to Dr. Chinmayee Hazarika, Assistant professor,

Department of Electronics and Communication Engineering, GIMT, Guwahati, for her valuable
inputs and suggestions in this work.
References
1. Vidya V, Dharmana MM (2017) International conference on circuit, power and computing

technologies, ICCPCT
2. Turnip A, Setiawan IR, Amri MF, Tamba TA (2015) International conference on technology,
informatics, management, engineering and environment, TIME-E
3. Ikonen E (2017) Active suspension control with state estimation using finite Markov chains.
Int J Adv Mechatronic Syst 7(3):183–192
4. Ikonen E, Najim K, Garca-Cerezo A (2016) MacPherson suspension system modeling and
control with MDP. In: IEEE symposium series on computational intelligence (SSCI)
5. Rankinen A, Ikonen E, Liedes T (2018) Validation of a nonlinear two-dimensional MacPherson
suspension system model with multibody simulations. In: 2018 14th IEEE/ASME international
conference on mechatronic and embedded systems and applications, MESA, pp 1–5. IEEE
6. Liang D, Zhang K, Jiang Q, Wang Y, Duan J, He H (2019) Mathematical model optimization of
electromagnetic suspension system based on additional constraints. In: 14th IEEE conference
on industrial electronics and applications, ICIEA, pp 2309–2314. IEEE
7. Neroula S, Sharma S (2019) Dynamic estimation of the tire-road frictional coefficient (TRFC)
for acceleration slip regulation in a lightweight electric vehicle. In: TENCON 2019 IEEE region
10 conference, TENCON, pp 2582–2586. IEEE
8. Podder P, Das K, Neroula S, Barman P, Sharma S (2017) Development of an automated setup for
dynamic evaluation of DC motor characteristics. In: 2017 IEEE Calcutta conference, CALCON,
pp 190–195
9. Gillespie TD (1992) Fundamentals of vehicle dynamics. Society of Automotive Engineers,
Warrendale, PA
10. Mohammadzadeh A, Haidar S (2006) Analysis and design of vehicle suspension system using
MATLAB and SIMULINK. American Society for Engineering Education
11. Grip HF, Imsland L, Johansen TA, Kalkkuhl JC, Suissa A (2009) Vehicle sideslip estimation:
design, implementation, and experimental validation. IEEE Control Syst Mag 36–52
12. Das K, Borah CK, Agarwal S, Barman P, Sharma S (2019) Road load model analysis for
eco-routing navigation systems in electric vehicles. In: 2019 IEEE 89th vehicular technology
conference, VTC2019, pp 1–5. IEEE, Spring
13. Sharma S, Pegu R, Barman P (2015) Electronic differential for electric vehicle with single wheel
reference. In: 2015 1st conference on power, dielectric and energy management at NERIST,
ICPDEN, pp 1–5
PIC Controller Based SVPWM
Technique for Induction Motor Variable
Speed Drive
Pabitra Kumar Guchhait, Suvadip Roy, and Abhik Banerjee
Abstract The Space Vector Pulse Width Modulation (SVPWM) technique is getting
more popular than other classical PWM techniques because of its cleaner harmonic
spectrum and higher level of output voltage. In this work, three phases of balanced
AC has been generated with four different frequency levels using an 8-bit microcon-
troller to improve the performance of the system and also to minimize the cost of the
inverter. Voltage/Hz (V /f ) control for induction motor drive is also done by devel-
oping a suitable algorithm for generating the control pulses. Experimental results are
provided to check the effectiveness of SVPWM technique.
Keywords Pulse width modulation · SVPWM · Inverter · Induction motor ·

Variable voltage and variable frequency control
1 Introduction
In the recent trend for the Variable Voltage as well as for the Variable Frequency
controlled induction motor drives, pulse width modulated Voltage Source Inverter
(VSI) techniques are universally used [1]. Among the various PWM techniques,
sinusoidal PWM (SPWM) is widely accepted. In SPWM the maximum output voltage
that can be reached up to 61.2% E dc in linear modulation region. To get maximum
output voltage, we have to keep the modulation index (m = 1) which will cause an
increased amount of harmonics at the output resulting in inferior system performance
[2]. In contrast to the SPWM technique, the space vector PWM (SVPWM) method
P. K. Guchhait (B) · A. Banerjee

Yupia, Arunachal Pradesh 791112, India
A. Banerjee
S. Roy
SMS India Private Limited, Gurugram, Haryana, India
542 P. K. Guchhait et al.
provides superior harmonic quality as well as an extended linear range of operation.

In SVPWM method, we can reach a maximum output voltage up to 91.2% E dc in the
linear modulation region [3]. There has been an upward trend in the use of SVPWM
technique for the induction motor drive because the technique is less trouble to
digital realization and also the better dc bus utilization. There are so many papers
[4–6] where different approach has been taken to implement the SVPWM technique
for induction motor drive using FPGA, DSP and microcontroller. Several literature
[7, 8] also discusses the output results between different PWM strategies that have
been carried out. In this present work, an SVPWM VSI fed V /f controlled induction
motor drive has been designed and an 8-bit microcontroller has been taken into
consideration to develop the inverter which may promise a very low cost-effective
and efficient than the SPWM.
2 Theoretical Background of the Work
Here, a brief explanation of the theory of SVPWM has been given. Space Vector
Modulation (SVM) deals with a special type of switching sequence based on the
three upper switches of the VSI. Theoretically, a sinusoidal voltage is treated as the
reference vector which rotates at a constant angular frequency and the average output
voltage track the reference voltage with the minimum level of the harmonic disorder.
2.1 Space Vector Modulation
The concept of Space Vector (SV) is explained here by deducing its mathematical
model with reference to three 120° phase-shifted sinusoidal signals. At first, three
vectors are considered R(t), Y (t) and B(t) which are 120° apart from each other in
space such that at any instant, the summation of R(t) + Y (t) + B(t) = 0 as shown in
the Fig. 1.
The resultant vector V can be written as
V = V x + j Vy (1)
π
Where, Vx = R − (Y − B)sin (2)
6
π
Vy = (Y − B)cos (3)
6
This is equivalent to the following matrix form
PIC Controller Based SVPWM Technique … 543
Fig. 1 Orientation of R, Y, B
and V
⎡ R ⎤
Vx 2 1 − 1
− 1
2 2 ⎣ Y ⎦
= (4)
Vy 3 0 23 − 23
B
where 2/3 in Eq. (4) represents the scaling factor.

The resultant vector V often known as the SV and also represents V = V e jθ .
Where,
Vy
θ = tan−1 (5)
Vx

V = Vx2 + Vy2 (6)
The three-phase sinusoidal balanced voltages R(t), Y (t), B(t) are as follows:
R(t) = Vm sin(ωt) (7)
Y (t) = Vm sin(ωt − 120◦ ) (8)
B(t) = Vm sin(ωt + 120◦ ) (9)
The voltage SV obtained from these equations is given by,
V = Vm e jωt (10)
Fig. 2 Trajectory of the

space vector
The SV rotates with constant speed and constant magnitude of ω rad/sec and V m ,
respectively. Figure 2 represents the locus of SV. The radius of the circular path
depends on the voltage amplitude.
2.2 Space Vector PWM in VSI
The schematic diagram of a typical VSI is shown in Fig. 3. S1, S3, S5 and S2, S4,
S6 are the upper and lower leg switches, respectively. The pairs S1–S4, S3–S6 and
Fig. 3 Three phase VSI
M1 M3 M5
Vdc
M4 M6 M2
Motor
S5–S2 are their complementary legs. So there will be eight possible switching states,
two of them will be zero switching states (V 7, V 8), and the rest six states will be
active switching states (V 1–V 6). The eight different switching states are tabulated
in Table 1. Here the switches can be ON or OFF, meaning 1 or 0. From Table 1,
we can see that magnitude of the inverter line voltage remains constant for different
switching conditions only the vectors are shifted by a phase angle 60° with them. The
SVs of the inverter line voltage are hexagonal in nature and depicted in Fig. 4 where
the zero vectors (V 7, V 8) reside at the origin. The locus of the reference voltage
space vector (V * ) (see Fig. 4) is circular in nature inside the hexagon. Let decides
Table 1 Switching states of the VSI

State Status of the switches Space voltage vector
S1 S3 S5
1 1 0 0 V1 = √2 Vdc ∠30◦
3
2 1 1 0 V2 = √2 Vdc ∠90◦
3
3 0 1 0 V3 = √2 Vdc ∠150◦
3
4 0 1 1 V4 = √2 Vdc ∠210◦
3
5 0 0 1 V5 = √2 Vdc ∠270◦
3
6 1 0 1 V6 = √2 Vdc ∠330◦
3
7 1 1 1 V7 = 0
8 0 0 0 V8 = 0
Fig. 4 Different sectors of

operation
Fig. 5 Realization of RVV
that at any instant, the reference voltage vector (RVV) makes an angle P w.r.t V1.
This RVV can easily be constructed using the corresponding vectors V1, V2, and
zero vectors V7, V8. For high switching frequency, the RVV can be considered to
be fixed at its previous position for one switching period.
Therefore, now the RVV can be formed by taking the average between V1, V2
for the time period of T 1 and T 2, respectively whereas the time period considered
for each zero vectors, i.e. V7, V8 as T 0 /2 such that total sampling time will be T s =
T 1 + T 2 + T0 . From Fig. 5, the following equation can be formed.
V ∗ · Ts = V1 · T1 + V2 · T2 + V0 · T0 (11)
In rectangular co-ordinate Eq. (11) can be expressed as

cos π6 + θ 2 cos π6 2 cos π2
π + T2 √ π
Ts M = T1 √ (12)
sin π6 + θ 3 sin 6 3 sin 2
From this equation, the following equations are derived

π
T1 = Ts M sin −θ (13)
6
T2 = Ts M sin(θ ) (14)
Ts = T1 + T2 + T0 (15)
T0 = Ts − (T1 + T2 ) (16)
Following these equations, the RVV (V * ) is constructed for 90° > θ > 30° as
shown in Fig. 5. The RVV is constructed for the other five sectors in the same way.
To eliminate the even harmonics in the output line voltage spectra, the line voltage
vectors are set for two sampling periods (2T s ) and the satisfied condition will be V z
Fig. 6 Experimental Setup
− V n − V n+1 − V z − V z − V n+1 − V n − V z (where V z will be anyone of the zero

vectors (V 7, V 8) which are used alternatively and the each V z is used for T 0 /2 time).
3 Drives Implementations
3.1 Hardware Descriptions
The Hardware setup is figured in Fig. 6. Here, the dc supply is needed for the VSI
which has been given from a 1230 V; 50 Hz supply using a variac and a diode
bridge rectifier. IGBT (H20R1203) is used as switching devices in three phases
VSI Circuit. The gate pulse signals are generated from PIC18F4520 microcontroller
device. Hybrid IC VLA542-01R is used for driving IGBT module. This amplifies the
gate pulses as well as isolates the control circuit from power circuit. We have used
3-phase/4-pole/415 V/50 Hz/4.7 A/squirrel cage induction motor as load. A voltage
divider circuit is used with the help of a potentiometer to provide 0–5 V dc to the
ADC pin of the microcontroller as the reference voltage.
3.2 Firm Ware Development
To buildup the SVPWM technique for VSI, an algorithm has been developed. The
programming is done on MPLAB C18 compiler, the details of the programming given
in flowchart as in Fig. 7. The program has been developed under the four different
frequency levels (20 Hz, 30 Hz, 40 Hz and 50 Hz, respectively). The frequency level
is changed by changing the ADC input reference voltage (0–5 V). V /f ratio is kept
constant for different frequency levels by varying modulation index (m a ). Here we
consider sampling angle θs = 10°. So the output voltage SV is approximated at every
Start
Defineeight
Different states
S1 to S8
No
Initialize Sx=S6, Sy= S1,

=10,i=1, Sx=Si, Is i >5?
i=1 Yes
Sy= S (i+1)
Increment the
Generate the value of
value of P using
ADC No
Is >60? i=i+1, =10

User input for Yes
variation of P
Choose particular Turn ON S7 state for

frequency f by T0/2 time
varying P
Turn ON S1 state for
Calculate Sampling T1 time
time Ts for particular
value of f Turn ON S2 state for
T2 time
Calculate values of T1, Turn ON S8 state for

T2 and T0 T0/2 time
Turn ON S7 state for

Calculate values of T1,
T0/2 time
T2 and T0
Turn ON Sx state for
T1 time Increment the value of
Turn ON Sy state for Turn ON S8 state for

T2 time T0/2 time
Fig. 7 Flowchart of the program

Fig. 8 State 1, 2, 3 and 4
10° of electrical rotation of reference voltage SV. Sampling time is calculated from
eq. (17).
103
Ts = ms (17)
f ×6× 60
θs
For every 10° of electrical rotation, the value of T 1 , T 2 and T 0 is obtained from
the Eqs. (13)–(16). We have used these equations in our program in MPLAB C18
compiler to calculate the values of T 1 , T 2 and T 0 . To make the output voltage follow
the reference voltage, state-7 (V 7 ) (S1, S3, S5) is activated for T 0 /2 times, state-1
(V 1 ) (S1, S6, S2) activated for T 1 times, state-2 (V 2 ) (S1, S3, S2) is activated for T 2
times and state-8 (V 8 ) (S4, S2, S6) is activated for T 0 /2 times. For next 10°, i.e. θ =
20°, the same procedure should be maintained to calculate T 1 , T 2 and T 0 and follow
the reverse sequence V 8 , V 2, V 1 and V 7 . To generate the output reference voltage for
the first 60° rotation, i.e. for sector-1 the same procedure should be followed three
times. For the remaining sectors, T 1 , T 2 and T 0 are calculated in the same way only
the line voltage vectors used to construct the reference voltage SV are changed and
the switching sequence is the same as discussed for sector-1. The switching states
and gate pulses are shown in the following Figs. 8, 9 and 10. The whole setup is
depicted in Fig. 11 schematically.
The line voltage and current waveforms are shown in Figs. 12 and 13 for 50 Hz
frequency. The output voltage is higher with SVPWM technique compared to Sinu-
soidal PWM (SPWM) technique for the same DC bus voltage. The line voltage nearly
Fig. 9 State 7 and 8
Fig. 10 Gate pulses
138 V for dc bus voltage 196 V whereas 105 V for the same dc bus voltage. The
line current is 0.45 A for SVPWM technique as compared to 0.34 A for SPWM
technique. It is clear from the figures that the harmonic content is low in output
voltage and current using SVPWM technique. The V /f control is easily done here
and the output line current remains almost constant for different frequency levels
after putting constant load on the induction motor. The V /f control curve is given
in Fig. 14. Therefore, it is evident from the results that SVPWM technique is very
much convenient to implement for different frequency levels and for V /f control.
5 Conclusions
In conclusion, the development and implementation of space vector pulse width

modulated VSI is successful. As we have encountered some problems for smooth
50Hz, 1Ф SUPPLY
+5V
ADC PWM1 PWM

PWM2
PWM3 INVERTER
GATE
PIC18F452 DRIVE
PWM1
PWM2
PWM3
VA VB VC
MOTOR
Fig. 11 Schematic diagram of whole set up
Fig. 12 Output line voltage
frequency variation from (0 to 50 Hz) with 8-bit microcontroller discreet frequency

levels have been used instead of a smooth variation. Still the results are good with
respect to other SVPWM industrial drives using DSP or FPGA. The V /f controlled
has been done very easily and preciously. Drives with V /f techniques solve many
applications in the area of compressor, pumps, fan, conveyor etc. and also save
Fig. 13 Output line current
Fig. 14 v/f control of SVPWM
energy to reduce the cost. V /f drive with close loop techniques is used, where the
load changing requires with constant speed.
References
1. Bose BK (2010) Power electronics and motor drives: advances and trends. Elsevier
2. Mohan N, Undeland TM, Robbins WP (2003) Power electronics: converters, applications, and
design. Wiley
3. Buasre A, Subsingha W (2013) Double-sided linear induction motor control using space vector
pulse width modulation technique. Energy Procedia 34:181–188
4. Dutta S, Chandra A, Chowdhury S (2016) Design and development of an 8 bit microcontroller
based space vector PWM inverter fed Volt/Hz induction motor drive. CIEC, Kolkata
5. Loongk LK (May, 2008) Micro-controller based space vector modulation (SVM) signal
generator. Faculty of Electrical Engineering University, Teknologi, Malaysia
6. Ali JA, Hannan MA, Mohamed A (2014) PSO algorithm for three phase induction motor drives
with SVPWM switching and V /f control. In: 2014 IEEE international conference on power and
energy (PECon), pp 250–254
7. Khan H, Miliani E-H, Drissi KEK (2012) Discontinuous random space vector modulation for
electric drives: a digital approach. IEEE Trans Power Electron 27(12)
8. Tian K, Wang J, Wu B (2016) A virtual space vector modulation technique for the reduction
of common-mode voltages in both magnitude and third-order component. IEEE Trans Power
Electron 31(1)
A Neural Network Based Strategic
Placement and Task Assignment
for a Multi-agent System
Mukund Subhash Ghole and Anjan Kumar Ray
Abstract To provide surveillance, relief, and support, a multi-agent system (MAS)

can be used. Therefore, task assignment and strategic placement become an essen-
tial part of the MAS. In this article, a method of attending the demands from the
workspace is proposed in two modes of operations. One is operating MAS from its
base camp through task assignment. Other is training the MAS using self-organizing
map to place at its strategic position and address the demands of attending task.
Distance and task execution overheads are included in the method for a more real-
istic situation. Extensive simulations are presented in 2D and 3D to validate the
proposed work and comparative studies are presented to highlight the effectiveness
of each mode of operation.
Keywords Task assignment · Self-organizing map (SOM) · Surveillance ·

Multi-agent system
1 Introduction
Natural disasters are painful and sudden events that come with little to no warning.
The recent examples of such disasters are the floods that occurred in 13 Indian states
in 2019 [1], the Nepal earthquake in 2015 [2], the Kedarnath flood in 2013 [3], and
the Sikkim earthquake in 2011 [4]. Such disasters result in loss of human lives and
resources. To provide relief and support, multi-agent system (MAS) can be used to
do various tasks simultaneously. Task assignment (TA) becomes an essential part of
MAS in pre and post-disaster events and for surveillance of several areas of interest.
MAS is a system composed of multiple interacting intelligent agents that can
solve tasks that are difficult for an individual agent. A MAS has several applications
M. S. Ghole (B) · A. K. Ray

Department of Electrical and Electronics Engineering, National Institute of Technology Sikkim,
Ravangla, Sikkim 737139, India
A. K. Ray
556 M. S. Ghole and A. K. Ray
such as support for pre/post-disaster events, warehouse management [5], robotic

soccer cup [6], multi-robot patrolling, traffic and transportation management [7].
TA also has applications related to rescue robots [8]. In [9], a crowd intelligence
based parcel delivery task assignment method was presented which includes ride-
sharing-based delivery mechanism. Some applications of self-organizing map (SOM)
include autonomous navigation in next-generation mars rovers [10], an attempt on
developing human–robot interactions to classify and select autonomous behavior of
the system [11]. In [12], a dynamic ambulance dispatching and redeployment system
to assist real-time decision making is developed to solve real-life instances arising in
Montreal. A multi-task allocation to participants while minimizing the total traveling
distance was presented in [13].
In this work, a method of attending tasks is proposed by authors for two modes
of operation. First, a multi-agent system is sent to the tasks directly from the base
camp to attend the tasks by taking overhead for distance traveled and task execution.
In the second mode, the agents are trained to place themselves strategically within
the workspace and then they attend the tasks considering both overheads. In order to
validate the proposed work, simulation studies are presented. A comparative study
between the above two modes is also presented in 2D as well as in 3D simulation
environments.
The paper is arranged as follows: in Sect. 2, the proposed method is presented. The
validation of the proposed work in a simulated environment is presented in Sect. 3.
In Sect. 4, the overall conclusion is made.
2 Methodology
A problem is considered based on the task to agent distance using a multi-agent multi-
task assignment method. Consider an open workspace having a set of tasks, and a
base camp is defined within the workspace, where the agents reside in the beginning.
The agents can be placed at their strategic locations to better serve the tasks coming
from the workspace by training these agents using SOM. The nomenclature used in
the proposed method is tabulated in Table 1.
2.1 Overhead Costs
In order to make this method look more realistic, two overhead costs are added during
auctioning as shown in Eqs. 2 and 3.
• Distance overhead cost
In order to attend a task, the agent has to travel a distance which could be directly
or via completing other task(s), therefore an overhead dia is added, which is zero
when the agent attends a task directly from its initial position.
A Neural Network Based Strategic Placement … 557
Table 1 Nomenclature used

N Total number of agents d j,k Lattice distance between winner and jth
agent
M Total number of demands/tasks hj,k(x) Neighborhood function
t time step σ (t) Width of the neighborhood function
t max Maximum number of time steps σ0 Initial value of width function
x Input position η(t) Learning rate
ω Weights of an agent η0 Initial value of learning rate
m Dimension of the search space τ1 Time constant of width function
k(x) Index of the winner agent τ2 Time constant of learning rate
A Set of all agents T Set of all tasks
Tr Set of all remaining tasks d ijr Distance overhead for Ai from task T jr
dia Distance overhead for assigned tasks d min Closest task to agent distance
for ith agent
Mr Number of remaining tasks δ to(i) Task completion overhead of ith agent
γ Task completion overhead value for
A(ir,assigned) to T (jr,assigned)
• Task completion overhead cost

Every agent that attends a task, requires some amount of work to be done to
complete that task, the more tasks an agent does, the task the completion overhead
costs of that agent increases.
2.2 Multi-agent Task Assignment
• The following relations are considered for modeling the task assignment:
Ai ∈ A, ∀i = 1, 2, . . . , N , T j ∈ T, ∀ j, T jr ∈ Tr , ∀ jr (1)
where, j is the ID of M number of tasks, jr is the ID of M r number of unassigned

tasks. Initially M r = M, T jr = T j , dia= 0 and δ to(i) = 0.
• Calculate the task to agent distance as

di jr = Air , T jr + dia + δto(i) , ∀Air , ∀T jr (2)
• Find the closest task for each agent

dmin = argmin(di jr ) ∀Air , ∀ jr (3)

ir, jr
then assign that task T jr to the agent Air , which will henceforth be referred as
T (jr,assigned) and A(ir,assigned) , respectively. Exclude that task to form the T jr ∈ T r .
• Update the remaining tasks and the position of agents to the already assigned task
as follows:
Mr = Mr − 1
A(ir,assigned) ⇐ T( jr,assigned)

dia = A(ir,assigned) , T( jr,assigned) + dia
δto(A(ir,assigned) ) = δto(A(ir,assigned) ) + γ (4)
• Repeat the Eqs. 2 to 4, till all the tasks are assigned.
2.3 Strategic Placement
Several applications of TA like area surveillance or quick response to the demand

for task execution, during emergency, require MAS to be placed strategically. This
provides advantages such as swiftness in response and less distance to travel. There-
fore, the SOM paradigm [14] is utilized here to train the MAS. Integration of overhead
costs emphasizes the simulation’s usefulness for real scenarios.
2.4 Lattice Structure of Agents and Their Weights
Every agent acts as a neuron in the SOM, and these agents are placed in a single-
dimensional lattice. For each demand from the workspace, the agents move and place
themselves in the vicinity of the demand. Every agent has a weight associated with
it, which represents the co-ordinates in the workspace. Consider an m dimensional
workspace, an agent i’s weight can be represented as:
ωi = [ωi1 , ωi2 , . . . , ωim ] for i = 1, 2, . . . , N (5)
Each agent’s initial weights are selected from the base camp.
2.5 Generation of Input Points
Simulated Environment: In a workspace of m-dimensions, an input is randomly

generated in the workspace and it is defined as:
x(t) = [x1 , x2 , . . . , xm ] (6)
where t is the sample. In this paper, for 2D and 3D workspaces, m is taken as 2 and
3, respectively.
2.6 Training of Agents Using SOM
Euclidean distance between each agent and input in a time step is calculated. The
winner agent is decided based on the following equation [14]:

k(x) = index minx(t) − ωi (t) for i = 1, 2, . . . , N (7)
i
The winner agent influences its neighboring agents to move towards the input
point. A Gaussian function is used as a neighborhood function for each step which
is given as
d2
− 2σ j,k
h j,k(x) (t) = e 2 (t)
, for t = 0, 1, . . . tmax (8)
As the training proceeds, the winner agent’s influence on its neighboring agents
should gradually decrease after each step, it is given by

− τt
σ (t) = σ0 e 1 (9)
Now, the weights of agents are updated in each step using the following equation
ωi (t) = η(t)h j,k(x) (t)(x(t) − ωi (t)) ∀i = 1, 2, . . . , N (10)
In Eq. 10, η(t) is the learning rate parameter and is given by

− τt
η(t) = η0 e 2 (11)
The weight of agent i is updated using the following equation:
ωi (t + 1) = ωi (t) + ωi (t) for i = 1, 2, . . . , N (12)

3 Results
The proposed method is tested for 2D and 3D workspaces and the effects of the
task assignment and strategic placement of a multi-agent system are presented and
compared among themselves. In both cases, γ = 20. For the training with SOM, the
parameters are t max = 1000, σ 0 = 5, η0 = 5, τ 1 = 5, τ 2 = 40.
• In TA mode only, agents are attending the demands which originate from the
whole workspace with the help of the proposed TA method. This demonstrated
in Figs. 1 and 3 for cases I.a and II.a, respectively.
• In SP and TA mode, the agents are strategically placed in the workspace and then
any demand that comes can be attended. A two-step process is followed. In order
to place the agents strategically, they are first trained using the data taken from
the whole workspace using SOM. After training, any demand that originates from
the workspace can be attended. This is demonstrated in Figs. 2 and 4 for cases
II.a and II.b, respectively.
Fig. 1 (Case I.a) agents attending tasks coming from the whole workspace; base camp is at a corner
place (TA mode only)
Fig. 2 (Case I.b) agents are placed at their strategic positions and then they attend the tasks coming
from the whole workspace (SP and TA mode)
3.1 Case I
In Case I, the task assignment method is presented in 2D workspace for two scenarios.
The TA mode only is shown in Fig. 1. The SP and TA mode is shown in Fig. 2. In
this case, N = 10 and M = 25 are taken. The results of both scenarios are tabulated
in Table 2.
3.2 Case II
In Case II, the task assignment method is presented in 3D simulation in two scenarios
as mentioned in Case I (Figs. 3 and 4). In this case, N = 4 and M = 40 is taken. The
results of both scenarios are tabulated in Table 3.
4 Conclusion
In this work, SOM based strategic placement, a task assignment method of multi-
agent system including two overhead costs are presented. From the results, it is
Table 2 2D workspace
Agent Case I
TA only SP and TA
(units) (units)
A1 89.6466 57.0805
A2 85.0095 27.5807
A3 74.1462 25.0475
A4 100.0458 25.9360
A5 130.0413 55.5508
A6 108.9334 30.0736
A7 99.7778 51.3964
A8 120.0514 39.2868
A9 65.5974 40.2478
A10 117.4837 36.8788
Total distance 990.7332 389.0793
Fig. 3 (Case II.a) agents attending tasks coming from the whole workspace; base camp is at a
corner place (TA mode only)
observed that the agents which are already at their strategic locations respond quickly
against demands. Incorporation of overhead costs exhibits its effect on realistic
scenarios of surveillance. This phase of the work is focussed on a task assignment
method for agents and comparison is provided. In the future, the inclusion of path
planning of these agents along with collision avoidance will be considered.
Fig. 4 (Case II.b) agents are placed at their strategic positions and then they attend the tasks coming
from the whole workspace (SP and TA mode)
Table 3 3D workspace
Agent Case II
TA SP and TA
(units) (units)
A1 272.49377 324.6839
A2 314.2017 222.0405
A3 256.5464 244.0955
A4 278.5442 285.3438
Total distance 1121.7862 1076.1639
Acknowledgements This work is supported by the Visvesvaraya Ph.D. Scheme, Digital India
Corporation (formerly known as the Media Lab Asia) for the project entitled “Intelligent Networked
Robotic Systems”.
References
1. https://en.wikipedia.org/wiki/2019 Indian floods. Accessed on 23 Dec 2019

2. Goda K, Kiyota T, Pokhrel RM, Chiaro G, Katagiri T, Sharma K, Wilkinson S (2015) The
2015 Gorkha Nepal earthquake: insights from earthquake damage survey. Front Built Environ,
Front 1(8)
3. https://en.wikipedia.org/wiki/2013 North India floods. Accessed on 23 Dec 2019
4. Prajapati SK, Kumar A, Chopra S, Bansal BK (2013) Intensity map of Mw 6.9 2011 Sikkim–
Nepal border earthquake and its relationships with PGA: distance and magnitude. Nat Hazards
69(3):1781–1801
5. https://azbigmedia.com/business/economic-development/here-are-aizonasbiggest-economic-
development-wins-of-2019. Accessed on 23 Dec 2019
6. Asada M, Stone P, Veloso M, Lee D, Nardi D (2019) RoboCup: A treasure trove of rich diversity
for research issues and interdisciplinary connections (TC spotlight). IEEE Robot Autom Mag
26(3):99–102
7. Burmeister B, Haddadi A, Matylis G (1997) Application of multi-agent systems in traffic and
transportation. IEE Proc Softw 144(1):51–60
8. Amaya S, Mateus A (2019) Tasks allocation for rescue robotics: a replicator dynamics approach.
In: Rutkowski L, Scherer R, Korytkowski M, Pedrycz W, Tadeusiewicz R, Zurada J (eds)
Artificial intelligence and soft computing. ICAISC 2019. Lecture Notes in Computer Science,
vol 11509. Springer
9. Wang F, Wang F, Ma X, Liu J (2019) Demystifying the crowd intelligence in last mile parcel
delivery for smart cities. IEEE Netw 33(2):23–29
10. Dlouhy S, Arneson E (2019) Applications of autonomous navigation in next-generation mars
rovers. In: International Telemetering Conference Proceedings
11. Jitviriya W, Hayashi E (2014) Application of self-organizing map for analysing robotic arm’s
action with consciousness-based architecture module. J Robot Netw Artif Life 1(2):135–139
12. Gendreau M, Laporte G, Semet F (2001) A dynamic model and parallel tabu search heuristic
for real-time ambulance relocation. Parallel Comput 27(12):1641–1653
13. Liu Y, Guo B, Wang Y, Wu W, Yu Z, Zhang D (September, 2016) TaskMe: multi-task allocation
in mobile crowd sensing. In: Proceedings of the 2016 ACM international joint conference on
pervasive and ubiquitous computing, pp 12–16
14. Kohonen T (1990) The self-organizing map. Proc IEEE 78(9):1464–1480
Review and Analysis of Charge-Pump
Phase-Locked Loop
Mriganka Gogoi and P. K. Dutta
Abstract A Charge Pump Phase-Locked Loop (CP-PLL) is one of the very impor-
tant circuits used in the communication system. Its main purpose is to lock the phase
and frequency of two signals one can be named as reference signal while other as
a feedback signal. It comprises a phase-frequency detector (PFD), a charge pump
(CP), a loop filter (LF), and a voltage control oscillator (VCO). In this paper, we
are presenting a review and analysis of various configurations of CP-PLL made by
different researchers. The analyses mainly emphasize on linear and nonlinear second
and third-order CP-PLL.
Keywords Phase-locked loop (PLL) · Charge pump · S-domain · Loop filter ·

Phase noise · Jitter · VCO
1 Introduction
The significant use of phase-locked loop (PLL) is to match an input signal which
is also considered as a reference signal with a feedback signal obtained from the
voltage control oscillator in terms of their frequency and phase. Initially developed
theories of PLL were analog-based [1], in which an analog mixer was used as a
phase difference detector to multiply reference and feedback analog signal. A PLL
can operate in three states, free-running state, capture state, and phase lock state. In
free-running state, the frequency of the reference signal is zero, while in capture state
the reference signal has a particular frequency and that affects the output frequency
of VCO and in phase lock state the difference in reference signal frequency and
that of output signal frequency is zero. Though analog PLL was faster, static error
in that structure was a major concern. As a result further analysis made to digital
M. Gogoi (B) · P. K. Dutta

Science and Technology, Nirjuli, Arunachal Pradesh, India
P. K. Dutta
566 M. Gogoi and P. K. Dutta
Fig. 1 A CP-PLL block diagram
PLL/CP-PLL. Hence one of the main reasons CP-PLL being widely used is that it
provides zero static phase offset theoretically [2].
Figure 1 represents the fundamental blocks of CP-PLL. PFD compares the
phase/frequency of reference (F ref ) and input signals (F in ). A charge pump is asso-
ciated with PFD to convert its digital states to analog signal so that it can be used
for controlling the VCO. Main function of LF is to remove and attenuate unwanted
components like noise and high-frequency signal components obtained from the
phase detector. The output signal of VCO is periodic with frequency ‘F out ’, its is a
linear function of the input control voltage. A frequency divider circuit is used to
divide the clock frequency.
The paper is organized as, Sect. 2 describes a detailed understanding of PFD,
Charge Pump, loop filter, and VCO along with their mathematical representa-
tion. Section 3 focuses on mathematical analysis of second-order CP-PLL, Sect. 4
describes the analysis of third-order modelling of CP-PLL, Sect. 5 gives the
concluding remark.
2 Description of PFD, CP, LF, and VCO
The falling or rising edge of reference (F ref ) or VCO (F in ) signal triggers the PFD
circuit. Its main task is to compare the phase or frequency of two signals and converts
into three discrete states namely −I P (negative state), +I P (positive state), and 0 (zero
state) that is shown in Figs. 2 and 3. A charge pump is associated with PFD to convert
its digital states to analog signal so that it can be used for controlling the VCO. In
1980 Gardner presented two different charge pumps with passive filters [1] as shown
in Fig. 4. Emphasis is given more on the charge pump with pump current ±I P , as the
analysis is simplified and also configuration is practical in many real-life scenarios.
A loop filter may be an active or a passive filter, one of such passive filter is shown
in Figs. 5 and 6, mathematical representation for the same is given in Eqs. (1) and
(2).
Review and Analysis of Charge-Pump Phase-Locked Loop 567
Fig. 2 PFD state diagram
Fig. 3 PFD operation
Fig. 4 Charge pump with passive loop filters a pump voltage ±V P b pump current ±I P [1]
Fig. 5 Low pass RC filter
Fig. 6 Input and output

signals of loop filter
1
H (s) = R + (1)
CS
H(s) is the response of the low pass RC filter in Laplace domain with R resistance
and C capacitance [2].
t
1
VF (t) = Vc (0) + Ri(t) + i(τ )dτ (2)
C
0
t
where VF (t) is the output of loop filter in volt and Vc (t) = Vc (0) + C1 0 i(τ )dτ
is capacitor voltage in time domain representation. Vc (0) is the initial value of the
capacitor.
Output signal of the loop filter VF (t) has control over the frequency of VCO. A
VCO is presented in Fig. 7 [2].
θ̄VCO (t) = WVCO (t) (3)
WVCO (t) = WVCO

fr
+ K VCO VF (t) (4)
where θ̄VCO (t) is the differentiation of θVCO (t) which is phase of VCO output signal,
fr
WVCO denotes VCO free-running angular frequency and K VCO its gain. θVCO (0) is
the initial value of phase of VCO.
Fig. 7 Voltage control oscillator
Differential equations based mathematical model of CP-PLL are achieved based

on the Eqs. (1), (2), (3), and (4) that are shown below [2].
1
V̄c (t) = i(t) (5)
C
θ̄VCO (t) = WVCO

fr
+ K VCO (Ri(t) + Vc (t)) (6)
3 Second-Order CP-PLL and Its Mathematical Analysis
In 1980 Gardner [1] presented a CP-PLL with a three-state phase detector, a charge
pump with passive filters, and a VCO, without a frequency divider circuit. A
continuous-time mathematical understanding of second-order CP-PLL filter was
presented and found that the transfer function was the same as the conventional
second-order PLL and it concluded that for any value of loop gain an analog, second-
order CP-PLL is unconditionally stable. Zero static phase error is an essential require-
ment in PLL which can be achieved using an active filter with large gain. However,
according to Gardener zero amplification is not a necessary condition for the charge
pump to have zero phase error [3].
A new model with a proper approximation of discrete-time state variables was
proposed in 1994 by Van Paemal [4] for CP-PLL of second order. Difference equa-
tions are used for representing discrete-time state variables without linearization
or approximation. The model estimates the transient waveform of the VCO control
voltage exactly and is valid for out of lock situation of PLL. However, the system gain
becomes unstable for higher loop gain. The limit derived for stability is approximately
10% of the frequency associated with PLL input signal.
Armaroli et al. [5] presented behavioural modelling in the time domain for CP-PLL
using the nonlinear elements. They have presented a simulated method by decom-
posing a network into a few number of one-way blocks. Proposed models for PFD
and VCO were new, charge pump model was obtained based on changes between its
input signals. For a second-order PLL they made assumptions for closed-loop natural
frequency, damping ratio, and calculated the cross over frequency. A simulated time
required for the proposed algorithms is much faster than the conventional ones, CPU
time required is 70 s. However, due to the discrete nature of the circuit, the system is
noisier. Acco et al. [6, 7] proposed a discrete linear, and nonlinear model of CP-PLL.
The analysis was made for CP-PLL with a frequency divider circuit. In discrete-
time linear model VCO, the N-divider and filter equations appear to be linear and
trivial while that of PFD characteristic is non-linear. The whole system is represented
in linear form by making an approximation of the relation that exists between the
phase error and the average current input. This model is suitable for locked state
of CP-PLL. The above-described model is event-driven and nonlinear that allows
to represent both unlock and overload state, it has high accuracy and requires less
simulation time. It is also observed that there is a similarity between the discrete
nonlinear model and event-driven model which estimates the state variables. They
have illustrated the second-order model by non-linear piecewise set of independent
difference equations based on small error signals.
Curran et al. [8] focussed on solving the stability issue of CP-PLL based on non-
linear and discrete-time model proposed by Paemal [4]. They emphasized on the
first approximation of the nonlinear system and observed that Gardener’s approach
of linearization that leads to first approximation is invalid, while paemal’s piecewise
approach accurately determines the first approximation and is highly conservative.
Their stability analysis gives a complete solution for the first approximation.
Leonov et al. [9] attempted to discuss and filled some of the spaces described by
the theory of control system, dynamic systems, and engineering concepts of PLL.
They have suggested an effective way to find a unique lock-in frequency that will
eventually help to solve this problem existed in Gardener’s PLL [1].
Homayoun et al. [10] presented a model of time-variant to calculate the small-
signal loop transmission of second-order CP-PLL exactly. They have analyzed both
small and large signal behaviour on the stability of the model. Observation was made
on translational nature of the loop and found that it degrades the system’s stability
due to overlapping of replica transfer functions. It was shown stability of CP-PLL
can be maintained with the loop bandwidth approaching half of the input frquency.
Niansong et al. [11] presented a design of a low jitter PLL for 802.11 a/b/g
applications. Performance of VCO affects phase noise and spur of PLL. To reduce
phase noise they replaced traditional tail current source by tail current source with
source degradation. Spur of the PLL is also affected by the gain of VCO, so a
6-bits digital controlled capacitor array that will maintain lower gain is used. A
comparison between phase noise with and without source degradation is found to be
−70.37 dBc/Hz and −66.6 dBc/Hz at 10 kHz, −113.9 dBc/Hz and −112.4 dBc/Hz
at 1 MHz. They also analyzed phase error occurred due to both static and dynamic
current mismatch nature of charge pump and thus made necessary modifications in it
to reduce the phase error. Further analysis also made to reduce phase error in Phase
frequency detector. Overall phase error reduction in the PLL at 10 kHz and 1 MHz
are −89 dBc/Hz and −118 dBc/Hz. RMS jitter in the same frequency band is found
to be 757 fs.
Zhao et al. [12] proposed a new nonlinear second-order PLL that has nonlinear
elements. It regulates the loop noise bandwidth adaptively through the non-linear
module adaptively. Their results demonstrate the increase of tracking speed and as
well as acquired bandwidth. Its bandwidth is 18.8 kHz which is much greater than the
traditional PLL whose bandwidth is just 4 kHz. Thus it provides a balance between
tracking, speed and accuracy.
4 Third Order CP-PLL and Its Mathematical Analysis
Gardner presented a mathematical model with linearized equations for third-order

CP-PLL [1]. To mitigate the ripple in second-order CP-PLL, an additional filtering is
required which is fulfilled by connecting a capacitor in parallel with RC impedance.
The transfer function (TF) ‘F(s)’ of the filter is given by Eq. (7) and the order of
the overall transfer function of CP-PLL changes to three. A third-order CP-PLL is
shown in Fig. 8, and its transfer function ‘P(S)’ is given by Eq. (8) [4].
Lu et al. [13] performed jitter transfer analysis in z- domain for a high order
giga-bit rate PLL. They have also used the impulse invariance method as used by
Hanumolu et al. to obtain the z-domain equations for 3rd order PLL. Stability and
frequency steps are analyzed using the discrete transfer function for 3rd order PLL
[4].

1+sτ2
Fout (s) K PDF K VCO C+C1
P(S) = = (7)
Fref (s)
K PDF KVCO τ2
s 3 + τ1 s 2 +
1
K PDF K VCO

N (C+C1 )s+ N (C+C1 )τ1
Fig. 8 A CP-PLL with loop filter of second-order [4]

where τ1 = C+C
CC1
1
and τ2 = RC.
The TF of the closed-loop is as given by Eq. (8) [4].
1 + s RC
F(s) =
s(S 2 R R2 CC1 C2 + S(RC(C1 + C2 ) + C2 R2 (C + C1 )) + (C + C1 + C2 ))
(8)
where K PFD PFD’s gain and K VCO is VCO’s gain respectively, N is the number of
times VCO clock is divided by the frequency divider circuit.
Hanumolu et al. [14] presented a third-order CP-PLL with exact mathematical
analysis and showed that z-domain representation of the state-space model in linear
form is similar to the one obtained from the impulse invariant transformation [6].
Nonlinearity linearity effect and current mismatch were assumed to be negligible as
a result the model cannot be used for simulation with modern-day PLL tools.
Xintian et al. [15] designed a new single-ended charge pump of low jitter with
an opamp of 76 dB voltage gain for clock multiplication in Low voltage differen-
tial signalling (LVDS) transmitter [7]. The proposed PLL model comprised of the
designed charge pump circuit, it is a full differential VCO, a PFD with dynamic
logic, and a loop filter of second order with passive nature. The output parameters
show optimal jitter-power consumption product, occupies small area, and large lockin
range when compared with some of the existing PLL designs. Power consumption
is 12 mW, jitter is 7.1 ps and 65 ps peak-peak at 350 MHz and lock-in range is
100–560 MHz.
Ji et al. [16] design a 1 GHz PLL based frequency synthesizer that is used in
IEEE 1394b PHY (physical). For the generation of 1 GHz frequency 4-stage ring
oscillator is employed. The final design is a third-order, type-2 charge pump PLL
with a differential low ripple charge pump and ring oscillator based VCO. 1 GHz
clock is divided into 200, 100, 50, and 25 MHz. Settling time is found to be 4 µs and
rms value of period jitter is 2.1 ps. The main focus of the proposed methodology is
on providing multi-frequency and multi-phase clock signals.
Armaroli et al. also extended its time-domain behavioural modelling of second-
order CP-PLL to third-order [17]. New models for PFD and VCO that were proposed
for second-order model were extended to third-order model. Shaktarin et al. [18]
proposed an algorithm for PLL with a current detector and second-order loop
filter describing the operation of the in-lock indicator. The proposed model of PLL
comprised of FPD (Frequency Phase Detector), current-controlled source (CCS),
filter (F), charge pump (CP) and controlled oscillator (CO). CCS and F together form
the CP. Mathematical analysis is done by representing the model with differential
equations, two types of third order CP-PLL are modelled by a different arrange-
ment of resistors and capacitors in the loop filter. The model will help in analysing
fine structure of signals in frequency and time ranges and solving stability issues of
CP-PLL.
Comparative analysis shows that nonlinear second-order model coupled with
nonlinear elements provides a balance between tracking, speed, and accuracy [12].
Third-order CP-PLL especially proposed in [18] solves stability issues of CP-PLL.
Many methodology are implemented by the research community to reduce measure-

ment parameters like speed, dead zone, jitter, phase noise, current mismatch, power
dissipation and area are also discussed here. Requirement of extensive analysis,
mathematical clarity and optimized design are essential for further development of
efficient CP-PLL design.
5 Conclusion
This paper describes the analysis of CP-PLL that has a PFD, a CP, a passive filter, and
an Oscillators. Description of these fundamental blocks from mathematical perspec-
tive is presented. It gives an idea of various measuring parameters of PLL for its
performance enhancement. In order to improve the efficiency, various examples
of PLL are proposed by various researchers at different times, however, with the
advancement of technology, improvement in the performance parameters is highly
essential. Analysis of linear and nonlinear second and third-order CP-PLL presented
by various researchers are discussed here. Further comparative analysis of different
CP-PLL with the various parameters are elaborately presented.
References
1. Gardner FM (November, 1980) Charge-pump phase-lock loops. IEEE Trans Commun COM-
28(11):1849–1858
2. Kuznetsov N, Yuldashev M, Yuldashev R, Blagov M, Kudryashova E, Kuznetsova O,
Mokaev T (2019) Charge pump phase-locked loop with phase-frequency detector: closed form
mathematical model. ArXiv:1901.01468, eess.SP
3. Garner FM (October, 1982) Phase accuracy of charge pump PLLs. IEEE Trans Commun
COM-30(10):2363
4. Van Paemel M (1994) Analysis of a charge-pump PLL: a new model. IEEE Trans Commun
42(7):2490–2498
5. Armaroli D, Liberali V, Vacchi C (August, 1995) Behavioral analysis of charge pump PLLs.
Proc Midwest Symp Circuits Syst 893–896
6. Acco P, Kennedy MP, Mira C, Morley B, Frigyik B (May, 1999) Behavioral modeling of charge
pump phase locked loops. Proc ISCAS’99 375–378
7. Acco P (2003) Study of the loop a phase lock: hybrid aspects taken into account. PhD thesis,
Toulouse, INSA
8. Curran PF, Bi Ch, Feely O (2013) Dynamics of charge-pump phase-locked loops. Int J Circuit
Theory Appl 41(11):1109–1135
9. Leonov GA, Kuznetsov NV, Yuldashev MV, Yuldashev RV (2015) Hold-in, pull-in, and lock-in
ranges of PLL circuits: rigorous mathematical definitions and limitations of classical theory.
IEEE Trans Circuits Syst–I: Regul Pap 62(10):2454–2464
10. Homayoun A, Razavi B (2016) On the stability of charge-pump phase-locked loops. IEEE
Trans Circuits Syst I Regul Pap 63(6):741–750
11. Niansong M, Yu S, Bo L, Yaohua P, Yumei H, Zhiliang H (2017) A low spur, low jitter 10 GHz
phase-locked loop in 0.13 µm CMOS technology. J Semicond 32(3):035004(1–5)
12. Zhao L, Shi L, Zhu C (2018) New nonlinear second-order phase-locked loop with adaptive
bandwidth regulation. Electronics 7:346
13. Lu J, Grung B, Anderson S, Rokhsaz S (May, 2001) Discrete z-domain analysis of high-order
phase locked loops. Proc ISCAS’01 260–263
14. Hanumolu PK, Brownlee M, Mayaram K, Moon UK (September, 2004) Analysis of charge-
pump phase-locked loops. IEEE Trans Circuits Syst I, Fundam Theory Appl 51(9):1665–1674
15. Shi X, Imfeld K, Tanner S, Ansorge M, Farine PA (2006) A low-jitter and low-power CMOS
PLL for clock multiplication. In: Proceedings of 32nd European solid-state circuits conference,
pp. 174–177
16. Ji JY, Liu HQ, Li Q (2012) A 1 GHz charge pump PLL frequency synthesizer for IEEE 1394b
PHY. J Electron Sci Technol 10(4):319–326
17. Best RE (2007) Phase locked loops: design, simulation, and applications. McGraw-Hill
Professional
18. Shakhtarin BI, Timofeev AA, Sizykh VV (2014) Mathematical model of the phase-locked loop
with a current detector. J Commun Technol Electron 59(10):1061–1068
Image Fusion: Challenges, Performance
Metrics and Future Directions
S. B. G. Tilak Babu, I. Chintesh, V. Satyanarayana, and Durgesh Nandan
Abstract Image fusion is a technique of fusing multiple images for better infor-
mation and a more accurate image compared to source images. The applications of
image fusion in the modern military, multi-focus image integration, pattern recog-
nition, remote sensing, biomedical imaging, etc. In this paper discussed, advantages
and drawbacks of newly arrived existing methods in the transform domain and spatial
domain image fusion, universal acceptable flowchart for image fusion obtained from
literature, different helpful datasets that are accessible to assess extensively image
fusion algorithms, many performance metrics used to measure the performance of
image fusion techniques and finally suggestions are made as per present research
necessities of respectable performance with acceptable effort in image fusion.
Keywords Image fusion · Data fusion · Performance metrics · Image integration ·

Multiscale fusion · Medical image fusion · Multi-focus image fusion
1 Introduction
Image fusion helps to extract more information from a single composite image over
two separate images [1–3]. The image fusion method having many applications like
multi-modality (e.g. visible and infrared) image fusion, multi-focus image integra-
tion, multi-exposure image fusion, remote sensing, biomedical imaging, etc. Proper
S. B. G. Tilak Babu (B) · I. Chintesh · V. Satyanarayana

Department of Electronics and Communication Engineering, Aditya Engineering College,
Surampalem, India
I. Chintesh
V. Satyanarayana
D. Nandan
Accendere Knowledge Management Services Pvt, Ltd, CL Educate Ltd, New Delhi, India
576 S. B. G. Tilak Babu et al.
Image
Preprocessing
1
Fusion Post Processing Fuse image
Process
image
Preprocessing
Image
2
Fig. 1 General image fusion process
design of fusion rule for a specific application like multi-modality is a very important
task in research.
Very few articles are identified in the literature on a fusion rule applied on multiple
image fusion applications. Moreover, the efficient design of a fusion rule that applies
to multiple fusion applications is much needed in the present scenario.
A simple fusion process is shown in Fig. 1, many image fusion methods use often
preprocessing for image registration, very few articles have also existed in literature
which does not require source image registration [4, 5]. Many of the available datasets
consist of registered images so the recently arrived fusion methods not discussing
registering of source images, but it is necessary to mention image registration as it
impacts the final fused image if source images are not of the same size. There are
many kinds of fusion rules proposed to fuse images such as Weight average, choose
maximum, coefficient grouping, etc. Few fusion methods existed in the literature that
uses more than one fusion rule in their algorithm based on applications [6, 7]. From
literature, it is observed that the fusion of images can be efficiently obtained in the
transform domain concerning the spatial domain. The simple three steps in the fusion
process based on transform domain are (1) Applying transform domain on source
images, (2) Fusing images in the transform domain, (3) Applying inverse transform
on fused image to get spatial domain fused image. Usually, in post-processing inverse
transform or final touches are given to get the final fused image.
Further, the paper is arranged as follows, in Sect. 2 various newly proposed image
fusion schemes are discussed. Quality assessment parameters for image fusion are
mentioned in Sect. 2. After the thorough study, the tabulated summary s presented
with its significance and comparison of various fusion methods in Sect. 3. Finally, a
review of the most significant conclusion and future directions are discussed. Articles
in the literature conclusion are given in Sect. 4.
2 Image Fusion: Literature Survey
The image fusion methods can be broadly classified into two groups which are as
follows:
1. Spatial domain fusion
2. Frequency domain fusion.
Image Fusion: Challenges, Performance Metrics … 577
Image Fusion
Spatial Frequency Domain
Statistical Subspace based Transform approach Pyramid description

approach
Fig. 2 Image fusion scheme categorization
Various existing approach categorization is as follows (Fig. 2).

As discussed earlier, image fusion having many applications such as multi-focus
image integration, pattern recognition, remote sensing, biomedical imaging, etc.
Many of the authors proposed to image fusion algorithms to a specific application
but a universal algorithm much needed that apply to all types of applications in image
fusion. Many algorithms have existed in literature with complexity varies from highly
complex transform domain to a simple weighted average of pixels. Jin et al. [8],
proposed a lightweight algorithm for fusing multi-focus images. This scheme works
on a lesser number of images with the help of adaptive pulse-coupled neural networks
local spatial frequency (PCNN-LSF) and Laplacian Pyramid Transform (LPT). Hou
et al. [9] presented visible and infrared image fusion performed in the NSST domain.
The shift-invariant technique NSST grabbed the properties from wavelet, source
images are directly decomposed by NSST into high-frequency and low-frequency
subbands. The fusion of low-frequency sub-band performed based on an improved
edge saliency map, where high-frequency subbands fusion done by the MSF map.
The multi-objective artificial bee colony (MOABC) algorithm with spinning cortical
mode utilized to optimize fusion results. Finally, inverse NSST applied to obtain a
time-domain fused image from the transform domain. The quantitative performance
of the algorithm shows robustness over other existing methods.
Karsanina et al. [10] proposed soil image resolution enhancement by using various
stochastic techniques, initially the soil images are rescaled by using bicubic interpola-
tion then immediately segmented into micro-units. Immediately Direction correlation
filters (DCF) are applied, the next superposition principle is used to fuse. As very few
papers available in the literature on soil image resolution enhancement through image
fusion, the authors mentioned key limitations to find solutions immediately. Lihua-
jian et al. [11] proposed image fusion based on Joint Bilateral filter (JBF), Kirsch
operator, Rolling Guidance Filter (RGF), Weight Average (WA) with global vari-
ance and multiplication of weight maps with detail layers. Initially, kirsch operator
is used to extracting salience map from source images, by inputting salience maps
to RGF, different subbands are obtained. In the iterative process, RGF can extract
edges. The salience maps are optimized by JBF and optimal correction. The perfor-
mance of fusion is good in objective and subjective evaluation on different types of
images, because of two separate fusion rules for base layers fusion and detail layers
fusion. Tested on different images viz., multi-modal images, multispectral images,
multi-focus images.
Huiqian et al. [12] proposed a fusion method based on two different decomposition
techniques name NSST, Moving Frame-based Decomposition Framework (MFDF).
Initially, the source images are applied to MFDF separately, and then it decomposed
into texture/edge coefficients and approximation coefficients. The maximum Selec-
tion Rule (MSR) is used to integrate the texture features, where NSST is utilized in
the integration of the approximation coefficients. The component synthesis process
is used to obtain the final fused image from approximation fused coefficients and
texture fused coefficients. From this proposed method, authors able to get good
visual effects and objective values in the fused image compared to source images.
In [13], Cartoon Texture Decomposition (CTD) technique results in a cartoon image
and texture image from a source image. The combination of proportional calcu-
lating and spatial fusion is called spatial based fusion rule which was used to fuse
cartoon coefficients. Structure information is preserved in cartoon components by
this energy-based fusion rule. Sparse Representation (SR) based fusion rule is used
to integrate texture components. The multi-modality image integration algorithm is
tested on visible - infrared images, Medical images for medical diagnosis, multi-
focus images. One pair or two pairs of MODIS and Landsat images can be fused
by the Huihui et al. proposed algorithm [14]. Based on Deep Convolutional Neural
Networks (DCNN) the proposed algorithm fuses the source images in two stages (1.
Training stage, 2. Prediction stage). Initially, learn the NLM model in between low
spatial resolution image and MODIS images. Usually, in the investigation of rapid
variations in land surfaces, the main problems are poor atmospheric conditions, and
frequent cloud cover is admitted in the proposed method. The authors mentioned that
the proposed algorithm is unable to estimate the lost information in LSR MODIS
images.
Robust Sparse Representation (RSR) and PCNN-based multi-focus image algo-
rithm proposed by Yond yang et al. [15], RSR helps to decompose sources images
into the reconstruction error matrix (residual matrix) and sparse coefficients matrix.
PCNN results initial decision map where morphological opening and closing results
final decision map. Weighted fusion rule is applied to get the final fused image.
Baihong [16], proposed a fusion algorithm with acceptable registration errors and
computational complexity for fusing Hyperspectral and multispectral images. The
fusion process done by depends on the Low Rank constrained Gaussian Mixture
Model (LR—GMM) (Table 1).
Table 1 Significant performance parameters for fused image

Name of metric Formula Value for best performance
RMSE RMSE = Lower (close to zero)

M N
1
MN Ir (i, j) − i f (i, j)2
i=1 j=1
SNR SNR = Higher value

M N
j=1 (Ir (i, j))
2
i=1
10 log10 M N
i=1 j=1 Ir (i, (
j)−I f (i, j)
2
)
PSNR PSNR = Higher value

L2
20 log10 M N
j=1 ( Ir (i, j)−I f (i, j))
2
i=1
CC CC =
2Cr f Higher value (close T to T + 1)
cr +c f
MI MI = Higher value

M N h I I (i, j)
h Ir I f (i, j) log2 r f
h I (i, j)h I (i, j)
r f
i=1 j=1
2σ
SSIM 2μ I I +C Ir I f +c2 Higher value (close to T + 1)
SSIM = r f 1
( μ2 Ir +μ2 I f +C1 )( σ 2 Ir +σ 2 I f +C2 )
QI QI =
4σx y x y

Higher value (close to +1)
σx2 +σ y2 (x)2 +(y)2

σ 1
2 Higher
σ = i − ī h I f (i)
i=0
HE
L Higher
HE = − h i f (i) log2 h i f (i)
i=0
FMI FMI = MI I p I f + MI Im I f Higher

FQI FQI = Higher (close to 1)

λ(w)QI(I p I f /w)+
C(w)
ωεε (I − λ(w)QI(Im I f /w)
3 Summary
In this section, it is the comparative analysis of all existing literature in the field of
image fusion and its possible applications (Table. 2).
3.1 Quality Assessment Techniques
Qualitative analysis and quantitative analysis are the two popular ways to estimate
the performance of fused images [18–21]. In the earlier analysis, the fused image is
Table 2 Significant measures for image fusion are shown
580
Refere nces Preprocessing Methods used Fusion rule Applications Comments

Jin et al. [8] 2018 NA LPT, PCNNLSF Combination of MAX Multi-Focus fusion on Authors claims, with the
and AVERAGE fusion gray and colour images help of a lesser number
rules of images the algorithm
achieves good
performance with
acceptable complexity
Hou et al. [9], 2018 NA NSST, SCM with Improved edge saliency Visible and infrared NSST, MOABC-SCM
MOABC algorithm map, MSF map image fusion based algorithm
performance is evaluated
quantitatively and shows
robustness over existing
methods
Karsani na et al. [10], Bicubic interpolation Bicubic interpolation, Superposition Soil images The key limitation is to
2018 (rescaling purpose) segmentation, DCF, find solutions
averaging immediately in the field
of soil image
enhancement through
image fusion
(continued)
S. B. G. Tilak Babu et al.
Table 2 (continued)
Jian et al. [11], 2018 Kirsch operator (saliency RGF (decomposition WA with global variance Multi-modal images, In the iterative process,
map extraction purpose) purpose), JBF (base layers), multispectral images, RGF can extract edges.
multiplication of weight multi-focus images The performance of
maps with detail layers fusion is good in
objective and subjective
evaluation on different
types of images, because
of two separate fusion
rules for base layers
fusion and detail layers
fusion
Liu et al. [12], 2018 NA MFDF (decomposition MSR—texture features, Medical diagnosis With the help of two
purpose), NSST, average fusion decomposition
component synthesis rule-approximate techniques, authors able
Image Fusion: Challenges, Performance Metrics …
components to get good visual effects

and objective values
infused image compared
to source images
Hongpe ng [13], 2017 NA CTD proportional calculating Visible and infrared Structure information is
and spatial fusion, image fusion, medical preserved in cartoon
SR-based method diagnosis, multi-modal components by
images energy-based fusion rule
(continued)
581
Table 2 (continued)
582

Song et al. [14], 2018 Downsampling of landsat NLM-CNN, SR-CNN SR-based method Spatiotemporal fusion, The main problems are
images landsat images, and poor atmospheric
MODIS images fusion conditions and frequent
cloud cover is admitted
in the proposed method.
The author directed that
the proposed algorithm
unable to estimate the
lost information in LSR
MODIS images
Yang et al. [15], 2018. The authors considered RSR, PCNN, Weighted fusion rule Multi-focus image The algorithm seems
two registered images morphological opening fusion complex and consumes
and closing and authors mentioned
average time around 73 s
Lin et al. [16], 2018 No need for image GMM LR-GMM based fusion Hyperspectral image Image registration is not
registration rule and multispectral image necessary. Acceptable
fusion computational
complexity and image
registration errors
Yongh et al. [17], 2018 NA SPCA, MTF SPCA based fusion rule MS, PAN image fusion With the help of spatial
structures, this method
efficient in the
predication of lost spatial
details
S. B. G. Tilak Babu et al.
compared with the input images and testifies the quality of the fused image using
various statistical parameters in terms of spatial information, geometric patterns,
object size, colour spectral information, etc. The objective evaluation indication is a
quantitative analysis that can overcome the influence of human’s inaccurate vision
judgment and make the indicators mathematically evaluate the effectiveness of image
fusion. Significant measures for image fusion are shown in Table 1.
4 Conclusion
Several methods are reported in the literature to address issues in image fusion,
many of them are on a specific application but a universal algorithm much needed
that apply to all types of applications in image fusion. In this paper, advantages and
drawbacks of image fusion literature, universal flowchart proposed to image fusion
and performance metrics of image fusion are discussed. The universal flow charts
are needed to test with various newly arrived feature extractor techniques, machine
learning/optimization techniques need to add for efficient fusion.
References
1. Ghamisi P, Rasti B, Yokoya N, Wang Q, Hofle B, Bruzzone L, Bovolo F, Chi M, Anders K,

Gloaguen R, Atkinson PM, Benediktsson JA (2019) Multisource and multitemporal data fusion
in remote sensing: a comprehensive review of the state of the art, vol 7. Institute of Electrical
and Electronics Engineers Inc., pp 6–39
2. Ma J, Ma Y, Li C (1 January, 2019) Infrared and visible image fusion methods and applications:
a survey. Inf Fusion 45:153–178
3. Meher B, Agrawal S, Panda R, Abraham A (1 August, 2019) A survey on region based image
fusion methods. Inf Fusion 48:119–132
4. Solanky V, Katiyar SK (2016) Pixel-level image fusion techniques in remote sensing: a review.
Spa Inf Res 24(4):475–483
5. Wang Z, Wang S, Zhu Y, Ma Y (1 December 2016) Review of image fusion based on pulse-
coupled neural network. Arch Comput Methods Eng 23(4):659–671
6. Yang B, Jing ZL, Zhao HT (2010) Review of pixel-level image fusion. J Shanghai Jiaotong
University (Sci) 15(1):6–12
7. Zhou T, Ruan S, Canu S (2019) A review: deep learning for medical image segmentation using
multimodality fusion. Array 3–4:100004
8. Jin X, Hou J, Nie R, Yao S, Zhou D, Jiang Q, He K (1 September 2018) A lightweight scheme
for multi-focus image fusion. Multimed Tools Appl 77(18):23501–23527
9. Hou R, Nie R, Zhou D, Cao J, Liu D (1 October 2019) Infrared and visible images fusion using
visual saliency and optimized spiking cortical model in non-subsampled shearlet transform
domain. Multimed Tools Appl 78(20):28609–28632
10. Karsanina MV, Gerke KM, Skvortsova EB, Ivanov AL, Mallants D (15 March, 2018) Enhancing
image resolution of soils by stochastic multiscale image fusion. Geoderma 314:138–145
11. Jian L, Yang X, Zhou Z, Zhou K, Liu K (1 June 2018) Multi-scale image fusion through rolling
guidance filter. Future Gener Comput Syst 83:310–325
12. Liu X, Mei W, Du H (1 February 2018) Multi-modality medical image fusion based on image
decomposition framework and nonsubsampled shearlet transform. Biomed Signal Process
Control 40:343–350
13. Zhang Y, Chen L, Zhao Z, Jia J (1 February 2016) Multi-focus image fusion based on cartoon-
texture image decomposition. Optik 127(3):1291–1296
14. Song H, Liu Q, Wang G, Hang R, Huang B (1 March 2018) Spatiotemporal satellite image
fusion using deep convolutional neural networks. IEEE J Select Topics Appl Earth Obs Remote
Sensing 11(3):821–829
15. Yang Y, Yang M, Huang S, Ding M, Sun J (2018) Robust sparse representation combined with
adaptive PCNN for multifocus image fusion. IEEE Access 6:20138–20151
16. Lin B, Tao X, Duan Y, Lu J (16 March 2018) Hyperspectral and multispectral image fusion
based on low rank constrained gaussian mixture model. IEEE Access 6:16901–16910
17. Kim Y, Kim M, Choi J, Kim Y (1 December 2017) Image fusion of spectrally nonoverlapping
imagery using SPCA and MTF-based filters. IEEE Geosci Remote Sensing Lett 14(12):2295–
2299
18. Tilak Babu SBG, Satyanarayana V, Srinivasarao C (2016) Shift invarient and Eigen feature
based image fusion. Int J Cybern Inf (IJCI) 5(4)
19. Li H, Qiu H, Yu Z, Li B (1 September 2017) Multifocus image fusion via fixed window
technique of multiscale images and non-local means filtering. Signal Process 138:71–85
20. Tilak Babu SB, Prasad KH, Gandeti J, Kadali DB, Satyanarayana V, Pavani K (2019) Image
fusion using Eigen features and stationary wavelet transform. Int J Innov Technol Explor Eng
8(8)Special Issue 3:38–40
21. Zhang Q, Liu Y, Blum RS, Han J, Tao D (1 March 2018) Sparse representation based multi-
sensor image fusion for multi-focus and multi-modality images: a review. Inf Fusion 40:57–75
Factors Influencing Financing
for Entrepreneurs
Raman Kumar, Rajasve Kaushik, Shivani, Rishi Kant Kumar,

and Prashant Kumar
Abstract Financing is always a constraint for any entrepreneurs, especially for

nascent entrepreneurs. They face various challenges and need to look towards various
factors and in order to gain finance from investors in any country. However, very few
studies have discussed what the factor which influences investors to provide finance
to entrepreneurs in any country. In this paper, a model has been proposed which
shows factors that influence financing for entrepreneurs. The model has been further
verified from the dataset provided by Global Entrepreneurs Monitor (GEM). The
result suggests that education and training, R&D and technology have a significant
impact on financing for the entrepreneur in a country while culture, infrastructure,
and service have a positive impact.
Keywords Financing · Government policies · Culture · Social support · GEM
1 Introduction
Traditionally, entrepreneurial activity was considered as a one-time process that

creates new offerings, or even a completely new business—an demonstration that
R. Kumar (B)
Lalit Narayan Mithila University, Darbhanga, Bihar, India
R. Kaushik
Pearl Consulting Services, Gurugram, India
Shivani
Centre of Rural Development, Indian Institute of Technology, Delhi, India
R. K. Kumar · P. Kumar
Department of Management Studies, Indian Institute of Technology, Delhi, India
P. Kumar
586 R. Kumar et al.
challenges or “creatively destroys” existing offerings and market relations. Yet,

entrepreneurship today is bound to be seen as an instrument established in the culture
of an enterprise than as an occurrence [1]. Therefore, in comparison to driving perfor-
mance independently, entrepreneurship integrates with other organizational variables
to achieve company results [2]. Rather than having a universal impact, entrepreneur-
ship can be assumed as a job in the quest for superior performance in different types
of organizations.
Entrepreneurship researchers are getting progressively interested in how region-
explicit, cultural, economic, political, demographic, and social factors that are outside
the ability to control individual specialists influence business activity [3]. Recently,
such a phenomenon has been identified as external enablers; however, both current
and prospective entrepreneurs are reliant on the ecosystem in which they work
in such a way that ecological changes can affect their behavior [4]. Indeed, few
growing research indicates that business formation is influenced by regional context
or environmental conditions.
It should be noted that access to external or foreign funding is critical for busi-
ness or enterprise. Studies on large-sized enterprises and small and medium-sized
enterprises (SMEs) around the world are readily available for financing and capital
structure. However, we know less about start-up companies’ external financing—
especially those who are in the preliminary or pre-angle stage [4, 5]. Indeed, the
literature often associates startups with angel investors or venture capital funding.
In the field of entrepreneurship, a well-documented argument is that
entrepreneurial practices or strategies are drivers of financial achievement or
economic development. Prior literature has invested a significant amount of effort into
understanding how entrepreneurial development is influenced by designing govern-
ment policies. In fact, institutions dictate entrepreneurial activities, like those that
set the rules of the game [2]. While government strategies shape various degrees of
the institutional system, those arrangements have a noteworthy capacity to impact
business exercises. Innovation encouragement, trade regulations, financing, and taxa-
tion could be included by the government while designing the entrepreneurs’ related
policies [6].
Government communities around the world are giving more consideration towards
research and development subsidization by venture capital programs. In several
knowledge-based high-tech markets, Venture Capitalists (VCs) put resources in the
financing and growth of startups and early-staged enterprises [7]. All such factors
generate a great deal of curiosity amongst researchers of entrepreneurship as to what
an extent these factors can affect the business environment and how to generate a
conducive environment for business development. So, in this paper, we have proposed
a model that supports the context of financing for entrepreneurs.
From this point forward, the flow of this paper is as follows. In Sect. 2, we have
reviewed the literature related to the area of study, followed by the hypothesis of
the study has been presented. Section 3 presents the data collection procedure and
method utilized to analyze the proposed model. In Sect. 4, we have presented the
results of our study, followed by a discussion of the result that has been stated.
Factors Influencing Financing for Entrepreneurs 587
Lastly, in Sect. 5, we have concluded the paper and suggested a few implications of
this research.
2 Literature Review
Financing is the biggest challenge for any entrepreneur especially the nascent one
who does not have access to various other resources that a professional one has. We
have presented different constructs through research available in the literature.
2.1 Education and Training
So, for these entrepreneurs, one way is the develop the expertise and skill through
various sources. In this context, [8] has suggested that for creativity, innovation, and
successful development of business, it is essential to go through education as well
as professional training from time to time. Besides training and skills, motivates
entrepreneurs to become more risk-averse and substantial move towards sustain-
able development [9]. Looking at the education and training in a country, we have
formulated the hypothesis
H1: Education and Training significantly lead to gaining finance for entrepreneurs.
Entrepreneurs who seek outside funding and support need to move towards devel-
oping a successful innovation technically or empirically from research and devel-
opment [5]. Because technical information sharing through a different medium,
techno-commercial articles [10], articles and information selection different review
techniques [11] help them in analyzing their various problems of business. This helps
entrepreneurs in decision-making strategy and moving towards more financial risk
advising path [12]. Besides, R&D helps them to move towards going various AI and
machine learning [13] based technique which helps them in building a better mindset
and predictive ability. Looking at the impact on R&D and technological development
in a country, we have proposed our second hypothesis.
H2: R&D and Technological development in a country impact access to financing
for entrepreneurs.
In this context, various entrepreneurs consider different ideologies towards
accessing training and education through technology and research and development
centers. So, we have formulated our third hypothesis as
H3: R&D and Technological development in a country significantly improves the
propensity of gaining education and training for entrepreneurs.
588 R. Kumar et al.
2.2 Cultural and Social Support
Hult et al. [14] studied the impact of four variables, i.e. organizational learning,
entrepreneurship, market orientation, and innovativeness on the cultural competitive
edge in an organization. They developed ten different models of four variables that
affect the performance of an organization and these models were examined on a
full and segmented sample of 764 organizations. They used a correlation matrix on
all four variables and Structural equation modeling (SEM) to test each of the ten
models. Casson and Della Giusta[15] entrepreneurship play an important role in
building cultural competitiveness in each type of organization but plays a different
role in the performance of small/old organizations and large/old organizations. So,
another hypothesis is as under:
H4: Cultural and social supports significantly impact the financing for
entrepreneurs.
2.3 Government Policies and Support
Dai and Si [16] analyzed the influence of different policies related to startups and busi-
ness, on the entrepreneurial orientation of Chinese private firms which is also majorly
promoted by the government. Also, the author has implemented strategic choice
theory to contemplate if and how entrepreneurial perceptions of positive national-
level government strategies result in a greater level of entrepreneurial orientation. It
was noted that a positive relationship exists between the entrepreneurs’ perception of
the adequacy of new policies and the commitment of their company in entrepreneurial
activities. This supports the voluntarist view of the relationship between organiza-
tions and the environment. Therefore, they concluded the integration of institutional
theory and strategic choice theory is helpful to understand the impacts of recently
planned national policies or strategies on the entrepreneurial orientation of private
firms.
Cumming [6] studied the comparison of Australian Innovation Investment Fund
(IIF) Program visibility and progression with government venture capital programs
in the US, UK and Canada Also, they analyzed its performance along various dimen-
sions such as the degree of risk undertaken by putting resources into the underlying
stage and cutting edge ventures; the probability to assess and increase the value of
investees through arranging syndication, and portfolio size per support chief and the
leave achievement [17]. We have presented our next hypothesis as under:
H5: Government policies and support significantly push entrepreneurs towards
gaining better financial assistance.
2.4 Infrastructure and Service Support
Bennett [18] studied the impact of infrastructure investments on entry rates and
exit rates of establishments. Also, they ascertained the impact on job creation if
new establishments enter and job destruction if establishments exit. They developed
a theoretical model that depicts the External Enabler/Disabler (EED) Model with
external factors such as technological change, policy change and institutional change.
Research suggests that private infrastructure investment resulted in the creation of
new businesses and increased job opportunities [19]. On the contrary, public infras-
tructure investment resulted in the destruction of business and permanent loss of jobs
[20].
Nofsinger and Wang [3] analyzed how product type, production technology and
experience of entrepreneurs affect the external investment made by the institutional
investors and the informal investors along with the problems of information asym-
metry and moral hazard faced by them. They used the survey data from the Global
Entrepreneurship Monitor from 27 countries. They used Logit regression on the pres-
ence of institutional investors and informal investors along with the external financing
ratio method. They concluded that problems of information asymmetry and moral
hazard can be reduced by product type and entrepreneurial experience. Also, it was
observed that the legal environment is also important for access to external financing.
Institutional investors prefer to invest in businesses of experienced entrepreneurs. So,
our last hypothesis is as follows:
H6: Infrastructure and various communication services significantly impact the
financing for entrepreneurs.
So finally, collecting all the hypothesis, we have presented or model which is
shown in Fig. 1.
Fig. 1 Proposed model for

hypothesis formulation
590 R. Kumar et al.
3 Data Collection and Methodology
For this study, we have used secondary data which is made available by Global
Entrepreneurship Monitor (GEM). These data are the assessment of different
nationals’ entrepreneur’s activity which comprises data related to various themes
from all around the globe. These data are publicly available after 3 years on their
website which could be accessed through. The main indicator of this data in total
early-stage entrepreneurs’ activity (TEA). The GEM data is collected using two large
surveys in which one is from the population from each individual and one is taken
from experts. For our research, the data from the expert survey of the year 2015
which is the latest dataset available to us. In this dataset, we have extracted questions
related to our hypothesis. A total of 34 questions from this dataset fall under our
research category.
Once the database is created, we move towards validating our model using these
data. We have selected Structural Equation Modeling (SEM) for analyzing our model.
“Structural equation modeling (SEM) is an extremely flexible linear-in-parameters
multivariate statistical modeling technique” [21]. This model is broadly used for
multivariate statistical analysis techniques which basically uses regression and factor
analysis in order to analyze the structural relationship between the constructs [22].
As we have a different set of questions and the set which is made available by the
secondary method, so data contains many outliers and missing segments. SEM has
features of manipulating those outliers and analyze the data with better accuracy
[23].
Now, for analysis of the data through SEM, we have chosen the IBM AMOS tool
[24] which a statistical software used to analyze SEM, path analysis, etc. through
a different method like maximum likelihood, scale-free least-squares, etc. [25]. For
our analysis, we have used the maximum likelihood to analyze our data.
4 Result and Discussions
The dataset consists of 34 elements from 2548 respondents. Once the model has been
prepared and run on the IBM AMOS tool, it showed a chi-square value of 7142.458.
The degree of freedom computed is presented in Table 1. Distinct sample moments
from table shows show the portion of the output which is used to show the internal
Table 1 Computation of
Number of distinct sample moments 629
degrees of freedom
Number of distinct parameters to be estimated 112
Degrees of freedom (629 − 112) 517
Chi-square 7142.458
Probability level 0.000
function of AMOS in order to analyze the desired output [24]. Another factor is the
degree of freedom which gives “each of a number of independently variable factors
affecting the range of states in which a system may exist, any of the directions in
which independent motion can occur”. Lastly, the value of chi-square which shows
the goodness of fit test of the model.
Tables 2 and 3 show the covariance and correlation between the constructs. Here
the covariance and correlation between financing for entrepreneurs and education and
training are 0.508 and 0.365 respectfully with standard error (S.E) of 0.038 which
shows the positive covariance with minimum error. This shows that the country
having better education and training for entrepreneurs shows a significant chance of
gaining better financial availability from investors as, education and training provide
skillset to build their business better. Now, for the second relationship, covariance
and correlation between financing and R&D and technology are 0.760 and 0.437
respectfully. This shows the strong relationship between financing and R&D and
technology in comparison to financing and education and training.
For our third hypothesis, education and training and research and technology
have the highest relationship with covariance and correlation of 0.915 and 0.522.
This shows a country having good research and development; technological support
constitutes better educations and training for entrepreneurs. Covariance and correla-
tion between financing and cultural and social support have a positive but weak rela-
tionship. Now, the government policies and support have some impact on financing
as this has a covariance of 0.295 and a correlation of 0.267. Lastly, the infrastructure
and service also have a weak relationship with the financing for entrepreneurs.
Table 2 Covariance relationship between the constructs

Factors Estimate S.E. C.R.
Financing ⇔ education and training 0.508 0.038 13.439
Financing ⇔ R&D and technology 0.760 0.056 13.682
Education and training ⇔ R&D and technology 0.915 0.058 15.901
Financing ⇔ cultural and social support 0.270 0.045 5.957
Financing ⇔ government policies and support 0.295 0.033 8.840
Financing ⇔ services and infrastructure 0.254 0.043 5.866
Table 3 Correlation between

Factors Estimate
the constructs
Financing ⇔ Education and training 0.365
Financing ⇔ R&D and technology 0.437
Education and training ⇔ R&D and technology 0.522
Financing ⇔ cultural and social support 0.142
Financing ⇔ government policies and support 0.267
Financing ⇔ services and infrastructure 0.147
592 R. Kumar et al.
So, the country should more focused towards developing research and develop-
ment facilities. The government should lean towards providing training and support
to their better opportunity to entrepreneurs. The results from Tables 2 and 3 validate
our proposed hypothesis and model. Out of six hypotheses, first three hypothesis
shows significant relationship while the last three have an impact but little less than
firs three.
5 Conclusion
In this paper, we have presented the effect of different factors like education and
training, R&D and technology, culture and social support, government policies and
infrastructure, and service on financing. Our finding suggests that R&D and techno-
logical development in a country have a significant impact on financing in compar-
ison to other factors. Other factors are also positively correlated with the financing,
but R&D and Technological innovation have the highest one. At the same time,
entrepreneurs should also focus on gaining better education and skills with more
technological advancement.
References
1. Kang H (2012) Essays on entrepreneurial finance: the role of corporate venture capital and its
performance implications. ProQuest Diss. Theses
2. Lerner J (2005) Venture capital and private equity: a course overview. SSRN Electron J
3. Nofsinger JR, Wang W (2011) Determinants of start-up firm external financing worldwide. J
Bank Financ
4. Lago R, Saurina Salas J, Lopez JA (2007) Determinants of access to external finance: evidence
from Spanish firms. SSRN Electron J
5. Scott JT, Scott TJ (2016) The entrepreneur’s idea and outside finance: theory and evidence
about entrepreneurial roles. Eur Econ Rev 86:118–130
6. Cumming D (2007) Government policy towards entrepreneurial finance: Innovation investment
funds. J Bus Ventur 22(2):193–235
7. Bacq S, Lumpkin GT (2014) Can social entrepreneurship researchers learn from family business
scholarship? A theory-based future research agenda. J Soc Entrep
8. Sarri KK, Bakouros IL, Petridou E (2010) Entrepreneur training for creativity and innovation.
J Eur Ind Train 34(3):270–288
9. Byrne J, Delmar F, Fayolle A, Lamine W (2016) Training corporate entrepreneurs: an action
learning approach. Small Bus Econ 47(2):479–506
10. Chatterjee S, Kumar P, Chatterjee S (2018) A techno-commercial review on grid connected
photovoltaic system. Renew Sustain Energy Rev 81(2):2371–2397
11. Kumar P, Das U, Chatterjee S (2016) A brief study on control structure of grid connected PV
inverter. In: 2016 international conference on energy efficient technologies for sustainability
(ICEETS), 2016, pp 577–582
12. Färnstrand Damsgaard E, Hjertstrand P, Norbäck PJ, Persson L, Vasconcelos H (2017) Why
entrepreneurs choose risky R&D projects—but still not risky enough. Econ J
13. Kumar P, Chatterjee S, Shah D, Saha UK, Chatterjee S (2017) On comparison of tuning method
of FOPID controller for controlling field controlled DC servo motor. Cogent Eng. 4(1)
14. Hult GTM, Snow CC, Kandemir D (2003) The role of entrepreneurship in building cultural
competitiveness in different organizational types. J. Manage
15. Casson M, Della Giusta M (2007) Entrepreneurship and social capital. Int Small Bus J Res
Entrep 25(3):220–244
16. Dai W, Si S (2018) Government policies and firms’ entrepreneurial orientation: Strategic choice
and institutional perspectives. J Bus Res 93:23–36
17. Tassabehji R, Hackney R, Popovič A (2016) Emergent digital era governance: enacting the
role of the ‘institutional entrepreneur’ in transformational change. Gov Inf Q 33(2):223–236
18. Bennett DL (2019) Infrastructure investments and entrepreneurial dynamism in the U.S. J Bus
Ventur 34(5):105907
19. Bian J et al (2016) Mining twitter to assess the public perception of the ‘internet of things.
PLoS One
20. Dourish P, Bell G (2007) The infrastructure of experience and the experience of infrastructure:
meaning and structure in everyday encounters with space. Environ Plan B Plan Des 34(3):414–
430
21. Golob TF (2002) Structural equation modeling. In: Transportation systems planning: methods
and applications
22. Asparouhov T, Muthén B (2009) Exploratory structural equation modeling. Struct Equ Model
A Multidiscip J 16(3):397–438
23. Gefen D, Straub D, Boudreau M-C (2000) Structural equation modeling and regression:
guidelines for research practice. Commun Assoc Inf Syst 4
24. IBM SPSS Amos (2012) IBM SPSS Amos. IBM Softw Bus Anal
25. Byrne BM (2013) Structural equation modeling with AMOS. Psychology Press
Simulation and Analysis of Hybrid
PEMFC and Solar PV Cell Model
for Commercial Load Profile System
Archana Thoudam, Pabitra Kumar Guchhait, and Abhik Banerjee
Abstract The performance and design optimization of a stand-alone PEM fuel cell
(PEMFC) and PV power system to be used for commercial purposes is investigated
in this paper. The main objective of the paper is to design a network of power systems
that consists of renewable energy sources for its environmental impact. This paper
shows the performance analysis of the hybrid PEMFC and solar PV cells under
residential load demand variation, after which simulation is carried out.
Keywords PEMFC · PV cell · Hybrid-power system · Load profile
1 Introduction
The increasing population has necessitated the rise in commercial energy demand
and the need for using the off-grid renewable power sources [1, 2]. Only one energy
source proves insufficient to provide all-day power to the residents due to intermittent
nature. The combination of fuel cell and solar PV is a possible option to meet the
goal amongst other possible combinations [3–5].
Fuel cells are hydrogen-based devices. They convert the chemical energy of the
fuel into electricity. In this case, we use hydrogen and oxygen as fuel. Fuel cells
provide an edge over the conventional batteries in residential and portable applica-
tions. The main advantages of fuel cells are its static nature and noise pollution-free
characteristics along with less atmospheric contamination of hazardous oxides. Fuel
cells can be used for numerous off- and on-grid applications [6, 7].
A. Thoudam (B) · P. K. Guchhait · A. Banerjee

Papum Pare, Arunachal Pradesh, India
P. K. Guchhait
A. Banerjee
596 A. Thoudam et al.
Out of all the varieties of fuel cells available in the market, proton exchange
membrane FCs, also referred to, as polymer electrolyte membrane FCs, are most
preferred due to their favourable characteristics for operation. A PEMFC is made up
of a solid electrolyte which is sandwiched between two conducting electrodes coated
with some catalyst. PEM fuel cell holds great promise for the use of small-scale
energy generation sources and offers high energy density, low corrosion, reliability,
rapid start-up, low operational temperature and no geographical limitations [8]. In
this paper, Simulink model based PEM fuel cell model has been developed.
Photovoltaic (PV) cells are p-n junction devices which use energy from sunlight
to produce power. This power is DC in nature. The main advantage of PV system is
that it does not generate noise during operation and requires very little maintenance
in comparison to other existing renewable energy sources. But we know that solar
irradiation is intermittent in nature and hence, hybridization of the solar PV system
with other renewable or non-renewable sources is necessary to provide continuous
and reliable power [9].
In [10], a hybrid power system (HPS) merging Solar, Fuel cell and Ultracapacitors
are created and modelled with controllers for power flow which can be applied for
stand-alone, residential, micro-grid consumers. In [11], a HPS of PV (Solar) and Fuel
cell is analyzed for applications in house-hold, i.e. micro-generation. The system
comprises of PV modules, PEM type water electrolyzer, batteries and PEMFC. For
a given location (Ankara) in Turkey, the working of the integrated hybrid system is
predicted and thereby exhibited for varying climatic conditions. This paper indicates
the strategy for power control of the PEMFC and solar PV cell-based-hybrid system
for residential load conditions.
2 Hybrid Model of PEMFC and PV Model Under Study
The schematic model of hybrid PEM fuel cell and solar PV cell for better power
quality service is presented in Fig. 1. We have a model of this hybrid power system
in Simulink which is presented in Fig. 2.
Fig. 1 Schematic diagram of hybrid PEM fuel cell and solar PV cell model
Simulation and Analysis of Hybrid PEMFC … 597
Fig. 2 Simulation diagram of PEM fuel cell
2.1 PEMFC Model
Various models of fuel-cells have been presented in the literature [12, 13]. But in
these models, the partial pressures of the fuels, i.e. hydrogen (PH2 ) and oxygen
(PO2 ) are taken as constant values, which are generally variable in nature in different
conditions. So for better accuracy, it has to be modified for better results.
The PEMFC power scheme can be implemented as follows.
The flow of hydrogen (molar) going through the opening is proportional to the
partial pressure of the hydrogen inside the channel which can be depicted as
q H2 K an
= = K H2 (1)
PH2 MH2
Considering various factors like input flow and output flow of hydrogen, and
the rate at which the hydrogen is flowing during the reaction, a relationship can be
established as follows:
d RT in
PH2 = qH2 − qHout2 − qHr 2 . (2)
dt Van
The molar flow of hydrogen used during the reaction according to the electro-
chemical reaction can be obtained as
N0 IFC
qHr 2 = = 2K r IFC (3)
2F
Substituting Eq. (3) in Eq. (2),
d RT in
PH2 = qH2 − qHout2 − 2K r IFC (4)
dt Van
Now after applying Laplace Transformation on both sides of the Eq. (4), yields
the following expression:
1
K H2 in
PH2 = qH2 − 2K r IFC (5)
1 + τH2 s
In accordance with the given expressions, the partial pressures of the input
variable, i.e. oxygen and water can be expressed as
Van
τH2 = (6)
K H2 RT
1/K O2 in
PO2 = qO2 − 2K r IFC (7)
1 + τO2 s
1/K H2 O
PH2 O = (2K r IFC ) (8)
1 + τH2 O s
The polarization curve of the PEMFC can be acquired by adding the three terms,
i.e. Nernst instantaneous voltage (E), Activation overvoltage (V ) and ohmic over-
voltage (V ohmic ). The mathematical expression for obtaining the polarization curve
of PEMFC is as follows:
Vcell = E + Vact + Vohmic (9)
where
Vact = −B ln(C IFC ) (10)

Vohmic = −R int IFC (11)
Nernst Voltage can be defined as in Eq. (12)

RT PH2 PO2
E = N0 E0 + log (12)
2F PH2 O
2.1.1 The Reformer Model with Its Controller
The fuel for the PEMFC power scheme should be pure hydrogen which will be fed
to the system through an external assembly. The supply of pure hydrogen depends
on the power requirement from the system and the reformer generates the same
by undergoing a reforming process of methanol [14]. In [14], the author depicts a
simple model. It is a transfer function of second-order and the model can be depicted
mathematically as follows:
q H2 CV
= (13)
qmeth. τ1 τ2 s + (τ1 + τ2 )s + 1
2
The flow of hydrogen is controlled in accordance with the power requirement. A

feedback stack current is required for the control. The correlation between the two
can be calculated as:
req N0 IFC
q H2 = (14)
2FU
A PI controller is used for control of the rate of flow of methane [14]. The hydrogen
which is utilized for this purpose is extracted from the reformer section. This can be
depicted by the mathematical expression as

k1 N0 IFC
qmeth. = k1 + − q H2
in
(15)
τ3 s 2FU
2.1.2 Power Conditioning Unit (PCU) for PEMFC
The generated power by the PEM fuel cell power scheme is DC in nature, so it has
to be converted to AC power for the use in residential applications as residential
domestic loads are AC in nature. In some papers [15], a DC/DC boost converter is
utilized to transform the DC power into AC power so that the system can gain better
voltage regulation. But to simplify the complexity of the proposed system, only a
DC/AC converter is employed in this paper. We assume that the effect of time constant
associated with the respective assembling units is negligible on the time response
accuracy of the overall system [15]. The system’s AC voltage and reactive power
flow are regulated by the inverter modulation index. The system’s active power flow
can be controlled by the phase angle or power angle. The terminal voltage at load is
assumed to have constant values. The inverter modulation index plays an important
role in keeping the terminal voltage value constant and PI controllers are put into use
for that.
The MI (Modulation Index) can be described mathematically as follows:
k5 + k6 s
m= (Vref − VAC ) (16)
s
The PEM fuel cell power control mechanism can be explained briefly as follows:
mVcell Vs
PAC = sin(δ) (17)
X
(mVcell )2 − mVcell Vs cos(δ)

Q= (18)
X
Considering a lossless inverter system following assumptions can be drawn:
PAC = PDC = Vcell IFC (19)
From Eq. (14), (16) and (18), we get
2FU X req
sin(δ) = q (20)
mVs N0 H2
The phase angle of o/p voltage and the rate of required hydrogen flow can be
described by the following relationship, let’s assume sin(δ) ∼
= δ, which yields
2FU X req
δ= q (21)
mVs N0 H2
The feedback current of PEMFC system can be derived by establishing the rela-
tionship between the modulation index, load current, load phase angle and phase
angle of the AC voltage.
I = m Il cos(θ + δ) (22)
where
Pl
Il = (23)
Vs cos(θ )
2.2 PV Cell Modelling
A solar PV cell is a device based on a p-n junction diode and it generates electricity
when the energy of photons is higher than the energy level of the semiconductor
bandgap. More than one PV cell is connected in a combination of series and parallel
connections for the formation of a PV module. Various models have been discussed
in the papers [16, 17]. Here, a single diode PV model structure has been used and is
given in Fig. 3 and the Simulink model for the solar PV cell are given in Fig. 4.
The current in the output terminal of the PV cell can be written as given by (24)
Fig. 3 Single diode solar PV cell
Fig. 4 Simulink model of solar PV cell


q(V0 + I0 Rs )
I0 = IPV − Ir exp −1 (24)
kTc n
where the symbols represent: q signifies charge of the electron, its value is 1.6 * 10−19
°C, V 0 is the voltage at the output terminal, and I 0 is the current generated in the PV
cell.
The photocurrent I PV is dependent on (1) the solar irradiance and (2) temperature
of the solar PV cell and may be written as
IPV = [k ∗ (Tc − Tr ) + Isc ] ∗ G (25)
The reverse saturation current Ir of the PV cell can be defined as

3 ( T1c − T1r )
q Vg ∗
Tc
Ir = Irs ∗ ∗e
n∗k
(26)
Tr
Here, the simulation for the PEMFC and PV hybrid system under different sources
and load conditions in Simulink (MATLAB) environment are presented. The param-
eters associated with the PEMFC and the PV cell and their considered values are
presented as shown in Table 1.
The considered load variation profiles are given in Fig. 5. The load profiles are
considered as some of peak load demand variations at time 0.4–0.6 s and 1.2–1.4 s.
The PEMFC output voltage according to the fuel cell feedback current is given in
Fig. 6. From the figure, we can see that the maximum electrical power generated from
the PEM fuel cell is 6 kW at the time of peak load demand. The flow of hydrogen and
methane is also demonstrated in Figs. 7 and 8, and it is shown that the flow varies
according to the load demand profiles.
The characteristic I-V graph and P-V graph of the PV cell are also given in Figs. 9
and 10 for an irradiance 1000 W/m2 (Fig. 11).
4 Conclusion
A hybrid model of PEMFC and PV cell has been developed in MATLAB Simulink.
Further, it can be observed from the simulation result that the electrical power demand
by the load is fulfilled with the aid of the proposed hybrid power system model. The
distribution of solar PV power and PEMFC power can also be achieved individually
with the help of MATLAB Simulink. Further, the convergence profile can also be
Table 1 The value of all the

Parameters of solar PV cell Parameters value
parameters of both the cells
Open circuit voltage of the solar PV 64.2
module (V OC )
Short circuit current of the module (I SC ) 5.96 A
Number of cells in series in a module 96
Number of modules in series and parallel 5 and 66
(N S and N P )
Boltzman constant (k) 1.38 * 10−23 J/K
Reference temperature (T r ) 25 °C
Series and shunt resistance of the module 0.18 and 360
(RS and RP )
Parameters of PEMFC Parameters value
Temperature of stack 343 K
Faraday’s constant depicted by F 96,484,600 °C/kmol
No. of cells per stack, N 0 and No 550 and 6
of stacks, N stack
The constant K r = N 0 /(4F) 1.4251 × 10−6 kmol/(s A)
Universal gas constant (R) 8314.47 J/kmol/K
No load voltage (E 0 ) 0.8 V
Fig. 5 Load power profiles
achieved. So, it may be a permissive hybrid power system (HPS) to fulfil the elec-
tricity demand from the consumers in isolated mode or in grid- integrated mode. The
control strategy of the output voltage can also be further improved with the help of
different controllers.
Fig. 6 PEMFC output

voltage
Fig. 7 Hydrogen gas flow

profile
Fig. 8 Methane gas flow

profile
Fig. 9 Current
characteristics of PV cell
Fig. 10 Power
characteristics of PV cell
Fig. 11 Power distribution

of the hybrid PEMFC and
PV cell for the considered
load power
References
1. Ghenai C, Janajreh I (2013) Comparison of resource intensities and operational parameters of

renewable, fossil fuel, and nuclear power systems. Int J Thermal Environ Eng 5(2): 95–104
2. Sen R, Bhattacharyya SC (2014) Off-grid electricity generation with renewable energy
technologies in India: an application of HOMER. Renew Energy 62:388–398
3. Lajnef T, Abid S, Ammous A (2013) Modeling, control, and simulation of a solar hydrogen/fuel
cell hybrid energy system for grid-connected applications. Adv Power Electron
4. Uzunoglu M, Onar OC, Alam MS (2009) Modeling, control and simulation of a PV/FC/UC
based hybrid power generation system for stand-alone applications. Renew Energy 34(3):509–
520
5. Ahmed NA, Miyatake M, Al-Othman AK (2008) Power fluctuations suppression of stand-alone
hybrid generation combining solar photovoltaic/wind turbine and fuel cell systems. Energy
Convers Manage 49(10):2711–2719
6. Khan MJ, Iqbal MT (2005) Dynamic modeling and simulation of a small wind–fuel cell hybrid
energy system. Renew Energy 30(3):421–439
7. Ganguly A, Misra D, Ghosh S (2010) Modeling and analysis of solar photovoltaic-electrolyzer-
fuel cell hybrid power system integrated with a floriculture greenhouse. Energy Build
42(11):2036–2043
8. Corrêa JM, Farret FA, Canha LN, Simoes MG (2004) An electrochemical-based fuel-cell model
suitable for electrical engineering automation approach. IEEE Trans Ind Electron 51(5):1103–
1112
9. Jayalakshmi NS, Gaonkar DN, Nempu PB (2016) Power control of PV/fuel cell/supercapacitor
hybrid system for stand-alone applications. Int J Renew Energy Res (IJRER) 6(2):672–679
10. Lagorse J, Paire D, Miraoui A (2009) Sizing optimization of a stand-alone street lighting system
powered by a hybrid system using fuel cell, PV and battery. Renew Energy 34(3):683–691
11. Özgirgin E, Devrim Y, Albostan A (2015) Modeling and simulation of a hybrid photovoltaic
(PV) module- electrolyzer-PEM fuel cell system for micro-cogeneration applications. Int J
Hydrogen Energy 40(44):15336–15342
12. Karami N, Moubayed N, Outbib R (2014) Energy management for a PEMFC–PV hybrid
system. Energy Convers Manage 82:154–168
13. Khanh LN, Seo JJ, Kim YS, Won DJ (2010) Power-management strategies for a grid-connected
PV-FC hybrid system. IEEE Trans Power Deliv 25(3):1874–1882
14. Hauer KH (2001) Analysis tool for fuel cell vehicle hardware and software (controls) with an
application to fuel economy comparisons of alternative system designs
15. Hatziadoniu CJ, Lobo AA, Pourboghrat F, Daneshdoost M (2002) A simplified dynamic model
of grid-connected fuel-cell generators. IEEE Trans Power Delivery 17(2):467–473
16. Natsheh EM, Albarbar A (2011) Photovoltaic model with MPP tracker for standalone/grid
connected applications. In: IET conference on renewable power generation (RPG)
17. Nguyen XH, Nguyen MP (2015) Mathematical modeling of photovoltaic cell/module/arrays
with tags in matlab/simulink. Environ Syst Res 4(1):24
How Does Perceived Desirability
and Perceived Feasibility Effects
the Entrepreneurial Intention
Rajasve Kaushik, Raman Kumar, Manishankar Datta, Rishi Kant Kumar,

and Prashant Kumar
Abstract The motive of this work is to study the entrepreneurial intention among
students using the Entrepreneurial Potential Model proposed by Krueger and
Brazil. For this analysis, questionnaires were distributed among students of NIT
Arunachal Pradesh. The questionnaire consists of six parts extracting informa-
tion about perceived feasibility, perceived desirability, entrepreneurial intention,
entrepreneurial capability, entrepreneurial ability and personal data. Five-point likert
type scale is used to collect the data. Simple random sampling process is used for
collection of the data. Confirmatory factor analysis is used to examine the reliability
and validity of the Questionnaire. Multiple regression analysis used for data analysis.
Findings show a positive impact of perceived desirability and perceived feasibility
on the entrepreneurial intention of the students. The work suggests some action plans
which can be executed for inspiring and strengthening the intent of graduate students.
Keywords Entrepreneurial intention · Skills · Perceived feasibility · Perceived

desirability · Entrepreneurship
R. Kaushik (B)
Pearl Consulting Services, Gurugram, India
R. Kumar
Department of Economics, G.D. College, LNMU, Darbhanga, India
M. Datta · R. K. Kumar · P. Kumar
Department of Management Studies, Indian Institute of Technology, Delhi, India
R. K. Kumar
P. Kumar
608 R. Kaushik et al.
1 Introduction
Entrepreneurship refers to the activity of an individual for executing a new idea

or endeavoring for a new opportunity. If any person wants to be an entrepreneur,
they should closely observe the problem statements of society and have the problem
identifying capability. This process of understanding what society wants helps to
find the solution to address that issue, and if potential exists, it can be converted into
a business opportunity. Entrepreneurship is a procedure of identifying and grabbing
the opportunity, integrating resources and the team which will implement the plan
[1]. Nurturing entrepreneurship is viewed as a key factor in promoting economic
development and sustainable growth. For achieving sustainable economic growth,
continuous harnessing of entrepreneurial activity is necessary else it will tend towards
saturation. Entrepreneurs are wealth creators; they nurture the new business, which
creates new employment, which results in solving the problems of unemployment
[2].
Entrepreneurs play an important role in the growth of any country. In comparison
to a developed country impact of entrepreneurship is more notified in a developing
country as it constantly enhances the economic situation of a growing economy
[3]. Entrepreneurship is continuously acquiring importance in India in the rouse of
economic liberalization, and globalization. The government of India has embarked
on ‘The Make in India’ initiative to boost up the entrepreneurial ecosystem in India
[4].
The growing young generation in India has a huge potential for embracing the
entrepreneurial ecosystem. The right place for inspiring, injecting the basic concept
of entrepreneurship, and further assessing their intention towards entrepreneurship
is the educational institutions [5]. Private, public and governmental organizations are
implementing various measures to encourage more and more entrepreneurial activity
in India. Educational institutions have started various programs on entrepreneur-
ship. The current Government policies are focused towards institutional structure
for ameliorating entrepreneurial skills, to hone entrepreneurial skills [6]. Voca-
tional education and training have extended the area for more inclusion of youths in
entrepreneurship. Many youths are looking towards different opportunities in control
of renewable energy [7] are facing various difficulties which require various support
in India [8].
The inclusion of entrepreneurship has taken such a deep route in our education
system, that it has been included in the school curriculum now. The inclusion of
entrepreneurship to the school level will help in the development of entrepreneurial
intention, in the youth generation [9]. A push that is given at a younger age adds more
power to any function. If there is a push from Government on Entrepreneurship, the
process to initiate a new business is made simpler, this will bring in more opportunities
and a new entrepreneur will find it more feasible to register a new business [10].
How much entrepreneurial intention will increase in the Product/Service industry
will depend upon the policy of the ruling government in the country.
How Does Perceived Desirability and Perceived Feasibility … 609
The literature on the entrepreneurial ecosystem in India is growing, and very few
studies are being conducted to assess the entrepreneurial intention in the north-eastern
part of India. Researchers widely endorsed that educational institutions must produce
a grilling academic curriculum, which can act as a booster for emerging ventures,
but still, a lot of improvement is required. After this, one of the most important
constructs, which has evolved in the entrepreneurship, literature over the last few
decades’ research is entrepreneurial intention. Intention-based research models tried
to give some good explanations of the amateur thought process of entrepreneurs and
proven to be good compared to models based on individual variables. Looking as
such intent it is necessary to find the impact of perceived desirability and perceived
feasibility on entrepreneurs’ intention.
From this point forward, the flow of this paper is as follows: Sect. 2 presents the
literature review section of this paper in which various constructs of this paper are
described Sect. 3
2 Literature Review
Hockerts [11] mentioned that major studies in the entrepreneurship area, before the
intention model assessment are mainly trait approaches, psychological characteris-
tics, ex-post conditions of entrepreneurs. Another lobby of researchers worked on
demographic factors. They are largely criticized due to conceptual, methodolog-
ical and lack of explanation. According to [12] the evolution of entrepreneurial
intent in a person depends upon his/her perception of entrepreneurial feasibility and
desirability which will later lead to entrepreneurial behaviour. Perceived desirability
depends upon the individual’s attitudes towards the result of any action. In the case of
entrepreneurial action, the outcomes include revenue generation, autonomy, owner-
ship, risk-taking ability, etc. Besides, entrepreneurs have desire to know various tech-
nocommercial [13] as well various new technology usage like Machine Learning [14]
and artificial intelligence which provide them future skill and feasibility to utilize
those technology [15].
Hill et al. [16] proposed Theory of Reasoned Action, which suggests that one’s atti-
tude is responsible for his/her behavioural intention towards that behaviour and along
with that subjective norms (i.e., family influence, the influence of role models, etc.)
are also responsible. It also suggested that individual’s motivation and surrounding
affects the constructs. Ajzen [17] proposed the theory of Planned Behaviour which is
an extension of the Theory of Reasoned Action. He worked with three conceptually
exclusive antecedents of one’s intention which are the attitude for a behaviour, the
degree of perceived level of behavioural control and subjective norms. According to
the author, behaviour is related to the beliefs relevant to the behaviour. It is these accu-
mulated beliefs that are reflected in a person’s attitudes, intentions and behaviour.
Performance of a behaviour is a connected function of intentions and perceived
behavioural control [17]. Subjective norms can be simply explained as to how an
individual perceives the external pressure of society to perform a specific behaviour.
Krueger and Brazeal [12] during his research on entrepreneurial intention

proposed Entrepreneurial Potential Model. The researcher highlighted two important
constructs in the study namely perceived feasibility and perceived desirability. During
the study, they found that in the Theory of Planned Behaviour and the Entrepreneurial
Event model few concepts are overlapping, and hence by integrating they proposed
the EPM model. Findings show that credibility comes from the perception of desir-
ability and with the understanding that this venture opportunity is feasible, which
leads to a behaviour depending on the potential of the person, wants to start the
entrepreneurial venture.
2.1 Entrepreneurial Intention
Entrepreneurship is an intention driven process. Intention of being self-employed

or having a business of own makes a person inclined towards starting a new busi-
ness. Entrepreneurial intent comes from one’s willingness to solve a problem, to
bring change, autonomy and from the present situation (Krueger and Brazeal [12]).
Researchers tried to predict the intensity of the individual’s entrepreneurial action
and to construct an individual’s intention to start a venture is being repeatedly used.
2.2 Perceived Desirability
Perceived desirability alludes to the intensity level to which a person is allured for
performing behaviour to become an entrepreneur. Moreover, previous research states
that cultural factors and social norms affect desirability. If one thinks that people
from their surrounding environment endorse to accomplish the behaviour, this will
augment one’s attitude towards the behaviour. These social burdens are a hindrance
to embark any entrepreneurial venture.
2.3 Perceived Feasibility
It can be defined as the degree at which one considers himself personally capable of
carrying out a behaviour. The presence of guidance, mentors, and role models helps
in establishing one’s perception towards feasibility that there is potential, and the
implementation can be done.
2.4 Entrepreneurial Capability
It refers to ones’ ability to perform a task that comes after the creation of trust on a
concept whose advantage and limitations have been understood properly. Literature
shows that previous researchers use this concept to explain entrepreneurial credibility.
It refers to one’s skill of understanding the entrepreneurial process which helps in
the implementation of a project.
Literature shows that previous researchers use this concept to explain the propen-
sity to act as an entrepreneur. So, in this paper, we have proposed a model in order
the find the objectives: (1) To study the entrepreneurial intention among students
of NIT, Arunachal Pradesh. (2) To assess the effect of perceived desirability and
perceive feasibility on entrepreneurial intention of students. (3) To study the effect
of entrepreneurial ability on entrepreneurial intention of students.
3 Methodology
This section constitutes of two subsections. In first the proposed model has been
discussed and in second data collection section has been described.
3.1 Proposed Model
Researchers like, [11, 17] stated that EPM meshes with two important perti-
nent antecedent models. The wide use of the Entrepreneurial Potential Model in
previous studies supports EPM as the best robust model for the measurement of
entrepreneurial intention. Entrepreneurial Potential Model [12] with little modifi-
cation, is chosen as the base model for the study. The model conceptualizes that
the occurrence of an entrepreneurial venture enquires potential and ability to start
a business that is explained by three important constructs: perceived desirability,
perceived feasibility and propensity to act. Here, the effect of the independent factors
such as perceived feasibility, perceived desirability, entrepreneurial capability and
entrepreneurial ability on the students’ intention to start an entrepreneurial venture
is tested.
Besides, Shapero [18] proposed Entrepreneurial Event Model, which is one
of the earliest contributions in the field of intention-based behavioural study of
entrepreneurs. The researchers irrespective of taking the entrepreneur as the unit
of analysis took an entrepreneurial event for the study purpose. The model depicts
the intention to embark on a new business as being dependent on three factors. (1)
Perceived Desirability—the thoughts and perceptions of desirability—one’s enthu-
siasm and attractiveness towards starting a new venture. (2) The propensity to Act—a
Fig. 1 Proposed research

model
person’s inherent qualities of mind and character strongly reflects in the decision-
making process of the individual. (3) Perceived Feasibility—the state to which one
feels to be personally competent to start a new venture. Based on these we have
proposed our hypothesis as under and the model is presented in Fig. 1.
H1 There is a significant impact of Perceived Desirability on Entrepreneurial
Intention
H2 There is a significant impact of Perceived Feasibility on Entrepreneurial
Intention
H3 There is a significant impact of Entrepreneurial ability on entrepreneurial
intention
H4 There is a significant impact of Entrepreneurial capability on entrepreneurial
intention.
3.2 Data Collection
Entrepreneurial Potential Model is used in the study to assess the intensity of

entrepreneurial intention among the undergraduate students of NIT Arunachal
Pradesh. Desirability and feasibility, which are the significant determinants of
entrepreneurial intention, are taken and to assess the impact on intention. Tested
scales used in the Entrepreneurial Potential Model is used with little modifica-
tion. The whole questionnaire is consistent of 06 (Six) portions, namely, perceived
feasibility, perceived desirability, entrepreneurial capability, entrepreneurial ability,
entrepreneurial intention, and personal data. Closed-ended questions with five-point
likert scales were used. Simple random sampling was adopted. Responses were gath-
ered from third year and final year students as they are on the verge of deciding on
their professional career. The questionnaire was distributed among 120 students of
NIT Arunachal Pradesh, out of which 100 successfully responded. In the beginning,
a pilot study was conducted by distributing the questionnaire to 30 respondents to
check the Cronbach’s alpha.
4 Result
Regression analysis is used to find out the significant impact or influence of indepen-
dent variables on the dependent variable [19]. So, equation of the proposed model is
given in Eq. (1).
EI = α + β1 PD + β2 PF + β3 EC + β4 EA + C
= (1)
where EI—Entrepreneurial intention, PD—Perceived desirability, PF—Perceived

feasibility, EC—entrepreneurial capability, EA—Entrepreneurial ability.
Table 2 represents the ANOVA analysis, which is computed to test whether the
proposed model is statistically significant or not.
H0 : β1 = β2 = β3 = β4 = 0(R square = 0) and

H1 : at least one β = 0 (R square = 0) (2)
Analysis confirms the overall model fit of the study, revealing a significant P-
value, which is 0.000. In the proposed model R Square value is 0.88 (from Table 1).
Interpretation of which means 88% variation in EI is explained by all the four vari-
ables, i.e. PD, PF, EC, and EA. Now the statistical significance of each regression
coefficient of the proposed model is checked (Table 5). Interpretation of the result
of the analysis shows that all the independent variables except EC are statistically
significant. The proposed model is shown below with the value of their respective
coefficient (Table 3):
Table 1 Proposed Model 1 Summary

R R square Adjusted R square Std. error of the estimate
0.940 0.883 0.878 0.370
Table 2 Annova table

Model Sum of squares df Mean square F Sig.
Regression 98.822 4 24.706 180.79 0.000
Residual 13.118 96 0.137
Total 111.941 100
Table 3 Revised model summary

Mode 1 R R Square Adjusted R square Std. error of the estimate
0.939 0.881 0.877 0.371
Table 4 Annova table

Model Sum of squares df Mean square F Sig.
Regression 98.611 3 32.870 239.20 0.000
Residual 13.329 97 0.137
Total 111.941 100
Table 5 Coefficients table for the regression analysis Dependent Variable: EI Predictors:
(Constant), EA, PD, PF
Model Unstandardized coefficients B Standardized coefficient beta t Sig.
(Stand. error)
(Constant) −1.097 (0.277) −4.047 0.000
PF 0.488.087 0.425 2.797 0.006
PD 0.309.081 0.259 4.526 0.000
EA 0.532.139 0.306 3.618 0.000
EI = −1.119 + 0.256 PD + 0.436 PF + 0.134 EC + 0.506 EA

(0.276) (0.091) (0.096) (0.108) (0.140) Standard error (3)
Values in brackets indicate individual standard error. A regression coefficient

is statistically significant if the standard error is less than half of the regression
coefficient. Regression analysis is again computed, and it is observed that the model
is statistically significant shown in Table 4. Further removing EC from the model,
R Square value is decreased by 0.002 (shown in Table 4), which is very negligible
and insignificant. This shows that removing EC has the least effect on the predictive
power of the model. The revised model is shown below (Table 3).
Since the coefficient of EC is found statistically insignificant, it has been removed
from the model and the model has been revised to as follows:
EI = α + β1 PD + β2 PF + β3 EA + C
=. (4)
EI = −1.097 + 0.488 PD + 0.309 PF + 0.532 EA (5)
The interpretation of the value of the coefficient of PF is that keeping PD and EA

constant if PF is changed for 1%, it brings a change in 48.8% in EI. Moreover, the
beta value indicates that PF influences EI most followed by EA and PD. Therefore,
PF is the most influencing factor among the independent factors taken in the study.
5 Discussion and Conclusion
This research used the EPM model of Krueger and Brazeal’s with suitable modifica-
tion, to study the entrepreneurial intention. Four constructs, such as perceived feasi-
bility, perceived desirability, entrepreneurial capability, and entrepreneurial ability,
have been taken as independent factors to study the entrepreneurial intent of the
students. This study shows a significant impact of perceived feasibility, perceived
desirability, and entrepreneurial ability on entrepreneurial intention.
The finding shows that perceived feasibility explains and influences most of the
entrepreneurial intention of the students. Entrepreneurial education, support from
the environment, understanding the potential of revenue generation of a new idea if
implemented impacts the feasibility. To improve the perceived feasibility institute
should give more exposure to the idea generation process, working process of start-
ups, successful technology transfer process so that students can develop an in-depth
understanding of the practicality of any project.
Desirability gets impacted by the student’s own will to become self-independent,
desire to build something of their own, which can generate revenue for them. The real-
ization of less campus placement also gives a boost for desirability towards starting
a new venture. It is also observed that due to less awareness, entrepreneurial intent is
not growing. On the other hand, those who have grasped the depth understanding of
the venture creation, skills to strategically implement, and have the propensity to act,
though understand hindrances, shows strong entrepreneurial ability towards venture
creation.
The entrepreneurship curriculum is creating an impact on the mindset of the
students in NIT, but the intensity of that towards leading to the implementation
behaviour is less. Absence of practical experience in incubators, less risk-taking skill,
less awareness of Government policies, poor engagement with the entrepreneurial
network is compelling them to step back from showing entrepreneurial behaviour.
Hence, it is recommended that the institute should follow up and grill the working
process of the incubation center. Providing exposure to the entrepreneurial networks
of the state and national level will boost the thought process of the students. Institute
should also provide knowledge about existing Centre and state Government policies
endorsing entrepreneurs and process to avail them. The curriculum designed should
be aligned to develop entrepreneurial intention among the students. A similar course
curriculum may be floated in other state control universities so that students will have
a proper understanding of the entrepreneurial process.
References
1. Burns P (2018) New venture creation

2. Neck HM, Meyer GD, Cohen B, Corbett AC (2004) An entrepreneurial system view of new
venture creation. J Small Bus Manage 42(2):190–208
3. Chaston I, Scott GJ (2012) Entrepreneurship and open innovation in an emerging economy.

Manage Decis 50(7):1161–1177
4. Dasannacharya BA (2016) Make in India. Curr Sci
5. Trivedi R (2016) Does university play significant role in shaping entrepreneurial intention? A
cross-country comparative analysis. J Small Bus Enterp Dev 23(3):790–811
6. Gupta VK, Guo C, Canever M, Yim HR, Sraw GK, Liu M (2014) Institutional environment
for entrepreneurship in rapidly emerging major economies: the case of Brazil, China, India,
and Korea. Int Entrep Manage J 10(2):367–384
7. Kumar P, Das U, Chatterjee S (2016) A brief study on control structure of grid connected PV
inverter. In: 2016 International conference on energy efficient technologies for sustainability
(ICEETS), pp 577–582
8. Surie G (2017) Creating the innovation ecosystem for renewable energy via social entrepreneur-
ship: Insights from India. Technol Forecast Soc Change 121:184–195
9. Premand P, Brodmann S, Almeida R, Grun R, Barouni M (2016) Entrepreneurship education
and entry into self-employment among university graduates. World Dev 77:311–327
10. Vilcov N, Dimitrescu M (2015) Management of entrepreneurship education: a challenge for a
performant educational system in Romania. Procedia Soc Behav Sci 203:173–179
11. Hockerts K (2017) Determinants of social entrepreneurial intentions. Entrep Theory Pract
41(1):105–130
12. Krueger NF, Brazeal DV (1994) Entrepreneurial potential and potential entrepreneurs. Entrep
Theory Pract 18(3):91–104
13. Chatterjee S, Kumar P, Chatterjee S (2018) A techno-commercial review on grid connected
photovoltaic system. Renew Sustain Energy Rev 81:2371–2397
14. Kumar P, Chatterjee S, Shah D, Saha UK, Chatterjee S (2017) On comparison of tuning method
of FOPID controller for controlling field controlled DC servo motor. Cogent Eng 4(1)
15. Duan Y, Edwards JS, Dwivedi YK (2019) Artificial intelligence for decision making in the era
of Big Data—evolution, challenges and research agenda. Int J Inf Manage 48:63–71
16. Hill RJ, Fishbein M, Ajzen I (1977) Belief, attitude, intention and behavior: an introduction to
theory and research. Contemp Sociol 6(2):244
17. Ajzen I (1991) The theory of planned behavior. Organ Behav Hum Decis Process 50(2):179–211
18. Shapero A (1975) The displaced, uncomfortable entrepreneur. Psychol Today 9(6):83–88
19. Hair JF, Ringle CM, Sarstedt M (2011) PLS-SEM: indeed a silver bullet. J Mark Theory Pract
19(2):139–152
A New Topology for Photo Voltaic
Energy Conversion System
Sudhir Maurya and Abhik Banerjee
Abstract A new topology for solar photovoltaic (PV) energy conversion system is
used in this paper. Various topologies are available nowadays which have several
problems associated with using several blocks and in its performance analysis of
the system from the stability point of view like DC-DC converter, low-frequency
transformer, maximum power point tracking [MPPT] block, unfold circuit like push-
pull circuit, etc. New types of topology for solar photovoltaic energy conversion
systems have been shown which bypass the MPPT separate block units, unfold circuit,
low-frequency transformer. By the use of current-controlled inverter integration of
MPPT along with the DC-DC to AC conversion takes place inside a single block
which reduces the number of blocks helping in getting a simple transformation block
for the stability analysis. With the help of Phase Locked Loop (PLL) quadrature axis
voltage can set to zero so that only direct axis voltage is used which corresponds
to a single PI controller which again reduces one set point type tracking controller.
Currents which are injected into the grid contains harmonics, THD generated form
inverter is approximately 10% only, which can be further reduced by 5% by using
simple suitable LC filter. MATLAB simulation depicts the whole work.
Keywords Current controlled inverter · Phase-locked loop · Integrated maximum

power point tracking · Set-point controller
1 Introduction
MPPT controller (DC-DC circuit) can be of any type either isolated or non-isolated
type. It ensures the load line attributes with respect to PV terminals so that maximum
power can be drawn from the photovoltaic cell. An old topology that includes various
S. Maurya · A. Banerjee (B)

Department of Electrical Engineering, National Institute of Technology, Yupia, Arunachal
Pradesh, India
S. Maurya
618 S. Maurya and A. Banerjee
Fig. 1 PV module
blocks like separate MPPT block, DC-DC converter (Buck–Boost), Unfold circuit,
push–pull circuit, low-frequency transformer is used. In the new topology, DC-DC
converter and unfolding circuit are replaced by a single power stage block called
current controlled inverter. Low-frequency transformer which is present at the right
side is shifted to the high-frequency side, i.e. into the dc-dc converter. Low frequency
transformer is bulky to use in this circuit. Isolation of transformer from the circuit
is necessary from any type of discharging of currents which leads to shock, for
this flyback converter topology is used which is an isolated version of buck-boost
topology. DC-DC converter is eliminated by the use of current-controlled inverter.
Inverter is doing current control with the help of inductor current which is actu-
ally being pumped into the grid is being controlled, then the MPPT algorithm can
be used for the terminal currents of the photovoltaic panel. The stability analysis
of the terminal voltage can be seen in [1–3], similarly the terminal voltage of the
photovoltaic can be measured and the power that is obtained can be pass through the
maximum power point tracking algorithm. The output of the reference current for the
inverter will be appropriately modulated in such a way that the desired current flows
through it and maximum power will be achieve from the photovoltaic panel [4]. In
[5, 6], the simulink model of hybrid system can be observed and also in this way
this topology works, where is only one power stage which is coming between the
photovoltaic module and the grid (Figs. 1 and 2).
2 Three-Phase Grid Connection
The three-phase inverter, there are six gates (let we are using IGBT or MOSFET,
for this six gate drive terminals are required, gate drive is getting the control signal
from PWM block for RYB phases [7]. Now the input to the PWM will be control
signals coming from the output of controllers. Now, to control the inductor current
that is being pumped into the grid. Again, with the help of a comparator an I g
reference can be compared. Now, it compare current with the value of ig reference
and then pass it to the controller block it could be PI controller. Each phase of the
A New Topology for Photo Voltaic Energy Conversion System 619
Fig. 2 PV module inside blocks
three phases are need to compare and control the output of controllers which will
give it to the PWM block. So let us say each of the controllers is connected to PWM.
Currents are sensed compared with a reference and based on the error the controller
output will send the signal to the PWM which compares with the triangle carrier and
then appropriately gives the PWM drive to each of the arms of the inverter with proper
∗ ∗ ∗
energy management [8]. These references i gR , i gY , i gB are obtained from the MPPT
algorithm. All the three currents are sensed by using a hall sensor or alternatively
resistive shunt and differential amplifiers, instrumentation amplifiers can be used but
hall sensor is better because it is contactless and has large bandwidth. V T and I T are
measured by another hall sensor for MPPT block. Input to MPPT is V T and I T and
∗ ∗ ∗
output are i gR , i gY , i gB (These are sinusoidal values).
3 Controller
However there are lots of drawbacks in this particular type of the circuit, first of all,
∗ ∗ ∗
the references i gR , i gY , i gB are sinusoidal. They are not DC or setpoint value. The
controller becomes a tracking controller it is no longer a setpoint controller reference
current signal and feedbacks current signal all are AC sinusoidal in general.
Tracking controller needs to handle large signal deviation as the consequence the
B.W of tracking controller will be lower compared to a similar setpoint controller, if
∗
i gR and feedback signal have DC. So, design of tracking controller will be much more
complex compare to a set point controller and DC to DC converter in power system
[9]. It can also able to achieve intra cycle dynamics if one user setpoint controller
but for the tracking controller only integral cycle dynamics can be achieved. Another
problem is here three separate controllers one for each phase and the dynamics are
coupled, so fining the three controllers simultaneously will become difficult as the
dynamics are coupled. So those issues should be short out for the controller for
three-phase photovoltaic grid-connected inverters [10] in the control methodology.
4 Current Controlled Inverter
Input to PWM is reference V a , V b , V c (as desired) which will be compare with

the carrier triangle which PWM has and then generate the necessary PWM signals
(pulse). For the gate drive to switch ON and OFF accordingly and supply a voltage
here such that ia , ib , ic flows according to this control such that error just before PT
is zero. Now to define ρ, angle between the d axis and α axis so that both coordinate
systems rotates synchronously. α, β coordinate system has current space vector id
get aligned along the current space vector that is iq is equal to zero because id is
aligned itself along grid current. We use V a , V b , V c voltage waveforms to obtain ρ
then id and iq will be such that it will be with the respect to the voltage space vector
as the d-axis and it has an advantage. Let we take angle ρ between V α and α axis
also V g is not along V d and V q is not equal to zero. Then we get some error the PT
activates and PI controller will work in such a way that error to PI tends to zero. Then
V q is equivalent to Vq∗ which is equal to zero and d-axis will go and aligned along
V g voltage space vector and such a value of ρ (ρ will increase while aligning) will
come up because of control action such that V q is equal to zero here. In this way, ρ
gives the value of difference between d q axis and αβ coordinate system such that
d-axis aligned along the voltage space vector.
So, it becomes a robust mechanism because it is a close loop system and then there
is PI component that has a history in it which will have a filtering effect on harmonics,
surges, spikes, and uncertainties. So by incorporating this modification into our entire
three phase grid-connected inverter block diagram, it will be a completely workable
solution like other PV inverters used in [11, 12]. So we modify ρ-generation block.
5 SPWM (Unipolar PWM Technique)
According to IEEE—5/19/992-for household appliances total harmonic distortion

(THD) is less than five percent. So we can just attach a simple AC filter like LC filter
in order to reduce harmonics which is injected into the grid without using complex
AC filter also we can make THD less than five percent or by using base works as a
filter. This can be achieved by using a sinusoidal PWM technique in which we should
use less filter (Fig. 3).
6 Detailed Comparisons of Older Photo-Voltaic Conversion

Topology with Respect to New One
Various topologies are available in the past in which separate DC-DC converter [9]
which can be of any type of isolated or non-isolated type are present whose work is
just to match load line to the PV terminals in such a way that maximum power [13,
14] can be drawn from the PV cell. Unfolding circuit which is basically a two-way
switch type which switches ON alternatively can be replaced by a more popular
bridge stage which is decoupled from the isolating stage. In the new topology, DC-
DC converter and unfolding stage are replaced by a single power stage which is
called the inverter. This leads to an increase in the efficiency of this model. Power
transformer is of low frequency which is bulky, heavy, and expensive, so in this
electronic model, the transformer is shifted to DC-DC converter so that it can work
as high switching frequency which thereby reduces its weight, size, and works with
increased efficiency. System becomes more compact. This leads to transformerless
interface. In this high-frequency isolation flyback converter topology which is the
isolated version of the buck-boost topology that can handle the entire first quadrant
of the curve of the PV module is used. Further, improve on this topology by the
use of current-controlled inverter, DC-DC converter is eliminated which is doing
the job of maximum power point tracking. MPPT is done by the current-controlled
inverter [15] itself, there is inductor current which is pumped into the grid is being
controlled, then the maximum point algorithm is used for the terminal currents and
the terminal voltage of the PV panel are measured and power that is obtained is pass
through the maximum power point tracking algorithm..
7 Proposed Model of PV Energy Conversion System

in MATLAB
See Fig. 4.
8 Simulated Result
See Fig. 5.
Fig. 3 Current-controlled inverter

Fig. 4 MATLAB simulation model
Fig. 5 PWM output
9 Parameters with Different Rated Values Used
See Table 1.
Table 1 Data sheet for PV module elements

Module type BLD240-60P
Peak power 240 Wp
Maximum power voltage (V mp ) 30.18 V
Maximum power current (I mp ) 7.96 A
Open circuit voltage (V oc ) 36.72 V
Open circuit current (I sc ) 8.99 A
Cell η 16.50%
Module η 14.56%
Maximum system voltage DC 1000 V
Temperature coefficient of I sc α I +0.045%/K
Temperature coefficient of V oc α v −0.34%/K
Temperature coefficient of Pmax α p −0.47%/K
Series fuse rating 15 A
Cells 6 × 10 pieces polycrystalline solar cells series (156 mm ×
156 mm)
Junction box With 3 bypass diodes
10 Conclusion
Setpoint controllers only two have been used (while we need three for AC tracking
type controller). Working with DC values is easy because it is difficult to compare
changing values of magnitude of AC which is changing with respect to time with the
reference values for comparison. MPPT is integrated into the invertors by removing
the DC-DC converter and power block is simple (just has one power stage) so that the
analysis from the stability point of view becomes easy and the efficiency of the model
is increased. Robust Phase Locked Loop (PLL) based on ρ determination is done as
the current is injected into the grid via current terminals contain harmonics. So getting
ρ we converted current terminals into voltage terminals with the help of proper trans-
formation. In PLL open-loop becomes closed-loop by doing certain modifications,
in this way it becomes robust a robust mechanism there is proportional-integral (PI)
component there is history in it which will filtering the effect on harmonics, surges
and spikes.
References
1. IEEE TF Report (1982) Proposed terms and definitions for power system stability. IEEE Trans
Power Appart Syst PAS-101:1894–1897
2. Banerjee A, Mukherjee V, Ghoshal SP (2013) Modeling and seeker optimization based simu-
lation for intelligent reactive power control of an isolated hybrid power system. Swarm Evol
Comput 13:85–100
3. Kundur P et al (2004, Aug) Definition and classification of power system stability IEEE/CIGRE
joint task force on stability terms and definitions. IEEE Trans Power Syst 19(3):1387–1401.
https://doi.org/10.1109/TPWRS.2004.825981
4. Dinga Z, Houa H, Yua G, Hub E, Duana L, Zhaoc J (2019) Performance analysis of a wind
solar hybrid power generation system. Elsevier
5. Onara OC, Uzunoglua M, Alama MS (2018) Modeling, control and simulation of an
autonomous wind turbine/photovoltaic/fuel cell/ultra-capacitor hybrid power system. Elsevier
6. Tsai HL, Tu CS, Su YJ (2008) Development of generalized photovoltaic model using
MATLAB/Simulink. In: Proceedings of the World Congress on Engineering and Computer
Science (WCECS’08), San Francisco, USA, 22–24 Oct 2008
7. Figueres E, Garcera G, Sandia J et al (2009) Sensitivity study of the dynamics of three-phase
photovoltaic inverters with an LCL grid filter. IEEE Trans Ind Electron 706–717
8. Dali M, Belhadj J, Roboam X (2010) Hybrid solar–wind system with battery storage oper-
ating in grid-connected and standalone mode: control and energy management–experimental
investigation. Energy 35(6):2587–2595
9. Banaei M, Ardi H, Alizadeh R et al (2014) Non-isolated multi-input-single output DC/DC
converter for photovoltaic power generation systems. IET Power Electron 7(11):2806–2816
10. Kroutikova N, Hernandez-Aramburo C, Green T (2007) State-space model of grid-connected
inverters under current control mode. IET Electr Power Appl 1(3):329–338
11. Fazeli M, Ekanayake J, Holland P et al (2014) Exploiting PV inverters to support local voltage—
A small-signal model. IEEE Trans Energy Convers 29(2):453–462
12. Fortunato M, Giustiniani A, Petrone G et al (2008) Maximum power point tracking in a one-
cycle controlled single-stage photovoltaic inverter. IEEE Trans Ind Electron 55(7):2684–2693
13. Patel H, Agarwal V (2008) Maximum power point tracking scheme for PV systems operating
under partially shaded conditions. IEEE Trans Ind Electron 55(4):1689–1698
14. Noguchi T, Togashi S, Nakamoto R (2002) Short-current pulse-based maximum-power-point
tracking method for multiple photovoltaic-and converter module system. IEEE Trans Ind
Electron 49(1):217–223
15. Feng J, Wang H, Xu J et al (2017) A three-phase grid-connected microinverter for AC
photovoltaic module applications. IEEE Trans Power Electron
A User-Centric Task Allocation
Approach in Mobile Crowdsensing
Moirangthem Goldie Meitei and Ningrinla Marchang
Abstract Mobile crowdsensing (MCS) refers to the process in which a collection of

mobile devices share sensing data, which can then be aggregated to develop policies
or applications that benefit a community. Traditionally, mobile crowdsensing has
been deployed using a platform-centric task allocation approach. These platform-
centric systems are primarily concerned with gaining profits for the platform by
minimizing rewards given to users. Instead of this approach, this paper focuses on a
more altruistic task allocation approach by proposing a user-centric task allocation
approach. In this regard, we have developed two task allocation schemes: Nearest
User Task Allocation (NUTA) and Nearest User Fair Task Allocation (NUFTA) in
order to facilitate an optimized and equitable workload for the users.
Keywords Mobile crowdsensing · User-centric · Task allocation
1 Introduction
The growth of technology and the increasing availability of Internet access has led
to large-scale production and consumption of data. Boom of consumer electronics,
specifically mobile devices, provides a unique opportunity to gather data. These
mobile devices come equipped with inbuilt sensors, e.g., camera, GPS, accelerom-
eters, etc. Mobile crowdsensing (MCS) refers to a broad range of community
sensing paradigms, which requires the active involvement of individuals to contribute
sensor data (e.g., taking a picture, reporting a road closure) related to a large-scale
phenomenon. Examples of such large-scale phenomena include traffic monitoring,
air pollution level monitoring, noise pollution level monitoring, etc. [1].
Task allocation is the selection of tasks to be assigned to different participants.
Task allocation can take on two approaches. The first approach can be referred to
M. G. Meitei (B) · N. Marchang

Department of Computer Science and Engineering, NERIST, Nirjuli, Arunachal Pradesh, India
N. Marchang
628 M. G. Meitei and N. Marchang
as a platform-centric approach. Prevalent architectures are mainly focused on the

platform-centric task allocation approach. These systems are primarily concerned
with profits gained by the platform, by minimizing rewards given to users. Examples
of platform-centric systems include ride-sharing apps. The second approach can be
defined as a user-centric approach. This user-centric system serves as the main focus
of this study.
In user-centric approach, unlike the platform-centric ones, the main concern is
not the financial benefit of the platform but rather, it follows an altruistic perspective
and seeks to fulfill the interest of the users. It removes financial competition among
users and instead places importance on collective work for the benefit of the envi-
ronment or society. Examples include volunteer work, namely by non-governmental
organizations (NGOs), or feedback regarding traffic or weather conditions.
For our proposed user-centric task allocation mechanism, the main goal is to
reduce the workload on the users by minimizing the overall distance traveled to
accomplish the allocated tasks. Additionally, we also seek to provide a fair mecha-
nism for task allocation so that no users get flooded with tasks while other users are
left underutilized.
This paper tackles the task allocation problem in MCS and proposes a user-centric
task allocation strategy. The rest of the paper is organized as follows: Sect. 2 discusses
the literature related to task allocation in MCS. Section 3 explains the user-centric task
allocation mechanisms that are followed in this study. Section 4 briefly describes the
experimental setup which yields the results in Sects. 5, and 6 provides the conclusion.
2 Related Work
The earliest work on mobile crowdsensing and the Internet of Things (IoT) was
done by Ganti et al. in [1]. In their work, they identified a category of IoT applica-
tions that rely on data collection from a large number of mobile devices. They defined
mobile crowdsensing and differentiated them from traditional IoT applications. They
also identified unique characteristics of mobile crowdsensing, described the mobile
crowdsensing process, and presented challenges of mobile crowdsensing and briefly
discussed their solutions. Further, they presented the mobile crowdsensing architec-
ture along with its features and problems and proposed the exploration of a unified
mobile crowdsensing architecture to solve its problems.
Some works in MCS which deal with task allocation are [2–14]. He et al. proposed
a methodology to address the problem of location-dependent task allocation and
designed a local ratio based algorithm (LRBA) to solve the task allocation problem
[2]. They designed the pricing mechanism based on the bargaining theory. Zhang
et al. proposed a participant selection framework for mobile crowdsensing known
as CrowdRecruiter [3]. They formulated the concept of piggyback crowdsensing in
order to reduce energy consumption by leveraging smartphone app opportunities to
perform sensing tasks and return sensed results. Liu et al. proposed TaskMe, a partic-
ipant selection framework for multi-task mobile crowdsensing environments [4]. In
A User-Centric Task Allocation Approach in Mobile Crowdsensing 629
particular, two typical multi-task allocation situations with bi-objective optimization

goals are studied: (1) For FPMT (few participants, more tasks), each participant is
required to complete multiple tasks and the optimization goal is to maximize the total
number of accomplished tasks while minimizing the total movement distance. (2)
For MPFT (more participants, few tasks), each participant is selected to perform one
task based on pre-registered working areas in view of privacy, and the optimization
objective is to minimize total incentive payments while minimizing the total trav-
eling distance. They adopted the Minimum Cost Maximum Flow (MCMF) model to
solve the FPMT problem, and proposed the MT-MCMF and MTPMCMF algorithms
to select the optimal participants. Also, they used the multi-objective optimization
model to solve the MPFT problem, and proposed the W-ILP and C-ILP algorithms
to select the best participants. Ding et al. defined the service node selection problem
(SNSP) when selecting proper service nodes in mobile crowdsensing and formulated
a multi-objective optimization model (MOOM), which converts SNSP to a multi-
objective optimization problem (MOOP) [5]. Duan et al. worked on producing a
joint pricing and task allocation system for mobile crowdsensing platforms [6]. They
used Walrasian equilibrium to find the optimal solution for maximization problem of
pricing and task allocation. They devised a distributed iterative algorithm based dual
decomposition which was proven to converge to the optimal solution of social welfare
maximization. Zhou et al. proposed an online primal-dual auction framework that
connected a series of one-round auctions into an online auction [7]. They formulated
an integer linear program (ILP) that characterized the winner determination problem
(WDP) for social cost minimization in the online mobile crowdsensing auction.
Amin et al. in 2016 looked to solve the task allocation problem in crowdsensing by
maximizing the net reward of the platform under the time budget constraints of smart-
phone users and different quality requirements of tasks [8]. They proposed a Harmony
Search (HS) based meta-heuristic allocation algorithm to solve the task allocation
problem. Cheung et al. proposed an Asynchronous and Distributed Task Selection
(ADTS) algorithm to help the users plan their task selections on their own [9]. They
focused on solving the distributed time-sensitive and location-dependent task selec-
tion problem, where users are heterogeneous in their initial locations, movement
costs, movement speeds, and reputation levels. Xiong et al. proposed an energy-
efficient piggyback crowdsensing task allocation framework called iCrowd to opti-
mize task allocation [10]. The goal of iCrowd is to achieve maximum coverage and
minimum incentive payment. Hassani et al. presented a Context-Aware Task Allo-
cation (CATA) approach for task allocation by selecting the best participant set [11].
CATA also provides energy efficiency and privacy. Guo et al. proposed a worker
selection framework called ActiveCrowd for multi-task allocation in MCS environ-
ments [12]. They studied the workers’ intentional and unintentional movements for
time-sensitive and delay-tolerant tasks, respectively, and proposed greedy genetic
algorithms to solve them. Wang et al. introduced sparse mobile crowdsensing to
make use of the spatial and temporal correlation among sensed data to reduce overall
sensing cost while preserving the quality of data [13]. It selects a small area for
sensing and infers the data of the remaining area using compressive sensing. Zhan
et al. investigated the incentive mechanism for platform-centric MCS [14]. They also
analyzed the interaction between the platform and the users using Nash bargaining
theory.
Some works that focused on the quality of sensed tasks are [15–18]. Kantarci et al.
proposed using social network theory to evaluate the trustworthiness of crowd sensed
data, as well as the mobile devices that provide sensing services [15]. They combined
centralized reputation-based evaluation with collaborative reputation values based
on votes and vote capacities. Wang et al. looked to solve the problem of scheduling
different sensing tasks assigned to a smartphone with the objective of minimizing
sensing energy consumption while ensuring Quality of SenSing (QoSS) [16]. They
formulated the Minimum Energy Single sensor task Scheduling (MESS) problem in
which each sensing task only requests data from a single sensor and also presented
a polynomial-time optimal algorithm for this problem. Further, they formulated the
Minimum Energy Multi-sensor task Scheduling (MEMS) problem in which some
sensing tasks request multiple sensors to report their measurements simultaneously
and also presented an Integer Linear Programming (ILP) formulation and two effec-
tive heuristic algorithms to solve it. Liu et al. introduced Quality of Information
(QoI) based satisfaction ratio to quantify the degree of how collected sensor data can
satisfy QoI requirements [17]. They proposed a participant sampling behavior model
to select participants in energy-efficient and QoI aware manner. Pouryazdan et al.
introduced a new metric, collaborative reputation scores, to quantify crowdsensed
data trustworthiness based on reputation scores [18].
Other works on MCS which do not directly deal with task allocation are as follows.
Krontiris et al. performed a study on privacy and trust, incorporating both technical
approaches and actual sharing behavior [19]. They highlighted three main chal-
lenges with respect to crowdsensing, namely privacy, sharing practices, and fairness
which also serve as areas of research in the future. Wu et al. proposed a Dynamic
Trust Relationships Aware Data Privacy Protection (DTRPP) mechanism for mobile
crowdsensing [20]. They combined key distribution with trust management, by eval-
uating the trust value of a public key according to both the number of supporters and
the trust degree of the public key. Wang et al. proposed ARTSense, a framework to
solve the conflicting problem of trust versus privacy in mobile sensing [21]. Their
solution consists of a privacy-preserving provenance model, a data trust assessment
scheme, and an anonymous reputation management protocol where both positive and
negative reputation updates can be enforced.
Whereas there are some existing task allocation schemes [2–18], all of them follow
the platform-centric approach to the best of our knowledge. On the other hand, the
proposed approach is user-centric. In several applications such as traffic monitoring,
noise monitoring, weather monitoring, and so on, users are usually not paid for their
participation in the sensing campaign [22]. Hence, task allocation must focus on
improving the participation experience of the users which will further encourage
better and long-term participation. This study attempts to exploit some motivating
factors that will achieve this goal.
3 User-Centric Task Allocation
For the purpose of the study, the following three task allocation schemes have been
developed:
• Nearest User Task Allocation (NUTA): Here, each task is assigned to the nearest
user.
• Nearest User with Fair Task Allocation (NUFTA): Here, each user is allocated an
approximately equal number of tasks while trying to assign a task to the nearest
user as far as possible.
• Random User Task allocation (RUTA): Here, each task is assigned randomly to
any user.
In each of the allocation schemes, a set of n users and m tasks are considered.
A location is associated with each task, which is the location where the task has to
be performed. Similarly, we also assume that a user has a location, from which she
sets out for carrying out the tasks assigned to her. In Nearest User Task Allocation
(NUTA), for each task, the nearest user is selected based on the Euclidean distance
between their locations, and the task is assigned to the user. For each user, the
minimum cost of performing all its assigned tasks can be found by solving the
traveling salesperson problem (TSP).
Algorithm 1: Nearest User Task Allocation (NUTA)
1: dist 0
2: i 0
3: while i < m do m = number of tasks
4: task[i].distance 10000
5: j 0
6: while j < n do n = number of users
7: dist Euclidean(task[i], user[j])
8: if task[i].distance > dist then
9: task[i].distance dist
10: task[i].user assigned j
11: end if
12: j j+1
13: end while
14: i i + 1
15: end while
Algorithm 2: Nearest User with Task Allocation (NUFTA)

1: dist 0
2: i 0
3: while i < m do m = number of tasks
4: task[i].distance 10000
5: j 0
6: while j < n do n = number of users
7: if user[j].numTasks < numFairTasks then
8: dist Euclidean(task[i], user[j])
9: if task[i].distance > dist then
10: task[i].distance dist
11: task[i].user assigned j
12: end if
13: end if
14: j j+1
15: end while
16: i i+1
17: end while
Algorithm 1 illustrates the working of NUTA. Each task i is initially assigned
a large distance value, say 10,000. This distance is then compared (step 8)
with the Euclidean distance between the task i and the user j, denoted by
Euclidean(task[i],user[j]). If this Euclidean distance so calculated is smaller, the
distance value belonging to the task is updated with the smaller value and the corre-
sponding user is also saved. This comparison is done till all users are covered and
finally, each task i is assigned to the user j with the shortest Euclidean distance among
all the users.
A demerit of NUTA approach is that it may result in the skewness of task allo-
cation, i.e., some users may get allocated a large number of tasks while other users
may get assigned very few number of tasks. To mitigate this, NUTA is modified to
incorporate a fairness task allocation policy. We call this modified scheme as Nearest
User with Fair Task Allocation (NUFTA). Using this fairness task allocation policy,
all users are assigned with approximately equal number of tasks. For this, once a
user reaches a certain allocation of tasks (denoted by numFairTasks in step 7 of algo-
rithm 2), it is not allowed to take on further tasks. For a task, if its nearest user has the
maximum number of tasks already assigned, the next nearest user is then checked if
it has exceeded its maximum number of tasks. If this user is eligible, then the task is
allocated to it. Otherwise, the algorithm searches for the next nearest user and so on.
This approach is illustrated in algorithm 2. Again, each task i is initially assigned
a large distance value, say 10,000. Now a user j is taken into consideration only if its
number of tasks (denoted as user[j].numTasks) is less than the number of fair tasks
that can be allocated to each user (denoted by numFairTasks). The number of fair tasks
is given by the value m/n. Once it has been made sure that the number of tasks held
by the user j does not exceed numFairTasks, we proceed to calculate the Euclidean
distance between the task i and the user j, denoted by Euclidean(task[i],user[j]). The
rest of the algorithm follows the same procedure as algorithm 1.
Lastly, Random User Task Allocation (RUTA) algorithm is also developed

wherein each task is assigned randomly to any user. Whereas this algorithm is of no
practical importance, it is used as a baseline algorithm for comparison with the other
algorithms.
4 Experimental Setup
The implementation has been done in C. The simulation setup consists of the
following arrangement:
• Area: 200 × 200 sq. units
• Setup A: Fixed number of tasks (m) and varying number of users (n)
• Setup B: Fixed number of users (n) and varying number of tasks (m)
For setup A, we first fix the number of tasks and vary the number of users. Then
the tasks are allocated to different users using NUTA, NUFTA, and RUTA schemes.
For each task allocation scheme, we measure five parameters which are explained
below. We make a number of observations by varying the number of users from 5
till 25 (in intervals of 5) while keeping the number of tasks fixed to 100.
For setup B, we vary the number of tasks while keeping the number of users
fixed. Then the task allocation is implemented according to NUTA, NUFTA, and
RUTA schemes, and we measure the five parameters in each task allocation strategy.
Observations are taken by varying the number of tasks in intervals of 20 while keeping
the number of users fixed at 10.
The parameters that are measured for the simulation are as follows.
4.1 Overall TSP Cost
The overall TSP cost is the aggregated TSP cost of all the users combined. Once
each task has been assigned to a user, we find the TSP cost for all users and add them
to get the overall TSP cost of the setup. This is given by:

n
Overall TSP cost = TSP cost of user i (1)
i=1
4.2 Standard Deviation of TSP Cost
This metric computes the standard deviation of the TSP cost of all users.
4.3 Load Fairness Index (LFI)
We define the Load Fairness Index as the fairness of the load or tasks that are placed
on a user. It measures the equitable distribution of tasks among the users. The higher
the LFI value is, the more evenly the tasks are distributed and vice versa. This is
represented as follows:
n
i=1 tmax − ti
LFI = 1 − 1/n (2)
tmax
where t i represents the number of tasks assigned to user i and t max is the maximum
number of tasks allocated to any user.
4.4 Standard Deviation of Load Fairness Index (SDLFI)
This parameter calculates the standard deviation of the LFI obtained above.
4.5 Average Traveling Distance (AvTD)
Average Traveling Distance is the average of the sum distance of all the users to their
assigned tasks. The less the AvTD value is, the less a user has to travel to accomplish
the allocated tasks. This is denoted as:
n
1
AvTD = Total distance of user i (3)
n i=1
This section gives the simulation results of the three different task allocation schemes
NUTA, NUFTA, and RUTA w.r.t. setup A and setup B as discussed in the previous
section. Each point in all the plots represents the average result of 30 simulation runs.
Figures 1, 2, 3, 4, and 5 show the results for setup A, while Figs. 6, 7, 8, 9, and 10
show the results for setup B.
Figure 1 shows the overall TSP cost of the system across the three task allocation
strategies. It can be seen that NUTA has the lowest cost as compared to the other
two, and RUTA incurs a high cost as the number of users increases. The reason
for this is because the traveling cost increases in RUTA as the tasks are located far
Fig. 1 Overall TSP cost
Fig. 2 Standard deviation of

TSP cost
Fig. 3 Load fairness index

LFI
Fig. 5 Average traveling

distance
Fig. 6 Overall TSP cost

TSP cost
Fig. 8 Load fairness index


LFI
Fig. 10 Average traveling

distance
apart from one another. Moreover, under NUFTA, due to the added constraint of
allocating an approximately equal number of tasks to each user, a task may not be
assigned to the nearest user. Thus, its overall TSP coat is higher than that of NUTA.
Figure 2 shows the corresponding standard deviation for the TSP cost. The standard
deviation illustrates the amount of variation in the TSP cost of the users. In NUTA,
it tends to decrease as the number of users increases, whereas it is more uniform for
NUFTA and RUTA across a varying number of users. This means that more uniform
distribution of the number of tasks results in a more uniform TSP cost for the users.
The Load Fairness Index graph is shown in Fig. 3. It is seen that NUFTA has the
highest fairness among the three schemes, with NUTA giving the lowest fairness.
This is because fairness (equal distribution of tasks) is incorporated in the NUFTA
algorithm itself, whereas it is not in both NUTA and RUTA. Figure 4 shows the
corresponding standard deviation for the Load Fairness Index. As seen in Fig. 2, a
more uniform distribution of tasks results in less variation in the LFI.
Finally, the Average Traveling Distance can be seen in Fig. 5. It is observed
that NUTA has the least amount of Average Traveling Distance, with RUTA having
the highest. Again, this is because of the proximity of the assigned tasks to the
corresponding users. However, it can also be seen that NUFTA performs closely
with NUTA in this regard.
The corresponding set of results for setup B are shown in Figs. 6, 7, 8, 9, and 10.
To summarize the results, it can be seen that NUTA gives the lowest overall TSP
cost and AvTD, which implies that it outperforms other task allocation schemes
regarding tasks that are allocated without fairness. However, NUFTA performs quite
close to NUTA in terms of AvTD. Also, NUFTA scores highest in LFI, thereby
indicating that it provides the highest amount of fairness in task workload among all
the three schemes considered.
6 Conclusion
In this study, we have tackled the lack of alternatives to platform-centric task allo-
cation strategies in MCS by proposing a user-centric approach to task allocation.
Our proposal includes two task allocation schemes: NUTA and NUFTA. We found
that NUTA performs better in situations when tasks are allocated without consid-
ering fairness. Additionally, we have seen that NUFTA may be more beneficial than
NUTA in situations where the task workload is intended to be spread equally. In
future research, we wish to examine the effects of introducing a time-sensitive task
allocation approach, where each task has a set time by which it needs to be completed.
Acknowledgements This work is partially supported by DST-SERB, Government of India under

grant EEQ/2017/000083.
References
1. Ganti RK, Ye F, Lei H (2011) Mobile crowdsensing: Current state and future challenges. IEEE
Commun Mag 49(11):32–39
2. He S, Shin DH, Zhang J, Chen J (2014) Toward optimal allocation of location dependent tasks
in crowdsensing. In: IEEE INFOCOM 2014-IEEE conference on computer communications.
IEEE, pp 745–753
3. Zhang D, Wang L, Xiong H, Guo B (2014) 4W1H in mobile crowd sensing. IEEE Commun
Mag 52(8):42–48
4. Liu Y, Guo B, Wang Y, Wu W, Yu Z, Zhang D (2016) TaskMe: multi-task allocation in mobile
crowd sensing. In: Proceedings of the 2016 ACM international joint conference on pervasive
and ubiquitous computing. ACM, pp 403–414
5. Ding S, He X, Wang J (2017) Multiobjective optimization model for service node selection
based on a tradeoff between quality of service and resource consumption in mobile crowd
sensing. IEEE Internet Things J 4(1):258–268
6. Duan X, Zhao C, He S, Cheng P, Zhang J (2016) Distributed algorithms to compute Walrasian
equilibrium in mobile crowdsensing. IEEE Trans Indust Electron 64(5):4048–4057
7. Zhou R, Li Z, Wu C (2017) A truthful online mechanism for location-aware tasks in mobile
crowd sensing. IEEE Trans Mob Comput 17(8):1737–1749
8. Amin RG, Li J, Zhu Y (2016) A new harmony search based allocation algorithm for location
dependent tasks in crowdsensing. In: 2016 international conference on progress in informatics
and computing (PIC). IEEE, pp 581–585
9. Cheung MH, Southwell R, Hou F, Huang J (2015) Distributed time-sensitive task selection in
mobile crowdsensing. In: Proceedings of the 16th ACM international symposium on mobile
ad hoc networking and computing. ACM, pp 157–166
10. Xiong H, Zhang D, Chen G, Wang L, Gauthier V, Barnes LE (2015) iCrowd: near-optimal task
allocation for piggyback crowdsensing. IEEE Trans Mob Comput 15(8):2010–2022
11. Hassani A, Haghighi PD, Jayaraman PP (2015) Context-aware recruitment scheme for oppor-
tunistic mobile crowdsensing. In: 2015 IEEE 21st international conference on parallel and
distributed systems (ICPADS). IEEE, pp 266–273
12. Guo B, Liu Y, Wu W, Yu Z, Han Q (2016) Activecrowd: a framework for optimized multitask
allocation in mobile crowdsensing systems. IEEE Trans Hum-Mach Syst 47(3):392–403
13. Wang L, Zhang D, Wang Y, Chen C, Han X, M’hamed, A (2016) Sparse mobile crowdsensing:
challenges and opportunities. IEEE Commun Mag 54(7):161–167
14. Zhan Y, Xia Y, Zhang J (2018) Incentive mechanism in platform-centric mobile crowdsensing:
a one-to-many bargaining approach. Comput Netw 132:40–52
15. Kantarci, B, Glasser PM, Foschini L (2015) Crowdsensing with social network-aided collab-
orative trust scores. In: 2015 IEEE global communications conference (GLOBECOM). IEEE,
pp 1–6
16. Wang J, Tang J, Xue G, Yang D (2017) Towards energy-efficient task scheduling on smartphones
in mobile crowd sensing systems. Comput Netw 115:100–109
17. Liu CH, Zhang B, Su X, Ma J, Wang W, Leung KK (2015) Energy-aware participant selection
for smartphone-enabled mobile crowd sensing. IEEE Syst J 11(3):1435–1446
18. Pouryazdan M, Kantarci B, Soyata T, Foschini L, Song H (2017) Quantifying user reputation
scores, data trustworthiness, and user incentives in mobile crowd-sensing. IEEE Access 5:1382–
1397
19. Krontiris I, Langheinrich M, Shilton K (2014) Trust and privacy in mobile experience sharing:
future challenges and avenues for research. IEEE Commun Mag 52(8):50–55
20. Wu D, Si S, Wu S, Wang R (2017) Dynamic trust relationships aware data privacy protection
in mobile crowd-sensing. IEEE Internet Things J 5(4):2958–2970
21. Wang XO, Cheng W, Mohapatra P, Abdelzaher T (2013) Enabling reputation and trust in
privacy-preserving mobile sensing. IEEE Trans Mob Comput 13(12):2777–2790
22. Mousa H, Mokhtar SB, Hasan O, Younes O, Hadhoud M, Brunie L (2015) Trust management
and reputation systems in mobile participatory sensing applications: a survey. Comput Netw
90:49–73
Fault Classification Based Approximate
Testing of Digital VLSI Circuit
Sisir Kumar Jena
Abstract Test data compression scheme plays a major role in testing digital VLSI
circuits. Approximate Computing technique is a design technique that consumes less
power, area, and provides better performance with a loss in the quality of the result.
This paper combines the test compression scheme and the idea of approximation to
test digital VLSI circuits. This testing technique is applied to the circuits that are
used in error-tolerant applications like image/video processing. The basic idea is to
construct a fault set that is more vulnerable to error and generate test patterns for
them. After generating the test patterns, a conventional compression technique is
applied to those patterns to reduce the pattern count drastically. The experimental
result shows that the compression percentage is 70–80% compared to the traditional
compression scheme. Finally, the circuit with the injected fault is used in the image
processing application, and the result is generous.
Keywords Approximate computing · Approximate testing · VLSI testing · Fault

classification · Output masking
1 Introduction
Reducing test application time (TAT) is the primary objective of a tester. Power
consumption during testing of digital circuits is another issue that needs to be focused.
Test data compression techniques play a major role in reducing test application time.
According to [1], test data compression technique is classified into three groups: (i)
Code-based scheme (ii) Linear decompression-based scheme and (iii) Broadcast-
based scheme. Code-based scheme like Run- Length coding focuses on consecutive
0s or 1s and replaces it with x0 s or x1 s, where x represents the number of 0s or
1s, respectively. Golomb Coding [2], Frequency-Directed Run-Length coding [3],
Alternate Run-Length coding [4] are some of the methods that follow Run-Length
S. K. Jena (B)
DIT University, Dehradun, India
642 S. K. Jena
coding scheme. In the linear decompression-based schemes, there is a linear rela-

tionship between the uncompressed data and the compressed data which has to be
generated by the help of LFSR (Linear Feedback Register) or XOR networks. Some
of the decent works based on linear decompression-based scheme were presented in
the literature [5, 6]. In broadcast-based scheme, the same values can be broadcasted
to multiple scan chains [7, 8].
To tackle the problem of power issue, a subset of researchers is moving towards
a different domain of research named as approximate computing. They are trying
to propose different design paradigm, which takes the benefit of inherent tolerance
property of some application like image processing, digital signal processing (DSP)
and machine learning. Their objective is to design hardware that may not produce an
accurate result but produce a result, which is acceptable for the specific application.
By doing this, they achieve a considerable amount of power reduction as well as
enhanced speed and area overhead [9]. Enhancing this idea, we define a new term
approximate test which may be well as test time and test cost. The basic idea is
we will not consider all fault to be tested rather we will test only those faults that
affect the output most. In another word, we will leave some of the faults untested.
By doing this, we may get some errors in the output, which we consider acceptable.
The objective of our approach is to reduce test power consumption and test time
drastically.
Let us consider a scenario where a semiconductor manufacturing company specif-
ically design and fabricate a huge quantity of adder circuit that is used in DSP appli-
cations. We know that DSP applications such as image processing can tolerate some
amount of errors. Assume that the fabricated adder circuit has a fault f . Due to the
existence of fault f, the circuit does not produce accurate result rather it produces a
result, which can be acceptable for image processing applications. If the traditional
testing process is applied to test these circuits, then the entire manufactured circuits
are considered as bad chip and discarded. By using an approximate testing approach,
these bad chips can be made usable. Not only the chips are usable but also it takes
very less time and power to test these circuits as compared to the traditional approach
of testing.
The contribution of this paper includes the following:
• Introduce a new concept known as Approximate Testing which can be defined as
the process of testing only those faults which are critical in nature and exclude
the testing of faults which do not affect the normal working of the circuit.
• In the process of approximate testing, we provide an efficient methodology that
can classify the faults into two groups: acceptable group and unacceptable group.
• Our process drastically reduces the most critical issues associated with testing that
includes test power, test time, and test data volume. This is achieved by the test
data compression technique applied to all the faults included in the unacceptable
group.
The remaining section of the paper is organized as follows. Section 2 tries to give
credit to some previous works that help in shaping our idea. Section 3 introduces
the proposed approach in detail. Section 4 describes the details of fault classification
Fault Classification Based Approximate Testing … 643
and automatic test pattern generation. Section 5 explains the test compression step
and Sect. 6 shows the experimental result and finally Sect. 7 conclude the approach.
2 Motivation and Previous Works
Many research publications have focused on identifying those areas of a circuit which
can be approximated. In [10], a gate-level pruning technique is used to design an
approximated adder. Depending on the application’s requirement the designer can
choose how many gates to prune. The more gate you prune the more error you get in
the output. In [11], presented an 8×8 multiplier using 4×4 multipliers and they called
it a recursive multiplier. This multiplier was used in image processing application
for sharpening images. All these approximation techniques are applied in designing
a circuit. The same concept can be employed in testing a conventional circuit so that
test time can be reduced. Consider an 8-bit Ripple carry Adder circuit shown in Fig. 1.
According to approximate adder circuit design techniques, the errors present in LSB
bits does not affect the result most. For an example, we want to add two numbers a =
01110011 and b = 01100111. In Fig. 1, Case-0 shows the actual output produced.
Suppose the LSB bit produces an error due to some fault present in the 1st FA (Full
Adder). Three different case arises. Case-1: For our example data a = 01110011
and b = 01100111, F0 = 1 and carry = 1. Calculated error_significance = 1.
Case-2: F0 = 1 and carry = 0. Calculated error_significance = 1. Case-3: F0 = 0
and carry = 0. Calculated error_ significance = 2. The formal definition of error-
significance is described in Sect. 3. From this example, we found that error in low
order bit of an adder circuit does not affect the output. Hence, the low order circuit
(FA1) does not require to be tested. Because this adder circuit will be used in error-
tolerant applications like image processing, some amount of error in the circuit will
be tolerated. So, all the fault associated with FA1 can be ignored and we don’t require
test patterns to test those faults. All these faults can be treated as acceptable faults.
In the next section, we will propose our approach of testing digital circuits.
Fig. 1 Motivating example

644 S. K. Jena
3 Proposed Testing Methodology
To implement an approximate testing approach, we need to classify the faults. In

[12] author introduced three measures (i) error-rate (ii) error-accumulation and (iii)
error-significance. These measures are well known as RAS measures. Error-rate:
The percentage of test vectors for which the output deviates from the corresponding
error-free value, or a given circuit C, let we apply N test vectors and out of which, k
test vector produces error, the error-rate for this circuit C:
k
error-rate(C) =
N × 100
Error-accumulation: It is defined as the change in error-rate over time. Error-

significance: The absolute difference between the error-free output and erroneous
output. For a given circuit C, let us apply a test pattern p, which produces output z.
Assume that the error-free output for the pattern p needs to be z p . Then the error-
significance of fault f for which the pattern p is applied:

error-significance = z p − z
Error-significance is generally represented with a numeric value. In this paper, we

use error-significance as the measure for the classification of faults. In our approach,
the faults are classified into two groups. (i) The first group (acceptable group) consider
the faults whose error-significance should not exceed the threshold value provided
by the application. Here threshold value is application-specific. For example, let
the circuit C is an 8-bit adder and we fix the threshold as ±3. Assume that there
is a fault f exist in the circuit and for any pair of the input vector, if the output
produces an addition result within these threshold ranges then this fault f is included
in an acceptable group. (ii) The second group (unacceptable group) considers the
faults, which are critical in nature, i.e., due to the existence of these faults the circuit
may behave abruptly. In our approach, we need to test these faults only. A fault
classification method is presented in [13]. The objective of this paper [13] is to
improve the yield. It is shown that the conventional ATPG which was used to generate
test sets for the detection of unacceptable fault can also detect some acceptable
faults. This is known as over-detection problem. So, the consequence is degradation
in yield improvement. To deal with the over-detection problem some literature has
been proposed [14–16]. Initially, we will generate the test set for unacceptable group,
and then using a test data compression scheme, we will compress the test data so that
it takes less amount of memory in our ATE. Hence, our approach provides a drastic
reduction in test data volume.
The overall procedure of our proposed method is shown in Fig. 2. Initially an
error-significance is specified by the designer depending upon the application where
these circuits are used. Error-significance is calculated based on a threshold value
specific to an application. For example, if our design is an adder that is to be used in
Fig. 2 Proposed testing

methodology Application Specific Error- Performance
Significance Evaluation
Error-Significance analysis
CUT
for all possible faults
Classify faults into two

Generated
groups according to Error-
Test Case
Significance analysis
Acceptable Unacceptable
ATPG
Faults Faults
image processing application then threshold value may be ranged between 1 and 5.
Generally, each color pixel of an image takes the value from 000 to 255, and any value
between ±5 does not change the perceptual view of a human eye. The next step of
our process is to analyze error-significance for all possible faults. Depending on this
analysis, fault classification is made. In the presence of a fault if the error-significance
lies between the specified threshold value, then that fault is treated as a member of an
acceptable set. The remaining faults are grouped into an unacceptable fault set. Our
approach does not generate test patterns for acceptable faults. Unacceptable faults
need to be tested properly. ATPG process is only applied to unacceptable faults. After
generating the test set, it goes through a test data compression technique to further
reduce the size of the test set. This step ensures a reduction in test application time.
The resulting test set is then applied to CUT (circuit under test) and the performance
is evaluated to see the performance gain in terms of test application time.
4 Fault Classification and ATPG
4.1 Fault Classification
We consider stuck-at-fault model for fault classification. Initially, we define a

application-specific threshold value (Th ). For our case, it will be (Th ) = 5, because
our circuit will be used for image processing applications. We take a 16-bit ripple
carry adder for our purpose and intentionally induce single stuck-at-fault at each net.
Our assumption here is either our circuit is fault free or it contains only a single
stuck-at-fault. For each stuck-at-fault, the circuit is tested by some random inputs,
646 S. K. Jena
01 0 0 1 1 0 0 0
b3 a3 b2 a2 b1 a1 cin b0 a0
f1
S-A-0
cout s3 s2 s1 s0
0 1 1/0 0 0
Fig. 3 4-bit Ripple carry adder
and error-significance is calculated. Consider a 4-bit ripple carry adder circuit (RCA)
shown in Fig. 3 with fault set denoted by F. The inputs ai and bi are the binary numbers
with carry-in (cin ).
The outputs si and carry-out (cout ) is the result of the adder which is defined as
R = cout × 2i + si − 1 × 2i−1 + si − 2 × 2i−2 + · · · + s0 × 20 . Here i represents total
number of bits in the adder. Similarly, the faulty circuit of the given circuit is given
as RCA f and the result of the fault circuit is denoted as R f . The error-significance
(E s ) is calculated as, E s = R f − R . For a given input {a3 , a2 , a1 , a0 } = 1010,
{b3 , b2 , b1 , b0 } = 0010, cin = 0, the output produced by the faulty circuit (shown in
Fig. 3) is {cout , s3 , s2 , s1 , s0 } = 01000 instead of 01100. Hence R = 12 and R f =
8. Thus, error-significance E s = |8 − 12| = 4. As per the previously mentioned
threshold value, this fault is an acceptable fault. This is just a simple test input to
describe acceptability of a fault. As mentioned previously, we took a 16-bit RCA.
In total, we tested 480 faults without considering fault collapsing for our circuit. 50
random number pairs are given as input in the presence of a single fault in the circuit
and the error-significance is calculated.
4.2 ATPG
In this section, we will discuss the test pattern generation procedure. The objective
is to target the predefined unacceptable faults and discard the acceptable fault. A
deterministic test pattern generation scheme is used in our case because the random
pattern may or may not detect all unacceptable faults. As we have already mentioned
the pattern that detects an unacceptable fault may also detect an acceptable fault. This
is called the over-detection problem. In our case, the over-detection problem though
not ignored but we did not guarantee that our approach is fully solved the over-
detection problem. In [16], both deterministic and random test pattern generation is
considered, and over-detection problem is fully solved to improve yield. In this paper
our objective to reduce test pattern count as much as possible. We use the following
notations for our discussion and the overall procedure is described in Fig. 4.
• UFS and AFS: The set of unacceptable and acceptable fault sets, respectively.
Fig. 4 Test set generation procedure

648 S. K. Jena
• u i and a j : Represents an element (s) of UFS and AFS, respectively. Where i and
j vary from 1 to n and 1 to m, respectively.
• TUi : Represents the set of a test pattern that only detects u i but not ai .
• TA j : Represents the set of test patterns that detects at least one u i and at least one
ai .
• T: Complete list of test patterns. T = TUi TA j
Initially, we have an unacceptable fault list—UFS. Our loop will continue until
we generate test patterns for all u i ∈ UFS. Two possibility arises. (i) The pattern can
detect only the unacceptable fault u i without detecting any a j . (ii) The pattern may
detect at least one u i and at least one a j . In the first case, we directly add the test
pattern to the test set TUi . The second case is slightly handling the over-detection
problem. For this case, we add the test pattern to the test set TA j and apply an
output masking technique presented in [13] to handle the over-detection. According
to [13], when a test pattern generated to detect an unacceptable fault still detects an
acceptable fault we use the output masking technique. Output masking means not
observing a particular output line.
Output Masking Techniques—In general, a single fault can be detected by
multiple test patterns and it is propagated through more than one output line. Hence,
we can observe only one erroneous output for one particular test pattern and mask all
other erroneous output lines which need not be observed and counted as redundant.
This way we can avoid detecting an acceptable fault.
Example: An example that explains the output masking technique is presented
in Table 1. Consider a CUT with two output lines O1 and O2 . In the first case, an
unacceptable fault u 1 is targeted and as shown in the table, there are three patterns
T1 , T2 and T3 which can detect this targeted fault at the different observable points.
By applying the test pattern T1 we can observe u 1 and a1 at output pin O1 . Similarly,
by applying the test pattern T2 we can observe a1 and a2 at output pin O1 , also we
can observe u 1 and a3 at output pin O2 . Finally, by applying the test pattern T3 we
can observe u 1 at O1 and a2 at O2 . All these three patterns are able to detect the
fault u 1 but they also detect some acceptable fault. By analyzing all these patterns,
we choose T3 for detection of u 1 by masking O2 , so that acceptable fault a2 will not
be observed. In the second case, u 2 is targeted and two test patterns T4 and T5 can
detect u 2 along with some unacceptable fault. In this case, the masking is difficult,
Table 1 Output masking

Targeted faults Test patterns Observed OP Remark
O1 O2
u1 T1 u 1 , a1 a3 , a2 Choose T3 and mask a2
T2 a1 , a2 u 1 , a3
T3 u1 a2
u2 T4 u 2 , a1 a3 Difficult to mask. Unacceptable fault detected
T5 a1 u 2 , a2
hence we can mask any of the output pins which detect more number of acceptable
faults.
After output masking step, finally, all the test patterns will be merged to form set
T, which is nothing but the complete list of test patterns that detect all unacceptable
faults but a very few acceptable faults.
5 Test Data Compression
This section basically describes the technique to compress the test set T obtained in
Sect. 4. We have used the Huffman coding technique described in [17] to compress
T. A Huffman code [18] is a statistical coding technique obtained by constructing
a Huffman tree. For each leaf, a codeword is constructed by traversing from root to
the corresponding leaf. A reader can refer [17] to know more about the construction
of a Huffman tree.
6 Experimental Evaluation
Our experimental result is two-fold. With the first case, we show the benefits in terms
of the compression ratio. For this purpose, we use iscas 85 benchmark circuits. First,
we generate test vectors for the circuits using FastScan Tool by Mentor Graphics.
Then we redesign those circuits using Verilog code and remove the gates that are
not affecting/less affecting the final result. Again we generate the test vectors for the
modified circuit and finally apply Huffman test compression code to the generated
test vectors. The result is shown in Table 2. Data compression percentage (CP) will
be calculated using the following equation:
Original Bits − Compressed Bits

CP = × 100
Original Bits
For the second case, we took a 16-bit ripple carry adder (RCA) and apply
that circuit for image processing application, particularly for image smoothing.
Table 2 Experimental result ISCAS 85

Circuit # Gate # Test pattern CP (Huffman) CP (proposed)
C432 160 27 46.58 48.34
C499 202 52 69.47 71.84
C1355 546 84 71.49 76.21
C1908 880 106 51.83 65.23
C3540 1669 63 66.07 66.74
650 S. K. Jena
Fig. 5 Experimental result. a Original image with added noise. b Smoothen with accurate adder.
c Fault injection-1st FA. d Fault injection-2nd FA. e Fault injection-5th FA. f Fault injection-8th
FA
Smoothing is used to reduce the noise present in an image. We took a gate-level

netlist of 16-bit RCA. We start by injecting fault from LSB bits and used that circuit
to smooth an image. The entire image processing application will be implemented
in Matlab R2018a. The result of the simulation is shown in Fig. 5. There are five
gates (two XOR gate, two AND gate, and one OR gate) that drives the first LSB bit
and having 16 fault sites where we can inject the fault. The second LSB (from right)
drives by second FA (full adder). Here we consider the stuck-at fault model. As you
can see when more and more fault injected in the circuit and propagating from left
to right the image quality it gets distorted. Figure 5b, shows the image processed
with the original accurate adder, and as we proceed from Fig. 5c–f, the image quality
degraded. One thing we can notice in Fig. 5c that, when we inject fault in the LSB
bit the image quality is very much similar to the original adder’s image.
7 Conclusion
A new era of testing direction is presented in this paper which uses the concept of
approximate computing. Though the proposed idea does not generate test patterns
for the entire circuits but can able test a conventional circuit with very fewer test
patterns which is objective a test engineer with a constraint that the circuit may
produce some amount of when used in applications. Hence the proposed approach is
only applicable to test those circuit which is to be used in error-tolerant applications.
References
1. Li X, Lee K-J, Touba NA, Chapter 6—test compression. In: Wang L-T, Wu C-W, Wen X (eds),
VLSI test principles and architectures. Morgan Kaufmann, pp 341–396
2. Chandra A, Chakrabarty K (2001) System-on-a-chip test-data compression and decompres-
sion architectures based on Golomb codes. IEEE Trans Comput Aided Des Integr Circ Syst
20(3):355–368
3. Chandra A, Chakrabarty K (2001) Frequency-directed run-length (FDR) codes with application
to system-on-a-chip test data compression. In: Proceedings 19th IEEE VLSI test symposium.
VTS 2001, Marina Del Rey, pp 42–47
4. Ye B, Zhao Q, Zhou D, Wang X, Luo M (2011) Test data compression using alternating variable
run-length code. Integr VLSI J 44(2):103–110
5. Bayraktaroglu I, Orailoglu A (2003) Concurrent application of compaction and compression
for test time and data volume reduction in scan designs. IEEE Trans Comput 52(11):14801489
6. Mrugalski G, Rajski J, Tyszer J (2004) Ring generators: new devices for embedded test
applications. IEEE Trans Comput-Aided Des 23(9):1306–1320
7. Hamzaoglu I, Patel JH (1999) Reducing test application time for full scan embedded cores. In:
Proceedings of international symposium on fault-tolerant computing, pp 260–267
8. Han Y, Hu Y, Li X, Li H, Chandra A, Wen X (2005) Wrapper scan chains design for rapid and
low power testing of embedded cores. IEICE Trans Inform Syst E88D(9):21262134
9. Xu Q, Mytkowicz T, Kim NS (2016) Approximate computing: a survey. IEEE Des Test 33(1):8–
22
10. Schlachter J, Camus V, Palem KV, Enz C (2017) Design and applications of approximate
circuits by gate-level pruning. In: IEEE transactions on very large scale integration (VLSI)
systems, 25(5), pp 1694–1702
11. Maheshwari N, Yang Z, Han J, Lombardi F (2015) A design approach for compressor based
approximate multipliers. In: 2015 28th international conference on VLSI design, Bangalore,
pp 209–214
12. Breuer MA, Gupta SK, Mak TM (2004) Defect and error tolerance in the presence of massive
numbers of defects. In: IEEE design and test of computers, vol 21(3), pp 216–227
13. Hsieh T-Y, Lee K-J, Breuer MA (2007) Reduction of detected acceptable faults for yield
improvement via error-tolerance. In: Proceedings of design automation test Europe, pp 1599–
1604
14. Shahidi S, Gupta SK (2007) ERTG: a test generator for error-rate testing. In: Proceedings of
international test conference, p 110
15. Shahidi S, Gupta SK (2008), Multi-vector tests: a path to perfect error rate testing. In:
Proceedings of design automation test Europe, pp 1178–1183
16. Lee KJ, Hsieh TY, Breuer MA (2012) Efficient overdetection elimination of acceptable faults
for yield improvement. IEEE Trans Comput-Aided Des Integr Circ Syst 31(5):754–764
17. Kavousianos X, Kalligeros E, Nikolos D (2007) Optimal selective Huffman coding for test-data
compression. IEEE Trans Comput 56(8):1146–1152
18. Huffman DA (1952) A method for the construction of minimum redundancy codes. In:
Proceedings of IRE, vol 40, pp 1098–1101
Image Processing Using Case-Based
Reasoning: A Survey
Rahul Barman, Saroj Kr. Biswas, Sunita Sarkar, Biswajit Purkayastha,

and Badal Soni
Abstract Case-based reasoning (CBR) is a technique which solves a problem using

past experiences, where a case base stores these past experiences called cases. CBR
is used to solve different kinds of problems where past information is available. With
the advent of modern and efficient digital tools, huge amount of image data is being
captured. Hence, huge amount of image data is stored in different application fields.
Different organizations use CBR to find valuable and useful information from this
image data. Therefore, in the recent times, there is a huge use of CBR in image
data applications. This paper discusses the framework of CBR in image processing
and reports different application models in different applications. From literature,
it is observed that the performance of CBR in image processing is significant and
acceptable.
Keywords CBR · Image processing · Artificial intelligence · Machine learning
1 Introduction
Digital image processing (DIP) is a method through which some useful informa-
tion is retrieved from an image. DIP is used in different fields like space image
R. Barman (B) · S. Kr. Biswas · B. Purkayastha · B. Soni

Department of Computer Science and Engineering, National Institute of Technology Silchar,
Silchar, Assam 788010, India
S. Kr. Biswas
B. Purkayastha
B. Soni
S. Sarkar
Department of Computer Science, Assam University, Silchar, Assam, India
654 R. Barman et al.
processing, computerized photography, medical/biological image processing [1],

fingerprint recognition [2], face recognition [3], remote sensing [4], etc. Prepro-
cessing, edge detection, segmentation, feature extraction, etc., are the main DIP
techniques used to retrieve information from a digital image.
In the artificial intelligence field, case-based reasoning (CBR) is a technique which
uses the past experiences and solves a new problem. CBR has been widely applied to
different fields such as medical diagnostics, industrial manufacturing and question
and answer systems. The working principle of CBR is very similar to the human brain,
which compares the previous problems with the new problem to propose a solution
to the new problem. CBR overcomes the problems of rule-based systems and inherits
advantages from knowledge-based system. Besides CBR can learn by acquiring new
cases through use, CBR does not need to create any rule to solve a problem and
solution to a new problem can also be given through precedent. Therefore, since
decades, there is trend of CBR in image processing to solve some image processing
problems. This research article focuses on different applications of image processing
using different CBR techniques. This paper also attempts to describe how CBR can
be fitted in solving image processing problems.
2 Case-Based Reasoning
CBR maintains a case base where each case consists of problem details and the
solution for the problem. Whenever a new problem arises, it takes the help of old
problems and tries to solve it by its previous experiences. Hence, it is also called an
experience-based problem. There are four stages of CBR: retrieve, reuse, revise and
retain. In the retrieve stage, the most similar cases are retrieved from the case base by
the system. In the reuse stage, the system uses the solution of the most similar case to
solve the new case. In the revise stage, the applicability of the new proposed solution
is tested by the administrator. And in the retain stage, the new proposed solution and
problem are updated as a case for the future purpose. The working principle of CBR
is schematically described in Fig. 1.
3 Framework of Image Processing Using Case-Based

Reasoning
Nowadays, images are used in various fields due to the availability of images in
abundance. Moreover, the human brain can recognize a familiar object only in 100 ms
by looking at an image as image contains huge information about an event. So, image
data are generally preferred to do certain jobs. Some features are extracted from an
image by feature extraction technique and stored in the case base along with its
known class. The combination of extracted features and the known class is called
Image Processing Using Case-Based Reasoning: A Survey 655
Fig. 1 Schematic representation of CBR method
a ‘case’ in the case-based reasoning technique. When the system receives a new
image, it extracts the same features from the new image and compares them with
the stored features. The system retrieves the most similar cases from the case base
after comparing the features. The retrieved cases are reused to predict the class of
the new image. If the new case is not present in the case base, then the features and
the predicted class of the image are stored in the case base for future purpose. The
framework of image processing using CBR is shown in Fig. 2.
Fig. 2 Framework of image processing using CBR

4 Current Status of Image Processing Model Using CBR
Nowadays, image data are highly used in medical and industrial applications. CBR is
one of the popular problem-solving techniques. CBR methodology allows different
powerful ways to extract rich information from an image. Therefore, there is a trend
of image processing using CBR in different application domains.
Surjandy et al. [5] have stated a solution for traffic congestion problem using
CBR, where congestion data are taken from CCTV cameras. The pictures from the
CCTV camera are converted into binary patterns using DECODE [17]. Each pattern
from DECODE is identified as low, medium and high. The case base contains the
binary patterns as the individual cases, where each case describes a street condi-
tion. The conditions are described as either empty (low), smooth (medium) or solid
(high). Further pictures captured by the CCTV cameras are treated as a new case. The
picture is converted into a binary pattern and compared with the present patterns in
the case base. The system will give a traffic light solution, according to the compar-
ison performed by DIFFNOW [18]. This proposed system shows that a solution to
traffic congestion problem is obtainable from CCTV footage by use of CBR. The
drawback being the image captured by the camera directly affects the performance
of the system, and hence, can deteriorate from bad weather conditions and lowlight
conditions.
Hongru et al. [6] have discussed a case representation method based on image
information. Color of an image is not a very important feature but it is one of the
perception characteristics. Although scaling, rotation and translation cannot affect
the color characteristic of an image. Hence, this paper uses color as a feature of the
image, and the case base is created by that feature. An image is represented either
in RGB, i.e., red, green and blue color space or HSV, i.e., hue, saturation and value
color space. HSV color model is very similar to the human vision where H (hue)
represents different colors, S (saturation) represents the depth of color, and V (value)
represents the degree of color shading. Therefore, the model converts the image to
HSV color space from RGB color space, where the range of H is in [0, 360], S is
in [0, 1], and V is in [0, 1]. Due to the range of H, S and V, the color dimensional
histogram is also a large number. In CBR, if the retrieval procedure is done by the
feature calculation, then the computation will be very large. Hence, using histogram,
this calculation amount is much smaller. So qualify the HSV space to reduce the
feature factor dimension before calculating the feature quantity. Hence, the color
histogram of HSV space is used for describing the image characteristics, which is
used as the attribute of case description, including the R, G, B, H, S and V of the pixel
in RGB and HSV color space, respectively. For example, if we want to determine
the fire pixel in a forest image, the R value should be greater than the G value and
also there is a certain value for the H, S and V. If the conditions are matched, then
we can say that the pixel is a fire pixel.
Chen et al. [7] have presented a paper on synthetic-aperture radar (SAR) image
classification using CBR. There are mainly four steps used for SAR image clas-
sification, including image preprocessing, construction of case library, case-based
classification and post-classification. In image preprocessing step, speckle removing
and geometric correction are included. Frost adaptive filter [8] is used to remove
the speckle which is harmful for image analysis, from the SAR image. During
the case library construction, image attributes are used. Sensor platform, backscat-
ters, entropy, polarization, SAR imaging time and variance of SAR systems are
the attributes. The last three attributes are used for the case filtering that is when
a new SAR image arrives for the classification, the sensor platform, polarization
and imaging time of the new image are matched with the cases in the library, then
the most similar cases are chosen for the classification. For case-based matching,
similarity is measured by using some information which are variance, entropy and
backscatters. Some isolated spots are present in the initial images which are already
classified. An object-oriented post-classification method [9] is used to remove that.
Thus, in this article, it is shown that the CBR method for SAR image classification
is very promising where the overall accuracy of that experiment is 76.843%.
Khanum [10] has discussed about facial expression recognition system using
CBR. Human emotion is a thing that machine is unable to identify correctly. Hence,
if machine can understand the human facial expression, then it is possible to the
machine for understanding human emotion. In this article, CBR is used to understand
the human emotion by the facial expression, where eye, chin, cheeks, eyebrows,
lips, teeth, nose and forehead are used to express mainly six emotions of human.
The emotions which are detected in this paper are fear, joy, disgust, sadness, anger
and surprise. Facial expression intensity is one of the main features to classify the
emotion, and one of the major problems is to understand this feature. Humans produce
different intensity to express variety of expressions. Hence, for the same expression
of face, it should express different emotion if the intensity level is different. The
eight facial elements are extracted from the facial image, where some range is there
for different elements. The features from facial image are extracted by the image
processing techniques. And that features are taken by the system as input and give
output of expression state. A case consists of eight elements of the face with a value
ranging from V 1 to Vn. There is also an efficiency index (e) and a frequency of past
occurrence (f) present in a case which helps to retrieve cases from the case base. For
the similarity measurement, three factors considered are frequency of occurrence,
efficiency index and difference between the retrieved case and current case. The
method is compared with the hidden Markov models (HMM), [11] and it is shown
that the performance of CBR is better than the HMM method. The overall results for
different facial expression are 92%.
Hijazi et al. [12] have presented an article where an approach based on histogram
is used for classifying retinal images and CBR method is used for representing the
histogram of retinal images to identify the age-related macular degeneration (AMD),
where CBR has two stages. A histogram is created by the green and saturation values
with blood vessels removed, which is present in the first case base and second case
base consists of same histogram with optic disk (OD) removed. For a new retinal
image, a dynamic time warping (DTW) technique is used for similarity matching
in the first case base, for the remaining that cannot be classified in the first case
base is repeated with second case base. In this paper, it is shown that a two-pass
CBR approach is giving a better result than one pass CBR approach. However, two-
pass CBR fails on a small dataset as the OD removal is not always correct, and
it misclassifies the image. 82% sensitivity is the best average performance by the
model.
Lee et al. [13] used the function of mask algorithm (Fmask) and case-based
reasoning to improve the simple image simulation algorithm. Fmask is used to detect
clouds, shadow of clouds and water areas. And CBR is used to select appropriate
reference imagery to compare with the target imagery, selected by image simulation.
There are four steps of improved simulation algorithm which are preprocessing, solar
radiation modeling, selection of reference imagery and top-of-reflection simulation.
Third step is done by using the CBR method which predicts the result of new case
considering the result of past cases. The improved simulation algorithm creates a
better result with 95.59% accuracy.
Abdool et al. [14] have stated a solution for dish recommendation. It is seen
that physical characteristics of a food can give a deep perception about the dish
tastes. Hence, from the dish image, the aesthetic score is calculated by using image
processing techniques. The aesthetic score is the weighted sum of three aesthetic
attributes which are visual appeal score, colorfulness score and color spreadness
score. After calculating the aesthetic score and taking some more attributes of the
dish like flavor, texture, cooking style, price, etc., the system creates a case base
using that features. The recommended system learns the food preferences of different
users. The CBR system will give suggestions for food according to the preferences
of individual users.
Waqar et al. [15] have discussed suggesting some wearable item for a person to
look attractive and well-dressed using case-based reasoning technique. The system
takes a photo of the user with a camera and extracts some facial features like skin
tone, eye color, face shape, etc., with the help of image processing techniques. Beside
this feature, the system also takes some more information about the user like height,
weight, etc., that will help in reasoning. After collecting all the features from the
user, the system finds the best matching case from the case base. All the cases stored
in the case base are generated through some experts and also some old cases.
Raman et al. [16] have presented an article where mammogram images are used
for detection tumor from it that provides support for the clinical decision to perform
a biopsy of the breast. There are four stages of the process which are—first, mammo-
grams are digitized and images are enhanced; second, suspicious areas are detected
and localized; third, shape-based features are extracted in the mammogram which
is already digitized; and fourth, using case-based reasoning techniques features are
analyzed. Here the region growing method is used to detect the mammographic
mass. After the segmentation of mass, some useful features are taken for the deci-
sion making. Using the features which are extracted, obtain a matrix which has real
values. This matrix is used in case-based reasoning to analyze the mammogram mass.
The model generates results with 69.45% sensitivity and 69.2% specificity.
5 Summary
As a whole, it is shown that CBR is very useful in image processing techniques.

Many of the work are done by the CBR, and in future many more work can be done
by the CBR method. The summary of the all papers is given below in Table 1.
6 Conclusion
This paper discussed the image processing techniques using case-based reasoning
method and showed fitness of CBR in image processing. It is observed from the
literature that CBR performance in image processing is significant and acceptable.
Different application domains of image processing using CBR are reported. In future,
CBR can be used differently in many fields to improve the performance of systems.
Digital image processing is also a very useful technique nowadays as the availability
of images is in a huge amount. Hence, combination of CBR and image processing
technique can make a promising system.
Table 1 Summary of survey papers in chronological order

SL No Name of the paper and author Year Findings
1. Khanum: [10] facial expression 2006 CBR is used to understand the human
recognition system using case-based emotion by facial expression where
reasoning eye, chin, cheeks, eyebrows, lips, teeth,
nose and forehead are used to express
mainly six emotions of human
2. Chen et al. [7]: SAR image 2007 SAR image is classified using CBR
classification using case-based where adaptive filter is used to remove
reasoning speckle
3. Raman et al. [16]: digital mammogram 2009 Region growing method is used to
tumor preprocessing segmentation detect the mammographic mass, after
feature extraction and classification that some useful features are taken for
decision making and then CBR is used
for classification
4. Hijazi et al. [12]: retinal image 2010 Two-phase CBR is used to identify
classification using a histogram-based age-related macular degeneration,
approach where green and saturation histogram
with blood vessels removed are used in
first CB and same histogram with OD
removed is used in second
5. Hongru et al. [6]: research on case 2013 Color histogram of HSV space is used
representation method based on image for describing the image
information characteristics, which is used as the
attribute of case description
6. Waqar et al. [15]: wearable items 2013 System takes a photo of the user
suggestion system using case-based through a camera and extracts some
reasoning feature from the image, after that using
CBR system suggests some wearable
item for a person to look attractive
7. Abdool et al. [14]: recommend my 2015 Aesthetic score of a dish image is
dish: a multi-sensory food calculated through some image
recommender processing techniques and the
recommended system learns the food
preferences of different users using
CBR
8. Lee et al. [13]: Improvement on image 2017 Using CBR and Fmask, improve the
simulation from multitemporal Landsat simple simulation algorithm. CBR is
images used in selection of reference imagery
to predict the result
9. Surjandy et al. [5]: CCTV traffic 2018 Congestion data are taken from CCTV
congestion analysis at Pejompongan cameras, and the system gives a traffic
using case-based reasoning light solution using CBR where the
pictures are taken as a case
References
1. Kumar S, kumar B (2018) Diabetic retinopathy detection by extracting area and number of
micro aneurysm from color fundus image. In: 2018 5th international conference on signal
processing and integrated networks (SPIN)
2. Shetter A, Prajwalasimha SN, Swapna H (2018) Finger print image enhancement using thresh-
olding and binarization techniques. In: 2018 second international conference on inventive
communication and computational technologies (ICICCT)
3. Musa P, Al Rafi F, Lamsani M (2018) A review: contrast-limited adaptive histogram equaliza-
tion (CLAHE) methods to help the application of face recognition. In: 2018 third international
conference on informatics and computing (ICIC)
4. Peng L, Zhao Z, Cui L, Wang L (2004) Remote sensing study based on IRSA remote sensing
image processing system. In: GARSS 2004. 2004 IEEE international geoscience and remote
sensing symposium
5. Surjandy, Anindra F, Soeparno H, Napitupulu TA (2018) CCTV traffic congestion analysis at
Pejompongan using case based reasoning. In: 2018 international conference on information
and communications technology (ICOIACT)
6. Hongru L, Xiaochen H, Yubin Z (2013) Research on case representation method based on
image information. In: 2013 sixth international symposium on computational intelligence and
design
7. Chen F, Wang C, Zhang H, Zhang B, Wu F (2007) SAR image classification using case based
reasoning. In: 2007 IEEE international geoscience and remote sensing symposium
8. Frost VS (1982) A model for radar images and its application to adaptive digital filtering of
multiplicative noise. IEEE Trans Pattern Anal Mach Intell 42(2):157–166
9. Benz U, Baatz M, Schreier G (2001) OSCAR-object oriented segmentation and classification
of advanced radar allow automated information extraction. Int Geosci Remote Sensing Symp
(IGARSS) 4:1913–1915
10. Khanum A (2006) Facial expression recognition system using case based reasoning. In: 2006
international conference on advances in space technologies
11. Zhou X, Huang X, Wang Y (2004) Real-time facial expression recognition in the interactive
game based on embedded hidden Markov model. In: International conference on computer
graphics, imaging and visualization (CGIV’04)
12. Hijazi MHA, Coenen F, Zheng Y (2010) Retinal image classification using a histogram based
approach. In: The 2010 international joint conference on neural networks (IJCNN)
13. Lee SB, La HP, Kim Y, Dalgeun L, Kim J, Park Y (2017) Improvement on image simulation from
multitemporal Landsat images. In: 2017 IEEE international geoscience and remote sensing
symposium (IGARSS)
14. Abdool H, Pooransingh A, Li Y (2015) Recommend my dish: a multi-sensory food recom-
mender. In: 2015 IEEE Pacific Rim conference on communications, computers and signal
processing (PACRIM)
15. Waqar MM, Aslam M, Usman MG, Martinez-Enrqiuez AM (2013) Wearable items sugges-
tion system using case based reasoning. In: 2013 10th international conference on electrical
engineering, computing science and automatic control (CCE)
16. Raman V, Then P, Sumari P (2009) Digital mammogram tumor preprocessing segmentation
feature extraction and classification. In: 2009 WRI world congress on computer science and
information engineering
17. https://www.decode.fr/binary-image
18. https://www.diffnow.com
Automatic Speech Recognition of Galo
Karter Nyodu and Samudra Vijaya
Abstract The development of spoken language systems for the tribal languages
of India is very beneficial to society. The details of the implementation of auto-
matic speech recognition for Galo language, spoken in the northeast Indian state
of Arunachal Pradesh, are presented here. A multi-speaker speech database of
continuously spoken Galo sentences was specifically created for this purpose. The
speech recognition system was implemented using Kaldi, a public domain software
toolkit. The automatic speech recognition system recognizes Galo sentences spoken
continuously by new speakers with an accuracy of about 80%.
Keywords ASR · Galo · Speech database · Kaldi · Speech-to-text
1 Introduction
Arunachal Pradesh is one of the states in the northeastern region of India. Being home
to a large number of tribes and subtribes, a large number of languages are spoken in
the state. Galo, a language of the Tani branch of the Tibeto-Burman language family,
is spoken by the people belonging to the Galo tribe of Arunachal Pradesh. Galo is
one of 12 tribal languages of Arunachal Pradesh, listed in the ‘The Scheduled Castes
and Scheduled Tribes Lists’ published by the census of India [1]. Although 29,246
Indians stated Galo as their native language in 2011 [2], it is in the UNESCO list
of ‘vulnerable’ languages [3]. Figure 1 shows the map of Arunachal Pradesh state
wherein the primary area of Galo speakers is shaded [4].
Even though 7 of the 99 major, non-scheduled languages of India belong to the
state of Arunachal Pradesh [2], study and technology development of the languages
of the state are limited. A speech database of English, Hindi, and local language
K. Nyodu (B) · S. Vijaya

Centre for Linguistic Science and Technology, Indian Institute of Technology Guwahati,
Guwahati, India
S. Vijaya
664 K. Nyodu and S. Vijaya
Fig. 1 Map of Arunachal Pradesh where the Galo speaking area is shaded. Sources [4, 5]
of Arunachal was created and used for speaker identification [6]. The same authors
prepared a similar, but a larger database of speech from 200 speakers. This Arunchali
language speech database was used for identifying the language of the input speech
as one of the English, Hindi, Adi, Nyishi, Galo, and Apatani [7].
Very little research work on the Galo language has been reported in the litera-
ture. A book introducing the Galo language was written in 1963 [8]. A descriptive
grammar of the Lare dialect of Galo was the theme of a recent Ph.D. thesis work
[5]. The sole work on spoken Galo is an investigation of the acoustic features of
Galo phonemes using formant frequencies and cepstral coefficients as features [9].
However, no spoken language system for Galo, whether it is speech-to-text or text-
to-speech, has been implemented so far. The primary objective of this research work
is to implement a speech-to-text system for Galo. A secondary objective was to
create spoken language resources necessary for achieving the primary objective.
Specifically, the following are the outcomes of the current research work:
1. A Galo speech database consisting of sentences read by many native Galo
speakers.
2. Transcription of the speech data using the Galo script, which is a modified Roman
Script.
3. A multi-speaker, continuous speech recognition system for Galo language.
The organization of the paper is as follows. The details of the spoken language
resources developed for Galo language are given in Sect. 2. The implementation of
ASR systems using various acoustic models and evaluation methodology is described
Automatic Speech Recognition of Galo 665
Table 1 Two examples of

Dialect Gloss
dialect-dependent
transcription of words using Lare Pugo
modified Roman Script ‘aci ‘asi brother
inrv inye to go
in Sect. 3. The results and discussion are also presented in Sect. 3. Section 4 draws
some concluding remarks.
2 Spoken Language Resources
This section describes the steps for the development of linguistics resources for
training and evaluating the Galo Automatic Speech Recognition (ASR) system. The
subsections contain detailed descriptions of text and speech corpora.
2.1 Text Corpus
The text corpus contains a total of 200 short sentences, selected from the Galo-
English dictionary [10]. The text corpus consists of 721 unique words. Galo script,
which is a variety of modified Roman Script (MRS) [10], was used for writing the
text.
M.W. Post, in his Ph.D. thesis, lists 6 dialects of Galo: lare, pugo, tai, gensi,
karko, and zirdo [5]. Dialect-dependent variations in lexical terms were taken into
account while writing the transcriptions of recorded speech. Table 1 shows two such
examples of dialect-dependent variations.
2.2 Recording of Speech Data
Thirty-five Galo speakers were asked to read sets of 30–50 Galo sentences written in
modified Roman Script. Speech data were collected at users’ locations using laptop,
PC, and an earphone. The speech data were recorded at the sampling rate of 44.1 kHz,
16-bit, mono and were stored in wav format. The statistics of the number of speakers
and the speech files is given in Table 2.
The set of 35 Galo speakers belonged to two broad categories of dialects: Lare
and Pugo. The dialect-dependent statistics of the speech corpus is given in Table 3.
The lexical transcriptions of the speech files were dialect-dependent.
Table 2 Statistics of the

Number of Male Female Total
Galo speech corpora
Speakers 20 15 35
Speech files 850 650 1500
Table 3 Distribution of
Number of Dialect Total
speakers and files according
to dialect Lare Pugo
Speakers 22 13 35
Speech files 930 570 1500
2.3 Pronunciation Dictionary
A pronunciation dictionary was created manually according to the format specified

by Kaldi [11], the software toolkit used for the implementation of ASR systems. The
entries in the dictionary specify the pronunciation of each word in the text corpus in
terms of the phones or phone-like acoustic units of the language.
The Galo script [5] lists 7 vowels and 17 consonants. The script does not seem to
distinguish between long and short vowels. In addition, diphthongs are also used in
spoken Galo. Further, geminated consonants are present in the spoken Galo language.
In order to have acoustic models of these acoustic–phonetic variations, we use a
list of 19 (7 short + 7 long + diphthongs) vowel-like labels and 31 (17 single +
14 geminate) consonant-like labels. The labels follow the ILSL12 [12] convention,
augmented with notations to mark tones of the language. Table 4 shows a few entries
in the pronunciation dictionary.
Table 4 First two columns of

Word Label sequence Gloss
the table show typical entries
in the pronunciation panam panam To cut
dictionary paanam p aa n a m To hover
kai k ai Big
alo alo Salt
allo a ll o Tomorrow
It illustrates the ILSL12 labels used to prescribe the pronunciation
of Galo words involving long vowels, diphthongs, and geminated
consonants
3 Implementation, Evaluation, and Results
The Kaldi software toolkit was used for the implementation of the ASR system.
In this section, the details of the implementation, the evaluation methodology, the
results, and the discussion are presented.
3.1 Feature Extraction
The default setting of the Kaldi toolkit was used for feature extraction. Thirteen mel-
frequency cepstral coefficients (MFCC) were computed from every speech frame of
25 ms duration at a frame rate of 100/sec. Further, the first and second derivatives
of 13 MFCCs were computed. As a result, a 39-dimensional feature vector was
obtained.
3.2 Acoustic Modeling
For training the acoustics model, the default setting of the Kaldi toolkit was used.
Six types of hidden Markov models (HMM) were trained. Each state of HMM is
characterized by either a Gaussian mixture model (GMM) or a subspace GMM or
a deep neural network (DNN). A brief description of each of these models is given
below.
The simplest acoustic model (called Mono) models the context-independent
phone-like units, also known as monophones. The second model (Tri1) is the first
of a series of HMMs that model context-dependent phones (triphone). The third
acoustic model (Tri2) uses linear discriminant analysis (LDA) and maximum like-
lihood linear transform (MLLT). The fourth acoustic model (Tri3) uses the speaker
adaptive training and feature-space maximum likelihood linear regression (fMLLR).
The fifth acoustic model (SGMM-HMM) employs subspace GMMs instead of
GMMs. The last model is the DNN-HMM-based model which uses the posterior
probability given by a DNN to compute the state-dependent likelihood of feature
vector.
3.3 Language Model
Bigram language model was used to model the syntax. The parameters of this model
are estimated using the transcription of the train data. IRSTLM software was used
to train the language model [13].
Fig. 2 WER (%) of the Galo ASR system for different types of acoustic models. Test data are the
same as training data (entire speech corpus)
3.4 Results and Analysis
This section presents the performance of the Galo speech recognition system using
various acoustic models. The word error rate (WER) is used as a measure of the
performance of the ASR; lower the WER, the better the system is. The WER is
computed as
WER(%) = 100(D + S + I )/N
where D is the number of deletion errors, S is the number of substitution errors, I

is the number of insertion errors, and N is the total number of words present in the
reference transcription.
3.4.1 Performance on Training Data
The first evaluation was carried out using the same data for both training and testing.
Here, the entire speech data were used for this purpose. The word error rates of the
ASR system with different acoustic models are shown in Fig. 2.
The word error rate for the monophone model is 22%. The WER decreases dras-
tically by using context-independent phone models (Tri*). The WER is the lowest
(2%) for SGMM-HMM model. The WER of the DNN-HMM-based ASR system is
slightly higher than that of the SGMM model. This is possibly due to the small size
of the speech corpus, as DNN demands a large amount of speech data to adequately
train a large number of parameters.
3.4.2 Performance on Test Data
In order to estimate the performance of the ASR system with respect to unseen speech
data, a threefold cross-validation methodology was adopted. Accordingly, the entire
Table 5 Statistics of the three subsets of speech files used for threefold cross-validation of ASR
Number of Fold 1 Fold 2 Fold 3
Train Test Train Test Train Test
Utterances 990 510 1000 500 1010 490
Male speakers 13 6 12 6 12 7
Female speakers 10 6 11 5 11 5
Total 23 12 24 11 23 12
speech corpus was divided into three threefolds (subsets). These subsets were divided
in such a way that each subset had an approximately equal number of speech files
from both the female and male speakers. One subset was reserved/labeled as the test
set. The system was trained with the remaining two subsets. The WER of the system
with respect to the unseen test data is computed. This process is repeated for all
three sets. Such a threefold evaluation is carried out for all the six acoustic models.
The characteristics of the three subsets employed in our experiments are shown in
Table 5.
The WERs of six types of acoustics models in threefold evaluation experiments
are listed in Table 6. The WER of the ASR systems is around 20% for unseen data
in all threefolds. This value of WER for unseen test data is an order of magnitude
higher than that for the training data. Also, the difference between the WER of
the context-independent (monophone) model and the best (tri3) model is negligible.
Even though triphone models are more powerful, their potential is yet to be exploited
due to lack of adequate amount of training data. The WER increases from 18% to
26% when SGMM-HMM acoustic model is used. A similar increasing trend in WER
is observed when a DNN is used instead of a GMM.
The WERs of the ASR systems using six types of acoustic models are shown in
the form of a bar chart in Fig. 3.
Table 6 WERs of various types of acoustic models in threefold experiments

WER (%)
Model Fold 1 Fold 2 Fold 3 Average
Mono 19 21 17 19
tri1 18 23 17 19
tri2 21 25 19 22
tri3 17 20 15 18
SGMM 25 29 22 26
DNN-HMM 21 23 20 22
The WERs are around 20%
Fig. 3 WERs of various ASR systems in a threefold cross-validation experiment
4 Concluding Remarks
An automatic speech recognition system was implemented for Galo, a zero-resource

language, spoken by Galo tribals in the state of Arunachal Pradesh. The system
was trained using a preliminary multi-speaker speech database. The effectiveness of
different types of acoustic models was investigated using a threefold cross-validation
methodology. While the recognition accuracy of the ASR system is good for training
data, the accuracy decreases significantly for all six types of acoustic models. This is
due to the limited amount of speech data that could be collected in this initial effort.
Future work includes expansion of the spoken language corpus and investigation of
the utility of using prosodic features for machine recognition of this tonal language.
References
1. List of notified scheduled tribes, census of India. Online: http://censusindia.gov.in/Tables_Pub

lished/SCST/ST%20Lists.pdf
2. Office of the Registrar General and Census Commissioner India (2011) Statement-1 part-b
languages not specified in the eighth schedule (non-scheduled languages). Online: http://cen
susindia.gov.in/2011Census/Language-2011/Statement-1.pdf
3. Moseley C (ed.) (2010) Atlas of the world’s languages in danger, 3rd edn. Paris, UNESCO
Publishing. Online version: http://www.unesco.org/languages-atlas/en/atlasmap/language-id-
988.html
4. Office of the Registrar General and Census Commissioner, India, State Maps. http://census
india.gov.in/maps/State_Maps/StateMaps_links/arunachal01.html
5. Mark William Post (2007) A grammar of Galo. PhD Thesis, La Trobe University
6. Bhattacharjee U, Sarmah K (2012) A multilingual speech database for speaker recognition. In:
2012 IEEE international conference on signal processing, computing and control. https://doi.
org/10.1109/ispcc.2012.6224374
7. Bhattacharjee U, Sarmah K (2013) Language identification system using MFCC and prosodic
features. In: 2013 international conference on intelligent systems and signal processing (ISSP).
https://doi.org/10.1109/issp.2013.6526901
8. Das Gupta SK (1963) An introduction to the Gallong language. Shillong, Northeast Frontier
Agency
9. Sora M, Talukdar J, Talukdar PH (January, 2013) Formant frequency and Cepstral method
estimation of Galo phonemes using acoustical cues. Int J Inf Electron Eng 3(1)
10. Rwbaa I, Post MW, Rwbaa I, Xodu M, Bagra K, Rwbaa B, Rwbaa T, Ado N, Keenaa D (2009)
Galo-English dictionary with English-Galo index
11. Povey D, Ghoshal A, Boulianne G, Burget L, Glembek O, Goel N, Hannemann M, Motlicek P,
Qian Y, Schwarz P, Silovsky J, Stemmer G, Vesely K (December, 2011) The Kaldi speech recog-
nition toolkit. In: IEEE 2011 workshop on automatic speech recognition and understanding.
IEEE signal processing society. Online: https://publications.idiap.ch/downloads/papers/2012/
PoveyASRU20-112011.pdf
12. Indian Language speech sound label set (ILSL12). Online: https://www.iitm.ac.in/donlab/tts/
downloads/cls/clsv2.1.6.pdf
13. Federico M, Bertoldi N, Cettolo M (2008) IRSTLM: an open source toolkit for handling large
scale language models. Proceedings of Interspeech, pp 1618–1621
Early Detection of Diabetic Retinopathy
Using Machine Learning Techniques:
A Survey on Recent Trends
and Techniques
Dolly Das, Saroj Kr. Biswas, Sivaji Bandyopadhyay, and Sunita Sarkar
Abstract Diabetic Retinopathy (DR) is a medical condition which occurs due to

diabetes mellitus. It causes severe blindness due to mutilation of the retina of human
eye. According to statistics, 80% of the people, especially the working-age people
have been suffering from this disease. Hence, currently, DR has become an important
issue which needs to be solved at an early stage such that blindness can be prevented
to a great extent among the working-age people. Various intelligent systems have
been designed for early detection of DR as manual diagnosis is time consuming and
error prone. Besides, the availability of ophthalmologist is not possible at any time
and everywhere. Thus, the need of a highly optimized computer-assisted intelligent
system is required that can be used for the early detection of DR. Various models
have been proposed by researchers across the globe since decades. This paper aims
to give an elaboration over the works that have been done earlier for the detection
of DR and the recent technologies that have evolved for the same. This paper thus
gives a state of the art on the features, causes, symptoms, various grades of DR and
models that have been proposed and implemented for the early diagnosis of DR.
Keywords Diabetic Retinopathy · Diabetes · Image Processing · Machine

Learning · Retinal Lesions
D. Das (B) · S. Kr. Biswas · S. Bandyopadhyay

National Institute of Technology Silchar, Silchar, India
S. Kr. Biswas
S. Bandyopadhyay
S. Sarkar
Assam University, Silchar, India
674 D. Das et al.
1 Introduction
Diabetic retinopathy is a chronic progressive disease which occurs due to diabetes

mellitus. It causes damage to the retina of the human eye which leads to severe blind-
ness. It is also known as diabetic eye disease. Statistically, 80% of the people who
have been suffering from prolonged diabetes such as for 15–20 years suffer from
DR [1]. DR is found to be highly prevalent among working-age people. Thus, such a
medical condition requires early detection and diagnosis to prevent the disease from
progressing into severe stages, thereby reducing the occurrence of blindness. DR can
be detected on the basis of the existence of features and/or retinal lesions such as
Microaneurysms (MA), Foveal Avascular Zone (FAZ), Exudates (EX) and Hemor-
rhages (HE) [2–6], in the rear view of the human eye, i.e., the fundus. Here, for the
purpose of diagnosing the disease, a comprehensive eye examination is performed in
which the human eye is dilated by injecting medically approved contrasting agents.
The fundus images are then obtained using a digitized fundus camera [7]. The
ophthalmologist examines these fundus images to identify features such as MA, HE,
EX and FAZ and determines the severity of the disease. Besides these, certain other
retinal lesions such as ruptured retinal blood vessels, Cotton Wool Spots (CWS),
Intra Retinal Microvascular Abnormalities (IRMA) [2–6] can also be helpful for
analyzing and identifying the severity of the disease. Of all the features mentioned
above, manual analysis is feasible but time consuming. Thus, for faster analysis and
detection, which can help experts as well as fill the absence of ophthalmologists
in certain cases, the requirement of an intelligent system arises which can analyze
and compute the severity of DR using fundus images. Hence, a system could be
proposed that takes fundus images as input and classifies them into various cate-
gories of DR. Based on the presence and absence of the features mentioned above
and also taking into consideration the severity level, the disease is identified and
classified into five categories of DR such as No DR-0, DR-1, Mild Proliferative
Diabetic Retinopathy (MPDR)-2, Proliferative Diabetic Retinopathy (PDR)-3 and
Non-Proliferative Diabetic Retinopathy (NPDR)-4 [3, 4]. Some other methods [8,
9] have classified DR into two categories—DR and No DR. Thus, an intelligent
system can be designed to identify DR using different features. Different kinds of
works have been done since decades, proposing various intelligent and computer-
assisted systems which can perform automated analysis of the disease. Various tech-
niques have been proposed earlier for the detection and diagnosis of DR, such as
fuzzy c-means clustering [5], Multilayer Perceptron (MLP) and Extreme Learning
Machine (ELM) [10], neural network [10], Support Vector Machine (SVM) [8, 9, 12,
13], meta-SVM [11], Naive Bayes (NB) classifier [10, 12], probabilistic classifier,
geometric classifier, K-Nearest Neighbor (KNN) classifier and tree-based classifier
[12], Bayesian classifier [6, 14], Mahalanobis classifier [14], KNN classifier [12–
14], Gaussian Bayes classifier [3, 4], genetic algorithm [16], AlexNet deep neural
network (DNN) [17, 18], Convolutional Neural Network (CNN) [18] and various
other Machine Learning techniques [10, 11, 13, 16, 19]. Figure 1 depicts different
retinal lesions in human retina which appear due to DR [20]. Figure 2 shows the
Early Detection of Diabetic Retinopathy Using … 675
Fig. 1 DR retinal lesions
Fig. 2 Severe grades of DR
severe classes of DR as PDR and NPDR [21]. Figure 3a depicts DR classification,

(b) depicts common method for DR detection and (c) [28] depicts DR detection using
deep learning and CNN.
2 Features of DR
There are different features which can be used to identify DR at an early stage
such that blindness can be prevented. A brief description on these features has been
explained below:
676 D. Das et al.
Fig. 3 a DR classification b. Common method for DR detection c. DR detection using deep learning
and CNN
2.1 Foveal Avascular Zone (FAZ)
FAZ is a region situated within the fovea of the retina devoid of retinal blood vessels.
It has a diameter of 0.5 mm and a Field of View (FOV) of 1.5°. The FAZ center also
known as the macula center is the point of interest which is a significant landmark in
Fluorescein Angiography (FA). Figure 4 shows the presence of FAZ in DR-affected
retina [22].
Fig. 4 FAZ in DR retina

Fig. 5 a Presence of MA in DR. b Grayscale image of Fig. 5a
2.2 Microaneurysms (MA)
MA, also known as dot hemorrhage, is identified as the earliest symptom for the
diagnosis of DR. They are the localized capillary dilations and saccular in structure.
They generally appear in clusters as small red dots but may also appear in isolation.
The region encircled black in Fig. 5a shows the presence of MA in DR retina.
Figure 5b is the grayscale image in which the white dots spotted are MA [23].
2.3 Cotton Wool Spots (CWS)
CWS or soft EX are bloodless grayish speckles of contusion in the nerve fiber layer
of the human retina. They are a consequence of ischemia causing disarray in the flow
of axoplasm. Multiple CWS may count to nearly 6 or more than 6 in one eye and
may specify pervasive ischemia in the retina. Figure 6 shows the presence of CWS
in DR-affected retina [23].
Fig. 6 CWS in DR retina

678 D. Das et al.
Fig. 7 IRMA in DR retina
2.4 Intra Retinal Microvascular Abnormalities (IRMA)
IRMA are the areas of capillary dilation and intraretinal new vessel formation, which
arise within ischemic retina. Figure 7 shows the presence of IRMA in DR-affected
retina [23].
2.5 Hemorrhages (HE)
HE is also known as intraretinal HE which may appear in shapes such blots, flames
or dot-like structures depending upon its depth in the retina. HE appears as flame
shaped when found in the nerve fiber layer of the capillary network, thereby following
the divergence of the axons. However, it appears as dot shaped or blot shaped when
found in the inner nuclear layer of the capillary network, aligned at right angles to
the retinal surface. It is not exactly possible to differentiate between MA and HE;
however, making use of Fluorescein Angiography (FA) shall help to discriminate
MA by lighting up the MA when the human eye is dilated. The appearances of both
these features do not affect vision. However, multiple blots HE may imply significant
pre-proliferative retinopathy. Figure 8 shows the presence of HE in a DR-affected
retina [22].
Fig. 8 HE in DR retina
Fig. 9 EX in DR retina
2.6 Exudates (EX)
EX or hard EX are prominent bloodless yellow intraretinal accumulations with varia-

tions in size such as from a tiny pinprick to wide-ranging speckle and may evolve into
ring-like structures called circinate. They are a consequence of leakage composed of
body fat or extracellular lipid (drusen) accumulation under the retina causing macular
degeneration, arising from the abnormal retinal capillaries. Such a condition may
lead to Diabetic Macular Edema (DME) [15] causing swelling and thickening of the
macula in the retina. Figure 9 shows the presence of EX in DR-affected retina [23].
3 Different Methods for DR and Its Features Detection
Various works have been done with respect to premature recognition, detection and
diagnosis of DR, thereby prohibiting the occurrence of blindness. Some of the impor-
tant works have been stated below, which make use of various machine learning
techniques.
Using Multilayer Perceptron (MLP), Extreme Learning Machine (ELM) and
Naive Bayes (NB), Asha et al. [10] have performed the detection of DR using
680 D. Das et al.
Fig. 10 Performance of MLP, NB and ELM in [10]
exudates. Experimental results show that the model built using ELM outperforms
other two models, i.e., MLP and NB, and effectively detects the presence of EX in
retinal images. Figure 10 depicts the performance of MLP, NB and ELM in [10].
Using meta-Support Vector Machine (meta-SVM), Li et al. [11] have simultane-
ously detected multiple types of lesions through fusion of classifiers and amalga-
mated the out-turn of every single SVM. It exploits the information using clustering
to improve detection of lesions [11]. An accuracy of 99.44% has been achieved
for detection of optic disk, and 93.49% has been achieved for detection of macula,
for the detection of DR. Using Support Vector Machines (SVM), various systems
have been proposed for detection of EX and non-EX in retinal lesions [8, 12], iden-
tification of HE trained on High-Resolution Fundus (HRF) and Standard Diabetic
Retinopathy Database (DIARETDB1) datasets [9], detection of EX and MA using
SVM and KNN classifier [13], detection of hard EX [24], for detection of DR and
its severe stages. A sensitivity of 100%, a specificity of 94.6% and an accuracy of
96.66% have been attained in [8]. The sensitivity and specificity of HE detection
are 94.76% and 99.85%, respectively, classification rate is 95% and Peak Signal-to-
Noise Ratio (PSNR) of 51.45 dB in [9]. Different variants of SVM such as Gaussian
SVM (GSVM), Cubic SVM (CSVM), Quadratic SVM (QSVM) and Linear SVM
(LSVM) kernel functions have been trained, validated and tested on datasets such
as e-ophtha, High-Resolution Image Set (HRIS), Methods to Evaluate Segmenta-
tion and Indexing Techniques in the field of Retinal Ophthalmology (MESSIDOR),
DIARETDB1, Vascular Disease Image (VDIS), Digital Retinal Images for Vessel
Extraction (DRIVE) and HRF, for detection of DR [12]. Figure 11 depicts the clas-
sification accuracy obtained for the variants of SVM to the corresponding dataset [8,
12]. Lachure et al. [13] have attained a specificity of 100% and a sensitivity of more
than 90%.
Using probabilistic, geometric, KNN and tree-based classifier [12], retinal lesions
have been identified and segregated as EX and non-EX region for detection of DR.
Five probabilistic-based classifiers, namely NB, Bayesian Net(BN), NB update-
able, Multinomial Naive Bayes (MNB) and Bayesian Logistic Regression (BLR)
and three KNN Kernels (KNNK), namely Fine KNN Kernel (FKNNK), Weighted
KNN Kernel (WKNNK) and Medium KNN Kernel (MKNNK) classifiers have been
successfully deployed for classification of retinal lesions in the process of detection
of DR. Figure 11 gives a representation of different variants of classifiers along with
Fig. 11 Different variants of classifiers and their accuracy to the corresponding dataset
their classification accuracy to the corresponding dataset [12]. Ege et al. [6] have
also proposed a DR screening system for identification of MA, HE, EX and CWS
as dark and bright abnormalities. Various statistical classifiers such as the Bayesian
classifier, the Mahalanobis classifier and the KNN classifier have been tested. The
KNN classifier achieved a sensitivity of 93% for MA, 51% for HE, 100% for EX
and 15% for CWS. Lachure et al. [13] have also proposed a methodology to detect
EX and MA, for detection of DR using KNN classifier.
Using Bayesian classifier [6], Mahalanobis classifier [6], KNN classifier [6, 12,
13], DR detection has been performed by Ege et al. [6]. The Bayes classifier and
Mahalanobis distance classifier have been used to estimate the covariance, mean and
prior probability on the learning set. KNN classifier performs classification of dark
and bright abnormalities on 134 retinal images. Mahalanobis classifier performed
better classification of bright abnormalities than Bayes and KNN classifier. Figure 12
lists the performances of the three classifiers with respect to the sensitivity parameter
[6]. Hsiao et al. [14] have proposed a detection scheme for optic disc in retinal images
eliminating unacceptable contour segmentation, for detection of DR. The Supervised
100
80
60 Mahalanobis
40 Bayes
KNN
20
0
MA in % HE in % EX in % CWS in %
Fig. 12 Performances of the three classifiers with respect to the sensitivity parameter
682 D. Das et al.
Gradient Vector Flow (SGVF) is deployed for the edge detection of optic disc (OD).
The Bayesian classifier identifies the correct and incorrect contour point and classifies
them as edge points or non-edge points. Using Gaussian Bayes classifier [16, 17], the
FAZ area is measured to detect DR. Frazil et al. [3, 4] performs pattern classification
upon classes having Gaussian distribution. A sensitivity of 95%, a specificity of 97%
and an accuracy of 98% have been obtained for various stages of DR using Log
Posterior Probability Ratio (LPPR). The classifier shows a high specificity of 97%
and a sensitivity of 84% for both mild and moderate NPDR. On an overall basis,
the DR system can detect DR with a higher sensitivity of 90.81%, a specificity of
98.29% and an accuracy of 97.46%, for all the stages of DR.
Using AlexNet DNN, Mansour, R.F. et al. [19] have proposed that features can
be extracted from Region of Interest (ROI) using Convolutional Neural Network
(CNN) [11]. It models deep learning methods to detect ascribable DR in Kaggle
dataset and MESSIDOR-2 database, exhibiting a better performance on the Kaggle
dataset. Linear Discriminant Analysis (LDA) and Principal Component Analysis
(PCA) have been used for feature selection and extraction of multidimensional
features, thus outperforming Spatial-Invariant Feature Transform (SIFT)-based DR
detection and other existing system with a highest classification accuracy of 97.93%
[19]. It achieves a sensitivity of 99% and a specificity of 71% and area under the
ROC curve of 0.97.
Various other ML [10, 13, 18, 20] techniques such as Adaptive Machine Learning
(AML) [21], CNN [25] for identification of MA, EX and HE, deep learning Arti-
ficial Intelligence(AI) [26] and heat map generation for ConvNets [27] have been
proposed for detection of DR. In [27], the area under the Receiver Operating Charac-
teristic (ROC) curve, Az of 0.954 and 0.949, for 2015 Kaggle Diabetic Retinopathy
competition and e-ophtha, respectively, for detection of referable DR (rDR). CNN is
trained using 128,175 images graded for DR and DME, using Eye Picture Archive
Communication Systems (EyePACS-1) and MESSIDOR-2 dataset.
4 Conclusion
The paper gives an elaborate idea in brief about DR, its symptoms, features, shape,
size and location of the features, various causes that lead to DR and how DR causes
blindness. This paper also gives a brief introduction about the works that have been
performed by various researchers and experts working on the domain of expert
systems for DR since decades. The paper shows different prominent features respon-
sible for DR detection. The models proposed by the researchers have introduced
different grades for classification of DR. Besides, this paper focuses on the study
and analysis of various machine learning techniques that have been deployed such
as fuzzy c-means clustering, MLP and ELM, neural network, meta-SVM, SVM, NB
classifier, probabilistic classifier, geometric classifier, KNN classifier and tree-based
classifier, Bayesian classifier, Mahalanobis classifier, KNN classifier, Gaussian Bayes
classifier, genetic algorithm, AlexNet DNN, CNN and various other ML techniques
to model systems for early DR detection and classification. Recent models such as
SVM, AlexNet DNN and CNN have evolved with efficient results compared to rest
of the models, thus making real-life application of such systems, easy and faster for
detection. Thus, this paper is very helpful for new and young researchers who are
working on the domain of DR.
References
1. Fong DS, Aiello L, Gardner TW, King GL, Blankenship G, Cavallerano JD, Ferris FL, Klein
R (2004) Retinopathy in diabetes. Diab Care 27(Suppl 1):s84–s87
2. Kumar PS, Deepak RU, Sathar A, Sahasranamam V, Kumar RR (2016) Automated detection
system for diabetic retinopathy using two field fundus photography. Procedia Comput Sci
93:486–494
3. Fadzil MHA, Ngah NF, George TM, Izhar LI, Nugroho H, Nugroho HA (2010) Analysis of
foveal avascular zone in colour fundus images for grading of diabetic retinopathy severity. In:
2010 annual international conference of the IEEE engineering in medicine and biology, pp
5632–5635
4. Fadzil MA, Izhar LI, Nugroho H, Nugroho HA (2011) Analysis of retinal fundus images for
grading of diabetic retinopathy severity. Med Biol Eng Comput 49(6):693–700
5. Sopharak A, Uyyanonvara B, Barman S (2009) Automatic exudate detection from non-dilated
diabetic retinopathy retinal images using fuzzy c-means clustering. Sensors 9(3):2148–2161
6. Ege BM, Hejlesen OK, Larsen OV, Møller K, Jennings B, Kerr D, Cavan DA (2000) Screening
for diabetic retinopathy using computer based image analysis and statistical classification.
Comput Methods Prog Biomed 62(3):165–175
7. Goh JKH, Cheung CY, Sim SS, Tan PC, Tan GSW, Wong TY (2016) Retinal imaging techniques
for diabetic retinopathy screening. J Diab Sci Technol 10(2):282–294
8. Senapati RK (2016) Bright lesion detection in color fundus images based on texture features.
Bull Electr Eng Inf 5(1):92–100
9. Raja SS, Vasuki S (2015) Screening diabetic retinopathy in developing countries using retinal
images. Appl Med Inf 36(1):13–22
10. Asha PR, Karpagavalli S (2015) Diabetic retinal exudates detection using machine learning
techniques. In: 2015 international conference on advanced computing and communication
systems, IEEE, pp 1–5
11. Li B, Li HK (2013) Automated analysis of diabetic retinopathy images: principles, recent
developments, and emerging trends. Curr Diab Rep 13(4):453–459
12. Amin J, Sharif M, Yasmin M, Ali H, Fernandes SL (2017) A method for the detection and
classification of diabetic retinopathy using structural predictors of bright lesions. J Comput Sci
19:153–164
13. Lachure J, Deorankar AV, Lachure S, Gupta S, Jadhav R (2015) Diabetic Retinopathy
using morphological operations and machine learning. In: 2015 IEEE international advance
computing conference (IACC), pp 617–622
14. Hsiao HK, Liu CC, Yu CY, Kuo SW, Yu SS (2012) A novel optic disc detection scheme on
retinal images. Expert Syst Appl 39(12):10600–10606
15. Agurto C, Murray V, Yu H, Wigdahl J, Pattichis M, Nemeth S, Barriga ES, Soliz P (2014) A
multiscale optimization approach to detect exudates in the macula. IEEE J Biomed Health Inf
18(4):1328–1336
16. Decencière E, Cazuguel G, Zhang X, Thibault G, Klein JC, Meyer F, Marcotegui B, Quellec
G, Lamard M, Danno R, Elie D (2013) TeleOphta: Machine learning and image processing
methods for teleophthalmology. Irbm 34(2):196–203
684 D. Das et al.
17. Mansour RF (2018) Deep-learning-based automatic computer-aided diagnosis system for

diabetic retinopathy. Biomed Eng Lett 8(1):41–57
18. Rakhlin A (2018) Diabetic retinopathy detection through integration of deep learning
classification framework. bioRxiv 225508
19. Porwal P, Pachade S, Kamble R, Kokare M, Deshmukh G, Sahasrabuddhe V, Meriaudeau F
(2018) Indian diabetic retinopathy image dataset (IDRiD): a database for diabetic retinopathy
screening research. Data 3(3):25
20. Fraz MM, Jahangir W, Zahid S, Hamayun MM, Barman SA (2017) Multiscale segmentation
of exudates in retinal images using contextual cues and ensemble classification. Biomed Signal
Process Control 35:50–62
21. https://www.glycosmedia.com/news/ 12–01–2020 to 13–01–2020
22. Kale P, Janwe N (2017) Detection of retinal hemorrhage in color fundus image. Int J Adv Res
Comput Commun Eng 6(3)
23. Kaur J, Mittal D (2018) A generalized method for the segmentation of exudates from
pathological retinal fundus images. Biocybernetics Biomed Eng 38(1):27–53
24. Tjandrasa H, Putra RE, Wijaya AY, Arieshanti I (2013) Classification of non-proliferative
diabetic retinopathy based on hard exudates using soft margin SVM. In: 2013 IEEE
International Conference on Control System, Computing and Engineering. IEEE, pp 376–380
25. Pratt H, Coenen F, Broadbent DM, Harding SP, Zheng Y (2016) Convolutional neural networks
for diabetic retinopathy. Procedia Comput Sci 90:200–205
26. Takahashi H, Tampo H, Arai Y, Inoue Y, Kawashima H (2017) Applying artificial intelligence
to disease staging: deep learning for improved staging of diabetic retinopathy. PLoS ONE
12(6):e0179790
27. Quellec G, Charrière K, Boudi Y, Cochener B, Lamard M (2017) Deep image mining for
diabetic retinopathy screening. Med Image Anal 39:178–193
28. Chandrakumar T, Kathirvel R (2016) Classifying diabetic retinopathy using deep learning
architecture. Int J Eng Res Technol 5(6):19–24
GUI-Based Secure Architecture Design
for Distributed Community Micro-grid
Praveen Tiwari, Maj Sunil Kumar Panwar, Bikram Paul,

Pinakeswar Mahanta, and Gaurav Trivedi
Abstract Power generation from solar photovoltaic (PV) plants depends on multiple
factors such as solar radiance and ambient temperature. Effective monitoring and
control of the energy generated with the solar PV plants may augment the complete
performance of the system. The local controllers regulate switching between solar
PV, local battery, community battery, and grid according to the availability of power
in each system. The sensor data and user information are stored in the cloud that
can be accessed with the Internet of Things (IoT) device for further computation.
A Web-based graphical user interface (GUI) has been developed for appropriate
monitoring and control of energy flow and also to enable the power pooling. The
app gets information from the cloud on which various sensor’s data regarding energy
pooling are stored by the local/community controllers, and it displays the information
that is stored in the cloud. The primary task is to design and develop a Web-based
application with sufficient data security. To secure the user data in the cloud, multi-
factor authentication and symmetric key encryption are used for authorized access
and data encryption. The basic model of the Web-based application is presented in
this paper.
Keywords Distributed community micro-grid · IoT · GUI · Identity access

management · Solar photovoltaic
P. Tiwari (B)
Center for Energy, Indian Institute of Technology Guwahati, Guwahati, India
M. S. K. Panwar · B. Paul · G. Trivedi
Department of Electronics and Electrical Engineering, Indian Institute of Technology Guwahati,
Guwahati, India
P. Mahanta
Department of Mechanical Engineering, Indian Institute of Technology Guwahati, Guwahati, India
686 P. Tiwari et al.
1 Introduction
The need for energy is rising rapidly in the developing countries [1]; consequently,
effective energy conservation and management is desired in the modern power system
operation [2]. This growing demand needs to be matched with additional energy
resources. Renewable energy resources such as solar photovoltaic (SPV) are being
promoted by the countries due to low production and maintenance costs, and also, it
creates a lower impact on the environment by reducing the emissions of CO2 . SPV
and fuel cells, which are the leading environment-friendly energies, produce a direct
current (DC). Due to the development of highly efficient power converters, DC power
distribution system is gaining interest over AC distribution [3]. Community smart
DC micro-grid can increase and improve the performance of the network in multiple
aspects, such as minimizing the power losses in transmission, reducing reliance on
the grid, and economizing the upgrading of smart grid infrastructure. Since micro-
grids do not send power over long distances, they require less metal indulge in heavy
metal lines, which can make the best use of low to medium voltage systems [4, 5].
This system has to be user-friendly, remotely monitored, and controlled, and also,
the appropriate energy trade model should be incorporated in it. Real-time moni-
toring of various parameters of energy generation with SPV plants is becoming
important to augment the complete performance of the SPV systems. Smart home
energy management system (SHEMS) augmented with IoT has been proposed for
effective usage of SPV [6–8], which enables the user to monitor energy consump-
tion. For the improved monitoring, graphical user interface (GUI) is to be designed,
which includes graphically plotted real-time data like temperature, power, current,
and voltage, and solar irradiance is reported [9–12]. The GUI applications must be
secure to avoid any type of cyber-attack. Various methods to improve the security
features of the GUI application are reported [13, 14]. However, there is scope for the
development of a Web application for SHEMS with enhanced security. In this paper,
the SHEMS is designed for the switching between SPV, battery, community storage,
and distribution utility (DU) according to the availability of power in each system.
The user information is stored in the ThingSpeak cloud (TC) [15] using IoT devices.
When this system is connected to the community (society), the information about
the energy consumption and the number of energy units consumed from the DU by
each house in the community is sent to the same cloud but from a separate hardware
system. The GUI reads information from cloud on which the data about the house
power are stored and displays the respective values. It will also be able to calculate
the total amount to be paid by the house owner until the current time. The Web-based
application is secured with identity access management (IAM) system and two-factor
authentication [14, 16]. The rest of the paper is organized as follows: Sect. 2 presents
the architecture for GUI design and development. Section 3 describes the security
architecture and its implementation. Section 4 presents the result/discussion on the
implementation. Finally, the paper concludes with Sect. 5.
GUI-Based Secure Architecture Design for Distributed … 687
2 Architecture for GUI Design and Development
This section elaborates on the basic model, design, and hierarchy of a secure Web-
based application. The framework is essentially flexible enough to fit in different
application scenarios. Web server-based access is viable for a larger user base and
remotely control ability or IoT/ FPGA (e.g., Raspberry Pi/Zed board)-based approach
where all the power management and transactions can be managed locally. In our
current work, we demonstrate our framework and features on an only Web-based
application. The GUI has real-time data like temperature, power, current, voltage,
and radiance which are (Fig. 1).
Graphically presented using MS Visual Studio (ASP.net). These data are being
collected by the sensors in real-time, and controllers push these data to particular
cloud storage (e.g., TC) server on a regular interval (N.B. displayed data on Web
GUI are a subset of data that are being collected by the storage server which is re-
configurable from utility only). By retrieving data from the cloud, the Web application
displays the most accurate percentage of power available in the SPV, community
battery, and the real-time status of the power. The program flow for developing the
Web application is as follows: (1) The first page of the application demands the user
with the email id, password, and the time-based one-time password (TOTP). If the
account does not exist, then it displays a popup menu saying that the account is
invalid than the user needs to do the registration. (2) The application allows the user
to create a new account by clicking the button. Do not have an account? Create new?
This page demands the name, email id, password, and TOTP; once the user submits
it, then these details get registered in the database. (3) If the account exists and user
successfully logins, then the data regarding power and voltage of the house(user) are
fetched from the ThingSpeak cloud and displayed. The information is then encrypted
and stored in the database by a suitable symmetric key. This GUI application has the
Fig. 1 Cloud-user data sharing framework

following security measures to counter any type of cyber-attack from any unwanted
elements. The developed GUI is a Web-based application that is secured through
HTTPS with the transport layer security (TLS) protocol for data transfer. My SQL is
used for the database as backend, whereas MS Visual Studio is used as a front end to
create the Web application. This model is working on identity access management
(IAM), which checks both the authentication and the authorization of the user before
providing access to the Web application and the data.
3 Security Architecture and Implementation
This section describes the security aspects of our proposed approach for DCM.
The first level of security can be incorporated by implementing TOTP on the user
login/registration portal. With two-factor authentication, the Web security enhances,
as the threat of password interception, a man-in-the-middle attack or eavesdrop-
ping drops dramatically. The fundamental structure of two-factor authentication is
described in Fig. 2 shown.
Fig. 2 Two-factor authentication

3.1 Data Sharing Security Using Robust Encryption

Algorithm
For building a secure database, it is imperative to encrypt the raw data with some
secure and robust encryption strategy. In this endeavor, we added the most commonly
used secure symmetric encryption scheme Advanced Encryption Standards (AES)-
128 bit [17]. For securing the local data server, symmetric encryption suffices whereas
for remote data storage and data processing, private–public scheme can be adopted.
The planned methodology is represented below, including the diverse components of
the system that is working to process the user data. The projected system represents
three major hierarchy units: distributed utility, community, and house. The user has to
share data over the open cloud or grid with no security for accessing data. Therefore,
in this paper, we have implemented protected data sharing methods on the micro-grid
that can be executed for three entities which are as follows:
1. Distributed Utility (DU) Portal
2. Community Portal
3. House (User).
Portal: These entities perform different tasks and respond effectively to their
assigned responsibilities. DU Portal: The distributed utility (DU) is the power distri-
bution unit that regulates the operations among various communities. The DU portal
acts as an administrator that can manage and grant access to data to various commu-
nities and the user. It is also responsible for simultaneously checking the reliable and
systematic functioning of all the community and house portals.
Community Portal: The community is freelance to the user portal. All user activ-
ities are handled by the community, which means that the users are uploading or
sharing their files on the cloud and have access rights given by the community.
During this panel, the community can approve or decline the user requests. If the
community granted permission to the user, then only the user accesses their account
and uploads or shares the data. All files stored by the user can be viewed by the
community.
House Portal: The house is the end-user who accesses the data on the cloud. A
user is the only affected one whose sensitive data are stored on the cloud; hence,
ensuring the safety of the user’s data is a must as at the user side, and many actions
need to be performed simultaneously like uploading, downloading, and updating of
information.
All entities, including DU, community, and the house(user), need to verify the
log in via user name/login Id, password, and TOTP, to get access to their respective
account. As the authentication is successful, then it is redirected to the home page
wherein the user can do various data operations. The security architecture for multi-
level encryption process is demonstrated in Fig. 3.
Figure 3 illustrates the hierarchy of encryption applied for encrypting the data of
the users. As and when the user uploads a file to the community, it gets encrypted using
a symmetric key, which is shared among users by the approval of the community.
Fig. 3 Data encryption model
Similarly, when the community uploads any data to distribution utility or cloud, it
gets encrypted again, using another symmetric key. In this way, the user data are
encrypted twice before getting saved to the cloud. Access to the keys also has the
same hierarchy. The house portal has only first level key; with this key, house can
encrypt its data. Then, community portal has two pairs of keys, both house and
community; with these pairs of keys, community can only encrypt data of the houses
registered and the same structure for distributed utility who have a super-set of keys
(three-layer keys) and have control over all the community registered under it. In the
same way, for decryption, the community and the users have their symmetric keys.
3.2 ThingSpeak Cloud Platform
In this paper, we have incorporated the ThingSpeak platform as it is an IoT analytic

platform that allows us to envisage, aggregate, and examine the live data streams in
the cloud. TC provides instant visualizations of data posted by IoT devices to TC
as it is often used for prototyping and proof of concept in IoT systems that require
analytic. Since it is used in both cross-platform, it is quite reliable. Here, we used
TC with Microsoft Visual Studio as a code editor. The system architecture consists
of the home page that contains a channel that displays the graphs containing various
values. The GUI is designed to take updated values from all sensors through the TC.
4 Implementation of GUI-Based Secure Architecture

Design for DCM
In this paper, a Web-based GUI has been developed for DCM. A scaled-down model
of DCM is designed to validate the working of the GUI platform. A sensor module
is connected to the individual household to find the desired load demand of the
respective house. The load demand of an individual house is then transferred to the
TC. Another IoT sensor module is connected to a battery, which is considered as the
community storage. An 8-channel relay is used for switching the power supply as per
the control action of the SHEMS. A set of four houses has been considered for the
present work. The setup of the rest houses is the same as the first one. The community
controller is common for the four houses, which decides the power trading among
the four houses. The detailed description of the experimental setup, including the
working of the SHEMS, community controller, etc., is beyond the scope of this paper.
For the illustration, two different graphs of user data retrieved from the ThingSpeak
platform are presented.
The graphs explain the variation of the required power for a user with variation
in time. By using the graph, we can find the values of the dynamic power demand
of the house. Figure 4 gives the value of the state of charge of battery storage of
community regulator.
The Web-based secure application is developed using Microsoft Visual Studio
and C#. It is tested for its reliability while working with many users. The user data
faces many security challenges before getting saved in the cloud. For securing user
identity and user data, firstly, an HTTPS-based Web application is developed that
uses transport layer security (TLS) for Web security. The screenshot of the Web
application is presented in Figs. 5, 6, 7 and 8. Secondly, an OTP based authentication
framework is used, which provides access to the user after checking its identity and
authorization (IAM). Therefore, authorized and authenticated users can only log in
to the environment using two-factor authentication.
Fig. 4 Related IoT graphs using ThingSpeak

Fig. 5 Home page
Fig. 6 Login/registration
page
For data flow and its security, an AES encryption [17] is used, which has consis-
tency and scalability and is easy to implement. A Java-based application is developed
to secure the user data by continuously monitoring the data at the server. When the
migration of data took place or user uploads any data through an application over the
cloud, the user data are then automatically encrypted Fig. 11 without user interven-
tion using the AES-128-bit algorithm and sent to the respective cloud environment.
Fig. 7 First time user registration page
Fig. 8 OTP authentication

page
If there are any requirements to read the data, it first is decrypted on the user’s end.
The plain text is not saved, written, or stored anywhere in the database or on to the
cloud. The key is also not saved on the cloud since the key might get compromised.
To store or save the keys, separate key management schemes are used. The results
are shown in Figs. 9, 10 and 11.
Fig. 9 Encryption at cloud

(server)
Fig. 10 Encryption at house

portal (user)
Fig. 11 Auto checking of non-encrypted files (cloud)
5 Conclusion
In this paper, a Web-based GUI has been developed for distributed community micro-
grid. Although the user data faces many security challenges before getting saved in
the cloud/grid, for securing user identity and user data, firstly, an HTTPS-based Web
application is developed that uses transport layer security (TLS) for Web security.
Secondly, an OTP-based authentication framework is used, which provides access
to the user after checking its identity and authorization (IAM). Therefore, autho-
rized and authenticated users can only log in to the environment using two-factor
authentication. For data flow and its security, an AES encryption is used that has
consistency and scalability and is easy to implement. AES encryption algorithm
shows immunity against various kinds of cyber-attacks such as middle man attack,
square attack, key, and critical retrieval attack. The user data are shared using multiple
encryptions before it gets uploaded to the cloud. The system introduced in this paper
provides decisions-based real-time data convention, thus eliminating manual data
collection using TC. The Web-based GUI is developed for implementing more func-
tionality in data management, visualization, and analysis to provide better interface
and experience to the user.
References
1. Tiwari P, Manas M, Jan P, Nemec Z, Radovan D, Mahanta P, Trivedi G (2017) A review on

microgrid based on hybrid renewable energy sources in south-Asian perspective. Technol Econ
Smart Grids Sustain Energy
2. Anvari-Moghaddam A, Monsef H, Rahimi-Kian A (2016) Optimal smart home energy manage-
ment considering energy saving and a comfortable lifestyle. In: 2016 IEEE power and energy
society general meeting (PESGM), July 2016
3. Lotfi H, Khodaei A (2017) Ac versus dc microgrid planning. IEEE Trans Smart Grid
4. Wu L, Ortmeyer T, Li J (2016) The community microgrid distribution system of the future.
Electr J 29(10)
5. Meena NK, Yang J, Zacharis E (2019) Optimisation framework for the design and operation
of open-market urban and remote community microgrids. Appl Energy 252:113399
6. Basit A, Sidhu GAS, Mahmood A, Gao F (2017) Efficient, and autonomous energy management
techniques for the future smart homes. IEEE Trans Smart Grid 8(2):917–926
7. Javaid N, Ullah I, Akbar M, Iqbal Z, Khan FA, Alrajeh N, Alabed MS (2017) An intelligent
load management system with renewable energy integration for smart homes. IEEE Access
5:13587–13600
8. Safari A, Fotuhi-Firuzabad M, Lehtonen M (2016) Optimal residential load management in
smart grids: a decentralized framework. IEEE Trans Smart Grid 7(4):1836–1845
9. Ramya AS, Periyasamy N, Vidhya A, Sundrapandian S, Raja P (2014) Design and development
of a GUI for an optimal hybrid energy system. In: 2014 Eighteenth national power systems
conference (NPSC) pp 1–6
10. Vallerio KS, Zhong L, Jha NK (2006) Energy-efficient graphical user interface design. IEEE
Trans Mobile Comput 5(7):846–859
11. Viljoen M, Pienaar C (2013) Graphical user interface control for a solar/battery/hydrogen/fuel
cell power plant. In: 2013 Africon, pp 1–5
12. Katole KR, Balwir SP, Badole M (2015) GUI based power management system with power
line communication for industries. In: 2015 IEEE international advance computing conference
(IACC), pp 1040–1043
13. Malviya S, Dave S (2018) Secure data sharing scheme using cryptographic algorithm for cloud
storage. Int J Appl Eng Res 13(20):14799–14805
14. Patil A, Zambare K, Yadav P, Wasulkar P, Kimmatkar N (2015) Integration of encryption of
file and one time password for secure file accession cloud. Int J Ad Comput Sci Cloud Comput
3(1):89–94
15. Mhatre L, Rai N (2017) Integration between wireless sensor and cloud. In: International
conference on I-SMAC (IoT in social, mobile, analytics and cloud) (I-SMAC), pp 779–782
16. Aloul F, Zahidi S, El-Hajj W (2009) Two factor authentication using mobile phones. In:
IEEE/ACS international conference on computer systems and applications, pp 641–644
17. Soliman SM, Magdy B, Abd El Ghany MA (2016) Efficient implementation of the AES algo-
rithm for security applications. In: IEEE international system-on-chip conference (SOCC), pp
206–210
A Hybrid Protocol for Stop and Wait
ARQ in Markov Two States Model
Sanjit Ningthoujam and Swarnendu K. Chakraborty
Abstract In this paper, a hybrid protocol is presented for improving the performance
of Stop and Wait (S/W) ARQ with Aggressive Packet Combining (APC) schemes
using Markov two states Model. APC protocols are not onlymimplemented to get
a correct copy but also to reduce retransmission of duplicate copies. To reduce the
retransmission of duplicate copies in conventional ARQ protocols, Packet Combining
(PC) and APC have been studied. Thus, the proposed protocol is considered under
two states of Markov model i.e., good and bad states channel. The simulation results
have shown that in the long period of bad state that the proposed protocol still
obtains better in probability of packet error, mean energy consumption, and packet
throughput over PC and APC schemes but lower throughput than S/W ARQ.
Keywords PC · APC · Throughput · ARQ · Energy
1 Introduction
Backward Error Correction (BEC) and Forward Error Correction (FEC) are used to
transfer data reliably from transmitter (T x ) to receiver (Rx ). It is well studied that
BEC is good for short distance communication that is for wired network commu-
nication. Thus, many researchers have conclusively established that BEC should be
attempted for wireless communication as it is cost effective. ARQ is the example of
BEC in which erroneous copies are discarded by the receiver but by retransmission
of duplicate copies, receiver uses to get correct copy. To reduce retransmission of
duplicate copies, PC [1, 2] and APC [3] are studied in literature. PC provides 50%
throughput which is greater than APC but lower error correction capability. However,
in long distance communication, FEC strategy is used but it consumes huge amount
S. Ningthoujam (B)
School of Computer Science and Engineering, VIT University, Amaravati, Andhra Pradesh, India
S. K. Chakraborty
Department of Computer Science and Engineering, NIT Arunachal Pradesh, Papum Pare, India
698 S. Ningthoujam and S. K. Chakraborty
of energy for error correction code. Many modifications like Multiple Route Packet
Combining [4], ARQs, and APC [5–17] have also studied for increasing throughput,
tackling various error syndromes and enhancing fast error correction. But, these
protocols do not address about energy efficiency. Therefore, to address stated above
problem of APC, PC, and modified protocols, we propose this protocol to improve
energy efficiency.
The paper is organized as follows. Reviews of conventional PC and APC are briefly
described in Sect. 2. Proposed protocol is discussed in Sect. 3. The simulation results
and conclusion are discussed in Sect. 4 and 5.
2 Related Works
2.1 Packet Combining Scheme (PC)
Let the first and second copies are erroneous and stored at Rx in order to perform
XOR operation to identify the erroneous bit location. The identified locations will
be applied bit inversion method by using brute force method. The example is shown
here:
Assume that original copy: “00001111”
• 1st erroneous copy: “10001111”
• 2nd erroneous copy: “00000111”
• 10001111
• 00000111
…………..
10001000 (erroneous bits are found at 1st and 5th bit positions from MSB)
• Bit inversion technique will be applied in 1st and 5th position from MSB (if 1
then 0 and vice versa)
• By this technique, original copy will be retrieved. The average number of brute
force application will be 2n − 1 in general if n bits are found as erroneous. However,
in PC technique, if two bits are identified in same position of two erroneous copies,
then PC technique will fail. The example is shown here:
Assume original copy: “00001111”
• 1st erroneous copy: “00000111”
• 2nd erroneous copy: “00000111”
• 00000111
• 00000111
………….
00000000.
A Hybrid Protocol for Stop and Wait ARQ … 699
2.2 Aggressive Packet Combining Scheme (APC)
In APC scheme, we have divided into four sections.

• 1st section: assume original copy is “11110000”
• 1st copy—“11110000”
• 2nd copy—“11110000”
• 3rd copy—“11110000”.
In this case, all three copies are error free. Correction probability will be (1 − P3 ),
where P is bit error probability. Therefore, correction is not required in this section.
• 2nd section: assume original copy is “11110000”
• 1st copy—“01110000”
• 2nd copy—“11110000”
• 3rd copy—“11110000”.
In this case, 1st copy is erroneous and 2nd and 3rd copies are error free. Correction
probability will be (1 − P2 ) × P, where P is bit error probability. Therefore, correction
is required and also possible.
• 3rd and 4th sections: assume original copy is “11110000”
• 1st copy—“01110000”
• 2nd copy—“01110000”
• 3rd copy—“11110000” in 3rd section and “01110000” in 4th section.
In these cases, correction probability will be given by (1 − P) × P2 and P3 in 3rd
and 4th section. Corrections are not possible in both sections.
3 Proposed Protocol
We assume that the proposed model will be Markov two states Model, good (G) and
bad (B) states. The length of the good and bad period will be observed by transmitter.
By independent geometrically distributed of random variables with parameter X and
Y. It is also assumed that time required to transfer a packet of ‘n’ bit size will be
having ‘m’ slots. Therefore, probability of good and bad period for ‘n’ packet size
which contain ‘m’ slots will be given by Eqs. (1) and (2) (Fig. 1).
Prob(G) = (1 − X )X m (1)
Prob(B) = (1 − Y )Y m (2)
So, the mean length of G and B periods will be given by Eqs. (3) and (4)
α = 1/(1 − X ) (3)
Fig. 1 Markov two states model
β = 1/(1 − Y ) (4)
/* Proposed protocol */
Input: Pi : the incoming packets, where i = 1, 2, 3…j
Output: correct or error packet
1. If (P(i=first packet) = correct copy) Then
2. Current protocol (C.P) = stop and wait ARQ;
a. If (test) = correct copy then
b. C.P = stop and wait ARQ;
c. Else
d. C.P = APC protocol;
e. End
3. C.P = APC protocol;
a. If (test) = correct copy then
b. C.P = stop and wait ARQ
c. Else
d. C.P = APC protocol
e. End
4. End
3.1 Mathematical Analysis
Let ‘μ’ be the bit error rate probability in wireless. Let single bit ‘1’ is sent from T x to
Rx and it is received erroneously as ‘0’. Then, single bit error correction probability
is given by Eq. (5).
Pec = 1 − μ (5)
The error correction probability for the n packet size can be given by Eq. (6)
Pec_n = (1 − μ)n (6)
Therefore, the probability of packet error for n bit packet size is given by Eq. (7)

Perr_n = 1 − (1 − μ)n (7)
It is assumed that energy (E) consumption for one packet of transmission from T x
to Rx is unit. The energy consumption for transmission of packet is attained higher
when retransmission of duplicate copies is occurred. Therefore, the mean energy (E)
consumption for proposed protocol will be given by Eq. (8).
1
E= (8)
[(1 − μ)n×i ]
where i is the number of duplicate copies.

By Markov two states model, total mean length of good and bad states will always
be unity.
Therefore, α + β = 1 (9)
In Stop and Wait ARQ, PC, and APC [18–21], the probability of packet error will
be given by Eqs. (10)–(12)
PEs/w = [1 − (1 − μ)n (10)
2
PEpc = 1 − (1 − μ)n (11)
3
PEapc = 1 − (1 − μ)n (12)
For the proposed protocol, the probability of packet error will be given by Eq. (13)
PEPro = PEs/w × PEapc (13)
Throughput (THpro ) of the proposed protocol will be given by Eq. (14)
THpro = THs/w × α + THapc × β (14)
where THs/w = (1 − PEs/w )/(i + PEs/w ), THapc = (1 − PEapc )/(i + PEapc )

Fig. 2 Mean energy consumption (α = 0.3, β = 0.7)
The proposed protocol is simulated using MATLAB. The probability of mean length
in Bad (B) and Good (G) channel is taken as unity. Therefore, the mean length of ‘G’,
having short periods α = 0.3 and mean length of ‘B’, β = 0.7 having long periods
have taken in simulation. The simulation results have shown in Figs. (2, 3 and 4).
In the results, the proposed protocol has shown higher throughput than PC and APC
but lower to S/W ARQ. Meanwhile, it also provides lower energy consumption by
lowering the probability of packet error when bit error rate varies from 10−2 to 10−4
and packet size of 4000 bits.
5 Conclusion
In this paper, we have studied a new hybrid protocol (APC + S/W ARQ) for
enhancing the performance of S/W ARQ using Markov two states model. The simu-
lation results have shown that the proposed protocol obtains better performance in
terms of mean energy consumption and probability of packet error. This protocol is
attractive for wireless network as the bit error rate is very high. The transmitter will be
switched from proposed protocol to S/W ARQ and vice versa as per the mean length
Fig. 3 Probability of packet error (α = 0.3, β = 0.7)
of ‘G’ and ‘B’ periods. The proposed protocol is found a bit lower throughput than
S/W ARQ but greater than PC and APC when long periods of ‘B’ and short period of
‘G’ states. Thus, little difference of throughput may not degrade the performance of
proposed protocol due to lower energy consumption and lower probability of packet
error.
Fig. 4 Throughput (α = 0.3, β = 0.7)
References
1. Chakraborty SS et al (1995) An ARQ scheme with packet combining. IEEE Comm Lett
2(7):200–202
2. Chakraborty SS (1999) An adaptive ARQ scheme with packet combining for time varying
channels. IEEE Commun Lett 3(2):52–54
3. Leung Y-W (2000) Aggressive packet combining for error control in wireless networks. Trans
Commun E83(2):38–385)
4. Hirayama Y, Okada H, Yamazato T, Katayama M (2005) Time-dependent analysis of the
multiple-route packet combining scheme in wireless multihop network Int J Wirel Inf Netw
42(1):35–44
5. Bhunia CT (2001) ARQ-review and modifications. J IETE Tech Rev 18:381–401
6. Sastry ARK (1974) Improving automatic repeat request (ARQ) performance on satellite
channels under high error rate conditions. IEEE Trans Commun COM-23:436–439
7. Chakraborty SK. Goswami RS, Bhunia A, Bhunia CT (2014) Investigation of two new protocols
of aggressive packet combining in achieving better throughput. J Inst Eng (India): Ser B 15
8. Chakraborty SK. Goswami RS, Bhunia A, Bhunia CT (2015) Studies of several new modifica-
tions of aggressive packet combining to achieve higher throughput, based on error correction
capability of disjoint error vectors. J Inst Eng (India): Ser B 30
9. Bhunia CT (2007) Packet reversed packet combining scheme. In: Proceedings of IEEE
Computer Society, CIT’07. Aizu University, Japan, pp 447–451
10. Ningthoujam S, Chakraborty SK, Bhunia CT (2014) New modified technique of aggressive
packet combining scheme with multiple routes selection to get high error correction and
throughput. Int J Ad Electron Commun Syst 3(5)
11. Towsley Don (1979) The shutter go back-N ARQ protocol. IEEE Trans Commun 27(6):869–
875
12. Wicker SB (1999) Adaptive rate error control through the use of diverse combining and majority
logic decoding in hybrid ARQ protocol. IEEE Trans Comm. 39(3):380–385
13. Chakraborty SK, Goswami RS, Bhunia A, Bhunia CT (2014) New protocol of aggressive
packet combining in gilbert two state model. IEEE, CIEC, University of Calcutta
14. Chakraborty SK, Goswami RS, Bhunia A, Bhunia CT (2013) Three new investigations of
aggressive packet combining to get high throughput. Int J Comput Appl USA (0975–8887)
81(5):41–44
15. Chakraborty SK, Goswami RS, Bhunia A, Bhunia CT (2013) Two new modified schemes of
aggressive packet combining schemes in achieving better throughput. In: IEEE, ITNG. IEEE
Computer Society, Las Vegas, pp 729–730
16. Ningthoujam S, Chakraborty SK. Bhunia CT (2015) New protocol for aggressive packet
combining in gilbert two state model using back up routes to achieve higher throughput. In:
Proceedings of ICARCSET-2015, ACM, ICPS, March, 6th–7th, Eluru, Andhra Pradesh, India
17. Ningthoujam S, Chakraborty SK, Bhunia CT (2015) New investigation of aggressive packet
combining scheme to reduce transmission delay and three states markov model using multiple
routes to increase throughput. Int J Future Gener Commun Netw 8(5):157–166
18. Ningthoujam S, Chakraborty SK et al (2017) PC-APC schemes in multipath diversity system
to get higher throughput. Int J Electr Comput Eng 7(1):337–343
19. Bhunia CT (2005) Information technology, networks and internet. New Age International
Publishers, New Delhi
20. Ningthoujam S, Chakraborty SK (2019) Analysis of the adaptive three-modes for PC + MPC +
APC techniques using retransmission cycle mechanism. J High Speed Networks 25(2):205–220
(IOS Press)
21. Ningthoujam S, Chakraborty SK (2019) A combined techniques of PC + MPC + APC to
achieve higher error correction probability and throughput over APC and MPC techniques.
Wirel Pers Commun 109(3):1–17
Comparative Study on a U-Slot
Miniaturized CPW-Fed Multi-band
Antenna Applicable for 5G
Communication
Kaushal Mukherjee, Amitesh Das, and Sahadev Roy
Abstract In this paper, a compact (17 × 20 × 1) mm3 coplanar waveguide-fed

multi-band antenna has been designed. This antenna consists of a U-Slot over the
rectangular patch structure and frequency shifting strips. Various U-Slot structures
have been studied in this work for the analysis of the antenna performances. A
comparison table between the proposed antenna and similar antennas has been given
in the result section. The proposed antenna resonates at 28 GHz, with a bandwidth
of 2.62 GHz. Design and simulation of structures are carried out with the help of
ANSYS HFSS software and other antenna parameters like field distributions, gain,
directivity, impedance bandwidth, resonant frequencies, return loss, and VSWR are
also reported in this study.
Keywords Coplanar waveguide (CPW) · Millimeter wave (mmW) · Reflection

coefficient (S11) · U-Slot · 5G antenna
1 Introduction
5G is the latest arena of wireless communication standard ahead of today’s 4G and

4.5G (LTE advanced) technology. This technology will facilitate users by improving
the latency, speed, battery life, and power consumption of the mobile devices. As
statistics show that the number of mobile devices in all over the world is expected to
cross the 5 billion mark by 2019, so researchers have always looked one step forward
beyond today’s technology [1]. 5G technology (Fifth generation) will make a massive
K. Mukherjee
Department of Physics, The University of Burdwan, Burdwan, India
A. Das (B) · S. Roy
Department of Electronics and Communication Engineering, National Institute of Technology,
Arunachal Pradesh, Yupia, India
S. Roy
difference with the existing 4G LTE (Long-Term Evolution) technology and will give
birth to a Wireless World Wide Web (WWWW) [2]. The major advantages of 5G
over existing 4G LTE communication system are wider bandwidth, the finest quality
of service, and a data transfer rate of more than 1 Gbps (Gigabits Per Second), etc.
[3]. Shannon in his channel capacity theorem explains that data rate is equal to the
bandwidth (BW) and the signal-to-noise ratio (SNR) value of the used modulation
technique [4].
Some techniques have been suggested by researchers to increase the BW of an
antenna for the 5G applications are: by increasing the substrate height [5], by using
different size and shapes of ground planes (such as elliptical, circular, etc.) [6],
by introducing partial ground planes [7], by the application of dielectric resonator
antenna (DRA) [8], or by using the photonic band gap structures [9], etc. Recent
researches have given some views on 5G antenna which are as follows: Ali et al.
proposed a broadband printed slot antenna applicable for 5G communication [10],
Zhang et al. proposed a DRA antenna based on MIMO technology for 5G appli-
cations [11], and Yassin et al. proposed a single fed antenna for both 4G and 5G
communication [12].
In the present work, the slot technique has been used to develop an antenna for 5G
applications. Further, a comparative study on the antenna performances for various
sizes and shapes of U-Slot has also been studied here.
Section 2 describes the geometry of the antenna, Sect. 3 describes the simulation
result and analysis, and Sect. 4 concludes this work.
2 Antenna Geometry and Parametric Analysis
The concept of bandwidth increment by the application of U-Slot [13] has been used
here. Lee et al. proved that a wider bandwidth can be achieved by the application of
a U-Slot and exciting the structure with co-axial feeding [14]. In his work, variation
in bandwidth and matching in impedance frequency were obtained by tuning the
probe position. In our proposed antenna, coplanar waveguide feeding has been used
to excite the structure (as co-axial feeding is not suitable for mobile communication)
[15] and instead of tuning the probe position, the sizes and shapes of U-slot have been
varied to achieve the wider bandwidth. The basic structure of the reference antenna
without strip and with strips are shown in Figs. 1 and 2, respectively.
The antenna without strip and U-Slot (Fig. 1) generates two resonant frequencies
at 1.72 GHz and 5.37 GHz which are applicable for GSM and WLAN with impedance
bandwidths of 0.25 GHz and 0.85 GHz, respectively. These narrow bandwidths do
not satisfy the minimum BW requirements of 5G communication as 5G requires BW
of almost 1000 times greater than the BW of today’s 4G technology.
Comparative Study on a U-Slot Miniaturized CPW-Fed Multi-band … 709
Copper Dimension (mm)

part
No. X Y Z
1 6.00 6.78 0.02
2 3.67 2.43 0.02
3 2.50 4.00 0.02
4 1.55 1.04 0.02
5 17.0 4.25 0.02
6 1.45 3.44 0.02
7 17.0 0.96 0.02
8 1.45 3.44 0.02
9 1.3125 5.04 0.02
10 1.3125 6.04 0.02
11 8.375 7.04 0.02
12 0.50 7.75 0.02
Fig. 1 Top view of antenna without strip
Fig. 2 Top view of antenna with additional strip (all dimensions are in mm)
2.1 Effect of Strip
The antenna without strip and U-Slot (Fig. 1) generates two resonant frequencies at
1.72 GHz and 5.37 GHz with impedance bandwidths of 0.25 GHz and 0.85 GHz,
respectively.
The frequency shifting strip (FSS) is responsible for the generation of multiple
bands and shifting of the frequency bands to the desired frequency range (Fig. 2). The
strip lines can shift the resonant frequency either toward the upper frequency band or
toward the lower frequency band, but it is not capable to increase the bandwidth of
the antenna. So, the proposed antenna has been modified by the etching of a U-Slot
over the patch.
2.2 Effect of U-Slot
This subsection describes the antenna performances due to the variations in size and
shape of the U-Slot. As 5G communication always requires a wider bandwidth, so
to increase the bandwidth of the proposed antenna, a U-Slot has been etched over
the patch. These variations of U-Slot differ the antenna performances in terms of
the amount of fringing field generation, the number of frequency band generation,
etc. [16]. In the absence of U-Slot, the antenna will generate resonate frequencies
within the desired bands, but does not satisfy the bandwidth requirements for 5G
communication. By the application of U-Slot over the patch and by tuning the size
and shape of the U-Slot, a wider bandwidth of 4.54 GHz is achieved in this work. In
the absence of the U-Slot, maximum bandwidth of 2.62 GHz can be achieved.
2.3 Combined Effects of Strips and U-Slot
The combined effect of strips and U-Slot are explained here. These strips are respon-
sible for the generation of multiple bands and shifting of the frequency bands to the
desired frequency range. The strip line can shift the resonant frequency either toward
the upper frequency band or toward the lower frequency band, but is not capable to
increase the bandwidth of the antenna. So, the proposed antenna has been modified
by etching of a U-Slot over the patch. In the presence of both the strips and U-Slot,
the proposed antenna (Fig. 3) generates a wider bandwidth of 4.54 GHz. So, strips are
used to shift the frequency range of the desired frequency band, and the combination
of U-Slot and strip are used to achieve a maximum amount of bandwidth within the
desired frequency band. The proposed antenna covers two 5G frequency bands, i.e.,
28 GHz and 37 GHz (as announced by FCC) and also applicable for Wi-Fi.
Fig. 3 Top view of proposed antenna
The antenna without strip and U-Slot (Fig. 4a) generates two resonant frequencies
at 1.72 GHz and 5.37 GHz with impedance bandwidths of 0.25 GHz and 0.85 GHz,
respectively.
The antenna with strips and without U-Slot (Fig. 4b) generates operating bands at
1.702 GHz and 5.802 GHz with reflection coefficients of −14.51 dB and −26.19 dB,
respectively. Impedance bandwidth obtained from this antenna are of 0.57 GHz
and 0.98 GHz. The antenna with U-Slot in reverse direction (Fig. 4c) generates
resonant frequencies at 4.82 GHz, 12.58 GHz, 15.02 GHz, 18.34 GHz, 21.28 GHz,
27.12 GHz, 33.52 GHz, and 38.30 GHz with reflection coefficients of −16.93 dB,
−16.46 dB, −16.25 dB, −19.54 dB, −22.37 dB, −15.61 dB, −16.87 dB, and −
22.74 dB. Impedance bandwidths of the generated resonant frequencies are 0.32 GHz,
0.54 GHz, 0.82 GHz, 0.98 GHz, 1.08 GHz, 3.42 GHz, 3.06 GHz, and 2.36 GHz,
respectively.
Antenna with narrow U-Slot (Fig. 4d) generates resonant frequencies at 5.18 GHz,
17.88 GHz, 20.58 GHz, 27.22 GHz, 33.74 GHz, and 38.40 GHz with reflection coef-
ficients of −13.17 dB, −15.65 dB, −10.88 dB, −13.57 dB, −14.56 dB, and −
11.22 dB respectively. Impedance bandwidths of the generated resonant frequencies
are 0.10 GHz, 0.82 GHz, 1.66 GHz, 3.50 GHz, 3.72 GHz, and 0.9 GHz, respectively.
The antenna is further investigated with Dual U-Slot (Fig. 4e), generates resonant
Fig. 4 Progress toward the proposed antenna a without strip and U-Slot, b with strip without
U-Slot, c reverse U-Slot, d narrow U-Slot, e dual U-Slot, f proposed antenna, and g fabricated
antenna
frequencies at 5.18 GHz, 16.46 GHz, 18.30 GHz, and 25.96 GHz with reflection coef-
ficients of −13.10, −18.78 dB,−36.92 dB, and −13.69 dB, respectively. Impedance
bandwidths of the generated resonant frequencies are 0.12 GHz, 0.82 GHz, 1.66 GHz,
and 3.49 GHz, respectively.
Proposed antenna (Fig. 4f) with single U-Slot generates resonant frequencies at
5.0 GHz, 17.62 GHz, 20.44 GHz, 28 GHz, and 37 GHz with reflection coefficients
of −10.02 dB, −23.59 dB, −28.20 dB, −21.01 dB, and −29.08 dB, respectively.
Impedance bandwidths are 0.02 GHz, 1.4 GHz, 1.28 GHz, 2.62 GHz, and 4.54 GHz,
respectively.
3 Result Analysis and Discussion
By analyzing the design variations, it is observed that maximum bandwidth of

4.54 GHz with a reflection coefficient of −29.08 dBcan be obtained from the
proposed antenna with single U-Slot which is having length = 1 mm, width =
3 mm, and height = 1 mm as dimension as shown in Fig. 4f. If the length of patch
increases, then the frequency will shift toward lower frequency band, and if the length
of the patch decreases, then the frequency will shift toward the upper frequency band
[17]. The mmW frequency band offers a wider bandwidth to the users which allow
Fig. 5 Reflection coefficient plot by U-Slot variation
the users to connect their home appliances by using IoT and also provides faster
transfer of data rate. The proposed antenna not only provides higher speed to the
mobile users but also will be used in the fields such as medical, agricultural, and
many others. This technology will facilitate people of urban and semi-urban areas
by improving the traffic control system with real-time data processing.
The design theory involves the concept of three neighboring resonances, one of
them is from TM01 mode of the patch, another one from the U-Slot, and the last one
is from the strips. Using two neighboring resonance, an increment in the BW up to
30% is achievable. If a third resonance appears (that will appear if co-axial feeding
is used instead of CPW feeding), then BW can be increased up to 40%.
In the presence of three resonances, BW is increased up to 42%. Achievable band-
width with only one resonance is about 0.85 GHz, when two neighboring resonances
appear, then the bandwidth is incremented up to 31% (2.62 GHz), and with three
resonances, bandwidth is raised up to 42% (4.54 GHz) with a VSWR value within 1
to 2 (Fig. 5). Figure 6 represents the reflection coefficient plot which compares the
simulated result with measured result.
The comparisons between proposed antenna and few relevant antennas have been
summarized and presented in Tables 1 and 2.
As shown in Fig. 7, radiation patterns of the proposed antenna illustrate that
far-field distribution and bandwidth are inversely related to each other. With the
increment in bandwidth, the area covered by fields (both E and H) reduces linearly.
Surface current distribution of the proposed antenna at 28 GHz in Fig. 8a shows
the charge accumulation and dissipation points. The far-field three-dimensional gain
plot is presented in Fig. 8b with a maximum gain of −30 dB.
Fig. 6 Reflection coefficient plot of simulated and measured result
4 Conclusion
A miniaturized coplanar waveguide-ed multi-band antenna applicable for 5G

communication is presented in this work. A comparative study of various sizes and
shapes of U-Slot has also been reported here. The proposed antenna operates at 28
and 37 GHz of the 5G frequency band with a wider bandwidth (up to 4.54 GHz).
An excellent far-field radiation pattern with a satisfactory amount of gain of 5 dBi
is achieved. For better performances, the substrate material can be replaced by RT
Duriod (dielectric constant of 2.2), but it will increase the cost of the antenna. The
concept U-Slot has been used over the dual-band antenna to achieve a wider band-
width up to near about 4.6 GHz. Again, mobile integration requires an antenna with
compound semiconductor (such as GaAs) as a substrate material, which will also
increase the cost of the antenna. To resolve these cost issues, the proposed antenna
can be applicable in the cell towers also, where an antenna with FR4 as a substrate
material, will overcome those problems. The proposed antenna is appropriate for
both cellular and wireless applications.
Table 1 Analysis of performances between few relevant antennas

Antenna Dimension Frequency band Center Bandwidth at Applications
(mm)3 at –10 dB (GHz) freq. –10 dB (GHz)
(GHz)
Ref. [7] —– 01.60–01.85 01.72 00.25 GSM
04.95–05.80 05.37 00.85 Wi-Fi/WLAN
Antenna S1 = 0.50 × 01.51–02.06 1.702 00.55 GSM1800/1900
with FSS 6.25 × 0.02 04.43–06.70 5.802 02.27 Wi-Fi/WLAN
only S2 = 4.175 × 5.2, 5.5 &
0.5 × 0.02 5.8 GHz
S3 = 0.50 ×
6.25 × 0.02
S4 = 1.50 ×
0.50 × 0.02
S5 = 0.50 ×
1.50 × 0.02
S6 = 0.50 ×
2.50 × 0.02
Proposed FSS 04.98–05.02 05.00 00.02 Wi-Fi
antenna S1 = 0.50 × 16.76–18.16 17.62 01.40 5G
6.25 × 0.02
S2 = 4.175 × 19.90–21.18 20.44 01.28
0.5 × 0.02 26.74–29.36 28.00 02.62
S3 = 0.50 × 34.26-38.80 37.00 04.54
6.25 × 0.02
S4 = 1.50 ×
0.50 × 0.02
S5 = 0.50 ×
1.50 × 0.02
S6 = 0.50 ×
2.50 × 0.02
U-Slot
A = 1.0 × 3.0
× 1.0
B = 5.0 × 1.0
× 1.0
C = 1.0 × 3.0
× 1.0
Table 2 Comparison between proposed antenna and few relevant antennas

Antenna Substrate Frequency Center BW at − Application VSWR
dimension band at − freq. 10 dB
(mm) 10 dB (GHz) (GHz) (GHz)
[7] 17 × 20 × 1.51–2.06 1.702 0.55 GSM1800/ 1.05
1 1900
4.43–6.70 5.802 2.27 Wi-Fi/WLAN 1.10
5.2, 5.5 & 5.8 GHz
[18] 6 × 6.2 × 37.03–38.97 38.00 1.94 5G 1.30
0.5 53–55 54.00 2.00 5G 1.64
[19] 11 × 8× 27.29–29.31 28.30 2.02 5G 1.53
0.5
Proposed 17 × 20 × 04.98–5.02 05.00 0.02 5G Data rate 1.95
antenna 1 16.76–18.16 17.62 1.40 06.24 Gbps 1.41
19.90–21.18 20.44 1.28
26.74–29.36 27.66 2.62 05.70 Gbps 1.08
33.26–37.80 35.92 4.54 11.68 Gbps 1.19
20.24 Gbps 1.07
Simulated E Field at 28 GHz Measured E Field at 28 GHz

(a) (b) (c)
Fig. 7 Radiation pattern plots: a simulated H Field at 28 GHz, b simulated E Field at 28 GHz, and
c measured E Field at 28 GHz
dB (Gain Total)
(a) (b)
Fig. 8 Electric field distribution a and far-field gain plot b of the proposed antenna at 28 GHz
References
1. Zhang S, Cai X, Zhou W, Wang Y (2018) Green 5 g enabling technologies: an overview. IET
Commun 13:135–143
2. Al-Turjman F (2019) 5G-enabled devices and smart-spaces in social-IoT: an overview. Future
Gen Comput Syst 92:732–744
3. Feng M, Mao S (2019) Dealing with limited backhaul capacity in millimeter-wave systems: a
deep reinforcement learning approach. IEEE Commun Maga 57:50–55
4. Paolo B, Stefano B, Giulio C, Andrea M, Rusek F, Alessandro U (2014) Modulation formats
and waveforms for 5G networks: who will be the heir of OFDM?: an overview of alternative
modulation schemes for improved spectral efficiency. IEEE Signal Process Maga 31:80–93
5. Colburn JS, Rahmat-Samii Y (1999) Patch antennas on externally perforated high dielectric
constant substrates. IEEE Trans on Antennas Propag 47:1785–1794
6. Soranun J, George K (2004) A class of shaped deficit round-robin (SDRR) schedulers.
Telecommun Syst 25:173–191
7. Koohestani M, Moghadasi MN, Virdee BS (2011) Miniature microstrip-fed ultra-wideband
printed monopole antenna with a partial ground plane structure. IET Microwaves Antennas
Propag 5:1683
8. Mashhadi SHH, Niaz MW, Jiao YC, Chen J (2019) Broadbeam coplanar-parasitic rectangular
dielectric resonator antenna. Progress Electromagnet Res 81:55–66
9. Alam MS, Islam MT, Misran N (2012) A novel compact split ring slotted electromag-
netic bandgap structure for microstrip patch antenna performance enhancement. Progress
Electromagnet Res 130:389–409
10. Abd El-Hameed AS, Barakat A, Abdel-Rahman AB, Allam A, Pokharel RK, Yoshitomi K
(2018) Broadband printed-dipole antenna for future 5G applications and wireless communica-
tion. In: IEEE radio and wireless symposium (RWS), pp 106–108
11. Zhang Y, Deng JY, Li MJ, Sun D, Guo LX (2019) A MIMO dielectric resonator antenna
with improved isolation for 5G mm-wave applications. IEEE Antennas Wirel Propag Lett
18:747–751
12. Mohamed ME, Abdallah HA, El-Hennawy EA, H. S. (2018). Single-fed 4G/5G multiband
2.4/5.5/28 GHz antenna. IET Microwaves Antennas Propag 13(3):286–290
13. Khodaei GF, Nourinia J, Ghobadi C (2008) A practical miniaturized U-slot patch antenna with
enhanced bandwidth. Prog Electromagnet Res 3:47–62
14. Chow YL, Chen ZN, Lee KF, Luk KM (1998) A design theory on broadband patch antennas
with slot. IEEE Antennas Propag Soc Int Sym 2:1124–1127
15. Chu H, Li P, Zhu XH, Hong H, Guo Y (2019) Bandwidth improvement of center-fed series
antenna array targeting for base stations in offshore 5G communications. IEEE Access 7:33537–
33543
16. Radavaram S, Pour M (2018) Wideband radiation reconfigurable microstrip patch antenna
loaded with two inverted U-slots. IEEE Trans Antennas Propag 67:1501–1508
17. Tighezza M, Rahim SKA, Islam MT (2018) Flexible wideband antenna for 5G applications.
Microwave Opt Tech Lett 60:38–44
18. Imran D, Farooqi MM, Khattak MI, Ullah Z, Khan MI, Khattak MA, Dar H (2018) Millimeter
wave microstrip patch antenna for 5G mobile communication. In: International conference on
engineering and emerging technologies, pp 1–6
19. Outerelo DA, Alejos AV, Sanchez MG, Isasa MV (2015) Microstrip antenna for 5G broadband
communications: Overview of design issues. In: IEEE international symposium on antennas
and propagation and USNC/URSI national radio science meeting, pp 2443–2444
An Experimental Investigation
of FPGA-Based LMS Algorithm
for Adaptive Noise Cancellation
Abinash Patnaik, Rajesh Kumar Patjoshi, and Rakhee Panigrahi
Abstract Desirable along with undesirable information are carried through signals.
Adaptive noise elimination is a substitute method for estimating signals spoiled
through additive interferences and noise. This process necessitates two inputs such
as primary input comprising polluted signal and reference input comprising desired
signal. The primary input is filtered adaptively and deducted from the desired signal
to attain error signal, and the adaptive algorithm performs in closed loop manner
to reduce the error signal. The adaptive noise cancellation has been designed via
two methodologies. One is model-centered methodology while another is FPGA-
based methodology. In case of model-centered methodology, MATLAB/Simulink is
employed, and the simulation results are observed through virtual oscilloscope. In
FPGA-based methodology, Quartus II IDE tool of Altera is utilized for hardware
implementation on Altera DE2 development board.
Keywords FPGA · Altera DE2 board · Matlab · Simulink · Adaptive filter · LMS
algorithm · Verilog design
1 Introduction
Usages of digital signal processing enforce significant checks upon power disrup-
tion, area, speediness, and price [1]. Accordingly, model tool should be thoroughly
selected. The most familiar tools for purpose of such applications are DSP, ASIC, and
FPGA. The DSP utilized for mathematical tasks cannot handle excessive sampling
A. Patnaik (B) · R. K. Patjoshi

Department of Electronics and Communication Engineering, National Institute of Science and
Technology, Berhampur, India
R. K. Patjoshi
R. Panigrahi
Department of Electrical Engineering, Parala Maharaja Engineering College, Berhampur, India
720 A. Patnaik et al.
rate appliances owing to its serial architecture. However, ASIC meets shortage of
flexibility and requires extended proposal cycle [2]. The FPGA can recompense the
shortcomings of DSP as well as ASIC [3]. Henceforth, FPGA has developed the
top platform for proposal of signal processing scheme owing toward their upper
bandwidth and better litheness through their parallel design [4]. The primary aim
of this paper is to explore usage of FPGA structure for adaptive noise cancellation
through model-based and FPGA-based approaches. Two approaches are considered
here, Model based and FPGA based. In model-built approach, MATLAB/Simulink
is employed to observe the different outputs as continuous verification of adaptive
noise cancellation system is achieved. In FPGA-based approach, there is a need of
developing Verilog code for LMS algorithm [5], FIR filter, I2C controller [6], ADC,
DAC, and mixed signal modules. These Verilog codes are compiled using Quartus
II IDE tool of Altera and generate the configuration file, which are dumped into the
FPGA to observe the operation [7]. The basic purpose of the paper is terminating the
noise from a corrupted or tainted signal and getting preferred signal as output.
The paper has been structured as given below. Here, Sect. 2 defines proposal
structure of the system. The system-level explanation is described in Sect. 3. Layout
for model- and FPGA-built design is provided in Sect. 4. Hardware implementation
process is described in this section. Conclusion and the forthcoming expansion of
the paper are prescribed in Sect. 5 followed by References.
2 Algorithm Description
Adaptive filters are computational tools that effort to do the modeling of the corre-
lation concerning dual signals occurring at actual time upon or using an iterative
method. These filters, which are basically compatibility tools, are accomplished as
an assembly of entailing program commands functioning on above of an arithmetical
handling appliance. These devices can be a microprocessor or a chip functioning DSP.
These devices can also be comprising of established logic operations, which has been
realized upon a field-programmable gate array (FPGA) or has been executed on a
VLSI integrated circuit [8]. Still, most important task of the above-mentioned filter
has been measured independently then of the particular physical attainment that it
receives [9, 10]. Because of this, the mathematical model of adaptive filter is decided
as contrasting to their particular accomplishments in software or as well as hardware.
An adaptive filter can be distinct through only four features:
1. The signals can be managed through filter.
2. The structure describes the way the output signal, that is coming out from the
computational tool, is calculated using its own input signal.
3. The affiliation among the input and output of the mentioned filter could be
changed, by altering the structure which consists different parameters. The
parameters are also very much accustomed from one particular time prompt
to another, which is designated by adaptive algorithm.
An Experimental Investigation of FPGA-Based LMS … 721
Fig. 1 Block diagram of

adaptive filter
Figure 1 displays a block diagram, where an individual sample, i.e., a digital

input signal M[n], is given to adaptive filter. This device calculates the resultant
output sample Z[n] at time n. The arrangement is not significant for adaptive filter,
excluding the point that it comprises modifiable factors whose amounts affect directly,
how Z[n] will be calculated. The output signal is related with another signal O[n],
which is known as desired signal, by deducting the two given sample signals at time
n. This difference response is specified as follows.
Q(n) = O(n) − Z (n) (1)
where Q(n) is an error indication. Here, mentioned signal can be served through the
way which modifies or adjusts the factors confined by filter which can be ranged in
time, i.e., starts from n and goes until (n + 1) through a precise method. Following
method of reworking has signified through a slanted arrow which goes through filter
block shown as above figure. When time catalog is increased, which is n, the output of
the adaptive filter will become improved match as compared with preferred response
signal is hoped via adaptation procedure, which may lead to decline of the value
Q[n] as time goes on. Here, in the prearranged framework, an adaptive process
is responsible for the automatic adjustment of filter weights. Interrelated with other
algorithms, LMS algorithm is comparatively easy and straightforward. It is not related
with the computation of correlation function as well as inversion of matrix.
The steps regarding LMS algorithms are presented below.
1. Filter output

k−1
Z (n) = M[n − l]wl n (2)
n=0
2. Estimation error
Q(n) = O(n) − Z (n) (3)
3. Adaptation weight
Wl [n + 1] = Wl [n] + u M [n − l]Q[n] (4)

where Z[n], O[n], Q[n], W [n], Wl [n + 1] and u represent output signal, desired
signal, error signal, old value of weight, updated value of the weight, and step
size of algorithm, respectively.
3 System-Level Description
Figure 2 displays a topmost level test arrangement, where peripherals have been
interfaced with Altera DE2 equipment. The above-mentioned peripherals include a
personal computer, an oscilloscope, and a waveform or signal generator. The pattern
file of FPGA, which is contained by the personal computer, is downloaded via JTAG
cable, and the input signal having noise is served from the signal generator [11, 12].
Also, the response of the filter is perceived through oscilloscope. This is a principal
arrangement necessitated in FPGA for any sort of signal processing.
The LMS design architecture in FPGA [13, 14] is presented in Fig. 2. The I2C
clock carrying a frequency of 10 kHz has been generated using 50 MHz clock. The
external codec supplies a clock, bearing a frequency of 27 MHz which gotten into a
FPGA clock PLL. This results in producing a AUDIO_XCK master clock having a
frequency of 18.4 MHz. The realization of audio bit clock and as well as audio left ~
right clock is done using the master clock. Digitalization of the external noise is done
with the help of audio codec, which goes into the FPGA as AUD_ADC_LRCK clock.
Addition to this, using rate of audio bit clock, the AUD_ADC_DATA signals are also
maintained. The AUD_ADC_DATA, which is a 16 bit serial data, is transformed as
parallel via SIPO module. This data is then delivered to MUX 1, to pick external noise
Fig. 2 Topmost level block diagram

mixed signal or internal noise mixed signal, which are the outputs of signal generator
and sine LUT, respectively. M[n], O[n], Z[n], and Q[n] are the four parameters
comprised by LMS adaptive filter block, which are applied upon MUX 2 to observe
the outcomes using the switching process. MUX 3 is utilized for choosing only one
signal, i.e., either adaptive output signal or external raw signal. After that, the output
of MUX 3 is further delivered into MUX 4, where it will choose the internal pure sine
wave signal or it will choose processed signal. PISO procedure is then performed for
witnessing the outputs serially through oscilloscope (Fig. 3).
The present work considers the noise signal generated through Verilog program
internally in FPGA, and hence, there is no external source connected to the FPGA
setup. There are 18 DPDT switches totally accessible through DE2 board. Switch
13 is utilized to choose the internal sine wave or processed internal noisy sine wave.
At the time, when switch 13 is made OFF, it will yield internal sine wave and yield
processed noisy sine wave as input when it is made ON. Switch 17 is utilized to pick
out internal sine signal or to pick out the adapted signal. Here, yielding of internal
sine signal and adapted sine signal depends upon the OFF and ON condition of switch
17. It yields internal sine signal, when the switch is in OFF condition and vice versa.
Fig. 3 Internal operation diagram

Switch 16 is used to choose the external signal or the adapted signal. When it is in
OFF condition, receives the external signal and receives adapted signal when it is in
ON condition as output. Since adapted output result is necessitated, switches 16 and
17 are remained to be ON. Switches 15 and 14 are used to realize all the outputs of
mentioned filter. M[n], which is the input signal, is displayed while switches 15 and
14 stay in OFF condition. When the prior switch is made OFF and the later switch
is made ON, we get the O[n] signal as output which is the desired signal, and we
will get the Z[n], which is the desired signal by using the vice versa condition. When
both switches stay as ON, we can get the error signal Q[n] in the oscilloscope. The
peripherals are described as follows:
Oscilloscope: Digital oscilloscope has been used for showing the results. Audio
codec: WM8731 is a stereo codec, which is also low power. It is also incorpo-
rated a headphone driver. This codec is generally considered for speech and MP3
players along with recorders. Stereo 24-bit multi-bit sigma delta ADCs and DACs
are employed for over sampling of decimation filter, as well as digital interpolation.
To maintain the digital audio input, the length of the word must stay between 16 and
32 bits, while the rate of sampling should be varying between 8 and 96 kHz. This
codec device is existing in a 28 pin SSOP package.
Altera Cyclone II FPGA is planned for all layer copper, adjusted for minimum
probable size of the die, and process containing 1.2 V SRAM. The Cyclone II FPGA
deals with brilliant performance, greater densities, and profits of programmable logic
under a price of an ASIC design.
MATLAB/Simulink: Simulink® is employed to execute model-built and multi-
domain design and simulation. Simulink® provides system ~ level scheme, embedded
system confirmation along with spontaneous generation of code and continuous test.
It also offers library consisting of customizable block, solvers, and an editor to
model dynamic structures. When assimilate with the MATLAB, it permits us to use
algorithms in different models and also allow us to export outcomes of different
simulations into MATLAB to do further analysis.
4 Design and Simulation
Figure 4 shows the adaptive filter model used for noise cancellation. This model is
done in system generator, using which we have got the VHDL code of the model that
will be directly used in the FPGA. In the model-based design approach, Simulink
model of adaptive noise elimination gives the simulation results of adaptive filter
parameters. Figure 5 shows the input signal denoted as M[n], which is the combi-
nation of both desired signal and noisy signal. Figure 6 shows the desired signal
denoted as O[n], which is removed from the M[n] through adaptive filter. Figure 7
shows the output signal denoted as Z[n], and Fig. 8 shows the error signal which we
got by comparing the desired signal O[n] and output signal Z[n]. Figure 9 shows the
output of wave scope, and Fig. 10 shows the simulation results of Quartus (Figs. 11
and 12).
Fig. 4 Simulink model of the adaptive filter for noise cancellation
Fig. 5 Input signal

waveform M[n]
Fig. 6 Desired signal

waveform O[n]
5 Hardware Implementation
In hardware setup (Fig. 13), the Altera DE2 kit is connected to the laptop by using
JTAG cable through USB blaster port of DE2 board. A 3.5 mm audio jack is connected
between line out and oscilloscope. The DE2 board is supplied via 9 V DC power
Fig. 7 Output signal

waveform Z[n]
Fig. 8 Error signal

waveform Q[n]
Fig. 9 Wave scope output
supply. After effective execution of the code, a soft file is created. This file desires
to be dumped into the FPGA via USB blaster from laptop using programmer option
given or provided by Quartus II. Figures 14, 15, 16, and 17 present the experimental
results for input waveform, desired waveform, output waveform, and error waveform,
respectively.
Execution of adaptive noise elimination by means of first-order LMS algorithm on

FPGA is an effective process to cancel the noise and obtain the output exactly same as
needed. LMS algorithm is very much easy, simple, and straightforward as compared
Fig. 10 Simulation results in Quartus environment
Fig. 11 Output signal

waveform Z[n]
Fig. 12 Error signal

Q[n] waveform and
corresponding weight update
signal waveform
Fig. 13 Hardware setup
Fig. 14 Experimental result

of input signal M[n]
Fig. 15 Experimental result

of desired signal O[n]
to other algorithms. This approach is more efficient toward energy saving and space
facility for FPGA. This effort can be expanded toward multi-order LMS to generate
output more perfect. Moreover, the external signal can be chosen as input for real-time
processing.
Fig. 16 Experimental result of output signal Z[n]
Fig. 17 Experimental result of error signal Q[n]
References
1. Mollaei Y (2009) Hardware implementation of adaptive filters. In: IEEE student conference
on research and development (SCOReD). IEEE, pp 45–48
2. Patjoshi RK, Mohapatra KK (2010) FPGA implementation of fuzzy logic elevator group
controller with traffic base system. 5th IEEE international conference on industrial and
information systems (ICIIS), pp 366–371
3. Patjoshi RK (2010) Design and implementation of embedded based elevator control system.
Ph.D. Diss. NIT Rourkela
4. Patjoshi RK, Panigrahi R (2019) Experimental investigation of microcontroller based elevator

positioning control system using fuzzy-logic. Int J Eng Adv Technol 8(5):88–94
5. Kim CM, Park HM, Kim T, Choi YK, Lee SY (2003) FPGA implementation of ICA algorithm
for blind signal separation and adaptive noise canceling. IEEE Trans Neural Netw 14(5):1038–
1046
6. Patjoshi RK, Suvarsha CH, Ali SKI (2017) Design and analysis of novel high performance
CMOS domino-logic for high speed applications. ARPN J Eng Appl Sci 12(17):5109–5114
7. Bahoura M, Ezzaidi H (2009) FPGA-implementation of a sequential adaptive noise canceller
using Xilinx system generator. In: International conference on microelectronics-ICM, pp 213–
216
8. Patjoshi RK, Vootla DH, Majety V (2017) A FDTD model for crosstalk analysis of dynamic
CMOS and bootstrap gate-driven coupled RLGC interconnects. Int J Pure Appl Math
115(6):25–30
9. Lan T, Zhang J (2008) FPGA implementation of an adaptive noise canceller. In: 2008
international symposiums on information processing, pp 553–558
10. Elhossini A, Areibi S, Dony R (2006) An FPGA implementation of the LMS adaptive filter
for audio processing. In: 2006 IEEE international conference on reconfigurable computing and
FPGA’s (ReConFig 2006), pp 1–8
11. Patjoshi RK, Mohapatra KK (2010) FPGA implementation of fuzzy logic controller for elevator
group control system. In: International conference on future engineering trends
12. Ramos R, Làzaro AM, Río JD (2007) FPGA-based implementation of an adaptive canceller for
50/60-Hz interference in electrocardiography. IEEE Trans Instrum Measur 56(6):2633–2640
13. Patjoshi RK, Mohapatra KK (2015) Real time implementation of sliding mode based direct and
indirect current control techniques for shunt active power filter. WSEAS Trans Syst Control
10(1):186–197
14. Diggikar AB, Ardhapurkar SS (2012) Design and implementation of adaptive filtering algo-
rithm for noise cancellation in speech signal on FPGA. In: 2012 international conference on
computing, electronics and electrical, pp 766–771
ICI Mitigation with Fixed-Number (FN)-
and Fixed-Region (FRgn)-Based BS
Coordination and Cooperation
in a Multi-cell Cellular Network
Janmoni Borah, Kshetrimayum Linthoinganbi Devi, and Joyatri Bora
Abstract The paper investigates the effect of inter-cell interference on the user’s
downlink performances at cell-boundary areas specifically when the user is immobile
at home cell or mobility occurs to adjacent and nearby cells. The regular and rear-
rangement of spectrums with well-known frequency reuse schemes in the 19-MCs
network are presented. The paper also illustrates FN- and FRgn-based BS selection
for coordination and cooperation among BSs. The simulated SINR and throughput
being obtained with FN- and FRgn-based BS selection approach are appreciable.
Keywords ICI · Frequency reuse · BS coordination · BS cooperation
1 Introduction
Recently, network densification with overlaying of small-cells (SCs) seems to be a

top priority for the network operator to meet the ever-growing demands of users. The
limited spectrum, cell densifications, and augmentation in mobile users have led to
some major network issues including inter-cell interference (ICI). ICI which typically
arises due to utilization and allocation of the full or same spectrum in nearby and
neighboring cells had been a basic cause of cell-edge user performance degradation
mostly in heterogeneous cellular networks (HetCNs). The path loss, diffraction,
reflection, and blockage of the signal have led to poor coverage of indoor and cell-
edge regions in macrocell (MC). To overcome that, the operators are deploying SCs
(e.g., picocells and femtocells) within MCs, particularly in cell-edge and indoor areas
J. Borah · K. L. Devi · J. Bora (B)

North Eastern Regional Institute of Science and Technology, Nirjuli, India
e-mail: [email protected]; [email protected]
J. Borah
K. L. Devi
732 J. Borah et al.
where MC is inadequate in providing services. SCs not only improves coverage of

cell-edge and indoor regions (e.g., in shopping malls, in hospitals, residential homes,
etc.) by reducing the serving distance from the base station (BS) to its user but also
are easy to deploy, economical, and energy-efficient. SCs also seem to be efficient in
providing services to isolated and outdoor places (e.g., mobile SCs with vehicles) [1–
6]. On the one side, SCs found themselves in network performance improvements;
on the other side, a relatively new approach specifically the Poisson point process
(PPP) has been brought up for modeling and analyzing the HetCNs entities based on
stochastic geometry (SG). The SG permits spatial and analytical modeling of network
entities (e.g., BS and user) without worrying about pre-existing configuration and
environments [7–10].
The diversifications from high transmit power MCs to low-powered femtocells
BSs results in severe ICI effects in HetCNs. SCs with low coverage range although
involve in ICI, but most dominant interferers are the high transmit powered MCs in
HetCNs. Typically researchers, therefore, consider MCs based networks during ICI
analysis and mitigation. The literature suggests various works related to ICI anal-
ysis and mitigation in MCs networks. Among well-known MC-based ICI mitigation
techniques, the frequency reuse schemes, BS coordination, and BS cooperation are
widely used. The static cell-edge user performance in a downlink network is analyzed
with BS coordination strategy and allocation of frequency reuse (FR) schemes among
cells [11]. The author proposes a fractional-coordinated fair scheduling scheme by
using fractional frequency reuse (FFR) with sectorization for improving the cell-edge
user’s performance in the network [12]. In [13], the cell-edge user’s performance in
a downlink network is analyzed with an interference suppression strategy and FFR
scheme of 1.5. In [14], the conventional centralized and distributed cellular network
is planned and the ICI effect on a cell-edge user is discussed.
In [11], the performance of a static cell-edge user at the boundary of three MCs is
analyzed with BS coordination along with FR methods. Motivated from work in [11],
here, authors tried to investigate the ICI effect on the typical cell-boundary user during
static or mobility in the 19-MCs network. In the previous work [16], authors present
ICI analysis with the planning of spectrums with different FR schemes and 120°
sectorization of MCs. The presented work is an extension of the authors’ previous
work [16] and illustrates ICI mitigation with FN- and FRgn-based BS selection in
the 19-MCs network.
The paper subsequent sections are ordered as: Sect. 2 presents a schematic of 19-
hexagonal cell planning with user distribution and FN- and FRgn-based BS selection
algorithms. The schematic of spectrum planning in 19-MCs is presented in Sect. 3.
The paper concluding remarks is given in Sect. 4.
ICI Mitigation with Fixed-Number (FN)-and Fixed-Region (FRgn)-Based BS … 733
2 System Model
Figure 1 depicts the planning of regular hexagonal cells in an unplanned geograph-

ical area to analyze the performance of heterogeneously distributed users. The user
heterogeneity so obtained is time-varying in nature which cannot be generated with a
single-layer PPP, thus modeled with the Cox process [11]. The Cox process brings an
additional layer to the Poisson point process (PPP) with a varying rate. The propaga-
tion is considered in the network with path loss, large-scale distances, and small-scale
channel variations. The network is set up with regular 19-MCs as√ hexagonal grid sites
of radius R = 300 m. The inter-site distance among MCs (D = R 3 m) is maintained
in the network. The work mainly analyzes the severe ICI effects on a typical mobile
user from adjacent MCs during mobility at the cell-edge and boundaries of two or
more MCs. For ease of analysis, a typical mobile user of MC1 is selected at the
boundary, and random mobility is assumed to MC4, and MC3 (Fig. 2a).
During analysis, it is pre-assumed that either user is static/mobile within serving
MC1 or moving to adjacent MC4 or MC3 . Assuming the round-robin algorithm,
the network downlink metrics are analyzed with an allocation of full bandwidth
(180 kHz) and 1-RB to each user. For the downlink scenario, say Pm is the transmitting
power of all 19-MCs, network large-scale fading factor is C, and FR factor is δ.
The downlink signal to interference plus noise ratio (SINR) received by the typical
cell-edge user k served by BS j is given as:
2
Pm h jk Cdk−α
j
SINRk = (1)
(Ik + N)δ
where d kj is the kth user distance from jth BS; hjk is rayleigh fading channel gain
from jth BS to kth user; N is the noise power associated with the network. The first
Fig. 1 Schematic of regular 5000

hexagonal cells with user
distribution 4000
3000
2000
B18
B17 B19
1000 B16 B2 B8
B7 B3
0 B15 B1 B9
B6 B4
B14 B5 B10
-1000
B13 B11
B12
-2000
-3000
-4000
-5000 -4000 -3000 -2000 -1000 0 1000 2000 3000 4000
734 J. Borah et al.
2500 2500
USER FR-1 among MCs USER
FR-1 among MCs
2000 2000
B 18 B18
1500 1500
B 17 B19 B17 B19
1000 B16 B2 B8
1000 B16 B2 B8
500 B7 B3 500 B7 B3
0 B15 B1 B9 0 B15 B1 B9
B6 B4 B6 B4
-500 -500
B14 B5 B 10 B14 B5 B10
-1000 -1000
B 13 B11 B13 B11
-1500 -1500
B 12 B12
-2000 -2000
-2500 -2500
-2000 -1500 -1000 -500 0 500 1000 1500 2000 -2000 -1500 -1000 -500 0 500 1000 1500 2000
(a) (b)
Fig. 2 Schematic of FR-1: a with FN-based BS selection, and b with FRgn-based BS selection
term in denominator Ik denotes the ICI effect on the kth user from adjacent MCs
allocates with the same spectrum and is given by:

Nm
Ik = Pm |h j k |2 Cdk−α
j (2)

j =1
j = j
Say, B is the bandwidth of each RB being allocated, then kth user downlink
throughput being received from jth BS can be written as:
γk = δ ∗ B ∗ log2 (1 + SINRk ) (3)
The effect of FN- and FRgn-based BS selection on user performance metrics

could be understood from the Algorithms 1 and 2, respectively.
736 J. Borah et al.
The section presents an analysis of ICI effects with the available and modified
arrangement of FR schemes [16]. The 19-MCs network with centralized macro-BS is
set up and a typical user with a single transceiver antenna is assumed to be stationary
or mobile to adjacent cells. The simulation parameters used during performance
analysis in MATLAB are tabulated in Table 1 [15].
The schematic of full-spectrum (i.e., FR-1) allocation in 19-MCs with FN-based
BSs selection is presented in Fig. 2a. The user during stationary at MC1 receives
severe interferences from adjacent MC BSs (B4 and B5 ). With FN-based coordination
Table 1 Parameters
Parameters Values
considered during simulation
Path-loss exponent (α) 2.7
Large-scale fading factor (C) 10−3.53
Bandwidth of channel 1.4 MHz
Number of available subcarriers 12
Subcarrier’s spacing 15 kHz
Bandwidth per resource blocks (B) 180 kHz
Total RB 6
RB allocation 1-RB/user
Power of noise (N) −174 dBm/Hz
Number of MCs 19
Typical user mobility pattern Random
Transmit power of MCs (Pm ) 46 dBm
and cooperation, the interferences from B4 and B5 could be mitigated to an extent.

Accordingly, when the user is mobile to adjacent cells, BSs are selected with the
FN algorithm and coordination and cooperation among them take place. Figure 2b
depicts an arrangement of FR-1 with FRgn-based BS selection. With FRgn, the
interferences to a stationary user coming from BS B3 to B6 could be mitigated with
coordination and cooperation among these BSs. The user during mobility within the
cell or adjacent cells could be analyzed accordingly with FRgn.
The regular FR-3 (RegFR-3) and rearrangement of FR-3 (RearFR-3) with 120°
sectorized 19-MCs are presented, respectively, in Fig. 3a, b. With FN-based Bs selec-
tion (Fig. 3a), it could be seen that the interferences to a stationary user coming from
2500 2500 0
0 Rearranged FR-3, 120 sectoring
USER Regular FR-3, 120 sectoring USER among MCs with subbands
among MCs with subbands
2000 fb2 fb1, fb2, fb3 in sectors 2000 fb1 fb1, fb2, fb3 in sectors
fb2 B18
fb3 B18
fb2 fb2 fb2
1500 fb1 fb2 1500 fb3
fb3 B17 fb3 B19 fb3 B17 fb3 B19
fb2 fb2 fb3 fb1 fb3 fb3
fb1 fb2 fb1 fb1
1000 B16 B2 B8
1000 fb1 B16 fb1
B2
fb1 B8
fb3 fb3 fb3
fb2 fb2 fb1 fb2 fb1 fb2
fb1 fb1 fb2 fb1 B7
500 fb3 B7 fb3 B3 500 fb2 fb2 B3
fb2 fb2 fb1 fb2 fb2 fb3 B1 fb2 fb3 fb2
fb1
fb3 B15 fb3 B9
0 fb3 B15 fb3 B1 fb3 B9 0 fb3 fb3
fb3
fb2 fb1 fb2 fb1 B6 fb1 fb1
fb1 fb1
fb3 B6 fb3 B4 fb1 fb1 B4
-500 fb2 fb1 fb2
-500 fb1 fb2 fb1 fb2 fb1
fb1 fb2 B5
fb3 B14 fb3 B5 fb3 B10 fb2 B14 fb2 fb2 B10
-1000 fb1 fb2 fb1 fb2 fb1 -1000 fb3 fb2 fb3 fb2 fb3
fb3 B13 fb3 B11 fb3 B13 fb3 B11
-1500 fb1 fb2 fb1 -1500 fb1 fb3 fb1

fb3 B12 fb1 B12
fb1 fb2
-2000 -2000
-2500 -2500
-2000 -1500 -1000 -500 0 500 1000 1500 2000 -2000 -1500 -1000 -500 0 500 1000 1500 2000
(a) (b)
Fig. 3 FN-based BS selection: a regular FR-3 (RegFR-3) planning, and b rearranged FR-3
(RearFR-3) planning
738 J. Borah et al.
Fig. 4 Rearranged FR-9 2500

0
Rearranged FR-9, 120 sectors
planning with FN-based BS USER fb1=fb11, fb12, fb13
2000 fb2=fb21, fb22, fb23
selection fb13
fb3=fb31, fb32, fb33
fb33 B18
fb33 fb31 fb11
1500
fb23 B17 fb31 B19
fb31 fb12
fb23 fb21 fb32
1000 B2 B8
fb13 B16 fb21 fb32
fb11 fb21 fb22 fb32 fb33
500 fb11 B7 fb22 B3
fb11 fb22 fb33
fb12 B1 fb23
fb31 B15 fb12 fb23 B9
0
fb12 fb13
fb32 fb23 fb21
B6
fb32 fb13 B4
-500 fb13
fb32 fb33 fb11 fb21
B5
fb22 B14 fb33 fb11 B10
-1000 fb23 fb33 fb31 fb11
fb12
fb23 B13 fb31 B11
-1500 fb21 fb31 fb32
fb21 B12
fb22
-2000
-2500
-2000 -1500 -1000 -500 0 500 1000 1500 2000
sectors of BS B2 , 7 allocated with fb1 sub-band could be mitigating with coordination

and cooperation among sectors of these BSs. The interferences to the user during
mobility within the home cell (i.e., to the sector with fb3 sub-band) or adjacent sectors
of MC3, 4 could be analyzed and mitigate with FN accordingly. The FN-based BS
selection for the typical user during stationary or mobility in RearFR-3 planning is
presented in Fig. 3b. It is observed that the user which is stationary at the home cell
in the sector with fb-1 sub-band face interferences from sectors with the same sub-
band of MC4, 5 . With FN-based BS coordination and cooperation, the interferences
in RearFR-3 spectrums arrangements could be mitigated accordingly.
The modified rearranged FR-9 (RearFR-9 [16]) with 120° sectorized 19-MCs
along with the planning of nine sub-bands is presented in Fig. 4. It is observed that
the interferences to a stationary user coming from sectors of BS B4 , 5 allocated with
fb13 sub-band could be mitigated with FN-based BS coordination and cooperation
among sectors of these BSs. As could be observed from Fig. 4, the interferences to
the user during mobility within the home cell (i.e., to the sector with fb12 sub-band)
or adjacent sectors of MC3 with fb22 sub-band and MC4 with fb13 sub-band could
be analyzed and mitigate with FN-based BS selections accordingly.
The simulated SINR of a typical cell-boundary user is plotted in Fig. 5. Figure 5a
depicts SINR of the stationary or mobile user obtained with FN- and FRgn-based BS
coordination. The SINR obtained with FN and FRgn based BS cooperation is shown
in Fig. 5b. The results depict that SINR being obtained with different FR schemes
(in [16]) improves with FN- and FRgn-based BS coordination and cooperation.
Figure 6 shows the simulated downlink throughput of the typical cell-boundary
user when it is static at home cell or moves to adjacent cells. The user throughput
during stationary or mobility obtained with FN- and FRgn-based BS coordination is
Fig. 5 Simulated SINR of 45

the typical user: a with FN- 40 stationary user
user move within home cell MC-1
and FRgn-based BS user move to adjacent cell MC-4
coordination, and b with FN- 35
user move to adjacent cell MC-3
and FRgn-based BS 30
cooperation
SINR (in dB)

25
20
15
10
-5
FR-1 FR-1 with FRgn RegFR-3 RearFR-3 RearFR-9
Frequency reuse with BSs coordination

(a)
45
stationary user
40 user move within home cell MC-1
35 user move to adjacent cell MC-3
30
SINR (in dB)
25
20
15
10
0
Frequency reuse with BSs cooperation
(b)
presented in Fig. 6a. The throughput obtained with FN- and FRgn-based BS coop-
eration is shown in Fig. 6b. The simulated results depict that the throughput of user
improves with FN- and FRgn-based BS coordination and cooperation in the network
in comparison to that obtained with different FR schemes (in [16]).
4 Conclusion
The ICI effect on stationary or mobile users near cell-boundary regions is inves-
tigated with the arrangement of different FR schemes (regular and rearranged of
sub-bands) in the 19-MCs network. The paper discusses ICI mitigation with FN-
740 J. Borah et al.
Fig. 6 Simulated downlink 700

throughput of the typical stationary user
600 user move within home cell MC-1
user: a with FN- and user move to adjacent cell MC-4
FRgn-based BS coordination
Throughput (Kbps)
500
and b with FN- and
FRgn-based BS cooperation 400
300
200
100
0
Frequency reuse with BSs coordination
(a)
700
stationary user
user move within home cell MC-1
600 user move to adjacent cell MC-4
Throughput (Kbps)
500
400
300
200
100
0
Frequency reuse with BSs cooperation

(b)
and FRgn-based BS coordination and cooperation in a multi-cell cellular network.

The simulated SINR and throughput being obtained with FN- and FRgn-based BS
selection approach are appreciable.
References
1. Claussen H, López-Pérez D, Ho L, Razavi R, Kucera S (2017) Small cell networks: deployment,

management, and optimization. Wiley-IEEE Press, London
2. Anpalagan A, Bennis M, Vannithamby R (2015) Design and deployment of small cell networks.
Cambridge University Press, Cambridge
3. Quek TQ, de La Roche G, Güvenç İ, Kountouris M (2013) Small cell networks: deployment,
PHY techniques, and resource management. Cambridge University Press, Cambridge
4. Andrews JG (2013) Seven ways that HetNets are a cellular paradigm shift. IEEE Commun
Mag 51(3):136–144
5. Sui Y, Vihriala J, Papadogiannis A, Sternad M, Yang W, Svensson T (2013) Moving cells: a
promising solution to boost performance for vehicular users. IEEE Commun Mag 51(6):62–68
6. Park Y, Heo J, Kim H, Wang H, Choi S, Yu T, Hong D (2014) Effective small cell deployment
with interference and traffic consideration. In: 2014 IEEE 80th vehicular technology conference
(VTC2014-Fall), Vancouver, Canada, pp 1–5
7. ElSawy H, Sultan-Salem A, Alouini M, Win MZ (2017) Modeling and analysis of cellular
networks using stochastic geometry: a tutorial. IEEE Commun Surv Tutorials 19(1):167–203
8. Trigui I, Affes S, Liang B (2017) Unified stochastic geometry modeling and analysis of cellular
networks in LOS/NLOS and shadowed fading. IEEE Trans Commun 65(12):5470–5486
9. Chen C, Elliott RC, Krzymień WA, Melzer J (2018) Modeling of cellular networks using
stationary and non-stationary point processes. IEEE Access 6:47144–47162
10. Heath RW, Kountouris M, Bai T (2013) Modeling heterogeneous network interference using
Poisson point processes. IEEE Trans Sig Process 61(16):4114–4126
11. Khan MHA, Chung Jin-Gyun, Lee MH (2016) Downlink performance of cell edge using
cooperative BS for multicell cellular network. EURASIP J Wirel Commun Netw 2016(1):1–12
12. Abiri M, Mehrjoo M (2014) Improving the performance of cell edge users with fractional
coordinated fair scheduling. IET Commun 8(13):2315–2321
13. Ghaffar R, Knopp R (2012) Interference suppression strategy for cell-edge users in the
downlink. IEEE Trans Wirel Commun 11(1):154–165
14. You X, Wang D, Zhu P, Sheng B (2011) Cell edge performance of cellular mobile systems.
IEEE Trans Sel Areas Commun 29(6):1139–1149
15. Chu NEX, Zhang J (2016) Small-cell deployment over existing heterogeneous networks.
Electron Lett 52(3):241–243
16. Borah J, Hussain MA, Bora J (2020) ICI effects on user mobility with different FR schemes
in a multi-cell cellular network. IET Commun 14(1):84–88
InGaAs-Based Square-Shaped
Nanopillar Array for High Photodetector
Performance
Smriti Baruah, Santanu Maity, and Joyatri Bora
Abstract Nanopillar arrays over the photodetector’s surface ensure a greater injec-
tion of incident light by enhancing light trapping phenomenon due to multiple
reflections from the nanopillars which contribute towards high-resolution optical
communication. This paper reports the analytical investigation of the Indium Gallium
Arsenide material based square-shaped nanopillar (NP) array over the detector’s
front surface for high-performance photodetection. The proposed structure exhibits
an optimum 98% overall absorption of the incoming light. High EQE of about 55%
and responsivity of 0.7 A/W has been obtained in 1.5 um operating wavelength range
with a minimum depletion width of 0.65 um.
Keywords Photodetector · Nanopillar · Multiple reflections · Quantum

efficiency · Responsivity
1 Introduction
Progressively, the majority of long-distance transmission nowadays operates in the

(1300–1550) nm wavelength range as this corresponds to the minimum optical loss
window for silica optical fibers [1]. In this respect, aggressive persuasion has been
done to enable high bandwidth, high speed and low noise photodetector application
in the efficient performance of the optoelectronic integrated circuit [2]. The proficient
opto-electrical conversion quality of photodetector placed at the end of the optical
path of an optical communication system greatly impacts the overall performance
S. Baruah (B) · J. Bora

North Eastern Regional Institute of Science and Technology, Nirjuli, Arunachal Pradesh, India
J. Bora
S. Maity
Indian Institute of Engineering Science and Technology, Shibpur, West Bengal, India
744 S. Baruah et al.
of the system [3]. However, the key issue that seriously limits the effective photode-
tector’s functioning is the inadequate coupling of light on the surface due to the
high refractive index mismatch at the air–semiconductor interface [4]. Conventional
antireflection coatings (ARC) consisting of single and multilayer thin films coated
on the semiconductor substrates operate only within a narrow range of wavelength
and small variations of light incident angles [5]. Also due to the unavailability of
the precise desired refractive index of the thin film materials, optimum reduction
of light reflection losses is not achievable in a practical sense [6]. All these phys-
ical and structural constraints could be overcome by the nanostructure deposition
to reduce reflectance losses [7]. The nanostructured material provides maximum
flexible space for reconstruction [8]. Also, the light absorption mechanisms in these
materials depend on the geometry of the structure which not only enhances the optical
path length due to multiple scattering phenomena taking place within these array of
structures but also produces an effective gradient refractive index [9]. Therefore,
could serve as an omnidirectional broadband ARC for photodetector applications
[10]. Intensive exploration for these nanostructures in minimizing optical losses
includes: Silicon nanowire, nanohole, nanopyramids reducing light reflection losses
and optical path enhancement owing to moderate filling ratios [11], Plasmonic silver
nanostructured array ARCs attributing light Raman scattering [12], Low aspect ratio
nano cylinders reducing reflection losses through forward Mie scattering [13], Broad-
band nanocone arrays mitigating reflection losses by providing effective refractive
index medium [14] and excitation of localized surface plasmon resonances through
InAsSb-coated Au nanopillar [15].
To date, the path followed by the incoming light during the multiple internal
reflection phenomenons taking place within the interpillar spacing of the nanopillar
arrays for enhancing photon reabsorption probability in the photodetector’s surface
has not been discussed in length.
This paper presents the analytical investigation of the multiple internal reflec-
tion phenomenons taking place within an InGaAs-based square-shaped nanopillar
(NP) array deployed over the photodetector’s surface of the same material for
enhanced optical coupling. Optical modeling of the square-shaped nanopillar array
as well as the mathematical analysis of the surface reflectance has been provided in
Sect. 2. Section 3 investigates the various photodetector performance metrics with
the nanopillar deployment through graphical representations.
2 Optical Modeling of the Nanopillar
Figure 1 provides the three-dimensional schematic representation of the InGaAs-

based square nanopillar array with each of the pillar edge being ‘s’ height ‘h’ and
interpillar spacing ‘z’ to be mounted over a photodetector’s surface of 500 nm ×
500 nm area. The pillar arrangement over the surface has been done in a manner to
achieve maximum absorption of the incoming light inside the device. As could be
demonstrated from the figure, each of the four faces of the nanopillar would assist in
InGaAs-Based Square-Shaped Nanopillar Array … 745
Fig. 1 Schematic of the

proposed square-shaped
nanopillar array
undergoing photon multiple internal reflections with their neighboring four square
NPs.
The cross-sectional view of the NP showing the entire path followed by the
incoming light during multiple internal reflections taking place within the interpillar
gap ‘z’ has been provided in Fig. 2. For the simplified analysis of the reflectance
pattern, five incoming photons are considered to capture inside the interpillar spacing.
The reflected portion of the three of the incoming photons striking the planar
detector’s surface with fixed incident angle ‘θ i ’ undergoes multiple reflections in
the NP interfaces with striking angle ‘θ p = 90 − θ i ’. Similarly, two of the incoming
Fig. 2 Cross-sectional view

of the NP array illustrating
the light interaction
mechanism within the NP
array
photons striking the NP interface directly with incident angle ‘θ p ’ get highly absorbed
as it faced an effective gradient refractive index inside the NP structure. Thus, an
overall enhancement of the light reabsorption probability due to the increased optical
path length results in achieving optimum light coupling efficiency.
The total number of square nanopillars (N p ) that could be mounted over a fixed
‘x × y’ area could be mathematically evaluated as,
√
tan(90◦ − θi ) area(x × y) + 2h
Np = (1)
s tan(90◦ − θi ) + 2h
The overall light reflectance R obtained with the deployed NP array structure
considering the negligible transmission rate could be induced by combining the
total reduced reflectance of the incoming photons striking initially the air-nanopillar
interface (Rp ) and the air- planar surface(Rs ), respectively. Thus,
R = Rp + Rs (2)

n
N
R = Rpi [(1 − Rs ) + (1 − (Rp − Rp ))] + Rsi [1 − (1 − Rp )] (3)
i=1 i=n+1
where Rp denotes the effective light reflectance obtained from the multiple
internal reflections inside NP array, respectively, ‘N’ denotes the total number of
incoming photons and ‘n’ denotes the total number of incoming photons striking the
‘N p ’number of nanopillars within a fixed detector’s surface area which is given by,

n = m[2 Np ( Np − 1)] (4)
where m indicates the incoming photons trapped inside a single interpillar spacing.
The InGaAs-based square-shaped nanopillar is mounted over the planar surface of a

photodiode of the same material with a filling factor of 0.22. The numerical, as well
as the graphical representation in this paper, has mainly focused on the investigation
of performance analysis of the detector’s front surface in efficient light trapping. The
numerical values of interpillar spacing ‘z’ and NP height ‘h’ have been calculated so
as to allow most of the incoming photons to get trapped inside the array for maximum
absorption without the shadowing effect. Figure 3 provides the amount of total light
absorption (in %) within the NP array structures of fixed height and interpillar gap of
0.4 um and 0.5 um respectively with varying photon incident angle (θ i ) over ‘1 um
× 1 um’ detector’s surface area. The figure shows an increased absorption up to
Fig. 3 Variation of light 1

absorption within a 0.4 um
high and with 0.5 um 0.9
interpillar gap NP array with 0.8
increasing angle of light
Absorption in %
incidence (θ i ) 0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 20 40 60 80 100
light incident angle (theta i)
98% with increasing incident angle up to a certain value beyond which there is a
gradual decrement in the overall light absorption due to the shadowing effect of the
nanopillar which inhibits photon trapping. The incident angle at which maximum
absorption took place depends on the distance between the nanopillar and striking
point of the incoming photon at the plane of incidence.
Figure 4 shows the variation of interpillar (‘z’) spacing required for maximum
light absorption inside the NP array structure for a different angle of incidence at
a fixed pillar height. For larger light angles, there is a wider interring pillar gap
required to trap as much as photon as compared to that of narrow incident angles
so as to achieve optimum absorption. As could be demonstrated from the figure, the
interpillar spacing required for the widest light incident angle is approximately five
times larger than the value of pillar height.
Fig. 4 Maximum NP 2.5

interpillar spacing required h=0.1 um
Spacing between pillars (z in um)
according to the light h=0.2 um

incident angle for pillar 2 h=0.3 um
h=0.4 um
height varying from 0.1 to
0.4 um for maximum
absorption 1.5
0.5
0
10 20 30 40 50 60 70
Incident angles (in degrees)
Fig. 5 Variation of 1.4

maximum pillar height z=0.1 um
required according to the 1.2 z=0.2 um
z=0.3 um
light incident angle with
Height of pillars (h in um)

z=0.4 um
different interpillar spacing 1
from 0.1 to 0.4 um
0.8
0.6
0.4
0.2
0
10 20 30 40 50 60 70
Incident angles (in degrees)
Figure 5 demonstrates the maximum number of nanopillars of fixed height and

spacing that could be mounted over a detector’s surface area of 1 um × 1 um w.r.t
to light incident angles. As could be evaluated from the figure, for wider incident
angles, there is a continuous reduction of the total number of NPs that could be
accommodated over a fixed surface area. Also, the number of longer NPs required
for maximum light absorption is comparatively very less than that of shorter NPs.
The maximum absorption (in %) of light with an oblique incidence at ‘30o ’
considering zero transmission rate within an NP array with varying interpillar gap at
different operating wavelength range (λ) is shown in Fig. 6. Within a certain spacing
for fixed NP height, the overall photon absorption rate goes on increasing as more
number of incoming photons could be trapped inside the structure. However, beyond
a certain limit, increasing the pillar gap would not be helpful in contributing multiple
Fig. 6 Total photon 98

absorption (in %) within the
0.4 um high NP array with 96
varying interpillar spacing
for operating wavelength 94
Absorption (in %)
range from 1.2 to 1.5 um

92
90
88
lambda = 1200 nm
86 lambda = 1300 nm
lambda = 1400 nm
84 lambda = 1500 nm
82
0.2 0.3 0.4 0.5 0.6 0.7 0.8
Pillar height (in um)
internal reflections and would affect the number of NPs that could be accommodated
over the detector/s surface which will reduce the light absorption efficiency. InGaAs
NPs operating at 1.5 um wavelength exhibits the highest absorption of 98% by 0.4 um
long NP arrays at 0.5 um fixed pillar spacing.
The variation of total photon absorption (in %) within the NP array with increasing
NP height is well represented in Fig. 7. As could be demonstrated from the figure,
longer NPs increases the probability of facing multiple internal reflections boosting
the total light absorption rate but beyond the limit increasing, the pillar height would
produce the shadowing effect which will inhibit the incoming photons to get trapped
inside the structure which ultimately reduces the detector’s performance. The 0.6 um
long NP array shows a maximum absorption of 98% with a fixed interpillar spacing
of 0.7 um at 1.5 um operating wavelength.
Figures 8 and 9 provide the comparison of the performance efficiency of the
photodetector with and without nanopillar array structure in terms of external
quantum efficiency (EQE) and responsivity (A/W) metrics. Enhanced carrier absorp-
tion with the 0.4 um long nanopillar array structure would provide a higher elec-
tron–hole pair generation. The higher quantum efficiency of a maximum 55% and
responsivity of 0.7A/W at 1.5 um operating wavelength for a narrow 0.65 um deple-
tion width has been obtained in comparison to that of a planar detector’s surface
which shows only a maximum of quantum efficiency and responsivity of 0.22 and
0.14A/W, respectively.
Fig. 7 Total photon 98

absorption(in %) within the
NP array with varying pillar 96
height and fixed interpillar
spacing of 0.7 um for
Absorption (in %)
94
operating wavelength range
from 1.2 to 1.5 um 92
90
88
lambda = 1200 nm
lambda = 1300 nm
86 lambda = 1400 nm
lambda = 1500 nm
84
0.3 0.4 0.5 0.6 0.7 0.8 0.9
Pillar height (in um)
Fig. 8 Measured spectral 0.7

responsivity of the PD with
and without NP array surface 0.6
for a detector’s depletion
width of 0.65 um and 1 mw 0.5
Responsivity (A/W)
input optical power within a
varying wavelength range of 0.4
1.2–1.8 um
0.3
with nanopillar array
with flat surface
0.2
0.1
0
1.2 1.3 1.4 1.5 1.6 1.7 1.8
Wavelength (in um)
Fig. 9 EQE comparison 0.7

curves of the PD with and with nanopillar array
without NP array surface 0.6 with flat surface
with a detector’s depletion
width of 0.45 um and 1 mw 0.5
optical power
0.4
EQE
0.3
0.2
0.1
0
1.2 1.3 1.4 1.5 1.6 1.7 1.8
Wavelength (in um)
4 Conclusion
A high-performance InGaAs material-based square nanopillar array has been

proposed to be deployed over the front surface of a photodetector of the same mate-
rial composition. The overall absorption of the incoming light has been boosted to
the greatest extent due to the multiple reflections of the light inside the NP array.
Proper manipulation of inter-pillar spacing and height of the NP structure results
in enhanced light trapping efficiency and thus minimizing the detector’s reflection
losses to a negligible level. The proposed structure shows a high light absorption
of about 98%, The EQE of 55% and 0.7 A/W responsivity with a narrow depletion
width of 0.4 um, which is higher than that of a photodetector without NP array.
References
1. Dawson JW, Kiani LS, Pax PH, Allen GS, Drachenberg DR, Khitrov VV, Chen D, Schenkel
N, Cook MJ, Crist RP, Messerly MJ (2017) E-band Nd 3 + amplifier based on wavelength
selection in an all-solid micro-structured fiber. Opt Express 25(6):6524–6538
2. Zhao Z, Liu J, Liu Y, Zhu N (2017) High-speed photodetectors in optical communication
system. J Semiconduct 38(12):121001–121005
3. Islam MM, Ahmed S, Islam A (2019) Performance analysis of 2.5 Gbps PIN and APD
photodiodes to use in free space optical communication link. Int J Thin Fil Sci Tec 8(2):53–58
4. Kang Y, Liang D, Mehra S, Huo Y, Chen Y, Christoforo MG, Salleo A, Harris JS (2015)
Efficiency enhancement of gallium arsenide photovoltaics using solution-processed zinc oxide
nanoparticle light scattering layers. J Nanomater 2015:1–6
5. Jackson EM, Nolde JA, Kim M, Kim CS, Cleveland ER, Affouda CA, Canedy CL, Vurgaftman
I, Meyer JR, Aifer EH, Lorentzen J (2018) Two-dimensional plasmonic grating for increased
quantum efficiency in midwave infrared nBn detectors with thin absorbers. Opt Express
26(11):13850–13864
6. Schubert MF, Poxson DJ, Mont FW, Kim JK, Schubert EF (2010) Performance of antireflec-
tion coatings consisting of multiple discrete layers and comparison with continuously graded
antireflection coatings. Appl Phys Express 3(8):082502–082503
7. Leung SF, Zhang Q, Xiu F, Yu D, Ho JC, Li D, Fan Z (2014) Light management with
nanostructures for optoelectronic devices. J Phys Chem Lett 5(8):1479–1495
8. Ilyas N, Li D, Song Y, Zhong H, Jiang Y, Li W (2018) Low-dimensional materials and state-
of-the-art architectures for infrared photodetection. Sensors 18(12):4163
9. Rani A, Reddy R, Sharma U, Mukherjee P, Mishra P, Kuila A, Sim LC, Saravanan P (2018)
A review on the progress of nanostructure materials for energy harnessing and environmental
remediation. J Nanostruct Chem 8(3):255–291
10. Kang JI, Kim H, Han CY, Yang H, Jeon SR, Park B, Cha HY (2018) Enhanced UV absorption
of GaN photodiodes with a ZnO quantum dot coating layer. Opt Express 26(7):8296–8300
11. Buczyński R, Klimczak M, Stefaniuk T, Kasztelanic R, Siwicki B, St˛epniewski G, Cimek J,
Pysz D, St˛epień R (2015) Optical fibers with gradient index nanostructured core. Opt Express
23(20):25588–25596
12. Yu R, Lin Q, Leung SF, Fan Z (2012) Nanomaterials and nanostructures for efficient light
absorption and photovoltaics. Nano Energy 1(1):57–72
13. Bandarenka HV, Girel KV, Bondarenko VP, Khodasevich IA, Panarin AY, Terekhov SN
(2016) Formation regularities of plasmonic silver nanostructures on porous silicon for effective
surface-enhanced Raman scattering. Nanoscale Res Lett 11(1):262
14. Zhang D, Ren W, Zhu Z, Zhang H, Liu B, Shi W, Qin X, Cheng C (2015) Highly-ordered silicon
inverted nanocone arrays with broadband light antireflectance. Nanoscale Res Lett 10(1):1–6
15. Lee WJ, Senanayake P, Farrell AC, Lin A, Hung CH, Huffaker DL (2016) High quantum
efficiency nanopillar photodiodes overcoming the diffraction limit of light. Nano Lett
16(1):199–204
Identification of Approximable Program
Components Using Edge Profiling
Priya Arundhati and Santosh Kumar Pani
Abstract Approximate computing is gaining popularity day-by-day due to its

growing application area. One of the areas where it is required to give attention is
program analysis techniques. Program analysis focuses mainly on program optimiza-
tion and the correctness of the program. For program optimization, some transforma-
tion in the source code is carried out. In this paper, we have proposed a novel idea for
transforming a normal program into an approximated program. The approximated
version of the program can run in the non-approximated hardware and can decrease
computation time, resource utilization, memory uses, and power consumption. Our
method finds the cold segment of the original program by using Eege profiling and
apply approximation as these parts have less effect on the final output value. The use
of the distortion metric verifies the output quality of our transformation.
Keywords Program optimization · Approximate program transformation · Edge

profiling · Distortion
1 Introduction
In today’s era, everything is managed by machines. These machines are designed in

such a way that they do not require human interaction, hence equipped with high
computing power. The high computing machines require high power and energy
consumption to produce the result. There are some applications where we may not
require the exact result but a “good enough” result. For providing an exact result,
the machine may require more energy, but for an acceptable result, it may require
less energy. The computation of in exact result with energy-saving is known as
“Approximate Computing.” Nowadays, approximate computing is an integral part
of the computing methodology. Approximate computing relaxes the correctness of
P. Arundhati (B) · S. K. Pani

School of Computer Engineering, KIIT University, Bhubaneswar, India
S. K. Pani
754 P. Arundhati and S. K. Pani
the system by producing an acceptable result along with resource-saving. Approxi-

mate computing is applied in the area where the system is error-tolerant like multi-
media processing, machine learning, signal processing, scientific computing, etc.
Except these traditional application areas of approximate computing, researchers are
thinking in a different way and trying to apply the concepts in the area where we can
optimize resources like software engineering, program analysis, optimization, etc.
However, most of the research in approximate computing is not related to software
testing research. At the code level, the components (variable, code, loop, etc.) of a
program need to be handled carefully and the section of code can be transformed
into approximate code. Generally, program transformation changes the program to
a different form without changing its semantics and used in reliability, maintenance,
and extensibility of software. On the other hand, approximate program transforma-
tion (APT) changes the semantics of the original program within a specified error
bound and this transformed program can take less computation time, less power to
execute, and less memory. To make a non-approximate program into an approximate
program, first, the components where the transformation should be applied must be
found out. After identification of the program components, then only, the approxi-
mate techniques can be applied in the program without crossing the tolerable error
limit. The transformed program is called as approximate program. Then the output
of the original program and the approximate program is compared and the error is
found out. If the error is within the specified bound, then the output is accepted.
In this paper, we will discuss the techniques by which we identify a suitable
program component for approximation. For the identification of the components, we
have taken the help of edge profiling algorithm [1] along with the edge selection
criteria for control flow graph (CFG) [2]. When one component is identified for
approximation, it can be replaced by some alternatives which produces approximated
result with a tolerable error. Graphically, it is shown in Fig. 1. The detailed explanation
of the process flow is discussed in Sect. 3. In Fig. 2, we are showing the error analysis
process of our methodology. Here, x is the variable in the original program and it is
replaced with x in the transformed program. It must be noted that x should be the
non-critical component (i.e., have less impact on the final output) of the program. It
has some approximated error err associated with the input. The exact value of x
can be replaced by x by dropping some part of the large computation. In the final
output, the error can be calculated as err + x − x .
Fig. 1 Process flow to

identify approximate
1 Original Program
program component and its 2 Construction of CFG
transformation
Cold Path identification using
3
Edge Profiling
4 Application of Approximation
5 Transformed Program
6 Output Analysis
Identification of Approximable Program Components … 755
Fig. 2 Error analysis process
The remaining section is organized as follows. Section 2 focuses on the discus-

sion of the related works. Section 3 describes the proposed method. Case study
and experimental result are discussed in Sect. 4, and finally, Sect. 5 concludes the
discussion.
2 Related Works
In [3], the authors introduce a new technique, named ApproxiMut, which combines
the power of mutation testing with the feature of approximate computing. In [4], the
approximate transformation is considered as the mutation operator and found that
the same program behaves differently as that of the traditional mutation operator
and the approximate transformation absorbs more mutants that are not done in tradi-
tional mutation operator. Another direction of approximate computing is approximate
instruction processing(AIP) [5] which includes the techniques in compiler, algorithm,
and code level. In [6], the authors propose a software framework which automati-
cally finds the approximable program components by using statistical method. They
find the range of the variables and by using binary instrumentation, the values of the
variables are settled. They measured the new output against the actual output and
which satisfied the QoS threshold they approximated that variable corresponding to
the QoS threshold. In [7], the authors propose a static framework which finds the
Degree of Accuracy (DoA) of program components and creates Component Influ-
ence Graph (CIG) to represent the relationship among various program components.
In [8], the authors have used the concept of slicing to find the non-critical components
of a program. After application of dependency slicing, they found the approximable
program components and after that they have measured the output by using assertion
metrics. Another new method is proposed in [9] where probabilistic reasoning is
used to justify the approximate program transformation.
For any program transformation, the first thing is the identification of the correct
component on which the conversion is applied to achieve profitable trade-off
without compromising much on the quality of the result. Our proposed methodology
comprises of the following steps:
1. Components identification using edge profiling
2. Application of approximation transformation on the identified components.
3. Result verification.
3.1 Component Identification Using Edge Profiling
A control flow graph (CFG) describes the behavior of a program graphically. By

the definition of a CFG, we know that a CFG is a directed graph (V, E) contains a
set of vertices (V) and a set of edges (E). In CFG, the vertices or nodes represent
the statements in a program, and edges represent the flow of control between the
statements. An edge x → y in a CFG means that the program control can flow from
node x to node y. Due to the simple structure of CFG, it is the most widely used
representation of the program for the program analysis technique. Profiling is used
to find the optimization candidate for a program. The profiler is used to monitor the
dynamic behavior of a program, by instrumenting it and executing the program with
some input and collects the execution traces. The output of a profiler is to give details
of the computation like time consumed, space consumed, and many more. A profiler
uses CFG for the analysis of the program. In this paper, we are using edge profiling
[1] to identify the components for transformation. The edge profile of a CFG can
accurately identify the cold paths in a program [10]. Before discussing the cold path,
we must explain the definition of the path. A path in a CFG is a sequence of n vertices
and n − 1 edges of the form (v1 e1 v2 e2 v3 . . . en−1 vn ) where for each edge ei , either
ei = vi tovi+1 or ei = vi+1 tovi [event counting algorithm of ball-larus]. The cold
paths of a CFG are the less frequently followed paths. Whereas the hot paths are the
most frequently followed paths. We are choosing the cold paths for transformation
because they do not have much impact on the final output of the program. A path is
said to be cold if any edge in the path is cold. So, to identify the cold path first, we
have to find the cold edges in the CFG. According to [2, 10], there are two criteria
on which an edge is said to be cold. These are local criteria and global criteria.
• Local criteria: According to [10], an edge is cold if the ratio of its frequency to
its source node frequency is below a threshold.
• Global criteria: An edge is cold if its frequency as a percentage of total program
Ffow falls below a threshold.
• If either of these conditions satisfies, then an edge is said to be cold [2].
3.2 Application of Approximate Transformation

on the Identified Components
After successful identification of the approximable component of a program, the next

step is the actual application of approximation, which converts the original code into
the approximated one. From the literature of approximate computing, we came to
know about the direction of application of approximation in the program [11]. The
directions are as follows: on the input of the program (Sampling), skipping some of
the computation, and replacement of code. The choice of the selection of the method
depends on the goal of the application. Under these three broad categories, there
are various software-based approximation transformation techniques available such
as as-loop perforation [12], task skipping or code perforation [13], memorization
[14–16], precision degradation, etc. In this paper, we are using the concept of code
perforation as approximation transformation. Code perforation is a method in which
the particular section of code from a program is skipped. According to our proposed
methodology, if the identified cold path contains a section of code, which has less
impact on the final output of the program, then without executing the section of code
we can discard or skip the execution. In [13], the authors have designed a compiler
called SpeedPress to perform code perforation. Example 1 explains the concept of
code perforation.
In the above example (Example 1), the final output is “r” and it depends on
the value of “product” and “s”. The final value of “s” is produced by the function
“fun()”. This function is again decrementing the value of “q”. For the application,
Example 1 Code
perforation
1. Read(p) CODE PERFORATION
2. q=p
3. s=10
4. product=1
5. if (q mod 2=0)
6. {
7. fun ()
8. {
9. s=q/s
10. q=q-1
11. write(s)
12. }
13. }
14. product=product*q
15. r=product + s
16. Write(r)
if the contribution of “s” produced by the function is very less then to increase the
efficiency of the program, we can skip the condition checking segment (if statement).
By doing this, we cannot get the accurate result of “r”. Some error will be associated
with “r” and the associated error must be within the tolerable limit, but the time
taken by this section of code to execute will be reduced. And this is the beauty of
approximate computing.
3.3 Output Analysis
By the application of approximation, the program produces the inexact result. But
the produced result must be within the user-defined threshold value so that the system
can behave adequately and tolerate the inaccurate result. For the verification of the
correctness of the system, several correctness metrics are used. From the litera-
ture [17], we came to know about various types of correctness verification metrics
depending on the output type. In our proposed method, assuming that our output
is of multimedia type, we are using PSNR [12, 18] as a quality metric. This model
provides the probabilistic bound on the deviated output. The user can use this bound
to find if the output is within the threshold or not.
4 Case Study and Experimental Evaluation
We implement our approach on the C code developed for image compression using
selective-Huffman coding technique. The overall steps involved in this approach is
shown in Fig. 3.
1 Image Read 2
Selec ve-Huffman Coding

4 5 6 7
12
Storing Pixel Build Backtracking
Matlab Code Code Image
intensity and Selec ve- and
For Assigning Compression
its Huffman Assigning
Decompression and Sor ng C Code
probability Tree Code
3
Compressed Code Cold Path Control Flow

Matrix Perfora on Iden fica on Graph
11 10 9 8
Proposed Approxima on
Fig. 3 Image compression process

For the image compression code block (highlighted in Image-3), we generate

a control flow graph and identified the cold path using edge profiling algorithm
proposed in [10]. After identifying the cold path, the corresponding C code is modi-
fied using code perforation technique and final compressed image matrix is generated.
The compressed matrix is then decompressed using traditional MATLAB code for
decompression and we get the original image. The entire process flow of encoding
and decoding.
4.1 Control Flow Graph and Cold Path Identification
In this section, we will show you the cold edge and cold path in our control flow graph
generated for image compression C code. The CFG for our code segment is shown
in Fig. 4. The dashed edges indicate the cold edges generated from the algorithm
described in [10].
Fig. 4 Control flow graph of 1

our C code for image
compression with dashed
line showing the cold edges 2
3
5
4
8 15
6 9 10 16
11 14 17
7 12
20 22
13
21 23
18
19 24
25
26
27
(a) Original Image (b) DCT (c) Selective Huffman (d) Proposed Method
Fig. 5 Comparison of processed Lena image
After we find the cold edge, we have modified the code by reducing most of
the loop to n/2 where n represents the number of iterations the loop executes. This
concept is called loop perforation technique to reduce the execution time. By doing
this, we will get a reduced version of the code and most of the loops in our code
execute lesser time. The final code takes lesser processing time as compared to the
original one. To compare the result of our process with the traditional process, we
took the universally used Lena image and check for the visual clarity as compared
with the traditional one. The overall result is shown in Fig. 5. These images will be
generated after the decompression method corresponding to their compression.
Figure 5a shows the original image. Discrete cosine transform (DCT) is the widely
used compression technique and it needs inverse DCT (IDCT) to obtain the orig-
inal image. The result of Lena image for traditional DCT is shown in Fig. 5b.
Figure 5c shows the result obtained which compressed using selective-Huffman
coding. Finally, Fig. 5d shows the result obtained using our proposed technique.
5 Conclusion
We propose a program approximation technique that drastically reduces the original

code. Reducing original code leads to higher processing power and take less time to
produce a result. We got the result with slight reduction in the accuracy, but that is the
theme of the approximate computing technique. We tested our approach on image
compression through C code and found that the image clarity is quite acceptable.
In the future, we are planning to apply this technique to other images and compare
the result with traditional methods. The future plan also includes the calculation of
peak signal-to-noise ratio (PSNR) of other images compressed using our approach,
so that we can get compete analysis of the result we got.
References
1. Ball T, Larus JR (1994) Optimally profiling and tracing programs. ACM Trans Prog Lang Syst
(TOPLAS) 16(4):1319–1360
2. Bond MD, McKinley KS (2005) Practical path profiling for dynamic optimizers. In: Proceed-
ings of the international symposium on code generation and optimization. IEEE Computer
Society
3. Gligoric M et al (2017) Mutation testing meets approximate computing. In: Proceedings of the
39th international conference on software engineering: new ideas and emerging results track.
IEEE Press
4. Hariri F et al (2018) Approximate transformations as mutation operators. In: 2018 IEEE 11th
international conference on software testing, verification and validation (ICST). IEEE
5. Barua HB, Mondal KC (2019) Approximate Computing: a survey of recent trends—bringing
greenness to computing and communication. J Inst Eng (India) Ser B 1–8
6. Roy P et al (2014) Asac: automatic sensitivity analysis for approximate computing. ACM
Sigplan Notices. vol 49, no 5. ACM
7. Roy P, Wang J, Wong WF (2015) PAC: program analysis for approximation-aware compilation.
In: Proceedings of the 2015 international conference on compilers, architecture and synthesis
for embedded systems. IEEE Press
8. Mitra S et al (2016) A verification guided approach for selective program transformations for
approximate computing. In: 2016 IEEE 25th Asian test symposium (ATS). IEEE
9. Misailovic S, Roy DM, Rinard MC (2011) Probabilistically accurate program transformations.
International static analysis symposium. Springer, Berlin, pp 316–333
10. Joshi R, Bond MD, Zilles C (2004) Targeted path profiling: Lower overhead path profiling for
staged dynamic optimization systems. In: Proceedings of the international symposium on code
generation and optimization: feedback-directed and runtime optimization. IEEE Computer
Society
11. Venkatagiri R et al (2018) Impact of software approximations on the resiliency of a video
summarization system. In: 2018 48th annual IEEE/IFIP international conference on dependable
systems and networks (DSN). IEEE
12. Sidiroglou-Douskos S et al (2011) Managing performance versus accuracy trade-offs with loop
perforation. In: Proceedings of the 19th ACM SIGSOFT symposium and the 13th European
conference on foundations of software engineering. ACM
13. Hoffmann H et al (2009) Using code perforation to improve performance, reduce energy
consumption, and respond to failures
14. Brandalero M et al (2018) Accelerating error-tolerant applications with approximate function
reuse. Sci Comput Prog 165:54–67
15. Sato Y et al (2015) An approximate computing stack based on computation reuse. In: 2015 3rd
international symposium on computing and networking (CANDAR). IEEE
16. Tziantzioulis G, Hardavellas N, Campanoni S (2018) Temporal approximate function memo-
ization. IEEE Micro 38(4):60–70
17. Akturk I, Khatamifard K, Karpuzcu UR (2015) On quantification of accuracy loss in approx-
imate computing. In: Workshop on duplicating, deconstructing and debunking (WDDD). vol
15
18. Rinard M (2006) Probabilistic accuracy bounds for fault-tolerant computations that discard
tasks. In: Proceedings of the 20th annual international conference on Supercomputing. ACM
19. Westbrook E, Chaudhuri S (2013) A semantics for approximate program transformations. arXiv
preprint arXiv:1304.5531
Bidirectional LSTM with Attention
Mechanism for Automatic Bangla News
Categorization in Terms of News
Captions
Md Shopon
Abstract The aim of any classification problem is to create a set of models that
can classify the class of different texts and objects. Text classification is known
as one such application. This problem can be used in various classification tasks,
e.g., news category classification, identifying language, classification of text genre,
recommendation systems, etc. In this paper, we propose a text classification method
using bidirectional LSTM with attention mechanism to classify Bangla news articles.
This news articles are collected from a renowned a news portal Prothom Alo. The
dataset consists of in total 383,304 news articles and there were total number of 12
different categories. Traditionally, news classification task is done in terms of news
content. But in our work, we have performed classification based on the news captions
which takes lesser amount of training time. We have achieved 91.37% accuracy using
our approach. This is the state-of-the-art result that has achieved on this dataset.
Keywords Text classification · Bangla news categorization · News classification ·

Bidirectional LSTM · Attention mechanism
1 Introduction
Text categorization refers to the problem of categorizing some given text into some
specific classes. The categories are mostly predefined and it depends on the text
content in which category the text would lay into. With the recent advancement of
textual data availability, processing these data automatically is now one of the most
crucial tasks in the sector of machine learning and pattern recognition. The approach
of using domain experts to classify and analyze the data to particular categories is
both time and cost consuming, also it has some limitations.
M. Shopon (B)
Department of Computer Science and Engineering,
University of Asia Pacific, Dhaka, Bangladesh
764 M. Shopon
One of the most needed uses of text categorization is in electronic newspapers. In

an electronic newspaper, a user is expected to read the articles from a computer
screen that interests him/her the most. A reader may be interested to read news from
different daily, weekly newspapers while being on move. It would most interest a
user to receive articles or news from his/her favorite sources. Every user expects
to have a custom personalized version of a newspaper with his/her favorable news
in the front pages. This kind of task is being done on different international online
news portals and blogging websites. So text categorization is a task that involves
both business prospects and human labor reduction prospects.
There are not many Bangla news portals that use such an approach to interest
their users in a more efficient way. In this work, we have introduced an approach to
classify news categories from news captions. Many works have already conducted
on text categorization based on the full news article. Text categorization from full
news article is computationally expensive; this is why we have worked on classifying
text category from news captions.
Several machine learning and statistical approach have been initiated for text classi-
fication. Among the most popular are regression model [1], support vector machine
[2], decision tree, N-Grams [3–5], Naive Bayes classifier [6, 7], neural networks
[8–10]. Most of this work is used on news class identification, authorship attribution,
text genre classification [11], etc. But these tasks are mostly used in English lan-
guages. There are very little amount of task that has been done on Bangla news text
classification. Bangla is known to be seventh most spoken language in this world.
Due to lack of Bangla text dataset, not much amount of work has been done on
Bangla text classification.
In this paper, we have introduced a bidirectional LSTM with attention mechanism-
based news categorization in terms of news captions. The next parts of the paper are
structured in the following manner: Sect. 2 discusses about the related works that
have been done on Bangla text classification, Sect. 3 deals with the background terms
for this work, Sect. 4 discusses about the corpus and working methodology that was
used in this work, Sect. 4 states our approach for classifying the texts, in Sect. 5, we
will discuss and compare our results with other works and finally, Sect. 6 will be the
conclusion of this work.
2 Related Work
In this section, we will be discussing some of the work that has been done on Bangla
text classification. Chy et al. [12] used a Naive Bayes classifier for classifying Bangla
news text. Naive Bayes (Naive Bayes) classifier is often used in text classification
experiments because of its effectiveness. The idea of Naive Bayes classifier is to adopt
the joint probabilities of categories and words to predict the probabilities of different
classes in a given document. In this work, they used their own developed RSS crawler
to extract news articles from different online newspapers. After crawling the articles,
they perform some preprocessing to make the text usable. Kabir et al. [13] Stochastic
Bidirectional LSTM with Attention Mechanism . . . 765
Gradient Descent (SGD) classifier to classify texts. For feature extraction, they have
used term frequency and inverse document frequency. Their used dataset consists
of a total of 9127 articles and each of the articles was categorized in nine different
categories. Mandal and Sen [14] used four different classification methods for his
work. They are
• Naive Bays classifier
• K-nearest neighbours (KNN) classifier
• Decision tree classifier
• Support vector machine (SVM).
They used 1000 articles as their corpus and their dataset was divided into five cate-
gories. They achieved the best result using SVM and decision tree. In text categoriza-
tion, SVM classifier converts the original data into a higher dimension; from there, it
can find a hyperplane for segregate the data by using important training details known
as support vectors and then categorize the document. Mansur [15] used an N-Gram-
based categorization method. They have used 1-year news articles from Prothom
Alo news which were divided into 6 categories. An N-Gram is a subsequence of n
number of items in the randomly given order.
3 Background
In this section, we will discuss the literature review of our used approach. We have
used neural networks for text classification this is why this section will introduce you
to the different methods and architectures we have used for classification. Neural net-
work and deep learning are gradually taking over other traditional text classification
algorithms. They have performed immensely well in the field of classification.
3.1 Fully Connected Neural Networks
Neural networks are inspired by mammalian cerebral cortex. Neural networks are
generally formed in layers. These layers are built up of some interconnected ‘nodes’
which goes through an ‘activation function.’ Patterns are represented to the network
via the first layer of neural network which is ‘input layer,’ it communicates to ‘hidden
layers.’ There can be one or multiple hidden layers present in a neural network. The
‘hidden layer’ finally communicates with the ‘output layer’ which gives the output
of the classification. Here, Layer L 1 is the input layer, where x1 , x2 , x3 , . . . , xn is
the input data. Layer L 2 and Layer L 3 is the hidden layer and Layer L 4 is the output
layer.
766 M. Shopon
3.2 Recurrent Neural Networks
Recurrent neural networks (RNN) are a variation of fully connected neural network
commonly used in different machine learning tasks such as speech recognition, image
recognition and natural language processing (NLP). RNNs are designed to classify
sequential characteristics and use specific patterns to predict the next possible sce-
nario. RNN’s input is not only the current input sample but also what they have
anticipated previously. The terminology behind RNNs is to use sequential informa-
tion. Usually, in a neural network, we consider that all the inputs and outputs depend
on each other. But for many tasks, this assumption is wrong. If we want to predict
the following word in a given sentence, it is necessary to know which word came
before it. RNNs are termed recurrent because they perform the exact task for every
component of a sequence, with the output being dependent on some of the previous
calculations. RNNs have a ‘memory’ which keeps the information that was computed
so far so that it can help to predict the next word.
3.3 Long Short-Term Memory (LSTM)
Hochreiter and Schmidhuber [16] proposed a modified version of a recurrent neural

network for reducing the long-term dependency problem. LSTMs perpetuate the
loss that is backpropagated through layers and time. By keeping a constant error, it
allows recurrent neural networs to learn in many time steps continuously. LSTMs
store information in a different way than a vanilla RNN. They store the information
in a gated cell. It can perform three operations on the stored information,
• New information can be stored
• Information can be read
• Information can be erased or write.
This gated cell takes a decision which operation to perform. These analog gates can
be either open or close. Being analog has some advantages over digital. As analog
gates are differentiable, it is suitable for the backpropagation.
The RNN architecture can be defined using the deterministic flow from the previous
to present hidden states. The deterministic flow is a function
RNN: HTl−1 , HTl −1 → HTl
For vanilla RNNs, this function is given by
HTl = f (Tn,n HTl−1 + Tn,n HTl −1 ), where f ∈ {sigm, tanh}
On the other hand, the LSTM has intricated architecture that allows it to easily store
information for a large number of timesteps. The ‘long-term’ memory is reserved in
a vector of memory cells ctl ∈ Rn .
LSTM in its core stores information from inputs that have already gone through it
using the hidden state. This is called unidirectional LSTM. It only stores features or
information of the past because the inputs it has seen are from the past. On the other
hand, bidirectional LSTM(BiLSTM) follows a different architecture. In BiLSTM,
the inputs run in both directions, one is from the past to future and the other one
is from future to past and what alters in this approach from unidirectional is in the
LSTM that runs in backward, we store information from the future and use the hidden
states to combined information from both past and future.
3.4 Attention Mechanism
Attention is one of the most prominent ideas in the field of deep learning. Bahdanau
et al. [17] first proposed the attention mechanism in their work for neural machine
translation (NMT). This mechanism is proven to be very effective in different natural
language processing tasks such as text completion, text generation and question–
answer generation. Also, this mechanism works excellent for different computer
vision tasks such as image captioning, image completion. Attention models imitate
human behavior to process information by concentrating on some specific parts or
segments of data that are most informative. To summarize attention networks, given
a target hidden state h 1 and source context vecto ct , we apply simple multiplication
layer to concatanate the information from the two vectors in order to produce an
attentional hidden state as follows:
ht = tanh(Wc [ct ; h t ])
The attentional vector h t is then passed to the softmax layer to generate the
predicted distribution which is formulated as:
p(yt |y<t , x) = softmax(Ws h t )
4 Corpus and Working Methodology
4.1 Corpus
The dataset we have used in this work is collected from Nabil [18]. This dataset
consists of in total of 437,948 news articles from 32 different categories. Each dataset
consists of ten different attributes. They are:
• Author
• Category Name in Bangladesh
• Category Name in Bangla
768 M. Shopon
• Published Date
• Modification Date
• Tag
• Comment Count
• Title
• URL
• Main Content.
4.2 Dataset Preprocessing
Some of the categories in the dataset are not informative. So we preprocessed the data
and removed 20 uninformative categories from the dataset. So in this work, we have
worked with 12 different categories. Figure 1 shows the histogram representation of
the dataset. Table 1 shows the category names and the number of samples in each
category. Figure 2 shows some headline examples of different categories.
After that, we performed word-level tokenization. The maximum number word
a headline has is 20. So we padded all the headlines to 20-word sequence and the
labels were converted into one-hot encoded vector.
4.3 Model Architecture and Methodology
Standard long short-term memory networks are not able to identify the important parts
for text classification. To resolve this issue, we introduced the attention mechanism
which can capture the most important parts in a given sentence. Let H ∈ R D be a
Fig. 1 Histogram
representation of the corpus
Table 1 Categories and number of samples

Category name Number of samples
Bangladesh 232,504
Sports 49,012
International 30,856
Economy 17,245
Opinion 15,699
Technology 12,116
Education 9721
Foreign life 7402
Mixed news 3443
Art and cultural news 2702
Fun and cartoon 2604
Total samples 383,304
Fig. 2 News headlines of different categories

770 M. Shopon
matrix of hidden vectors [h 1 , h 2 , .....h N that the LSTM network produced. Here D is
hidden layer size and N is the sentence length. The attention mechanism will result
in an attention weight vector α and a hidden weighted representation r. The final
sentence representation is given by:
h∗ = tanh(Wp r + Wx h N ) (1)
where h∗ ∈ R D and wp and wx are the projection of the parameters that were learned
during the training phase. The attention mechanism helps the model to learn the most
crucial part of a sentence when various aspects are considered.
The layer architectures are given below
1. Input Layer: The input layer receives a sentence.
2. Embedding Layer: The input vectors are then mapped into a lower-dimensional
vector, where the maximum vocabulary size is considered 150,000.
3. Bidirectional LSTM Layer: We utilize bidirectional LSTM to get higher-
dimensional features from the embedding layer. We have used two 100-dimensi-
onal BiLSTM layers.
4. Attention Layer: This attention layer is used for producing a weight vector and
joins the world level features from every time step into a sentence-level feature
vector by performing multiplication operation between the weight vector.
5. Flattening Layer: These sentence-level features are then converted into a one-
dimensional vector.
6. FCNN Layer: After flattening, we fed the results into two 128-dimensional fully
connected layers with 0.2 rate of dropout.
7. Output Layer: A probability distribution is produced in the output layer which is
a softmax layer and it produces 12 different classes.
We have used categorical cross-entropy loss function and root mean squared propa-
gation as the optimizer of our model.
5 Result Analysis
We have divided the dataset into two parts. One is for testing purpose and other one
is training purpose. The dataset was divided into 70:30 ratio. Table 2 illustrates the
division of the dataset.
We have conducted three different experiments. The first experiment was done
only by using LSTM networks, bidirectional LSTM networks and bidirectional
Table 2 Division of the dataset

Training sample Testing sample
270,000 113,304
Table 3 Result and analysis

Method Accuracy (%)
LSTM 86.06
Bidirectional LSTM 88.56
Bidirectional LSTM with attention mechanism 91.37
Fig. 3 Accuracy graph
Fig. 4 Loss graph
LSTM networks with an attention mechanism. We have trained all the experiments
for 100 epochs. Table 3 shows the result of the different experiments we have con-
ducted.
From the above table, we can see that bidirectional LSTM with attention mecha-
nism performs better than other experiments. Figures 3 and 4 show the accuracy and
loss graph for our experiment BiLSTM with an attention mechanism. This is the first
work that has been done on this dataset. We consider our result as the state-of-the-art
result for this dataset.
772 M. Shopon
6 Conclusion
Text classification is a core research area in the deep learning field. Different
approaches have been taken for English to achieve better automated classification
performance. In this work, we evaluated the efficiency of bidirectional LSTM with
attention mechanism-based text classification based on a news corpus of a renowned
online news portal Prothom Alo. For Bangla, analyzing the accuracy of our approach
has much better performance for text classification for Bangla. Most of the news clas-
sification tasks are done based on the news content. But in our work, we have done
our work in terms of news captions. As the length of news captions is smaller than
its content, the training time is lesser than training on news content.
Acknowledgements Thanks to Institute of Energy, Environment, Research and Development

(IEERD), University of Asia Pacific for providing funding for this research.
References
1. Genkin A, Lewis DD, Madigan D (2007) Large-scale bayesian logistic regression for text
categorization. Technometrics 49(3):291–304
2. Joachims T (1998) Text categorization with support vector machines: learning with many
relevant features. In: European conference on machine learning. Springer, pp 137–142
3. Fürnkranz J (1998) A study using n-gram features for text categorization. Aust Res Inst Artif
Intell 3(1998):1–10
4. Náther P (2005) N-gram based text categorization, institute of informatics. Comenius Univer-
sity
5. Sebastiani F (2002) Machine learning in automated text categorization. ACM Comput Surv
(CSUR) 34(1):1–47
6. Mooney RJ, Roy L (2000) Content-based book recommending using learning for text catego-
rization. In: Proceedings of the fifth ACM conference on Digital libraries. ACM, pp 195–204
7. Forsberg M, Wilhelmsson K. Automatic text classification with Bayesian learning
8. Lai S, Xu L, Liu K, Zhao J (2015) Recurrent convolutional neural networks for text classifica-
tion. In: Twenty-ninth AAAI conference on artificial intelligence
9. Zhang X, Zhao J, LeCun Y (2015) Character-level convolutional networks for text classifica-
tion. In: Advances in neural information processing systems, pp 649–657
10. Zhou C, Sun C, Liu Z Lau F (2015) A c-LSTM neural network for text classification. http://
arxiv.org/abs/1511.08630
11. Cavnar WB, Trenkle JM, et al (1994) N-gram-based text categorization. In: Proceedings of
SDAIR-94, 3rd annual symposium on document analysis and information retrieval, vol 161175.
Citeseer
12. Chy AN, Seddiqui MH, Das S (2014) Bangla news classification using naive Bayes classifier.
In: 16th International conference on computer and information technology. IEEE, pp 366–371
13. Kabir E, Siddique S, Kotwal MRA, Huda MN (2015) Bangla text document categorization using
stochastic gradient descent (SGD) classifier. In: 2015 International conference on computing
and information processing (CCIP). IEEE, pp 1–4
14. Mandal AK, Sen R (2014) Supervised learning methods for Bangla web document categoriza-
tion. http://arxiv.org/abs/1410.2045
15. Mansur M (2006) Analysis of n-gram based text categorization for Bangla in a newspaper
corpus. Ph.D. thesis, BRAC University
16. Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):1735–1780
17. Bahdanau D, Cho K, Bengio Y (2014) Neural machine translation by jointly learning to align
and translate. http://arxiv.org/abs/1409.0473
18. Nabil ZAN (2019) Bangla newspaper dataset
Sentence Relation Classification Using
Deep Learning Experiments
B. Haripriya and Kishorjit Nongmeikapam
Abstract There are existing problems of finding relation between sentences for quite
a few years. A system that can find the relation between a phrase and a document can
be used for search engines, article finder and many more. This paper is focused on the
problem statement of deciding whether given two sentences or two set of sentences
are unrelated or related and if related whether it is mutually agreeing, disagreeing
or neutral relation will be the classification outcome. Specifically, an attempt to find
the title body consistency of a news article is done in this paper. For which, a deep
transfer learning-based approach is proposed where the problem of detecting title
body consistency is taken from the viewpoint of textual entailment (TE) where the
title is considered as a hypothesis and news body is treated as a text. The framework
used is bi-directional long short-term memory (LSTM) network which is a type of
Recurrent Neural networks in deep learning with experiments with transfer learning
technique.
Keywords Classification · Textual entailment · Bi-directional long short · Term

memory · LSTM · Recurrent neural networks · Transfer learning
1 Introduction
The problem which is being taken into consideration deals with the study of the
credibility and authenticity of information increased on the Web these days. For
example, if a user inputs a phrase or claim in the Web app like “the US is the most
powerful country in the world” or “Barak Obama is a gentleman” or “steps to follow
to study abroad with scholarship.” These claims are taken and searched for many
articles concerning that subject. If there are many reputed (and well known) sources
Supported by TEQIP-III.
B. Haripriya (B) · K. Nongmeikapam

Indian Institute of Information Technology Senapati, Senapati, Manipur, India
K. Nongmeikapam
776 B. Haripriya and K. Nongmeikapam
which all “agree” with this claim, then the particular claim is most probably true.
More concisely, the task can be defined as given a claim and a body of text (like news
article), the system has to decide whether the body of the text generally “agree,”
generally “disagree,” generally “neutral” or is completely “unrelated” to the claim.
This problem is typically called a stance-detection problem. Given two pieces of
texts, one being the premise (P) and another one is the hypothesis (H), the system
has to decide whether:
• the H is the logical consequence of the P or not
• H is true in every circumstance (possible world) in which P is true.
The term entailment is used in the context of logical reasoning [1]. Entailment
refers to when two statements are there and when one of them is true then the other
is true and supports the former statement. For example, P: Johns assassin is in jail
entails H: John is dead.
The main approach used is transfer learning which uses supervised features that
have been used in several computer vision applications like face recognition, question
answering, sentiment analysis and many more. These applications need to compar-
atively fewer data to be trained and tested. The features or the semantic information
is stored in a neural network and will be used for a similar task but on a different
corpus.
For this, standard recurrent models like long short-term memory (LSTM) models
are used. Many basic concepts of natural language processing and deep learning were
studied and this approach was picked. The different recurrent neural networks and
making of those models were learnt in depth. The data set is having the title and body
pairs with its labels (agree, disagree, discuss, unrelated). The back-end classification
algorithm used is logistic regression with the word to vector embedding. Using the
knowledge from an external embedding can enhance the precision of an LSTM
because it can use the information that has been trained and distilled on a very large
corpus of data. The pre-trained embedding used here is from fastText.
1.1 Recurrent Neural Networks
Figure 1 shows one unit of recurrent neural network (RNN). From bottom to top:
input state (x), hidden state (h), output state (o). U, V and W are the weights of the
network. Compressed diagram on the left and the unfold version of it on the right.
Basically, RNNs are best fit for sequential data but they do have some drawbacks.
Firstly, since they process inputs in temporal order, their outputs tend to be mostly
based on previous context; secondly they have trouble learning time-dependencies
more than a few time steps long [2]. As a solution, LSTM and bi-directional LSTM
[3] were discovered which solves these drawbacks of RNNs.
Sentence Relation Classification Using Deep Learning Experiments 777
Fig. 1 Recurrent neural networks interior mechanism
1.2 LSTM
LSTM [4] is a varient of RNN architecture introduced to deal with long time-
dependencies. It was motivated by an analysis of error flow in existing RNNs, which
had a problem of long time lags and were inaccessible to existing architectures,
because the back-propagated error either blows up or decays exponentially. LSTM
contains hidden layers known as memory blocks that remember the input to the cells.
Each of the memory cell consists of an input gate, output gate and forget gate. These
can be imagined as write, read and reset operations for the cells.
1.3 Bi-Directional LSTM
Bi-LSTM connects two hidden layers of opposite directions to the same output. With
this the output layer can get information from past (backwards) and future (forward)
states simultaneously. This feature finely fits the problem statement of this paper.
In Fig. 2, O1, O2, O3, O4 are the output layers of each LSTM layer and X1, X2,
X3, X4 are the input layers. Each LSTM layer is connected with the neighbor LSTM
layer.
2 Related Works
The task of finding relation between any two entities has been done using various
neural network architectures. The paper by Zeng et al. [5] a convolutional deep neural
network(CNN) was used to extract lexical and sentence level features. Firstly, the
word tokens are transformed to vectors by taking up word embedding. Then, lexical
level features are extracted according to the given nouns. Meanwhile, sentence-
level features are learned using a convolutional approach. These two-level features
are concatenated to form the final extracted feature vector. Finally, the features are
Fig. 2 Bi-directional LSTM interior mechanism
fed into a softmax classifier to predict the relationship between two marked nouns.
Another paper by Augenstein et al. [6] focuses on stance detection as a task of classi-
fying the attitude expressed in a text toward a target. Another paper [7] also utilizes 2D
convolution to sample more meaningful information of the matrix. Experiments are
conducted on six text classification tasks, including “sentiment analysis,” “question
classification,” “subjectivity classification” and “newsgroup classification.”
3 Mechanisms
3.1 Initial Setup
The data sets taken for this project are Stanford Natural Language Inference (SNLI),
Fake-News corpus, Facebook word to vector file (fastText-300dim). Stanford Natural
Language Inference (SNLI) is a corpus created by Stanford NLP group which is avail-
able for free to download and use. It is a collection of 570 k human-written English
sentence pairs manually labeled for balanced classification with the labels entail-
ment, contradiction, and neutral, supporting the task of natural language inference
(NLI), also known as recognizing textual entailment (RTE). News corpus is derived
from the emergent data set created by Craig Silverman. The fastText vector file is
freely available for download.
3.2 Text-Pre-processing
Generally, text pre-processing is done for transferring text from human language
to machine-readable format for further processing. Removing punctuation, accent
marks and other diacritics are only done because the special characters are least
important in this paper and considering these might decline the flow of data analysis.
Firstly, the SNLI data set is split into hypothesis and text/body. Then the length
of the longest sentence is found. Next, the classes will be encoded with sequence
of 0 and 1 s. Later FastText vector representation is used to map each word to the
corresponding vector representation in the data set and a new embedding matrix that
contains only the embedding for words of the SNLI data set is present. These same
initial steps are taken for News corpus also.
3.3 Experiments
The embedding matrix of the input data set is reshaped into three dimensions and
given as input to the Bi-LSTM network. Bi-LSTM works on two passes, i.e., forward
and backward. The outputs obtained from such networks for premise/document and
title/hypothesis are concatenated. The concatenated vector is further given to feed-
forward neural network (dense layer in this case) with ReLu [8] activation function.
Later the output will be connected to the output layer, adding a fully connected hidden
layer which is a usually an effective way of learning nonlinear combinations of these
features.
Domain Adaption: This approach is a Naive approach where the training corpus
and the testing corpus are different. The output obtained from the Bi-LSTM layer
represents high level features of the data, while that output will be connected to the
output layer, adding a fully connected hidden layer which is a usually an effective
way of learning nonlinear combinations of these features. Since this was done for
testing purpose, it is not that much efficient.
Transfer Learning: In this approach, two models with Bi-LSTM architecture are
built and one of them is trained on SNLI corpus and the weights of the dense layers
of this model were freezed and transferred to the second model. The second model
is being trained and tested on Fake-News corpus. This way, the semantics learned
from the SNLI corpus is used to detect stances (unrelated, discuss, agree, or disagree)
between an unknown example pair. The basic idea of this approach is that the pre-
trained embedding used has certain features that are common for all text based
classification, like spam detection and many more. Those featured are freezed and
used according to the needed classification task is transfer learning technique. This
approach is well known in the problem condition where less data is available to train
and test the system for new application. So for this, a standard data set is used to get
the common features among the new examples.
Fig. 3 Architecture of the hybrid model. Combined weights of two models are tested with the
News corpus
Hybrid(Combined): At first, the models are built following the same process as
described earlier. The first model is trained on SNLI corpus which is having four
dense layers. The weights of the first two lower layers are saved and transferred to
the third model. Then transfer of the two lowers layers is done, as the SNLI corpus is
very large, the vanishing gradient problem in the lower layers will be less compared
to upper layers. The second model is trained on Fake-News corpus which is also
having four dense layers. Here, transfer of the weights of upper two dense layers
is done as shown in Fig. 3 to the dense of third model. Following this strategy, the
benefits of both SNLI and Fake-News corpus are leveraged which are further utilized
to train and test another model.
4 Results
Implementation includes the use of few necessary Python libraries for mathematical
calculations and open source API for construction of deep learning models. The
following are the verified results given by the proposed model:
1. Domain Adaption
Label/Label Agree Disagree Discuss Unrelated

Entailment 1 0 0 1
Contradiction 12 18 12 34
Neutral 1890 679 4452 18,314
Accuracy: 17.72%
2. Transfer Learning

Agree 23 0 1739 141
Disagree 17 2 572 170
Unrelated 3 0 18,176 170
Discuss 4 0 3355 1105
Accuracy: 76%
3. Hybrid (Combination)

Agree 502 98 1150 153
Disagree 207 108 307 75
Unrelated 135 25 17,520 669
Discuss 47 29 2241 2147
Accuracy: 79.789%
From the hybrid results, it has achieved the best accuracy among the results
obtained from the three proposed models. In this model, the knowledge of seman-
tically enriched corpus’s (SNLI) is incorporated. Here, transfer of the weights from
the upper two layers is done, because as the training corpus (Fake-News corpus) is
comparatively less in size, the vanishing gradient problem in the upper layers will be
less as compared to the lower layers. Then transfer of the weights of four-way final
(with softmax activation) layer of the second model to the third model’s final layer is
done as the actual classification is performed in the third model. The justification of
transferring this final layer’s weights is empirical. Using these weights training and
testing of the third model on News corpus is done.
No system is correct in all respect. All systems have their own advantages and
disadvantages too. The proposed system is not the exception of that the state of
the art is 82.04% but the result got was 79.789% which could be due to less effi-
cient pre-processing or any problem in the internal layers of model. The proposed
system is fully deep learning-based approach which avoids any hand-crafted feature
engineering and also we mitigate the data paucity problem of the target task by
incorporating the concept of TE/NLI with this task which is conceptually similar to
this task and making use of NLI benchmark data (SNLI) through transfer learning
techniques.
5 Conclusion
In this paper, there was an attempt to investigate the role of textual entailment in news
title body consistency through stance detection. As the problem of stance detection is
quite similar with the problem of textual entailment. Framing this problem with the
problem of textual entailment and utilizing the notion of it is done. For that purpose,
various deep neural network-based transfer learning approaches were fostered. Typi-
cally, such kind of analysis leads to many researches upcoming in the field of artificial
intelligence and machine learning. Future direction of this research includes enrich-
ment of the best performing model by incorporating the relevance score between
the title and body texts which is awaited to be resolved by future upcoming papers.
Furthermore, testing of the best performing model in other kinds of corpus is also
one of the future research scope of this paper.
References
1. Sammut C, Webb GI (2011) Encyclopedia of machine learning. Springer, Boston

2. Hochreiter S, Schmidhuber J (1997) Long short-term memory. Neural Comput 9(8):17351780
3. Hochreiter S, Bengio Y, Frasconi P, Schmidhuber J, Kremer SC, Kolen JF (2001) Gradient ow
in recurrent nets: the difficulty of learning long-term dependencies. A field guide to dynamical
recurrent neural networks. IEEE Press, New York
4. Graves A, Schmidhuber J (2005) Framewise phoneme classification with bidirectional LSTM
and other neural network architectures. Neural Netw 18(5–6):602–610
5. Zeng D, Liu K, Lai S, Zhou G, Zhao J (2014) Relation classification via convolutional deep
neural network. zeng2014relation
6. Augenstein I, Rocktschel T, Vlachos A, Bontcheva K (2016) Stance detection with bidirectional
conditional encoding. arXiv:1606.05464
7. Zhou P, Qi Z, Zheng S, Xu J, Bao H, Xu B (2016) Text classification improved by integrating
bidirectional LSTM with two-dimensional max pooling. arXiv:1611.06639v1 [cs.CL]
8. Cui J, Qiu S, Jiang M, Pei Z, Lu Y (2017) Text classification based on ReLU activation function
of SAE algorithm. In: Cong F et (eds) Al advances in neural networks—ISNN 2017. Lecture
notes in computer science, vol 10261. Springer, Cham
Spectral Analysis of Consonants
in Arunachali Native Language—Adi
Sajal Sasmal and Yang Saring
Abstract In this article, an attempt has been made to analyze the spectral properties
of consonants of Adi, which is one of the most common native languages of Arunachal
Pradesh, India. The formant frequencies and spectral properties of consonants in Adi
language like /t/,/k/, /p/, /d/, /g/, /b/, /n/, /ng/, /m/, /ny/, /r/, /s/, /h/, /j/, /y/, /l/ were
studied. Also, the speech samples were analyzed in a wideband spectrogram. The
first formants frequencies (F0) were observed to be varied from 538 Hz to 1228 Hz
and the second formants frequencies (F1) range from 1190 to 2822 Hz for /k/, where
as for /d/, we found the ranges of F0 and F1 to be in the range from 303 to 1621 Hz
and 1710 to 2732 Hz, respectively. It was observed that the formants frequencies of
consonants (unvoiced part) are much higher than the formants frequencies of vowels
(voiced part).
Keywords Adi · Formant · Phonemes · Spectrogram · ASR · Pitch
1 Introduction
Automatic speech recognition (ASR) system allows a machine to recognize the

words spoken by an individual and converts it into text. ASR system is found in
different types of applications like handsfree control and operation (as in airplanes
and modern cars), telephone communication with information systems, automatic
query answering system, automatic dictation, government information system, etc.
Owing to advances in computing devices and architectures, the success of algorithm
development and availability of large quantities of data ASR system has seen a rapidly
rising figure of practical applications as well as commercial markets [1].
S. Sasmal (B) · Y. Saring

Department of Electronics and Communication Engineering, National Institutes of Technology
Arunachal Pradesh, Yupia 791112, India
Y. Saring
784 S. Sasmal and Y. Saring
Since the first ASR system was developed in 1952 for small vocabulary digits,
ASR systems have been developed in several languages like English, Arabic, German,
Mandarin, Spanish, and French, etc. India is linguistically very rich with eighteen
constitutional languages written in ten diverse scripts [2]. In India, machine-oriented
interfaces are restricted to the computer usage to a miniature fraction of the citizens,
who are both computer skilled as well as familiar with on paper English. So there is a
particular necessity for the ASR system to be developed in different languages spoken
in India [2]. In a special reference to northeastern part of India, 350 languages have
been identified in this region, and several types of research are going on in different
northeastern languages such as Assamese [3], Mizo [4], Manipuri [5], Bodo [6], and
Khasi [7].
Linguistically, Arunachal Pradesh is very rich and diverse; around 50 distinct
languages are spoken in along with innumerable dialects and sub-dialects thereof.
However, at a time of rapid globalization when languages are dying at the rate of one
language every fortnight, many of the languages in Arunachal Pradesh are endan-
gered; Adi is one of the endangered languages [8] with around only 100 thousand
people speaking this language. Therefore, for the preservation of such endangered
languages technical intervention such as digital archiving, implementing ASR in
such language is of utmost importance. To implement ASR in any language, the
analysis of its vowels and consonants are of utmost importance. Legoh et al. had
created “Speech Corpus of Adi” and studied linguistic phoneme characteristics [9].
In 2018, Bordoloi et al. performed a spectral analysis of “vowels of Adi” in their
work [10].
There are approximately 50 phonemes in Adi language with 16 consonants, 7
long vowels, 7 short vowels, 19 diphthongs (transition between two vowels), and 1
triphthong [9–11]. The Adi consonants [9, 11] are shown in Table 1.
In the Adi language, there is no dental and labio-dental fricative, retroflex, and
aspirated types sound. For example, [s] and [h] are fricatives but one can swap in
Adi speech without any variation in meaning.
Table 1 List of Adi consonants

Alveolar Glottal Velar Bilabial Palatal
Stops (v) /t/ /k/ /p/
(uv) /d/ /g/ /b/
Nasals (v) /n/ /ng/ /m/ /ny/
Roll (v) /r/
Fricatives (uv) /s/ /h/
Affricates (v) /j/
Approximant (v) /y/
Lateral approximant (v) /l/
v voiced part, uv unvoiced part
Spectral Analysis of Consonants in Arunachali Native Language—Adi 785
2 Formant Frequency of Speech
Format frequencies were computed using the linear predictive coefficient technique
[10]. Different vowels and consonants have a different set of formant frequencies.
Similar sounds /e/ and /a/ are close by similar formant frequencies but distinct sounds
have a different range of formant frequencies. So formant frequencies can be consid-
ered as the signature of vowels and consonants to recognize them. High-pass filtering
of log power spectrum yields pitch harmonics, i.e., useful for pitch estimation and
low pass filter is used to measure formants information of the speech. The pitch
frequencies may be ignored and formant frequencies may be considered at the time
of typical speech sound recognition. In this article, an attempt has been made to find
the range of the formant frequencies of different consonants of Adi, which would be
useful in the design of an ASR system of Adi language.
In this paper, an analysis was done for the first 3–4 formants for different conso-
nants, and scatter plots are shown. The formant frequencies were measured using
“WaveSurfer” and “Praat-6.1.09”.
3 Data Set
For spectral analysis of Adi language, we have recorded 684 Adi sentences from 7
native Adi speakers. Among the 7 speakers: 3 speakers are male and 4 are females.
The speech samples were collected with a sampling rate of 8 kHz and were stored
in .wav format.
4 Spectral Analysis
The Adi sentence “no kape aidun” uttered by an Adi female speaker (AD-F-01) is
shown in Fig. 1. The total duration of the sentence is 5 s but the speaker response starts
from 1.05695 s and ends at 2.08816 s. So the speaker’s involvement was 1.03121 s
only. The word “no” exists from 1.05695 s to 1.2370 s and among this, the consonant
/n/ lies between 1.0569 s and 1.1118 s.
Time is represented along the X-axis and amplitude along the Y-axis of the speech
signal. We look into the spectral properties to compute formant frequencies, pitch,
and energy intensity of different consonants. In Fig. 1, all four formants are shown
by red dots for the Adi sentence “no kape aidun.” The first formant exists just above
X-axis then second formant, third formant, and fourth formant are spread above all.
From the spectrum in Fig. 1, it is observed that the formants follow a sequence for
the speech part but there is no uniformity for noise part of the sound. Pitch is shown
by the blue line in the spectrum, whereas the variation of intensity in the dB scale is
represented by the yellow line. So we can observe that the speech part intensity has
Fig. 1 Sound waveform and spectrogram with formant plot of Adi sentence “no kape aidun”
high value but it is very low for the noise section. Here, /n/ was uttered for 0.0549 s.
In this duration, first formant frequency (F0), second formant frequency (F1), third
formant frequency (F2), forth formant frequency (F3) have been calculated using 8
formants position in the spectrogram. F0 lies between 261 and 365 Hz, i.e., mean in
this section is 306 Hz. Mean of F1, F2, F3 in the selected part are 1608 Hz (1576–
1665 Hz), 2795 Hz (2717–2909 Hz), 4095 Hz (3855–4437 Hz), respectively. The
minimum pitch in this section is 204 Hz and the maximum is 243 Hz but the mean
pitch is 213 Hz. Mean energy intensity in the section is 53.88 dB (51.83–56.31 dB).
In the word “kape” consonant /k/is found in between 1.2652 s and 1.2996 s, so /k/was
uttered for 0.0344 s only. For /k/the mean of formant frequencies F0, F1, F2 is 748 Hz
(483–1328 Hz), 1425 Hz (1311–1968 Hz), 2739 Hz (2594–3031 Hz), and 4318 Hz
(4155–4523 Hz), respectively. The minimum pitch in the section is 96 Hz and the
maximum is 101 Hz giving mean to be 98 Hz. Mean energy intensity in the section
is 46.02 dB (44.68–46.71 dB). The details of formant frequencies of /k/ are shown
in Table 2.
Table 2 Details of formant frequencies of /k/in the word “kape” by “AD-F-01”

Time (sec) F0 (Hz) F1 (Hz) F2 (Hz) F3 (Hz)
1.266827 1328 1968 3031 4523
1.273077 680 1304 2595 4488
1.279327 651 1311 2675 4271
1.285577 696 1358 2754 4245
1.291827 714 1388 2760 4233
1.298077 483 1341 2682 4155
748(mean) 1425(mean) 2739(mean) 4318(mean)
The formants frequencies: F0, F1, F2, and F3 of consonant /k/computed for all
seven speakers by considering more than twenty /k/phoneme uttered by each speaker
is shown in Table 3.
The formant frequencies F0–F3 of consonant /b/were computed and the scatter
plot of F0 versus F1 is shown in Fig. 2. We observed that the F0 lies in the range
from 235 Hz to 492 Hz, F1 lies in the range from 1266 Hz to 1782 Hz, F2 in the
range from 2428 Hz to 3750 Hz and F3 in the range from 3749 Hz to 4341 Hz. We
notice that the range of F0 is 235–492 Hz and F1 is 1266–1782 Hz considering all
four speakers. The frequency range of F2 and F3 are 2428–3750 Hz and 3749–4341
Hz respectively.
The scatter plot for F0 versus F1 for consonant /k/ is shown in Fig. 3 and it is
observed that the formant frequencies F0 lies between 538 Hz and 1228 Hz and F1
lies in the range from 1190 Hz to 2822 Hz.
In Fig. 4, we have plotted F0 versus F1 for consonant /n/and it is observed that F0
lies in the ranges from 261 to 485 Hz and F1 lies in the range from 1256 to 1922 Hz,
whereas for consonant /s/the variation of F0 and F1 are in the range 343–1738 Hz
and 1802–3120 Hz, respectively, as shown in Fig. 5.
The range of formants frequencies F0, F1, F2, and F3 of different Adi consonants:
/b/, /k/, /n/, /s/, /ng/, /t/, and /d/ were computed and listed in Table 4.
Table 3 Average of formant frequencies of consonant /k/

Speaker Average F0 (Hz) Average F1 (Hz) Average F2 (Hz) Average F3 (Hz)
AD-F-01 846.25 1806.94 2991.94 4247.50
AD-M-01 717.67 1608.33 2783.75 4154.00
AD-F-02 910.20 2052.53 3022.67 4412.92
AD-M-02 825.67 1782.27 2974.27 4154.00
Fig. 2 F0 versus F1 plot for

consonant /b/of 4 speakers

consonant /k/of 4 speakers

consonant /n/of 4 speakers

consonant /s/of 4 speakers
5 Conclusion
In this article, we have studied the formant frequencies of consonants in Adi language.
For this work, we recorded 684 Adi sentences from 7 Adi speakers and detected
different consonants from those speech samples. The first four formants frequencies:
F0–F3, pitch period and mean energy intensities were computed. By analyzing the
Table 4 Range of formants frequencies of different Adi consonants

Adi consonant Range of F0 (Hz) Range of F1 (Hz) Range of F2 (Hz) Range of F3 (Hz)
/b/ 235–492 1266–1782 2428–3750 3749–4341
/k/ 538–1228 1190–2822 2149–3698 3817–4826
/n/ 261–485 1256–1922 2525–3277 3579–4549
/s/ 343–1738 1802–3120 2739–4554 4163–5123
/ng/ 280–542 1289–1664 2533–2932 3915–4357
/t/ 658–1470 1524–2715 2501–3613 4077–5061
/d/ 303–1621 1710–2732 2672–4099 4057–5667
spectrogram, we detected all formant frequencies of consonants like /b/, /k/, /n/, /s/,
/ng/, /t/, and /d/. The range of F0 is from 235 Hz to 492 Hz and F1 is from 1266 Hz to
1782 Hz for /b/. The ranges of first two formant frequencies: F0 and F1 for consonants
/k/,/n/, /s/, /ng/, and /t/ are 538 Hz to 1228 Hz and 1190 Hz to 2822 Hz; 261 Hz to
485 Hz and 1256 Hz to 1922 Hz; 343 Hz to1738 Hz and 1802 Hz to 3120 Hz;
280 Hz to 542 Hz and 1289 Hz to 1664 Hz; 658 Hz to1470 Hz and 1524 Hz to
2715 Hz; 303 Hz to 1621 Hz and 1710 Hz to 2732 Hz, respectively. We observed
that the formants frequencies of consonants (unvoiced part) are much higher than
the formants frequencies of vowels (voiced part).
References
1. Baker JM, Deng L, Glass J, Khudanpur S, Lee C, Morgan N, O’Shaughnessy D (2009) Devel-
opments and directions in speech recognition and understanding, Part 1 [DSP Education]. IEEE
Signal Process Mag 26(3):75–80
2. Chandrasekar M, Ponnavaikko M (2008) Tamil speech recognition: a complete model. Electron
J (Tech Acous)
3. Sarma MP, Sarma KK (2011) Assamese numeral speech recognition using multiple features
and cooperative LVQ-architectures. Int J Electr Electron 5(1):27–37
4. Lalhminghlui W, Das R, Sarmah P, Vijaya S (2017) A Mizo speech database for automatic
speech recognition. In: Conference of the oriental chapter of international committee for coor-
dination and standardization of speech databases and assessment technique (O-COCOSDA),
pp 302–306
5. Ningthoujam N, Prathim VR (2016) Designing of a feature extraction system for Manipuri
language. Int J Sci Res Dev 4 (2)
6. Thakuria LK, Acharjee P, Das A, Talukda PH (2014) BODO speech recognition based on
hidden Markov model Toolkit. Int J Sci Eng Res 5(1)
7. Arjunasor S, Deka GK, Ismail T, Singh LJ (2016) Khasi dialects identification based on
Gaussian mixture model. Int J Eng Sci Comput 6(4):3882–3885
8. Post M (2011) A documentation of the upper belt variety of Minyong (Adi), Arunachal Pradesh,
North East India. Endangered Languages Archive
9. Legoh K, Bhattacharjee U, Tuithung T (2013) Development of multi-variability speech corpus
of Adi language for speech recognition researches. Int J Adv Res Comput Sci Softw Eng
(IJARCSSE) 3(10):604–610
10. Bordoloi S, Saring Y (2018) Spectral analysis of vowels of Adi language of Arunachal Pradesh.
In: 2nd international conference on electronics, materials engineering and nano-technology
(IEMENTech), pp 1–4
11. Lalrempuii C (2005) Morphology of the Adi language of Arunachal Pradesh (Doctoral
dissertation)
Double Gate Tunnel FET Versus Double
Gate MOSFET: Electrical Properties
Comparison
Menka Yadav
Abstract In this manuscript, an investigation, with the help of extensive device

TCAD Sentaurus simulations, is presented for comparative analysis to understand
the effects of variations in gate and drain potential on the device electrical properties
of such silicon double gate tunnel field effect transistor (DG TFET) as well as DG
MOSFET. The study is mainly focused at electrical properties like electrical channel
potential, electric channel field, electron density, electron quasi-fermi potential in
channel, drain current, and threshold voltage calculation. From the TCAD simulated
study, it is found that electrical properties in the channel region of DG TFET are
quite different from DG MOSFET. It is observed that the central channel potential
of DG TFET is not pinned to a fixed potential even after threshold voltage (as in
case of DG MOSFET occurs), but it initially increases and later on decreases with
increasing gate voltage. It is also observed that the threshold voltage extracted with
maximum transconductance method or linear extrapolation (LE) and electron quasi-
fermi potential of DG TFET are much higher than the DG MOSFET’s one. It is
also observed that just on-set of inversion is not sufficient condition for DG TFET
threshold voltage. These differences are explained in this paper with proper physics
reasoning.
Keywords BTBT · DG tunnel FET · Electric potential · Electron density ·

Electric field · Threshold voltage · Electron quasi-fermi potential · Drain current
1 Introduction
Nowadays, we are living the era of More-Moore and More than Moore, this is because
of the continuous down scaling of the benchmarked device—metal–oxide–semicon-
ductor field effect transistor (MOSFET) which inevitably leads to fundamental phys-
ical limits that can no longer be overcome by technology innovation alone. Limits
for conventional MOSFETs are: subthreshold leakage which limits the subthreshold
M. Yadav (B)
Department of Electronics and Communication Engineering, Malaviya National Institute of
Technology Jaipur, Jaipur, Rajasthan, India
792 M. Yadav
swing (SSMOSFET ) to 60 mV/dec at room temperature and higher OFF state device
current, I off , at lower device dimensions. Hence, new device concepts are needed.
The tunnel FET, based on the interband tunneling effect, has recently been subject
to a variety of both theoretical and experimental studies [1–4]. It is being demon-
strated that the device has several superior properties as compared to the benchmarked
MOSFET device.
TFETs have the possibility to overcome the drawbacks of MOSFET; therefore,
they are widely studied in recent years [5, 6]. TFETs may have SSTFET lesser than
MOSFET at room temperature because SSTFET is independent of temperature, there-
fore, TFETs can be a good option to operate at higher temperatures, without any
compromise on reliability. TFETs works on band to band tunnelling (BTBT) prin-
ciple; therefore, it could be scaled down without any degradation in its electrical
properties [7]. For example, I off and threshold voltage (V Th ) in MOSFET depend
on the gate length scaling but in case of TFET, these properties do not change with
further gate length scaling [8]. Because of its superior properties, TFETs can be used
in lower power applications such as SRAMs.
In this manuscript, electrical properties of DG TFET like electric potential, elec-
tric field, electron density, electron quasi-fermi potential, etc., in channel region
are comparatively studied with DG MOSFET. Threshold voltage for both devices
using maximum transconductance method or linear extrapolation method (LE) is
also presented, as explained in [9].
The organization of the paper is as follows: Sect. 2 describes the device structure
and simulation setup, Sect. 3 describes simulation results and discussions and Sect. 4
concludes the work.
2 Device Structure and Simulation Setup
The DG TFET device structure under study is presented in Fig. 1. Here, a transverse
cut AA is taken. Cut AA is used to explore the results at point A, which is termed
as surface or Si-oxide interface and at the mid of the channel, i.e., at mid of cut
AA, which is termed as the mid-channel point. The device parameters are listed in
Table 1. For DG MOSFET also same device parameters are used, except phosphorus
source doping is used.
Fig. 1 Double gate tunnel

FET device structure
Double Gate Tunnel FET Versus Double Gate MOSFET … 793
Table 1 Double gate tunnel

Parameter Value
FET device parameters
1. Gate length, L g 50 nm
2. Source doping, p-type 1020 /Cm2
3. Channel doping-p-type 1017 /Cm2
4. Drain doping. n-type 5 × 1018 /Cm2
5. Gate oxide, HfO2 , ox 22, o
6. Gate oxide thickness, T ox 3 nm
7. Silicon body thickness, T Si 10 nm
Two-dimensional TCAD device simulations are done using TCAD device simu-
lator Sentaurus [10]. To model BTBT, electric field dependent Kane’s [11]- model is
used for physics of the device. Since the source and drain regions are heavily doped
and tunneling current is strongly dependent on band gap, band-gap narrowing model,
OldSlotboom, is also included. High field saturation mobility model for electrons and
holes is included. Shockley Reed Hall (SRH) model is used as carrier recombina-
tion model. Simulations are done for various gate voltage and drain voltages. The
doping in the n-drain, p-source, and p-layer is kept constant at the optimum value and
doping profiles used are abrupt. For n-type impurity, phosphorus, boron as a trivalent
impurities are used and gate work function is set as 4.25 eV.
3 Simulation Results and Discussions
In this section, the simulated results for DG TFET at the surface, i.e., at point A and
at the mid of the channel are presented and compared with DG MOSFET results.
Electric properties at Silicon- HfO2 interface of AA cut, i.e., at point A are termed
as surface electric properties and at mid of AA, i.e., center of the channel are termed
as mid-channel properties.
3.1 Drain Current and Gate Threshold Voltage
Gate voltage, V gs versus drain current, I ds is shown in Fig. 2 for DG MOSFET

and DG TFET. In Fig. 3, gate threshold voltage versus drain voltage is shown for
both devices. We observed that DG TFET has higher threshold voltage than DG
MOSFET. Here, the threshold voltage is extracted using maximum transconductance
method, as given in [9], for MOSFETs. DG MOSFET has lower threshold voltage
as compared to DG TFET because in DG MOSFET after on-set of inversion device
becomes ON, while in case of TFET, on-set of inversion is not sufficient condition
to make the device ON, as given in [12]. A saturation in the drain current is seen
794 M. Yadav
Fig. 2 Drain current versus applied gate voltage for DG TFET and DG MOSFET
Fig. 3 Threshold voltage for DG TFET and DG MOSFET

Fig. 4 Energy band diagrams for DG tunnel FET and DG MOSFET at V ds =1.0 V
after a particular applied gate voltage, in DG MOSFET. This is because velocity

saturation is obtained by electrons, hence the (drift) current starts to saturate. While
in DG, TFET current is increasing with increase in applied gate voltage, V gs , this
is because after threshold voltage (which is higher than DG MOSFET), the band to
band tunneling (BTBT) generation rate, Gbtbt , increases at a rapid pace and this rate
is also controlled by applied drain voltage, therefore, as we increase drain voltage
the tunnel barrier width, wb decreases, as shown in Fig. 4; hence, average electric
field across the tunnel junction gets increased and we get
-larger current for larger drain voltages [13–15]. The tunneling generation rate is
given in Eq. (1).
2
E avg
G btbt = A Exp B E g3/2 /E avg (1)
Eg
where
A 3.5 × 1021 eV0.5 /cm s V2
B 22.6 × 106 V/cm-eV3/2
Eg Silicon energy band gap
E avg Average electric field along tunnel path
Here tunneling is occurring at source end denoted by circles in Fig. 1.
796 M. Yadav
3.2 Electron Quasi-Fermi Potential
In Fig. 5, electron quasi-fermi potential (eQFP) versus drain voltage is shown for
DG TFET. For DG MOSFET, eQFP although varies with applied drain voltage, V ds
its range is between 0 and 0.1 V even when V ds is varied between 0 and 1.0 V.
This is because the applied drain voltage sees a reversed biased p-n junction at the
drain/channel junction, therefore, whole drain voltage is applied across this junction
and channel eQFP remains zero always.
In case of DG TFET, p-n junction is formed at source/channel junction, therefore,
the applied drain voltage can be seen at the junction and this sets channel eQFP equal
to drain voltage, below threshold voltage as shown in Fig. 5. After gate threshold,
there is enough tunneling current to make the device ON and hence the channel/drain
region will act as a series connected resistance, since now channel is a resistive
region so there will be a potential drop equal to the drain current times of the channel
resistance. This will make the channel eQFP = V ds ; hence, there is a drop as seen at
higher gate voltages, and this pattern is also followed at higher drain voltages with
same physics.
But, in subthreshold regime, a hip hop change at higher drain voltages is seen.
This is because for higher drain voltages, the whole channel p-type (since inversion
is not set for lower gate voltages, as is clear from Fig. 6a) will work like a potential
barrier, therefore, a fall in eQFP seen for higher drain voltages, which after onset of
inversion follows the same physics as described in above paragraph.
Fig. 5 Electron quasi-fermi potential at AA cut for DG TFET

Fig. 6 Electric potential at surface and the mid-channel point for DG TFET and DG MOSFET.
a Electric potential at surface and the mid-channel point for DG TFET. b Electric potential at surface
and the mid-channel point for DG MOSFET
798 M. Yadav
3.3 Electric Potential
Surface electric potentials and mid-channel potentials for 50 nm gate length DG

TFET and DG MOSFET are shown in Fig. 6. For DG MOSFET, the surface potential
always monotonically increases with increasing gate voltage, while mid-channel
potential gets pinned to a fixed value for a fixed drain voltage, after threshold voltage
as shown in Fig. 6b. It is also observe that the surface potential is always greater than
mid-channel potential at any drain or gate voltage. This pattern of potential is also
seen in literature [16, 17]. The shape of the pattern of potential will not change with
the applied drain voltage, but for higher drain voltages, higher channel potentials
are obtained. Since source/channel and drain/channel junctions in DG MOSFET are
symmetric, therefore, there will be attraction of electrons from both source and drain
regimes hence the symmetric pattern in the mid-channel and surface potential is
found even if we change the drain voltage from 0.1 to 1 V.
In case of DG TFET as shown in Fig. 6a, surface potential and mid-channel
potentials for lower V ds (upto 200 mV) is seen constant, however, when we increase
the drain voltage V ds to 0.5 V and beyond this value, a monotonic decrease in -surface
and mid-channel electric potentials is seen, this property of DG TFET is different
from the potential pattern of DG MOSFET. At the same time, the surface potential is
always greater than the mid-channel potential in DG TFET (i.e., the parabolic shape
of the potential profile along AA cut line). However, it is observed that the potential
profiles for both DG TFET and DG MOSFET are parabolic in shape along AA cut
line, but still we are observing a different pattern in surface potentials and mid-
channel potentials in both the devices. This is because a crossover point in potential
is observed in case of DG TFET in oxide regime after gate voltage is further increased
beyond threshold voltage as is clear from Fig. 6a.
It is also observed that surface potentials and mid-channel potentials for DG TFET
are always much higher then surface and mid-channel potentials of DG MOSFET
respectively. Potentials for DG TFET increase much more at higher applied drain
voltages as compared to DG MOSFET, this happens because of sharper energy band
bending in case of TFET, as shown in Fig. 3.2.
The possible reason for this decrease in both surface and mid-channel potential
is the asymmetric device structure of TFET. Because of asymmetric device structure
when the applied gate voltage is increased (below threshold voltage), since there is no
tunneling from source to channel regime, therefore, the charge from drain/channel
junction is attracted at surface (this is because of drain/channel tunneling at very
low gate voltages or direct electrons are attracted from drain after inversion layer is
formed), since there is no tunneling below threshold voltage, both surface and mid-
channel potentials will increase monotonically as in case of DG MOSFET but after
gate voltage goes beyond threshold voltage of the DG TFET device, the tunneling
junction width at the source/channel junction gets lowered and more electrons will
come to the channel, and these are related by Eq. (2) for both DG MOSFET and DG
TFET. Since in DG MOSFET, there is zero eQFP hence surface potential increase,
but in case of DG TFET eQFP is non zero and decreasing after threshold voltage (as
mention in Sect. 3.2), but electron density increases due to excess electrons generated
from BTBT. This will decrease the channel potential as the gate voltage is increased.
In case of DG TFET, the nonzero eQFP will make its channel potential higher
than the DG MOSFET’s one.
3.4 Electric Field
In this subsection, total electric field (E = (E 2x + E 2y )0.5 ) versus applied gate voltage is
presented for various drain voltages, V ds for both devices. As shown in Fig. 7, surface
electric field for DG MOSFET is almost independent of the applied drain voltage,
while it decreases for DG TFET with applied drain voltage. While the mid-channel
electric field increases with applied drain voltage in both the devices.
Because of symmetry in Double gate structures, the y-component of electric field
will be zero at the mid channel and from x-component is more depending on the
applied drain voltage. Hence, as the drain voltage increases, a corresponding increase
in total electric field is solely due to E x -component in both the devices. In DG TFETs,
for lower gate voltages, the shape of the surface potentials as shown in [18] is not
constant in the mid channel, due to sharp energy band bending in the device, as
shown in Fig. 3.2, therefore higher mid-channel electric field is seen for lower gate
voltages for DG TFETs. In case of surface electric field, a stronger E y —component
for lower drain voltages makes the total electric field higher. As the drain voltage is
increased, it will make E x stronger and E y will become a bit weaker, since higher
the drain voltage means, we need to apply higher gate voltage to make the surface
potential constant, as shown in [12], therefore, the curves are shifted downwards for
surface potential.
3.5 Electron Density
In this subsection, electron density for both devices is comparatively studied.

Analytical expression for electron density in the channel is given by Eq. (2).
2
n i,eff ψ(x,y)−φFn (x)
n(x, y) = e VT
(2)
Na
where
n(x, y) channel electron density/cm3
ni,eff effective intrinsic concentration of channel region per cm3
Na channel doping per cm3
ψ(x, y) channel potential at any point (x, y), V
ϕ F,n (x) channel eQFP at x, V
800 M. Yadav
Fig. 7 Electric field at surface and the mid-channel point for DG TFET and DG MOSFET. a Electric
field at surface and the mid-channel point for DG TFET. b Electric field at surface and the mid-
channel point for DG MOSFET
VT thermal voltage, V
It is clear from Fig. 8, DG MOSFET’s electron densities are almost independent
on applied drain voltage but DG TFET electron densities are strong function of drain
voltages. This is because of higher quasi-fermi potentials seen in DG TFETs. As
Fig. 8 Electron density at surface and the mid-channel point for DG TFET and DG MOSFET.
a Electron density at surface and the mid-channel point for DG TFET. b Electron density at surface
and the mid-channel point for DG MOSFET
802 M. Yadav
shown in Sects. 3.2 and 3.3, an increase in drain voltage increases channel potentials
and eQFP. But, we see, increase in channel potentials is lesser as compared to eQFP,
hence from Eq. (2), electron density of DG TFET decreases with increasing drain
voltages. As the applied gate voltage is increased, more electrons will be generated
(due to BTBT) at the tunneling junction and will be attracted by vertical electric
fields near the surface hence electron density will always increase with gate voltage,
but mid-channel electron density will always be lesser than surface density.
4 Conclusion
Comparative study based on TCAD simulations for DG TFET and DG MOSFET

devices with 50 nm gate length is carried out in this paper. It is found that although the
channel (transverse) potential shape is parabolic in nature, its quality and quantity
are different for both devices. In DG TFET, we found that the surface as well as
mid-channel potential decreases after a particular gate voltage, which is different
for different drain voltages. Unlike DG MOSFET, the mid-channel potential is not
pinned to a fixed value after threshold voltage, but it decreases with increasing V gs.
Unlike DG MOSFET, a nonzero e—quasi-fermi potential is seen for DG TFET,
which increases with applied drain voltage, but decreases at higher gate voltages.
Gate threshold voltage for tunnel FETs is very high as compared to DG MOSFET
and it is not depending on the on-set inversion charge theory. DG TFETs have very
high I ON/I OFF, ratios, and steeper subthreshold swing. Electron density and electric
field properties are also different for both the devices.
References
1. Anghel C, Gupta A, Amara A, Vladimirescu A (2011) 30-nm tunnel FET with improved
performance and reduced ambipolar current. IEEE Trans Electron Devices 58(6):1649–1654
2. Boucart K, Ionescu AM (2008) A new definition of threshold voltage in tunnel FETs. Solid-State
Electron 52(9):1318–1323
3. Gupta SK, Kulkarni JP, Datta S, Roy K (2012) Heterojunction intra-band tunnel FETs for
low-voltage SRAMs. IEEE Trans Electron Devices 59(12):3533–3542
4. Boucart K, Ionescu AM (2007) Double-gate tunnel FET with high-k gate dielectric. IEEE
5. Born M, Bhuwalka KK, Schindler M, Abelein U, Schmidt M, Sulima T, Eisele I. Tunnel FET:
a CMOS device for high temperature applications. In: 2006 25th international conference on
microelectronics 14 May 2006, pp 124–127
6. Vandenberghe WG, Verhulst AS, Sorée B, Magnus W, Groeseneken G, Smets Q, Heyns M,
Fischetti MV (2013) Figure of merit for and identification of sub-60 mV/decade devices. Appl
Phys Lett 102(1):013510
7. Boucart K, Ionescu AM (2007) Length scaling of the double gate tunnel FET with a high-k
gate dielectric. Solid-State Electron 51(11–12):1500–1507
8. Li YC, Zhang HM, Hu HY, Zhang YM, Wang B, Zhou CY, Lou YL (2013) Effect of high-
k material on gate threshold voltage for double-gate tunnel FET. Appl Mech Mater 275–
277:1984–1987 (2013)
9. Ortiz-Conde A, García-Sánchez FJ, Muci J, Barrios AT, Liou JJ, Ho CS (2013) Revisiting
MOSFET threshold voltage extraction methods. Microelectron Reliab 53(1):90–104
10. Sentaurus TCAD Manual, Mar 2010
11. Kane EO (1960) Zener tunneling in semiconductors. J Phys Chem Solids 12(2):181–188
12. Anand B, Dasgupta S (2012) A TCAD approach to evaluate channel electron density of double
gate symmetric n-tunnel FET, pp 577–581 (2012)
13. Yadav M, Bulusu A, Dasgupta S (2013) Two dimensional analytical modeling for asym-
metric 3T and 4T double gate tunnel FET in sub-threshold region: Potential and electric field.
Microelectron J 44(12):1251–1259
14. Bulusu A, Dasgupta S (2015) Insights into channel potentials and electron quasi-Fermi
potentials for DG tunnel FETs. J Semicond 36(1):014005
15. Yadav M, Bulusu A, Dasgupta S (2015) Super-threshold semi analytical channel potential
model for DG tunnel FET. J Comput Electron 14:566–573
16. Taur Y (2000) An analytical solution to a double-gate MOSFET with undoped body. IEEE
17. Taur Y (2001) Analytic solutions of charge and capacitance in symmetric and asymmetric
double-gate MOSFETs. IEEE Trans Electron Devices 48(12):2861–2869
18. Bardon MG, Neves HP, Puers R, Van Hoof C (2010) Pseudo-two-dimensional model for double-
gate tunnel FETs considering the junctions depletion regions. IEEE Trans Electron Devices
57(4):827–834
Wireless Access Support for Distribution
Management System (DMS) to Microgrid
Communication in Power Network
N. Himabindu, Rajashekar P. Mandi, Santoshkumar Hampannavar,

and Swapna Manasani
Abstract With indispensable and rapid ingress of renewable energy sources in

power systems, the communication between the distributed energy sources has
become very crucial to have a reliable distribution management system (DMS).
In this paper, wireless access support is proposed for Distribution Manage-
ment System (DMS) and microgrid control center (MGCC) using WiMAX protocol.
MATLAB/Simulink was used to model the WiMAX physical layer. The bit error
rate (BER) v/s signal-to-noise ratio (SNR) curves were plotted and promising results
were obtained.
Keywords Microgrid · Distribution management system (DMS) · Distribution

generation (DG) · Distributed energy resource (DER)
1 Introduction
The policy of deregulation in electric power systems has allowed participation of

private players in generation of power and this resulted in rapid penetration of
distributed energy resources (DER) in the power grid at medium and low-voltage
networks. The distribution system plays a vital role in smartgrid/microgrid which
includes: (i) DER integration to the grid; (ii) load demand control; and (iii) efficient
usage of network assets. Distribution system (DS) is too complex in nature and it is
difficult to control, monitor, and handle it. The difficulties associated with DS are:
N. Himabindu (B) · R. P. Mandi · S. Hampannavar

School of Electrical and Electronics Engineering, REVA University, Bengaluru, India
R. P. Mandi
S. Hampannavar
S. Manasani
National Institute of Technology Silchar, Silchar, India
806 N. Himabindu et al.
Graphical User Interface
Syste m monitoring, System operation,

Ne twork topology Outage
Network mangeme nt,
and state
configuration, Fault de te ction,
monitoring,
DER state Volt/VAR Fault diagnosis,
monitoring, management, Fault location,
Syste m and DER ope ration, Supply re storation,
Fie ld C re w O pe rator training Eve nt analysis &
Manage me nt simulator re cording
Modeling & Analysis Tools
SCADA
Fig. 1 Components of DMS
(i) often meshed but operate radically at times (because of the faults and expansion
of the network) and (ii) difficulty in predicting time varying load demand. Distribu-
tion management system (DMS) is a set of applications adopted by the distribution
network operators (DNO) to control and enhance the performance and manage its
complexity. Figure 1 gives the key components of the DMS.
DER greatly impacts network performance and depends on DER-DS interface. It
can be connected to the grid through an power electronic interface/microgrid (MG).
2 Microgrid
A MG is a network of small modular distributed generation (DG), storage, and

loads. Microgrid central controller (MGCC) acts as an interface between DMS and
microgrids. MG integration with DMS through MGCC [1] is shown in Fig. 2.
There are two operating modes: (i) grid-connected and (ii) islanded mode.
Grid-connected mode: In grid connected operating mode, the power is supplied
to loads connected to MG and excess power is transferred back to the grid. The total
MG system load is supported by MG and the grid.
Wireless Access Support for Distribution Management … 807
MG1
MGA1
MV LV
MG2
DistribuƟon
Management
System (DMS) MGA2
PCC
WiMAX MGCC
MGAn
MGn
Fig. 2 MG integration to DMS via MGCC
VPCC = Vgrid
Pload = PDG + Pgrid
Q load = Q DG + Q grid
where
VPCC is voltage at point of common coupling
Vgrid is grid voltage
PDG , Pload ,Pgrid , Q load , Q DG , Q grid are inverter, grid, load active/reactive powers in
kW and KVAr, respectively.
If DGs are generating excess power than the MG system load, then the excess
power will be transferred to the grid.
PDG = Pload + Pgrid
Q DG = Q load + Q grid
Standalone/Islanded mode:
In this mode, grid does not have any role to play and MG starts energizing the
small network segment and supplies the power to the critical loads to be addressed.
The voltage and frequency levels are maintained as per the IEEE 1547-2003
standard and PCC voltages are equal to the inverter front end voltage.
VPCC = VDG
Pload = PDG
Q load = Q DG
If the MG system load is greater than the DGs generated power, then the system
will enter the unstable region.
Pload >PDG
Q load >Q DG
The recent literature elucidates DG integration issues, MG integration and control

strategies, long-term evolution (LTE) protocol for MGCC to DMS communication,
grid control center to the aggregator, and an aggregator of EV communication [2–5].
The MGCC to DMS communication is very crucial in deciding the MG integra-
tion to the grid. WiMAX protocol for MGCC’s communication with DMS in the
power network has been proposed in this paper. Also, the corresponding path loss
models are simulated. The physical layer of the WiMAX downlink is modeled in
MATLAB/SIMULINK. The curves pertaining to bit error rate (BER) and signal to
noise (SNR) are also plotted.
3 WiMAX Protocol
WiMAX standard works in the frequency range of 2–6 GHz. The standard uses
OFDM, 256-FFT, and handles 70 Mbps of data flow and coverage area of up to 50 km.
The coverage area is 5–7 km at 7 Mbps. An OFDM system has three subcarriers: (i)
pilot, (ii) data, and (iii) null.
The transmission of physical layer data vectors requires (randomization, inter-
leaving, RS coding, and error correction) for modulation schemes (M-QAM) and for
distinct coding rates (convolution coding and concatenated Reed–Solomon). OFDM
symbols, demodulate waveform, deinterleave decode (RS and Viterbi), and bit scram-
bling process inversion [6–9] are used to retrieve the receiver data. WiMAX physical
layer is shown in Fig. 3.
The fixed WiMAX protocol is chosen for the communication and the parameters
used for simulation are shown in Table 1.
Signal-to-noise ratio (SNR) is represented as
SNR = Psignal (dB) /Pnoise (d B)

Data
Error CorrecƟon MQAM OFDM

RandomizaƟon
Coding Modulator Modulator
CommunicaƟon
Channel
Error CorrecƟon MQAM OFDM

DerandaomizaƟon
Decoding Demodulator Demodulator
Data
Fig. 3 WiMAX physical layer block diagram
Table 1 IEEE 802.16-2004

Parameters Value
(fixed WiMAX)
Frequency band 2–11 GHz fixed
OFDM carriers OFDM: 256
OFDMA: 2048
Modulation 16QAM, 64QAM, QPSK
Channel bandwidth (1, 4, 8, 12) × 1.25 MHz
(1, 2, 4 ,8) × 1.75 MHz
8.75 MHz
Transmission rate 1–75 Mbps
Duplexing TDD/FDD
In dB it is represented as
SNRdB = Psignal (d B) − Pnoise (d B)
MATLAB/SIMULINK is used to model and simulate the physical layer of WiMAX

protocol. Figure 4 shows the plotted curves for BER v/s SNR. In order to achieve
BER of 10−2 both BPSK ½ needs 7.5 dB, QPSK ½ needs 12 dB, QPSK ¾ needs
about 15 dB, and 16QAM ½ needs about 16.5 dB. The QPSK ½ and 16 QAM ½ are
less susceptible to errors than other modulation schemes [10–13].
The transmitter and receiver antenna heights are 30 m and 10 m, respectively.
Figure 5 shows the results of transmitter power which is 43 dBm. The SUI model
0 BER v/s SNR

10
-1
10
BER
-2
10
BPSK1/2
QPSK1/2
QPSK3/4
-3 16-QAM1/2
10
-4
10
0 2 4 6 8 10 12 14 16 18
SNR
Fig. 4 BER v/s SNR for GCC to aggregator Link
Fig. 5 BER v/s SNR for GCC to aggregator link

shows a lower path loss of 142 dB and 162 dB, respectively, for suburban and
urban environments. The COST WI, COST HATA, and ECC-31 show, respectively,
a path loss of 162 dB, 158 dB, and 164 dB, respectively. Ericsson model depicts
comparatively higher path loss of 204 dB for suburban environment but interestingly
in case of urban environment, it shows path loss of 160 dB which is less when
compared to all other models. Compared to other models in urban environment, the
COST HATA and COST WI, respectively, exhibit higher path loss of 176 and 178 dB
for 10 m height receiver antenna.
5 Conclusion
The wireless access support for MGCC to DMS communication incorporating

WiMAX protocol is discussed. For various bandwidth and path loss models BER v/s
SNR graphs were plotted. This study helps the researchers associated in developing
wireless infrastructure model for MG communication in distribution network.
References
1. Ekanayake J, et.al (2012) Smart grid technology and applications, 1st edn. Wiley Ltd
2. Kumar S, Udaykumar RY (2015) Modeling and comprehensive analysis of WiMAX protocol
for grid control center to aggregator communication in V2G. Int J Renew Energy Res (IJRER)
5(1):133–138
3. Kumar S, Udaykumar RY (2013) IEEE 802.16-2004 (WiMAX) protocol for grid control center
and aggregator communication in V2G for smart grid application. In: IEEE international
conference on computational intelligence and computing research
4. Kumar S, Udaykumar RY (2013) IEEE 802.16 protocol for grid control center and aggre-
gator communication in V2G for smart grid application. In: IEEE international conference on
computational intelligence and computing research, Madurai, (IEEE-ICCIC), India
5. Ali-Yahiya T (2011) Understanding LTE and its performance. Springer Science and Business
Media, Heidelburg, Germany
6. Kumar S, Udaykumar RY (2015) Performance analysis of LTE protocol for EV to EV commu-
nication in vehicle-to-grid (V2G). In: 2015 IEEE 28th Canadian conference on electrical and
computer engineering (CCECE), Halifax, Canada, pp 1–6
7. Kumar S, Udaykumar RY (2013) Long term evolution protocol for grid control center and
aggregator communication in V2G for smart grid application. In: IEEE international conference
on computational intelligence and computing research, Madurai, India
8. Hamid M, Kostanic I (2013) Path loss models for LTE and LTE–a relay stations. Univers J
Commun Netw
9. Zyren J (July, 2007) Overview of the 3GPP long trem evolution physical layer. Freescale
Semicond
10. Hampannavar S, Chavhan S, Yaragatti U, Naik A (2017) Gridable electric vehicle (GEV)
aggregation in distribution network to support grid requirements: a communication approach.
Int J Emerg Electr Power Syst
11. Santoshkumar H, Udaykumar RY, Selvakumar S, Teshale A (2018) A framework for microgrid
to distribution management system (DMS) communication in smart distribution network. In:
IEEE international conference on computational intelligence and computing research
12. Kumar S, Chavhan S, Udaykumar RY (2019) A stochastic model based on markov chain
to support vehicle-to-grid (V2G) operation in smart distribution network. Int J Emerg Electr
Power Syst
13. Kumar S, Chavhan S, Swapna M, Udaykumar RY (2020) Electric vehicle traffic pattern analysis
and prediction in aggregation regions/parking lot zones to support V2G operation in smart grid:
a cyber physical system entity. Int J Emerg Electr Power Syst
Implementation of a Charge Controller
for Hybrid Electric Vehicle
Gautam and Brajagopal Datta
Abstract A hybrid electric vehicle (HEV) is a vehicle which is the combination

of more than one supply source. In this HEV, two batteries each of 24 Vs are used,
one battery is connected to load which provide power to drive motor, and another
battery is connected to charging system which takes power for charging. In the
present work, the main controller of HEV is based on Arduino software which is
used by the microcontroller. The system uses a battery swap circuit which increases
the performance of the electric vehicle. Battery swap circuit measures the voltage
level of both batteries, if the voltage level of load connected battery gets lower
than the changing connected battery, then the connection of both the batteries get
interchanged. The recharging system contains BLDC and solar panel which provides
constant 24 Vs to the battery.
Keywords Hybrid electric vehicle · Brushless DC generator · Solar panel ·

Battery swap circuit · Microcontroller · Controller behavior · Battery management
1 Introduction
In present time, electric vehicles are the need for the transportation system. The
demand of non-renewable energy (petrol, diesel, coal, etc.) is increasing day by
day, so it is necessary to find out other energy resources. Renewable energy like
solar is the most important and available in the large amount on earth which is very
useful for hybrid electric vehicles. It provides an alternative source of energy which
is used to feed extra energy to charge the battery or directly used to feed motor.
Investigation of PV-Fed Hybrid Electric Vehicles with Various Power Electronic
Circuits (2017) [1]. In this paper, the hybrid electric vehicle is based on conventional
Gautam · B. Datta (B)

Gautam
814 Gautam and B. Datta
and non-conventional energy sources. It is using solar PV modules and petrol- or

diesel-based engine, which is not fully renewable energy-based electric vehicle. The
demand for non-conventional energy is increasing day by day, so it is necessary to
find out another alternative source of energy which should be renewable based and
it must be eco-friendly.
Energy Management in a Fuel Cell-based Hybrid Electric Vehicle using a Fuzzy
Logic Approach (2017) [2]. An energy management strategy is provided for an HEV.
This is based on a fuel cell as well as a Li-ion battery. These energy management’s,
based on the fuzzy logic approach, aim is to increase the performances of the hybrid
electric vehicle. Here, the fuzzy logic-based controller has three inputs: the power
error, load power, and battery.
A Rapid Charging System with Bidirectional Power Flow for Plug-in Hybrid
Electric Vehicle (HEV) (2017) [3]. This paper shows the capability of the rapid battery
charging technology with bidirectional power flow for HEV. It is used rectangular
phase shift method with PWM (RPPWM) for zero voltage switching based on the
dual-active bridge.
It controls the capability of power flow in both directions and reduces the
complexity. Here, boost converter is used to regulate the constant DC output voltage
which also improves the power factor. Here, boost converter plays another important
role, and it provides the bidirectional power flow.
Real-time Power Hardware-in-the-Loop Emulation of a Parallel Hybrid Electric
Vehicle (2017) [4]. This paper shows the control and coordination of several sources
of the hybrid electric vehicle. This paper put light on a real-time emulator for parallel
hybrid electric vehicle. Here, every system is modeled and controlled to balance the
performance and control as per torque and speed.
Energy Management And Strategies For Hybrid Electric Vehicle Using Photo-
voltaic, Ultra-capacitor, And Battery (2017) [5]. In this paper, three sources are using
PV, battery, an ultra-capacitor. During daytime, when solar panel provides maximum
voltage, then at that point of the time, PV runs the motor as well as it also charges
the battery and capacitor. When the power output of the PV get reduced, then battery
and an ultra-capacitor provide power to the motor.
Here, regenerative braking technology is also applied which is very useful for
hybrid electric vehicles. In this method, when the brakes are applied, then at that
point in time, motor acts as a generator and starts feeding supply to the battery and
an ultra-capacitor. As compared to others, this type of electric vehicle provides high
performance, and battery backup will be improved. But, this is not a permanent
solution. If the vehicle is running in the night, then the solar panel will be unable
to provide power, and then, after sometime, battery will be low which is the major
disadvantage of the electric vehicle.
From literature survey it is evident that more than one input sources can be used to
drive an electric vehicle [7–11]. The sources are internal combustion engine (ICE)
and electric motors. But, both the sources are separately connected to the motor.
When battery gets discharged, ICE starts providing supply to the motor.
The pollution from the vehicles is one of the main cause of air pollution [12]. Day
by day, the graph of pollution in traffic area is increasing very sharply. To prevent this,
Implementation of a Charge Controller for Hybrid Electric Vehicle 815
the electric car must run in the city, so that the maximum pollution will be reduced.
The amount of greenhouse gases as well as the global warming will be decreased.
2 Working of Proposed Recharging System
To increase the overall performance of the hybrid electric vehicle, it is necessary to

increase the battery performance or its backup. According to this charging system,
this hybrid vehicle uses two battery, so the charging of one battery and discharging
of second battery which is connected to load occurs at the same time. This controller
has recharging unit which is divided into two phases. In phase 1, two input sources
like solar panel and brushless DC generator are connected. Both sources are again
connected to relay via voltage sensor, and this voltage sensor is connected to the
microcontroller, as shown in Fig. 1.
Phase 1
As shown in Fig. 2, another microcontroller is used in phase 2 with a voltage
sensor, relay, and boost converter. The coming voltage is directly connected to relay
Solar
Panel
BLDC
Generator
Fig. 1 Circuit diagram of phase 1

Fig. 2 Circuit diagram of phase 2
via voltage sensor. If the coming voltage is less than or equal to 23 V, then microcon-
troller feeds the voltage to boost converter via a relay. If the voltage is 24 V or more,
then microcontroller bypasses the boost converter and directly feeds to the battery.
The ATMEL 8-bit microcontroller contains 32 kB flash memory with the capa-
bility of read-while-write, 1 kB of EEPROM, 2 kB of SRAM, 23 general purpose
I/O lines. The device operates between 1.9 and 5.5 V. The programming is written
inside the microcontroller in such a way that, which input sources provide the higher
voltage, the microcontroller will proceed that voltage source to next level via a relay.
If both the incoming sources provide the same amount of voltage, then at that point
of time, microcontroller will proceed the voltage of BLDC generator because the
preference is given to the BLDC generator in the programming. Here, BLDC gener-
ator and solar panel both deliver voltage which is shown in Fig. 3. The value of
the voltage of both sources continuously varies according to time. Here, the micro-
controller compares both voltages and selects the higher voltage with the help of
relays.
The response of controller with different conditions.
Condition 1
If the vehicle is in running condition (During Day).
Here, BLDC generator and solar panel both deliver voltage which is shown in
Fig. 3. The value of the voltage of both sources continuously varies according to time.
Here, the microcontroller compares both voltages and selects the higher voltage with
the help of relays.
Condition 2
If the vehicle is in stationary condition (During Day Time).
Fig. 3 Different input

values during running
conditions (day)
As shown in Fig. 4, solar panel delivers supply, and output of BLDC generator is
almost zero. The supplied value of BLDC generator is less than solar panel, so the
microcontroller selects only higher supply value.
There are two pairs of the relay which are connected with solar panel and BLDC
generator, and these relays are controlled by the microcontroller as per programming
is installed in it. Here, V 1 stands for the supply of BLDC generator, and V 2 stands
for solar panel’s supply. Relay 1 and 2 are connected with BLDC generator, and relay
3 and 4 are connected to the solar panel. If the conditions (V 1 > V 2 and V 1 = V 2)
are as given in Table 1, then relay 1 and 2 will on and relay 3 and 4 will be off and

values during a traffic jam or
in parking conditions
Table 1 Switching mode of the relay according to the supply voltage

Condition Relay 1 Relay 2 Relay 3 Relay 4
V1 > V2 On On Off Off
V1 = V2 On On Off Off
V1 < V2 Off Off On On

values during running
conditions (night)
vice versa.
Condition 3
If the vehicle is in running condition (During Night Time).
In this condition, only BLDC generator provides supply, and the output of the
solar panel is almost zero as shown in Fig. 5.
3 Battery’s Connection Swapping System
This technology is useful for HEV which increase the battery backup, range, and
the performance. Here, one battery, which is connected to the charging system, and
another battery, which is connected to a load, both are also connected to the micro-
controller via voltage sensor. The voltage sensor measures the voltage level of the
batteries and sends the information to the microcontroller. Here, the microcontroller
will decide that which battery is going to connect the load and which will connect to
the recharging system via a relay. The circuit diagram of battery swap technology is
given in Fig. 6.
The switching mode of the relay of battery swap circuit is controlled by the
microcontroller as given in Table 2.
The output of recharging system with respect to variable input source for the battery
is 24 V which is shown in Fig. 7. The image of a hardware implementation of battery
swap system is shown in Fig. 8. The image of a hardware implementation of battery
swap system is shown in Fig. 9. This system is controlled by Arduino.
• Recharging system and battery connection swap method increases the perfor-
mance and efficiency of the HEV.
• The recharging system provides constant output to the battery as shown in Fig. 9.
Fig. 6 Circuit diagram of batteries swap system
Table 2 Switching mode of the relay with respect to the voltage level of batteries
Condition R1 R2 R3 R4 R5 R6 R7 R8
V1 > V2 Off On On Off Off On On Off
V1 = V2 Off On On Off Off On On Off
V1 < V2 On Off Off On On Off Off On
• It increases the range of HEV.
5 Conclusion
The main source of pollution is the internal combustion-based vehicles which is the
main cause of air pollution. The traffic area is more polluted. So, it is prevented by
electric vehicle. So, the level of pollution will be low. Both greenhouse gases, as well
as global warming, will decrease, and it also improves the quality of the environment.
Due to addition of recharging system with battery, the overall range of the vehicle
Fig. 7 Output of recharging system to feed battery
Fig. 8 Image of batteries connection swap system
will increase. The solar panel is used as a source in this system. As recharging system
provides the sufficient amount of power to the battery in small period of time, it also
improves the positive characteristics as well as driving range of the whole system.
Hence, with this phenomenon, a large amount of energy saving is possible. The cost-
effectiveness of this system is increased due to large number of components used,
but the cost impact of recharging system will increase fuel economy and decrease
harmful emissions. It also saves the overall cost because of long life cycle.
PHASE-2
CIRCUIT
PHASE-1
CIRCUIT
Fig. 9 Image of recharging system
References
1. Kaleeswari M, Madheswaran M, Vijayakumar K (2017) Investigation on PV fed hybrid electric

vehicles with various power electronic circuits Kaleeswari. In: IEEE international conference
on advances in electrical technology for green energy, pp 49–55
2. Saib S, Hamouda Z, Marouani K (2017) Energy management in a fuel cell hybrid elec-
tric vehicle using a fuzzy logic approach. In: The 5th international conference on electrical
engineering
3. Naresh M, Bai VS, Pandey AK (2017) A rapid charging system with bi-directional power flow
for plug-in hybrid electric vehicle. In: International conference on innovations in power and
advanced computing technologies, pp 01–05
4. Kaarthik RS, Pillay P (2017) Real-time power hardware-in-the-loop emulation of a parallel
hybrid electric vehicle drive train. IEEE transportation electrification conference
5. Liu M, Huang J, Cao M (2017) Handling stability improvement for a four-axle hybrid electric
ground vehicle driven by in wheel motors. IEEE Access
6. Pancholi G, Yadav DK, Chaturvedi L (2017) Energy management strategies for hybrid electrical
vehicle using PV, ultracapacitor, and battery. IEEE transportation electrification conference
7. Senthilnathan A, Manohar R, Mohanavel J, OmeshHeman Kumar A, Saravana Kumar R (2017)
Smart hybrid electric vehicle. In: IEEE international conference on Innovations in green energy
and healthcare technologies, pp 01–06
8. Shaikh MZ, Kodad SF, Jinaga BC (2008) Performance analysis of Piezoresistive MEMS for
pressure measurement. J Theor Appl Inf Technol 227–231
9. Yan J, Li C, Xu G, Xu Y (2009) A novel on-line self-learning state-of-charge estimation
of battery management system for hybrid electric vehicle. IEEE transportation electrification
conference, pp 1161–1166
10. http://www.circuitstoday.com/working-of-electric-cars
11. https://www.arduino.cc/
12. https://www.wikipedia.org/
Series-Stacked Non-parallel Multipath
Differential Inductor for C Band
Applications
Sunil Kumar Tumma and Bheema Rao Nistala
Abstract Radio frequency-integrated circuits built on silicon require an inductor

with high-quality factor and miniaturized on-chip area. This paper introduces series-
stacked non-parallel multipath differential inductor to resolve this problem. The
proposed inductor decreases the parasitic capacitance among the adjoining metal
layers and also reduces the conductor’s current crowding effects at high frequencies,
leading to higher quality factor and self-resonance frequency. The proposed inductor
manages to achieve a raise of 35% in quality factor and 25% gain in self-resonance
frequency as compared to conventional series-stacked multipath inductor occupying
the equivalent on-chip area for uniform inductance. High-frequency structural simu-
lator (HFSS) is used to design and simulate the proposed inductor using 0.18 µm
CMOS technology.
Keywords Quality factor · Self-resonance frequency · Series stack · Multipath
1 Introduction
On-chip inductors are used in wireless communication systems, micro-

electromechanical systems (MEMS), and in few RFIC circuits such as mixers,
voltage-controlled oscillators, and low-noise amplifiers. With an increased demand
for personal mobile communications, RFIC requires high-performance on-chip
inductors to reduce the total system size and assembly cost.
A CMOS radio frequency (RF) circuit design benefits from one or more on-
chip inductors having a high-Quality factor (Q), a small occupied on-chip area,
and a high self-resonance frequency (f SR ). At high frequency of operation, the skin
and proximity effects and substrate losses degrade the performance of the on-chip
S. K. Tumma (B) · B. R. Nistala

B. R. Nistala
824 S. K. Tumma and B. R. Nistala
inductor [1]. Several methods were proposed to overcome these drawbacks which
include implementation of the inductor in multilayers [2], usage of high thickness
and high conductivity metal strip to form an inductor [3], usage of patterned ground
shield underneath the inductor [4], and employing variable width in inductor turns
[5]. Multipath inductor lowers the proximity effect and increases the Q without
altering inductance (L) [6, 7]. Symmetric (differential) inductor structure which is
excited differentially is less affected by substrate parasitics which improves both Q
and f SR for miniaturized on-chip area [8]. The symmetric trace differential-stacked
spiral inductor (SDSSI) further enhances the Q and f SR of the symmetrical inductor;
however, it is unable to achieve higher inductance value [9]. Series-stacked spiral
consists of two or more spirals overlaid one on top of the other increases the induc-
tance value to a large extent for the miniaturized on-chip area for small degradation
in Q value [10, 11]. Series-stacked multipath spiral inductors with crossover inter-
connection method improves the Q and f SR without affecting the inductance for
the miniaturized on-chip area as the current crowding effects are very less [12].
The performance of the series-stacked multipath inductors is further enhanced by
using the proposed series-stacked non-parallel multipath differential inductor. In
conventional series-stacked multipath inductors, the divided paths (segments) are
parallel which forms the parasitic capacitance. However, in the proposed inductors,
the segments are not parallel that decreases the parasitic capacitance and improves
the Q and f SR of the inductor to a significant extent. The proposed inductor is centered
in C band which ranges from 4–8 GHz. C band is used for frequency allocation for
satellite communications, cordless phones, and surveillance systems.
The construction of single-layer multipath symmetrical inductor is discussed
in Sect. 2. The construction of the proposed series-stacked non-parallel multipath
symmetrical inductor is depicted in Sect. 3. Section 4 describes simulation results,
and conclusions are given in Sect. 5.
2 Multipath Differential Inductor
In a planar multipath differential inductor, the conductor width (metal strip) is parti-
tioned into various current paths based on skin depth and minimum design rule as
shown in Fig. 1. The path width is selected as less than the skin depth of the conductor.
At very high frequencies (>1 GHz), the skin depth for good conductors becomes
tiny, and it is less than a micrometer. The skin depth (δ) is expressed as:

2ρ
δ= (1 + (ρωε)2 + ρωε (1)
ωμ
where ρ is the resistivity, ω is the angular frequency, and ε and μ are the permittivity
and permeability of the free space, respectively.
Series-Stacked Non-parallel Multipath … 825
Fig. 1 Single-layer
multipath differential
inductor
The multipath structure reduces the current crowding effects and contributes to
the high Q [6]. The conductor is partitioned into multiple segments. The width of
the segment is selected as below or equal to the skin depth.
3 Series-Stacked Non-parallel Multipath Differential

Inductor
The proposed series-stacked non-parallel multipath differential inductor is shown in

Fig. 2. The top and bottom layer each consists of one differential inductor, and these
Fig. 2 Proposed
series-stacked non-parallel
multipath differential
inductor
Fig. 3 a Partition of conductor width b non-parallel metal layers
are connected in series using vias. As shown in Fig. 3a, the conductor width is 10 µm,
and it is partitioned into two segments, each segment has 2 µm width and spacing
between the segments is 6 µm. The top layer consists of two segments denoted as
S 1 and S 2 , and the bottom layer consists of two segments denoted as S 3 and S 4 as
shown in Fig. 3a. The segments in the top and bottom layers are not parallel to each
other as shown in Fig. 3b.
The non-parallel segments arrangement reduces the parasitic capacitance, and the
multipath technique lowers the current crowding effect. Hence, the proposed inductor
achieves high Q, high f SR for almost same inductance value compared to existing
series-stacked multipath inductor. The Q is often used to give an indication of its
performance in an RF or another circuit is represented as:
ωL s
Q = (2)
RS
L s is the inductance, and Rs is the D.C. series resistance of the inductor. The
multipath structure also reduces the D.C. resistance (Rs ) of the inductor which leads
to a desirable increase in the inductor’s Q. Self-resonance frequency (f SR ) is charac-
terized as the point where the inductor ceases to offer inductive reactance, and it is
represented in (3). The bandwidth of any designed application depends on f SR .
1
f SR = (3)
2π L s Ceq
From (3), C eq is the sum of interlayer capacitance (C s ), metal to substrate capac-

itance (C p ), inter-winding capacitance (C in ) which are shown in Fig. 4a, b. The
proposed inductor reduces the C s and C p , which leads to increase in f SR . The perfor-
mance of the proposed inductor is improved by modifying the number of segments,
turn width, and the spacing between the turns throughout the spiral length.
Fig. 4 a Interlayer
capacitance (C s ) and metal
to substrate capacitance (C p )
b inter-winding capacitance
(C in )
The proposed inductor is constructed for an outer diameter of 180 µm and has 3
turns. The design and simulation of the proposed inductor is performed in HFSS
using 0.18 µm CMOS technology. The differential impedance [13] of a differential
inductor is obtained by using (4). The simulated L and Q values for the proposed
symmetrical inductor are obtained by using (5) and (6).

Y11 + Y12 + Y21 + Y22
Z Diff = (4)
Y11 Y22 − Y12 Y21
Im(Z Diff )
Q Diff = (5)
Re(Z Diff )
Im(Z Diff )
L Diff = (6)
2 ∗ pi ∗ freq
The Q and inductance plots of the proposed inductor are shown in Figs. 5 and 6.
From the results shown in Fig. 5, the proposed inductor with non-parallel segments
achieves 25% improvement in f SR for the same inductance value over conventional
series-stacked multipath inductor. From Fig. 6, it is noted that the Q-factor of the
proposed inductor is 33% greater than the conventional series-stacked multipath
inductor. The maximum Q frequency (f Qmax ) of the proposed inductor is 4.8 GHz.
The proposed inductor f SR is 11.5 GHz which covers the entire C band.
Q-factor
inductance
The simulation results for series-stacked multipath inductor, conventional series-

stacked inductor, and the proposed inductor for an on-chip area of 180 µm are
tabulated in Table 1.
5 Conclusion
In this paper, series-stacked non-parallel segmented multipath differential inductor

has been proposed and simulated using 0.18 µm CMOS Technology. The proposed
inductor gains 35% advancement in Q and 25% enhancement in f SR over conventional
series-stacked multipath inductor with an on-chip area of 180 µm for the inductance
Table 1 Comparison of performance with state-of-the-art inductors

Inductor L (nH) Qmax f Qmax f SR
Conventional series-tacked inductor [11] 7.09 22.52 3.6 9
Series-stacked multipath inductor [12] 7.8 25.07 4.1 9.2
Proposed inductor 7.9 29.87 4.7 11.5
The bold is used to represent the proposed inductor to differntiate it from the other works in the
literature
of 7.9 nH. The peak Q frequency (f Qmax ) is 4.7 GHz and f SR is 11.5 GHz, which are
within the C band. As the proposed inductor covers the entire C band, it is suitable
for applications which include satellite communications, Wi-Fi, and radar systems.
References
1. Yue CP, Wong SS (2000) Physical modeling of spiral inductors on silicon. IEEE Trans Electron
Dev 47:560–567
2. Tang C-C, Wu C-H, Liu S-I (2002) Miniature 3-D inductors in standard CMOS process. IEEE
J Solid-State Circuits 37:471–480
3. Vanukuru VNR (2015) High Q inductors utilizing thick metals and dense-tapered spirals. IEEE
Trans Electron Dev 62:3095–3099. https://doi.org/10.1109/TED.2015.2458772
4. Yue CP, Wong S (1998) On-chip spiral inductors with patterned ground shields for Si-based
RF IC’s’. IEEE J Solid-State Circuits 33:85–86
5. Pei S, Wanrong Z, Lu H, Dongyue J, Hongyun X (2011) Improving the quality factor of an
RF spiral inductor with non-uniform metal width and non-uniform coil spacing. J Semicond
32:064011
6. Xu X, Li P, Cai M, Han B (2012) Design of novel high-Q-factor multipath stacked on chip
spiral inductors. IEEE Trans Electron Dev 59:2011–2018
7. Vanukuru VNR, Chakravorty A (2014) Design of novel high-Q multipath parallel-stacked
inductor. IEEE Trans Electron Dev 61:3905–3909
8. Danesh M, Long JR (2002) Differentially driven symmetric microstrip inductors. IEEE Trans
Microwave Tech 50:332–341. https://doi.org/10.1109/22.981285
9. Kim J, Kim H (2010) CMOS symmetric trace differential stacked spiral inductor. Electron Lett
46:1005–1006
10. Akhendra Kumar P, Bheema Rao N (2017) Fractal series stacked inductor for radio frequency
integrated circuit applications. Electron Lett 53:1387–1388
11. Zolfaghari A, Chan A, Razavi B (2001) Stacked inductors and transformers in CMOS
technology. IEEE J Solid-State Circuits 36:620–628
12. Vanukuru VNR, Chakravorty A (2015) Series stacked multipath inductor with high self resonant
frequency. IEEE Trans Electron Dev 62:1058–1062
13. Teo TH, Choi YB, Liao H et al (2004) Characterization of symmetrical spiral inductor in
0.35 µm CMOS technology for RF application. Solid State Electron 48:1643–1650
Implementation of DCGAN to Generate
Gamocha Design Patterns
Adarsh Pradhan, Arunav Buragohain, Uddipta Pathak, Sahil Ansari,

and Manali Baruah
Abstract Generative adversarial network (GAN) [1] can learn to mimic any kind
of distribution of data. The main motive of GAN is to generate images of a partic-
ular class on its own. One of the most popular and successful network of GAN
is DCGAN [2], termed as deep convolutional generative adversarial network. It
consists of deconvolutional and convolutional layer in the generator and discrimi-
nator network instead of dense layers. The generation and discrimination of DCGAN
is more precise compared to the other GANs, as it features the minute details of the
image. In this paper, we implement the DCGAN network to generate new Gamocha
(traditional towel of Assamese tribe) design patterns. The design patterns made by
DCGAN are as indistinguishable and unique as the human-made Gamocha. This
GAN can follow the pattern structure of art to create new designs.
Keywords GAN · DCGAN · Gamocha
1 Introduction
The vivid art of creating and generating new images, patterns as well as designs
plays a significantly important role in the day-to-day life of an artist. As the rate
of development of artificial intelligence is blooming like an orchid these days, the
A. Pradhan (B) · A. Buragohain · U. Pathak · S. Ansari · M. Baruah

Department of Computer Science and Engineering, Girijananda Chowdhury Institute of
Management and Technology, Azara, Guwahati, India
A. Buragohain
U. Pathak
S. Ansari
M. Baruah
832 A. Pradhan et al.
work pressure and the amount of work for humans is decreasing. But these changes
failed to put shadow on the luminosity of lucid artistic nature of human creation.
In the context of the algorithm of generative adversarial network proposed by Ian
Goodfellow [1], the generation of artificial image took an enormous leap by providing
realistic and specific output of the required class. With this very algorithm, images
and arts can be created like a symphony by training the generative network with the
training data provided.
The GAN consists of two different networks, i.e., generator and discriminator
network. The generator network provides random images, and the discriminator
distinguishes whether the image of that given class is fake or real. Here both the gener-
ator and the discriminator are trained simultaneously. The generator and discrimi-
nator consist of simple multi-layer perceptrons in Vanilla-GAN. The algorithm in
Vanilla-GAN is simple, and it optimizes the mathematical equations using stochastic
gradient descent.
For semi-unsupervised learning of GAN, Alec Radford and Luke Metz proposed
the paper for DCGAN [2] where the classes consist of CNN called deep convolu-
tional generative adversarial network. It is one of the most precise and successful
implementation of GAN. It consists of ConvNets in place of multi-layer perceptrons.
Here, the generator architecture consists of deconvolutional layer which upsamples
the image and the discriminator architecture consisting of convolutional layer which
downsamples the image.
In this paper, we implement the deep convolutional generative adversarial network
to generate artificial texture of Gamocha design patterns. Gamocha is one of the most
easily recognizable cultural symbol of Assam. ‘Ga’ means ‘body’ and ‘Mosa’ means
‘wipe.’ Gamocha is mainly a rectangular-shaped cloth with two colors, i.e., white
in the body and red in the borders. Gamocha has historic significance in Assamese
culture and is presented as a token of appreciation and respect. It is also used as a piece
of water-absorbent cloth in daily household. By using the DCGAN, we recognize and
design patterns from the Gamocha dataset and generate new design artificially. The
generator here will try to produce new designs for Gamochas by generating random
noises and deconvoluting, while the discriminator will distinguish it between real or
fake by comparing and convoluting with the real dataset.
2 Related Works
Over the past few years, GANs have been implemented in various fields of art. Since
GANs have the ability to learn from a data distribution and then create new data
that is related to that data distribution, it can easily learn the underling patterns of
an art or an image of clothing design. Various different models of GANs have been
used to generate astonishing new art, for example, by training the network to learn
from variety of other art, like CAN [3], where they use a modified version GAN to
generate new artworks that are different from the training samples but are flawless to
the human eye. GANs have also been implemented to generate new texture mosaics,
Implementation of DCGAN to Generate Gamocha Design Patterns 833
GANosaic [4]. There is also the work of Zalando Research on texture synthesis using
SGAN [5] which is used for learning various repeated patters in an image.
3 Generative Adversarial Network
Generative adversarial network is a class of unsupervised deep learning networks

used to generate image proposed by Ian Goodfellow [1]. This neural network consists
of two networks generator and discriminator. The generator generates images, while
the discriminator distinguishes between fake or real of that particular image. Both
the networks learn simultaneously.
In Fig. 1, blue box denotes the generator which provides random noises as the
sample and this is sent to the red box, i.e., discriminator which discriminates the
image to be real or fake and learns the features from the real-world dataset (blue
cylindrical box) while the generator learns to generate more realistic images based
on the output of the discriminator.
In deep convolutional generative adversarial network (DCGAN), the generator
consists of deconvolutional layer [6] and the discriminator consists of convolutional
layer [7] instead of dense layers. In DCGAN, the CNN architecture [7] is adopted
and modified with three changes. The architecture of DCGAN is:
• In the pooling layers, both strided convolutions (discriminator) and fractional
strided convolutions (generator) are used.
• Batchnorm is used by both the generator and discriminator.
• The removal of fully connected hidden layers are done for deeper architectures.
• Generator in all the layers uses ReLU activation function except the output as it
takes Tanh.
Fig. 1 Generative Adversarial Network architecture diagram

• In all the layers of discriminator, LeakyReLU activation is used.

Figure 2 represents the architecture of the generator in DCGAN. This diagram
shows how an image is generated of a resolution of 64 × 64 × 3. The generator
consists of 10 layers.
Figure 3 represents the architecture of discriminator in DCGAN. Discriminator
is a CNN containing 10 layers that takes 64 × 64 × 3 and downsamples it using 2D
convolution.
Fig. 2 Generator layer architecture diagram
Fig. 3 Discriminator layer architecture diagram

4 Experiments
In this section, analysis and evaluation of the experiment is approached. The details
of the dataset and implementation are presented.
4.1 Dataset
We have collected the design pattern dataset from Europeana collection [8] and
combined it with some of the design patterns scraped from Google image and
Pinterest to increase the variety of designs. A total of 318 images (Fig. 4) are collected
with unique styles. These design patterns are then resized to a dimension of 64 × 64
to implement in the DCGAN.
Fig. 4 Sample of the real dataset

Fig. 5 Graph after 10 epoch
4.2 Design Pattern Generation
Training and testing of the model is performed on the environment configuration as

follows: OS-Windows 10, CUDA10.0, cuDNN7.6.5, GPU (NVIDEA 1060) 6 GB
memory, Python 3.7, 8 GB memory. Model parameter settings are: batch 64, image
shape (64, 64, 3), epochs 190. Deep learning library used is: Keras [9].
In the generator network, the activation function used is ‘ReLU’ [10] with batch
normalization momentum of 0.5. The last layer of the generator function uses ‘tanh’
activation function. In the discriminator network, the activation function used is
‘LeakyReLU’ [11] with batch normalization momentum of 0.5. After the flattening
layer, the discriminator network uses ‘sigmoid’ activation function. Adam optimizer
[9] is used in both the generator and discriminator network.
For training the model, the generator function and the discriminator function
are run simultaneously loading the real dataset. During the process of training, the
generated images are saved after an interval of (batch number % 50).
The graphs of Figs. 5 and 6 show the difference of results produced by the loss
and batch iteration in DCGAN model after 50 epochs and 190 epoches.
Figure 7 is the result of the images produced by the DCGAN model after 190
epochs. It took almost six (6) hours to completely execute the program on our system
with the configuration as mentioned earlier. Here, we see that the generated images, as
shown in Fig. 6, have distortions and the design pattern cannot be distinctly observed,
but they show uniqueness in them. Our dataset has only 318 unique images, and if
we had larger number of images in the dataset, then the new design pattern generated
will also be of better quality. These new design patterns can be used to craft new
Gamocha with a different look and feel. With these uniquely designed patterns, every
Gamocha can have a new appearance and will enhance the beauty of the traditional
Gamocha.
Table 1 shows the results of generator loss and discriminator loss after several
epochs.
Fig. 6 Graph after 190

epoch
Fig. 7 Design patterns produced by DCGAN
Table 1 Shows the results of

Epoch Generator loss Discriminator loss
generator loss and
discriminator loss after 10 3.769013 0.9742602
several epochs 100 2.111452 0.7276244
190 1.363392 0.6452213
The results produced by the DCGAN are distorted but have uniqueness in it. As
our training dataset has only 318 unique images, so due to the lack of a larger training
dataset, the DCGAN model could not generate new noise-free images. We expect
that as the size of dataset will increase, the model will also obtain better results.
5 Conclusion
Generative adversarial networks are a stimulating and redefining recent innovation

in deep learning. In this paper, we pragmatically focused on generating new and
innovative design patterns for Gamocha implementing effective deep convolutional
generative adversarial network. The generation of textures, different fibers and fabrics
and designs for clothing using GANs can benefit and acknowledge the artificial
intelligence to learn more and practically contemplate about the art patterns. The
generation of new and variable designs produced by the DCGAN can be more unique
if fed with proper design datasets.
Acknowledgements We would like to thank the developers of NVidia and TensorFlow for the
immensely useful hardware and software which were used to implement our project.
References
1. Goodfellow I, Pouget-Abadie J, Mirza M, et al (2014) Generative adversarial nets. NIPS

2. Radford A, Metz L, Chintala S (2015) Unsupervised representation learning with deep-
convolutional generative adversarial networks (CNN). arXiv preprint arXiv:1511.06434
3. Elgammal A, Liu B, Elhoseiny M, Mazzone M (20–22 June, 2017) Eighth international
conference on computational creativity (ICCC). Atlanta, GA
4. Jetchev N, Bergmann U, Seward C (2017) 31st conference on neural information processing
systems (NIPS 2017). Long Beach, CA, USA
5. Bergmann U, Jetchev N, Vollgraf R (2017) Learning texture manifolds with the periodic spatial
GAN. In: Proceedings of the 34th international conference on machine learning
6. Zeiler DM, Krishnan D, Taylor GW, Fergus R Deconvolutional networks. https://arxiv.org/abs/
1311.2901
7. Zeiler MD, Fergus R Visualizing and understanding convolutional networks. https://arxiv.org/
pdf/1311.2901.pdf
8. Europeana collection. https://www.europeana.eu/portal/en
9. Chollet F, et al (2015) Keras. https://keras.io
10. Ramachandra R, Zoph B, Le QV Searching for activation function. https://arxiv.org/abs/1710.
05941
11. Nwankpa CE, Ijomah W, Gachagan A, Marshall S Activation functions: comparison of trends
in practice and research for deep learning. https://arxiv.org/pdf/1811.03378.pdf
Performance Evaluation of On-Grid
Rooftop Solar PV System in GMCH,
Guwahati, Assam
Israfil Hussain, Utpal Chandra Baro, Rupa Chhetri, Vanita Agrawal,

and Abdul Latif
Abstract In this paper, 500 KWp On-Grid Rooftop Solar PV plant installed and
commissioned in 2018 at Guwahati Medical College Hospital (GMCH), Guwahati,
Assam which is one of the largest in Assam is studied based on the information
or the data in the report of Assam Energy Development Agency (AEDA) and its
performance evaluation is carried out using PVsyst. Then we have compared the
parameters with the one simulated by PVsyst for varying seasonal tilt i.e 20° in
summer and 60° in Winter season and also for the fixed-tilt of 22°. Here the azimuth
angle is taken 0° subject to the direction of the panel facing towards the south. The
mean global horizontal radiation for the site is 4.72 KWh/m2 /day with the average
annual temperature is 24.6 °C. This study includes the design and performance
analysis of the plant and then compares it with the obtained simulation values. The
study uses different techniques and methods of a PV system for performance analysis
and optimizes the operational behavior of PV systems.
Keywords PV system · On-grid · Performance ratio · Annual energy yield
1 Introduction
In a tropical country like India, the possibility of solar power generation is huge.
In most parts of the region, it receives around 300 days of sun annually with 7–
8 h of sunshine every day. The average incident radiation in India is between 4 and
7 kWh/m2 /day [1]. The solar system received around 5000 Tera KWh of energy
per year. The Ministry of Power (MoP), under Indian Government in November
2009 has launched a solar mission in the name of Jawaharlal Nehru with an aim for
I. Hussain (B) · V. Agrawal

Royal Global University (RGU), Guwahati, Assam, India
U. C. Baro · R. Chhetri
EE Department, RGU, Guwahati, Assam, India
A. Latif
National Institute of Technology Silchar, Silchar, Assam, India
840 I. Hussain et al.
sustainable growth of solar power generation. The aim of the mission is to make the
country strong in the fields of solar power by 2020 [2].
Government of India aims at installing 40 GW of Solar Rooftop grid-connected
systems throughout the country by the year 2022. In fact, 250 MW Grid-tied Solar
Rooftop projects have been provided in the state under the supervision of the Ministry
of New and Renewable Energy (MNRE) and will be completed by the year 2022. Out
of 250 MW, MNRE has sanctioned 14 MW grid-connected Solar Rooftop Program to
Assam Energy Development Agency to implement the project in the state of Assam
for the year 2017–2018 with 70% subsidy for Residential, Institutional (Non-Govt.)
and Social Sectors Institutes. The “Smart City” initiatives by GOI have provided an
unprecedented choice for the municipalities to use the modern technologies for better
livelihood and better connectivity of the cities where we live. Before undertaking the
program, the decision-maker predicts the solution to some of the key questions like
“What we are planning to do?” and “What are the priorities for the city?”
The ‘Solar City’ development program initiated by the MNRE under India
Government has aimed at developing the Guwahati city as the ‘Solar City’ in
Assam by Assam Energy Development Agency (AEDA) through Guwahati Munic-
ipal Corporation in 2010. The 14 MWp Solar Rooftop Project has been sanctioned
to Assam Energy Development Agency (AEDA).
Due to various land variations, the insertion of solar PV system on the geometrical
area results in less efficiency. Hence, the rooftop area is emerging as a good option
for energy generation. This rooftop grid-connected network will not only solve the
failure of power but simultaneously it minimizes the environmental hazards generated
by traditional fossil fuel based generation systems [3].
The plant performance work is described into two stages as follows:
1. Manual extraction of the parameters for electricity generation at a fixed-tilt.
2. Then, comparing the parameters with the one simulated by PVsyst in varying
seasonal tilt and fixed-tilt. PVsyst is software that deals with the analysis of data,
PV system sizing for grid-tied, independent, solar PV pumping as well as DC
grid system.
2 Literature Review
Schaefer et al. [4] studied the performance, availability, and maintenance of 10

solar PV plants in U.S. Performance ratios, capacity factors are discussed here. The
photovoltaic energy cost is mentioned which is followed by investment/maintenance
costs, and capacity factor, and depends on the location, availability of plant, and the
tracking system. Marion et al. [5] studied the performance indicators for grid-tied PV
networks. The research focused on the importance of four performance indicators
required for the total system performance assessment corresponding to the produc-
tion of energy, solar reserve, and overall outcome of system losses which includes the
absolute PV unit yield, base yield, performance factor, and PVUSA grading. Tudo-
rache et al. [6] investigated the performance evaluation of sun-tracking PV panels of
Performance Evaluation of On-Grid Rooftop Solar PV System … 841
single-axis type. The efficiency of sun-tracking PV system is then equated with the
fixed PV system. The device identifies the fittest PV unit position with reference to
the sun via a DC motor which is governed by a smart drive unit that will receive input
signals from LDR light sensors. Cristaldi et al. [7] studied the economical evaluation
of PV system losses that occurs because of dust and pollution. This article presents a
technique for the maintenance of a PV plant that includes the cleaning expenses of the
panels, then compares them to the economic losses in order to increase the efficiency.
Kumar et al. [8] illustrated the performance evaluation of a 20 kWp grid integrated
solar photovoltaic which is installed in an industry, located in Tiruchirappalli, India
that describes some important features of the plant and the performance as well as
economic analysis has also been done perfectly. Finally, the results are obtained for
energy generation per month, different factors (performance and capacity), economic
as well as maintenance features, etc. Bharathkumar et al. [9] studied the performance
assessment of grid-tied 5 MW Solar PV unit in the regions of Karnataka where,
the grid-tied solar PV unit set up by the Karnataka Power Corporation Limited, is
described, and the evaluation of performance is done correspondingly. Sundaram
et al. [10] studied the performance assessment of 5 MWp grid-tied PV units in
the Southern part of India wherein the real-time analysis of the plant parameters
is done and is verified by RET screen plus software environment. Bahaidara et al.
[11] reported the analyses of the performance of PV unit for the climatic scenario of
Dhahran in Saudi Arabia. EES software is leveraged to compute the parameters like
cell temperature, maximum power position current/voltage/efficiency, and power.
The values were then equated with the simulated results extracted by placing the PV
modules to its environmental condition. Khalid et al. [12] studied the importance
of performance ratio in grid-tied PV units efficiency. It determines the efficiency
and output power. It has highlighted some economic and environmental benefits of
using PR by framing a scenario example from the project named ‘SolMap project’ in
India. Vasisht et al. [13] studied the performance of solar PV installation in seasonal
variations. This attempt emphasizes the performance analyses of PV units leveraging
the Capacity Utilization Factor (CUF) and Performance Ratio (PR). Sharma et al.
[3] studied the performance of a grid-tied PV unit (capacity: 11.2 kWp) in Eastern
India. This paper presents the parameter results for the month in between September
2014 to August 2015 of the installed plant. PV panels/inverter efficiency, PV array
yield, final yield, and PR of the unit are also described perfectly. Nirwan et al. [14]
addressed the performance analysis of grid-tied solar PV unit employing PVsyst soft-
ware. The performance evaluation has been done with PVsyst for 1 MW grid-tied
power plant set up at PEC University of Technology. He studied the performance
evaluation at the given tilt angle for which the plant is installed and compared it with
the optimum tilt according to the site with and without the Horizon considerations.
Hussain et al. [15–20] studied the performance analysis of Autonomous Hybrid
System using renewable energy sources. Hereby the author uses the Artificial Intel-
ligence techniques for optimization of the parameters of the different controllers.
Finally, the comparative performance of Artificial Intelligence controllers for the
hybrid model is presented.
3 Depiction of Solar PV Grid Scheme
The installed rooftop grid-tied PV plant employs solar panels, dc to ac conversion

system (inverters), a power conditioning unit, and grid-tied equipment. The system
has no energy storage. When the grid-connected PV system generates surplus amount
then it supplies the excess power to the utility grid (Fig. 1; Tables 1 and 2).
According to data available on the NASA website, the monthly average isolation
falling on a horizontal surface at the location is 4.72 kWh/m2 /day and the average
annual temperature is 24.6 °C.
3.1 Plant Layout of the Established PV System
The total capacity of the proposed GMCH rooftop PV plant is of 500 kWp with a
total of 3567 m2 rooftop area. The plant is characterized into three different blocks
of the hospital building. Each building is characterized by different solar generation
capacity in accordance with the rooftop area available, as shown below
Fig. 1 Site location—GMCH, Guwahati
Table 1 Site information and

Site name Rooftop of Guwahati Medical College Hospital
Meteorological data
Coordinates 26° 10 34 N, 91° 45 46 E
Elevation 72 m above mean sea level
Altitude 43 m above ground level
Table 2 The table below showing the meteorological data of the located plant
Jan Feb Mar Apr May Jun July Aug Sep Oct Nov Dec
Hor. global 123.0 125.1 168.1 164.6 163.5 144.1 145.2 151.9 147.8 135.4 136.8 132.4
Hor. diffuse 42.4 53.2 73.0 85.9 93.8 86.6 80.5 82.5 69.7 70.3 34.5 23.3
Extraterrestrial 203.6 219.0 284.4 311.8 343.1 338.6 346.2 330.3 289.4 256.8 207.1 191.7
Clearness index 0.604 0.571 0.591 0.528 0.477 0.426 0.419 0.460 0.511 0.527 0.660 0.691
Amb. temp. 16.8 19.4 23.1 24.8 27.2 28.0 28.9 29.0 27.8 26.2 22.0 18.6
Wind velocity 0.3 0.6 0.9 1.3 0.9 0.7 0.7 0.7 0.7 0.6 0.5 0.3
Performance Evaluation of On-Grid Rooftop Solar PV System …
843
1. Main building (rooftop)—370 kW (operating)

2. Main building (5th floor)—80 kW (operating)
3. Tin roof—50 kW (operating).
So, the total plant generation capacity as 500 kW (Fig. 2).
A total of 1516 solar panels and 10 string inverters that include two inverters
of 60 kW, seven inverters of 50 kW, and one inverter of 30 kW. The efficiency of
the inverter is 95%. These inverters convert the DC power to AC power and the
extractable output is supplied to the 11 kV grids utility.
The solar modules used are of “SOVA SOLAR” of the type SS330P polycrystalline
with open-circuit voltage (V oc ) given as 45.27 V and short circuit current (I sc ) given
as 9.29 A (Table 3).
Fig. 2 Picture of 500 kWp grid-tied rooftop PV unit
Table 3 PV module and Inverter specification (Report of AEDA)

PV module parameters Specification Inverter parameters Specification
Type of modules Polly crystalline Model BG50KTR INVT
Pmax 330 W Maximum DC Input power 53,000 W
I mp 8.89 A Maximum DC Input voltage 1100 V
V mp 36.78 V MPPT range 200–900/750
I sc 9.29 A No. of MPPT/strings/MPPT 2/5
V oc 45.27 V Starting voltage 200/150 V
Cell temperature 25 °C Rated output power 50,000 W
Array area 50,000 sqft Max. AC output current 72 A
No. of modules 1516 Frequency of grid 50 Hz
Efficiency 15% Max. efficiency 98.60%
Weight 21.20 Kgs
Accessibility Data Parameters
Location Total Load
System Designing
System Parameters Orientation Losses
Generation Data
Monthly Energy o/p Performance Ratio
System Cost
Provides access to variables that define the installation and
operating cost of the PV plant.
Simulation
Performance comparison in PV syst between the fixed and
seasonal tilt angle for obtaining optimization.
Fig. 3 Flowchart of the research work
3.2 Flowchart of the Work
See Fig. 3.
3.3 Plant Orientation
The tilt angle for the solar panels is inclined at 22°. The azimuth degree is 0° since
the panels are facing towards the south. For optimization, the tilt angle is taken as 22°
for summer and 60° for winter with azimuth of 0° that means the panels are facing
towards the south (Figs. 4 and 5).
Monthly energy generation of the installed plant for tilt angle 22° is shown below
for the last four months Since the RMS is activated in July 2019.
Figure 6 is collected from TrackSo [21], an IoT based energy management plat-
form to track the performance of the remote plants provide a full control without
actually being present there. The fig. shows that in the month of October has the
record for the highest solar energy yield, by considering 4 months from August 2019
to November 2019.
Fig. 4 Original tilt angle
Fig. 5 Seasonal tilt angle
GENERATION REPORT
Energy Performance Ratio Solar Irradiation (W/m2)
4760
4310
4100
3360
429.35
363.74
322.53
316.77
99.99
80.83
80.18
99.9
AU GU S T SEPTEM BER OC T OB E R NOVE M B E R DE C E M B E R
Fig. 6 Monthly energy generation and performance ratio

Table 4 Average energy yield for the specified months

Months Monthly energy o/p Avg. monthly extractable Avg. annual extractable
energy o/p (MWh) energy o/p (MWh)
August 2019 4.1 4.1 49.2
September 2019 4.3
October 2019 4.7
November 2019 3.3
Total 16.4 (MWh)
The average annual energy yield can be calculated by the product of the average
monthly extractable energy output with the number of total months throughout the
year. Thus, the annual energy yield is tabulated as in Table 4.
4 Performance Analysis
The solar PV plant performance depends on the various parameters that result in
the true possibility of solar power generation variability. The calculation of the
power, energy, and yield of the grid-tied PV power plants requires some important
parameters. The required parameters given by [3] are shown below
1. PV array yield
The PV array yield is referred to as the PV array extractable output, divided by
its rated power. The output represented (in a day, month or year) is given by [3]
Ya = E DC.d /Ppv rated
2. PV final yield
The PV final yield is referred to as the output energy of the entire PV unit divided
by the maximum (rated) power of the established PV array. Thus, the final yield
is given by [3]
YFd = E AC.d /Ppv rated
3. PV segment efficiency
The efficiency of PV segment (module) is given as
ηpv = (PDC /G t ∗ Am ) ∗ 100
4. DC-AC conversion (Inverter) Efficiency

The dc-ac conversion efficiency (η) is given as
ηinv = PAC /Pdc

5. Efficiency of system
The overall system efficiency is given by
ηsys = ηpv ∗ ηinv
6. Performance Ratio (PR)

PR is referred to the energy supplied to the network or grid divided by the energy
produced by the unit in DC power, for the maximum peak hours/day [12]. The
expressed is given as-
PR = YF /YR
7. CUF
It is a ratio of energy produced in a year by the PV unit to the aggregate of energy
the PV unit will produce when operated under full rated power/day/year is known
as the Capacity Utilization Factor (CUF). It is given as [13]
CUF = (E ac.a /Ppv rated ∗ 24 ∗ 365) ∗ 100
5 Simulated Results
The simulation is carried out for the comparison, using the PVsyst software by
considering the total installed capacity, PV array area, field parameters, system sizing,
etc.
Case 1 For fixed tilted plane field parameters include 22° angle with azimuth angle
as 0° as the panels are south facing.
Case 2 For Seasonal Tilt the summer tilt is given as 20° and winter tilt as 60°. The
azimuth is 0° as well.
As seen from Figs. 6 and 7, the optimized energy generated from the PV array with
seasonal tilt is 4.54 kWh per kWp per day and the energy generated from fixed-tilt is
4.43 kWh per kWp per day. But the PR of fixed-tilt system shows quite a significant
increase as compared to the performance ratio of the optimized tilt system (Fig. 8).
5.1 Loss Diagram Analysis
Figs. 9 and 10.

Fig. 7 Normalized production and performance ratio of original tilt
Fig. 8 Normalized production for seasonal tilt
6 Cost Analysis
As per the Electricity Regulatory commission for Assam the cost per watt of solar
power is Rs. 53/Watt. The total cost of the plant is 2.65 crore with the MNRE capital
subsidy of 70% of the project cost. The overall energy fed into the grid from the
solar PV (SPV) plant is given as 81.1 MWh or 81,100 units per annum. The total
time for the payback period is 7 years for the installed 22° tilt and south orientation
with azimuth angle 0°.
7 Conclusion
The installed 500 kWp grid-tied PV unit at the rooftop of Guwahati Medical College
Hospital, Assam was commissioned and monitored since August 2019 and its perfor-
mance parameters are studied month wise. Then the PV system performance has been
Fig. 9 Loss diagram for fixed-tilt angle
compared with the simulated values. Thus, the important features from this study are
provided below-
1. The installed solar PV plant by using PVsyst shows that the 500 KWp power
plant generates more energy for seasonal tilt angle equal to 22° for summer and
60° for summer.
2. The important factor to consider for designing the plant is that the energy
produced is 1.92 ≈ 2% more in the case of a seasonal tilt than the original
tilt angle.
3. The energy supplied to the grid is 2000 kWh more for seasonal tilt than that of
the fixed tilted system.
4. Moreover, the mounted whole PV unit in GMCH has reduced about 25.22 ton of
CO2 from the atmosphere per year.
Fig. 10 Loss diagram for optimized tilt angle
Acknowledgements The authors are very much grateful to the authority of the Assam Energy
Development Agency (AEDA) for giving such facilities for carrying out this research study.
References
1. Mahtta R, Joshi PK, Jindal AK (2014) Solar power potential mapping in India using remote
sensing inputs and environmental parameters. Renew Energy 71:255–262
2. Akoijam AS, Krishna VV (2017) Exploring the Jawaharlal Nehru National Solar Mission
(JNNSM): impact on innovation ecosystem in India. Afr J Sci Technol Innov Dev 9(5):573–585
3. Sharma R, Goel S (2017) Performance analysis of a 11.2 kWp roof top grid-connected PV
system in Eastern India. Energy Reports 3:76–84
4. Schaefer JC (1990) Review of photovoltaic power plant performance and economics. IEEE
Trans Energy Convers 5(2):232–238
5. Marion B et al (2005) Performance parameters for grid-connected PV systems. In: Conference

record of the thirty-first IEEE photovoltaic specialists conference, 2005. IEEE
6. Tudorache T, Oancea CD, Kreindler L (2012) Performance evaluation of a solar tracking PV
panel. Univer “Politehnica” Bucharest Sci Bull Ser C: Electr Eng 74(1):3–10
7. Cristaldi L et al (2012) Economical evaluation of PV system losses due to the dust and pollu-
tion. In: 2012 IEEE international instrumentation and measurement technology conference
proceedings. IEEE
8. Kumar KA, Sundareswaran K, Venkateswaran PR (2014) Performance study on a grid
connected 20 kWp solar photovoltaic installation in an industry in Tiruchirappalli (India).
Energy Sustain Dev 23:294–304
9. Bharathkumar M, Byregowda HV (2014) Performance evaluation of 5 MW grid connected
solar photovoltaic power plant established in Karnataka. Int J Innov Res Sci Eng Technol 3(6)
(2014)
10. Sundaram S, Babu JSC (2015) Performance evaluation and validation of 5 MWp grid connected
solar photovoltaic plant in South India. Energy Convers Manag 100:429–439
11. Bahaidarah H et al (2015) Performance evaluation of a PV module under climatic conditions
of Dhahran, Saudi Arabia. Energy Explor Exploit 33(6):909–929
12. Khalid AMohd et al (2016) Performance ratio—Crucial parameter for grid connected PV
plants. Renew Sustain Energy Rev 65:1139–1158
13. Vasisht MS, Srinivasan J, Ramasesha SK (2016) Performance of solar photovoltaic installa-
tions: effect of seasonal variations. Solar Energy 131:39–46
14. Nirwan D, Thakur T (2017) Performance evaluation of grid connected solar PV plant using
PVsyst. Int Res J Eng Technol (IRJET) 4:3190–3194
15. Hussain I, Das DC, Sinha N (2017) Reactive power performance analysis of dish—Stirling
solar thermal—Diesel hybrid energy system. IET Renew Power Gener 11(6):750–762
16. Hussain I, Ranjan S, Das DC, Sinha N (2017) Performance analysis of flower pollination algo-
rithm optimized PID controller for wind-PV-SMES-BESS-diesel autonomous hybrid power
system. Int J Renew Energy Res (IJRER) 7(2):643–651
17. Hussain I, Das DC, Latif A, Sinha N, Ranjan S (2020) FPA optimized controller for frequency
control of dish stirling/diesel/fuel cell/energy storage based autonomous hybrid system. In:
Dawn S, Balas V, Esposito A, Gope S (eds) Intelligent techniques and applications in science
and technology. ICIMSAT 2019. Learning and analytics in intelligent systems, vol 12. Springer,
Cham
18. Latif A, Pramanik A, Das DC et al (2018) Plug in hybrid vehicle-wind-diesel autonomous
hybrid power system: frequency control using FA and CSA optimized controller. Int J Syst
Assur Eng Manage 9:1147–1158. https://doi.org/10.1007/s13198-018-0721-1
19. Latif A, Das DC, Ranjan S, Hussain I (2018) Integrated demand side management and gener-
ation control for frequency control of a microgrid using PSO and FA based controller. Int J
Renew Energy Res (IJRER) 8(1):188–199
20. Latif A, Das DC, Ranjan S, Barik AK (2019) Comparative performance evaluation of WCA-
optimised non-integer controller employed with WPG–DSPG–PHEV based isolated two-area
interconnected microgrid system. IET Renew Power Gener 13(5):725–736
21. Solar.trackso.in
Forced Alignment Method for Detection
of Bodo Word Boundary
Raja Jwbthasa Daimari and Aniruddha Deka
Abstract Word Boundary Detection is one of the most important problems in the
field of speech processing. There may be many methods for word boundary detection
developed to date. But most of them are in languages like English, German, Hindi,
etc. We know that language features are different for different languages, so a single
method cannot be considered efficient for all languages. This paper proposes a new
approach to Word Boundary Detection in Bodo Language (tonal language) in a
noise-free environment. In a tonal language, a slight variation in pitch may change
the meaning of the word completely. From the analysis of the results of male and
female data, it is found that 80.56% of word boundaries are correctly detected for
male data and 81.11% of word boundaries are correctly detected for female data.
The proposed algorithm is developed using Forced Alignment Method using Hidden
Markov Model (HTK) toolkit.
Keywords Word boundary detection · Tonal language · Forced alignment · Pitch ·

HTK toolkit
1 Introduction
word Boundary Detection can be defined as identifying the start and end of a spoken
word in an utterance. The detection of a word boundary is one of the most challenging
works in the area of speech processing. The accurate detection of word boundaries
is challenging in the sense that the speaker does not pause consciously between the
words. Also, the method developed for a particular language may not work for other
languages as different languages possess different features. It is more challenging in
the case of a tonal language. Tonal languages are unique in the sense that they have
R. J. Daimari (B) · A. Deka

Department of Computer Science and Engineering, Royal School of Engineering and Technology,
Guwahati, India
A. Deka
854 R. J. Daimari and A. Deka
two tones for certain words. Bodo [1] is a tonal language and the tonal language
features for Bodo will also be explored to a certain extent. In this paper, we propose
a method to detect word boundary for Bodo Language using the Forced Alignment
Method in common speech signal environments. Hidden Markov Model Toolkit
(HTK) is used in the proposed system for the detection of word boundaries.
2 Previous Research
Jain et al. [2] developed an algorithm for Word Boundary Detection in the Hindi
Language which was based on intensity and pitch. Their algorithm was designed to
work in a noisy environment by considering three prosodic parameters which are
“defined pitch contour”, “undefined pitch on silence zone” and “intensity contour”.
Mandal et al. [3] proposed a method of detecting word boundaries by using
Suprasegmental features for the Bengali language, which is a stress bound language.
While detecting word boundaries for the Bengali language, the authors’ Intonation,
Stress and rhythm of the language. After their research, they found that 87.8% of
word boundaries have been correctly detected. The drawback of this method is that
it can be used only for stress bound language.
Kacur and Rozinaj [4] discussed a method for Word Boundary detection using
cepstral matrices. They proposed that the method is efficient for detecting word
boundaries in a noisy environment. In this method, speech signals are divided into
consecutive segments, which are overlapped by a certain part of their length to keep
the feature vectors behaviour smooth in the time and moreover not losing any speech
information.
Ramana Rao and Srichand [5] proposed a method for word boundary detection
using pitch variation, which determines the highness and lowness of a tone in a spoken
utterance for Hindi, Bengali, Marathi, Telugu and German. They used the frequency
of pitch for each word and found that more than 85% of word boundaries were
correctly determined for the Indian languages and about 65% word boundaries were
correctly determined for German language. The authors attributed the low detection
rate in German to the fact that the misses occurred more often between words with
only one vowel in that particular word and the next word.
3 Proposed System
3.1 Data Preparation
Data Preparation, which is also known as the preprocessing step is the first step in the
detection of word boundaries as shown in “Fig. 1”. For generating a set of HMMs,
the speech data files and the transcriptions associated with those files are required.
Forced Alignment Method for Detection of Bodo Word Boundary 855
Fig. 1 Block diagram of the proposed model
Therefore, data preparation is essential and before the data can be used for training
it should be converted into the appropriate parametric form along with its associated
transcriptions in the correct format and use the required labels. Thus, before the
data is recorded, a phone set for the words is defined, a dictionary is constructed to
cover both training and testing which consists of the list of unique words present
in the sentences and a task grammar is to be defined. Our database consists of 10
sentences which contain a total of 41 unique words in which certain words possess
tonal features. Therefore, the dictionary consists of 41 words, and they are kept in
the standard HTK format [6].
3.2 Data Recording Plan
The training and test data are recorded using the tool Wavesurfer. Wavesurfer is a
waveform recording and trimming tool. For the purpose of recording, 10 sentences
are taken in Bodo where certain words have tonal features. The data is recorded
from 10 speakers where the number of male speakers is 5 and the number of female
speakers is 5. Each sentence is recorded from each speaker 10 times. The resulting
database consists of 1000 chipped files (500 male and 500 female utterances). In our
research, four male speakers and four female speakers are used for training and one
male and one female speaker is used for testing. This means 800 files (Male: 400,
Female: 400) are used for training. And the remaining 200 will be used for testing.
The word boundaries will be detected from the test data using the Forced Alignment
Method.
3.3 Creating Transcription Files
To train a set of Hidden Markov Models (HMM) [6], every file of training data
must have an associated phone level transcription. As a result, transcription files
are necessary. To do this, a set of phone transcriptions are required. Each word
should be written in a single line and each utterance should be terminated by
a single period of its own. These transcriptions are kept in a file known as
Master Label File (MLF) which is a single file containing a complete list of
transcriptions. When HTK processes speech files, it expects to find a transcrip-
tion with the same name but a different extension. For example, when the file
/root/raja/speech_test_new/bodo_female_1.wav was being processed, HTK would
look for a label file called /root/raja/speech_test_new/bodo_female_1.lab. When
MLF files are used, HTK will look for a pattern in a file that matches the required
label file name. In our research MLF files are created for both training and testing
data for each chipped wav file.
In the detection of word boundaries in Bodo Language with our data, Mel Frequency
Cepstral Coefficient (MFCC) is used for feature extraction. MFCC is chosen for
feature extraction as compared to other methods because MFCC is derived on the
concept of logarithmically spaced filter bank which resembles the concept of the
human auditory system and hence produces better results as compared to Linear
Predictive Coding (LPC). The MFCC parameters are given in the configuration file
where the Fast Fourier Transform (FFT) should use a Hamming window and the
signal should have first-order pre-emphasis applied using a coefficient of 0.97. The
filter bank has 28 channels and 12 MFCC coefficients will be the output. This is
specified in the configuration file. In HTK, the tool HCopy is used to automatically
convert its input into MFCC vectors. In order for HCopy to work, some of the source
file with its respective output file is to be used (Fig. 2).
3.5 Word Recognizer Using HMM
The word recognizer is built-in HTK with the help of HMM. The significance of a
recognizer is to map between sequences of speech vectors and the wanted underlying
speech utterance. A word recognizer is built using 41 unique words as the dataset
consists of 10 sentences containing 41 unique words.
Fig. 2 Construction of MFCC feature vectors
3.6 Forced Alignment
In the determination of word boundaries, we have used the Forced Alignment Method.
With the help of this method, the timestamp information of the sentence is found out.
This method involves spotting the words which are recognized with the help of HMM
and thereby marking the boundaries at the appropriate places. Forced Alignment
Method is an effective method for labelling the word boundaries for a large speech
database. In this method, the words are checked and compared with the language
model which is developed and if the word matches the language model, the timestamp
for that particular word is marked and if it does not match the language model which
we have developed, it proceeds to check in the word. After forced alignment is
performed on the test files, a file with the word boundaries is created by HTK.
In our research, two measures are defined for the detection of word boundaries. They
are the hit rate and false alarm rate. In our hypothesis, a hit occurs when the hypothesis
contains the correct word along with the correct timestamp information. The hit rate
determines the percentage of word boundaries correctly detected. On the other hand,
False Alarms in the word boundary hypothesis occur when the hypothesis does not
contain the actual word boundaries to be detected. It occurs when the word boundary
is detected but the detection is wrong. It specifies the percentage of error in the word
boundary hypothesis. Hit Rate and False Alarm rates are defined for a single test
speech file as follows.
No. of Word Boundaries Correctly Hypothesized

Hit Rate =
No. of Word Boundaries Present
No. of error in Word Boundary Hypothesis
False Alarm Rate =
No. of Word Boundary Hypothesis
In “Fig. 3” for the utterance “Musua nokorni fichijagra junat”, the waveform
for the sentence is shown along with the word boundaries after each word. In the
sentence, the word “Musua” has its word boundary from 0.05 to 0.82 s, the word
“nokhorni” has its word boundary from 0.82 to 1.27 s, the word “fichijagra” has its
word boundary from 1.27 to 1.91 s and the word “junat” has its word boundary from
1.91 s to 2.3 s. As can be seen from “Fig. 3” for this sentence, the words along with
the word boundary information are correctly detected thereby giving a hit rate 4/4,
i.e. 100% and no false alarm is generated as no miss occurred in the detection of
boundaries.
In “Fig. 4” for the utterance “Bharatni rajdhania gwdan delhi”, the waveform for
the sentence is shown along with the word boundaries. In the sentence, the word
“Bharatni” has its word boundary from 0.04 to 0.56 s, the word “rajdhania” has its
word boundary from 0.56 to 1.12 s, the word boundary for the word “gwdan” is
wrongly detected as “gandani” along with its boundary from 1.12 to 1.69 s and the
word “delhi” has its word boundary from 1.69 to 2.12 s. As it can be seen from
Fig. 3 Waveform of the utterance “Musua nokhorni fichijagra junat” with timestamp
Fig. 4 Waveform of the utterance “Bharatni rajdhania gwdan delhi” with timestamp
“Fig. 4”, the word “gwdan” is wrongly detected as “gandani”. This results in a
miss, thereby generating a false alarm. As we have highlighted that Bodo is a tonal
language, the false alarm may be attributed to the tonal features of the language.
Therefore, the hit rate, in this case, is 3/4, i.e. 75% and the false alarm is 1/3, i.e,
33.33%.
For the whole corpus which consists of 1000 chipped wav files, the overall results
from the test data files are shown in “Table 2”. The hit rate is marginally higher in the
case of male data as compared to female data. This can be because of the fact that the
pitch of a female voice is higher as compared to male and pitch is also a measure to
analyze word boundaries. Therefore, the variation of pitch is related to the difference
in hit rate for male and female data. The result is represented graphically in “Fig. 5”.
As it is already emphasized that Bodo is a tonal language, therefore a slight change
in variation of pitch in certain words can change the meaning of the word completely.
After analysis of the results, it is found that the false alarms are mainly due to the
tonal variations. Basically, in our study false alarms have been raised for the words
which are highlighted in “Table 1”.
Fig. 5 Graphical
representation of hit rate and
false alarm rate
Table 1 Variation in actual

Actual word Detected word
and detected word boundary
asama asomia
mache munche
harini gandani
mache sache
gwdan gadani
musua mache
asama rama
mache gongshe
Table 2 Overall results

Sex Hit rate (%) False alarm rate (%)
Male 80.56 24.14
Female 81.11 23.28
Fig. 6 Speech analysis of the word “mairong”
In “Fig. 6” shows the speech analysis of a single word “mairong”. The first part of
the diagram shows the waveform of the word which ranges from −3193 to 2551 Hz.
The second part is the spectrogram of the same word plotted in a log scaled spec-
trogram varying amplitude with time (ms). Here, we can divide the word in smaller
frames say 0.05 ms and then we can eliminate the frames which do not reach the
required frequency range.
5 Conclusion
In this paper, it is shown that word boundaries are detected in Bodo Language with
the help of HTK. Moreover, a few tonal features of Bodo Language have been
explored. The hit rate for male and female data is found out to be 80.56% and
81.11% respectively. Also, the false alarm rate is found out to be 24.14% and 23.28%
respectively. Moreover, the reasons for variation in hit rate for male and female data
have been mentioned. Also, false alarm rates have been discussed. The hit rate can be
increased by improving the recording environment and as such a studio environment
can be chosen. Also, due to the variation in speaker’s pronunciation and accents, the
hit rate is affected.
6 Future Scope
The future scope of this work is that after the word boundaries have been successfully
detected, the individual words with accurate boundaries can be extracted and it can
be used for detecting regions out of vocabulary words and also for Keyword Spotting.
References
1. Sarmah P (2004) Some aspects of the tonal phonology in Bodo. Central Institute of English And
Foreign Languages, Hyderabad
2. Jain A, Agrawal SS, Prakash N. Performance evaluation of word boundary detection for Hindi
speech database. GGSIP University, Delhi, India
3. Mandal SKD, Gupta B, Datta AK (2006) Word boundary detection based on Suprasegmental
features: a case study on Bangla speech. Int J Speech Technol
4. Kacur J, Rozinaj G. Word boundary detection in stationary noises using cepstral matrices.
Department of Electrical Engineering and Information Technology, Slovak University of
Technology
5. Ramana Rao GV, Srichand J (1996) Word boundary detection using pitch variations. Department
of Computer Science and Engineering, IIT Madras
6. Young S, Kershaw D (2000) The HTK book. Microsoft Corporation, pp 27–49
Fault-Tolerant Cascaded Asymmetrical
Multilevel Inverter for the Solar-Based
Electric Vehicle
Hillol Phukan, Tamiru Debela, and Jiwanjot Singh
Abstract Fault-tolerant asymmetrical multilevel inverter (FTAMLI) has been

performed. Multicarrier selective harmonic elimination pulse width modulation
(MC-SHEPWM) has been adopted to generate the output voltage levels for an
FTAMLI. The fault-tolerant algorithm has been adopted. Both during and after fault,
the voltage remains constant by using this algorithm; to analyze the implementation
of the propound algorithm with MC-SHEPWM scheme, faults have been assumed
as open circuit and short circuit in semiconductor switches. Evaluation of output
voltage, total harmonic distortion (THD), blocking voltage, and efficiency has been
performed for both normal and fault condition. Further, improvement due to fault-
tolerant algorithm in blocking voltage, THD, and efficiency has been demonstrated.
In this paper, simulation is done in MATLAB/Simulink software.
Keywords Blocking voltage · Fault-tolerant asymmetrical multilevel inverter ·

Selective harmonic elimination pulse width modulation · Total harmonic distortion
1 Introduction
Due to the major growth of semiconductor power devices, the power electronic
systems are now extensively used in medium and high voltage. Multilevel inverters
(MLI) are power electronic equipment used to create high power quality for renew-
able power systems, photovoltaic systems, STATCOM, marine electric systems, and
for industrial application [1, 2]. The first multilevel inverter has been invented as a
neutral point clamped (NPC) to create three levels in the output voltage. Successively,
the flying capacitor multilevel inverter (FCMLI) and cascaded H-bridge inverter
(CHBI) become mature [3]. The NPC and FCMLI use extra diodes and capacitors in
its structure unlike CHBI. CHBI has modular structure using semiconductor switches
and it does not need any extra passive components. But, the disadvantage of CHBI is
that it uses ample quantity of semiconductor switches and DC sources as the power
H. Phukan · T. Debela · J. Singh (B)

Department of Electrical Engineering, National Institute of Technology, Silchar, Assam, India
864 H. Phukan et al.
quality increases in the output voltage [4]. Further, ample number of switches and DC
sources decreases the reliability due to increase in the fault probability. To increase
the power quality using n DC sources, a cascaded asymmetrical MLI using 2(n + 1)
bidirectional switches has been implemented [5]. Further to increase the reliability,
complexity, and cost, a sub-MLI-based topology has been reported [6].
First segment of asymmetric multilevel inverter (AMI) is to shoot up the levels in
output voltage by reducing overall cost of AMI and to maintain the power quality at
load end has been described [7–17]. In addition, another segment of an AMI is relia-
bility and the AMI has to maintain the power quality in normal and fault condition. To
increase the reliability of multilevel inverter, new topology has been proposed with
fault-tolerant ability [18]. Reconfiguration of gating signal of semiconductor device
for multilevel inverter in faulty condition has been reported [19]. The open-circuit
fault is assumed, and fault tolerant strategy has been adopted for T-type 3-L inverter
[20].
In this paper, fault-tolerant asymmetric topologies of MLIs have been proposed.
In Sect. 2, basic structure of topology has been given and principle of incorporation
and diminution of DC sources has been described. In Sect. 3, details of the proposed
modulation scheme have been explained. Section 4 includes proposed fault-tolerant
algorithm for fault-tolerant asymmetrical multilevel inverter (FTAMLI). Section 5
includes the simulation results of output voltage, THD, and blocking voltage in
normal and fault condition using three DC source FTAMLI. The simulation results
of two DC source FTAMLI have been discussed in detail. Section 6 summarize the
results.
2 Basic Structure of FTAMLI
The proffer topology of FTAMLI is shown in Fig. 1. This topology is based on H-

bridge. Several H-bridges have been concatenated in cascade to acquire the elevated
number of levels and hence the total harmonic distortion in output voltage is reduced.
If number of level increases, the THD becomes decreases. This topology has advan-
tage that it produces ample number of levels using less DC sources or semiconductor
devices. This topology uses unidirectional devices instead of bidirectional devices.
The proposed topology of asymmetrical MLI is shown in Fig. 1. The magnitude of
the output voltage levels in DC-DC converter are different, due to which the number
of levels in the output voltage is high. If N numbers of DC sources are used:
Numbers of unidirectional semiconductor devices = 4N (1)
The number of levels produced by the proposed topology elevates with the
escalation in number of basic units.
No. of levels in output voltage = 3 N (2)

Fault-Tolerant Cascaded Asymmetrical Multilevel Inverter … 865
LOAD
S11 S31
S21
S32
S12
S22
E1 E2 E3
S23 S24 S33 S34
S14
S13
DC to DC DC to DC DC to DC
converter converter converter
with MPPT with MPPT with MPPT
Fig. 1 Structure of FTAMLI for PV application
3 MC-SHEPWM for FTAMLI
Selective harmonic elimination (SHE) PWM technique is utilized to minimize

selected harmonics levels by optimizing the conduction angles of the output voltage
[21, 22]. Many PWM techniques are popular for MLI to minimize the THD [23].
These techniques are classified by two categories as a high frequency and funda-
mental frequency techniques. Under the category of high-frequency-based sinusoidal
PWM (SPWM) is used for seven-level cascaded MLI. It has been reported that phase
shifted SPWM gives better results than other SPWM techniques [24].
MC-SHEPWM scheme is employed to acquire output voltage levels of FTAMLI.
The FTAMLI uses MC-SHEPWM to create twenty-seven levels of output voltage.
The levels in positive half-cycle are obtained by comparing constant wave C 1 , C 2 , …,
C 13 with sinusoidal wave and levels in negative half-cycle have also been achieved
by using the similar principle. The carrier wave has a constant level and reference
wave is a sinusoidal wave. To obtain L number of levels in output voltage, L − 1
carrier waves are required.
G1 , G2 , …, G13 have been defined as pulse modes (PM). Corresponding to each
pulse mode, the semiconductor switches will be in ON state to create desired level at
output voltage of FTAMLI as shown in Table 1. Pulse modes G1 –G13 can be obtained
by following Boolean operations:
G 1 = C1 ⊕ C2 (3)
G 2 = C2 ⊕ C3 (4)

Table 1 Switching table of three DC source FTAMLI

PM DC source Switch states 1 = ON, 0 = OFF
combination S11 S12 S13 S14 S21 S22 S23 S24 S31 S32 S33 S34
E1 > E2 > E2
G1 −E 2 − E 3 + 1 0 0 1 0 1 1 0 0 1 1 0
E1
G2 +E 3 1 0 1 0 0 0 0 1 1 0 0 1
G3 −E 2 + E 1 1 0 0 1 0 0 1 0 0 1 0 1
G4 +E 2 − E 3 1 0 1 0 1 1 0 1 0 1 1 0
G5 −E 2 + E 3 + 1 0 0 1 0 0 1 0 1 0 0 1
E1
G6 +E 2 1 0 1 0 1 1 0 1 0 1 0 1
G7 +E 1 − E 3 1 0 0 1 0 1 0 1 0 1 1 0
G8 +E 3 + E 2 1 0 1 0 1 0 0 1 1 0 0 1
G9 +E 1 1 0 0 1 0 0 0 1 0 1 0 1
G10 +E 1 + E 3 1 0 0 0 0 0 0 1 1 0 0 1
G11 +E 2 − E 3 + 1 0 0 1 1 1 0 1 0 1 1 0
E1
G12 +E 2 + E 1 1 0 0 0 1 1 0 1 0 1 0 1
G13 +E 2 + E 3 + 1 0 0 1 1 1 0 1 1 0 0 1
E1
G0 0 0 0 0 0 0 0 0 1 0 0 0 0
G1 +E 2 + E 3 − 0 1 1 0 1 1 0 1 1 0 0 1
E1
G2 −E 3 0 1 0 1 0 0 0 1 0 1 1 0
G3 −E 1 + E 2 0 1 1 0 1 1 0 1 0 1 0 1
G4 −E 2 + E 3 0 1 0 1 0 0 1 0 1 0 0 1
G5 −E 2 + E 2 − 0 1 1 0 0 0 0 1 0 1 1 0
E3
G6 −E 2 1 0 1 0 0 1 0 0 1 0 1
G7 −E 1 + E 3 0 1 1 0 0 0 0 1 1 0 0 1
G8 −E 2 − E 3 0 1 0 1 0 0 1 0 0 1 1 0
G9 −E 1 0 1 1 0 0 0 0 1 0 1 0 1
G10 −E 1 − E 3 0 1 1 0 0 0 1 0 0 1 0 1
G11 −E 2 − E 2 + 0 1 1 0 0 0 1 0 1 0 0 1
E3
G12 −E 2 − E 2 0 1 1 0 0 0 1 0 0 1 0 1
G13 −E 2 − E 2 − 0 1 1 0 0 0 1 0 0 1 1 0
E3
G 3 = C3 ⊕ C4 (5)

G 12 = C11 ⊕ C12 (6)

G 12 = C12 ⊕ C13 (7)
Here, ⊕ stands for EX-OR operation and C 1 , C 2 ,…, C 13 are carrier waves.
C1 , C2 , . . . , C13

are the logical outputs. When sinusoidal wave is greater than that
of carrier waves, logical output will be high. The MC-SHEPWM is combination of
fundamental frequency and higher frequency for the proposed topologies.
∞

4Vdc
Vout (wt) = [cos(nα1 ) + cos(nα2 ) + · · · + cos(nα13 )] (8)
nπ n=1,3,5...

40
n=2 (Vn )
2
THD = (9)
v12
In Eq. (8), α 1 , α 2 , …,α 13 are the angles of output voltage V out . By varying these
angles, the magnitude of fundamental output voltage (v1 ) and THD in output voltage
V out given in Eq. (9) can be varied. These angles have been changed by varying sinu-
soidal wave or carrier wave in MC-SHEPWM technique. From the above discussion,
for this topology, modulation index (MI) is defined as:
Magnitude of reference wave

MI = (10)
Magnitude of carrier wave, C13
Figure 2 shows that by changing the modulation index from 0 to 2, the magnitude
of the fundamental output voltage v1 and THD could be varied. It has been observed
that with the elevation of modulation index, v1 elevates. The THD of output voltage
is more at low and high modulation index and THD is less than 5% between when
the modulation index ranges from 0.48 to 1.2.
4 Proposed Fault-Tolerant Algorithm
Figure 3 shows fault-tolerant scheme for FTAMLI. The role of fault-tolerant algo-
rithm is to maintain the voltage at load side in fault condition or in normal condition
without changing the structure of FTAMLI. To show the effectiveness of fault-tolerant
scheme, two types of faults have been considered, i.e., open-circuit fault and short-
circuit fault. Open-circuit fault has been considered by opening of different semicon-
ductor switches in the FTAMLI. The open-circuit fault has been assumed when one or
Fig. 2 Variation in THD and magnitude of fundamental output voltage accompanied by the variation
of modulation index
Fig. 3 Fault-tolerant algorithm for FTAMLI

two semiconductor switches are opened simultaneously. Short circuit is considered

when one or two semiconductor switches have been shorted simultaneously.
In this algorithm, first the output voltage is computed, now if the fundamental
output voltage is not equal to the V desired then v1 goes to the look up table. The look
up table or decision unit has preselected values of voltage obtained by simulation by
taking open-circuit and short-circuit faults for FTAMLI. The v1 is compared with the
preselected values and decision is taken about type of fault that has been initiated.
According to the type of fault, the MC-SHEPWM changes its switching strategy
to decrease the total harmonic distortion in the output voltage V out . After reducing
THD, the modulation index is selected according to V desired . This process will be
stopped when V out is approximately equal to V desired .
In this paper, fault analysis of FTAMLI has been done using three basic units of
FTAMLI connected as shown in Fig. 4. To examine the capability of proposed algo-
rithm for FTAMLI, the fault is assumed as open circuit in semiconductor switch S32 ,
S34 and short circuit in semiconductor switch S32 , S34 as shown in Fig. 4.
The circuit simulation has been done using MATLAB/Simulink software by
connecting resistive load. The frequency of the output voltage is 50 Hz and
RONswitch of each semiconductor switch is selected as 1 milliohm. Here, V desired is
assumed as 15.3 V by choosing E 1 as 8.1 V, E 2 as 5.26 and E 3 as 1.78 V. Figure 5a
shows the 27-level output voltage and its THD by using three DC source FTAMLI
LOAD
S11 S31
S21
S32
S12
S22
E1 E2 E3
S23 S24 S33 S34
S14
S13
FAULTY SEMICONDUCTOR
DEVICES
Fig. 4 Three DC source FTAMLI

Fig. 5 Simulation results: a Twenty-seven-level output voltage and THD of voltage in standard
condition, b output voltage and THD in switch fault condition (switch S34 , open), c output voltage
and THD in fault condition (switch S34 , S32 , shorted), and d output voltage and THD using the
proposed algorithm in fault condition (switch S32 , S34 open or shorted)
in normal condition. Figure 5b shows that without application of the proposed algo-
rithm, the fundamental voltage decreases to 13.05, THD increases to 30.05% when
semiconductor switch S34 is opened. Figure 5c shows that without application of the
proposed algorithm, fundamental voltage decreases to 14.41 V and THD increases
to 13.06% when S32 and S34 is shorted.
Figure 5d shows that when the proposed algorithm has been applied, the funda-
mental output voltage remains constant under fault condition and THD in output
voltage also reduces than that of without algorithm.
It has been observed from Fig. 5, the lower-order harmonics are less by utilizing
the propound algorithm. The detailed results of fundamental voltage, THD of output
voltage, efficiency, and blocking voltage in fault condition have been compared in
Table 2 using the proposed algorithm and without proposed algorithm. It has been
observed from Table 2 that fundamental voltage remains constant is equal to 15.13 V,
but in fault condition the voltage does not maintain its desired voltage.
THD and blocking voltage remain constant with proposed algorithm. The effi-
ciency increases in fault condition by using the proposed algorithm which is equal
to 99.9% whereas in short circuit of switch S32 and S34 condition, it reduces to 0.6
and 0.35%, respectively.
6 Conclusion
Fault-tolerant asymmetrical multilevel inverter topology using principle of addition

and subtraction of DC sources has been presented in this paper. The output voltage and
THD have been calculated for the three DC source FTAMLI by varying modulation
index. It has shown that lower-order harmonics are eliminated under normal and
fault condition. Fault-tolerant scheme has been adopted for FTAMLI. The FTAMLI
has been analyzed by taking open-circuit condition and short-circuit condition in
semiconductor switches. It has been analyzed that the propound algorithm gives less
THD and it maintains desired voltage at load side without any change in structure of
FTAMLI under normal and fault condition. It has been concluded that the blocking
voltage of FTAMLI is less and efficiency is more when the proposed algorithm has
been applied in fault conditions than that without proposed algorithm.
872
Table 2 Detailed simulation results by incorporating open-circuit fault and short-circuit fault in semiconductor switches of FTAMLI
Faculty Without algorithm With algorithm
semiconductor Fundamental THD of V out Total max. Efficiency η Fundamental THD of V out Total max. Efficiency η
switch open output voltage (%) blocking voltage output voltage (%) blocking voltage
circuit (OC) (v1 ) (v1 )
short circuit
S32 (OC) 11.07 49.96 7.24 99.9 15.13 13.44 53.44 99.9
S34 (OC) 13.04 30.32 7.24 99.9 15.13 13.44 53.44 99.9
S32 , S34 (OC) 9.05 64.03 80.16 99.9 15.13 13.44 53.44 99.9
S32 (SC) 15.12 8.19 57 0.6 15.13 13.44 53.44 99.9
S34 (SC) 14.42 9.05 57 0.6 15.13 13.44 53.44 99.9
S32 , S34 (SC) 14.41 13.06 53.44 0.35 15.13 13.44 53.44 99.9
Normal 15.13 3.85 63.56 99.9 – – – –
condition
H. Phukan et al.
References
1. Kouro S, Malinowski M, Gopakumar K, Pou J, Franquelo LG, Wu B, Rodriguez J, Perez MA,

Leon JI (2010) Recent advances and industrial applications of multilevel converter. IEEE Trans
Ind Electron 57(8):2553–2580
2. Nabae A, Takashi I, Akagi H (1981) A new neutral point clamped multilevel inverter. IEEE
Trans Ind Appl 17(5):518–523
3. Rodriguez J, Lai JS, Peng FZ (2002) Multilevel inverter: a survey of topologies, controls and
application. IEEE Trans Ind Electron 49(4)
4. Ruderman A, Schlosberg S (2008) A hybrid asymmetric cascaded multilevel inverter
comprising high resolution and symmetric low resolution parts. In: IEEE 25th convention
of electrical and electronics engineers in Israel, Dec 2008, pp 21–25
5. Babaei E, Hosseini SH, Gharehpetian GB, Haque MT, Sabahi M (2007) Reduction of DC
voltage sources and switches in asymmetrical multilevel converters using a novel topology.
Electr Power Syst Res 77(8):1073–1085 (Elsevier)
6. Babaei E, Hosseini SH (2009) New cascaded multilevel inverter topology with minimum
number of switches. Energy Convers Manage 50(11):2761–2767 (Elsevier)
7. Hinago Y, Koizumi H (2016) A single-phase multilevel inverter using switched series/parallel
DC voltage sources. IEEE Trans Ind Electr 57(8)
8. Gupta KK, Jain S (2013) A multilevel voltage source inverter (VSI) to maximize the number
of levels in output waveform. Electr Power Energy Syst 44:25–36 (Elsevier)
9. Gupta KK, Jain S (2014) A novel multilevel inverter based on switched DC sources. IEEE
Trans Ind Electron 61(6)
10. Alishah RS, Nazarpour D, Hosseini SH, Sabahi M (2014) Novel topologies for symmetric,
asymmetric and cascade switched-diode multilevel converter with minimum number of power
electronic components. IEEE Trans Ind Electron 61(10)
11. Babaei E, Alilu S, Laali S (2014) A new general topology for cascaded multilevel inverters
with reduced number of components based on developed H-bridge. IEEE Trans Ind Electron
61(8):3932–3939
12. Banaei MR, Salary E (2011) New multilevel inverter with reduction of switches and gate driver.
Energy Convers Manage 52(2):1129–1136 (Elsevier)
13. Babaei E, Kangarlu MF, Mazgar FN (2012) Symmetric and asymmetric multilevel inverter
topologies with reduced switching devices. Electric Power Syst Res 86:122–130 (Elsevier)
14. Babaei E, Dehqan A, Sabahi M (2013) A new topology for multilevel inverter considering its
optimal structure. Electr Power Syst Res 103:145–156 (Elsevier)
15. Kangarlu MF, Babaei E, Sabahi M (2013) Cascaded cross-switched multilevel inverter in
symmetric and asymmetric conditions. J IET Power Electron 6(6):1041–1050
16. Esfandiari E, Mariun NB (2013) Experimental results of 47-level switch-ladder multilevel
inverter. IEEE Trans Ind Electron 60(11):4960–4966
17. Suresh Y, Panda AK (2013) Investigation on hybrid cascaded multilevel inverter with reduced
DC sources. Renew Sustain Energy Rev 26:49–59 (Elsevier)
18. Chen A, Hu L, Chen L, Deng Y, Yao G, He H (2004) A multilevel converter topology with fault
tolerant ability. IEEE applied power electronics conference and exposition, vol 3, Feb 2004,
pp 1610–1616
19. Ma M, Hu L, Chen A, He X (2007) Reconfiguration of carrier based modulation strategy for
fault tolerant multilevel inverters. IEEE Trans Power Electron 22:2050–2060
20. Choi V, Blaabjerg F, Lee KB (2015) Reliability improvement of a T-type three level inverter
with fault tolerant control strategy. IEEE Trans Power Electron 30:2660–2673
21. Kumar J, Aggarwal P, Das B (2010) Implemetation of cascade multilevel inverter-based
STATCOM’. IETE J Res 56(2):119–128
22. Shen K, Zhao D, Mei J, Tolbert LM, Wang J, Ban M, Ji Y, Cai X (2014) Elimination of harmonics
in a modular multilevel converter using particle swarm optimization-based staircase modulation
strategy. IEEE Trans Ind Electron 61(10):5311–5322
23. Colak I, Kabalci E, Bayindir R (2011) Review of multilevel voltage source inverter topologies
and control schemes. Energy Convers Manage 52:1114–1128 (Elsevier)
24. Rushiraj GJ, Kapil PN (2016, March) Analysis of different modulation techniques for multi-
level inverters. In: 2016 international conference on electrical, electronics, and optimization
techniques (ICEEOT). IEEE, pp 3017–3024
New H-Bridge for Symmetrical
and Asymmetrical MLI with Reduced
Number of Devices
Tamiru Debela, Hillol Phukan, and Jiwanjot Singh
Abstract A novel symmetrical and asymmetrical single-phase five-level (5LSA)

converter has been presented in this paper. Along with one dc source and one capac-
itor, a bidirectional switch is used in order to obtain the five levels in the output
voltage. The proposed multilevel inverter has two circuits, namely single H-bridge
circuit or main circuit and special combination of switching devices called auxil-
iary circuits. Main advantages of the proposed MLI are it constitutes single DC
source, reduced number of switching devices, low switching system scores, reduces
the complexity, low cost and enhances overall performance. The effectiveness and
feasibility of the proposed topology are validated by simulations under different
condition.
Keywords Isolated DC sources · Multilevel inverter · Multi-carrier pulse width

modulation · Reduced switches · Symmetrical topology
1 Introduction
Inverter at multilevel (IM) plays a vital role to achieve superior power standard
with appropriate control technique and tie of various semiconductor devices. These
inverters use low-rated components, generate high-quality staircase voltage, and
reduce the electromagnetic interference (EMI) to a large scope [1–3]. The major
disadvantage of these types of inverters (MLI) is requirement of isolated DC source,
increased cost, and hence decreased reliability. Therefore, a lot of effort has been
made with different topologies to reduce the number component while improving
performances [4–6]. Multilevel inverter is effectively employed in applications, such
as active power filters [7], AC motor drive [8–10], and linked of renewable energy
T. Debela · H. Phukan (B) · J. Singh

Department of Electrical Engineering, National Institute of Technology, Silchar, Assam, India
J. Singh
876 T. Debela et al.
with the grid [11, 12]. Growth of multilevel has played a vital role in the field photo-
voltaic (PV) [13, 14], topologies [15–17], static compensator [18, 19], and control
and modulation techniques [20–25]. Figure 1 indicates different topology by which
the device count of multilevel inverter reduced.
Out of all multilevel (ML) topologies, the CMI stands out because of voltage
rating devices, possibility of using low, its modular structure, and its developed high-
magnitude voltages. The topology suggested is for both symmetrical and asymmet-
rical multilevel inverter (MLI) structures. The recommended topology minimizes the
number of devices, cost, and complexity as estimated with conventional topologies.
Reduced Device Count Multilevel

Inverter (RDC-MLI) Topologies
With H-bridge Without H-bridge
Cascaded Half Bridge T-type Inverter [15]

Multilevel DC link Inverter
(MLDCL) [18]
Cascaded Biopolar
Switched Series Parallel Switched Cell
Sources Based MLI [19] (CBSC) [16]
Pcked- U Cell
Series Cascaded Topology (UPC) [17]
Switched Sources
(SCSS) based MLI [20]
Multilevel Module (MLM)

based [21]
Topology wich required Isolated DC Level

Reversing Voltage (RV)
Topology [22]
Topology wich required nonIsolated DC
Level
Two Switched Enabled
level Generation (2SELG)
based MLI [23]
Fig. 1 Classification of reduced device count multilevel inverter topologies

New H-Bridge for Symmetrical and Asymmetrical MLI … 877
The characteristics of propound topology and modulation schemes used to

1. Decrease total power count.
2. Increase reverse voltage at each switches limited to V in .
3. Less power losses because number of components, the number of switch on state
is reduced.
4. The voltages of the capacitor are independent of load power, and the voltages are
self-balanced through the condensers.
2 Proposed MLI Topology
The suggested topology is mostly used for improvement of self-balancing, reduced

components and uniform voltage blocking ability at all switches. Figure 2a indicates
the configuration of conventional three-level cascaded H-Bridge which consists of 4
switches and one separate DC source. Figure 2b is the proposed topology five levels,
which is also composed of four switches and one separated DC source, capacitor,
and an auxiliary switch.
The auxiliary switches are attached in series with the capacitor; the capacitor
is discharged at the beginning. It slowly starts to charge 0.5 V when the inverter
switch becomes ‘ON.’ The capacitor maintains 0.5 V at each cell. Figure 2c shows
the general framework designed to achieve higher output levels. Figure 3 shows the
modes of operation of the proposed symmetric topology for 5L performance. The
highlighted red indicates the current path in all from (i) to (x).
2.1 Proposed Symmetrical Topology
In the proposed symmetrical only one DC source is required for new five levels.
Extended category of the propound cascaded configuration is shown in Fig. 2c.
Table 1 indicates the comparison between different five-level topologies, and mode
of operation is given in Table 1.
For ‘N’ number of cells, the general proposed symmetric topology is given by
(VS1 = VS2 = VS3 = . . . = VS N )

VDC total = Capacitortotal = N (1)
VDC total = Capacitortotal = N (2)
Diodestotal = 4 × N (3)
Outputlevel = 4 × N + 1 (4)
S1 S1
S3 S3
D1 D2
LOAD LOAD
Vdc D4
Vdc D3
C
S2 S4 S2 S4
(a) (b)
S1
S3,1
D1,1 D2,1
Vdc D4,1
D3,1
C
S2,1 S4,1
L
O
A
D
S1,n
S3,n
D1,n D2,n
Vdc n D4,n
D3,n
C n
S2,n S4,n
(c)
Fig. 2 a Three-level structure of conventional CHB. b Five-level structure of propound topology.
c Extended proposed topology
2.2 Proposed Topology for A Symmetrical MLI
For asymmetrical topology, the DC sources are more and have different values.
In this asymmetrical MLI, the values of DC source are assigned in three different
algorithms as shown in Table 2. For conventional asymmetric MLI, value of DC
source is increasing in binary combination (2:1) or tertiary combination (3:1). Table 2
shows that in addition to binary and tertiary the new and novel algorithm is proposed
to achieve highest number of output (Table 3).
S1 S1
S3 S3
D1 D1
D2 D2
LOAD LOAD
Vdc Vdc D4
D4
D3 D3
C C
S2 S2 S4
S4
(i). Vo = CV (iL>0) (ii). Vo = CV (iL<0)
S1 S1
S3 S3
D1 D1
D2 D2
LOAD LOAD
Vdc D4 Vdc D4
D3 D3
C C
S2 S4 S2 S4
(iii). Vo = VS (iL>0) (iv). Vo = VS (iL<0)
S1 S1
S3 S3
D1 D1
D2 D2
LOAD LOAD
Vdc D4 Vdc D4
D3 D3
C C
S2 S4 S2 S4
(v). Vo = 0V (iL>0) (vi). Vo = 0V (iL<0)
S1
S1
S3
S3
D1
D2 D1
D2
LOAD LOAD
Vdc D4 Vdc D4
D3 D3
C C
S2 S4 S2 S4
(vii). Vo = - (VS-CV) (iL<0) (viii). Vo = - (VS-CV) (iL>0)
S1
S1
S3
S3
D1
D1 D2
D2
LOAD
LOAD
Vdc Vdc D4
D4 D3
D3
C C
S2 S4
S2 S4
(ix). Vo = - (VS) (iL<0) (x). Vo = - (VS) (iL>0)
Fig. 3 Proposed symmetrical topology operation modes for five-level performance

Table 1 Comparison between different five-level inverter topology for symmetrical MLI
Components NPC CHB FC [26] [27] [28] [4] [5] Proposed
PS 8 8 8 6 6 6 8 6 4
AS 0 0 0 0 0 0 0 0 1
DC sources 1 2 1 2 1 2 1 1 1
Capacitors 4 – 10 0 1 – 1 1 1
Diodes 20 8 8 6 7 6 8 6 4
NPC- Neutral Point Clamped, CHB- Cascaded H-bridge, FC- Flying Capacitor, PS- Power Switches,
AS- Auxiliary Switch
Table 2 Mode of operation states for Fig. 3

Output voltage Switches state ‘ON’ Capacitors State when “iL” Capacitors state when “iL”
<0 >0
(C v ) S4 and S5 Capacitor charges (Fig. 2 Capacitor discharges
(ii)) (Fig. 2 (i))
Vs S1 and S4 No effect (Fig. 2 (iv)) No effect (Fig. 2 (iii))
V0 S2 and S4 No effect (Fig. 2 (vi)) No effect (Fig. 2 (v))
(V s − C v ) S3 and S5 Capacitor charges (Fig. 2 Capacitor discharges(Fig. 2
(vii)) (viii))
(V s ) S2 and S3 No effect (Fig. 2 (i)) No effect (Fig. 2 (x))
Table 3 The proposed

Algorithm DC source Number of Configuration
topology with different level
values output level
incorporates with ‘N’ number
of cells 1st VS1
= VS2
= 2(N + 1) − 1 Asymmetrical
20 21
VS N
··· = 2 N −1
2nd VS1
= VS2
= 2N Asymmetrical
30 31
VS N
··· = 3 N −1
3rd VS1
= VS2
= 5N Asymmetrical
50 51
(Proposed) VS N
··· = 5 N −1
2.3 Conduction Losses
The one important part during design of the system is calculation of losses. MLI oper-
ates in three different modes, namely blocking mode, switching mode, and conduc-
tion mode. In blocking mode, no current will flow across the devices. The transition
of switch is called switch losses, i.e., transistors from ‘ON’ to ‘OFF,’ and conduction
mode is the power losses during ON state. At instant ‘t,’ the device losses are given.
CT (t) = [VT + RT × i β (t)] × i(t) (5)
CD (t) = [VD + RD × i β (t)] × i(t) (6)
where
CT transistor conduction losses
CD Diode conduction losses
VT ON state of voltage transistor
VD ON state of voltage diode
RT Forward resistance of transistor
RD Forward resistance of diodes
β Constant.
2.4 Power Losses
Power loss is directly proportional to blocking voltage and switching frequency
PLOSS, i αVB &Ploss, i α f S (7)
For symmetrical 5-level topology loss is given by
PLOSS = C × VB × f S (8)
Let f S is switching frequency and is fundamental frequency, and the above

equation can be written as (Table 4)
PLOSS, PROPOSED, 5 - LEVEL = C

(VDC × f + 2 × 2 × VDC × f S + 2 × 2 × VDC × f o ) (9)
But (f S f o ) So Eq. (9) becomes
PLOSS,PRO,5 - LEVEL = C × VDC (5 × f S + 4 × f o ) (10)
Table 4 Parameters of the

Components Value Unit
simulated model
Input voltage 20 V
Output voltage 20 V
Capacitors 3300 µF
Frequency 50 HZ
PLOSS,PRO,5 - LEVEL = 5 × C × VDC × f S (11)
Figure 4a–c displays the output parameters for 5L-MLI, respectively. Figure 5a, b
indicates the output parameters under dynamic change of load. Note as the load
inductive load is added in parallel to the resistive load, only the magnitude of current
will change.
(a) (b)
(c)
Fig. 4 Simulation results. a Output voltage under resistive condition, b output current under
resistive load, c Result for five-level inverter charging and discharging capacitor
(a) (b)
Fig. 5 Simulation results. a Output voltage under dynamic load. b Inverter output current for five
level inverter under dynamic load in the proposed topology
4 Conclusion
Both symmetrical and asymmetrical MLIs are discussed. Supported topology was
connected in a cascade manner for medium to high voltage. At the beginning, the
symmetrical five-level inverter is presented. The operation of propound topology
and generalized structure of symmetrical description has been described in detail.
At last, the asymmetrical proposed topology has been presented. The propound
topology utilizes lesser number of device count when compared with recent topolo-
gies, thus reducing the cost. Additional superiority of the propound topology is
self-balancing, minimum number of switches, complexity is reduced. Finally, the
developed proposed topology confirmed the feasibility and has been verified through
simulation.
References
1. Akagi H (2011) Classification, terminology, and application of the modular multilevel cascade
converter (MMCC). IEEE Trans Power Electron 26(11):3119–3130
2. Franquelo LG, Rodriguez J, Leon JI, Kouro S, Portillo R, Prats MAM (2008) The age of
multilevel converters arrives. IEEE Ind Electron Mag 2(2):28–39
3. Nami A, Zare F, Ghosh A, Blaabjerg F (2011) A hybrid cascade converter topology with series-
connected symmetrical and asymmetrical diode-clamped H-bridge cells. IEEE Trans Power
Electron 26(1):51–65
4. Ebrahimi J, Babaei E, Gharehpetian GB (2012) A new multilevel converter topology with
reduced number of power electronic components. IEEE Trans Ind Electron 59(2):655–667
5. Najafi E, Yatim A, Samosir AS (2008) A new topology—reversing voltage (RV)—for multi
level inverters. In: Proceedings of IEEE 2nd international power energy conference, 1–3 Dec
2008, pp 604–608
6. Kangarlu MF, Babaei E (2013) A generalized cascaded multilevel inverter using series
connection of sub multilevel inverters. IEEE Trans Power Electron 28(2):625–636
7. Alepuz S, Busquets-Monge S, Bordonau J, Gago J, Gonzalez D, Balcells J (2006) Interfacing
renewable energy sources to the utility grid using a three-level inverter. IEEE Trans Ind Electron
53(5):1504–1511
8. Gupta KK, Ranjan A, Bhatnagar P, Kumar Sahu L, Jain S (2016) Multilevel inverter topologies
with reduced device count: a review. IEEE Trans Power Electron 31(1):135–151
9. Hagiwara M, Nishimura K, Akagi H (2010) A Medium-voltage motor drive with a modular
multilevel PWM inverter. IEEE Trans Power Electron 25(7):1786–1799
10. Cheng Y, Qian C Crow, ML, Pekarek S, Atcitty S (2018) A comparison of diode-clamped and
cascaded multilevel converters for 2168-6777 (c) 2018 IEEE
11. Barrena JA, Marroyo L, Vidal MAR, Apraiz JRT (2008) Individual voltage balancing strategy
for PWM cascaded H-bridge converter-based STATCOM. IEEE Transactions Ind Electron
55(1):21, 29
12. Choi W, Kang F (2009) H-bridge based multilevel inverter using PWM switching function. In:
Proceedings of 31st international telecommunications energy conference, 18–22 Oct 2009, pp
1–5
13. Ristow A, Begovic M, Pregelj A, Rohatgi A (2008) Development of a Methodology for
improving photovoltaic inverter reliability. IEEE Trans Ind Electron 55(7):2581–2592
14. Rodriguez J, Lai J-S, Peng FZ (2002) Multilevel inverters: a survey of topologies, controls,
and applications. IEEE Trans Industrial Electron 49(4): 24–738
15. Gupta KK, Jain S (2013) Multilevel inverter topology based on series connected switched
sources. IET Power Electron 6(1):164–174
16. Manjrekar MD, Steimer PK, Lipo TA (2000) Hybrid multilevel power conversion system: a
competitive solution for high-power applications. IEEE Trans Ind Appl 36(3):834–841
17. Lai J-S, Peng FZ (1996) Multilevel converters-a new breed of power converters. IEEE Trans
Ind Appl 32(3):509–517
18. Liu C-H, Hsu Y-Y (2010) Design of a self-tuning PI controller for a STATCOM using particle
swarm optimization. IEEE Trans Ind Electron 57(2):702, 715
19. Walker GR, Sernia PC (2004) Cascaded DC-DC converter connection of photovoltaic modules.
20. Gupta KK, Jain S (2014) Comprehensive review of a recently proposed multilevel inverter.
IET Power Electron 7(3):467–479
21. Ruiz-Caballero DA, Ramos-Astudillo RM, Mussa SA, Heldwein ML (2010) Symmetrical
hybrid multilevel DC–AC converters with reduced number of insulated DC supplies. IEEE
Trans Ind Electron 57(7):2307, 2314
DC voltage sources. EEE Trans Ind Electron 57(8):2643–2650
23. Najafi E, Yatim AHM (2012) Design and implementation of a new multilevel inverter topology.
IEEE Trans Ind Electron 59(11):4148–4154
24. Rahim NA, Chaniago K, Selvaraj J (2011) Single-phase seven-level grid connected inverter
for photovoltaic system. IEEE Trans Ind Electron 58(6):2435–2444
DC voltage sources. IEEE Trans Ind Electron 57(8):2643–2650
26. Babaei E (2008) A cascade multilevel converter topology with reduced number of switches.
27. Ounejjar Y, Al-Haddad K, Gregoire LA (2011) Packed U cells multilevel converter topology:
Theoretical study and experimental validation. IEEE Trans Ind Electron 58(4):1294–1306
28. Gui-Jia S (2005) Multilevel DC-link inverter. IEEE Trans Ind Appl 41(3):848–854
Different EDGE Detection Techniques:
A Review
Chinmoy Ghosh, Suman Majumder, Sangram Ray, Shrayasi Datta,

and Satyendra Nath Mandal
Abstract An image is represented by an important feature called EDGE which is

very much useful for understanding the image properly. Edge can be considered as the
discontinuity from one image pixel to another. So, edges are the boundaries between
textures and accordingly edge detection is very much useful for segmentation, object
detection and recognition purpose. Several methodologies exist using which the
information about the edges can be extracted. This paper presents a brief study
of different edge detection methodologies along with their performance issues and
limitations. A detail discussion related to benefits and drawbacks of each algorithm
are also discussed.
Keywords Edge detection · Fuzzy Inference System (FIS) · Genetic Approach

(GA) · Line Segment Detector (LSD) · Xilinx System Generator (XSG) · Adaptive
Network-based Fuzzy Inference System (ANFIS)
C. Ghosh
Department of Computer Science & Engineering, Jalpaiguri Government Engineering College,
Jalpaiguri 735102, India
S. Majumder (B) · S. Ray
Department of Computer Science and Engineering, National Institute of Technology Sikkim,
Ravangla, Sikkim 737139, India
S. Ray
S. Datta
Department of Information Technology, Jalpaiguri Government Engineering College, Jalpaiguri
735102, India
S. N. Mandal
Nadia, India
886 C. Ghosh et al.
1 Introduction
Edge detection procedure is utilized to identify an edge of an image [1–10]. For

solving the issues related to lack of stability, differences and distinguishing opera-
tions of an image, edge detection technique is used mainly in applications of image
processing for different organizations to symbolize the meaningful features and
significant information of edges [1–5]. This procedure is mainly used to keep only
the appropriate structural features of an image and diminish the size of an image
after filtering the unnecessary information. At the time of detection and reconstruc-
tion of edges, the redundancies are removed that present within most of the images.
To segregate the images into respective regions using the image segmentation proce-
dure, the edge detection procedure is used as it represents the boundary of objects
[1, 5–8, 11, 13].
2 Literature Review
In this section, different edge detection techniques are analyzed and the details are
mentioned below.
One of the widely used edge detection algorithm is Canny Edge detection method
invented by Canny [1]. Canny Edge detection has more engineering and mathematical
concept using gradients. It is an optimal edge detection technique with important
criteria such as good edge detection. It is kept in mind at the time of detection of
edges that is as close as possible to the true values and maintaining the accuracy of
the localization. On the other hand, there should not be multiple responses and the
detector must return one point only for each edge. For doing any operation using
canny edge detection, at first it is required to make the edge as smooth applying
Gaussian filters. Thereafter computations are performed for the derivative of the
filtered image in both X and Y coordinate using Sobel operator. Further, magnitude
and direction is obtained using the equation |G| = |Gx | + |Gy | and theta = invtan
(Gy /Gx ) followed by applying suppression procedure of non-maximum. The main
concept is to suppress the pixels which are not local maximum. Finally “hysteresis
threshold” are applied. There are two types of thresholds (high and low) and based
on the two threshold values the gradient is calculated. If the gradient of the pixel is
higher than the upper threshold then this is an edge pixel otherwise it is a non edge
pixel. If the threshold of the pixel is between high and low and connected to an edge
pixel, then it is also considered to be an edge pixel.
In 2012, Sriramakrishnan and Shanmugam [2] proposed an algorithm based on
neuro-fuzzy system where Sobel and Laplacian methods are applied on input images.
Further, fuzzy rules are applied on the respective output. Neural network is used on
input layer. Accordingly, Dhir [3] proposed another algorithm based on Neuro-Fuzzy
Inference system. Using the method of global thresholding, the binary images in term
of either numerical zero (0) or one (1) are derived from the gray label. On the other
Different EDGE Detection Techniques: A Review 887
hand, an adaptive network-based fuzzy inference system (ANFIS) can be used with
3 × 3 masks to prepare the network to higher level.
Rani and Tanwar [4] proposed another algorithm based on neuro-fuzzy system.
Using the back propagation method, an adaptive network-based fuzzy inference
system (ANFIS) can be constructed based on FIS and its respective membership
function values are adjusted accordingly. It can be derived from the 4 inputs and 1
output based FIS system. Using the gray levels which is denoted as L, an intensity
of edge is normalized between the range [0 to (L − 1)] and the relevant pixels under
the edge are partitioned to two different sets S 0 and S 1 .
Later, Petrova and Hostalkava [5] proposed an edge detection technique using
wavelet transformation. This technique can be used in medical images for efficient
edge detection. It is the most appropriate method for non stationary signal. The
fundamental functionality may vary both in frequency and spatial range. In this case,
low frequency (average intensity values) and high frequency component (for the
edges) are obtained after designing the wavelet transformation. On the other hand,
for an original image, a low pass filter in rows as well as columns is applied. As a result
approximated image is achieved. The simplest method is to replace approximated
image by zero and remove all low frequency components, and reconstruct the image.
The remaining component is of high frequency (refers an edge). Another method is to
pass the approximated images through a good edge detector (especially canny) where
modified approximated pixels and the remaining pixels are used in reconstruction
which leads to an edge.
In 2014, Bar et al. [6] proposed an application based on Mumford–Shah model for
implementation of image segmentation and image restoration process. Segmentation
is the process of labeling the images and paints it with gray values. While detecting
an edge there maybe two types of penalties—(1) Penalty for edges and (2) Penalty
for errors. The basic concept of this model is to write basic formula to compromise
between representations of image to a simplified image which is not too far from the
original image. The Mumford–Shah model is also used in image restoration. Due
to simplicity and efficiency, Mumford and Shah model is also used in Chan Vase
segmentation [6, 7]. In 2012, Getreuer [7] proposed the performance of Chan-Vese
segmentation. Although many methods are reliant on the Chan-Vese segmentation
but some of them do not work properly on edges. On the other hand, this scheme
affords an efficient two-phase model for image detection. This segmentation tech-
nique is very much efficient than Snake methods [7, 9] and works even if the topology
has transformed and derived from Mumford–Shah model. However, the technique
introduced by Osher and Sethian [7], can crack this problem applying the level set
function as the zero crossing rather than declaring the boundaries explicitly [7, 8].
In 2012, Chan and Vese [8] proposed an active contour model without edges.
The process of labeling the location in the image where change rate is higher than
adjacent locations is known as edge detection. With the gradient of the image we
get the information about the location where major change in intensity in the image
exists. So, the gradient gives the local information. The aim of active contour is
to integrate the local measurements. Edge integration is the process of combining
local preferably sparse and non contiguous edges into a meaningful edge curve for
888 C. Ghosh et al.
Original Gaussian Hysteresis Canny Edge VHDL

Image Filtering Thresholding detection Synthesis
Fig. 1 VHDL based canny edge detection
segmentation. There is also a relationship between local edges and active contours
that are going to compute it. Based on the curve evolution, this new model picks up the
respective objects from an offered image and Mumford–Shah function is utilized to
segment and position the levels of the corresponding object whose boundaries are not
distinct by gradient. This, this model attempts as active contour models combining
different lively models along with Mumford–Shah model.
Alvarez et al. [9] proposed the Snake algorithm that is based on real-time system.
For modeling the snake evolution, a level set approach is used. To solve using level
set methods for the partial differential equations (PDE), the Euler-Lagrange equation
is exercised. Using the standard methods for estimation of contour distance function,
various problems regarding the PDE is solved. Morphological operators associated
with PDE components and an advanced morphological approach that is based on
simple and fast numerical analysis has been proposed for the evolution of the PDE
snakes approach. In this method, distance function of a contour cannot be estimated
because the level set is just a binary constant function.
In 2014, Rashid et al. [10] recommend an algorithm which is VHDL based Canny
edge detection method. The aim of this algorithm is to detect the edge automatically
of a digital image and works based on the concept of canny edge detector algorithm
followed by the Gaussian filtering on Xilinx System Generator (XSG). After that
the VHDL synthesis is done for detecting the edge of the digital images efficiently.
It mainly used for edge detection of brain tumors. The detail procedure is given in
Fig. 1.
Further, Priyadarshini and Sahoo [11] proposed a new technique of edge detection
based on the method of addition and division. The method is simple and based
on the automatic computation of threshold during edge detection. It uses single
dimensional array for storing the computations those are performed on the image
pixels for calculation of the image threshold. The proposed method was applied on
a number of images giving good and satisfying results compared to Sobel method
[2, 11]. Later, Biswas and Sil [12] suggested an improved canny edge detection
algorithm that stands on type-2 fuzzy sets. The algorithm is effective in handling the
uncertainties arising in automatic threshold value selection for image segmentation
using canny algorithm. Using type-2 fuzzy logic, it selects single threshold from
image histogram and it efforts on gradient histogram produced by Sobel operator.
The algorithm was found effective for many images especially hand radiography
image. Moreover, Rashmi et al. [13] discussed an algorithm called Bilateral Filtering
which calculates the weighted average intensity values of neighboring pixels of an
image to replace with the intensity value of the pixel. The algorithm has better noise
removing capability as well as better edge preserving capabilities than traditional
canny operator.
In 2014, Shamlee [14] proposed an algorithm that removes the entire dependency
between the blocks and it is referred to as a split canny detection of edges. For the
distribution of gradient of the block and compute the block type based on predefined
threshold values, the above said algorithm is used. To calculate the thresholds values
of hysteresis related to block, histogram is used. This algorithm has reduced latency
and can possibly be pipelined with other block based codes. The algorithm is scal-
able and has high edge detection performance even full HD videos. In 2014, Liu
et al. [15] proposed a modified canny edge detection iterative algorithm. Using arith-
metical iteration, it calculates optimal high and low dual threshold and then using
mathematical morphology the images are made with sharp quality edges. It has good
noise suppression and interference reduction capability. The algorithm is effective
for edge detection for medical cell image to detect characteristics of roundness, size
and number of cells but not good for Lena that needs extraction of image details.
Further, Gao et al. [16] proposed an algorithm based on alpha weighted quadratic
filter. The positive alpha power is known for its ability to enhance image contrast and
the negative alpha power is also used for suppressing the non edge portion. Thus,
positive/negative power law can be utilized in the filter design for further enhancing
the desired portions. In 2014, Dollár and Zitnick [17] proposed a new model using
structured forests method that is used for the detection of edges rapidly. It tries to
patch of edges as straight lines or T-junctions to form the local structure such are
detected clearly and applied in discrete space to get real time.
Later, Maladakara and Vanamala [18] proposed an algorithm to design an
improved Canny Edge detection algorithm. It is same as the actual canny algo-
rithm but the only difference is it is applied at the block level. Blocks are classified
into smooth, texture, hybrid and strong. This classification helps to minimize the
computation using fuzzy logic for block classifications. The required steps of canny
edge detections are all performed at first. Then fuzzy logic for block classification
is applied. Later hysteresis threshold is applied and this is calculated on the basis of
percentage of strong edges present in a block.
In 2017, Yu et al. [19] proposed a method called Deep Category—Aware Semantic
Edge Detection (CASENet) which is based on deep semantic learning based edge
detection method. In this method two or more semantic classes are connected by
each edge pixel and follow an architecture where edge activation is performed based
on different category at the top convolution layer. These layers share and fuse with
the same set of bottom layer feature using convolution neural network (CNN). On
the other hand, from the given K numbers of semantic edge categories, CASENet
maintains a deep network. The network generally constructs K numbers of separate
edge maps where each map indicates the edge probability of a certain category.
CASENet also maintains similar network like deeply supervised network (DSN).
Later, Liu et al. [20] proposed a convolutional neural network (CNN) based edge
detection technique which is called richer convolutional features (RCF). Based on
the information obtained by different convolution layers in edge detection, a network
is formed to produce outputs of intermediate layers using VGG16 network which
has five numbers of convolutional layers. Further the layers are segregated into five
numbers of stages and within the stages pooling layers are connected after each stage.
890 C. Ghosh et al.
The RCF generally takes an image with any arbitrary size as an input and produce
an edge as output with same size.
In 2018, Agrawal and Rosepreet [21] proposed a method to detect edges of dental
radiographs using three different techniques—(1) Sobel type (2) Fuzzy Logic System
Type 1 (T1FLS), and (3) Interval Type 2 Fuzzy logic System (T2FLS). To obtain
image information or pixels as a form of dental radiographs or X-ray of different
patients regarding the healthy and effective teeth, the aforesaid three techniques
are used for edge detection. Sobel type mainly works on the gray scale values and
computes gradients of every pixel for digital image. T1FLS is mainly used in less
contrasted images related to biomedical application and easy detection of edges.
However, T1FLS faces many difficulties where edges are not continuous, not very
clear or blurred. But T2FLS is more efficient than T1FLS. Based on some predefined
selected threshold values, a comparison is done among the three types and it is found
that more pixels related to edges are obtained from T2FLS and analyzed easily than
Sobel and T1FLS methods [1, 2, 11, 21].
In 2019, Halder et al. [22] used a novel approach for the detection of edges
based on Richardson’s extrapolation technique to quantify the strength of edges.
This extrapolation is used to pick up convergence rate which is related to some
sequence and a numerical analysis is done using this aforesaid method. In this method,
the strength based on the neighboring pixel value is calculated in relation to the
formula obtained from Richardson’s extrapolation technique. On the other hand, to
calculate the strength of gradient in different directions related to detection of edges,
Richardson’s extrapolation is utilized. For the identification of respective edges of a
binary image this technique is also used.
A novel spatial FFT-based filtering approach as well as a boundary point detection
algorithm is introduced in 2019 by Mineo et al. [23] for direct production of low
noise supported surfaces from point cloud data which is not based on predefined
threshold values. For this purpose, two methods are used—(1) Detection of boundary
points and (2) Edge reconstruction. To detect boundary points, the BPD algorithm
is used and the main benefit of this algorithm is that it does not dependent on any
predefined threshold values. The algorithm identifies maximum numbers of boundary
points from the region under every cloud application. The algorithm permits and
merges the unlabeled surface points under the cloud into two numbers of groups—
(1) internal points and (2) boundary points. This algorithm is mainly used to detect the
boundary points and performs better than the existing methods. The boundary points
are detected by BPD algorithm are further clustered and grouped together to order the
points under the respective cluster. Later on, from every cluster, the closed boundary
paths are collected. To resemble the real surface borderlines, the boundary curves are
used to make it smooth. For this purpose and further reconstruct of the surface points
under the borderline of the respective cloud, a novel boundary smoothing algorithm
termed as RBS was introduced. Afterward, the closed boundaries are created using
the clustering of the boundary points and the relative edges are created using the
RBS algorithm after identifying each closed boundary.
Later, a novel methodology is applied to measure system Modulation Transfer
Function (MTF) directly from natural or live scenes like camera or CCTV, using the
Table 1 ROI parameters

Parameters ISO 12233 Proposed scheme
ROI size >64 × 80–500 27 × 4 Pixels
Pixels
Angle <45° 2.5–42.5°
Constant <10:1 No such constant
limit
Number of edge Not applicable 15
points in ROI
Slanted Edge Method and ISO 12233 algorithm [24]. In this method edge detection
techniques are used to select and extract suitable step edges from pictorial images.
The system performance scheme-Spatial Frequency Response (SFR) is used in ISO
12233. Different edge selection techniques are proposed in ISO 12233 for digital
imaging systems for SFR. However, in this case slanted edge SFR is used that works
mainly on edge charts as an input. The chart is captured including at least one gray
scale that is taken as a photograph with an angel. Another term used in slanted edge
SFR is Region of Interest (ROI) that is a pre-processing activity which includes
image linearization with respect to input luminance and edge selection. In this case,
two edge detection methods are used—(1) Canny approach and (2) a matched filter.
The maximum MTF in the database should be optimum performance of the imaging
system derived from a perfect edge. The ROI parameters are described in Table 1.
The advantage of this type of system is that it can derive the system performance
in real-time condition instead of any laboratory condition.
In 2020, Shi et al. [25] proposed a novel method related to edge detection for polari-
metric synthetic aperture radar (PolSAR) images to pull out all the features related
to edge. This method creates a limitation for the speckle noise within heterogeneous
areas for weak and strong edges. For this purpose two methods are implemented—
(1) To detect weak edges, polarimetric constant false alarm rate (IP_CFAR) is used
and (2) To detect edges in heterogeneous areas but it is unable to maintain some
of the weak edges, weighted gradient-based (WG) detector is used. But it some-
times produces false alarm rate due to speckle noises. There are three numbers of
methods are used for PolSAR images—(1) feature-based edge detection methods,
(2) multiple-channel fusion based methods and (3) statistics based edge detection
methods. For this first method, some edge details are lost as it omits some weak
edges. For the second method, it is unable to detect the edge details in heterogeneous
terrain types such as the thin roads in the urban area. To overcome these limitations,
a wavelet-based hybrid edge detection method is used [5, 25].
In 2020, Bausys et al. [26] proposed a multi-criteria based decision-making
(MCDM) platform called suitability for satellite images. MCDM works mainly on
the concept of visual features of edge detection algorithms. Using the neutrosophic
weighted aggregated sum product assessment (WASPAS) method, the algorithms
892 C. Ghosh et al.
ranked according to their suitability and there algorithms are selected based on prim-
itive six numbers of criteria- contrast, roughness, density, number of different regions,
size of the individual image regions and processing time.
3 Tabulated Review
In this section a tabulated review on several methodologies using which the informa-
tion about the edges can be extracted is given. The tabulated review is summarized as
Table 2 that includes different methodologies, algorithms used and performance anal-
ysis related to benefits and drawbacks. There is no perfect methodologies regarding
edge detection are found so far but it depends on the requirements and adaptability.
4 Conclusion
Edge detection is the basic concept that needs to be focused in most of the image
processing work. In this paper, an effort has been made to provide the brief overview
about some of the robust, well known and successful edge detection techniques. It is
observed that most of the recent algorithms are based on/modification of well known
Canny edge detection algorithm. However, after blooming success of deep learning,
the deep learning based edge detection techniques is gaining attention nowadays.
Table 2 Tabulated review

Schemes Algorithm (s) used Performance analysis
Canny [1] Canny edge detection It is one of the basic algorithms. It
was the first edge detection
algorithm which was based on
second order derivation. It is
motivated by human vision and
has the ability to detect edges
within wider ranges in all
direction. However, this scheme is
sensitive to noise and there is
localization problem
Sriramakrishnan and Hybrid approach of For the input purpose, both the
Shanmugam [2] neuro-fuzzy-1 Laplacian and Sobel methods are
imposed. On the other hand, for
the output purpose, Fuzzy rules
are incorporated. Neural network
is applied on input layer
Dhir [3] Hybrid approach of Using the method of global
neuro-fuzzy-2 thresholding, the binary images in
term of either numerical zero (0)
or one (1) are derived from the
gray label. Further, Adaptive
network-based fuzzy inference
system (ANFIS) can be used
along with 3 × 3 masks to train
the network to higher level
Rani and Tanwar [4] Hybrid approach of Using the back propagation
neuro-fuzzy-3 method, an adaptive
network-based fuzzy inference
system (ANFIS) can be
constructed based on FIS. It can
be derived from the 4 inputs and 1
output based FIS system.
Moreover, using the gray levels
denoted as L, an intensity of edge
is normalized between the ranges
[0 to (L − 1)]. The respective
pixels under the edge are
partitioned into two different sets
S 0 and S 1
Petrova and Hostalkava Edge detection in medical images Modification of approximation
[5] using the wavelet transform coefficient by zero gives false
edges but approximation with
canny edge detectors gives better
result. This scheme is appropriate
for non stationary signal
(continued)
894 C. Ghosh et al.
Table 2 (continued)
Bar et al. [6] Mumford and Shah Model is Through this algorithm we can
used for image segmentation and obtain an image near about the
image restoration actual image which has less noise
and less errors. This scheme is
widely used in image restoration
Getreuer [7] Chan-Vese segmentation Using the set function of any
level, the boundary of
segmentation is generally depicted
where the segmentation technique
is very much efficient than Snake
methods and works even if the
topology has changed
Chan and Vese [8] An active contour model without This model is inspired by level
edges sets and it has wide use in medical
fields. The objects whose
boundaries are not defined by
gradient are mainly detected by
this system. As the boundaries
related to the location are detected
very healthy, there is no need to
formulate the image in a smooth
form though some noise
component is present within the
image
Alvarez et al. [9] A real-time morphological snake Here the level set is binary
algorithm piecewise constant and no need of
defining the contour distances,
i.e., no need of geodesic distance
computation. So, this approach
becomes easier to implement
Rashid et al. [10] VHDL based Canny edge This edge detection technique
detection algorithm automatically detects edges of
digital images. This is also very
efficient and used in medical
image detection
Priyadarshini and Sahoo Additions and divisions based This algorithm computes
[11] edge detection method threshold by simple calculations
which is easy compared to Sobel
operator. This algorithm needs
fewer computations
(continued)
Table 2 (continued)
Biswas and Sil [12] Improved canny edge detection The algorithm considers the
using type-2 fuzzy logic uncertainties in the image and
computes the threshold to the
segment of the gradient image.
The algorithm helps in selecting
threshold besides the uncertainties
in an image
Rashmi et al. [13] Bilateral filter based adaptive The algorithm computes weighted
Canny edge detector average of intensity of a pixel
using its nearby pixels and
reduces noise in a digital image.
The algorithm works well for high
resolution images and has noise
reduction capacity but it is not
good for low resolution images
Shamlee [14] A split canny edge detection To reduce the loss of edges within
the specific regions and compute
the threshold values, a histogram
is used. The algorithm has low
memory requirements, low
latency, high throughput and great
edge detection performance
Liu et al. [15] Canny edge detection based on The algorithm computes dual
iterative algorithm threshold—high and low by
iterative computation. The
detected edges are sharpen using
this algorithm. It reduces the noise
interference and gives the best
threshold for image segmentation
using mathematical morphology
but cannot involve large
calculations if needed
Gao C et al. [16] Alpha weighted quadratic filter Positive alpha is used for
based edge detector enhancing image contrast whereas
negative alpha is used for
suppressing the non edge portion.
This scheme safeguards contrast
while concurrently get rids of
noise
Positive/Negative power law can
be utilized in the filter design for
further enhancing the desired
portions
(continued)
896 C. Ghosh et al.
Table 2 (continued)
Zitnick and Dollár [17] Using the Structured Forests Applied in discrete space to get
method, detection of edges is real-time performance. Local
done image patches are used to form the
structure. It is a computationally
efficient edge detector
Maladakara and Improved canny edge detection The canny is used at the block
Vanamala [18] algorithm level and the total image is
differentiated into several blocks
and thus reduces the
computational time and helps in
more appropriate detection of
edges
Yu et al. [19] Deep semantic learning based Category wise edge detection is
edge detection method performed
Liu et al. [20] Richer convolution features This RFC method produces high
(RCF) based edge detection quality edges and also exploits
method multilevel and multi scale
information of objects for
prediction
Agrawal and Rosepreet Three different techniques are To obtain image information or
[21] utilized—(1) Sobel type (2) pixels as well as detection of edges
Fuzzy Logic System Type 1 and regarding dental radiographs or
(3) Interval Type 2 Fuzzy logic X-ray of patients regarding the
System (T2FLS) healthy and effective teeth, those
three techniques are used
Halder et al. [22] Richardson’s extrapolation It is used for the identification of
technique is used to measure the respective edges of a binary image
strength of edges
Carmelo et al. [23] BPD algorithm, RBS algorithm This algorithm does not depend
and spatial FFT-based filtering on any predefined threshold
values and reduces the chance of
selecting any polynomial function
for curve fitting
Van Zwanenberg et al. Modulation Transfer Function This scheme can derive the system
[24] (MTF) and ISO 12233 algorithm performance in real-time
condition instead of any
laboratory condition
Shi et al. [25] Polarimetric constant false alarm This scheme is very effective to
rate edge detector and weighted suppress the noises and enhance
gradient-based (WG) detector the edge information
Bausys et al. [26] Adaptive selection based edge It is useful to get visual
detection algorithms characteristics like—size of the
individual regions, contrast,
roughness, density, and the
number of different regions.
References
1. Canny J (1986) A computational approach to edge detection. IEEE Trans Pattern Anal Mach
Intell 679–698
2. Sriramakrishnan C, Shanmugam A (2012) A fuzzy neural approach for medical image
retrieval. J Comput Sci 8(11):1809–1813. https://doi.org/10.3844/jcssp.2012.1809.1813
3. Dhir R (2012) Novel edge detection using adaptive neuro-fuzzy inference system. Int J Comput
Appl 2(5)
4. Rani P, Tanwar P (2013) A novel hybrid approach for edge detection. Int J Comput Sci Eng
Surv 4(2):27
5. Petrova J, Hostalkova E (2011) Edge detection in medical image using the Wavelet transform.
Report of Research, Department of Computing and Control Engineering, Czech Public
6. Bar L et al (2014) Mumford and Shah Model and its applications to image segmentation and
image restoration. In: Handbook of mathematical methods in imaging, pp 1–52
7. Getreuer P (2012) Chan-vese segmentation. Image Process Line 2:214–224
8. Chan TF, Vese LA Active contours without edges. IEEE Trans Image Process 10(2):266–277.
https://doi.org/10.1109/83.902291
9. Alvarez L et al (2012) A real time morphological snakes algorithm. Image Process Line 2:1–7
10. Rashid SS et al (2014) VHDL based Canny edge detection algorithm. Int J Curr Eng Technol
4(2)
11. Priyadarshini S, Gadadhar S (2010) A new edge detection method based on additions and
divisions. Int J Comput Appl 9(10):10–13
12. Biswas R, Sil J (2012) An improved canny edge detection algorithm based on type-2 fuzzy
sets. Procedia Technol 4:820–824
13. Rashmi MK et al (2013) Performance analysis of adaptive canny edge detector using bilateral
filter. Int J Innov Res Comput Commun Eng 1(4):931
14. Shamlee VJ (2014) A split Canny edge detection: algorithm and its FPGA implementation. Int
J Sci Res (IJSR) 3:1198–1205
15. Liu X et al (2014) Canny edge detection based on iterative algorithm. Int J Secur Appl 8(5):41–
50
16. Gao C et al (2011) New edge detection algorithms using alpha weighted quadratic filter. In:
IEEE international conference on systems, man, and cybernetics. IEEE
17. Dollár P, Zitnick CL (2014) Fast edge detection using structured forests. IEEE Trans Pattern
Anal Mach Intell 37(8):1558–1570
18. Maladakara DK, Vanamala HR (2015) Design of improved Canny edge detection algorithm
19. Yu Z, Feng C, Liu MY, Ramalingam S (2017) Casenet: deep category-aware semantic edge
detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition,
pp 5964–5973
20. Liu X et al (2017) Richer convolutional features for edge detection. In: Proceedings of the
IEEE conference on computer vision and pattern recognition, pp 3000–3009
21. Agrawal A, Rosepreet KB (2018) Edge detection techniques in dental radiographs (Sobel,
T1FLS & IT2FLS). In: International conference on communication, networks and computing.
Springer, Singapore
22. Halder A, Bhattacharya P, Kundu A (2019) Edge detection method using Richardson’s
extrapolation formula. In: Soft computing in data analytics. Springer, Singapore, pp 727–733
23. Carmelo M, Pierce SG, Summan R (2019) Novel algorithms for 3D surface point cloud
boundary detection and edge reconstruction. J Comput Des Eng 6(1):81–91
24. Van Zwanenberg O, Triantaphillidou S, Jenkin R, Psarrou A (2019) Edge detection techniques
for quantifying spatial imaging system performance and image quality. In: Proceedings of the
IEEE conference on computer vision and pattern recognition workshops
898 C. Ghosh et al.
25. Shi J, Jin H, Xiao Z (2020) A novel hybrid edge detection method for polari metric SAR
images. IEEE Access 8:8974–8991
26. Bausys R, Kazakeviciute-Januskeviciene G, Cavallaro F, Usovaite A (2020) Algorithm selec-
tion for edge detection in satellite images by neutrosophic WASPAS method. Sustainability
1(2):548
Design and Analysis of an Improvised
Fully Differential Amplifier
Swagata Devi, Koushik Guha, Naushad Manzoor Laskar, Sourav Nath,

and Krishna Lal Baishnab
Abstract This work proposes the design of an amplifier intended for low power
applications, operating at an ultra-low voltage and exhibiting a low input-referred
noise. The suggested structure implements a flip voltage follower (FVF) implemented
in its biasing circuit and thereby demonstrates improvised design parameters as
compared to the conventional single-stage fully differential amplifier (FDA). The
design is an attempt to enhance the transconductance, slew rate and gain-bandwidth
product (GBW). All the devices are operating in a weak inversion region such that the
entire design is appropriate for low power applications. The circuit simulations have
been performed in 180 nm CMOS SCL Cadence environment and are working at
0.6 V supply voltage. The simulation results report an increased gain by 7.81 dB and
an enhanced phase margin by 3.85°. The√ design consumes a power of 1.07 µW with
a total input-referred noise of 6.56 µV/ Hz @1 Hz. The gain-bandwidth product
and slew rate are measured to be 485.28 kHz and 26.85 V/ms.
Keywords Flip voltage follower (FVF) · Weak inversion region · Gain-bandwidth

(GBW) · Input-referred noise · Slew rate
1 Introduction
In the modern era; there is a demand for compact and portable gadgets. These criteria
arise the need for implantable battery-operated devices, which should have low power
consumption and low noise. Low power ensures longer battery lifetime, while the
requirement for low noise is to minimize the disturbances caused by skin-electrode
contact as well as external noises due to EEG recording electrodes attached to the
human body. Furthermore, low area is also a major concern in device implanta-
tion, to avoid damage to tissues during surgery [1]. The EEG signals have very low
amplitude, in order of few microvolts to millivolts, with a frequency range of a few
S. Devi (B) · K. Guha · N. M. Laskar · S. Nath · K. L. Baishnab

Silchar, Silchar, Assam, India
900 S. Devi et al.
Fig. 1 Block illustration of a neurostimulator
hundred hertz. So, the primary idea of this work is to design a power noise effi-
cient amplifier, which can be used as a preamplifier in neurostimulators of epilepsy
seizure detection systems [2]. Epilepsy is a neurological disorder in which a person
experiences periods of abnormal brain activity, and sometimes loss of awareness.
The Center for Disease Control and Prevention (CDC) has described epilepsy as
a “common neurological condition” characterized by recurrent seizures [3] and it
has affected 50 million people worldwide as reported by World Health Organization
in 2016. The pictorial illustration in Fig. 1 is the analog block representation of a
neurostimulator. It comprises of a preamplifier and a band-pass filter. The preampli-
fier continuously records the EEG signals, by maintaining a low input-referred noise.
The frequency range from the amplifier stage can be further narrowed down to the
required bandwidth with the help of a band-pass filter if necessary.
Researchers worldwide have proposed various preamplifiers and, as reported by
Qian et al. [4], a two-stage OTA low noise amplifier has been put forward which incor-
porates the multiple stages, thereby increasing of die area. Another work presented
by Niranjala et al. [5] provides a low input-referred noise, but the power consumed
is extremely high; which therefore acts as a constraint as it impacts the battery-life
of the design. Similar work by Bano et al. [6] reports very high input-referred noise
but a low power consumption. Wattanapanitch et al. [6] have mentioned in the article
that the input-referred noise of the amplifier should be lower than the background
noise of neural recording site, i.e., (5–10) µV. The power of an amplifier is propor-
tional to v12 , where vn is the input-referred noise of the amplifier, which establishes
n
that low noise is achievable at the expense of power. Therefore, power becomes a
controlling criterion for a complete neural amplifier system. Thus, the power noise
trade-off is also a major challenge for designers considering the demand of wearable
and implantable devices [7].
The circumvention of multiple stages can resuscitate area but the output swing
will be degraded. Therefore, this work suggests a flip voltage follower in the biasing
circuit of a fully differential folded cascode amplifier (FDA), to enhance its transcon-
ductance, gain-bandwidth product and slew rate. The paper follows up with a brief
description of FDA with CMFB circuit in Sect. 2, while Sect. 3 enumerates the
Design and Analysis of an Improvised Fully Differential Amplifier 901
proposed structure. Section 4 is an articulation of the results and observations,

followed by Sect. 5 which concludes this paper.
2 Fully Differential Folded Cascode Amplifier
Fully differential amplifiers have advantages in contrast to single-ended amplifiers.

They are less susceptible to common-mode noise, provide larger output voltage
swing and do not have even order nonlinearities. The FDA with a CMFB ensures
proper functioning of the circuit and stabilizes common-mode output voltages at all
operating range of frequencies. The output V OC from the FDA block is fed into the
+ −
CMFB block. V OC is the average of VOC and VOC . The difference of V OC and desired
V CM is amplified by a factor of ACM and then fed back to the op-amp, which is shown
through pictorial depiction in Fig. 2 [8]. The CM block senses the common-mode
voltage, which is expressed in Eqs. (1) and (2) as:
V01 +V02
VOC = (1)
2
Vcms =acms (VOC +VCM )+VCS bias (2)
Fig. 2 Block representation of FDA with CMFB circuit

902 S. Devi et al.
where, V CS bias is the additional dc bias voltage. The V cms output of CM sense block
is required to drive the V cmc input of op-amp. Thus, by varying V cms and V OC can
be changed but V od is not affected if the circuit is perfectly balanced. Hence, both
V cmc and V cms have equal values for voltages.
3 Proposed Topology
In this work, flip voltage follower (FVF) is introduced in the structure, as illustrated
in Fig. 3, while the schematic of its CMFB circuit is shown in Fig. 4. The circuit
operates in the weak inversion region; hence the intrinsic gain becomes limited but
the threshold voltage does not change. The low power supply voltage will restrict the
voltage swing of the amplifier. A solution to overcome the issues is to use multistage
op-amp, but it will consequence in a larger area which is not desirable. An alternative
is to use a flip voltage follower which can enhance the driving capability of the
structure, as shown in this work. The flip voltage follower circuit is a combination
of two alike transistors, which is a modification of a conventional source follower.
1
It provides a very low output impedance at the output node: gmA r0A gmB
and hence is
capable of sinking a large amount of current, and is unaffected by noise [9]. These
Fig. 3 Schematic of the proposed amplifier circuit

Fig. 4 Schematic of the CMFB
circuits are capable of functioning at a very low voltage and are therefore efficient
for low power applications. The use of the FVF in the differential amplifier design
results in increased transconductance and gain-bandwidth, as well as an enlarged
voltage swing. The load capacitor will discharge by (2k−1)
CL
I for positive output and
I
charge by CL for negative output. Thus, the total slew rate is given by Eq. (3):
6Id
SR = (3)
CL
Again, as depicted in Fig. 3 the gate of MB is connected to the input transistor

M2, therefore this current will change with respect to the input voltage, and therefore
will remain self-biased.
Id =ID +i d (4)
The increased current which is proportional to the input voltage correspondingly

result in a slight rise in consumed power. Now, when an ac input is being applied to
source terminals of input transistors, it will result in doubling of the transconductance
of its input stage, when compared to its conventional design. Thus, the ac current of
its input stage is shown in Eq. (4).
904 S. Devi et al.
gmOAroOA − 1
Id = 1+ (5)
gmOAroOA + 1
Equation (5) states that the ac current is twice as compared to its conventional
structure without FVF. Thus, its slew rate improves. The FVF circuit also introduces
an additional pole but since the devices have very low impedance hence this pole is
forced to high frequency, without affecting the frequency response [10].
The entire design is realized specifically for low power applications. Therefore,
all the devices are operating in the weak inversion region, where V GS is somewhat
less than V th . This signifies that V GS – V th generates a smaller value than zero, but
it allows a detectable drain current to flow through, which results in a degradation
in frequency response and an increase in flicker noise. The degradation in frequency
is due to very low currents, while flicker noise increases as it is the dominant noise
at low frequencies [11]. Flicker noise reduction can be achieved by increasing the
length of MOSFET devices, but at the expense of low current. This increase can
be compensated either by increasing transistor width or gate-source voltage. Again,
with an increase in transistor width, there is degradation in phase margin, which
means more power will be consumed due to increased parasitic capacitances, also
silicon area will be increased. On the other hand, an increase in gate-source voltage
will increase input-referred flicker noise, but the impact of increase in transistor
length is much more, hence noise can be effectively reduced without the increased
cost of die area [12]. The expression of drain current for devices working in weak
inversion region is shown in Eq. (6), where, V GS is the gate-source voltage, V DS is
the drain-source voltage, V th is the threshold voltage with a value of 26 mV at room
temperature and I S is the current with constant dimensions at V GS = V th , and the
width and length of a transistor device is denoted by W and L.
V −V
W GS th −VDS
ID =IS e nVT 1 − e VT (6)
L
Thus, the overall transconductance Gm of the amplifier can be calculated to be as

gm1, which is shown in Eq. (7) as:
δ ID ID
G m =gm1 = = (7)
δVGS nVth
where gm1 is the transconductance for the input transistor: M1. The dominant pole
ωp1 for the amplifier is given by the high output resistance and high capacitance that
occurs at the output node, while the other pole ωp2 have a frequency much greater
than dominant pole frequency shown by Eq. (8):
gm5
ωp2 = − (8)
Cgs5
Again, at low frequencies, flicker noise is the dominant noise over thermal noise.
The additional silicon and silicon dioxide energy states contribute to an increase
in flicker noise making it dominant. The overall input-referred noise is denoted by
Eq. (8) where the first term is the total thermal noise and the second term is the entire
flicker noise.
2
K F gm
i o2 = 4K B T γ gm + f (9)
COX W L f
The overall flicker noise can be determined by expanding the second term in
Eq. (9) and then replacing terms gma3 and gm9 with their respective expressions of
transconductances as depicted in Eq. (7). Equation (10) is the illustration of total
flicker noise in the amplifier:

2
K FP 1 K Fn gm3 Is3 2 W3 2 VGS3 −Vth
Vif, proposed = 2 f
2
+ e n 3 VT
COX (W L)a1 K Fp gm1 2 n 3 VT L 33

2
gm9 Is9 2 W9 2 VGS9 −Vth
+ 2 e n 9 VT
f (10)
gm1 n 9 VT L 39
Table 1 is a comparative summary of all the design parameters between a fully

differential amplifier circuit with and without a flip voltage follower, and with a
few recent works. The simulation has been performed in the Cadence 180 nm SCL
technology @0.6 V supply voltage. The measured gain has improved by 7.81 dB is
due to the doubling of transconductance and the phase margin by 3.85˚, as compared
to the conventional design. There is an increase in power consumption as a result
Table 1 Comparison of design parameters

Parameters Conventional Proposed [12] [5]
Supply voltage (V) 0.6 0.6 1 0.8
Overall gain (dB) 40.16 47.97 82 44.6
Phase margin (°) 85.67 89.52 84 –
GBW (kHz) 243.22 485.28 768 –
Power consumption (µW) 0.593 1.07 0.834 18.74
√
Input-referred noise (µV/ Hz) 13.45 6.56 80 1.24
Slew rate (V/ms) 15.49 26.85 – –
Technology (nm) 180 180 150 180
906 S. Devi et al.
of the increased current due to the implementation of the FVF. However as there
exists a trade-off between noise and power, so the increased power reflects in the
decrease in noise. As for the intended application the power does not exceed the
requisite criterion, so is acceptable. The frequency response of both conventional
and proposed design has been represented in Fig. 5.
The increase in transconductance also impacts the flicker noise by bringing it down
by a significant value, as depicted in Eq. (10). Figure 6 shows the variation in input-
referred flicker noise with frequency, and exhibits decreased value as compared to
conventional design. The step response for V out is illustrated in Fig. 7. The change in
its biasing circuit puts an impact upon average slew rate and gain-bandwidth product
Fig. 5 Frequency responses

of proposed and
conventional structure
Fig. 6 Spectral density of

the input-referred noise for
amplifiers
Fig. 7 Step response for

V out
Table 2 Comparison of
Devices Sizing (W /L) Devices Sizing (W /L)
design parameters
MA MB 45/1 M9 M10 10/1
M1 M2 60/1 M21 M24 5/1
M3 M4 1/2 M22 M23 5/0.18
M5 M6 15/1 M25 M26 ½
M7 M8 30/1 M27 M28 1.2/1
due to the increased transconductance, that can be observed. The device sizing values
have been listed in Table 2 accordingly.
5 Conclusion
The proposed preamplifier is designed in weak inversion region √ to exhibit a low

power of 1.07 µW with an input-referred voltage of 6.56 µV/ Hz @1 Hz, an open-
loop gain of 47.97 dB and a phase margin of 89.52°. The overall improvement in
gain, slew rate, and the gain-bandwidth product is by 7.81 dB, 242.06 kHz and
11.35 V/ms due to the FVF implementation. The use of FVF is to avoid the use of a
multistage amplifier and enhance both increase voltage swing and gain. The designed
prototype has been simulated in 180 nm SCL CMOS technology: higher technology
can be incorporated based on availability. The design is an open-loop topology, while
a closed-loop topology can be used for further analysis and comparison. Hence,
the results suggest that the design is an improvisation over a conventional fully
differential amplifier designed under same conditions.
908 S. Devi et al.
References
1. Sarpeshkar R, Wattanapanitch W, Arfin SK, Rapoport BI, Mandal S, Baker MW, Andersen
RA (2018) Low-power circuits for brain–machine interfaces. IEEE Trans Biomed Circ Syst
2(2):173–183
2. Qu H (1994) Self-adapting algorithms for seizure detection during EEG monitoring. Depart-
ment of Electrical Engineering, McGill University, Canada
3. Amudhan S, Gururaj G, Satishchandra P (2015) Epilepsy in India I: epidemiology and public
health. Ann Ind Acad Neurol 18(5):263–277
4. Qian, C, Parramon. J, Sanchez-Sinencio E (2011) A micropower low-noise neural recording
front-end circuits for epileptic seizure detection. IEEE J Solid St Circ 46(6):1392–1405
5. Niranjala M, Mohana D (2018) Epileptic seizure prediction using fully integrated neural signal
acquisition amplifier. Int J Res Dev Eng Sci Mng 1(12):826–830
6. Bano S, Narejo GB (2017) Low power fully differential folded cascode OTA with CMFB
circuit. J Multi Eng Sci Technol 4(7)
7. Wattanapanitch W, Fee M, Sarpeshkar R (2018) An energy-efficient micropower neural
recording amplifier. IEEE Trans Biomed Circ Syst 1(2):136–147
8. Gray PR, Hurst PJ, Lewis SH, Meyer RG (2005) Analysis and design of analog integrated
circuits, 4th edn. Wiley
9. Sarkar A, Panda SS (2017) Design of a power efficient, high slew rate and gain boosted
improved recycling folded cascode amplifier with adaptive biasing technique. Microsyt Technol
23(9):4255–4262
10. Vij S, Gupta A, Mittal A (2014) A highly adaptive operational amplifier with recycling folded
cascode topology. In: International conference on computer science and application engineering
pp 135–145
11. Akbari M, Biabanifard S, Hashemipour O (2014) Design of ultra-low-power CMOS amplifiers
based on flicker noise reduction. In: 2014 22nd Iranian conference on electrical engineering
12. Rajput SS, Singh A, Chandel AK, Chandel R (2016) Design of low-power high-gain operational
amplifier bio-medical applications. In: IEEE computer society annual symposium on VLSI
Hyperspectral Band Selection Based
on Variants of Rough Set Theory
Barnali Barman and Swarnajyoti Patra
Abstract To deal with the arduous task of hyperspectral image classification band
selection is a prominent approach broadly used in the literature. Rough set (RS)
theory is a paradigm suitable for handling uncertain, incomplete, and vague data. By
incorporating various concepts, classical RS has been extended for band selection
capable of handling those problems that classical RS cannot deal with. Four forward
greedy hyperspectral band selection algorithms are empirically studied in this paper
achieved using the rough set, variable precision rough set, tolerance rough set, and
neighborhood rough set. The usefulness of these state-of-the-art techniques is gauged
in terms of average overall classification accuracy, average kappa accuracy, and
standard deviation acquired by using support vector machine classifier on two real
hyperspectral data sets.
Keywords Hyperspectral image · Rough set theory · Variable precision rough set ·
Tolerance rough set · Neighborhood rough set · Support vector machines
1 Introduction
Presence of a large number of bands in the hyperspectral image (HSI) has driven
the attention of the researchers to the area of dimensionality reduction. HSI with
hundreds of bands provides the better discriminating capability of the land cover
objects compared to the multispectral images. Due to the availability of limited label
samples, ‘curse of dimensionality’ is a significant problem to be dealt with. It also
suffers from high computational cost and storage issues. These problems are resolved
by reducing the dimensionality of HSI. Out of two traditional dimensionality reduc-
tion techniques, feature or band selection is found to be more advantageous for its
semantics preserving nature over feature extraction. Band selection that selects a
B. Barman (B) · S. Patra

Department of Computer Science & Engineering, Tezpur University, Tezpur, Assam 784028, India
S. Patra
910 B. Barman and S. Patra
subset from original band set is divided into supervised [5] and unsupervised [12]
categories. Supervised band selection techniques generally produce better accura-
cies compared to unsupervised band selection techniques as they employ the label
information. Several supervised wrapper and filter methods for HSI band selection
are presented in the literature [5, 11].
Rough set (RS) theory proposed by Pawlak [9] is a tool used to compute data
dependencies to select a feature subset having the same predictive capability as that
of the entire data set without any additional information [8]. Application of RS is
found to be rare in the field of remote sensing [3, 8]. RS uses the crisp concept of
equivalence classes. Before using continuous band values, it requires discretization
due to its ability to dealt with only discrete values. Besides, it cannot find the extent of
similarity between the attributes. Therefore, based on relaxation on the subset oper-
ator, similarity, tolerance relation, or capable of handling continuous values, variants
of the classical rough set (RS) have been proposed. Variable precision rough set
(VPRS) provides modification on the subset operator of RS with additional allowed
error information while selecting the feature subset [3, 14]. RS and VPRS need data
to be discretized before performing feature selection which can cause loss of valuable
information. This problem is mitigated with the help of tolerance rough set (TRS)
[13]. Instead of indiscernibility relation of RS, TRS uses a similarity relation and a
manually defined similarity threshold value to handle continuous data. Discretization
problem can also be mitigated using another variant of RS, i.e., neighborhood rough
set (NRS) with neighborhood granulation concept. It needs a neighborhood threshold
and distance metric function as the extra information during feature selection [1, 2].
This empirical study focuses on such variants of the rough set for HSI band selection.
2 Basic Notions of Variants of Rough Set Theory
2.1 Rough Set (RS)
A data set can be expressed in a tabular form, where each row identifies an object and
each column an attribute is known as information system denoted by (U, A, V, f ). The
tuples represent the universe, attribute set, value domain, and information function,
respectively. RS uses the concept of equivalence or indiscernibility relation [9]. A
pair of objects belongs to the indiscernibility relation signifies that the knowledge of
the used attribute set cannot make the two objects discernible.
Given (U, A, V, f ), B ⊆ A and X w ⊆ U . Let IND(B) define the equivalence
classes generated using indiscernibility relation denoted as [x i ]B (x i ∈ U). On the
basis of the knowledge in B, all the objects can be partitioned into two sets. The
first set, i.e., B-lower approximation contains only the objects that surely belong to
X w also known as positive region (denoted by (POSB (X w )), and the other set, i.e.,
B-upper approximation contains all the objects that possibly belong to X w [9]
Hyperspectral Band Selection Based on Variants … 911
B X w = ∪{ [xi ] B |[xi ] B ⊆ X w } (1)
B X w = ∪{[xi ] B |[xi ] B ∩ X w = φ} (2)
The inclusion operator for computing both the approximations using RS some-
times becomes too restrictive as in real-life application decision making the objects
of classes often overlap. Therefore, a variant of RS known as variable precision rough
set is proposed.
2.2 Variable Precision Rough Set (VPRS)
Variable precision rough set proposed by Ziarko [15] provides relaxation on the
subset operator. It allows partial classification rather than complete classification
required by RS. Given X, Y ⊆ U , where X, Y = φ. The relative classification error
of X in Y is given by [15]

|X ∩Y |
1− |X |
, |X | > 0
e(X,Y) = (3)
0, |X | = 0
Given an information system (U, A, V, f ), B ⊆ A and X w ⊆ U . Considering β

to be the allowed level of error, the lower and upper approximation are defined as
[15]
B β X w = ∪{[xi ] B ∈ U |e([xi ] B , X w ) ≤ β} (4)
β
B X w = ∪{[xi ] B ∈ U |e([xi ] B , X w ) < 1 − β} (5)
RS and VPRS cannot deal with continuous band values directly without applying
a suitable discretization technique which can cause valuable information loss. This
problem is mitigated by using the following two techniques.
2.3 Tolerance Rough Set (TRS)
Tolerance rough set (TRS), another variant of RS, can be applied successfully on
real-valued data. Instead of using indiscernibility relation, TRS employs the concept
of similarity relation. According to a specified threshold value, tolerance classes
can be achieved. TRS requires a suitable similarity relation defined for the features.
Given (U, A, V, f ), a ∈ A, xi , x j ∈ U . A standard similarity measure is given by
[10]
|a(xi ) − a(x j )|
SIMa (xi , x j ) = 1 − (6)
|max(a) − min(a)|
For B ⊆ A feature set that contains more than one feature, the combined similarity
measure is given by any of the following equation

(xi , x j ) ∈ SIMτB ⇔ SIMa (xi , x j ) ≥ τ (7)
a∈B

SIMa (xi , x j )
(xi , x j ) ∈ SIMτB ⇔ a∈B
≥τ (8)
|B|
where τ is known as a global similarity threshold representing the required similarity

level for insertion within the tolerance classes. An object xi uses a similarity relation
for generating the tolerance classes is defined as SIMτB (xi ) = { x j ∈ U |(xi , x j ) ∈
SIMτB } .
Given (U, A, V, f ), B ⊆ A, X w ⊆ U and τ . The lower and upper approximation
are defined as
B τ X w = {xi |SIMτB (xi ) ⊆ X w } (9)
τ
B X w = {xi |SIMτB (xi ) ∩ X w = φ} (10)
2.4 Neighborhood Rough Set (NRS)
Neighborhood rough set (NRS), an another variant of RS, is also capable of handling
real-valued attributes effectively. It mitigates the loss of valuable information that
occurs during the discretization of continuous band values using RS. Given an arbi-
trary sample x i ∈ U and feature set B ⊆ A. Neighborhood granule generated by
sample xi for the B feature space is denoted by δ B (xi ) = {x j |x j ∈ U, B (xi , x j ) ≤ δ}
[7], where δ is a threshold value, and is a distance metric function. The above equa-
tion shows that δ B (xi ) the neighborhood information granule centered with xi is influ-
enced by δ and , determining the shape and size of neighborhood granules, respec-
tively. With the increment in δ value, the number of samples in the neighborhood of
xi will also increase. Given an information system (U, A, V, f ), B ⊆ A, X w ⊆ U
and δ B (xi ). The lower approximation and upper approximation are defined as [7]
B δ X w = {xi |δ B (xi ) ⊆ X w , xi ∈ U } (11)
δ
B X w = {xi |δ B (xi ) ∩ X w = φ, xi ∈ U } (12)
Hyperspectral band selection using RS, VPRS, TRS, and NRS

Let the set of n patterns associated to the n pixels of the HSI with label be U =
{x1 , x2 , . . . , xn } and the set of m bands be C = {b1 , b2 , . . . , bm } that indicate each
pixel of the image. An HSI data set can be represented by H = {h i j |i = 1, . . . , n; j =
1, . . . , m} matrix, where h i j is the value of pattern xi ∈ U for band b j ∈ C. A
hyperspectral data set can be expressed by a decision system (U, C ∪ D, V, f ),
where C, D, V and f indicate condition attribute, decision attribute, value domain,
and information function, respectively.
Given a decision system (U, C ∪ D, V, f ) and B ⊆ C. Let X w1 , X w2 , . . . , X wc
be the object subsets in U with decision classes w1 , w2 , . . . , wc , respectively. The
dependency degree of D w.r.t. B is defined as [9]
|POS B (D)|
γ B (D) = (13)
|U |
where POSB (D) denotes the positive region computed using Eqs. (1), (4), (9) and
(11) for RS, VPRS, TRS, and NRS, respectively. |.| represents the number of objects
present in a set. Informative attributes can be selected from the conditional attribute
set using the above dependency measure. If the bands are chosen based on depen-
dency criterion only, it may contain redundant information. Therefore, a significance
criterion is defined to remove the irrelevant attributes as follows [7]
σ B (D, a) = γ B∪a (D) − γ B (D) (14)
An attribute with the highest significance value is considered to be the most useful
attribute. If the significance of an attribute is zero, then it is uninformative.
Algorithm (1) introduces a greedy forward search technique by exploiting the
properties of RS, VPRS, TRS, and NRS for selecting an optimal hyperspectral band
subset. Initially, an empty subset is considered, and in each iteration bands with the
highest significance value are added. This process continues until a criterion to abort
is encountered.
Algorithm 1: Hyperspectral band selection using RS, VPRS, TRS, and NRS
1. Initialize P = {φ}, C = {b1 , b2 , . . . , bm } and S = C.
2. Repeat
3. For each and every band bi ∈ S, calculate significance σ P (D, bi ) using (14).
4. Select informative band b j = arg max {σ P (D, bi )}.

bi ∈S
5. Upgrade P ← {P ∪ b j } and S ← {S − b j }.
6. Until a criterion to abort is encountered.
Table 1 Basic details of Botswana and KSC data sets

Data set Sensor type Spatial resolution Image size Total spectral bands Total classes
(m)
Botswana Hyperion 30 1476 × 256 145 14
KSC AVIRIS 18 512 × 614 176 13
3 Details of Data Sets
Two hyperspectral data sets used in our experiment to compute the effectiveness of
the variants of rough set theory based band selection techniques described in Sect. 2
are Botswana [6] and KSC [6]. The details of these two data sets are listed in Table 1.
Effectiveness of the three different variants of rough set theory is investigated on

two aforementioned hyperspectral image data sets. To scale the spectral band values
between 0 and 1 initially, normalization is performed. The variants of RS, i.e., VPRS,
TRS, and NRS, are dependent on manual setting of additional parameters β, τ and δ,
respectively. β value can be chosen from range [0, 0.5), whereas τ and δ from [0, 1].
β = 0,τ = 0 and δ = 0 forces VPRS, TRS, and NRS to behave completely like RS,
respectively. To train the SVM classifier, we have selected 50% labeled samples as
the training set, and the rest are considered as the test set. The classifier is executed
ten times for different training sets. Thus, it reduces the irregularities present in the
result.
Three measures, i.e., average overall classification accuracy (OA), its standard
deviation (STD), and average kappa accuracy (KA) are used for evaluating the perfor-
mance of all the variants of RS for band selection. The percentage of the number of
test samples correctly classified by the classifier is the overall classification accuracy
(OA). Standard deviation is the variation of OA produced on ten runs. Using Kappa
accuracy (KA) also, we can compute the accurate classification of samples in the test
set.
In the experiment, we have used (7) to compute the combined similarity between
bands and Euclidean distance to calculate the neighborhood granule of the patterns
for TRS and NRS, respectively. In this paper, we have displayed results for bands
5–35 in the interval of 5 bands. After selecting the required number of bands from
the band selection techniques, we tried to find out how much accurate they are.
The standard one-against-all (OAA) architecture of support vector machine (SVM)
classifier is used for this purpose. It assumes the radial basis function (RBF) as the
kernel function. The variance parameter and the penalization factor of SVM classifier
for various data sets are approximated by applying a grid search via a ten-fold cross-
validation technique. The LIBSVM library is used in implementation [4]. This paper
has adopted the MATLAB environment for presenting all the results.
4.2 Results
The techniques are analyzed with the help of Botswana, and KSC data set is described
in Sect. 3. Tables 2 and 3 show the results in terms of OA, KA and STD of classical
rough set (RS), variable precision rough set (VPRS), tolerance rough set (TRS), and
neighborhood rough set (NRS) for Botswana and KSC data sets, respectively. From
the tables, it is visible that with the increment in band numbers, the classification
accuracy also increases. Result of the VPRS technique is dependent on a manual
parameter (β) that allows a certain level of error. This β value is manually consid-
ered to be 0.1 from the range [0, 0.5). We have observed that for β = 0.1, VPRS
Table 2 Average overall classification accuracy (OA), average kappa accuracy (KA), and standard
deviation (STD) of RS, VPRS, TRS and NRS for Botswana data set (for manually considered
parameters)
Number RS VPRS (β = 0.1) TRS (τ = 0.7) NRS (δ = 0.1)
of OA KA STD OA KA STD OA KA STD OA KA STD
bands
5 86.02 0.849 0.49 86.51 0.854 0.562 79.91 0.783 0.305 85.7 0.845 0.371
10 89.93 0.891 0.316 90.53 0.897 0.235 86.37 0.852 0.434 90.04 0.892 0.237
15 90.72 0.899 0.285 90.97 0.902 0.263 93.13 0.926 0.308 94.25 0.938 0.402
20 91.44 0.907 0.349 91.61 0.909 0.351 94.3 0.938 0.274 95.04 0.946 0.361
25 92.3 0.917 0.297 92.49 0.919 0.369 94.79 0.944 0.218 95.92 0.956 0.398
30 93.87 0.934 0.264 94.24 0.938 0.283 95.51 0.951 0.316 96.17 0.959 0.217
35 95.32 0.949 0.203 95.87 0.955 0.306 96.23 0.959 0.351 96.76 0.965 0.252
Table 3 Average overall classification accuracy (OA), average kappa accuracy (KA) and standard
deviation (STD) of RS, VPRS, TRS, and NRS for KSC data set (for manually considered parameters)
Number RS VPRS (β = 0.1) TRS (τ = 0.7) NRS (δ = 0.1)
of bands OA KA STD OA KA STD OA KA STD OA KA STD
5 79.67 0.773 0.37 84.28 0.824 0.286 73.15 0.699 0.569 69.11 0.653 0.714
10 85.59 0.839 0.497 89.21 0.88 0.224 87.62 0.862 0.294 76.32 0.735 0.438
15 87.93 0.865 0.326 90.27 0.891 0.234 92.38 0.915 0.204 84.55 0.827 0.282
20 89.97 0.888 0.212 91.74 0.908 0.192 93.67 0.929 0.226 86.52 0.85 0.33
25 91.25 0.902 0.332 92.61 0.918 0.217 94.24 0.936 0.236 90.54 0.895 0.465
30 92.11 0.912 0.368 93.09 0.923 0.226 94.75 0.942 0.243 93.31 0.925 0.383
35 93.01 0.922 0.246 93.48 0.927 0.207 95.06 0.945 0.314 93.76 0.93 0.321
produces better result compared to RS for both the data sets. VPRS produces better
result compared to RS as it provides relaxation on the subset operator of RS. Here, the
similarity threshold (τ ) value is manually considered to be 0.7 from the range [0, 1].
In the case of TRS, τ = 0.7 produces better result compared to RS and VPRS while
having more than ten bands for both the data sets. It is due to the fact of using contin-
uous band values directly without applying any discretization method. It exploits the
concept of similarity between objects. For NRS, results are shown by considering δ
value manually to be 0.1. When the number of selected bands are more than ten, it
produces better result compared to all the other three techniques for Botswana data
set. Thus, the concept of neighborhood granule for band selection is more useful for
Botswana data set. But for KSC data set when the number of selected bands are less
than 30, it produces low accuracy compared to RS and VPRS. Afterward, accuracy
improves compared to them. NRS could not produce a better result compared to
TRS. Thus, TRS is more suitable compared to all the other techniques for KSC data
set.
5 Discussion and Conclusion
Rough set theory is capable of handling vague, incomplete, and uncertain data. To
handle the drawbacks of classical RS, several variants exist in the literature. The
effect of these variants is discussed in terms of hyperspectral image data sets. Since
VPRS provides relaxation on the subset operator of RS, it produces a better result
compared to RS. But VPRS is dependent on a predefined parameter β to be set
manually. RS and VPRS cannot deal with continuous band values directly without
applying a suitable discretization technique. TRS can handle continuous band values
directly using the concept of similarity. It provides a better result compared to RS
and VPRS. It also became dependent on the similarity measure and global similarity
threshold value. Another variant of RS, i.e., NRS, uses the concept of neighborhood
granule to select the most informative bands. It produces a better result compared
to all the remaining techniques presented in this paper except KSC data set. NRS is
also dependent on predefined parameter (δ and ). From this empirical study, we
can conclude that the variants of rough set theory capable of handling continuous
band values directly without discretization significantly outperforms.
Although the variants, i.e., VPRS, TRS, and NRS of RS are capable of handling
several problems that classical RS cannot, all of them requires additional information
during band selection. Automatic computation of the manual parameters for VPRS,
TRS, and NRS can be considered as the future scope of this paper.
References
1. Barman B, Patra S (2019) Empirical study of neighbourhood rough sets based band selection
techniques for classification of hyperspectral images. IET Image Process 13(8):1266–1279
2. Barman B, Patra S (2019) A novel technique to detect a suboptimal threshold of neighborhood
rough sets for hyperspectral band selection. Soft Comput 23:13709–13719
3. Barman B, Patra S (2020) Variable precision rough set based unsupervised band selection
technique for hyperspectral image classification. Knowl-Based Syst 193:105414
4. Chang CC, Lin, CJ (2012) Libsvm: a library for support vector machine. Software available at
http://www.csie.ntu.edu.tw/~cjlin/libsvm
5. Habermann M, Fremont V, Shiguemori EH (2019) Supervised band selection in hyperspectral
images using single-layer neural networks. Int J Remote Sens 40(10):3900–3926
6. Ham J, Chen Y, Crawford MM, Ghosh J (2005) Investigation of the random forest framework
for classification of hyperspectral data. IEEE Trans Geosci Remote Sens 43(3):492–501
7. Hu Q, Yu D, Liu J, Wu C (2008) Neighborhood rough set based heterogeneous feature subset
selection. Inf Sci 178(18):3577–3594
8. Patra S, Modi P, Bruzzone L (2015) Hyperspectral band selection based on rough set. IEEE
Trans Geosci Remote Sens 53(10):5495–5503
9. Pawlak Z (1991) Rough sets, theoretical aspects of reasoning about data. Kluwer, Dordrecht,
The Netherlands
10. Skowron A, Stepaniuk J (1996) Tolerance approximation spaces. Fundamenta Informaticae
27(2, 3):245–253
11. Taşkın G, Kaya H, Bruzzone L (2017) Feature selection based on high dimensional model
representation for hyperspectral images. IEEE Trans Image Process 26(6):2918–2928
12. Xie F, Li F, Lei C, Yang J, Zhang Y (2019) Unsupervised band selection based on artificial bee
colony algorithm for hyperspectral image classification. Appl Soft Comput 75:428–440
13. Yun O, Ma J (2006) Land cover classification based on tolerant rough set. Int J Remote Sens
27(14):3041–3047
14. Zhang C, Sargent I, Pan X, Gardiner A, Hare J, Atkinson PM (2018) VPRS-based regional
decision fusion of CNN and MRF classifications for very fine resolution remotely sensed
images. IEEE Trans Geosci Remote Sens 56(8):4507–4521
15. Ziarko W (1993) Variable precision rough set model. J Comput Syst Sci 46(1):39–59
Performance Analysis of Interference
Alignment-Based Precoding
Amlan Deep Borah, Md. Anwar Hussain, and Joyatri Bora
Abstract When it comes to the modern multiple-inputs and multiple-output system,

i.e., MIMO, the wireless communication system faces a complex challenge in the
form of interference mitigation to provide an efficient service. In order to mitigate
such challenges, the technology of interference alignment (IA) is an important factor
to limit the inter-cell interference of a cellular system. IA technique can increase
the system throughput with perfect channel state information in high signal-to-noise
(SNR) case. The interference alignment technique also requires feedback technology
at least within a cell. This paper discusses the topic of precoder design, user, and
antenna scheduling techniques based on IA in a cellular system of two- and three-
cell clusters. By obtaining the user’s antenna combining vector, the IA scheme aligns
the interfering signal from each transmitter of the neighboring cell. We have used
zero-forcing beamforming (ZFBM) and minimum mean square error (MMSE)-based
precoding to simulate our proposed scheme. On the conclusion of the simulation
result, it is tried to showcase the proposed scheme gives better sum-rate performance
than the traditional IA-based precoding scheme.
Keywords MIMO · Interference alignment · System throughput · Antenna

combining · Sum rate
1 Introduction
In wireless communication system, interference is a major hurdle as mobile users

with single and multiple antennas receive signals form single and multiple transmit-
ters with single and multiple antennas of a communication system. To increase the
A. D. Borah (B) · Md. A. Hussain · J. Bora

North Eastern Regional Institute of Science and Technology, Nirjuli, Arunachal Pradesh, India
Md. A. Hussain
J. Bora
920 A. D. Borah et al.
system performance, MIMO can provide diversity gain multiplexing or array gain
by using multiple antennas both at the receiver and transmitter. An increased number
of simultaneously serving users that results in the increase of multiuser diversity
gain can also increase the system throughput. This can be achieved through linear
precoding, a technique for obtaining better system performance by optimizing the
interference of a multiuser MIMO system.
Interference is the primary challenge of a communication system [1]. In recent
studies, IA is the new aspect to deal with the reduction of inter-cell interference in
multicell MIMO system [2]. Firstly, IA was introduced as a coding technique to
achieve multiplexing gain [3]. IA was also used as a way to minimize interference-
free space in an interference channel for the desired signal [4]. The basic idea of
the IA scheme is to compress all interference signals to a predefined dimension of
the transmitted signal by the proper design of transceiver to maximize the degree
of freedom of the system (DoF). Channel state information (CSI) is important for
ideal interference alignment. The precoding techniques use to mitigate the inter-user
interference and an IA scheme works on the reduction of the inter-cell interference
[5].
The alignment of the interference signal into a subspace dimension of a transmitted
signal can provide maximum DoF [6]. Article [7], [8] made a proper study on the
feasibility of the IA system and DoF for a downlink multiuser MIMO system. In
two-cell cluster, scenario BS can recover desired V symbols using V + 1 dimension
which a DoF of VV+1 per cell [9]. In [10], IA base zero-forcing (ZF) for multiuser
MIMO was proposed. Authors in [9] have discussed the implementation of an IA
scheme in two-cell interfering systems with the receive antennas number larger than
the transmitting antennas and reciprocal. [11] discussed a two-user (in each cell)
two-cell clusters. A joint design of the beamforming vector at the transmitter and
user antenna combining factor design to align the interference to a specific direction
can eliminate the inter-cell and intra-cell interferences [12]. Sum rate of a two- and
three-cell network has been discussed using IA-based precoding aspect for cell-edge
users with variable user’s position. Orthogonal property is used to select the best
users from a large group that increases diversity gain. Antenna scheduling technique
is also performed at both the user and the transmitter side that get a better efficiency
of the system.
2 System Model
In this work, we consider two different clusters of two and three cells for downlink
multiuser MIMO as shown in Figs. 1 and 2. Antenna and user scheduling are used
to select the best antennas and users from a larger group to get better system perfor-
mance. At the transmitter and user, we consider X and Y numbers of antennas in
this system, respectively. In a cell, R base station (BS) antennas are selected from
X numbers. A total of U numbers of users exists in a cell, where it can serve V
numbers of users (V ≤ U). Users are selected by using both random and orthogonal
Performance Analysis of Interference Alignment-Based Precoding 921
Fig. 1 Two cells network of

a downlink multiuser MIMO
Fig. 2 Three cells network of a downlink multiuser MIMO
user selection techniques. We consider the cell-edge users in the work, where users
receive equal power from its own BS and neighboring BS.
The mth user’s receive signal in the lth cell can be written as
yl,m = Rl,m sl + Tl,m s3−l + n l,m (1)
The term sl can be written as

V
sl = Fl dl = fl,V dl,V (2)
v=1
In Eq. (1) above, Rl,m is the direct channel between transmitter and user and Tl,m
is the cross-channel between the user and the interfering transmitter. sl represents
the transmitted signal and n l,m is the additive
white Gaussian
noise with variance
C N (0, 1). Whereas shown in (2), Fl = fl,1 . . . fl,V ∈ C X ×V is the precoding
with fl,m as the beamforming vector of user m in cell l, that satisfies fl,m = 1.
matrix
E xl2 = P is the power signal, here transmitter allocates equal power to each user.
T
dl = dl,1 . . . .dl,V is the data vector with dl,V data symbol.
Since we have considered the case V ≤ U , the cross-channel is not invertible for
this, which results that cross-channel cannot be matched to the reference direction.
The Euclidean distance can be minimized by multiplying the mth user of the lth
cell received signal with a vector ql,v . The proper design of the antenna combining
vector can provide this minimization by aligning the interference signal to a reference
direction. ql,v can be written as [9]
ql,V = argmin q H Tl,V − vref

H
s.t. q = 1 (3)
The antenna combining factor can be expressed as [12]

−1
H
Tl,m Tl,m Tl,m vref
ql,V = −1 (4)
H
Tl,m Tl,m Tl,m vref
Using the above, received combining signal can be written as
yl,m = ql,m
H
Rl,m xl + ql,m
H
Tl,m x3−l + ql,m
H
n l,m
By using Eq. (2), the above equation can be obtained as

V
yl,m = ql,V
H
Rl,V fl,k dl,k + ql,V
H
Tl,V
k=1

V
f 3−l,k d3−l,k + ql,V
H
n l,m (5)
k=1

Sum rate of V users is obtained as Ri = V log2 1 + SINRi, j . So, signal-to-
interference plus noise ratio (SINR) of (l, m)th user can be written as

P H
2
q R f
K l,V l,V l,V
SINRl,m =
K H 2 P
K H 2 (6)

k=1 ql,V Rl,V f i,k + K
+1
k=1 ql,V Tl,V f 3−i,k
k= j
H 2
where in the numerator KP ql,V Rl,V fl,V is the average signal power, in the denom-

V H 2
2
inator k=1 q Rl,V fl,V is intra-cell interference and P V q H Tl,V f 3−i,k
l,V K k=1 l,V
k=V
is the inter-cell interference. Using the receive antenna combining, (l, m)th user’s
detected signal can be written as
yl,m = ql,V
H
Rl,V sl + ql,V
H
Tl,V s3−l + ql,V
H
n l,m
eq eq
= rl,V xi + u l,V s3−l + n l,m
eq
In the equation above, rl,V is the effective direct channel and interfering channel
eq eq
u l,Vcan be decomposed into two parts having a parallel to vref (u vref ) and perpen-
eq eq H
dicular to vref (u ⊥vref ) [12]. Let the interfering channel can be written as β = u l,V
and also β = βvref + β⊥vref . Here, β⊥vref causes interference, which can be computed
as

H BfB H

vref vref
β⊥vref = Projvref ((B ) ) =
eq H
I− (7)
vref
2 (B f ) H
vref
Interfering channel average power is

2 P 2
PICI = E β⊥vref s3−l = β⊥v (8)
K ref
Transmitter constructs the channel matrix as given bellow [12]

⎡ eq ⎤
rl,1
⎢ . ⎥
⎢ ⎥
⎢ ⎥
⎢ . ⎥
⎢ ⎥
⎢ . ⎥
⎢ eq ⎥
⎣ rl,V ⎦
vref
H
Form the expression above, the precoding matrix can also be written as shown
below,
−1
Fl = βlH βl βlH τ
⎡√ ⎤
γ i,1 0 0
−1 ⎢ ⎥
= βiH βi βiH ⎣ . . . ⎦
√
0 0 γ i,k+1
⎡√ ⎤
γ l,1 0 0
⎢ ⎥
= wl,1 ..wl,V wl,V +1 ⎣ . . . ⎦
√
0 0 γ l,V +1
√ √ √
= γ l,1 wl,1 .. γ l,V wl,V γ l,V +1 wl,V +1

= fl,1.. fl,V f L ,V +1 (9)
In the equation above, γl,V = 1

wl,V
2 is the power control factor. F i , the precoding
matrix uses to take the first V column from the Fl [12]. In our scheme, here we
have considered (V + 1)th column as the interfering element for the calculation of
SINR. We used the F i value to obtain the ZFBM and MMSE precoding to analyze
the system performance for various user’s locations as mentioned below.
3 Result and Analysis
The sum-rate performance of the system has been evaluated in this section. The
system with two cells and three cells network is considered as shown in Figs. 1 and
2. We have considered that users are nearer to an edge of a cell, which is also closer to
the neighboring cell, considering that users are getting equal power from its own BS
and in the interfering BS. User and transmit antenna selection techniques are used to
obtain better system performance. Traditional IA-based scheme and the proposed IA-
based scheme with ZFBM and MMSE precoding having user and antenna scheduling
techniques with low rate-limited feedback have been simulated using MATLAB.
Each cell has a total of U users and V are selected for communication. We used both
random and orthogonal user selection techniques and observed the system sum-rate
variation. R numbers of the antenna are selected from the X number of BS antennas
in each cell, and each user has Y numbers of antennas. In the traditional IA scheme,
the first V columns of the matrix obtained from Eq. (9) are considered. But in our
scheme, (V + 1)th column is considered as the interfering element for sum-rate
calculation. This consideration makes changes in the weighting matrix for precoder
design in our proposed scheme due to which the sum rate is increased as shown in
Figs. 3 and 4. Also, because of the scheduling applied the sum rate is increased in
our scheme.
Figure 3 compares the system performance of a two-cell cluster for both the
ZFBM and MMSE precoding. It compares the sum-rate variation with orthogonal
and random user selection techniques for both the IA-based approach as mentioned
above. Here, X = 4 is selected form R = 80, where Y = 2. Taking similar consideration
Randomly selected user with proposed scheme

7 Orthogonally selected user with proposed scheme
Randomly selected user with traditional IA based precoding
6 Orthogonally selected user with traditional IA based precoding

Sumrates (Bit/Sec)
0
ZFBM Precoding MMSE Precoding
Various precoding aspect
Fig. 3 Sum-rate performance of ZFBM and MMSE precoding with randomly and orthogonally
selected user with traditional and proposed IA-based precoding scheme for two microcell
15
Randomly selected user with proposed scheme
Orthogonally selected user with proposed scheme
Randomly selected user with traditional IA based precoding
Sumrate (Bit/Sec) Orthaganally selected user with traditional IA based precoding
10
0
ZFBM Precoding MMSE Precoding
Various precoding aspects
Fig. 4 Sum-rate performance of ZFBM and MMSE precoding with randomly and orthogonally
selected user with traditional and proposed IA-based precoding scheme for three microcell
as before, Fig. 4 shows the system sum-rate performance for three-cell network as
shown in Fig. 2. Simulated values for both the traditional and proposed IA-based
ZFBM and MMSE precoding are shown in Table 1. Sum-rate variation with respect
to the random and orthogonal user selection technique is also shown in Table 1.
Table 1 Sum rate obtained for orthogonal and random user selection technique with MMSE and
ZFBM precoding for traditional and proposed IA-based scheme
Network Schemes ZFBM MMSE
types Random user Orthogonal Random user Orthogonal
selection user selection selection user selection
Two-cell Proposed 4.638 5.21 5.0685 5.2415
network IA-based
precoding
Traditional 0.5383 0.6405 0.8568 0.8722
IA-based
precoding
Three-cell Proposed 8.9629 10.3071 10.138 10.4289
network IA-based
precoding
Traditional 1.1349 1.4825 1.5837 1.7319
IA-based
precoding
4 Conclusion
A multiuser MIMO system with two cells and three cells network are used to analyze
in this work, where users are considered to be placed at the cell edge as shown
in Figs. 1 and 2. This paper studies the sum-rate variation using a proposed IA-
based scheme and traditional IA-based scheme with ZFBM and MMSE precoding. A
comparative study has been made on sum-rate variation using random and orthogonal
user selection techniques separately. From the bar diagram plotted in Figs. 3 and 4,
it has been proved that for a three cells cluster, sum rate is higher than the two cells
network. It has also been concluded that the proposed scheme gives better sum-rate
performance than the traditional IA-based ZFBM and MMSE precoding. From the
bar diagram, it can be also added that the orthogonal user selection technique gives
better outcomes than the random user selection technique and MMSE precoding
works better than the ZFBM.
References
1. Jungnickel V, Manolakis K, Zirwas W, Panzner B, Braun V et al (2014) The role of small cells,
coordinated multipoint and massive MIMO in 5G. IEEE Commun Mag 52(5):44–51
2. Zhou Y, Liu L, Du H, Tian L, Wang X et al (2014) An overview on intercell interference manage-
ment in mobile cellular networks: from 2G to 5G. In: 2014 IEEE international conference on
communication systems. IEEE, Macau, pp 217–221
3. Maddah-Ali MA, Motahari AS, Khandani AK (2008) Communications over MIMO X channel:
interference alignment, decomposition, and performance analysis. IEEE Trans Inform Theory
54(8):3457–3470
4. Cadambe VR, Jafar SA (2008) Interference alignment and the degrees of freedom for the K
user interference channel. IEEE Trans Inform Theory 54(8):3425–3441
5. Jindal N (2006) MIMO broadcast channels with finite-rate feedback. IEEE Trans Inf Theory
52(11):5045–5060
6. Fadlallah Y, Ferrand P, Cardoso LS, Gorce J-M (2016) Interference alignment for downlink
cellular networks: joint scheduling and processing. In: IEEE 17th international workshop on
signal processing advances in wireless communications (SPAWC), Jul 2016, Edinburgh, UK,
pp 1–5
7. Liu T, Yang C, On the degrees of freedom of asymmetric MIMO interference broadcast chan-
nels. In: 2014 IEEE international conference on communications. IEEE, Sydney, Australia, pp
1971–1976
8. Sridharan G, Yu W (2013) Degrees of freedom achieved using subspace alignment chains for
three-cell networks. In: 2013 Asilomar conference on signals, systems and computers. IEEE,
California, USA, pp 1869–1874
9. Bayesteh A, Mobasher A, Jia Y (2011) Downlink multi-user interference alignment in two-cell
scenario, In: 2011 12th Canadian workshop on information theory. IEEE, Kelowna, Canada,
pp 182–185
10. Suh C, Ho M, Tse DNC (2011) Downlink interference alignment. IEEE Trans Commun
59(9):2616–2626
11. Shin W, Lee N, Lim J-B, Shin C, Jang K (2011) On the design of interference alignment
scheme for two-cell MIMO interfering broadcast channels. IEEE Trans Wireless Commun
10(2):437–442
12. Zhu Y, Ou Y (2016) Interference alignment-based precoding and user selection with limited
feedback in two-cell downlink multi-user MIMO systems. J Eng Sci Technol Rev 9(2):115–120
Structural and Optoelectronic Properties
of Spin-Coated CH3 NH3 PbCl3 Thin Film
Using Non-halide Source of Lead
Paramita Sarkar, S. K. Tripathy, and K. L. Baishnab
Abstract In recent times, methylammonium lead chloride (CH3 NH3 PbCl3 )

perovskites have great significance as the optoelectronic material. Besides all the
promising advantages, there are some problems which need to be resolved, one of
such problem is to synthesize well-crystalline, highly homogeneous and the less-
defective surface of the perovskites by reducing the synthesis complexity. In this
continual search of better properties, we have synthesized of CH3 NH3 PbCl3 film
by replacing traditional halide source of lead with a non-halide source of lead. We
have also analysed their structural, morphological and optical properties of the film
using XRD, FEG-SEM and DRS UV-Vis spectroscopy techniques. The detailed
study confirms the cubic structure of the as-prepared film without any contamina-
tion. Surface morphology shows the formation of homogeneous spherical particles
with an average size of 0.165 μm. DRS UV-Vis data shows a wide bandgap of 3.7 eV
and due to that high transparency in the visible region of light. These outcomes open
up the door for CH3 NH3 PbCl3 film as a vital transparent conductor.
Keywords Spin coating · Urbach tail · Hybrid perovskite · Optical properties ·

Particle size
1 Introduction
Recently, organic–inorganic lead halide (OILH) perovskites APbX3 (A=CH3 NH+ 3;

CH(NH2 )+ 2 , etc., and X=I −
; Br −
; Cl −
) have gained the attention of researchers. These
materials show extraordinary optoelectronic properties, and moreover, they can be
easily synthesized via simple cost-effective solution-processable methodologies. The
exceptional properties of OILH perovskites include tunable bandgap, high mobility
of the charge carriers, higher absorption coefficient, larger diffusion length, etc. These
are the reasons which make these materials that fascinating and enable the horizon
P. Sarkar (B) · S. K. Tripathy · K. L. Baishnab

Silchar, Silchar, Assam 788010, India
930 P. Sarkar et al.
of its use in the field of optoelectronics like solar cells, UV photodetectors, organic
light emitting diodes (LEDs), transparent conductors and lasers [1–5].
Besides all the promising advantages, there are some problems which require
special attention before their use in large-scale production. One of such problem is
to synthesize well-crystalline, highly homogeneous and the less-defective surface
of the perovskites for the better performance of devices. Henceforth, to synthesize
high performing devices, the properties of perovskites need to be tuned either by
the careful selection of starting materials, preparation methodologies, temperatures,
solvents, etc. [6, 7].
The properties of CH3 NH3 PbX3 change dramatically when halide changes from
one to another [8, 9]. It has also been observed by the researchers that the proper-
ties can be significantly changed by substituting the lead halide precursor salts (i.e.
I− , Br− , Cl− sourced) with lead non-halide precursor salts (CH3 COO− , NO− 3 , etc.,
sourced) [10]. There are number of reported works on the synthesis of non-halide
lead sourced CH3 NH3 PbI3 film, and researchers have attained smooth defect-free
surface together with well-crystallinity. Works of the literature suggest that well-
crystalline and defect-free surfaces are the utmost requirements for effective device
fabrication. One of such notable works has been reported by Moore et al. [11], they
have witnessed favourable crystallization kinetics of CH3 NH3 PbI3 using the lead
acetate (Pb(C2 H3 O2 )2 ) precursor, compared to lead iodide (PbI2 ) and lead nitrate
(Pb(NO3 )2 ) [9]. Another work shows that acetate ion-based lead salt accelerates
the growth of perovskite crystal, which leads to achieving homogeneous and less-
defective films [12]. Another group of researchers reported a low-temperature proce-
dure to synthesize uniform films using again the acetate salt, and they have attained
12.5% power conversion efficiency. Similar behaviour of changed morphology and
growth mechanism has been stated by many other reports [13–15].
Mostly, a one-step spin-coating methodology is an efficient way for CH3 NH3 PbX3
thin film deposition. Precursor solutions are typically synthesized by dissolving lead
halide and methylammonium iodide (CH3 NH3 I) into the selected [16, 17]. On the
other hand, to acquire high-grade CH3 NH3 PbI3 films using lead iodide, we need
an anti-solvent treatment, which is very difficult to control practically [18]. To skip
the anti-solvent step, again we need thermal annealing for a longer time. This time-
consuming process may limit its use soon [19, 20]. In recent times, non-halide source
of lead particularly lead acetate (Pb(C2 H3 O2 )2 ) has been commonly used lead source
as it does not require an anti-solvent step or thermal annealing for a long time [21–23].
All of the works mentioned above pave the way to modify the properties as
well as the crystal growth kinetics by introducing different lead salts. Pieces of the
literature suggest that regarding the performances of CH3 NH3 PbX3 perovskite thin
films using a non-halide source of lead, CH3 NH3 PbI3 and CH3 NH3 PbBr3 perovskites
have been studied mostly till this date, whereas the effect is still under investigation
for CH3 NH3 PbCl3 perovskite. As far as we know, there has been no reported work on
the thin films of methylammonium lead chloride (CH3 NH3 PbCl3 ) using lead acetate
as the non-halide source of lead via one-step spin-coating methodology. In this work,
CH3 NH3 PbCl3 perovskite film has been deposited onto the glass substrate, and their
structural, optoelectronic and morphological properties have been studied in detail.
Structural and Optoelectronic Properties … 931
2 Materials and Methodology
2.1 Materials
The starting materials used in this work were in analytical grade purchased from
commercial sources and were used without any further refinement. Lead acetate trihy-
drate (Pb(C2 H3 O2 )2 .3H2 O, 99%), methylamine hydrochloride (CH5 N.HCl, ≥98%),
n,n-dimethylformamide (C3 H7 NO, 99.8%) were procured from Sigma-Aldrich.
Dimethyl sulfoxide (C2 H6 OS 99%) was purchased from Fisher Scientific.
2.2 Methodology
Methylammonium chloride precursor solution was prepared by dissolving

Pb(C2 H3 O2 )2 .3H2 O and CH5 N.HCl in 1:3 molar ratio in 5 ml n,n-
dimethylformamide and dimethyl sulfoxide solvent mixture in 3:5 ratio. The acquired
solution was stirred in a hot plate at 60 °C for 8 h. Afterwards, the solution was coated
onto the glass substrate at 6000 r.p.m. for 1 min in ambient conditions. Before the
deposition of thin film, the substrate was pre-cleaned with soap solution, deionized
water, ethanol, and acetone and finally dried in an oven. After the completion of the
film coating, the coated substrate was annealed at 70 °C for 5 min in an oven and at
60 °C on a hot plate, for 5 min each.
2.3 Characterizations
The structural properties of the film were investigated by Phillips, X’Pert pro diffrac-
tometer using CuKα radiation of 1.5406 Å within the Bragg’s angle (θ ) range of
10–60° at a step size 0.02° in the ambient conditions. The morphology of the sample
was captured by field emission gun scanning electron microscope, FEG-SEM, JEOL
(JSM-7600F). Optical properties have been studied using diffuse reflectance spec-
troscopy (Analytikjena SPECORD). Horiba Jobin Yvon Fluorolog-3 was used to
study the photoluminescence (PL) study of the as-prepared perovskites.
X-ray diffractometer (XRD) study was performed to observe the crystallinity of the
synthesized CH3 NH3 PbCl3 perovskite film. Figure 1 shows the XRD pattern of the
film. All diffraction peaks are in fair agreement with the reported works for cubic
CH3 NH3 PbCl3 perovskite structure, along with space group pm3m [2, 6]. The crystal
Fig. 1 XRD pattern of

as-synthesized
CH3 NH3 PbCl3 film
planes associated with the diffraction peaks at 15.49°, 31.36° and 47.88° are (100),
(200) and (211) (300), respectively. Observed lattice constants are a = b = c =
5.717 Å, and the values are in good agreement with other testified works [24, 25].
Absence of any diffraction peaks related to any of the precursor salts confirms the
formation of single-phased CH3 NH3 PbCl3 perovskite film. Scherrer’s equation has
been employed to find out the average crystallite size of the material, and the equation
is expressed as follows:
0.9λ
D= (1)
βhkl cosθ
Furthermore, XRD data has been used to calculate some other structural properties
[26] such as lattice strain and dislocation density, and the exact values are briefed in
Table 1.
Figure 2 illustrates the microstructures of CH3 NH3 PbCl3 film on the glass
substrate. The surface of the film is homogeneously covered with spherical-shaped
particles. Shapes are regular throughout the surface of the film, and the average
particle size is 0.165 ± 0.02 μm, which has been calculated with the help of ImageJ
software by considering 20 particles. It can be seen from micrographs that some
pinholes are present on the surface. A similar type of pinholes for CH3 NH3 PbCl3
thin films has been observed in a few other reported works [4, 27].
Figure 3 represents the reflectance (R) spectra with respect to wavelength (nm)
of as-synthesized CH3 NH3 PbCl3 film using diffuse reflectance UV-Vis analysis.
Reflectance has been logged within 300–900 nm wavelength. Optical bandgap of
the synthesized film can be calculated from reflectance data obtained from diffuse
reflectance spectrometer. It is an important parameter to know the material’s use
as optoelectronic material. Using the well-known Kubelka—Munk relationship, the
Table 1 Properties of as-synthesized CH3 NH3 PbCl3 film

Parameters This work (using non-halide Other reported works (using halide
sourced lead-based sourced lead-based CH3 NH3 PbCl3
CH3 NH3 PbCl3 film) crystal/film/powder)
Lattice constants (Å) 5.717 5.67 [2], 5.67 [25], 5.683 [26], 5.68
(a = b = c) [28]
Average crystallite size 106.3 115.72 [4], 50.16 [5]
(nm)
Dislocation density 9.606 × 1013 –
(lines/m2 )
Lattice strain 3.35 × 10−4 –
bandgap (eV) 3.7 2.88 [2], 3.8 [4], 2.97 [5], 3.11 [28],
2.94 [25]
absorption coefficient (α) can be obtained as follows [5]:
(1 − R)2
α ∝ F(R) = (2)
2R
Here, F(R) is the Kubelka—Munk function. Moreover, the optical bandgap has
been estimated by using Tauc’s plot method. According to Tauc’s methodology:
αhν = A1 (hν − E g )n (3)
where A1 is the constant, hν is the photon’s energy in eV, n is the exponent whose
value is dependent on the type of electronic transition occurred in the supporting
material. Electronic transitions take places such as direct allowed, direct forbidden,
indirect allowed and indirect forbidden for the values of ‘n’ are equal to 1/2, 3/2, 2
and 3, respectively. As we have seen in Eq. (2) that α is proportional to F(R) and
by taking n = 2 (as CH3 NH3 PbCl3 is direct bandgap material), hence Eq. (3) can be
written as

(F(R).hν)2 = A2 hν − E g (4)
Extrapolation approach on (F(R).hν)2 versus (hν) plot (shown in Fig. 4) gives the
exact value of bandgap, and the calculated value is 3.7 eV. The obtained value of
bandgap is higher than the previously reported values of CH3 NH3 PbCl3 films [29,
30]. The probable reason for achieving higher bandgap may be the thickness of the
deposited film due to the experimental procedure.
When film thickness reduces, then size quantization effect comes to the picture,
and bandgap may reduce. The film is highly transparent in the visible range of light
due to its high bandgap, which is very much appreciable for the materials to be used
as a transparent conductor. Previously, many perovskite oxides have shown wide-
bandgap nature and have been used as transparent conductors. Few notable works
Fig. 2 FEG-SEM
micrographs of
as-synthesized
CH3 NH3 PbCl3 film
are indium doped Cd3 TeO6 (bandgap is 3.8 eV), CaTiO3 (~3.46 eV) and PbTiO3
(~3.4 eV) [31–34].
4 Conclusion
To find better perovskite film as a transparent conductor, CH3 NH3 PbCl3 film has been
fabricated onto a glass substrate using one-step spin-coating methodology using a
non-halide source of the lead (lead acetate). Synthesized CH3 NH3 PbCl3 film is a
Fig. 3 Reflectance (%R)

versus wavelength (nm) plot
of as-synthesized
CH3 NH3 PbCl3 film
Fig. 4 Tauc’s plot for

bandgap determination
cubic crystalline structure with space group pm3m. The average crystallite size,
particle size, optical bandgap and few other structural parameters such as lattice
strain and dislocation density have been measured. The obtained values have been
compared with the available reported works on lead halide sourced CH3 NH3 PbCl3
films/powders/single crystals, due to the lack of available literature of lead acetate-
based CH3 NH3 PbCl3 work. All the values are in good agreement with the compared
results. Lead acetate sourced CH3 NH3 PbCl3 film shows regular-shaped spherical
grains on the surface, and film surface is fully covered, although few pinholes like
defects are also seen. Considering the disorders raised due to lead acetate, it can
be improved by solvent annealing which can effectively increase the crystallinity of
powders and crystallite size. It can be identified as the future scope of work.
Acknowledgements This work was supported by CSIR, New Delhi and ECR scheme (File No.
ECR/2016/001404) of SERB, Government of India.
References
1. Wang Q, Lyu M, Zhang M, Yun J, Chen H, Wang L (2015) Transition from tetragonal to cubic
phase of organohalide perovskite: The role of chlorine in crystal formation of CH3 NH3 PbI3
on TiO2 Substrates. J Phys Chem Lett 6:4379–4384
2. Maculan G, Sheikh AD, Abdelhady AL, Saidaminov MI, Haque MA, Murali B, Alarous E,
Mohammed OF, Wu T, Bakr OM (2015) CH3 NH3 PbCl3 single crystals: inverse temperature
crystallization and visible-blind UV-photodetector. J Phys Chem Lett 6:3781–3786
3. Tian Y, Ling Y, Shu Y, Zhou C, Besara T, Siegrist T, Gao H, Ma B (2016) A Solution-processed
organometal halide perovskite hole transport layer for highly efficient organic light-emitting
diodes. Adv Electron Mater 2:1600165
4. Sarkar P, Tripathy SK, Baishnab KL, Palai G (2019) Structural, optoelectronic, and morpho-
logical study of indium-doped methylammonium lead chloride perovskites. 125:580
5. Sarkar P, Mayengbam R, Tripathy SK, Baishnab KL (2019) Cubic methylammonium lead
chloride perovskite as a transparent conductor in solar cell applications: an experimental and
theoretical study. Ind J Pure Appl Phys 57:891–899
6. Xiao Z, Dong Q, Bi C, Shao Y, Yuan Y, Huang J (2014) Solvent annealing of perovskite-induced
crystal growth for photovoltaic-device efficiency enhancement. Adv Mater 26:6503–6509
7. Nie W, Tsai H, Asadpour R, Blancon JC, Neukirch AJ, Gupta G, Crochet JJ, Chhowalla M,
Tretiak S, Alam MA, Wang HL (2015) High-efficiency solution-processed perovskite solar
cells with millimeter-scale grains. Science 347:522–525(2015)
8. Mashiyama H, Kurihara Y, Azetsu T (1998) Disordered cubic perovskite structure of
CH3 NH3 PbX3 (X=Cl, Br, I). J Kor Phys Soc 32:S156–S158
9. Zhang Z, Ren L, Yan H, Guo S, Wang S, Wang M, Jin K (2017) Bandgap narrowing in bi-doped
CH3 NH3 PbCl3 perovskite single crystals and thin films. J Phys Chem 121:17436–17441
10. Wakamiya A, Endo M, Sasamori T, Tokitoh N, Ogomi Y, Hayase S, Murata Y (2014) Repro-
ducible fabrication of efficient perovskite-based solar cells: X-ray crystallographic studies on
the formation of CH3 NH3 PbI3 layers. Chem Lett 43:711–713
11. Moore DT, Sai H, Tan KW, Smilgies DM, Zhang W, Snaith HJ, Wiesner U, Estroff LA (2015)
Crystallization kinetics of organic–inorganic trihalide perovskites and the role of the lead anion
in crystal growth. J Am Chem Soc 137:2350–2358
12. Zhang W, Saliba M, Moore DT, Pathak SK, Hörantner MT, Stergiopoulos T, Stranks SD,
Eperon GE, Alexander-Webber JA, Abate A, Sadhanala A (2015) Ultrasmooth organic–inor-
ganic perovskite thin-film formation and crystallization for efficient planar heterojunction solar
cells. Nat Commun 6:6142
13. Forgács D, Sessolo M, Bolink HJ (2015) Lead acetate precursor based pin perovskite solar
cells with enhanced reproducibility and low hysteresis. J Mater Chem A 3:14121–14125
14. Qing J Chandran HT, Xue HT, Guan ZQ, Liu TL, Tsang SW, Lo MF, Lee CS (2015) Simple
fabrication of perovskite solar cells using lead acetate as lead source at low temperature. Org
Electron 27:12–17
15. Zhao Q, Li GR, Song J, Zhao Y, Qiang Y, Gao XP (2016) Improving the photovoltaic
performance of perovskite solar cells with acetate. Sci Rep 6:38670
16. Shen C, Courté M, Krishna A, Tang S, Fichou D (2017) Quinoidal 2,2 ,6,6 -tetraphenyl-
dipyranylidene as a dopant-free hole-transport material for stable and cost-effective perovskite
solar cells. Energy Technol 5:1852–1858
17. Wolff CM, Zu F, Paulke A, Toro LP, Koch N, Neher D (2017) Reduced interface-
mediated recombination for high open-circuit voltages in CH3 NH3 PbI3 solar cells. Adv Mater
29:1700159
18. Xiao M, Huang F, Huang W, Dkhissi Y, Zhu Y, Eheridge J, Gray-Weale A, Bach U, Cheng Y-
B, Spiccia L (2014) A fast deposition-crystallization procedure for highly efficient lead iodide
perovskite thin-film solar cells. Angew Chem Int Ed 53:9898–9903
19. Liu Z, Lee E-C Solvent engineering of the electron transport layer using 1,8-diiodooctane for
improving the performance of perovskite solar cells. Org Electron 24:101–105
20. You J, Yang YM, Hong Z, Song T-B, Meng L, Liu Y, Jiang C, Zhou H, Chang W-H, Li G, Yang
Y (2014) Moisture assisted perovskite film growth for high performance solar cells. Appl Phys
Lett 105:183902
21. Qing J, Chandran H-T, Cheng Y-H, Liu X-K, Li H-W, Tsang S-W, Lo M-F, Lee C-S (2015)
Chlorine incorporation for enhanced performance of planar perovskite solar cell based on lead
acetate precursor. ACS Appl Mater Interfaces 7:23110–23116
22. Li C, Guo Q, Wang Z, Bai Y, Liu L, Wang F, Zhou E, Hayat T, Alsaedi A, Tan ZA (2017) Efficient
planar structured perovskite solar cells with enhanced open-circuit voltage and suppressed
charge recombination based on a slow grown perovskite layer from lead acetate precursor.
ACS Appl Mater Interfaces 9:41937–41944
23. Liu Y, Liu Z, Lee E-C (2018) Dimethyl-sulfoxide-assisted improvement in the crystallization
of lead-acetate-based perovskites for high-performance solar cells. J Mater Chem C 6:6705–
6713(2018)
24. Zhang Z, Zheng W, Lin R, Huang F (2018) High-sensitive and fast response to 255 nm deep-UV
light of CH3 NH3 PbX3 (X=Cl, Br, I) bulk crystals. R Soc Open Sci 5:180905
25. Hsu HP, Li LC, Shellaiah M, Sun KW (2019) Structural, photophysical, and electronic
properties of CH3 NH3 PbCl3 single crystals. Sci Rep 9:13111
26. Studenyak I, Kranjec M, Kurik M (2014) Urbach rule in solid state physics. Int J Opt Appl
4:76–83
27. Wang ZK, Li M, Yang YG, Hu Y, Ma H, Gao XY, Liao LS (2016) High efficiency Pb–In binary
metal perovskite solar cells. Adv Mater 28:6695–6703
28. Nandi P, Giri C, Swain D, Manju U, Topwal D (2019) Room temperature growth of
CH3 NH3 PbCl3 single crystals by solvent evaporation method. Cryst Eng Commun 21:656–661
29. Zheng E, Yuh B, Tosado GA, Yu Q (2017) Solution-processed visible-blind UV-A photode-
tectors based on CH3 NH3 PbCl3 perovskite thin films. J Mater Chem C 5:3796–3806
30. Zhang Z, Ren L, Yan H, Guo S, Wang S, Wang M, Jin K (2017) Bandgap Narrowing in bi-doped
CH3NH3 PbCl3 perovskite single crystals and thin films. J Phys Chem C 121:17436–17441
31. Tetsuka H, Shan YJ, Tezuka K, Imoto H, Wasa K (2005) Transparent conductive In-
doped Cd3 TeO6 thin films with perovskite structure deposited by radio frequency magnetron
sputtering. J Mater Res 41:2256–2260
32. Balachandran U, Odekirk B, Eror NG (1982) Electrical conductivity in calcium titanate. J Solid
State Chem 41:185–194
33. Piskunov S, Heifets E, Eglitis RI, Borstel G (2004) Bulk properties and electronic structure of
SrTiO3 , BaTiO3 , PbTiO3 perovskites: an ab initio HF/DFT study. Comput Mater Sci 29:165–
178
34. Kumar A, Balasubramaniam KR, Kangsabanik J, Vikram, Alam A (2016) Crystal structure,
stability, and optoelectronic properties of the organic-inorganic wide-band-gap perovskite
CH3 NH3 BaI3 : candidate for transparent conductor applications. Phys Rev 94:180105–180110
Robust Gunshot Features and Its
Classification Using Support Vector
Machine for Wildlife Protection
Vivek Singh, Kailash Chandra Ray, and Somanath Tripathy
Abstract In the present study, automatic gunshot sound event detection in forest
areas for wildlife protection has been considered. The feature extraction method
used is robust to any length of gunshot sound events, which also takes care of the
burst of multiple gunshots. First, low-level DWT-based features of large dimension
were extracted which projected onto high-level histogram feature vector of small
dimension using the bag-of-words approach. Support vector machine (SVM) classi-
fier was considered to classify input audio signals into gunshot or forest ambience.
The obtained results are highly reliable with a classification accuracy of 96.04% and
area under ROC curve of 0.9866 indicating low false alarming rate. Automatic audio
event detection can effectively expand the overall consistency of forest surveillance
systems.
Keywords Wildlife protection · Gunshot detection · Discrete wavelet transform ·

Support vector machine
1 Introduction
Gunshot sound is a vital component of the acoustic world which conveys a situation
of emergency for both legal and medical teams. In this paper, a wildlife protec-
tion system using automatic detection of gunshot sound event has been considered.
Poaching using a firearm is pervasive all over the world. Continuous monitoring of
forest areas for such criminal activities using a piece of audio information is the main
aim of this study. Gunshot being an impulsive sound event which lasts only for few
milliseconds is more challenging to detect than the sustained sound events.
V. Singh (B) · K. C. Ray (B) · S. Tripathy

Indian Institute of Technology, Patna, India
K. C. Ray
S. Tripathy
940 V. Singh et al.
In [1], Clavel et al. detected alarming events using GMMs in two steps: in the
first step, detected whether sound is alarming or normal and in the second step,
it classifies alarming sound into category of shot (e.g. gunshot, riffle shot, etc.),
but the results obtained cannot be generalized. Gerosa et al. in [2] also used two
classifiers-based approach, respectively, for scream and gunshot against background
sound and achieved an accuracy of 93% with very high false rejection rate of 5%.
In [3], Chacon et al. compared six gunshot detection algorithm and concluded that
feature obtained from discrete wavelet transform performs best. George et al. in [4]
reported the advantages of audio analysis using DWT for non-speech signals. In [5],
Rabaoui et al. used one-class wavelet base audio feature with SVM and achieved an
accuracy of 96.89%. Pittner and Kamarthi in [6] reported how wavelet coefficients
could be used for feature extraction of variable length audio events. In [7], Istrate
et al. extracted audio feature using wavelet decomposition for telemonitoring medical
application and reported a false alarm rate of only 3%. In [8], Oltean et al. extracted
audio features based on DWT for detecting alarming event for forest application,
reported an accuracy of 95%; however, size of dataset used was very small to be
generalized.
Reliable detection and classification of gunshot sound against the background
need a robust feature extraction method which can be classified with low complexity
machine learning models. In the proposed methodology, the discrete wavelet
transform-based audio features were extracted which projected onto high-level
histogram features. The classification is achieved using support vector machine clas-
sifier. The result shows quite an improvement compared to previously reported work
on similar sound classes.
1.1 Existing Methods
Numerous feature extraction and classification methods have been worked upon in
the field of automatic recognition of acoustic events. The feature extraction methods
are used to highlight the similar features of intraclass sound, whereas discriminative
features of inter-class sound. The goal is to achieve both by means of less coefficients.
The feature extraction comprises the temporal domain [9], spectral or frequency
domain [9], low-level audio descriptors [9], time–frequency [10], cepstral [10], Mel-
scaled [5] for getting the prominent features from the acoustic signal. These features
were used to train the machine learning model such as deep neural networks [11],
RNN [11], CRNN [10], non-negative matrix factorization (NMF) [12], GMM-hidden
Markov model (HMM) [13] and support vector machines (SVM) [14] to achieve
automatic recognition of acoustic events. Mel-scaled-based audio feature extraction
forms a valuable feature for audio classification and is the leading feature extraction
technique in reported literature [13].
In our study, low-level time–frequency DWT-based features projected onto high-
level histogram features using the bag-of-words approach are used as an audio feature.
The proposed feature extraction technique has not been used for gunshot detection
Robust Gunshot Features and Its Classification … 941
task in any of the reported literature for comparable studies. The feature classi-
fication method was achieved using support vector machine (SVM), as SVM is
lightweight classifier from hardware point. This study has been done in view of
resource-constrained devices.
Paper is organized as follows: Sect. 2 describes the proposed methodology in
detail, followed by the experimental procedure in Sect. 3. Gotten results are discussed
in Sect. 4. Finally, we conclude the paper, followed by future suggestions in Sect. 5.
The objective of the proposed methodology is to classify gunshot sound against the
background sound, i.e. forest ambience. In this work, a sound that is assumed to be
composed of the small atomic audio unit is considered. A bag-of-words approach, in
which the audio to be classified represented by the histogram of occurrences of each
atomic audio unit, is employed to form a high-level feature vector from low-level
discrete wavelets transform (DWT) coefficients
The proposed architecture for gunshot classification consists of four stages:
(a) DWT-based low-level audio features extraction, (b) unsupervised learning of
basic audio units using K-means clustering algorithm, (c) high-level feature vector
construction and (d) feature classification.
2.1 Dataset Pre-processing
Gunshot data considered in this paper is the subset of airborne [15] gunshot database.
The considered gunshot data consists of 290 files of 21 different firearms. The number
of gunshots present in each record varies from one event to ten events. Some record-
ings also contain a burst of multiple gunshots which ranges from two to eight in a
row. Hence, this type of gunshot events is also taken care of in the study. Gunshot
sound events were isolated to form a dataset of individual gunshot event. In total,
the dataset contains 811 individual gunshot events. For background ambience noise,
forest sound from freesound.org [16] database is considered, having a total ambience
sound of 35 min. Impulsive sound events from forest ambience were extracted and
segmented into the small-signal; in total, there were 1113 individual impulsive sound
events belongs to forest ambience. Detail of dataset used is summed in Table 1.
Table 1 Detail of dataset

Class Gunshot Forest ambience
used for this study
Total 811 1113
942 V. Singh et al.
As audio data considered are obtained from different sources, standardization of

data was required. All the dataset converted into single-channel resampled to 16 kHz
and normalized in the range [−1, 1].
2.2 DWT-Based Low-Level Features Extraction
An acoustic signal may show sudden changes contained in a few milliseconds.

Accordingly, in order to take into account its short-time changeability, the audio
stream is fractioned into small frames with 50% overlapping. Each frame is then
decomposed using fast DWT into L sub-bands. The DWT analyses acoustic signal
at several scales with a different resolution by decomposing the signal into a
coarse approximation and detail information. The decomposition of the signal into
different scales is achieved by sequential high-pass and low-pass filtering of the
time–domain signal. In the fast DWT, only one-sided, i.e. approximated coefficients
are decomposed further into sub-bands, whereas detailed coefficients remain intact.
Mathematically, a single level of decomposition is expressed by Eq. 1a and 1b.

Mhigh [k] = x[n].h[2k − n] (1a)
n

Mlow [k] = x[n].l[2k − n] (1b)
n
where M high [k] and M low [k] are the output of the highpass, h[n] and low pass, l[n]
filters, respectively; after decimation by 2, the low-level feature vector of a single
frame is obtained by appending coefficients of all levels.
2.3 Unsupervised Learning
Fixed set of audio units is achieved by using K-means clustering algorithm. The
outcome of K-means clustering is a set of K points that relate to the centroids of the
clusters. Each centroid represents a group of similar low-level vectors, and the set C
= {c1 , …, cK } of the centroids is regarded as the dictionary of basic audio units.
2.4 A High-Level Histogram Feature Vector
Figure 1 depicts the method of building a high-level feature vector, given a dictionary
C of cluster’s centroids, for every low-level vector vi , the nearest audio unit cj is
determined. The frequency of each word cj in a time-limited interval is used to
Low-level vector Dictionary C, of clusters High-level feature vector H

(Histogram)
v1 c2 3
c1
v2
c9 2
c7
v3
c3 c5
1
Ö ck c6
vn-1 c4
c10 c8
vn
ck-1
(a) (b) (c)
Fig. 1 Building of the high-level histogram feature vector. Every low-level vector vi (a) is mapped
to its closest atomic audio unit cj in the dictionary (b). Then, the rate of the single atomic audio unit
is stored in a histogram, having bins as hj (j = 1, …, K) that contains the high-level feature vector
(c)
construct a high-level feature vector. Such vector corresponds to the histogram H =

(h1 , …, hK ), whose bins are computed as:

M
hj = δ(bi , j), j = 1, . . . , K (2)
i=1
where δ (·) is the Kronecker delta, and bi is the index of a word within the set C,
determined as:
bi = arg min d(vi , c j ), j = 1, . . . , K (3)

j
where d(vi , cj ) is a dissimilarity measure between the vector cj and the prototype vi
(the Mahalanobis distance is considered).
2.5 Classification
Support vector machine (SVM) is a binary discriminative classifier. In the SVM

classification scheme, the selection of the kernel function parameter is entirely data-
dependent and chosen empirically. This paper considers the radial basis function
(RBF) kernel for implementing SVM approach to classify the gunshot sound events
against
the forest ambience signals. The RBF kernel is represented by the expression
−γ xi −x j
2
as K (xi , x j ) = exp . The regularization parameter C and kernel function
argument parameter γ of the classifier are varied in the range of [2−5 , 210 ] and [2−10 ,
25 ], respectively, and optimized using PSO [14] technique. In this paper, two classes
have been considered, i.e. gunshot and forest ambience, which required only one
SVM to be trained. The high-level histogram feature vectors were utilized for training
944 V. Singh et al.
SVM model. The SVM classifier is particularly suited for the employed acoustic
representation since it is able to learn the words that are relevant for a particular class
of events and discard that do not contribute to an useful classification, by giving them
a little weight. The parameters of the classifier, i.e. C and γ , are slowly tuned using
the PSO algorithm in the training phase. The fivefold cross-validation is performed
on the training dataset, with the preeminent learning parameter values of C and γ
to yield the maximum classification performance. The testing is conducted on the
testing dataset, and the performance is presented in the form of the confusion matrix
3 Experiment
The experiment was conducted on detecting gunshot sound events against the back-
ground noise, i.e. forest ambience sound. Our approach consists of six stages, viz.
pre-processing, low-level DWT feature extraction, K-means clustering, high-level
histogram feature extraction, training SVM model and testing. The pre-processing
step is discussed in Sect. 2.1.
The training dataset was utilized to obtain the cluster centres using K-means
clustering algorithm. First, low-level features of small audio segment, i.e. frame, were
extracted. The population of frames kept very high as each frame act as an atomic
audio unit. Next paragraph discusses how these atomic audio units are extracted.
Every signal in training dataset undergoes proceeding steps.
One signal at a time is taken from the training dataset and segmented into small
overlapping (50% hopping with adjacent frames) frames of 16 ms duration, i.e. 256
sample points for 16 kHz sampled signal. Next, each frame undergoes five-level DWT
decomposition using Daubechies4 wavelet. As the number of coefficients varies with
levels, a uniform low-level feature vector was obtained by appending coefficients of
all five levels together to form a single long feature vector of a frame. This procedure
was repeated for all the frames of the signal (Fig. 2).
After performing the above procedure to all the signals from a training dataset
of, we get a considerable number of data for unsupervised learning. Unsupervised
Low-level
feature vector of K-means
all training clustering
Train SVM model
dataset -5 10
(RBF Kernel, C=2 ó 2 ,
−10 5
γ=2 − 2 , 5− Fold cross
Training
validation)
data
Repeat for all training and testing dataset Training

Training and Testing
data
Repeat for all frames Gunshot
Database
Segmentation Dictionary High-level

dataset
Append Trained SVM

(16ms frames, Decomposition of atomic Bag-of-words histogram
coefficients classifier model
50% overlapping) (DWT) audio unit feature vector
of all levels Forest ambience
Train/Test Train/Test Train/Test
Test data Prediction
data data data
(Prediction)
Fig. 2 Block diagram representation of the proposed method

learning treats all the frame-level feature vectors as an individual signal; i.e. K-means
clustering algorithm sees all frames as individual signals having a feature vector of
length 288. The database is split into a 50/50 manner for training and testing dataset,
respectively. K (= 20) number of clusters centre were learned out of the given dataset.
For training the SVM model in a supervised manner, the same training dataset
containing individual gunshot events and forest ambience sound was considered.
High-level histogram features were extracted out of individual signal. The dimension
of the high-level feature vector is the same as the number of clusters. High-level
histogram features were obtained by first decomposing the signal using DWT and
preparing a feature vector out of it the same as for unsupervised learning. After getting
a low-level feature matrix of an individual signal of dimension (a × 288); a may
vary with the length of the audio event, each a vector of dimension 288 is predicted
for its cluster number. The output of the prediction has a dimension of (a × 1); i.e.
each a vector is now representing by its cluster number. The histogram is formed
for a percentage of weight-age each cluster has in the particular audio event. The
final dimension of the feature vector or the high-level histogram feature is (20 × 1).
After obtaining a high-level histogram feature vector for the whole training dataset,
supervised learner, i.e. a support vector machine, is trained as a binary classifier.
For testing purpose, the test dataset is considered. High-level histogram features
were extracted and fed into the SVM classifier. The output shows a promising
result with a respectable detection accuracy of 96.98%. The main concern in the
surveillance-based application is a false alarming rate. The proposed method shows
a false alarming rate of only 1.34%, which can be considered as a significant result.
For training the SVM model, 50%, i.e. 964, of the total available dataset was used. In
terms of the number of signals for each class, i.e. forest ambience and gunshot event,
576 and 388, respectively. Fivefold cross-validation was taken into consideration. An
SVM model is optimized using particle swarm optimization algorithm. The value for
C = 8 and γ = 7 shows an optimum result. Classifier model is trained with an accuracy
of 97.19%. Testing data consists of 547 signals for forest ambience and 414 signals for
gunshot events. The testing accuracy obtained was 96.04%, which can be considered
as remarkable for the amount of dataset. Receivers operating characteristic (ROC)
curve as shown in Table 2 under 50/50 data split is also computed; the area under
the ROC curve of 0.9884 indicates a low false alarm rate and high true positive rate.
The experiments were also conducted for different sizes of training, and testing
dataset splits from the original database. Table 2 sums the results, i.e. confusion
matrix and ROC curve of all the different splits. From Table 2, it is observed that, as
we increase the size of training data, the accuracy of the proposed methodology does
not vary much. Also, improvement in area under ROC curves is not significant for a
different number of training examples. Hence, conclusion from the above observation
946 V. Singh et al.
Table 2 Confusion matrix and ROC curve for different dataset splits
Train/Test Confusion matrix ROC curve
50/50 F G
F 0.96 0.04
G 0.04 0.96
60/40 F G
F 0.96 0.04
G 0.06 0.94
70/30 F G
F 0.97 0.03
G 0.06 0.94
80/20 F G
F 0.98 0.02
G 0.04 0.96
is that, the proposed feature extraction technique is robust to the size of the training
dataset.
The proposed method is independent of the length of the audio events. In gunshot
dataset, the length of events varies from 100 to 1200 ms. The proposed model sees
all gunshot events as equal length by extracting histogram features, as discussed in
Sect. 2.4. The dataset is split in such a way that 8 firearm models are not included
in the training set. Though the above constraint, the proposed methodology robustly
detects the event into the gunshot class. Thus, our model can be generalized for the
completely unseen firearm sounds. In training as well as in testing datasets, a burst
of gunshot firing sound is also included. In the testing phase, the proposed system
successfully detected all the ten signals of burst gunshot sounds.
This paper proposes a gunshot detection system against the background sound of
forest ambience; for wild areas, surveillance is based on DWT features projected onto
high-level histogram features and recognized with support vector machine classifier.
In our approach, we have considered that an event has already been detected blindly
from the continuous recording, and the audio segment is fed into the proposed system.
The detection accuracy of 96.04% with area under ROC curve of 0.9866 was achieved
for 50% - 50% split of an available dataset for training and testing, respectively. As
future work, authors are working on detection of gunshot events with different SNR
from forest ambience.
Acknowledgements MeitY, Government of India, New Delhi, partially supports this work through
the project titled ‘Special Manpower Development Program for Chips to System Design’ underneath
project reference no. R&D/SP/EE/DEITY/SMD/2015-16/126 and Visvesvaraya PhD Scheme.
References
1. Clavel C, Ehrette T, Richard G (2005) Events detection for an audio-based surveillance system.
In: 2005 IEEE international conference on multimedia and expo, Amsterdam, pp 1306–1309
2. Valenzise G, Gerosa L, Tagliasacchi M, Antonacci F, Sarti A (2007) Scream and gunshot detec-
tion and localization for audio-surveillance systems. In: 2007 IEEE conference on advanced
video and signal based surveillance, London, pp 21–26
3. Chacon-Rodriguez A, Julian P, Castro L, Alvarado P, Hernandez N (2011) Evaluation of gunshot
detection algorithms. IEEE Trans Circ Syst I Regul Pap 58(2):363–373
4. Tzanetakis G, Essl G, Cook P (2001) Audio analysis using the discrete wavelet transform. In:
Proceedings of conference on acoustics and music theory applications, Sept 2001
5. Rabaoui A, Davy M, Rossignol S, Ellouze N (2008) Using one-class SVMs and wavelets for
audio surveillance. IEEE Trans Inf Forensics Secur 3(4):763–775
6. Pittner S, Kamarthi SV (1999) Feature extraction from wavelet coefficients for pattern
recognition tasks. IEEE Trans Pattern Anal Mach Intell 21(1):83–88
7. Istrate D, Castelli E, Vacher M, Besacier L, Serignat J (2006) Information extraction from
sound for medical telemonitoring. IEEE Trans Inf Technol Biomed 10(2):264–274
8. Oltean G, Grama L, Ivanciu L, Rusu C (2015) Alarming events detection based on audio signals
recognition. In: 2015 International conference on speech technology and human-computer
dialogue (SpeD), Bucharest, pp 1–8
9. Mesaros A, Heittola T, Virtanen T (2016) Metrics for polyphonic sound event detection. Appl
Sci 6(6), Article 162
10. Mesaros A et al (2018) Detection and classification of acoustic scenes and events: outcome of
the DCASE 2016 challenge. IEEE/ACM Trans Audio Speech Lang Process 26(2):379–393
11. Adavanne S, Politis A, Nikunen J, Virtanen T (2019) Sound event localization and detection
of overlapping sources using convolutional recurrent neural networks. IEEE J Sel Top Signal
Process 13(1):34–48
12. Komatsu T, Toizumi T, Kondo R, Senda Y (2016) Acoustic event detection method using semi-
supervised non-negative matrix factorization with a mixture of local dictionaries. In: DCASE
2016 challenge, technical report, Sept 2016
948 V. Singh et al.
13. Sigtia S, Stark AM, Krstulovi S, Plumbley MD (2016) Automatic environmental sound recog-
nition: performance versus computational cost. IEEE/ACM Trans Audio Speech Lang Process
24(11):2096–2107
14. Raj S, Ray KC (2017) ECG signal analysis using DCT-based DOST and PSO optimized SVM.
IEEE Trans Instrum Meas 66(3):470–478
15. https://www.airbornesound.com/
16. https://freesound.org/
Effect of pH Variations on Zinc Oxide
(ZnO) Nanowire Dispersions
Vimal Kumar Singh Yadav and Roy Paily
Abstract This work analyses the effect of pH variations on the morphological prop-
erties and dimensional modulation of ZnO nanowires (NWs) aqueous dispersions in
de-ionized (DI) water. To observe the effect of lower pH values, different sets of ZnO
nanowire dispersions (NWD) have been prepared and a certain amount of drop-casted
dispersion has been added with acetic acid, hydrochloric acid (HCl) and sulphuric
acid respectively. Further, ammonium hydroxide has been added in the same propor-
tion as the aqueous ZnO NWD to observe the effect of higher pH values. Such a
study can be instrumental in the development of ZnO NWDs consisting of well sepa-
rated, sharp-edged and long nanowires for the fabrication of single-nanowire based
electronic devices, which can be applied as sensors, energy harvesters or display
units.
Keywords Zinc oxide · Nanowire · pH · Dispersion
1 Introduction
Zinc oxide (ZnO) nanowires (NWs) have been deeply researched due to their unique
and attractive material properties, which are being employed in the field of elec-
tronics, optics, energy harvesting, and so on [1–3]. ZnO NWs are generally synthe-
sized over the substrate using vapour–liquid–solid (VLS) method with the help of
chemical vapour deposition (CVD), laser ablation method, atomic layer deposition
(ALD), hydrothermal or solvothermal techniques and so on [4–8]. However, most of
the above-mentioned methods use some kind of catalyst or auxiliaries that leads to
residual contamination of ZnO NWs [4].
V. K. S. Yadav (B) · R. Paily

Guwahati, Assam, India
R. Paily
950 V. K. S. Yadav and R. Paily
Therefore, more efficient method of ZnO NW synthesis requires a contamination-

free process that does not alters the original properties of a NW. Shih et al. have
reported the synthesis of a ZnO single NW (SNW) using a Ti-assisted CVD in order
to avoid catalytic contamination which can control the dimensions of NWs with the
help of a diffusion phenomenon taking place during the fabrication process [4]. Lupan
et al. have reported the low-temperature solution-based synthesis of Au-modified
ZnO NWs with a diameter varying in the range of 10–100 nm [8]. Similarly, Galan
et al. have used atomic layer deposition (ALD) to grow ZnO seed layers followed
by chemical bath deposition (CBD) of ZnO NWs to produce uniform single crystal
ZnO NWs in low-cost [9].
In this work, the effect of acid (lower pH variations) and base addition (higher pH
variations) to DI water-based ZnO nanowire-dispersion on the morphology of ZnO
NWs, especially on ZnO single-nanowire (SNW) (width or diameter of ZnO wire
<500 nm) or single micro-wire (SMW) (width or diameter of ZnO wire >500 nm) is
investigated. For this purpose, a commercial ZnO NW powder has been purchased
and an optimized dispersion has been prepared utilizing DI water as a dispersion
medium and several drops of the prepared dispersion have been drop-casted in µL
volume on a cleaned SiO2/Si substrate and different acids and base have been added to
this drop-casted dispersion drop in varied proportions. The optimized acid proportion
helps to form sharp-edged and longer ZnO SNW and SMW which can be efficiently
used to fabricate discrete electronic components such as micro-resistors, Schottky
diodes, and so on [10].
2 Materials and Methods
2.1 Dispersion Preparation
For all the experiments, ZnO Nanowire powder is procured from Sigma Aldrich
(Product Id: 773999) and a low surface-tension DI (LST-DI) water is prepared simply
by adding 1 mg of sodium dodecyl sulfate (SDS) to 5 mL of DI water to improve
the dispersibility of ZnO NWs in the aqueous dispersion as shown in Fig. 1. ZnO
NW dispersion is prepared by adding 2 mg of ZnO NW powder in 1200 µL of LST-
DI water. The reason to experiment with a very small concentration of ZnO NW
dispersion in ‘µL’ level as compared to ‘mL’ level is to keep the density of ZnO
NWs high so that the effect of acids can be observed clearly. 2 µL of above prepared
ZnO NW dispersion is drop-casted on to a cleaned silicon substrate and the initial
image is captured with the camera of a Material Printing System (MPS) [11].
Effect of pH Variations on Zinc Oxide (ZnO) Nanowire Dispersions 951
Fig. 1 Optical image showing a portion of ZnO NW dispersion drop in a DI water only b DI water
+ SDS after heating. ZnO NWs are in black color
3.1 Effect of Acid Addition to ZnO NW Dispersion in LST-DI

Water
The drop-casted ZnO NW dispersion is in white color as seen from naked eyes. Then,
2 µL of acetic acid, hydrochloric acid (HCl) and sulphuric acid (H2 SO4 ) is added to
the already drop-casted ZnO NW dispersion to keep the proportion same (1:1). It is
observed that (H2 SO4 ) dissolves ZnO NWs in all the proportions. HCl also partially
dissolves ZnO NWs, however, these NWs turned into micro-particles when heated
at ~220 °C as shown in Fig. 2.
When acetic acid is added in 1:1 proportion with NW dispersion, it is observed
that within a few seconds, the white color of the dispersion changes to colorless and
transparent like LST-DI water as shown in Fig. 3a, b. The ZnO SNW with length
<10 µm is almost invisible at this scale as shown in Fig. 3b. Further, acetic acid is
Fig. 2 Optical image showing a HCl partially dissolving ZnO NWs b NWs converted to micro-
particles at high temperatures
Fig. 3 Optical image showing a portion of 2 µL drop-casted ZnO NW dispersion a without acetic
acid b with acetic acid in 1:1 proportion
added in different volume proportions such as 1:1 (0.5 µL acetic acid in 0.5 µL of
NW dispersion), 1:10 (0.5 µL acetic acid in 5 µL of NW dispersion) and 1:40 (0.5 µL
acetic acid in 20 µL of NW dispersion) as shown in Fig. 4. It is observed that acetic
acid addition to ZnO NW dispersion in different proportions is resulting in long and
sharp-edged NWs and therefore has the capability to modulate the morphology of
already synthesized ZnO NWs due to its different pH range as compared to H2 SO4
or HCl [12–14].
It is also observed that acetic acid dissolved all the ZnO NWs in those dispersions
in which SDS is not added. This is strong evidence that the, it is the combination
of acetic acid and SDS in a proper ratio is mainly responsible for the creation of
Fig. 4 Optical image showing sharp and long ZnO micro-wires formation after acetic acid addition
in a 1:1 b 1:10 c 1:40 proportion d 1:10 drop periphery e zoomed-in image of encircled region in
d, f ZnO nanowires converting to micro-particles at higher temperatures. AA: acetic acid NWD:
NW dispersion
Fig. 5 Optical image showing the formation of ZnO NW mat-cluster ring after NH4 OH addition
to ZnO NW aqueous dispersion in a 1:20 b 1:40 after heating the drop at 180 °C
long, sharp-edged, and highly distinguishable ZnO NWs [15, 16]. However, the
dispersion containing acetic acid also converts nanowires and micro-wires into ZnO
micro-particles when exposed to high temperatures near ~220 °C as shown in Fig. 4f.
3.2 Effect of Base Addition to ZnO NW Dispersion in LST-DI

Water
To observe the effect of adding base (higher pH substance) on the morphology of ZnO
NWs, 0.5 µL of ammonium hydroxide (NH4 OH) was added in 0.5 µL (1:1), 10 µL
(1:20) and 20 µL (1:40) of ZnO aqueous dispersion (1 mL DI water + 3 mg ZnO NW
powder). As observed from the optical images taken from MPS microscope, NH4 OH
shows no tendency to either dissolve the NWs or separate them. It forms a ring of
ZnO NW mat-cluster of width near 25–50 µm with small, single, and unmodified
ZnO NWs scattered randomly along the periphery of the ring as shown in Fig. 5.
The width of NW cluster ring increases from ~32 to 60 µm with a dilution ratio, as
shown in Fig. 5a, b.
4 Conclusion
This work studies the effect of acetic acid, HCl, H2 SO4 (acids), and NH4 OH (base),
on the morphological changes of ZnO NWs in aqueous dispersions. It is observed
that acetic acid with SDS in 1:40 to 1:1 proportion with ZnO NW aqueous dispersion,
for ZnO NWD drop-casted volume up to 5 µL, is optimized for ZnO single-nanowire
or single micro-wire formation. It is observed that HCl turns NW to micro-particles
at high temperature and H2 SO4 dissolve ZnO NWs completely. ZnO NW micro-
mat is formed by the addition of ammonium hydroxide in ZnO NW dispersion in
Fig. 6 Optical image showing the effect of annealing temperature on ZnO NWs at a 180 °C b 220 °C
LST-DI water. The ZnO NWs show appreciable, thermal, and structural stability, at
temperatures ~220 °C as shown in Fig. 6a, b.
References
1. Zhang Y, Ram MK, Stefanakos EK, Goswami DY (2012) Synthesis, characterization, and
applications of ZnO nanowires. J Nanomater 2012:22
2. Pan ZW, Dai ZR, Wang ZL (2001) Nanobelts of semiconducting oxides. Science
291(5510):1947–1949
3. Poulin-Vittrant G, Dahiya AS, Boubenia S, Nadaud K, Morini F, Justeau C, Alquier D (2019)
Challenges of low-temperature synthesized ZnO nanostructures and their integration into nano-
systems. Mater Sci Semicond Process 91:404–408
4. Shih P-H, Wu S (2017) Growth mechanism studies of ZnO nanowires: experimental
observations and short-circuit diffusion analysis. Nano-materials 7(7):188
5. Shih HY, Chen TT, Chen YC, Lin TH, Chang LW, Chen YF (2009) Size-dependent photoelastic
effect in ZnO nanorods. Appl Phys Lett 94(2):021908
6. Jung M-H, Chu M-J (2014) Synthesis of hexagonal ZnO nanodrums, nanosheets and nanowires
by the ionic effect during the growth of hexagonal Zno crystals. J Mater Chem C 2:6675–6682
7. Chu Y, Wan L, Wang X, Zhang J (August, 2012) Synthesis and characterization of ZnO
nanowires by solvothermal method and fabrication of nanowire-based ZnO nanofilms, pp
366–369
8. Lupan O, Postica V, Wolff N, Su J, Labat F, Ciofini I, Cavers H, Adelung R, Polonskyi O, Faupel
F, Kienle L, Viana B, Pauport T (2019) Low-temperature solution synthesis of au-modified ZnO
nanowires for highly efficient hydrogen nanosensors. ACS Appl Mater Interfaces 11(35):32
115–32 126
9. Galan-Gonzalez A, Gallant A, Zeze DA, Atkinson D (May, 2019) Controlling the growth of
single crystal ZnO nanowires by tuning the atomic layer deposition parameters of the ZnO seed
layer. Nanotechnology 30(30):305602
10. Yadav VKS, Natu G, Paily R (2020) Fabrication and electrical characterization of printed micro-
resistors of silver nanoparticles us-ing micro-cantilever based printing technology. IEEE Trans
Compon Packaging Manuf Technol 10(1):57–64
11. BioForce Nanosciences. Accessed April 10, 2019. Available: www.bioforcenano.com
12. S. Edinger, J. Bekacz, M. Richter, R. Hamid, R. Wibowo, A. Pei, and T. Dimopoulos.: Influence
of the acetic acid concentration on the growth of Zinc oxide thin films prepared by spray
pyrolysis of aqueous solutions,” Thin Solid Films, vol. 594, pp. 238 – 244, (2015)
13. Jiao B, Zhang X, Wei C, Sun J, Huang Q, Zhao Y (2011) Effect of Acetic acid on ZnO:
in transparent conductive oxide prepared by ultrasonic spray pyrolysis. Thin Solid Films
520(4):1323–1329
14. Biswal R, Maldonado A, Vega-Perez J, Acosta D, De La Luz Olvera M (2014) Indium doped
zinc oxide thin films deposited by ultrasonic chemical spray technique, starting from zinc
acetylacetonate and indium chloride. Materials 7(7):5038–5046
15. Wa Yu H, Wang J, An Yan X, Wang J, Fei Cheng P, Juan Xia C (2019) Effect of surfactants on
the morphology and photocatocatalytic properties of ZnO nanostructures. Optik. 185:990–996
16. Usui H (2009) The effect of surfactants on the morphology and optical properties of precipitated
wurtzite ZnO. Mater Lett 63(17):1489–1492
Intensity Modulated U-Shaped Bent
Tapered Optical Fiber Concentration
Sensor
Putha Kishore, D. Dinakar, Manchineellu Padmavathi, and L. Obulapathi
Abstract In this paper, a liquid concentration (viscosity) sensor is designed and

developed at a low cost using a clad removed U-shaped bent plastic optical fiber
(POF). A small portion of POF cladding is removed and bends to form a U-shaped
probe for an optimized radius to improve the sensitivity of parameter measurement.
The probe is immersed in a viscous liquid, then the interaction of the evanescent field
with the ambient viscous liquid in both forward and reverse directions movements
of the liquid were recorded. As the probe dipped slowly in and out of the liquid an
offset is observed in the response of the sensor, it can be measured as the hysteresis
loss in the loop. The offset has been measured in both horizontal (Distance) and
vertical (Intensity) directions. The experimental results reveal that the sensor is highly
linear with about 0.98 linearities. From the results, It is observed that intensity offset
measurement is more suitable than distance offset owing to high sensitivity. Thus, the
concentration (or viscosity) of the liquid can be measured using the proposed simple
sensor in terms of the offset height and hysteresis area. This sensor is easy to make,
simple, low cost, and had inherent advantages of the optical fiber like Immune to EMI,
corrosive, small size, flexibility in length. The sensor may found many applications in
measuring the viscosities of fluids such as motor oils, petroleum products, chemical
solvents, and many other industrial and biological fluids.
Keywords Plastic optical fiber · U-shaped bend · Intensity modulated ·

Concentration · Viscosity · Hysteresis area
P. Kishore · L. Obulapathi
Department of H&S, Annamacharya Institute of Technology and Science, Rajampet, Kadapa,
India
D. Dinakar (B)
Department of Physics, National Institute of Technology Warangal, Warangal, India
M. Padmavathi
Department of Physics, SKR and SKR Government Degree College for Women, Kadapa, India
958 P. Kishore et al.
1 Introduction
Analysis of the liquid viscosity or concentration enables to open new era applica-
tions especially in the fields of Biomedical, chemical, and petroleum industries. The
viscosity of a fluid is basically a measure of how sticky it is, and for liquid, it corre-
sponds to the informal concept of thickness also called concentration. Liquid concen-
tration measurement would provide indirect clues for liquid quality monitoring for
safeguarding human health. The salinity of liquid is a very important parameter
for Mariculture, prediction of climatic changes, Oceanic, marine liquid monitoring,
drinking liquid monitoring similarly, blood glucose monitoring an essential task in
monitoring the health condition of diabetic patients [1–3]. Thus, the measurement of
viscosity has great significance for estimating liquids quality in biotechnology appli-
cations, blood protein concentration, salinity of urine, measuring concentrations of
contagious chemicals, petrochemicals, and various industrial fluids. Viscometers
which are well-established instruments in measuring the viscosity, their usage is
limited by constraints such as their voluminous size, inability for remote monitoring
applications, aging effects, and highly cost-effective [4, 5].
Several optical methods can be used to detect substance concentrations in an
aqueous solution such as absorption, refraction, or a combination of both absorp-
tion and refraction of light by the solution. Substance concentration can be inves-
tigated more accurately by asborption principle ranther than refraction principle in
the solution. Here, when a viscous liquid interacts with the fiber guiding mechanism
effects the light propagating in the fiber, which results in the change in the profile
of the sensor response in forward and reverses directions owing to the viscosity of
the liquids. Therefore, this principle of absorption often used to detect substance
concentrations.
For the last few decades, fiber optic sensors are well advanced and becoming a
substitute for conventional existing sensing technologies. Fiber optic-based sensors
have been well known for their inherent advantages like low cost, immunity to elec-
tromagnetic interferences, adaptability to other systems, fast response [6, 7]. Many
physical and chemical parameters such as displacement, pressure, high tempera-
ture, pH, and others were reported to be sensed by various fiber optic techniques.
Many special designs in fiber optic sensors are reported for the dedicated application
of liquid properties measurement [8–11]. Fiber optic sensor detection of substance
concentration by utilizing changes of the substance concentration in solution which
will change the refractive index of the solution resulted in the change of refraction
direction. This principle has been used to detect the refractive index of chlorinated
water and the concentration of the uric acid utilizing side micro bend fiber optic and
tapered fiber. Based on the displacement sensor, detection of liquid refractive index,
the concentration of liquid can be measured. Based on the displacement sensor, U-
bent tapped optical fiber can also be applied to detect the substance concentration in
a solution using the principles absorption light by liquid.
Among the many proposed methods, a U-shaped bent tapered fiber-optic intensity-
modulated sensor is one of the new methods for sensing the viscosity of the liquids
Intensity Modulated U-Shaped Bent Tapered Optical … 959
[10, 12–16]. This sensor yields a switching of its output response when it is dipped
into the liquid and taken out from the liquid. If it is closely observed, it gives a
hysteresis-like nature of response during the dipping and taking out of the liquid
process. From this, the concentration/viscosity of a given liquid can be measured. In
this paper, we have explored the hysteresis nature of tapped U shaped bent fiber optic
sensor to measure various liquid parameters like concentration, refractive index, and
viscosity from the hysteresis nature. The liquid parameters give a linear response
which can be taken as a novel way of sensing the properties.
2 Experimental Setup
A simple plastic optical fiber is used as a sensing element and it is bent to U-shape
with a calibrated radius. Before using in the experiment the diameter of the sensor
probe is calibrated for getting high sensitivity. The schematic experimental setup of
the proposed system is shown in Fig. 1. Here, one end of the fiber is coupled with
fiber-coupled LED source of 650 nm, ½ watt power, and another end of the fiber is
coupled with fiber-coupled matched silicon photodiode (PD). The output from the
PD is connected to a digital multimeter via a photodetection circuit to convert the
light intensity into an equivalent voltage signal. To enhance the sensing capability
of the sensing probe for the applied viscous liquid with various concentrations, the
cladding of the sensing area of the optical fiber is removed. The plastic optical fiber of
dimensions of around 1470/1500 nm is used in this experiment so that the thickness
of the clad is around 30 nm is very small. Clad is removed manually, enough care
is taken for the precise dimension and uniformity, and a microscope is also used to
verify the uniformity of the optical fiber. At the center of the length of POF, a small
portion of about 2 cm the clad is removed. The prepared liquid is filled in a beaker
and then it is placed under the sensor probe using a support that is holed with the
help of a micro translation stage of precision 0.01 mm.
In this experiment, Glycerin liquid is prepared to study the response of the sensor
with various concentrations for the measurement viscosity of the liquid. The Glycerin
is mixed in the distilled water from 10 to 50% in volume with an increment of 10. The
experiment is conducted with water in a forward direction (downward) that is dipped
into the liquid from a fixed position for a distance of 12 mm in steps of 0.05 mm.
After a few seconds the probe is removed in the reverse direction (upward) in the
same manner as mentioned above. During both forward and reverse directions the
intensity of light is measured through the multimeter in terms of voltage for 12 mm
distance in steps of 0.05 mm. The experiment is repeated with various concentrations
of Glycerin liquid and the sensor response is measured for the movement of the liquid
beaker for both directions.
Fig. 1 Schematic and photograph of the experimental setup
The liquid sample attached to the tip of the sensor head because of the stickiness of
the liquid owing to the viscosity. If the concentration (refractive index) of the glycerin
solution is higher, then the liquid will stay more time with the sensor end of the u-bent
fiber. As the liquid directly interacts with the light at a tapered portion of the fiber by
glycerin solution with higher concentration, causing the light refraction or scatter out
from the fiber did not return to the sensing port of the fiber, so the intensity of light (in
this case the value of the output voltage detectors) will be reduced. Thus, an increase
in glycerin concentration will lead to a change in the value of the peak voltage in
forward and reverse directions proximity measurement called intensity offset. Also,
change in the distance for change of the intensity also called distance offset. It is
clearly showing a hysteresis area between the two paths gives as a measurement of
the concentration of liquid (glycerin). The response of the sensor is tested for water
and glycerin with different concentrations by recording the sensor output with respect
to the displacement of the sensor probe as shown in Fig. 2. It is clearly showing that
owing to the viscosity of the liquid, the sensor response does not follow the same path
while forward (dipping) and reverse (rising) movement of the liquid from the sensor.
If we observe the sensor response for water, there is an offset in both coordinates of
the axis. It represents the intensity of light originates with small change and fall of
light intensity at different positions.
The offset of displacement for the glycerin liquid of various concentrations
(viscosity) has been measured is illustrated in Fig. 3a. It is observed that the sensor
follows a linear response with a linearity of about 97% but the sensitivity of the sensor
is very less. It has been observed that the response of the sensor with intensity offset
for both reverse and forward shown in Fig. 3b. The sensor shows a high linearity
response of about 98% and also exhibits high sensitivity. The hysteresis area versus
concentration of liquid is plotted as shown in Fig. 3c. It is clearly shown that the
sensor following linear response for the concentration with respect to the hysteresis
area, with a linearity of about 99% and having high sensitivity.
It is observed that intensity offset or hysteresis area measurements are more suit-
able than distance offset owing to their high linearity and sensitivity. From the
obtained results, it is illustrated that the concentration (or viscosity) of the liquid
can be measured in terms of the offset height and/or hysteresis area.
Fig. 2 Response of the sensor for rise and fall of the liquid with different % of glycerine and water
Fig. 3 Experimental results of the sensor a distance offset b height offset c hysteresis area with
respect to the viscosity (concentration) of the liquid
The viscosity of the glycerine is calculated from the following empirical formula
from the concentration of the liquid [17].

L2 L4 2
η = η0 1 + k 1 2 c + k 2 4 c + . . . (1)
a a
where η0 is the viscosity of a pure liquid, k 1 and k 2 are the constants, L is the length
of the molecular distance and a is the radius of the molecule, and c is the % of
concentration of the liquid. From the above Eq. 1, the relation between the glycerine
concentration and viscosity is drawn along with the hysteresis area measured with
respect to the concentration of the liquid. From this, it is easy to find the viscosity
of the glycerine liquid as shown in Fig. 4. The simulated viscosity results are almost
with the data given in the paper [18]. It represents the measurement of the viscosity
of the liquid in terms of hysteresis area. This sensor can also be used to measure the
refractive index of the liquid as the concentration is directly related to the refractive
index of the liquid.
Thus the proposed sensor may be used to replace many existing optical and chem-
ical viscometers because of the inherent and voluminous advantages of the optical
fibres. The sensor many find applications for measuring the viscosities of fluids such
Fig. 4 Response of the sensor for concentration and viscosity
as motor oils, petroleum products, chemical solvents, and many other industrial and
biological fluids.
4 Conclusions
The design and development of intensity-modulated plastic optical fiber (POF)

viscosity sensor is successfully reported in this paper. The sensor probe is tested
for a viscous liquid like Glycerin, by recording the interaction of evanescent field
with the ambient viscous liquid in both forward and reverse directions. An offset
is observed in the sensor response as the sensor probe dipped slowly in and out of
the liquid and measured it in terms of hysteresis loss. The offset has been measured
in both horizontal (Distance) and vertical (Intensity) directions. The experimental
results reveal that the sensor is highly linear with about 0.98 linearities. From the
results, It is observed that intensity offset measurement is more suitable than distance
offset owing to high sensitivity. Thus, the concentration (or viscosity) of the liquid
can be measured using the proposed simple sensor in terms of the offset height
and hysteresis area. This sensor is easy to make, simple, low cost, and had inherent
advantages of the optical fiber like Immune to EMI, corrosive, small size, flexibility
in length. The sensor may found many applications in measuring the viscosities of
fluids such as motor oils, petroleum products, chemical solvents, and many other
industrial and biological fluids.
References
1. Gholamzadeh B, Nabovati H (2008) Fiber optic sensors. World Acad Sci Eng Technol 42
2. Yu F, Yin S, Marcel-Dekker (2002) Fiber optic sensor
3. Krohn DA Fiber optic sensors-fundamental and applications. Instrument Society of America
4. Wang Jian-Neng et al (2010) Sensors 10:11174–11188
5. Alexander I, et al (2012) Proceedings of ISOT’12, IEEE. Intl Symp Optomechatronic Technol
6. Udd E et al (1991) Fiber optic smart structures. Wiley INC, New York
7. Dakin, Culshaw b (1988) Optical fiber sensors-principles and components, vol I, Artech, House,
Boston
8. Int J Electron Commun Comput Technol (IJECCT) 3(4) (July, 2013)
9. Jesus Castrellon-Uribe W (February, 2012) Optical fiber sensors: an overview. Center for
Research in Engineering and Applied Sciences, CIIC Ap. Autonomous University of Morelos
State, UAEM México, ISBN 978-953-307-922-6
10. Golnabi Hossein (2004) Opt Lasers Eng 41:801–812
11. Razani M, Golnabi H (2007) Iranian Phys J 1–2:61–66
12. Natha P, Singh HK, Datta P, Sarma KC (2008) Sensors Actuat A148:16–18
13. Fidanboylu K, Efendioglu HS (2009) Fiber optic sensors and their applications. In: 5th
international advanced technology symposium (IATS-09) May 13–15, Karabuk, Turkey
14. Hizhuoyin S, Ruffin PB, Francis TSY (2008) Fiber optic sensors, 2nd edn. Edited by CRC
Press ISBN 13: 978-1-4200-5365-4
15. Barry Arkles|GelestInc, Morrisville PA (October, 2008) Hydrophobicity, hydrophilicity and
silanes. Paint Coating Magz
16. Pekka Raatikainena U, Ivan Kassamakov b, Roumen Kakanakov b, Mauri Luukkala (1997)
Sensors and Actuat A 58:93–97
17. Simha R (1949) J Res Natl Bur Stand RP1981 42:409–418
18. Segur JB, Helen E (1951) Oberstar, industrial and engineering chemistry 43(9):2117–2120
Performance Enhancement
for Scattering Effect in Perovskite Solar
Cell with Distinct Cathode Materials
Sagar Bhattarai, Arvind Sharma, and T. D. Das
Abstract A comprehensive simulation study has been demonstrated to study

perovskite solar cells with imposing scattering interface between TaTm (HTL) and
methylammonium lead tri-iodide (MAPI) active layer with a distinct combination of
cathodes. The active layer acts as a light-absorbing layer with a bandgap of 1.5 eV.
Among the distinct device structures, the highest efficiency is 10.49% at active layer
width of 500 nm for Ag as a cathode. Precisely, the present simulated device is
also offering the maximum open-circuit voltage (V oc ) up to 0.83 V, short circuit
current density (J sc ) about 16.1 mA cm−2 and fill factor (FF) of 0.804, respectively,
in wavelength spanning from 380 to 780 nm. This simulation results demonstrated
an optimal result with imposing of scattering interface in the perovskite layer and it
affords several significant strategies for the feasible invention of more efficient PSC.
Keywords HTL · PCE · Active layer · V oc · J sc · FF
1 Introduction
Among the various renewable energy resources offering better device performance
and cost-effectiveness, the perovskite solar cell (PSC) is the futuristic substitution
for conventional energy resources. The recent advancement in organic-inorganic
perovskite solar cells, the methylammonium (MAPI) lead halide has revolutionized
due to their high light-harvesting characteristics [1, 2]. Apart from a great absorp-
tion characteristic, the MAPI in the PSC has gained extensive attention because of
their superior optoelectronic properties like ideal bandgap, high charge mobility,
large absorption coefficient, and small exciton binding energy, respectively [3–6].
The MAPI has also fascinated the attention due to outstanding absorbing property
which offers a very high PCE of the PSC device. To enhance the efficiency of the
device, the flexibility in its device structures is very important. Thus, we introduced
a scattering interface in the active layer which may cause to enhance the absorption
S. Bhattarai · A. Sharma · T. D. Das (B)

Department of Basic and Applied Science, National Institute of Technology Arunachal Pradesh,
Yupia, Arunachal Pradesh 791112, India
966 S. Bhattarai et al.
in the device. Carrying forward, it will also impact the stability of the device. So,
the better thoughtful knowledge in the device characteristic machinery is mandatory
to optimize efficiency [7, 8]. In that prospective Ag, cathode material offers high
device characteristics at a very cheap cost. Meanwhile, the superior electrical prop-
erty of MAPI organometal halide can be used in solid-state sensitizer as well [9]. As
it started in the year of 2009 when Kojima et al. [10] used perovskite in photovoltaic
and reported the value of efficiency rates 3.81 and 3.13% for MAPI based PSC with
different halide materials [11–13]. Since then numerous study in the field of PSC has
been carried out. Thus, the structural property behind the understanding of perovskite
solar cell is crucial which is denoted as ABX3 , where A structure is considered as an
organic-inorganic ion, B is a divalent small metallic cation, and X is a halide which
binds with cations. By varying the A, B, and X ions, optical and electrical properties
of the ABX3 perovskite structure can be tuned.
In the simulation study, the proposed perovskite solar cells structure of glass
/ITO/TaTm/MAPI/C60 /Ag or Au, where TaTm acts as P-type material, C60 as N-
type material, MAPI functions as an absorber layer, and glass used as a substrate.
Also, ITO used as an electrode to the front contact, and Ag used as the back
contact. Then the influences of incident light wavelengths and the layer thickness
of perovskite absorbers on photovoltaic parameters were predominantly investigated
by introducing a scattering layer in the MAPI layer with different cathode materials.
2 Device Simulation Parameters
The present PSC devices are simulated by three input layers; material TaTm acts as
ETL, C60 is functioning as HTL, and MAPI is an active layer. ITO is an electrode used
in the front contact and silver (Ag) is a cathode in the back contact in the simulated
device structure, which is represented in Fig. 1a. Simultaneously, the corresponding
energy levels in the corresponding layer of the device level are also represented in
Fig. 1b. Perovskite absorber absorbs photons having energy higher than 1.5 eV which
creates excitons within this photoactive layer. The parameters utilized for the present
study of MAPI based PSC are shown in Table 1, where E g is bandgap energy in eV
scale, N c and N v are called effective density of states (EDS) of the conduction band
and valence band in m−3 , µn , and µp are the electron and hole mobility in cm2 /V s
scale, respectively. The band diagram of the PSC is represented in the simulated
device structure.
2.1 Parameters Used During the Simulation of MAPI Based

PSC
The simulated PSC device structures are represented as shown below.

Performance Enhancement for Scattering Effect … 967
LIGHT -3.5 -3.5

-3.9
-4.0
ITO -4.0 -4.0
MAPbI
-4.5 -4.5
TaTm
Energy (eV)
-5.2
-5.0 -5.0
MAPI -
C 60
-5.5 -5.4 -5.5
+
TaTm
C 60
ITO
-6.0 -6.0
Ag
(a) Au -6.5
-6.6 -6.5
-6.8
-7.0 -7.0
(b)
Fig. 1 a The schematic layer of simulated PSC device with b corresponding energy band diagram
with CB and VB for the present study
Table 1 The input parameters in MAPI based PSC

Parameter Term ETM (TaTm) MAPI (CH3 NH3 PbI3 ) HTM (C60 )
d (nm) Thickness 60 500 40
E g (eV) Bandgap 1.4 1.5 2.4
N c (m−3 ) EDS in CB 1 × 1027 1 × 1027 5 × 1026
N v (m−3 ) EDS in VB 1 × 1027 1 × 1027 5 × 1026
µn (cm2 /V s) Mobility of electrons 8× 10−4 0.01 0.08
µp (cm2 /V s) Mobility of holes 8 × 10−4 1 1 × 10−5
D1(nm) ITO (120)/TaTm (50)/MAPI (without scattering) (500)/C60 (50)/Ag (100)

D2(nm) ITO (120)/TaTm (50)/MAPI (with scattering) (500)/C60 (50)/Ag (100)
D3(nm) ITO (120)/TaTm (50)/MAPI (without scattering) (500)/C60 (50)/Au (100)
D4(nm) ITO (120)/TaTm (50)/MAPI (with scattering) (500)/C60 (50)/Au (100).
In depicted Fig. 2a, profile for the absorbance describes how much light absorbs out
of the harvesting light inside the different device structures D1–D4, respectively. The
overall absorbance rate is due to consecutive layers of the device is depicted in the
figure, out of which the absorbance spectrum reaches a maximum in the absorber
layer interface and reduces at a very fast rate due to the minor diffusion span in
charge carriers. Hence, by imposing a scattering layer in the perovskite active layer,
the optical path of the incident photon can be increased which results in enhancing
the absorbance rate between the wavelength range of 380–780 nm, respectively.
0.55
1.0 1.0 (c)
(a) (b) 0.50
0.9 0.9
0.45
0.8 0.8
0.40
0.7 0.7 0.35
Transmittance
Reflectance
Absorbance
0.6 0.6 0.30
0.5 0.5 0.25
0.20
0.4 0.4
0.15 D1
0.3 0.3 D1 D2
D1 0.10
0.2 D2 0.2 D2 D3
0.05
D3 D3 D4
0.1 0.1 0.00
D4 D4
0.0 0.0 -0.05
400 500 600 700 400 500 600 700 400 500 600 700
Wavelength, nm Wavelength, nm Wavelength, nm
Fig. 2 The variation in a absorbance profile measured at different PSC structure over the wavelength
at 1.5AM solar spectral irradiance and b reflectance and c transmittance profile for the PSC devices
Similarly, in the device D2, the reflective index of the anode material Ag is smaller
than anode material Au which results in more light to transmit through the device.
Thus, the absorbance profile for device D2 increased up to 90–95% for the perovskite
solar cell, due to the triggering the optical pathway by scattering layer throughout
the spectrum region.
Figure 2b shows the reflectance spanning over the wavelength for differently struc-
tured perovskite solar cell structure D1–D4. It can be noted that the less reflectance
results in more effective perovskite solar cell device. Taking into account, in the
present study D2 achieved lesser reflectance about 10% since the active layer contains
the scattering interface which reduces the reflectance in the active layer influencing
the incident light to enter through the device. In D1 and D3 structures, since the
optical path is lesser than D2 and D4, it causes more reflectance up to 20 and 30%,
respectively, in the PSC. Precisely in D3, the Au layer has a more refractive index
than Ag in D1, for which reflectance is more in D3 than D1.
Figure 2c gives the transmittance of the perovskite solar cell for D1–D4 device
structures. As more absorbance results in less transmittance rate of the incident light
decrease with the scattering layer since it traps more photons in respective active
layers of the device. Resulting in decreasing the transmittance in the other layers of
the perovskite solar cell. For a 500 nm thick active layer in the perovskite solar cell,
we achieved maximum absorbance and minimum reflectance and transmittance for
the D2 structure.
The simulated photovoltaic characteristics (J-V curve) are represented through
Fig. 3. In the device structure D2, due to imposing a scattering layer in the perovskite
active layer gives the best outcome of J sc and FF, respectively. Since J sc is propor-
tional to the field intensity, the active layer which imitates that J sc may be moder-
ated from imposing the scattering layer which enhances the light absorption due to
attaining the longer optical path. The lower mobility of charge carriers, regardless
of higher absorbance, J sc and V oc value increases for thicker perovskite layer at an
externally applied voltage of 1 V. Therefore, in our device structure, we used an opti-
mized thickness in the PSC active layer of 500 nm as reported by Adhikari et al. [14].
Figure 3 shows the dependency of J sc and V oc on device structures D1, D2 under
the illumination of AM1.5. Whereas, the PCE of the device in different perovskite
Fig. 3 The photovoltaic 10

characteristics of J-V D1
D2
characteristics at an applied D3
5
voltage of 1 V across the D4
junction
Current (mAcm-2)
0
-5
-10
-15
-20
0.0 0.2 0.4 0.6 0.8
Applied Voltage, V
solar structures D1–D4 is calculated, respectively, as shown in Fig. 4. For device D2

of MAPI based perovskite solar cell, we get maximum FF and PCE about 10.49%.
As the above figure implies, V oc , J sc , and FF of the PSC improved by introducing
the scattering interface in the active layer since escalating the layer thickness absorbs
photons with longer wavelengths. As the increment in the device parameters for D2
improves the conversion of input power at 1 V of potential. Usually, the results in the
generation of excitons further improve the electrical parameters of the PSC device.
This relates to more generations in carriers and further optimizes in enhancing device
efficiency.
The comparison of different structured PSC structure is given below.
Fig. 4 The power 0.12

conversion efficiency of PSC D1
D2
structure at AM1.5
power conversion efficiency
0.10 D4
illumination and applied D3
voltage 1 V in the PSC
0.08
0.06
0.04
0.02
0.00
0.0 0.2 0.4 0.6 0.8
applied voltage, V
16.0
(a) 0.830 (b)

15.5
0.829
15.0
0.828
14.5
Jsc
Voc
0.827
14.0
13.5 0.826
13.0 0.825
D2 D1 D4 D3 D2 D1 D4 D3
0.11
(c) 0.834
(d)
0.832
0.10 0.830
Fill factor
0.828
PCE
0.826
0.09
0.824
0.822
0.08 0.820
D2 D4 D1 D3
D4 D3 D2 D1
Fig. 5 The simulated result comparison of the different parameters of the MAPI based solar cell
In Fig. 5a, the comparative statement for the different device structure D1–D4 are
shown at the AM1.5 illuminance. Current density (J sc ) for the device D2 is depicted
to be the maximum about 15.7995 mA cm−2 and open-circuit voltage (V oc ) of 0.8298
volts is also obtained in the study. Similarly, the D2 device offered optimum power
conversion efficiency of 10.49%. Since the scattering effect enhances the voltage
and current for higher absorbance in the active layer region triggering the generation
of excitons in the simulation study of PSC since a higher generation of excitons
enhances the current and voltage in the PSC device [15]. In D2, the effect of more
voltage production leads to improving the FF up to 0.834.
In summary, the study of effect due to the scattering layer interface of perovskite
solar cell which is introduced between the HTL and MAPI layer with distinct cathode
combinations is carried out using a commercially available software package. It
observed that the power conversion efficiency is higher for the perovskite layer with
500 nm thickness. It is also evident from this investigation that with the optimum
perovskite layer thickness, the device parameters like open-circuit voltage (V oc ) about
0.8298 V, short circuit current (J sc ) as 15.7995 mA cm−2 , and fill factor (FF) nearly
0.8035, with PCE reaching significantly high up to 10.49%, respectively, corre-
sponding to the incident light with a wavelength range 350 nm - 800 nm. This result
provides pathways for future improvement of PSC with a better suitable optimization
mechanism.
Acknowledgements The authors thank the third phase of the Technical Education Quality
Improvement Programme (TEQIP-III) under seed grant, NIT Arunachal Pradesh, for financial
support. Author Sagar Bhattarai would like to express his deep and sincere gratitude to TEQIP-III for
providing fellowship to carry out his research work.
References
1. Wu HQ, Xue C, Li Y, Zhou P, Liu W, Zhu J, Dai S, Zhu C, Yang S (2019) Simulation of
perovskite solar cells by using CuSCN as an inorganic hole-transport material. Mater Res
Express 095527
2. Adhikari KR, Gurung S, Bhattarai BK, Soucase BM (2016) Comparative study on MAPbI3
based solar cells using different electron-transporting materials. Physica Status Solidi (c),
13:13–17
3. Brabec CJ, Cravino A, Meissner D, Sariciftci NS, Fromherz T, Rispens MT, Sanchez L,
Hummelen JC (2001) Origin of the open-circuit voltage of plastic solar cells. Adv Funct
Mater 11(5):374–380
4. Tang CW (1986) Two-layer organic photovoltaic cell. Appl Phys Lett 48(2):183–185
5. Brabec (2013) Concepts and realization, vol 60. Springer
6. Granström M, Petritsch K, Arias AC, Lux A, Andersson MR, Friend RH (1998) Laminated
fabrication of polymeric photovoltaic diodes. Nature 395(6699):257–260
7. Zhiliang K, Rong Y, Mi X, Liu T, Han H (2013) Full printable processed mesoscopic CH3 NH3
PbI3 /TiO2 heterojunction solar cells with a carbon counter electrode. Sci Rep 3:3132
8. Shi J, Dong J, Lv S, Yuzhuan X, Zhu L, Xiao J, Xin X, Huijue W, Li D, Luo Y, Meng Q (2014)
Hole-conductor-free perovskite organic lead iodide heterojunction thin-film solar cells: high
efficiency and junction property. Appl Phys Lett 104(6):063901
9. Burschka J1, Pellet N, Moon SJ, Humphry-Baker R, Gao P, Nazeeruddin MK, Grätzel M (2013)
Sequential deposition as a route to high-performance perovskite-sensitized solar cells. Nature
499(7458):316–319
10. Kojima A, Teshima K, Shirai Y, Miyasaka T (2009) Organometal halide perovskites as visible-
light sensitizers for photovoltaic cells. J Am Chem Soc 131(17):6050–6051
11. Torabi N, Behjat A, Shahpari M, Edalati S (2015) Development of a silver/polymer
nanocomposite interconnection layer for organic tandem solar cells. J Nanophon 9(1):093049
12. Aharon S, Gamliel S, El Cohena B, Etgar L (2014) Depletion region effect of highly efficient
hole conductor free CH3 NH3 PbI3 perovskite solar cells. Phys Chem Chem Phys 16(22):10512–
10518
13. Tress W (2014) Photovoltaic energy conversion in organic solar cells. Springer, Cham, pp
15–65
14. Adhikary P, Venkatesan S, Adhikari N, Maharjan PP, Chenb OAJ, Qiao Q (2013) Enhanced
charge transport and photovoltaic performance of PBDTTT-CT/PC 70 BM solar cells via
UV–ozone treatment. Nanoscale 5(20):10007–10013
15. Zhou D, Tiantian Zhou Y, Tian XZ, Yafang T (2018) Perovskite-based solar cells: materials,
methods, and future perspectives. J Nanomater 2018:8148072
Need of Techno-Preneurship
and Innovations for Community-Level
Cancer Screening in Arunachal Pradesh
M. S. Momocha and K. Chaitan
Abstract It is estimated that noncommunicable diseases cause 71% of all death

in the world (World Health Organization in India: first to adapt the global moni-
toring framework on noncommunicable diseases (NCDs). World Health Organization
(WHO), Geneva, 2015, [1]). Key noncommunicable diseases are cancer, cardiovas-
cular diseases, and respiratory-related. 61.8% of death in India is due to NCDs.
NFHS 4 data shows one-third population faces catastrophic health expenditures
and faced distress financing mainly because of hospitalization (Key indicators of
social consumption in India: health. NSS 71st round. GoI, Ministry of Statistics and
Programme Implementation, New Delhi, 2014, [2]). Out of Pocket Expenditures on
hospitalization especially on cancer is highest in India. State diseases burden study
2016, highlighted that in Arunachal Pradesh premature death was 63.7% (SDBS,
2016) and NCDs causing 52.9% death (Indian Council of Medical Research in A
report on the cancer burden in North East India, 2012–14, [3]). ICMR 2012–14
reported that in per 100,000 female population Papumpare district of Arunachal
Pradesh had the highest Cancer incident (249.0) in the Population-Based Cancer
Registration (PBCR) in India. Whereas In per 100,000 male population Papumpare
district of Arunachal Pradesh had the second-highest incident of Cancer (230.4) after
Aizawl (270.7) in Mizoram (Indian Council of Medical Research in A report on the
cancer burden in North East India, 2012–14, [3]). The footfall for NCD screening in
the public health system in Arunachal Pradesh during April 2018–Sept 2019 reported
7003 and highest screening was on Oral Cancer (2875) (Ayushman Bharat Health and
Wellness Centers (ABHWC) in first 18th months of implementation (April 2018–
Sept 2019), [4]). The screening demand generation can be increased if more afford-
able and accessible technology-based screening services are provided at community-
level especially at doorstep along with proper behavior change communications on
the risk factors. The paper tried to investigate the Cancer burden and screening
services in Arunachal Pradesh and had also objective to oversee the community-level
M. S. Momocha · K. Chaitan (B)

Department of Management & Humanities, National Institute of Technology Arunachal Pradesh,
Itanagar, India
M. S. Momocha
974 M. S. Momocha and K. Chaitan
demand and supply gap of screening services in Arunachal Pradesh. Secondary data
analyzed to meet these objectives. There is high demand for tech-entrepreneurship
to meet the demand for screening services of cancer diseases in the community.
Keywords Community cancer screening · Techno-entrepreneurship · Innovations
1 Introduction
Noncommunicable disease (NCD) are chronic diseases mainly cancer, chronic respi-
ratory disease, cardiovascular disease, diabetes, and injuries, and mental health. 80%
NCDs are related to lifestyles and risky behaviors like tobacco use in any form,
alcohol consumption, inappropriate diet, and physical inactivity which cause physi-
ological changes in the body manifested as high blood pressure, high blood glucose
increased, raised body mass index, and waist circumference. The behaviors are devel-
oped in various socio-cultural conditions. Study shows in India NCDs are raising at
alarming rate and huge economic loses are expected which will hamper the overall
development goals if timely preventive majors are not taken. If we achieve a 2%
reduction in these diseases our economic growth can increase by one percent. ICMR
Estimates that 2.25 million people live with Cancer in India. 1,157,294 new patient
registers every year. 413,519 men and 371,302 women died of Cancer in 2018 [3].
Government of India health insurance schemes ABPMJAY data (Sept 2018–July
2019) reflects that 34% of tertiary care claims are related to Oncology [5]. The
ABPMJAY schemes are designed for approx. 40% poorest population in India.
Seeing the national burden and vulnerability in Arunachal Pradesh the researcher
adopted a secondary literature review method to analyze the supply and demand
situation of screening services of Cancer diseases in Arunachal Pradesh. Arunachal
Pradesh has 98 Operational AB Health and Wellness Centers (Sept 2019) which
provided screening of noncommunicable diseases [4]. The state has a population of
13.83 Lakh (Census 2011) and the screening population on NCDs shows 7003 for
the period April 2018–Sept 2019 [4]. The screening is being done at HWCs through
trained Lab technicians and tablet-based technology is used for MIS data generation
and reporting. Tech-entrepreneurship and innovation for doorstep services will help
fill the gap in the shortage of screening services.
2 Methodology
The study was done with a literature review, analysis of health department data,
reports of research institutions, Government of India Policy documents, reports of
Ministry of Health and Welfare were consulted.
Need of Techno-Preneurship and Innovations for Community-Level … 975
3 Key Findings
3.1 Risk Factors/Life Style Diseases
WHO recognized key risk factors on NCD are alcohol use, insufficient physical
activity, excess salt intakes, tobacco smoking, raised blood pressure, raised blood
glucose, obesity, and air pollution. Hence, NCDs are also called lifestyle diseases.
3.2 Global Disease Burden
World Health Assembly 2013 adopted a global action plan for the prevention and
control of NCDs 2013–2020 and recommended member countries to apply 25
comprehensive indicators for positive outcomes of the programs [6].
NCDs cause 71% death worldwide and estimated that 52 million death might
occur by 2030 [1]. 50% such death is in lower and middle-income countries. Death
due to NCDs is an obstacle for other sustainable development goals and poverty
reduction.
The family having NCD patient has a double burden on one side health care cost
and on the other had wage loss, time, and income of caring members.
Catastrophic health spending and distress health financing are the two ways
economic study of the diseases found in the various literature.
Sustainable development goals targets to reduce one-third premature mortality
from noncommunicable diseases through prevention and treatment and promote
mental health and wellbeing by 2030.
3.3 Indian Situation
NCDs in India is an increasing trend. 62% of hospitalization is in private hospitals.

Fifty percentages of the family having Cancer patient has to suffer catastrophic
spending and 25% of such family fall into poverty [2].
The report of the Global Adult Tobacco Survey (GATS) says hat that 28.6% of
adults consume tobacco in various form, it is estimated that 266.8 million tobacco
users are in India [7].
3.4 Government of India Initiative to Reduce Out of Pocket

Expenditures
Government of India’s Ayushman Bharat Pradhan Mantri Jan Arogya Yojana an

insurance scheme, covering 40% of the poorest and most vulnerable individuals in
the country for secondary and tertiary care.
On the other hand, 150,000 Ayushman Bharat Health Wellness Centers (AB
HWCs) are planned for primary care services including screening of NCDs. NCD
screening has been planned through these HWCs including outreach activities
[8]. This transformation is planned to facilities delivering high-quality, efficient,
equitable, and comprehensive care.
ICMR Estimates that 2.25 million people live with Cancer in India. 1,157,294
new patient registers every year. 413,519 men and 371,302 women died of Cancer
in 2018.
Government of India health insurance schemes ABPMJAY data (Sept 2018–July
2019) reflects that 34% of tertiary care claims are related to Oncology [5]. The
ABPMJAY schemes are designed for the poorest approx. 40% poorest population.
3.5 Vulnerability in Arunachal Pradesh
State diseases burden study 2016, highlighted that in Arunachal Pradesh premature
death was 63.7% and among premature death, NCDs caused 52.9% [9].
Risk Factors and their contribution to NCDs are highlighted in Table 1.
ICMR report 2017 highlighted the following figures for Arunachal Pradesh for
the Reporting Year 2012–2014.
Table 1 NCD risk factors

Risk factor Percentage of contribution in
death (%)
Malnutrition 14.80
Air pollution 5.1
High blood pressure 5.00
Tobacco use 4.9
Dietary risks 4.7
High fasting plasma glucose 4.00
Wash 3.9
Alcohol and drug use 3.90
Occupational risks 2.6
Impaired kidney function 2.0
In per 100,000 male population Papumpare (230.4) district of Arunachal had the
second-highest incident of Cancer after Aizawl (270.7) in Mizoram in the Population-
Based Cancer Registration (PBCR) in India.
Whereas in per 100,000 female population Papumpare (249.0) district of
Arunachal had the highest incident in the Population-Based Cancer Registration
(PBCR) in India.
The report highlighted a total of 1773 (910 male and 863 female) cancer patients
registered out of which total 406 (251 Male and 155) death cases recorded [3]. This
shows that approx. 23% of Cancer patients could not survive in the reporting period.
The report also highlighted that 61% of Cancer incidents are mostly in four
tribes—Nishi 33%, Galo 14%, Monpa 7%, and Apatani 7% [3].
National Family Health Survey (2015–16) says that in Arunachal Pradesh 60%
male and 17% female population consume tobacco in the various form [2].
59% of male consumes alcohol. 18–20% population has overweight (male—20%,
female—18%), 5.9–8.5% female does screening for breast or Cervical Cancer [2].
Male has high Stomach and liver Cancer whereas the female has high cervical and
breast cancer in the state in comparison with the rest of India. The cancer incident
and death ratios from 2014 to 2016 has been estimation has been highlighted in Fig. 1
[10].
It shows that the State has Low survival chances of Cancer cases. The death
percentage with respect to reported cases during 2014–2016 in Arunachal Pradesh
and National average has been shown in Fig. 2.
Medical investigations reported that more than fifty percent of male Cancers and
more than twenty-five percent female cases are due to the use of tobacco.
Fig. 1 Cancer incidence versus death in Arunachal Pradesh

Fig. 2 Death percentage India versus Arunachal Pradesh
3.6 Cancer Screening Situation in the Public Health System

in Arunachal Pradesh
During the period April 2018–Sept 2019 total footfall of screening on NCDs at
public health facilities namely Ayushman Bharat Health and Wellness Centers (AB
HWCs) in Arunachal Pradesh reported as 7003 [4]. The NCDs are chronical diseases
and screening required multiple times. Only three Cancer namely cervical, oral, and
breast cancer is being screened across functional health and wellness Centers. The
summary of the analysis of demand and supply of screening services is shown in
Table 2. An Analysis of Cancer and other diseases screening are shown in Fig. 3.
There is a high gap in the demand and supply of screening services in the
population. The NCDs screening gaps analysis is shown in Fig. 4.
The state having 1,383,727 population (Census 2011) and high disease prevalence
required more innovation and technological solutions for grass root level screening
for Cancers and other NCDs. The interaction with officials of the health departments
highlighted the vulnerability and risk factors to be known to the community and
health workers. Geographic challenges due to the hilly state and socio-economic
characteristics of the community lead to high-risk factors. Community behavior
change is a hard task to avoid risk factors in the state. The screening will help to
early detection of the diseases and precaution from further risk factors.
Table 2 Population screened on NCD diseases during April 2018–September 2019

Diseases Screened Suggested rate Adult population % of population
population (April from the national above 30 years to having a shortage
2018–Sept 2019) programme be screened based of screening at
on 2011 census HWC
population
Hypertension 508 37% annually 511,979 99.90
Diabetes 194 37% annually 511,979 99.96
Breast cancer 1442 51% of adult 261,109 99.45
population in
5 years
Survical cancer 215 51% of adult 261,109 99.92
population in
5 years
Oral cancer 2875 37% annually 261,109 98.90
Total foot fall 7003 Total foot fall 511,979 98.63
estimated
Fig. 3 Diseseas wise

screening percentages
Fig. 4 Gap in NCD screening services in the community
4 Conclusion
NCD prevention framework suggests that multisectoral engagement is required along

with community knowledge, attitude, and practice changes to prevent NCD in the
community. The data reflects that the risk factors for Cancer burden are high in the
state and massive population screening is required at the community-level.
The techno-entrepreneurship for Cancer screening is highly required for afford-
able and accessible services to fill the shortage of public health Cancer screening
system. Further detailed field study is required to find out effective, affordable, and
innovation screening solutions for the maximum adult population in the state with
geographic and socio-cultural context. Cause for lifestyle behavioral change to be
studied for future solutions for reducing risky behaviors in the population. Screening
services at private hospitals are to be integrated for policy-level decisions. Current
screening policy in the public sector is mainly focused on three high prevalence
Cancers. There might be a demand for other Cancer site screening as well.
References
1. World Health Organization (2015) India: first to adapt the global monitoring framework on
noncommunicable diseases (NCDs). World Health Organization (WHO), Geneva
2. Key indicators of social consumption in India: health (2014) NSS 71st round, Jan–June 2014.
GoI, Ministry of Statistics and Programme Implementation, New Delhi
3. Indian Council of Medical Research. A report on cancer burden in North East India, 2012–14
4. Ayushman Bharat Health and Wellness Centers (ABHWC) in first 18th months of implementa-
tion (Apr 2018–Sept 2019). https://abhwc.nhp.gov.in/download/document/340b49eb2c0937e
7b79ad8c1d6b975ad.pdf. Accessed 18 Jan 2019
5. Can Ayushman Bharat help to tackle cancer burden in India? https://www.livemint.com/
science/health/can-ayushman-bharat-help-tackle-Cancer-burden-in-india-11576396813549.
html. Accessed 18 Jan 2020
6. WHO (2013) Global action plan for the prevention and control of noncommunicable diseases
2013–2020. WHO, Geneva
7. Global adult tobacco survey India fact sheet 2016–17
8. IHME Report (2017) Comprehensive health study in India finds rise of non-communicable
diseases. http://www.healthdata.org/news-release/comprehensive-health-study-india-finds-
rise-non-communicable-diseases
9. India State-Level Disease Burden Initiative Collaborators (2017) Nations within a nation:
variations in epidemiological transition across the states of India, 1990–2016 in the Global
Burden of Disease Study. Lancet 390:2437–2460. https://doi.org/10.1016/S0140-6736(17)328
04-0
10. Press Information Bureau, Government of India, 18 Mar 2018
Phrase-Based Machine Translation
of Digaru-English
Rushanti Kri and Koj Sambyo
Abstract The Natural Language Processing (NLP) is mainly associated with the
interactivity of computers and natural languages in order to process and analyze
natural language data. The machine translation is a subdomain of the NLP which
helps us settle language inconceivable issues. The high resourced language such
as the European language has been widely studied and used within the discipline of
computational linguistics and for Machine Translation whereas several low resourced
language such as Digaru spoken by the Tawra Mishmi Tribe of Arunachal Pradesh
has yet to be introduced in the Machine Translation field of NLP due to unavailability
of data in this data-driven world. In this paper, we have provided freshly collected
Digaru-English translation of about 5057 aligned sentence pairs and measured the
translation accuracy of the freshly collected data through Statistical Approach of the
Machine Translation and provided the unigram score 0.4063 calculated by the BLEU
metrics.
Keywords Natural language processing · Statistical machine translation · BLEU

score · METEOR · Human evaluation (5–6)
1 Introduction
The Natural Language Processing is an approach based mainly on the set of theories
and technologies for analyzing texts. The Machine Translation being a subfield under
the NLP has in the past few years garnered the attention of many researchers as well as
the public due to the increasing interactions of people from different linguistic back-
grounds from across the Globe. This paper is an attempt to provide a comprehensible
translation of the Digaru Language used by the Tawra Tribe which is one among the
R. Kri (B) · K. Sambyo

Department of Computer Science and Engineering, National Institute of Technology Arunachal
Pradesh, Yupia, India
K. Sambyo
984 R. Kri and K. Sambyo
three of the Mishmi group residing mainly in the eastern side of Arunachal Pradesh
with as few as 35,000 native speakers as per the 2001 census. The Digaru language
is also considered among the family tree of perhaps the Sino-Tibetian language
and also shares the linguistic relationship with the Idu which also belongs to the
Mishmi group, they may share few words in common but their linguistic connection
is quite puzzling [1]. While many highly resourced languages which have a rich
collection of data benefits in providing almost perfect translations from a trained
system to a human translator, the low resourced language such as Digaru might in
further years disappear if it is not preserved now. Through the computational and
linguistic perspective, the structure of Digaru language does not permit us to create
a very long sentence due to its unnatural structure as it gets claque from the Hindi
language, also due to being a highly tonal language and also due to the involvement
of several homographs it gets complicated to write it down in the simpler form [2].
The collection of a completely new corpus and applying several computations on
the corpus will provide us with the performance metric of several Machine Trans-
lation model for a completely new corpus as well as preserve the corpus of a low
resource language such as Digaru. Machine Translation has been present for many
years but gained attention only recently. Several Machine Translation approaches
have been proposed which includes the RBMT which stands for the Rule-Based
Machine Translation, SMT for Statistical Machine Translation, and NMT for the
Neural Machine Translation. While many research works are being done based on
the NMT approach which is the latest among the three, here in this paper we will
use the Statistical Approach for our Corpus with the help of MOSES an opensource
Statistical Machine Translation Toolkit and evaluate the translation accuracy that the
SMT provides for a completely new corpus with the help of the BLEU, the Bilin-
gual Evaluation Understudy which is an algorithm for calculating and judging the
translated text quality of the machine-translated corpus from Digaru-English.
2 Related Works
The natural language processing was introduced in the late 1940s where MT was the
first computer-based application developed in order to achieve human-like language
processing [3]. The Digaru-English corpus is first among the three Mishmi languages
which have been introduced to machine translation through this paper. Since it is a
completely new corpus there is no history of computations therefore we have selected
the phrase-based MT system which lays its priority on meaning preservation as it aims
to provide adequate translation [4]. The meaning preservation of the phrase-based
MT system is preferable for a low resource language like Digaru as the focus tends
towards the meaning instead of the fluency parameters [4]. The SMT system assigns
a probability to all set of sentences as it needs several ways to compute the language
model probability as well as the translation model probability and also searching for
the best source sentence which will provide the highest value [5]. The SMT system
makes use of n-gram model to find out the highest probability and rank the words
Phrase-Based Machine Translation of Digaru-English 985
from the bag of translation where every sentence is chopped into words and every
word count provides probability measures for a translation output during decoding
[6]. The SMT system uses alignment models such as giza++, MGIZA to provide
translation correspondences between a source sentence and its reference translation
[7]. The MOSES toolkit has been widely used for SMT by several researchers as well
as the industry people to provide faster and reliable translation through a statistical
approach [8]. The working order of Moses toolkit starts by first installing all the
supporting tools like the MGIZA and mkcls, SRILM, IRSTLM, etc. which are the
main building block used during the pre-processing, language and translation model
training through Moses and similar translation work has been provided in Ref. [9]
where they have given detailed information about their approach to statistical machine
translation for their corpus using Moses. The trained models providing the predicted
translations are calculated by an automatic metric BLEU which analyzes and counts
the n-grams to rank the translated output based on its precision [10].
3 The Statistical Advancement Towards Machine

Translation
The main pretension of the machine translation is to provide translation between two
languages basically input and an output language. In the past few years Machine
Translation has shifted from the Classical approach, i.e., RBMT to the Statistical
approach towards machine translation, where the Rule-Based Translation offers a
reasonable translation accuracy but is exhausting during the pre-designing set of
translation rules and generation based on the form of language structure, according
to the syntax of the language and the meaning behind the language of both the source
as well as the target language [6]. The SMT creates translations with the underlying
support of the statistical models where the parameters are obtained from the analysis
of bilingual text corpora [11]. A statistical model does not rely on language features
such as the grammar and the language-specific models of translation which provides
an advantage over the Rule-Based Translation system and also allows us to build
multiple language pairs with minimal modification. Statistical model can be trained
on small as well as large data set and increasing the dataset will eventually lead to
higher quality translation and capture more of the linguistic aspect of the language.
There are more approaches to machine translation including the Example-Based
Machine Translation system and the Neural Machine Translation system which
are also corpus-based machine translation system but we have used the statistical
approach as it makes use of the predictive algorithms in order to train a translation
model that formulates the translation problem with regard to the conditional prob-
ability. The SMT model requires methods to compute translation probabilities and
also for computing language model probabilities. The implementation of the Bayes
Theorem provides with better modelling of the probability distribution p(e|f) where
e in the target language is the translation of string f in the source language. The Bayes
Theorem, p(e|f) α p(f|e)p(e), where the translation model p(f |e) is the probability
that the source string is the translation of target string, and language model p(e) is
the probability of seeing that target language string [7].
4 System Description
4.1 Corpus Preparation
The Corpus preparation of the Digaru-English [12] parallel corpus includes Tokeniza-
tion which is very helpful for Statistical Machine Translation (SMT) as it divides
a sentence into tokens which are separated by spaces, Truecasing which restores
case information to badly-cased or non-cased text and avoids irrelevant difference
between lowercased and uppercased words and Cleaned which removes long and
empty sentences in order to formulate it and make the corpus ready for training,
tuning as well as testing purpose.
4.2 The Language Model
To build a reliable language model for SMT system there is the need for time and
human cost which results in several difficulties. The language model in Statistical
Machine Translation helps with reordering of words and also helps with the word
choice which requires less time and human cost compared to the human evaluation
of a language model. Also, the prediction is always difficult, especially about the
future. In our case, it is the prediction of the next word where assigning probabilities
to a sequence of words is essential in machine translation. Language models can be
defined as the model that allocates probabilities to a sequence of the word and the
n-gram is a model that allocates probabilities to a sequence of word and sentences.
The n-gram is a series of n words whereas the bigram or 2-gram is a two-sequence
of word, trigram or 3-gram is a three-sequence of word and so on. The main instinct
of n-gram model is to approximate the record by just the last few words instead
of calculating the probability of a word given its entire record in the history. The
presumption that the probability of a word depends exclusively on a previous word
is known as Markov model, we can generalize the bigram to the trigram and so on
to the n-gram.
We can calculate the unigram probabilities p(w1) = f(w1)/N where f(w1) is the
number of times w1 is seen in some corpus and N is the total number of words seen in
the corpus (by token). The relative frequency estimation, in this case, can be shown
to be an instance of maximum likelihood estimation which is an intuitive way to
estimate probabilities.
Fig. 1 Word alignment with Muwa aa cha naara

MGIZA [7]
translation step
He is sick
Alignment step
He is sick
An open-source machine translation project with very few resources such as our
Digaru-English corpus always restricts the quality of the parallel corpus. Word align-
ment plays an important role in SMT system which can be obtained by increasing
language model training data.
4.3 The Translation Model
The translation model is also called as the models of Translation equivalence because
they aim to predict whether expressions in different languages have equivalent
meanings.
We train a translation model on sentences that have the same definition in each
language and the translation quality depends on the size of dataset, i.e., high quantity
of the dataset will give better translation quality. The first step to build a translation
model is by building an alignment model between the input and the output language
[7]. The alignment models provide the best set of alignment links for each sentence
in the training corpus. The MGIZA from the Moses toolkit has been used to align our
Digaru-English training corpus. Alignment models are a significant part of machine
translation as a word-aligned text is mandatory in order to initialize a translation
model. The main objective of alignment is to provide an aligned model resulting in
the best translation for a particular word or a sentence adding with the fact that it is
difficult to corresponds a given word in target sentence to word in a source sentence.
We refer to a picture such as that shown in Fig. 1 as an alignment.
4.4 Decoding
The output of the Language model and the output of the Translation model as well as
the source sentence is the input for the decoder. The Decoder implements a search for
a good translation candidate and tries to find the hypothesis with the highest model
score. The computational complexity is a large problem in the decoding process. The
inputs may translate in different ways as the search space can become quite large due
Table 1 Corpus description

Corpus type No. of instances
for Digaru-English and
German-English parallel Training data 3246
corpus Tuning data 811
Testing data 1000
to a few reasons such as output word which does not provide direct translation of
either word in an input sentence. Decoder customarily utilizes the dynamic program-
ming algorithm and potent search methods like the beam search strategy which is a
heuristic-based algorithm for reducing the search space which has a set of possible
hypotheses during translation predicted by the combined attempt of the language and
translation model.
5 Experimental Design
We have divided the Digaru-English corpus into training, testing and tuning files
and also provided a monolingual file for building the language model with English
sentences.
The Moses toolkit builds the language model with IRSTLM which promotes
algorithm and data structures fit for storing and accessing large n-gram language
models, the same is also used for building the recasing model [13]. The language
model parameters for our corpus consists of 1 language model builder with IRSTLM
= 1, RandLM = 5 where the RandLM uses Bloom Filter for testing and checking
whether an element is a member of the set. And the IRSTLM parameters consist of
1 distributed LM training which splits the dictionary into balanced n-gram lists, 20
dictionary number parts, improved-kensor-ney smoothing technique, and 1 memory
mapping. The training corpus of Digaru as well as English is trained along with the
language model with mkcls and MGIZA for word alignment on 4 cores of the CPU
described in Table 1.
6 Result and Analysis
6.1 Scoring Phrases with BLEU/NIST Metrics
The BLEU/NIST score is computed based on the precision of n-grams. Both metrics
require a test-set for machine translation system. The Translation from a source
sentence or reference sentence to a target sentence or candidate sentence depends on
the word orderings and word choices and the results can vary ranging from a perfect
translation to a totally unrelated translation [10]. The BLEU metric ranges from 0 to
Table 2 Unigram score for

No. Corpus type NIST score BLEU score
Digaru-English and
German-English corpus 1 Digaru-English 2.3982 0.4063
2 German-English 4.0430 0.6793
1 and the equation for calculating the BLEU score is,
n n1
candidate-length
min 1, precisioni
reference-length i=1
where BLEU/NIST uses the n-gram precision approach to compute the resemblance
between the reference translation and the candidate translation, and if a match is found
between the reference and the candidate translation then the n-gram of reference
exhausts for next match. The candidate-length and reference-length from the above
equation denote the length of the candidate and the length of the reference translation,
respectively and precision score is calculated for the ith gram matching [4].
We have calculated the unigram BLEU/NIST score of our relatively small Digaru-
English parallel corpus as well as for the German-English corpus consisting equal
number of instances with the MOSES an open-source machine translation toolkit.
The unigram score for both corpora is provided in Table 2.
The score of Digaru-English corpus is low compared to its German-English coun-
terpart mainly due to the high word ambiguity of the Digaru language which directly
affects the n-gram count of the Digaru-English corpus.
6.2 About the Automatic Evaluation Metrics: METEOR
METEOR which stands for Metric for Evaluation of Translation with Explicit
Ordering is an automatic scoring metrics that generates scores of a translation by
calculating the alignments with respect to its exact word, synonym, stemmed word
matching and paraphrase matching of the word and sentences. The above three
modules provide alignment between the reference and the predicted translation.
METEOR provides additional features like the synonym matching, stemming and
exact word match which is not a part of the BLEU metrics [14]. METEOR makes
use of the precision as well as the recall, unlike the BLEU and NIST metrics which
only depends on the precision. Implementation of the precision along with recall
provides higher association with the human judgement at sentence level [14]. Let us
consider a few notations,
rt No. of unigrams in reference translation
pt No. of unigrams in predicted translation
mt No. of matching unigrams between rt and pt.
The unigram precision can be obtained by the formula P = mt/pt where P stands
for the precision.
The unigram recall can be obtained by, R = mt/rt where R stands for recall.
And, F-measure which is generated during the computation of METEOR score
gives the mean of P and R and can be obtained using, F-measure = 2 * P * R/P +
R [4].
6.3 Human Evaluation
The translation quality of output is basically judged relative to its reference translation
and the criteria of the judgment are based on its fluency and adequacy where fluency
provides the quality of a language and adequacy provides the measurement of the
meaning of the reference translation with respect to their hypothesis translation. The
human evaluation is, however, quite expensive because of the time constraint and the
financial expense. Linguistic expert in our experiment is my mother and a cousin as
they are very well-versed in Digaru Language and is a native of the Tawra Tribe of
Arunachal Pradesh. Further, on our paper, we have performed comparative analysis
on a few samples from Digaru-English Translation and examined their quality from
different perspectives.
6.4 Analysis of Digaru-English Translation
To further examine our translated output from the SMT system we will use adequacy
and fluency as the two main factors to judge the quality of our predicted output. We
have taken a few sample sentences from predictions and examined the same for
their quality from a different perspective. We have taken the Digaru sentences as
the source sentence and on that basis; we will examine the translation quality of the
resultant English translation which is the predicted output by MOSES with the actual
English translation of the Digaru sentences. Digaru Test (DT) will stand for the test
sentence, English Gold (EG) will stand for the reference sentence in English and
English Predicted (EP) for the predicted sentence in English.
1. DT: “tya jina”
EG: “Do it now”
EP: “Do it now”
The SMT system has made a perfect prediction for the above sentence and the
predicted translation for the Digaru sentence is adequate and fluent.
2. DT: “boya”
EG: “Don’t go”
EP: “go”
The SMT system has made prediction which has predicted the unigram “go” and
therefore it is not adequate but is fluent.
3. DT: “Tom kasa”
EG: “Tom knew”
EP: “Tom knows”
The SMT system has made prediction which is partially adequate and is perfectly
fluent.
4. DT: “haa halyo ya kanum jidi”
EG: “I just want to be happy”
EP: “I don’t happy a about”
The SMT system has made prediction which is neither adequate nor fluent.
5. DT: “tom naara”
EG: “tom’s sick”
EP: “tom is sick”
The SMT system has made prediction which is adequate as well as fluent but here
the SMT system automatically interrogates and adds ‘is’ to the predicted sentence.
From the above analysis of the SMT system, we can conclude that the predicted
translation gives relatively good translation for a completely new corpus and also
gives nearly perfect translations. But for a few sentence pair which has more than
three words, it gives a very bad translation which is neither adequate nor fluent.
The idea of machine translation is to provide a better and faster communication

bridge for people belonging to different linguistic backgrounds. Through this paper,
we have contributed a small portion of our work to the machine translation commu-
nity by providing a new corpus Digaru-English with few performance analyses and
comparisons with the German-English corpus based on the statistical approach of
MT. We have analyzed the translation output based on its adequacy and fluency. The
further work in progress aims to increase the Digaru-English corpus and apply more
of the statistical methods and also apply a neural approach to machine translation
and generate its performance metric which will improve the adequacy as well as the
fluency believing the fact “more the data better will be the results”.
References
1. Blench R (2017) The ‘Mishmi’ languages, Idu, Tawra and Kman: a mismatch between
cultural and linguistic relations. McDonald Institute of Archaeological Research University
of Cambridge
2. Sastry GDP (1984) Mishmi grammar, CIIL grammar series-II
3. Liddy ED (2001) Natural language processing. In: Encyclopedia of library and information
science, 2nd ed. Marcel Decker, Inc., New York
4. Pathak A, Pakray P, Bentham J (2019) English-Mizo machine translation using neural and
statistical approaches. Neural Comput Appl 31:7615–7631
5. Bhattacharyya P (2015) Machine translation, xxv, 234 pp. CRC Press, Boca Raton. ISBN:
978-1-4398-8
6. Brown PF, Cocke J, Pietra SAD, Pietra VJD, Jelinek F, Lafferty JD, Mercer RL, Roossin PS
(1990) A statistical approach to machine translation. Comput Linguist 16(2):79–85
7. Brunning J (2010) Alignment models and algorithm for SMT. University of Cambridge
8. Koehn P, Hoang H, Birch A, Callison-Burch C, Federico M, Bertoldi N, Cowan B, Shen W,
Moran C, Zens R et al (2007) Moses: open source toolkit for statistical machine translation. In:
Proceedings of the 45th annual meeting of the ACL on interactive poster and demonstration
sessions. Association for Computational Linguistics, Stroudsburg, pp 177–180
9. Jabin S, Samak S, Sokphyrum K (2013) How to translate from English to Khmer using Moses.
Int J Eng Invent 3(2):71–81. e-ISSN: 2278-7461, p-ISNN: 2319-6491
10. Papineni K, Roukos S, Ward T, Zhu W-J (2002) BLEU: a method for automatic evalua-
tion of machine translation. In: Proceedings of the 40th annual meeting of association for
computational linguistic (ACL), Philadelphia, pp 311–318
11. Dabre R, Cromieres F, Kurohashi S, Bhattacharyya P (2015) Leveraging small multilingual
corpora for SMT using many pivot languages. In: Proceedings of the conference of the
North American chapter of the association for computational linguistics: human language
technologies, Denver, pp 1192–1202
12. Blench R (2017) A dictionary of Tawra, a language of Arunachal Pradesh, orthographic edition.
McDonald Institute of Archaeological Research University of Cambridge
13. Koehn P, Hoang H (2010) Moses, statistical machine translation system. User manual and code
guide
14. Denkowski M, Lavie A (2014) Meteor universal: language specific translation evaluation for
any target language. In: Proceedings of the EACL 2014 workshop on statistical machine
translation
EMT Analysis of Heavy-Duty EVs
in Charging Station
Swati Shukla and Gaurav Trivedi
Abstract The catastrophic global energy crisis, which is gripping the world’s
progress voraciously, requires stringent and long-term solutions. Sooner or later,
there will be a complete shift from IC engine based heavy-duty vehicles to elec-
tric powered ones. These vehicles will draw a high magnitude of current from the
charging stations. This paper presents EMT analysis for heavy-duty electric vehicle
(EV) charging stations. A constant current EV charging system has been designed.
Simulation is performed for the 24-s duration, considering the various charging
scenarios (CS).
Keywords Heavy-duty electric vehicle · Fast charging · Grid-tied-micro-grid ·

Constant current
1 Introduction
The EV market today is witnessing sharp rise in sales due to a variety of factors
like improved technology, better vehicle design, and dipping battery prices. The ever
increasing demand for environmentally safe means of transport and increased govern-
ment spending in EV infrastructure are the other reasons to drive this growth. Big
international companies are playing a significant role in popularizing the EV bene-
fits over conventional internal combustion (IC) engine based vehicles. EV market
faces several discrepancies in terms of standardization of EV batteries and ratings of
charging stations. These discrepancies make it difficult for government and private
entities to provide support to the EV manufacturers and customers. Table 1 presents
a few types of heavy-duty EV batteries available in the market.
S. Shukla (B)
Center for Energy, Indian Institute of Technology Guwahati, Guwahati, India
G. Trivedi
Guwahati, India
994 S. Shukla and G. Trivedi
Table 1 Market specifications of heavy-duty EVs

Serial number Company Truck model name Power rating (kWh)
1 Emoss EMS 18 Series 240
2 DAF CF TRUCK 170
3 Daimler E-FUSO Vision ONE 300
4 Volvo FL Electric 300
Fig. 1 Block diagram of the proposed system
Figure 1 shows the block diagram of the distribution network studied in this paper.
A grid supply is used to run the dynamic loads of the system, meanwhile charging
stations are operated at a stepped-down output voltage in a parallel topology. The
analyses of power system behavior due to various real-life situations is carried out
in this study, and the detailed result discussion is presented in the later sections of
the paper.
The paper is organized as follows. Section 2 discusses the related work carried
out in the domains of EV charging. Practical simulation models of a heavy-duty
EV battery and modeling of the constant current based EV charger is discussed in
Sect. 3, and Sect. 4 presents results/discussion, followed by a conclusion of the entire
analysis in Sect. 5.
2 Related Work
Charging station design plays a vital role in determining the popularity of the EV
market. Therefore, a lot of studies have been done in this domain. In [1], vehicle
to grid (V2G) and vehicle to micro-grid (V2M) implementation have been studied.
It can be noted that the power quality and reliability improves when this system is
deployed. In another study [2], integration of wind energy system has been done in an
active distribution network. Zheng et al. [3] propose an optimal design for the battery
EMT Analysis of Heavy-Duty EVs in Charging Station 995
charging/swapping station. Based on the life cycle cost of a battery, an assessment

is made for the project, and an improved differential algorithm is used to solve the
problem of reinforcing the distribution network. However, the study assumes the EV
bus growth forecast to be 25 per year, which can vary significantly for the practical
scenario.
In [4], a mathematical model of a single load EV charging has been devised,
and actual data of harmonic current has been studied. The internal parameters are
considered to be constant. Another study in [5] proposes a heuristic micro-grid for a
commercial building, to reduce the dependency of EV charging on the power grid.
Marcincin and Medvec [6] check the impact of an active charging station, on micro-
grid based geographical locations. In [7], an impact analysis to study the effect of
plug-in electric vehicle (PEV) charging on a grid network with medium voltage, has
been performed. Three separate charging technologies with three different charging
methods have been used to study the impact. In an analysis performed in [8], a study
about the penetration of EV charging in residential establishments is shown.
In [9], an analysis to study the impact of EV charging has been performed. The
results show that EV charging casts a negative impact on the grid voltage profile,
power factor and also leads to increased energy losses. In paper [10], a situation
with a micro-grid in islanding operation is considered, where it becomes neces-
sary to perform EV load shedding. In [11], an analysis to study the impact of PEV
charging using a low voltage distribution grid has been completed. Various penetra-
tion levels have been considered, and the scenarios of slow and fast charging have
been discussed. A study in [12] uses Monte Carlo simulation to estimate the power
consumption due to fast EV charging. In [13], an analysis to study the impact of EV
charging on the distribution network in Los Angeles has been done using real data.
The report estimated the usage of charging stations by EV owners under various
situations. The proposed study has been performed for 10, 20, and 30% penetration
rates of EVs in the area. Paper [14] discusses the possibility of designing a charging
station park in a high population density urban area, which consists of slow and fast
charging facilities. In [15], it has been studied that EV charging at a commercial
building, adds up to the power consumption of the micro-grid and increases the peak
demand energy requirement. The analysis in [16] studies the impact of a single line
to ground fault condition occurring as a result of EV charging. It has been observed
that the node connecting the EV to the grid witnessed an increase in the voltage
drop and current rise. In [17], a study of the feeder for the charging station has been
performed. In [18], a Monte Carlo based simulation is performed to study the impact
of EV charging in power substations of commercial and residential areas. Another
paper [19] discusses the impact of an EV charging station on the distribution grid by
analyzing the power demand, voltage and current harmonics, voltage swelling and
sag, and power loss in distribution transformers. All the studies above, have under-
lined drawbacks, which this proposed analysis aims to address. The assessments
carried out in this paper closely resemble practical situations and are carried out for
a range of scenarios. The main objective here is to study the impact of functional EV
charging stations on the distribution network.
3 Modeling of the System
3.1 Estimation of the EV Battery Capacity Model
Energy rating of the battery can be estimated by assessing the energy required by the
vehicle to travel a distance of 1 km [20]. The mass of the unloaded vehicle is taken
to be 24 tons (24,000 kg) and its length, breadth, and height to be 6.7 m, 2.286 m,
and 2.286 m, respectively. As per our assumptions, we have considered a vehicle
which transports heavy loads between cities, on interconnected high-ways. Hence,
once the vehicle achieves its nominal speed, it does not experience any retardation
till the destination is reached or the vehicle runs out of charge. It can be assumed that
the vehicle with these characteristics will take 15 s to attain a speed of 100 kmph.
The resistive forces opposing the truck’s motion can be calculated by the following
equations (assuming road angle to be zero).
Force due to rolling resistance (F roll ):
Froll = f r Mg
Force due to aerodynamic resistance (F aero ):
Af Cd ρV 2
Faero =
2
Force due to acceleration resistance (F acc ):
dV
Facc = α M
dt
Table 2 shows the values considered to solve the above equations. Work done by
the vehicle to oppose these forces can be calculated by the equation:

W = F · dS
Table 2 Physical quantities

Serial number Physical parameter Value
for battery design
1 Rolling resistance coefficient (f r ) 0.01
2 Mass of vehicle (M) 24,000 kg
3 Acceleration due to gravity (g) 9.8 m/s2
4 Frontal area (Af ) 5.225 m2
5 Drag coefficient (C d ) 0.5
6 Rotational inertia constant (α) 1.1
7 Density of air (ρ) 1.184 kg/m3
Therefore, the work done by the vehicle against rolling force is 0.049 kWh, against
aerodynamic force is 0.0089 kWh and against acceleration, force is 1.0176 kWh.
Hence, the total work done by the vehicle to cover a distance of 1 km is 1.0755 kWh.
It can thereby be concluded that to cover a distance of 300 km in one charge, a battery
with a capacity of approximately 320 kWh needs to be installed.
The battery considered in this paper is a lithium-ion battery for a heavy load
EV. Lithium-ion battery is extensively used in EVs due to its property of high energy
density and low weight as compared to other battery types of similar capacity. Trucks
require batteries with high capacity which can help them to achieve greater travel
distances with lesser charging stops. The lithium-ion batteries fulfill this criterion
without adding extra weight to the truck.
3.2 Modelling of the EV Charging System
With the EVs dominating the automotive market, charging station infrastructure is
still in its niche stage and needs improvement. Various topologies and methodologies
have been proposed and tested to make EV charging more efficient. This paper
proposes a constant current charging model for an EV battery, suitable for heavy-duty
vehicles. The proposed charger can produce a range of constant charging currents,
and thus the charging time of the battery can be regulated. The current is controlled
by the controller unit presented in Fig. 2.
The charging circuit consists of a universal bridge that acts as a rectifier, a DC
link capacitor that produces a constant DC voltage, and a controller. The controller
consists of a voltage regulator unit, a current regulating inductor, and a switch. The
specification of DC-link capacitance, limiting resistance, and inductance is 58.9 µF,
10 k, and 100 mH, respectively. The rated voltage of the zener diode is considered
Fig. 2 Controller unit of charger

to be 30 kV for the simulation. Dynamic impedance, knee voltage and the forward
voltage drop across the zener diode are considered 0.001 , 400 , and 0.7 V,
respectively.
In practical situations, switching between EVs on charging stations may lead to

switching transients, which will affect the performance of the charging station. These
switching transients might lead to sudden spikes in power consumption or cause
high tension faults. Hence, protection devices must be designed and installed at the
charging station, to prevent any unnecessary transients in the system. This requires
an assessment of the distribution network performance profile, during the time of
EV switching. In this study, various charging scenarios (CS) as shown in Table 3,
consisting of different charging points and battery topologies, have been assumed. In
this study, 4 charging stations (with 12 charging points each), on an 11 kV node of a
distribution network have been considered. Table 3 shows, the pattern of switching
followed during the analysis.
The EV charger has been designed to give a range of constant current outputs,
depending on the requirement. Defining the standard charging current of the charger
as 146 A, it will charge the heavy load EV battery in 3.5 h. With this proposed
charger model, charging currents between 50 and 200% of the standard value can
be achieved. The following graphs depict the charging behavior (SoC) of the battery
for the corresponding charging current values.
It can be seen from Fig. 3a, b, that the proposed charger can be used to charge the
battery to different levels of SoC for the same charging period. A range of charging
Table 3 Switching pattern for EMT analysis

S. No. Time in s EVs on charging stations
CS1 CS2 CS3 CS4
1 0–1 0 0 0 0
2 1–5 10 10 10 0
3 5–7 0 0 0 0
4 7–8 6 6 6 6
5 8–9 0 0 0 0
6 9–11 6 6 6 6
7 11–15 0 0 0 0
8 15–18 10 10 10 0
9 18–20 0 0 0 0
10 20–24 12 12 12 12
(a) Current characteristic of the EV charging system
(b) SoC of the EV battery for different charging current
(c) Voltage characteristic of the EV (d) SoC of the EV battery

battery
Fig. 3 Characteristics of the EV charging system

(a) PV output current characteristicat DC (b) Battery charging current characteristics

Link during load shifting
Fig. 4 DC link current characteristic
currents can be achieved by controlling the charger parameters. An increase in the

charging current across the battery leads to higher charging rate for the same charging
time.
A photo-voltaic (PV) based captive power plant has been modeled to meet the
outing of the grid supply. Output current characteristic of PV system at DC link is
shown in Fig. 4a. In Fig. 4b, a sharp rise in the EV charging current may be observed
during the load shifting (i.e., from grid to PV system). Similarly, transients during
short interruption across EV’s charging system may be observed in Fig. 5a, b.
5 Conclusion
The design of the charging station is critical for the secure and reliable operation of
power distribution network. Various critical design parameters of the charging station
can be determined by the EMT analysis. In this paper, a heavy-duty EV battery model
has been presented. Design of constant current based EV charging system has also
been discussed. A wide range of constant current can be achieved which can regulate
the charging time of the battery according to the requirement. The EMT simulation
has been performed.
(a) Battery charging current characteristic during switching/interruption
(b) Voltage characteristics across battery during switching/interruption
Fig. 5 Switching analysis across the EV’s battery
References
1. Villafáfila-Robles R et al (2011) Electric vehicles in power systems with distributed generation:

vehicle to microgrid (V2M) project. In: 11th international conference on electrical power quality
and utilisation
2. Shi R et al (2012) Dynamic impacts of fast-charging stations for electric vehicles on active
distribution networks. In: IEEE PES innovative smart grid technologies
3. Zheng Y et al (2014) Electric vehicle battery charging/swap stations in distribution systems:

comparison study and optimal planning. IEEE Trans Power Syst 29(1):221–229
4. Li Q, Xiao X (2013) EV charging and its applications in active distribution systems. In: 2013
IEEE PES innovative smart grid technologies conference (ISGT)
5. Liu N et al (2015) A heuristic operation strategy for commercial building microgrids containing
EVs and PV system. IEEE Trans Ind Electron 62(4):2560–2570
6. Marcincin O, Medvec Z (2014) Concept of charging stations for electric cars. In: Proceedings
of the 2014 15th international scientific conference on electric power engineering (EPE)
7. Bosovic A, Music M, Sadovic S (2014) Analysis of the impacts of plug-in electric vehicle
charging on the part of a real medium voltage distribution network. In: IEEE PES innovative
smart grid technologies, Europe
8. Torquato R et al (2014) Impact assessment of the two-way operation of plugin electric vehicles
on distribution networks. In: 2014 IEEE PES general meeting | conference exposition
9. Joshi K, Lakum A (2014) Assessing the impact of plug-in hybrid electric vehicles on distri-
bution network operations using time-series distribution power flow analysis. In: 2014 IEEE
international conference on power electronics, drives and energy systems (PEDES)
10. Sengnongban M, Dechanupaprittha S (2014) Control strategy of PEVs charging for reducing
the impact of power fluctuations in micro grid. In: 2014 international electrical engineering
congress (iEECON)
11. Bosovic A, Music M, Sadovic S (2015) Analysis of the impacts of plug-in electric vehicle
charging on the part of a real low voltage distribution network. In: 2015 IEEE Eindhoven
PowerTech
12. Pea-Da B, Dechanupaprittha S (2015) Impact analysis of fast charging to voltage profile in
PEA distribution system by Monte Carlo simulation. In: 2015 7th international conference on
information technology and electrical engineering (ICITEE)
13. Jiang Z et al (2016) Statistical analysis of electric vehicles charging, station usage and impact
on the grid. In: 2016 IEEE power energy society innovative smart grid technologies conference
(ISGT)
14. Grackova L, Oleinikova I (2016) Impact of electric vehicle charging on the urban distribution
network. In: 2016 57th international scientific conference on power and electrical engineering
of Riga Technical University (RTUCON)
15. Moghimi M et al (2016) Impact of PEV behavior on peak demand reduction in a commercial
microgrid. In: 2016 Australasian universities power engineering conference (AUPEC)
16. Kumar R, Saxena D (2017) Fault analysis of a distribution system embedded with plug-in
electric vehicles. In: 2017 recent developments in control, automation power engineering
(RDCAPE)
17. Gómez JR et al (2017) Impact of electric vehicle charging on the distribution system in Cuenca,
Ecuador. In: 2017 IEEE PES innovative smart grid technologies conference—Latin America
(ISGT Latin America)
18. Longhi G, Borges C, Gruosso G (2018) A model to estimate the impact of electrical vehicles
displacement on medium voltage network. In: IECON 2018—44th annual conference of the
IEEE industrial electronics society
19. Karmaker AK, Roy S, Ahmed MR (2019) Analysis of the impact of electric vehicle charging
station on power quality issues. In: 2019 international conference on electrical, computer and
communication engineering (ECCE)
20. Kumar M et al (2012) Analysis of implementing the electric city bus and its coordination with
the grid. In: 2012 annual IEEE India conference (INDICON)
Analysis of the Transient Stability
of a Multi-machine System with SVC
Srishail K. Bilgundi, M. Shivu, H. Pradeepa, and M. V. Likith Kumar
Abstract Interconnection of remote electric power systems with abundant gener-

ation quantity and system load is significantly widespread due to the advancement
of the power trade among countries and within countries in various regions of the
world. In long-distance AC power transmission, as in interconnected electric power
systems, attention to be paid for retaining synchronism besides steady system volt-
ages, notably in combination with faults in the system and line switching. The aim
of this paper is to examine the ability of FACTS controller for power flow control
and to strengthen the transient stability. A two area two machine systems equipped
with SVC is considered for this study, Matlab software is used for simulation.
Keywords Static var compensator · Shunt compensation · Multi-machine system ·

Transient stability
1 Introduction
Transient stability is vital from the perspective of retaining system security. The
development of a fault should not prompt the tripping of generating and the likelihood
of successive outages will lead to complete system blackout. Transient stability limit
S. K. Bilgundi (B) · H. Pradeepa

Department of E&EE, The National Institute of Engineering, Mysuru, Karnataka 570008, India
H. Pradeepa
M. Shivu
Data Analyst, Unilog Content Solutions, Mysuru, Karnataka, India
M. V. Likith Kumar
Department of E&EE, GSSS Institute of Engineering and Technology for Women, Mysuru,
Karnataka 570016, India
1004 S. K. Bilgundi et al.
is based on the location and magnitude of disturbance and the type of disturbance in
the system and also the fault clearance time [1].
For Power system operation, stability and security are critical factors. The power
system stabilizers (PSS) are extensively implemented to damp low-frequency oscil-
lations in the range of 0.1–2 Hz. However, PSS may fail to determine the appro-
priate damping in present-day complicated power systems. Broadly, it is valuable to
identify that as loading changes the power parameters of the machine also change
accordingly. Hence, PSS adds stability to the discrepancy in loading conditions, spec-
ifications and structures of the system. Although PSS provides auxiliary feedback
signals, they endure an imperfection of being accountable to produce a considerable
discrepancy in the system voltage profile. The utilization of FACTS controllers is
highly prominent in power systems [2].
Transient disturbances in power systems have led to the introduction of FACTS
controllers like SVC, SSSC, STATCOM, UPFC and IPFC. In stable power systems,
when synchronous machines are disturbed, the equilibrium point will either go back
to its original state or will reach a new equilibrium state, without loss of synchro-
nism. To increase overall system performance, a lot of research work is already
carried out on the cooperation among PSS and FACTS controllers in damping power
oscillations [2–7], the parameter tuning of PSS and FACTS controllers is a compli-
cated activity as heavy-handed regional control of PSS and FACTS may lead to
destabilizing interactions [2].
2 Proposed Approach
2.1 Introduction to Static Var Compensator (SVC)
The SVC standardizes voltage at its endpoints by consuming reactive power at system
voltage beyond the threshold value or by feeding reactive power at system voltage
is below the reference value. The SVC and a simplified block diagram of its control
system are represented in Fig. 1.
2.2 Description of the Proposed System
A 1000 MVA hydro generation area-1 is interconnected to a remote load area over a
500 kV transmission line of length 700 km. The load Centre consists of 5000 MW
+ j1500 MVAR inductive load. The power to load is supplied from the remote
1000 MVA area-1 generation and a regional hydro generation area-2 of 5000 MVA.
In order to retain system stability for a post fault condition, the ±200 MVAr SVC
is connected at midpoint of transmission line. Figure 2 represents the single line
diagram of the proposed system.
Analysis of the Transient Stability of a Multi-machine System … 1005
Fig. 1 Single-line diagram of an SVC with its control system block
Fig. 2 Single line diagram of the proposed system with SVC at the midpoint of the line
Area-1: 1000 MVA area-1 generator is connected to Bus 1 through a step up

transformer-1 (13.8/500 kV, 1000 MVA).
Area-2: 5000 MVA area-2 generator is connected to Bus 3 through a step-up
transformer 2 (13.8/500 kV, 5000 MVA).
The two power plants located on the 500 kV systems generate a total power of
6000 MW and is fed to a 5000 MW, 1500 MVAr inductive load connected at bus
B3. The two machines are designed with a hydro turbine and governor (HTG), IEEE
Type-I excitation system, and PSS.
2.3 Cases Considered for Performance Evaluation
The following cases are simulated to evaluate the performance of the system under
consideration with SVC connected to transmission line at midpoint.
1. Occurrence of a symmetrical fault
2. Sudden increase of load.
For a stable system, usually, the oscillatory transients that occur after system
perturbation will be damped towards a new quiescent condition. These oscillations
anyhow cause power flow fluctuations over the transmission lines. If surplus power
flow fluctuations in a line connecting to two areas may lead to tripping of protective
devices thereby detaching or separating two areas, which may disturb the stability
of the tie line. The shunt converter is can produce or consume administrable reactive
power based on the requirements [1].
3 Simulation Results of the Proposed System Under

Transient Disturbances with SVC Controller
The SVC uses local measurements as input signals to control systems. The Simulink
model of the system is shown in Fig. 3.
3.1 Case 1: Occurrence of a Symmetrical Fault
A three-phase to ground symmetrical fault is generated in the system at bus B1 at 1 s

for a duration of 6 cycles, i.e., for a time period of t = 0.1 s, the following cases are
investigated.
1. System with PSS only
The plots of rotor speed of each machine, and tie-line power flow, BSVC a measured
voltage in pu, are shown in Figs. 4, 5, 6, 7, 8 and 9. The demagnetizing current rises at
the time of fault and remains high even though the fault is cleared, due to increase in
the rotor angle. High initial response of the excitation systems with huge ceilings (for
field forcing) contributes to 1st swing stability. However, they also affect the damping
of subsequent swings. The system becomes unstable even after the clearance of fault.
From Figs. 4 and 5, it is observed that during fault the maximum swing of machine
1 and machine 2 are 68° and 38°, respectively. It also observed that the rotor angle
swing is not getting damped and as a result machine 1 goes out of synchronization.
The plots of rotor speeds of machine 1 and machine 2, tie line active power, tie line
reactive power and terminal voltage are shown in Figs. 6, 7, 8 and 9, respectively.
Fig. 3 MATLAB/SIMULINK system model for a symmetrical fault without and with SVC at bus
B2
Fig. 4 Plot of rotor angle of

machine 1 versus time
[system with PSS only]
Fig. 5 Plot of rotor angle of

Fig. 6 Plot of rotor speeds

of machine 1 and machine 2
versus time [system with
PSS only]
Fig. 7 Plot of tie-line active

power versus time [system
with PSS only]
Fig. 8 Plot of tie-line

reactive power versus time
Fig. 9 Plot of terminal

voltage versus time
2. System with PSS and SVC

The system has a SVC connected to transmission line at bus B2 i.e., at a transmission
line length of 350 km from each of the buses B1 and B3. The plots of rotor angle
(with respect to Centre of Inertia), rotor speed of each machine and tie line power
flow are shown in Figs. 10, 11, 12, 13, 14 and 15. From the plots it is observed that
the system is stable for t = 0–1 s with normal operating conditions (steady state). The
system oscillates for short time and become stable with the application of SVC. The
primary objective of applying a SVC in transmission network is the fast regulation
of voltage at a load or an intermediate bus. The SVC injects a reactive current in
phase quadrature with the SVC bus voltage. In the transmission system, the voltage
at the middle of the line or some other suitable intermediate point can be maintained
with the help of SVC, in case of power flow increment that comes upon instantly
Fig. 10 Plot of rotor angle

of machine 1 versus time
[system with PSS and SVC]


PSS and SVC]

active power versus time

Fig. 15 Plot of bus voltage

and reactive power injected
by SVC versus time
after clearing the fault, Hence, terminal voltage is improved after clearance of fault
thus the transient stability of the system with SVC is improved considerably.
As observed from Figs. 10 and 11 within few cycles the system regains the stability
after fault clearance whereas, with only PSS the system loses its stability as seen in
the previous case, it is also observed that there are no power swing oscillations when
SVC is present along with PSS. The plots of rotor speeds of machine 1 and machine 2,
tie line active power, tie line Reactive power, bus voltage and reactive power injected
by SVC are shown in Figs. 12, 13, 14 and 15, respectively.
3.2 Case 2: Sudden Increase of Load
The load in the system is suddenly increased by 20% of normal load at 1 s at bus B3
for a duration of 25 cycles i.e., for a time period of 0.4 s. The Simulink model of the
system is shown in Fig. 3.
1. System with PSS only
The plots of rotor angle (with respect to Centre of Inertia), rotor speed, terminal
voltage of each machine and tie-line power flow are shown in Figs. 16, 17, 18, 19, 20
and 21. From the plots, it is inferred that the system is stable from 0 to 1 s with normal
load conditions (steady-state). When a load connected to the system is increased
suddenly, it results in a transient disturbance. The system becomes unstable even
after the disconnection of additional load as there is a large decrease in the system
voltage.
From Figs. 17 and 18, it is observed that during step increment of load there is a
maximum rotor angle swing of 14° and 16° in machine 1 and machine 2, respectively.
It is also observed that the rotor angle swing is not getting damped and as a result
machine 1 goes out of synchronization. The plots of rotor speeds of machine 1 and
machine 2, tie line active power, tie line reactive power, and terminal voltage are
shown in Figs. 19, 20, 21 and 22, respectively.

[system with PSS only for
sudden increment in load]


PSS only for sudden
increment in load]



voltages of machine 1 and

[system with PSS and SVC
for sudden increment in load]
2. System with PSS and SVC

The system has a SVC connected to transmission line at bus B2, i.e., at a transmission
line length of 350 km from each of the buses B1 and B3. The plots of rotor angle
(with respect to Centre of Inertia), rotor speed of each machine and tie line power
flow are shown in Figs. 22, 23, 24, 25, 26 and 27, respectively. From the plots, it is
inferred that the system is stable from 0 to 1 s with normal load conditions (steady-
state). When the load connected to the system is suddenly increased it results in a


PSS and SVC for sudden
increment in load]



voltages of machine 1 and
Fig. 28 Plot of bus voltage and reactive power injected by SVC versus time [for sudden increment
in load]
transient disturbance. The system oscillates for short period of time and becomes
stable with the application of SVC. Whenever an additional load is connected to the
system, voltage decreases below the rated value. Thus, the SVC will inject a reactive
current ‘ISVC ’ in negative (capacitive mode operation of SVC). The plots of SVC
bus voltage and reactive power injected by SVC are shown in Fig. 28.
As observed from Figs. 22 and 23 within a few cycles the system regains the
stability during step increment of load. It is also observed that there are no power
swing oscillations in this case. The plots of rotor speeds of machine 1 and machine
2, tie line active power, tie line Reactive power, terminal voltages of machine 1 and
machine 2, bus voltage and reactive power injected by SVC are shown in Figs. 24,
25, 26, 27 and 28, respectively.
Summarizing the contributions of this paper; the model developed subjected to
different types of transient disturbances is analyzed. The stability of the proposed
system with and without SVC is examined by connecting the SVC to transmission
line at mid-point. From the results, it is observed that the system effectively damps
the power oscillations under transient disturbances.
4 Conclusion
A fast-acting power electronics-based voltage impedance type converters are devel-

oped to improve the transient stability to a large extent. In this paper, dynamics
of the power system performance are compared with and without SVC in the
event of transient disturbances. Proposed FACTS controller is implemented in
MATLAB/SIMULINK. Simulation results indicate that a two area two machine

system provides better damping characteristics in the presence of SVC and PSS than
PSS alone.
References
1. Patel AD (2015) Comparison of shunt FACTS controllers for the improvement of transient
stability of two machine power system using MATLAB modelling. Int J Sci Res Dev 3(10)
(2015). ISSN (online): 2321-0613
2. Nayak S, Mohapatra SK (2015) Coordinated design of PSS and STATCOM based power oscil-
lation damping controller using MOL algorithm. Int J Appl Innov Eng Manag (IJAIEM) 4(12).
ISSN 2319 – 4847
3. Zhou E (1993) Application of static VAr compensators to increase power system damping. IEEE
Trans Power Syst 8(2):655–661
4. Singhal P, Madotra PK (2015) Transient stability enhancement of a multi machine power system
using SVC controller. Int J Emerg Trends Technol Comput Sci 4(1)
5. Banga A, Kaushik S (2011) Modeling and simulation of SVC controller for enhancement of
power system stability. Int J Adv Eng Technol
6. Kushwah R, Gupta M (2016) Modelling & simulation of SVC and statcom for enhancement of
power system transient stability using Matlab. In: 2016 international conference on electrical,
electronics, and optimization techniques (ICEEOT), Chennai, pp 4041–4045
7. Virk PS, Garg VK (2013) Power system stability improvement of long transmission line system
by using static var compensator (SVC). Int J Eng Res Appl 3(5):1–3(3)
Projection-Based Throughput
Computation of Three-Dimensional
Geographical Routing Protocol
for Multi-hop Wireless Sensor Networks
Afruza Begum and Md. Anwar Hussain
Abstract Three-dimensional multi-hop wireless sensor networks consist of a large

number of wireless devices having sensing and communication capabilities to
perform a collaborative task for supporting a wide range of applications. In the
majority of the existing routing techniques are proposed considering 2D environ-
ments, maybe because of working with 3D space is more difficult than 2D space, but
in various real-life scenario nodes need to be deployed in 3D space also. An efficient
3D routing technique that gives maximum overall throughput becomes very essential
because traditional existing 2D routing techniques are not perfectly applicable in 3D
environment. In this paper, we have proposed a projection technique to project the
routing paths obtained by our earlier proposed three-dimensional routing protocol
and computed the throughput and other performance metrics based on the projec-
tion technique to simplify the working difficulties with the route in 3D space and
for enhancing the overall network performance. The performance of the proposed
technique is evaluated through extensive simulation in Matlab platform in terms of
end-to-end throughput, energy consumptions, and end-to-end delay.
Keywords Routing · End-to-end delay · Throughput · Three-dimension ·

Protocol · Projection
1 Introduction
Multi-hop wireless sensor network consists of number of a wireless device having

sensing and communication capabilities to perform a collaborative task. In most of
the multi-hop wireless network, communication between wireless devices is done
through radio propagation range, for long-distance communication, when the desti-
nation is not within the transmission range of sender, one or more intermediate node is
A. Begum (B) · Md. Anwar Hussain

Department of Electronics and Communication Engineering, NERIST, Itanagar 791109, India
Md. Anwar Hussain
1018 A. Begum and Md. Anwar Hussain
used as a next-hop relay/forwarder to reach the destination, and this is called multi-
hop routing. In last few decades, a large number of geographic routing technique
has been proposed, almost all of those techniques are designed for two-dimensional
space [1], however, in various real-life scenarios sensor nodes need to deployed in 3D
space for important applications like underwater ocean monitoring, forest fire moni-
toring, weather monitoring, environment, pollution monitoring, etc. 3D embedding
in routing gives more accurate network behavior and results. Geographic routing is a
very popular and efficient routing technique; it uses location information for routing
instead of IP address. Considered that each node in the network knows the location
of itself, location of its neighbor and destination using Global Positioning System
(GPS) or other centralized location management techniques. In the network of nodes,
some of those nodes are considered as *sender or *user and one of those nodes as a
*base station or *destination.
End-to-end throughput, delay, and energy consumption are very crucial network
parameter for 3D multi-hop wireless sensor network when a huge number of nodes are
distributed over an area in a 3D space, it becomes very essential to design routing and
scheduling algorithm, which is energy efficient, needs minimum delay and hops to
reach the destination, and gives maximum overall throughput. When a user (sender) in
multi-hop wireless sensor network is sending packets to the base station (destination),
the user node should select such an intermediate node/path to go forward, such that
information can be sent in less time (delay) with minimum energy consumptions and
gives maximum throughput. Dealing with and implementing 3D routing protocol is
much more complicated than dealing with two-dimensional routing. Projecting the
path obtained in 3D space into a plane may simplify the task up to some extends.
In this paper, we have presented a projection technique to project the path obtained
by our earlier proposed 3D routing and scheduling technique [1] on to a particular
plan to simplify the working difficulties with the route in 3D space and enhancing
the overall network performance. Performance metrics such as end-end delay and
throughput and energy consumption are computed by the proposed projection tech-
nique and compared with the performance metrics of origin route in 3D space. Here
in the proposed projection technique, each intermediate distance (path) travel by
proposed protocol in a 3D multi-hop wireless network is projected onto a particular
plan in 3D space, which simplifies the complications of graphing on 3D space. The
route obtains from the proposed projected technique consumes less energy, requires
a minimum delay, and gives maximum throughput compared to the original route
and some existing popular routing protocol. Each sender selects the minimum angle
IM node within the transmission range, the line joining the sender and the selected
IM is projected on to a particular plane, for each sender/user this process of selection
of minimum angle IM and then the projection of selected node on to the plane is
repeated till it reached the destination (base station).
The rest of the paper is divided into different sections as follows-Description of
related existing works is done in Sect. 2, the methodology and the proposed technique
is described in Sects. 3 and 4 is detail about result analysis and performance evaluation
and the Sect. 5 is the conclusion and future works.
Projection-Based Throughput Computation of Three-Dimensional … 1019
2 Related Works
Research in three-dimensional wireless networks is gaining popularity because of

many applications like natural resource monitoring, pollution monitoring, forest fire
monitoring, prediction of natural resources, and many more. In [2], Mariam Akbar
et al. introduced a mobile sink AUV and courier node for 3D underwater sensor
networks for minimizing energy consumptions. In [3], to improve the performance
of the sensor network, the author Ganglia et al. proposed a high-availability data
collection scheme based on multiple autonomous underwater vehicles. They have
also proposed malfunction discovery and repair mechanisms when the AUV fails to
communicate with each other.
In [4], the authors proposed a Robust and Efficient Routing Protocol (RERP)
to obtain a static route for the communication between two nodes in the network.
They have assigned a rank to every node by a stable network by formulating some
important aspects of the sensor. They have selected the least rank sensor nodes for
efficient communication. In [5] this paper, Ghoreyshi et al. proposed energy-efficient
cluster-based novel mobile data gathering (CMDG) scheme for large-scale UWSNs.
The authors have claimed that to deal with the sensors mobility in UWSNs, CMDG
is the first AUV tour planning scheme, this technique is applicable in both connected
and disconnected networks.
In [6], the authors proposed a prediction-based delay optimization data collection
algorithm (PDO-DC). To obtain the corresponding prediction models, delay opti-
mization algorithms, Kernel Ridge Regression (KRR) is utilized via cluster member
nodes. Here to reduce the delay, the AUV obtains all the cluster data it traverses by
less cluster head nodes. In [7] using the adaptive packet forwarding region (PFR)
and selecting fast-forwarding nodes in the PFR, Rubeaai et al. have proposed a three-
dimensional real-time routing protocol (3DRTGP) for wireless sensor networks. In
3DRTGP they have used the PFR technique to limits the number of nodes forwarding
towards the destination and they also given heuristic solution to the multi-hop wireless
networks.
Shah and Kim [8] analyses various issues related to 3D wireless ad hoc and sensor
networks, they have discussed mainly 3D underwater sensor networks (UWWSN)
and 3D airborne ad hoc networks (AANET’s). Design requirements and challenges
in 3D AANET’s and UWWSN is discussed thoroughly, different researchers have
highlighted different issues associated with sensor networks and 3D wireless ad hoc
networks.
In [9] Entezami and Christos Politis have proposed a real-time, position-based and
energy-efficient protocol for WSNs based on the spanning tree method. When looking
for a path to the destination they have used a greedy forwarding approach. To reduce
the minimum number of node forwarding they have uses unique restricted parent
forwarding regions (PFR), to cover only the parent node, it reduces the radio range.
To overcome from dead-end route problem they have used a rainbow mechanism.
In the paper [10], optimized depth-based routing (iAMCTD) has been proposed,
which is based on improved adaptive mobility of courier hubs. This protocol is used
for distance communication and mainly for the fisherman to identify the country’s
border while fishing. In [11], Ahmed et al. proposed two-level redundant transmission
control (RTC) approach using some statistical test that ensures the data-similarity,
the data similarity from the original data removed before forwarding to the next
level by the cluster head and the regional head. Redundant transmission between the
mobile sink and cluster head is controlled by the region head. The data-similarity and
redundant inner-network transmissions between clusters members are controlled by
the cluster head.
In this paper, we have calculated the performance metrics throughput, energy
consumption, delay of the proposed 3D routing protocol with the proposed projection
method, here the line joining the sender/user node and the selected minimum angle
intermediate node of 3D space is projected on to a particular plane to simplify the
calculations and working difficulties with the 3D space, this process is repeated till
reaches the destination, performance metrics are compared and it is found that the
performance of the projection method outperforms the performance of the origin
route found in 3D space and the popular typical routing protocol Most Forward
Routing (MFR).
3 Protocol Design
In the design of this proposed protocol for multi-hop wireless networks, we have
considered that each node knows its location, location of the destination, and loca-
tions of neighbors through GPS or some other centralized location management tech-
niques. It is considered that all the nodes in the networks are homogeneous and every
node has the transmission range of Tr, nodes are scattered randomly in a uniform
manner in a 3D space. Consider a set of U stationery nodes in three-dimensional
planes X, Y, Z. When sending information to the base station which is not in the
transmission range of sender node, sender node sends information’s through one or
more intermediate nodes. Any two nodes P and Q can communicate with each other
if the Euclidean distance between P and Q is less or equal to its transmission range
(Tr), i.e., |PQ| <= Tr.
In the implementation of the proposed protocol, first a reference line from *source
to destination is drawn, then within the transmission range, the node which forms
a minimum angle with this drawn reference line is considered as an intermediate
node to transfer information towards a destination. Suppose P is the source node,
and within the transmission range Tr, among all the neighbors the node Q makes
minimum angle with the reference line, suppose the node P has location information
(Px , Py , Pz ), the location of the destination D is (Dsx , Dsy , Dsz ) and the location of a
neighbor node Q is (Qx , Qy , Qz ).
−→ −→
The vector P D is given by (Px − Dsx , Py − Dsy , Pz − Dsz ) and P Q is given by
(Px − Qx , Px − Qy , Px − Qz ). The angle < QPD, i.e., θ formed by node Q with the
reference line drawn from P to D is calculated by using Eq. 1 [1].
⎡ ⎤
−→ −→
P Q · P D
θ = cos−1 ⎣ −→−→ ⎦ (1)

P Q P D
−→ −→ −→ −→ −−−→ −→

where ( P Q · P D) is the dot product of P Q and P D. P Q and P D is the
Euclidean vector norms [7].
The projection of the line joining the sender node and the selected minimum angle
intermediate node is done on considering a particular plane (there are infinite number
of plans), this minimum angle node Intermediate Node (IM) acts as the new sender
node, this new sender node(IM) a reference line is drawn to the destination, then
from neighbor list of this new sender node minimum angle node is selected again
and projection of the line joining the new sender and the new selected IM is projected
on to the plane. The same procedure will be repeated for the selection of IM and
project it on to the plane till it reaches the destination/base station.
If the line is perpendicular to the plane, its projection on to the plane is simply the
intersection point on the plane, and if the line is parallel to the plane then it is also
parallel with its projections onto the plane and orthogonal to the normal vector of the
plane. If the line on the 3D space is not parallel also not perpendicular to the plane
then the projection of a line onto a plane is the line one end joining the intersection
points of line and the plane and the intersection of the normal through the other ends
of the line and the plane, or the line joining the two intersections through the normal
of the two ends of the line and the plane.
As shown in the Fig. 1, S is the sender and if IM is the selected minimum angle
intermediate node, for projecting the selected IM onto the projection plane, need
to consider a plane in 3D space, suppose the plan is ax + by + cz + d = 0, then
calculation of equation of the line is done by joining the sender node and the IM
node.
Intersection point of the line and the plane is calculated or the intersections of
the normal through ends of the line, S and IM is the intersection point of S and IM
on to the plane, the projection of line S I M on to plane is S I M . If S(sx 1 , sy1 , sz1 ),
is the sender and the Intermediate node is IM(mx 1 , my1 , mz1 ) the intersection point
S (ipx , ipy , ipz ) of the line and the plane is calculated from the equation of line by the
following equations:
i px = a ∗ t + sx1 (2)
i p y = b ∗ t + sy1 (3)
i pz = c ∗ t + sz 1 (4)
where a, b, c is the normal vectors and the t is the parameter describes a particular
point on the line, sx1 , sy1 , sz1 is the location of the node which needs to project on to
the plane. Now putting the values of the equations intersection points (ipx , ipy , ipz )
on the equation of the plane (in x, y, and z).
a(at + sx1 ) + b(bt + sy1 ) + c(ct + sz 1 ) + d = 0 (5)
Solving the Eq. 5, obtained the value of the parameter t.
asx1 + bsy1 + csz 1 + d

t =− (6)
a 2 + b2 + c2
where d refers to the transition or sliding of plane, so that the plane doesn’t pass
through the origin a, b, c is the normal vectors, the parameter t describes the particular
point on the line. Solving the Eq. 5, obtained the value of the parameter t.

I px = sx1 − a · (7)
a 2 + b2 + c2
I p y = sy1 − b · (8)
a 2 + b2 + c2
I pz = sz 1 − c · (9)
a 2 + b2 + c2
Obtained S (Ipx , Ipy , Ipz ) is the intersection point, i.e., the projection of node S
(Fig. 1) of 3D space onto the plane. Similarly, projection of the other end of the line
(IM) as shown in Fig. 1 is calculated, it is the intersection of the normal through the
Fig. 1 Projection of the selected path of 3D space on to a plane

point IM. Now joining the two intersection points S IM on the plane is a projection
of the line SIM of 3D space on to the plane. Now the IM will select another minimum
angle IM to go towards the destination. The same procedure of selection of minimum
angle node on 3D space and project is on to the 3D plane we have considered will
be repeated till it reached the destination node.
4 Performance Evaluation and Result Discussion
Performance of the proposed technique is evaluated by comparing the output metrics

of the origin route obtain in the 3D space, the route obtains after projection on to
the plane and with route obtain from the popular the typical protocol Most Forward
Routing (MFR). We have just considered the route selection techniques of MFR. We
have executed in the Matlab platform considering a 3D network environment, the
number of the nodes we have increased from 100 to 500, by adding 50 number of
nodes every time and each node in the network has the transmission range 2.5 km. For
accuracy, we have we run each network 100 times and take the average for statistical
constant result. We have considered a plane 3x − 2y − 2z + 2 = 0 for projecting the
line joining the sender node and the considered intermediate node (IM) to forward
toward the destination. In the networks, all the nodes are scattered randomly, among
all the scatted nodes some are considered as a sender/user (S), and one them is
considered as a base station (BS). We have considered here the three performance
metrics end-to-end throughput, energy consumption, and the end-to-end delay. The
commons simulation parameters are tabulated in Table 1 [1].
In unit time, successfully transmitted data amount from source to destination
is measured as end-to-end throughput (ETH). Figure 2 depicts the changes of the
end to end throughput for an increasing number of nodes deployment for the route
obtained by our proposed protocol in the 3D space, route obtained due to projection
of the selected route onto the plane and the route obtained by the popular typical
protocol MFR. From Fig. 2, it is seen that end-end throughput of the route obtained
due to projection on to the plane, gives better performance than the origin route
obtained on the 3D space and the MFR because routing in 3D space is much more
complicated than routing in 3D plane. Projecting the routes on a 3D plane simplifies
Table 1 Network simulation

Network parameters Values
parameters
Data packet size 1 MB
Network area 5 × 6 × 5 km
Number of nodes 100–500
Range of node 2.5 km
E elec 50 nJ/bit
E amp 100 pJ bit m2
the routing process and improve the performance of the routing. From Fig. 2 it is
also observed that with increasing network density, average end-to-end throughput
increases very slightly, as the number of nodes deployed in the network does not
create any considerable changes in the selection process of intermediate nodes (IM) to
forward packets toward destination. It is also observed, end-end throughput increases
slightly in case of plane compared to 3D space and MFR whenever we calculate
routes, with an increasing number of nodes in the networks, this because find route
on a plane is easier than challenging 3D space.
The amount of time needed for transmitting information from source to destination
is the end-end delay. The Fig. 3 shows the variations of end-to-end delay under
increasing number of node deployment for the proposed techniques with the popular
existing protocol most forward routing (MFR). From the Fig. 3, it can be observed that
end-to-end delay for the route obtain through the projection on the plane is much
less than the route obtain in the 3D space and the route obtains by most forward
routing(MFR), less delay means better performance. This is because projection on
the 3D plane simplifies the complexities of working with 3D space. From Fig. 3,
also we have observed that the end-to-end delay decreasing very slightly with the
increasing number of node deployment. Because when the transmission range is
constant, number of nodes deployed do not create any issue in the position of selected
IM nodes to forward packets toward destination.
Fig. 2 End-end throughput variation with different number of nodes deployment

Fig. 3 End-to-end delay variation with different number of nodes deployment
Energy consumption is a very important issue for the nodes deployed in an under-
water multi-hop wireless network. Figure 4 depicts the variation in consumption
of energy for an increasing number of nodes deployment for the route obtained by
the proposed technique in 3D space and the projected route on to the 3D plane,
and the typical protocol most forward routing (MFR). From Fig. 4, it is seen that
energy consumption for a route in 3D space and the route obtained on the projected
3D plane by proposed technique is considerably less than the existing protocol the
Most Forward Routing (MFR) because, in 3D space, selection of the IM node is
done based on minimum angle node with the reference line, it chooses the IM nodes
which are not too far away from source and destination, so the energy consumption
is decreased. It is also observed that with the increasing number of node deployment
energy consumption decreases when there is less number of nodes are there in the
network, the route through 3D space consumes less energy than the route through
3D plane, but when more number of nodes are there in the network, route though 3D
plane consumes less energy than the 3D pace and most forward routing (MFR), this
is because with an increasing number of node better IM can be found to go forward
towards the destination.
In this paper, we have computed the performance metrics for the proposed three-
dimensional routing protocol based on the projection technique, first route discovery
Fig. 4 Energy consumption variation with different number of nodes deployment
is done in 3D space based on the neighbor node that makes a minimum angle with
the reference line from sender to destination or base station, then the path obtained
on the 3D space is projected on a particular plane to enhance the performance of the
route obtained in challenging 3D space. Performance of the proposed techniques is
evaluated by calculating the end-end delay and throughput, and energy consump-
tion for each user at the time of computing routes from source to the base station.
Output metrics is compared with popular existing protocol, from the result anal-
ysis it is observed that the performance metrics of proposed techniques outperform
the existing techniques. In future work, our focus will be on developing efficient
scheduling techniques so that at a time more than one user can send data without any
interference, considering node mobility of the nodes in a network.
*Sender, user, source are used synonymously and the base station and destination
are used synonymously.
Acknowledgements This work is fully supported by the Visvesvaraya Ph.D. Scheme, MeitY,
Government of India. The authors are thankful to the MeitY, GoI.
References
1. Begum A, Hussain MA, Bora J, Askari SMS (2018) A simple three-dimensional joint routing
and scheduling protocol for multi-hop wireless networks. In: 2018 4th international conference
on recent advances in information technology (RAIT), pp 1–7. IEEE
2. Akbar M, Javaid N, Khan A, Imran M, Shoaib M, Vasilakos A (2016) Efficient data gathering
in 3D linear underwater wireless sensor networks using sink mobility. Sensors 16(3):404
3. Han G, Long X, Zhu C, Guizani M, Zhang W (2019) A high-availability data collection scheme
based on multi-auvs for underwater sensor networks. IEEE Trans Mobile Comput
4. Jain R, Rakesh N (2018) Robust and efficient routing in three-dimensional underwater sensor
network. In: Proceedings of first international conference on smart system, innovations and
computing, pp 107–116. Springer
5. Ghoreyshi SM, Shahrabi A, Boutaleb T, Khalily M (2019) Mobile data gathering with hop-
constrained clustering in underwater sensor networks. IEEE Access 7:21118–21132
6. Han G, Shen S, Wang H, Jiang J, Guizani M (2019) Prediction-based delay optimization data
collection algorithm for underwater acoustic sensor networks. IEEE Trans Veh Technol
7. Al Rubeaai SF, Abd MA, Singh BK, Tepe KE (2016) 3D real-time routing protocol with tunable
parameters for wireless sensor networks. IEEE Sensors J 16(3):843–853
8. Shah B, Kim K (2014) A survey on three-dimensional wireless ad hoc and sensor networks.
Int J Distrib Sens Netw 10(7):616014
9. Entezami F, Politis C (2015) Three-dimensional position-based adaptive real-time routing
protocol for wireless sensor networks. EURASIP J Wireless Commun Netw 2015(1):197
10. Paramesh J, Robin CR (2018) A novel and efficient routing protocol for fishermen using
underwater wireless sensor network. J Comput Theor Nanosci 15(4):1226–1232
11. Ahmed G, Zhao X, Fareed MMS, Fareed MZ (2019) An energy-efficient redundant transmis-
sion control clustering approach for underwater acoustic networks. Sensors 19(19):4241
Digital Image Processing Based
Automatic Fabric Defect Detection
Techniques: A Survey
Vaishnavi Raut and Indrasen Singh
Abstract The fabric defect detection is done truly with manual inspection. Auto-
matic fabric investigation is basic to keep up the idea of texture. Fabric examina-
tion is present on the reason of modernized pictures of the fabric. The detector
gains modernized fabric pictures by picture securing gadgets and transfers it to a PC
program to forms the got image. The PC program makes a material assessment to
check whether the surface is sans defect or slipped off using Digital picture getting
ready systems. Notwithstanding unlimited estimations available, the investigation is
up ‘til now testing one. This paper shows the requirement, provocation, and methods
of automatic fabric defect-recognition program Also, moreover, the paper shows all
the chance of available advances related to the fabric defect-recognition program.
1 Introduction
To keep up the nature of the fabric, automatic fabric investigation is significant. It is

attractive to deliver the best products in the most brief measure of time is conceivable.
Almost 85% of the imperfections are found by the object of the textile industry.
Workers improve just 45–65% of their imperfections from seconds or off-standard
items. It is basic, to keep these imperfections from reoccurring. A computerized
investigation system comprises of a PC based vision system. As they depend on PC,
these systems don’t endure the downsides of manual examination.
V. Raut (B) · I. Singh

Department Electronics and Telecommunication Engineering, G H Raisoni College of
Engineering, Nagpur 440016, India
I. Singh
1030 V. Raut and I. Singh
1.1 Fabric Inspection
Material review is a significant element in modern producing factories, for example,

computerized, mechanized and curative factories. Essentially, Material recognition
has two different potential outcomes. First item disconnected review wherein the
produced fabric must be examined through the fabric inspection system. The subse-
quent chances are an online review in which the fabricating procedure can be
continually checked for the presence of defects.
1.1.1 Limitation of Manual Inspection
1. It is tough to realize the faults by mortal specialists manually.

2. Manual method needs instruction and they take some period to evolve such
expertise.
3. Visible investigation is unvarying or hard, even for the top instruct specialist.
4. The machines give the result faster and accurate than humans.
5. Human inspectors can cause eye fatigue.
1.2 Automated Fabric Inspection
Automated inspection programs are proposed to accelerate absconds distinguishing

proof and precision fabric fabricating procedures to diminish work costs, improve
the idea of things and augmentation creating an order. In part 2 the requirement and
issues of Automated inspection system is sorted, the requirement and issues of an
Automated inspection systems have about. In part 3, a chart of Automated material
audit program describing the points in levels is introduced. In part 4, classification
of each and every believable method open for Automatic inspection systems has
been immediately inspected. In part 5, the output of the utilization of different point
periods of Automated inspection systems is displayed. At last it is concluded in part
6.
2 Online Fabric Defect Detection System
The accessible material investigation is known as present material investigation

where generation and creation limit the works at the same time progressively. The
requirement for this Program is from the way that the real-time material investigation
technique which is done physically is a difficult task. The significant necessities for
an online accessible material investigation Program are as per the following:
1. The framework must work continuously with great outcomes.
Digital Image Processing Based Automatic … 1031
2. It should decrease get away from price.

3. It should decrease blunder.
4. It should be hearty and adaptable.
5. It should be of minimal effort and effective.
2.1 Challenges and Difficulties
1. Because of the enormous number of fabric defects classes, the task is challenging.
2. There are between class correspondence and between class varieties of deformi-
ties.
3. The depiction of deformities in rugged substance is normally not explained
distinctly.
4. There is a huge assortment of texture designs and fabric types.
5. There are arbitrary varieties in scute.
6. The issue of evaluating optical suspicion in difficult times.
7. This assignment has a high information stream.
8. It endures from disturbance effects.
2.2 Deformities in Fabric
Many faults in material happen during fabricating. A portion of these material faults
is obvious, while others are definitely not. Be that as it may, some fabric defects
might be redressed during weaving and in the wake of fabricating while others are
most certainly not. Surface of fabric depends on the experience of the texture. They
can be named unpleasant, smooth, smooth, delicate, luxurious, shiny, and so forth.
The various surfaces of the fabric preserve upon the kinds of fabricate utilized. The
various surfaces of the fabrics rely on the kinds of material used. Fabric materials
are utilized to get ready various kinds of texture items in the factories viz. cotton,
silk, wool, leather, and linen. Characteristic cloth and manufactured cloth are the two
unique kinds of textile cloth. The fabric defects can happen because of machine flaws,
opening, Color dying, yarn issues, scratch, poor completing, soil spot, excessive
stretching, and crack point.
3 Automatic Fabric Inspection System
The advancement of an automatic, i.e., PC visual program for material fault

investigation has a few stages as appeared in Fig. 1.
Fig. 1 Methodology of
various levels in an
automatic fabric inspection Scene Analysis and Feature
system Extraction
Image Aquisition
Image Processing
Feature Extraction
Classification
Decision Making
3.1 Image Acquisition
The basic time of any vision structure is Image procurement. There can be different
sorts of cameras utilized for this application, for example, charged coupled gadget
camera, CMOS camera, digital camera, and so on. The pixel estimation of these
cameras is around 320 × 420 pixel.
3.2 Image Preprocessing
In picture pretreating levels the collection of procedures that are utilized for the
improvement of the visual aspect of an image and furthermore it is utilized to change
over the image to a structure, from which it very well may be more qualified for
additional survey in the forthcoming levels by a human or instruments.
Feature Extraction is a phase wherein different strategies can be utilized to repro-

duce the visual substance of images for ordering and recovery reasons. There is a
numeral of properties characterized by a picture and there are strategies for figuring
every one of these properties. The properties which are more qualified for specific
implementation chosen for additional investigation.
3.4 Classification
There are some profitable classifiers, for example, Artificial Neural Network, support
vector machines, bunching, and measurable deduction. By revealing whether the
fabric is deserted or desert free the characterization organize gives the final product
of the whole fabric defect detection process. The preparation stage and testing stage
are two stages that requires a classifier utilizing neural systems. The neural system
makes the best possible change for its loads in the preparation stage.
4 Classification of Automated Fabric Defect Inspection
The Statistical, unearthly, model-based methodologies are the grouping of Struc-

tural methodologies. The surfaces which are made out of natives are the Structural
methodologies. These natives’ strategy for basic methodologies is as basic as single
pixels, a locale with constant gray stages, or line portion. In this way, right off the bat
to separate surface natives, and also to show or sum up the spatial situation rules are
the primary targets of these methodologies. Geometric connections between natives
or taking in measurable properties from surface natives can be gotten by the arrange-
ment administers through demonstrating. Be that as it may, these methodologies are
infertile on material fault recognition, fundamentally because of the flexibility of
yarns, texture movement, fiber load, commotion, etc.
4.1 Statistical Approaches
Computation of the dimensional dissemination of elements ethics, the fundamental

article is to isolate the picture of the reviewed material in the area of different analyt-
ical nature. A significant suspicion right now the action is that the insights of defect-
free area are unmoving and that these locales reach out above a critical bit of review
image. In view of the amount of pixels characterizing the neighborhood highlights,
they are grouped these methodologies into first request, second request, what’s more,
higher request measurements. The principal request measurements gauge effects like
the normal and difference of singular pixel esteem, disregarding the dimensional
collaboration between picture elements, next and elevated request measurements
then again estimate properties of at least two pixels esteems happening at explicit
areas comparative with one another.
4.1.1 Gray-Level Thresholding Approach
To recognize high differentiation fabric defects these are immediate and straightfor-
ward mean methodologies. The guideline relies upon the peak or trough that is signal
variety because of the appearance of high differentiation defects.
4.1.2 Normalized Cross-Correlation Approach
An image that shows up in another and the relationship coefficient can create a
connection plan for defect statement is utilized to find the features in correlation.
The straight and proper measure of similar nature between the two images provides
the cross-correlation function. The presence of a defect in the merit of this estimate
shows any significant variation.
4.1.3 Statistical Moments Approach
The factual details, for instance, Mean, standard deviation, skewness, and kurtosis
are given above an area and these characteristics are pre-owned for picture division.
4.1.4 Rank-Order Functions Approach
Based on histogram analysis a picture rank-purpose is a basic measurable strategy

used for defects recognition. It is given by the concatenation of grey levels in the
scatter diagram and this arrangement is arranged in the arising order. The same
information is provided by the histogram and the rank function.
4.1.5 Edge Detection Approach
For image analysis border identification is a conventional strategy. The significant

element in the textured images is the appropriation of the edge sum per unit area.
The measure of grey level advances in the material picture has address bars, fringe,
tip absconds, and other dimensional disruption. Consequently, it has characteristics,
as it were, hold for likeness testing, get together assessment, and material fault
recognition.
4.1.6 Morphological Operations Approach
The coherent morphology helps depicting the geometrical and essential properties of
pictures. Morphological pictures dealing with have significance to molding, marking,
gathering, extricating, and coordinating procedures on pictures. The morphological
activities are one of the perfect devices for exhausting clamor, in spatially isolated
pictures of textures.
4.1.7 Local Linear Transforms Approach
This methodology is emphatically identified with channel bank examination strate-

gies. It gives quantifiable help for the extraction of surface properties by systems for
convolution executives (covers). These spreads might be considered as neighborhood
discoverers’ principal structures, for instance, deserts.
4.1.8 Artificial Neural-Networks Approach
The artificial neural-framework is among the speediest and most flexible classifiers
utilized for insufficiency ID due to their non-parametric nature and capacity to portray
complex choice change. In the event that the window work is Gaussian, the windowed
zones made out of various near rudimentary handling units (neurons) related together
into a framework. These neurons are arranged in layers with the data instating the
planning at the info layer.
4.2 Structural Approaches
The spectral methodologies involve a major part of the most recent PC vision research
work, in view of spatial-frequency area highlights which are delicate to disturbance
and force variant than the highlights extricated from a dimensional domain. Incor-
poreal methodologies are prescribed to be used uniquely for PC visual of constant
rugged stuff like material as they require a high level of periodicity.
4.2.1 Fourier Analysis Approach
The Fourier investigation is a worldwide methodology that describes the fabric

picture as far as recurrence segments. Fourier approaches have alluring effects of
clamor invulnerability, interpretation unchanging, and the ideal portrayal (upgrade)
of the occasional highlights. To actualize Fourier examination for material fault
recognition, different strategies are used; Optical Fourier Transforms acquired in the
visual area by utilizing focal points and spatial channels can be utilized, in any case,
most strategies, carefully actualized, are inferred from Discrete Fourier Transforms
and additionally it is Converse which recuperates the pictures in the dimensional
area.
4.2.2 Gabor Filters Approach
The Fourier analysis is presented by its old way of spatial reliance through
the windowed Fourier change. The windowed Fourier transform turns into the
outstanding Gabor change if the window work is Gaussian, which can probably
achieve ideal restriction in the dimensional and recurrence areas.
4.3 Model-Based Approaches
The issue of discovering potential bunches in an information set (picture) is a tedious

one with a past history. Campbell et al. consolidated images preparing methods with a
ground-breaking new measurable procedure to investigate denim fabric. The method-
ology utilizes model-based grouping to distinguish moderately blackout adjusted
deformities. In request to survey the proof for the present of an imperfection, Bayesian
data establishment (BIC) is used.
4.3.1 Gauss Markov Random Field (GMRF) Model Approach
The Markov arbitrary fields utilize an exact model of this dependence, the image
has essentially random noise. The local logical data in an image is able to capture.
The power at every element in the picture relies on the powers of essentially the
neighboring elements that are anticipated from this model. This gives a supportive and
dependable way for showing setting subordinate components, for instance, pixels,
through portraying common impacts among such elements utilizing situation MRF
circulation.
4.3.2 Poisson’s Model Approach
The stochastic models of some self-decisively present-day finished materials depend

upon the chance of the collecting system. One occurrence of such material is the
wiry, non-woven material utilized for air filtration that is produced through adhesive
advancement.
Defects Missed thread Missed thread Greece Ink Tear
Original images
Filtered Images
Segmented Images
Defected Area
Fig. 2 Results of the images after applying different filtering techniques on original images
The need to recognize the best techniques is significant as there are different fabrics
and fabric defects. The relative examination is significant, as it empowers the scien-
tists to learn and comprehend the distinction dependent on its chance and precision.
The computerized pictures of both defects and defect-free are caught by the advanced
camera; at that point, this image is preprocessed and the noise is removed and then
image is noise-free. By applying the PSO middle separating system, the image appear
as shown in Fig. 2. The entropy-based segmentation applied with respect to the sepa-
rated picture, sectioning out the surrendered area if any in the picture. By then these
picture is sent to the Artificial Neural Network classifier, it arranges the pictures
in deserted and defect-free groups. To improve more the subsequent advance is
executed utilizing PSO based feature selection. Here in the proposed system, five
sorts of imperfections are recognized correctly, for example, Tear, Oil Stain, Ink
Stain, missed thread (Wept and Warp). The objective here is to demonstrate the
advantage of the strategy to identify defects if there should arise an occurrence of
genuine fabric and simulated one.
6 Conclusion
In this paper the challenges and need of Automatic Fabric Inspection system is given,
a blueprint of the system shows the computerized fabric review which shows the
defects in fabric and the cause of the defect is examined. This procedure orchestrates
85% of defects in texture and finds the blemish in the average fabric at a commendable
rate and gives 80% gathering precision. A present system and the output has been
given. We have introduced promising outcomes for a computerized texture exami-

nation for multi-class imperfection identification and portrayal in texture using both
geometric and surface features to get the visual properties.
References
1. Banumathi P, Nasira GM (2012) Fabric inspection system using artificial neural networks. Int
J Comput Eng Sci (IJCES) 2(5):20–27
2. Banumathi P, Nasira GM (2013) Fourier transform and image processing in automated fabric
defect inspection system. Int J Comput Int Inf (IJCII), 3(1):61–64
3. Bodnarova A, Bennamoun M, Kubik K (2000) Suitability analysis of techniques for flaw
detection in textiles using texture analysis. Pattern Anal Appl 3(3):254–266. https://doi.org/10.
1007/s100440070010
4. Brzakovic D, Vujovic N (1996) Designing defect classification systems: a case study. Pattern
Recogn 29(8):1401–1419. https://doi.org/10.1016/0031-3203(95)00166-2
5. Cho CS, Chung BM, Park MJ (2005) Development of real-time vision-based fabric inspection
system. IEEE Trans Ind Electron 52(4):1073–1079. https://doi.org/10.1109/TIE.2005.851648
6. Conci A, Proença C (2000) A Computer vision approach for textile inspection. Tex Res J 70(4)
7. Fatemi-Ghomi N, Palmer PL, Petrou M (1996) Performance evaluation of texture segmentation
algorithms based on wavelets. In: Proceedings of the workshop on performance characteristics
of vision algorithms (ECCV), Cambridge
8. Gonzales R, Woods R (2008) Digital image processing. Prentice Hall
9. Jansi S, Subhashini P (2013) Particle swarm optimization total variation filter for image
denoising. J Theor Appl Inf Technol 57(2):169–173
10. He J, Jiang Q (2012) Research on the fabric defect detection method based on improved PSO
and NN algorithm. Int J Dig Cont Technol Appl (JDCTA) 6(8):177–184. https://doi.org/10.
4156/jdcta.vol6.issue8.21
11. Kumar TA, Paul V, Priya S (2011) A novel approach to fabric defect detection using digital
image processing. Signal Process Commun Comput Netw Technol (ICSCCN). https://doi.org/
10.1109/icsccn.2011.6024549
12. Mak KL, Peng P, Yiu KFC (2009) Fabric defect detection using morphological filters. Image
Vis Comput 27(10):1585–1592. https://doi.org/10.1016/j.imavis.2009.03.007
13. Nasira GM, Banumathi P (2014) Automatic defect detection algorithm for woven fabric using
artificial neural network techniques. Int J Innov Res Comp Comm Eng (IJICRCCE) 2(1):2620–
2624
14. Vyas P, Kakhani M (2015) Fabric fault processing using image processing techniques. Int J
Mult Res Dev 2(2):29–31
15. Sengottuvelan P, Wahi A, Shanmugam A (2008) Automatic fault analysis of textile fabric using
imaging systems. Res J Appl Sci 3(1):26–31
16. Srinivasan K, Dastor PH, Radhakrishnaihan P, Jayaraman S (1992) FDAS: a knowledge-based
frame detection work for analysis of defects in woven textile structures. J Text Inst 83(3):431–
447
A Balanced Expert System to Manage
Parkinson’s Disease by Identifying Major
Risk Factors: B-TESM-PD
Arpita Nath Boruah, Saroj Kr Biswas, Sivaji Bandyopadhyay,
and Sunita Sarkar
Abstract With changing lifestyles and emergence of modern ways of living, people
fail to take enough care of their physical health and increase their chance of falling
prey to fatal and non-fatal diseases. With tremendous advancement in science and
technology, it is expected to have a system that can help to manage this problem. One
of such prevailing diseases is Parkinson’s disease (PD). The part of the brain that
controls movement, posture, and also emotional imbalance is affected in PD. The
symptoms of PD may differ from one individual to another. PD is detected basically
in elderly people but now it has turned out to be a chronic disease affecting all age
groups. As of its severity it is better to have an expert system which can identify
the major risk factors of PD in the early stage and accordingly some measures may
be taken to control those risk factors. It would, therefore, be really a great help to
the world to have one such expert system for managing a PD. Therefore, this paper
proposes an expert system named balanced-transparent expert system for managing-
PD (B-TESM-PD) that can identify the major risk factors by generating rules from
decision tree (DT). B-TESM-PD encompasses of five phases: preprocessing, rule
generation, rule selection, rule pruning and merging, and risk factor identification.
The imbalance nature of the PD dataset is treated by the preprocessing stage using
a dominant undersampling technique, Tomek Link. Rule generation phase gener-
ates rules using decision tree. Rule selection stage selects the transparent rules, rule
pruning and merging removes the redundant and inefficient rules from the rule set
and then merges the pruned rule set to a single rule, and lastly the major risk factor of
A. N. Boruah (B) · S. K. Biswas · S. Bandyopadhyay

Computer Science and Engineering Department, National Institute of Technology Silchar, Silchar,
Assam 788010, India
S. K. Biswas
S. Bandyopadhyay
S. Sarkar
Computer Science Department, Assam University, Silchar, Assam, India
1040 A. N. Boruah et al.
PD is identified using the risk factor identification phase. The model is validated with
the PD dataset collected from UCI repository. Standard decision tree performance is
compared with the proposed model.
Keywords Parkinson’s disease · Decision tree · Expert system · Hill climbing ·

Attribute selection · Undersampling
1 Introduction
Parkinson’s disease (PD) is among one of the prevailing neurodegenerative disor-

ders around the globe. Conferring to the American Parkinson Disease Association
(APDA) [1] more than 10 million individuals have PD. PD can be defined as a
progressive neurodegenerative disorder that basically effects the motor as well as
the emotional imbalance. The principal symptoms of the disease are formation of
tremor, rigidity and stiffness, dysphonia and slowness in motion [2–4].
Along with these, a PD patient may also suffer from emotional disorder of depres-
sion, anxiety, panic attacks, tiredness, and constipation [5] which affects day-to-day
life. PD mostly occurs among the elderly people of age above 50 [6, 7], due to the
demise of particular group of brain cells that are responsible for producing neuro-
transmitters like dopamine, acetylcholine, serotonin, and also norepinephrine [1, 8].
The degeneration of dopaminergic neurons cause the physical and mental imbalance
in the human being leading to PD. The actual reason of PD whether environmental or
genetic is still unknown. Several researches have been made for diagnosis of PD and
it is about 1–3% of the world population is being suffered from PD. In accordance
to the study World Health organization (WHO) [9], the PD growth is increasing and
soon it will be one of the most commonly occurring chronic diseases among all age
groups. Considering the severity of PD, it is very much essential to prevent it by
identifying the major risk factor of the disease.
Extracting the useful and hidden information and their relationship becomes a
decisive part of making a data-driven intelligent system. Various data mining (DM)
techniques are available which can extract the hidden information and make them
use for an expert decision-making system. However, when it aims to recognize the
most important risk feature of a disease, it becomes essential to have a transparent
system which can make its decision-making process into human understandable
form. Among all the classification tools, decision tree (DT) is the most popular
and a white box technique which generates a user understandable decision rules.
Therefore, DT can be used to find the major risk factors responsible for PD by
generating transparent rules. Once the major risk factors are identified, they can be
used to manage PD and prevention method can be discovered with very less effort.
Since decades DT, especially, has received a lot of attention because of their
universal approximation property and its transparent nature. However, DT assumes
balanced training set which is not correct always in case of medical data and hence
the classification models that are inclined toward the overrepresented class. Thus, this
A Balanced Expert System to Manage Parkinson’s Disease … 1041
paper proposes a decision support system named balanced-transparent expert system

for managing-PD (B-TESM-PD) which extracts the transparent decision rules from
DT from balanced PD data. The most prominent undersampling technique, Tomek
Link [10], is used to make the PD data balanced. Therefore, DT can learn the patterns
in both of the classes maintaining uniformity in generating transparent rules. The
proposed B-TESM-PD also prunes the generated rules and factors to make the rules
more transparent and to identify the major risk factors for PD.
The paper is arranged as follows: Sect. 2 shows the background of PD and
various works done on it, Sect. 3 gives the methodology in details, Sect. 4 evalu-
ates the experimental results of the model and comparison, and finally Sect. 5 draws
conclusion.
2 Literature Review
Different researchers used different techniques and data to identify PD. Bisalia et al.
[11] described PD elaborately. Stuart et al. [12] explained PD as a progressive
neurodegenerative disease which causes motor symptoms and visual deficits. Yuvaraj
et al. [13] stated that PD has emotional symptoms of depression, sleep disorder, and
autonomic impairments. Wang et al. [14] stated that exposure to pesticides may
increase the rate of PD. Lee et al. [15] investigated the different applications of
digital image processing method to detect the neurological disorder. Stuebner et al.
[16] particularly focused on the blood pressure and heart beat rate because in advance
stage of PD a patient may have an imbalance blood pressure and thereby can predict
any new symptoms of the PD or some other problems. Ramani and Sivagami [17]
proposed a classifier model for PD with the higher classification accuracy for vocal
data set. Sama et al. [18] used SVM classifier to classify the signals obtained from
12 PD patients using a waist worn triaxial accelerometer. Kim et al. [19] designed an
automatic PD detection algorithm using convolution neural network (CNN). Joshi
et al. [20] proposed spontaneous non-invasive PD classification by using gait analysis.
Little et al. [21] proposed algorithms to measure dysphonia and put forward anal-
ysis methods for speech to study voice disorders on the PD patients. The outcomes
of the research offer a wide range of voice disorder classes of Parkinson’s dataset.
Arjmandi and Pooyan [22] proposed an improved algorithm for pathological voice
quality assessment. They recognized different voice disorders of the vocal folds and
suggested linear discriminant analysis over other detection approach. It is evident
from the literature that PD has been studied a lot and different techniques are being
proposed for its detection.
Fig. 1 Balanced-transparent expert system for managing-PD (B-TESM-PD)
3 Proposed Balanced-Transparent Expert System

for Managing-PD (B-TESM-PD)
This section describes B-TESM-PD model, which identifies the major risk factors
of PD. The proposed B-TESM-PD encompasses of five stages: preprocessing, rule
generation, rule selection, rule pruning and merging, and risk factor identification.
Rule generation step generates rule set using decision tree (C4.5), rule selection stage
selects the transparent rules, rule pruning and merging removes the redundant and
incompetent rules from the rule set by using a proposed rule pruning algorithm and
then merges the pruned rule set into a single rule, and finally the major risk factors
of PD are identified using the risk factor identification phase. The schematic layout
of proposed B-TESM-PD is shown in Fig. 1.
This section is divided into three subdivisions: First subdivision shows the experi-
mental setup and a brief explanation of the PD dataset used in the experiment, and the
second subdivision gives the results. The last section discusses about the important
observations.
4.1 Experimental Setup and PD Dataset
The experimentation of the proposed model is performed in PYTHON 2.7 version

in Ubuntu environment. The PD dataset [21] is taken from the UCI repository. It
contains 197 samples, 23 attributes, and 2 classes. Each column in the dataset is a
particular voice measure, and each row corresponds to a patient. The imbalance ratio
of PD dataset is 3.06. The purpose of the dataset is to distinguish the healthy and
those with PD, according to “status” column which is set to 0 for healthy and 1 for
PD. The experiment of the paper is done with 10-fold cross-validation with 70% for
training and 30% for testing. The dataset comprises voice perturbation and nonlinear
dynamic parameters measured from the acoustic signals of the patients.
x1 MDVP:F0 (Hz)—Average vocal fundamental frequency
x2 MDVP:Fhi (Hz)—Maximum vocal fundamental frequency
x3 MDVP:Flo (Hz)—Minimum vocal fundamental frequency
x4 MDVP:Jitter(%)—MDVP jitter in percentage
x5 MDVP:Jitter(Abs)—MDVP absolute jitter in ms
x6 MDVP:RAP—MDVP relative amplitude perturbation
x7 MDVP:PPQ—MDVP five-point period perturbation quotient
x8 Jitter:DDP—Average absolute difference of differences between jitter cycles
x9 MDVP:Shimmer—MDVP local shimmer
x10 MDVP:Shimmer (dB)—MDVP local shimmer in dB
x11 Shimmer:APQ3—Three-point amplitude perturbation quotient
x12 Shimmer:APQ5—Five-point amplitude perturbation quotient
x13 MDVP:APQ11—MDVP 11-point amplitude perturbation quotient
x14 Shimmer:DDA—Average absolute differences between the amplitudes of
consecutive periods
x15 NHR—Noise-to-harmonics ratio
x16 HNR—Harmonics-to-noise ratio
x17 RPDE—Recurrence period density entropy measure
x18 D2—Correlation dimension
x19 DFA—Signal fractal scaling exponent of detrended fluctuation analysis
x20 Spread1—Two nonlinear measures of fundamental
x21 Spread2—Frequency variation
x22 PPE—Pitch period entropy.
4.2 Results
Rule Generation B-TESM-PD model uses C4.5 algorithm to generate decision

rules for each training set, which uses information gain ratio for attribute selection.
Each rule is assigned a score function called as worth of rule (WOR).
The rule set of a particular fold along with its WOR value is given in Table 1.
Similarly, the rule sets of the remaining folds are generated along with their WOR
values
Rule Selection The rule with the highest WOR is selected from each fold as the best
rule for that particular fold. It is seen from Table 1 that rule no. 8 has the highest
WOR value and thus is selected. Similarly, the best rules for the remaining folds are
selected and thus the primary transparent rule set is formed.
The primary transparent rule set is shown in Table 2 and the accuracy obtained
by this is 84.61%.
Table 1 Rules and WOR values of the first fold

Rule no. Rule WOR
1 x22<=0.134 and x2<=229.18 and x17<=0.409 6.068
2 x22<=0.134 and x2<=229.18 and x17>0.407 4.59
3 x22<=0.134 and x2>229.18 34.00
4 x22>0.134 and x12<=0.013 and x1<=118.141 6.068
5 x22>0.134 and x12<=0.013 and x1>118.141 and x10<=0.171 23.50
6 x22>0.134 and x12<=0.013 and x1>118.141 and x10>0.171 2.83
7 x22>0.134 and x12>0.013 and x20<=0.143 7.33
8 x22>0.134 and x12>0.013 and x20>0.143 102.33
Table 2 Primary transparent

(1) if (x22>0.134 and x12>0.013 and x20>0.143) then
rule set
Class==1
(2) if (x19>−6.318 and x10>0.19 and x20>0.143) then
Class==1
(3) if (x20>0.208 and x15>0.006) then Class==1
(5) if (x19>−6.318 and x20>0.207 and x4>0.00) then Class==1
(6) if (x22>0.134 and x12>0.013 and x13>0.019) then Class==1
Class==1
Rule Pruning A sequential hill climbing algorithm is applied on the primary trans-
parent rule set. Every time a rule is dropped and the accuracy is computed for that
particular rule set and compared with the accuracy of the original rule set. If the accu-
racy increases or there is no change then that dropped rule is permanently pruned. This
process continues till there is no more rules to be pruned which increases the accu-
racy of the rule set. Rule no. 1 is removed and accuracy is calculated. It is observed
that the model produces 84.61% accuracy after removing Rule no. 1 and hence Rule
no. 1 is pruned permanently. It is observed that the model produces 81.71% accuracy
when Rule no. 2 is removed and hence Rule no. 2 is not pruned. Similar process is
done for remaining rules and Table 3 is obtained after one complete iteration.
In the next iteration, Table 3 is treated as input and similar process is done for
rule pruning. Finally, Table 4 is obtained by sequential hill climbing procedure and
is treated the absolute transparent rule set.
The accuracy of the absolute transparent rule set is 85.52%.
Rule Merging Table 4 gives a minimized rule set after rule pruning phase. So, in
the rule merging phase, the minimized rule set is merged to form a single rule but the
features x20 have values with different ranges. The particular feature value is pruned
one by one if their absence from the rule increases the accuracy. Feature x20 > 0.143
and feature x20 > 0.207 are pruned as by dropping individually it rises the accuracy
of classification of positive patterns. Ultimate rule is shown in Table 5.
Risk Factor Identification In the single merged rule obtained in Sect. 4.2, it can
be seen that the rule contains seven features x19, x10, x20, x15, x4, x22, x13. In this
Table 3 Rule set after first

(2) if (x19>−6.318 and x10>0.19 and x20>0.143) then
iteration
Class==1
(3)if (x20>0.208 and x15>0.006) then Class==1
(4)if (x22>0.134 and x13>0.02) then Class==1
(5) if (x19>−6.318 and x20>0.207 and x4>0.00) then Class==1
(6) if (x22>0.134 and x12>0.013 and x13>0.019) then Class==1
Class==1
Table 4 Absolute
1 if (x19>−6.318 and x10>0.19 and x20>0.143) then
transparent rule set
Class==1
2 if (x20>0.208 and x15>0.006) then Class==1
3 if (x19>−6.318 and x20>0.207 and x4>0.00) then
Class==1
4 if (x22>0.134 and x13>0.02) then Class==1
Table 5 Ultimate rule

If((x19>−6.318 and x10>0.19) or (x20>0.208 and x15>0.006)
or (x19>-
6.318 and x4>0.004) or ((x22>0.134 and x13>0.02)) then
Class=1
Table 6 Misclassification rates (in %) when condition reversed (one attribute)

Testing set X19 X10 X20 X15 X4 X22 X13
Testing set 1 85.0 95.0 75.0 70.0 85.0 85.0 100
Testing set 2 75.0 90.0 60.0 85.0 85.0 85.0 100
Testing set 3 60 85.0 80.0 65.0 85.0 85.0 80.0
Testing set 4 80 95.0 90.0 100 95.0 100 100
Testing set 5 90.0 90.0 100 100 100 90.0 90
Testing set 6 68.42 63.16 47.37 63.16 63.15 63.16 68.42
Testing set 7 63.16 73.68 68.42 73.68 78.95 84.21 94.74
Testing set 8 100 100 94.74 94.74 100 100 100
Testing set 9 78.95 73.68 73.68 84.21 89.47 94.74 68.42
Testing set 10 57.89 63.16 15.79 78.95 84.21 73.68 57.89
Average 75.84 82.86 70.5 81.47 86.57 86.07 85.94
step, the misclassification rate of each feature is checked by reversing its individual
ranges. The higher the misclassification rate the more important is the feature. The
misclassification rates are shown in the table below.
It is observed from Table 6 that x4 (MDVP jitter in percentage) has the highest
misclassification rate, hence this feature is considered as the major risk factor for
PD. So if it is the case of selecting a single major risk factor responsible for PD, then
x4 is selected as the one. By controlling this single feature, PD can be managed up to
86.57%. This process can be further continued to find more risk factors which may
be used control PD.
Performance and Comparison of B-TESM-PD with Simple Decision Tree

In addition to accuracy, comprehensibility of the model is also taken into consider-
ation for comparison with simple DT and the same model without using the prepro-
cessing step. Global comprehensibility, which is the total number of deciding rules
and local comprehensibility is nothing but the total number of features in the deciding
rules.
The experimental results of Table 7 show the performance of the B-TESM-PD
with the simple decision tree model.
Table 7 Performance measures (accuracy in %, comprehensibility)

Methods Accuracy Global comprehensibility Local comprehensibility
B-TESM-PD 85.52 4 7
B-TESM-PD (without the 84.10 4 10
preprocessing step)
Decision tree 77.86 7 23
4.3 Observations
The B-TESM-PD calculates the rate of misclassification of individual feature by

reversing its ranges for the positive patterns only. It identifies the feature with highest
misclassification rate as the major risk factor for PD. It is observed that by controlling
that particular feature PD can be prevented up to 86.57%. The prevention rate may
increase with the increase in the number of risk factors. Although the performance
increases with increase in risk factor(s) but at the same time it may be difficult to
control more factor(s) responsible for PD. Moreover, from the analysis it is obvious
that, even if it is the identification of one risk factor, the disease can be managed
significantly. B-TESM-PD is further compared with a simple decision tree and it is
clearly evident that the classification accuracy of the B-TESM-PD is much higher
than the simple decision tree along with a very high rate of comprehensibility both in
terms of global and local. For classification accuracy, both the positive and negative
instances are taken into consideration.
5 Conclusion
This paper proposes a rule generation method called B-TESM-PD, to determine

important and significant rules and to make a rich transparent system and thereby
identifies the major risk factors of PD. The proposed B-TESM-PD extracts the deci-
sion rules from the decision tree which are further pruned to increase the transparency
of the rule set without sacrificing accuracy. Later a sequential hill climbing method
is proposed to prune the rule set to obtain a highly transparent system. Performance
of the proposed B-TESM-PD system is measured by global and local comprehen-
sibility, accuracy, precision, recall, and f-measures. Performance of the proposed
B-TESM-PD is evaluated with simple decision tree. The overall performance of the
B-TESM-PD system is much better than decision tree in terms of accuracy, preci-
sion, f-measures, and comprehensibility. The transparent rule finally obtained may
contain some attributes or factors which are common with same or different ranges
of deciding values. The major factors for PD have been identified by merging and by
calculating the misclassification rate of individual features for the positive patterns
by reversing its ranges.
Thus, it can be concluded that the B-TESM-PD is a rich transparent system with
less decision rules which makes a system efficient, user convincing, and manageable
to a great extent for easy prevention of PD and take necessary precautionary measures
to control the risk factor thus identified.
The proposed model uses decision tree for rule generation; however, some other
machine learning algorithms such as neural network can be used for generating
production rules. Moreover, some other optimization techniques can be used in
place of hill climbing algorithm in future for pruning the rule set and to get more
transparency in the rule set.
References
1. Standaert DG, Saint-Hilaire MH, Thomas CA (2015) Parkinson’s disease handbook. American
Parkinson Disease Association, New York, USA
2. Lang AE, Lozano AM (1998) Parkinson’s disease. New Zealand J Med 339:1044–1053
3. Elbaz A, Bower JH, Maraganore DM, McDonnell SK, Peterson BJ, Ahlskog JE, Schaid DJ,
Rocca WA (2002) Risk tables for parkinsonism and Parkinson’s disease. J Clin Epidemiol
55:25–31
4. Van Den Eeden SK, Tanner CM, Bernstein AL, Fross RD, Leimpeter A, Bloch DA, Nelson
LM (2003) Incidence of Parkinson’s disease: variation by age, gender and race/ethnicity. Am
J Epidemiol 157:1015–1022
5. Yuvaraj R, Muragappan M, Ibrahim NM, Sundaraj K, Omar MI, Mohamad K, Palaniappan R
(2014) Detection of emotions in Parkinson’s disease using higher order spectral features from
brain’s electrical activity. Biomed Signal Process Control 108–106
6. Davie CA (2008) A review of Parkinson’s disease. Br Med Bull 86(1):109–127
7. Sutherland M, Dean P (2017) What is Parkinson’s disease? Neuro challenges. Foundation for
Parkinson’s. http://www.parkinsonsneurochallenge.org
8. Jankovic J (2008) Parkinson’s disease: clinical features and diagnosis. J Neurol Neurosurg
Psychiatry 79(4):368–376
9. Organization WH (2011) Global health and ageing. Technical report. US National Institute of
Aging
10. Tomek I (1976) Two modifications of CNN. IEEE Trans Syst Man Cybern 769–772
11. Bisaglia M, Filograna R, Beltramini M, Bubacco L (2014) Are dopamine derivatives implicated
in the pathogenesis of Parkinson’s disease? Ageing Res Rev 13:107–114
12. Stuart S, Alcock L, Galna B, Lord S, Rochester L (2014) The measurement of visual sampling
during real world activity in Parkinson’s disease and healthy controls: a structured literature
review. J Neurosci Methods 222:175–188
13. Yuvaraj R, Muragappan M, Ibrahim NM, Sundaraj K, Omar MI, Mohamad K, Palaniappan
R. Optimal set of EEG features for emotional state classification and trajectory visualization
in Parkinson’s disease. Int J Psychophysiol 94(3):482–495. http://doi.org/10.1016/j.ijpyscho.
2014.07.014
14. Wang A, Costello S, Cockburn M, Zhang X, Bronstein J, Ritz B (2011) Parkinson’s disease
risk from ambient exposure to pesticides. Eur J Epidemiol 547–555
15. Lee H, Guan L, Lee I (2008) Video analysis of human gait and posture to determine neurological
disorders. EURASIP J Image Video Process 1–12
16. Steubner E, Vichayanrat E, Low DA, Mathias CJ, Isenmann S, Haensch C-A (2013) Twenty-
four hour noninvasive ambulatory blood pressure and heart rate monitoring in Parkinson’s
disease. Front Neurol 4
17. Ramani RG, Sivagami G (2011) Parkinson’s disease classification using data mining algo-
rithms. Int Comput Appl 32(9): 17–22
18. Sama A, Perez-Lopez C, Rodriguez-Matrin D, Catala A, Moreno-Arostegui JM, Cabestany J,

de Mingo E, Rodriguez-Molinero A (2017) Estimating bradykinesia severity in Parkinson’s
disease by analysing gait through a waist-worm sensor. Comput Biol Med 84:114–123
19. Kim HB, Lee WW, Kim A, Lee HJ, Park HY, Jeon HS, Kim SK, Jeon B, Park KS (2018) Wrist
sensor-based tremor severity quantification in Parkinson’s disease using convolutional neural
network. Comput Biol Med 95:140–146
20. Joshi D, Khajuria A, Joshi P (2017) An automatic non-invasive method for Parkinson’s disease
classification. Comput Meth Progr Biomed 145:135–145
21. Little MA, McSharry PE, Roberts SJ, Costello DA, Moroz IM (2007) Exploiting nonlinear
recurrence and fractal scaling properties for voice disorder detection. Biomed Eng Online
6(1):23
22. Arjmandi MK, Pooyan M (2012) An optimum algorithm in pathological voice quality assess-
ment using wavelet-packet-based features, linear discriminant analysis and support vector
machine. Biomed Signal Process Control 7(1):3–19
Bioinformatics Advancements
for Detecting Epidemic Disease Using
Machine Learning Approaches
Bikash Baruah and Manash Pratim Dutta
Abstract In the twentieth century, many researchers have started working on bioin-
formatics for disease biomarker detection using genetic information, i.e., DNA
microarray dataset and RNA sequencing dataset with machine learning approaches.
The journey of this concept starts with the classification technique on DNA
microarray dataset by comparing it with reference genome or by deNovo (without
reference genome) technique, and lots of different tools were published in different
publications. Later, with the availability and advancement of computational power
many researchers started working on large RNA sequencing dataset and some tools
are published again with significant features. Nowadays, also this area is like a
newborn baby and several challenges are still not solved, but it does not have a proper
guideline for new researchers to face those challenges. After analyzing so many tools
on DNA as well as RNA, we are able to summarize these works with a common
workflow, and in this paper, we have proposed a generalized workflow for detecting
epidemic diseases like HIV-AIDS, Cancer using machine learning approaches.
Keywords DNA microarray · RNA sequencing · Genome · Sanger · NGS ·

Differential co-expression
1 Introduction
A recent advancement of bioinformatics [1, 2] is in trend where machine learning

approaches are used on DNA microarray [3–6] and RNA-seq dataset [7–10] to iden-
tify the progression of epidemic diseases. The effectiveness and reliability of this
approach are far better than the traditional techniques. Researchers are working
continuously to develop cost-effective and robust algorithms. To obtain the input
B. Baruah (B) · M. P. Dutta

Department of Computer Science and Engineering, National Institute of Technology Arunachal
Pradesh, Papum Pare, India
M. P. Dutta
1052 B. Baruah and M. P. Dutta
sequence, basically there are two sequencing methods, i.e., Sanger [11, 12] and NGS
[13, 14], and these can be applied either in extracted DNA or RNA of any living
being and result obtained will be microarray and RNA-seq, respectively. Now co-
expression analysis [15–17], differential expression analysis [16], and differential
co-expression analysis [18–22] or hybridized analysis (combination of these anal-
yses) can be used through bi-clustering [23–25] or tri-clustering [26] techniques to
detect the disease biomarker [27]. Here, we are proposing one generalized workflow
which is fitted in almost all researches of this area. Further, we will extend our work
to design robust and cost-effective algorithm to apply Cancer and HIV progression
human dataset to identify the highly affected genes.
2 Proposed Model
We have proposed a model given in Fig. 1 which gives a complete workflow starting
with DNA and RNA extraction from living cells or tissues followed by sequencing
and co-expression analysis. In each step, we try to explain different available tech-
niques. The workflow discussed in this model will certainly help the new researchers
of this field, because it was never explained before in such a simple, systematic and
step-by-step manner how wet lab and dry lab processes are combined together for
detecting disease affected genes. Once sample extraction followed by sequencing is
being completed in wet lab, the output of sequencing is taken as the input for dry lab
for data analysis. In the following sections, different modules are explained.
2.1 Sample Extraction
Sample can be of two types: DNA and RNA and their extraction process from living
or conserved cells, tissues, or virus particles are also different. Though, nowadays,
many advanced kits are available for high-quality DNA and RNA extraction [28],
the basic steps are almost similar to each other.
2.1.1 DNA Extraction Procedure
Step 1 Cell lysis to release the DNA.

Step 2 Centrifuge the sample to separate the DNA from other cellular debris and
proteins.
Step 3 Use chilled isopropanol to precipitate the DNA.
Step 4 Wash DNA properly with ethanol.
Step 5 Gel electrophoresis for quality and quantity check of DNA.
Bioinformatics Advancements for Detecting Epidemic Disease … 1053
Sample Extraction
Select
One Sanger NGS
Dataset
to
Analyse
DNA RNA
Microarray Sequencing
Dataset Dataset
Sequence Analysis Method
Choose
One
Analysis
Method
CoExpression Differential Differential
Analysis Analysis CoExpression
Analysis
Detection of Disease Biomarker as a Cluster

(If Present)
Fig. 1 Proposed model

2.1.2 RNA Extraction Procedure
Step 1 Cell lysis and dissolution.

Step 2 Denaturation of proteins and DNA.
Step 3 RNases inactivation.
Step 4 Removal of cellular components.
Step 5 Precipitation of RNA.
Step 6 Gel electrophoresis for quality and quantity check of RNA.
Once, high-quality sample is extracted from cells, tissues, or virus particles; it

becomes ready for sequencing either by Sanger or NGS.
2.2 Sanger Sequencing
Frederick Sanger and his colleagues developed Sanger sequencing in 1977, which
is known as the first generation sequencing. In complete Human Genome Project,
Sanger sequencing is used and completed in 2003. The output of Sanger sequencing
gives high-quality data with low noise and robustness. Sanger sequencing method

uses dideoxynucleotides (ddNTP) with a hydrogen atom instead of 3 hydroxyl group
to sequence the deoxyribonucleic acid (DNA). These modified ddNTPs are able to
terminate the polymerization of DNA. Here, DNA sample is divided into four sepa-
rate samples and each of the four samples contains DNA polymerase and deoxynu-
cleotides (dATP, dGTP, dCTP, and dTTP). In each sample, one of the four dideoxynu-
cleotides (ddATP, ddGTP, ddCTP, or ddTTP) is added. When the bases of unknown
strand bind with its complementary ddNTPs, the polymerization will stop as ddNTPs
and fragments of DNA are produced. Further, the samples from all four vessels are
collected and determine their size by agarose gel electrophoresis. Unknown sequence
is determined by arranging the size of fragments from lowest to highest in 5 to
3 order. Many copies of DNA fragments with different lengths are generated to
compute the final DNA Microarray dataset which will be further analyzed by applying
expression analyzing algorithm to detect the disease biomarker.
2.3 Next-Generation Sequencing
Next-generation sequencing (NGS), a more cost-effective and efficient sequencing

technology compared to Sanger sequencing, is used to sequence both DNA and
RNA. Different machines are available for next-generation sequencing (NGS), but
basically it follows three steps to complete the process.
2.3.1 Sample Preparation
It needs custom adapter sequence either by ligation or amplification. These adapter

sequences provide universal primers for library hybridization to the sequence chip.
2.3.2 Sequencing Through Machines
Each library fragment is amplified and attached with DNA linkers to hybridize the
adapters. This creates clusters of DNA, and each cluster is an individual sequencing
read.
2.3.3 Collect Output Data
At the end of sequencing, raw data in the form of reads will be available which can
further be analyzed to retrieve the more meaningful and informative result.
NGS can be implemented for both DNA microarray and RNA sequencing. In
microarray datasets, gene intensities are in normal distribution; whereas in RNA-
seq, it follows either Poisson or negative binomial distribution. The major advantages
of RNA-seq over microarray dataset are: DNA microarray has very less sensitivity
to gene expressions compared to RNA-seq dataset. RNA-seq can measure approx-
imately 70,000 non-coding [29] RNAs which have an important role in disease
biomarker detection; but, it is not possible in microarray.
2.4 Co-expression Analysis
Co-expression analyses are done generally in three steps.

Firstly, individual relationships among genes have to be calculated based on mutual
information on each pair of genes. These information are stored in a matrix to describe
the similarity or co-expression among the expression patterns of different genes
across all the samples. Let us consider an example of five gene co-expression matrix
as shown in Table 1.
In Table 1, we can see that maximum and minimum values are one and zero,
respectively, for completely identical and complete dissimilar gene pair. Different
ways of correlation measures, i.e., (1) Spearman’s or Pearson’s correlations [30, 31],
(2) least absolute error regression [32], and (3) Bayesian algorithm [33] can be used
to derive the co-expression matrix shown in Table 1. The values of Table 1 are only to
describe the pattern of a co-expression matrix. Bayesian algorithm and least absolute
error regression have the advantage to identify causal links.
Secondly, co-expression network has to be constructed using genes as nodes and co-
relation between the nodes as edges. Edge can be either weighted versus unweighted
Table 1 5 × 5 co-expression matrix

Gene1 Gene2 Gene3 Gene4 Gene5
Gene1 1 0.50 0.67 00 0.32
Gene2 0.50 1 0.39 0.80 0.45
Gene3 0.67 0.39 1 0.20 0.17
Gene4 00 0.80 0.20 1 0.78
Gene5 0.32 0.45 0.17 0.78 1
and signed versus unsigned. Thickness of edge shows the weight of the edge, and
the value lies between zero and one.
Weighted Versus Unweighted

Unweighted edged network is the simplest way of constructing co-expression
network where interaction between node pairs is binary, i.e., either 0 or 1. By consid-
ering the correlation of all gene pairs or node pairs above, a certain threshold to be
connected (i.e., 1) and all others be disconnected (i.e., 0).
In a weighted edged network, all nodes are connected to each other with a weighted
edge consists of continuous values determining the co-relation between the nodes
where value determines the strength. Weighted edge can be of two types: signed and
unsigned edges.
Signed Versus Unsigned

In a signed network, edge correlation values lie between −1 (perfect negative corre-
lation) and 1 (perfect positive correlation). An unsigned edge network assigns the
correlation values between 0 and 1 so that values less than 0.5 indicate negative
correlation and values greater than 0.5 indicate positive correlation.
Thirdly, co-expressed genes are clustered using bi-clustering or tri-clustering tech-
nique to group the genes with similar expression patterns across multiple samples.
Some well-known clustering algorithms are K-means clustering [34], hierarchical
clustering [34], THD-tricluster [26], shifting-and-scaling correlation clustering [35],
etc. The clusters can be interrogated to identify regulators, functional enrichment,
and hub genes for a potential disease gene by using guilt-by-association (GBA) [36]
approach, while differential co-expression analysis gives the advantage of comparing
modules in different conditions for better identifying disease regulators.
2.5 Differential Expression Analysis
Differential analysis has been done by comparing gene expression datasets of

different conditions. For disease detection, minimum two samples have to be consid-
ered; one dataset of normal or healthy conditions and another in unhealthy or disease
affected conditions. Different statistical tests like t-test, z-test, chi-square test are
applied to analyze whether expressions are in up-regulation (disease is growing)

or down-regulation (in control) states. In unhealthy conditions, if more number
of samples with a fixed interval can be collected, then differential analysis gives
more informative result. Bi-clustering and tri-clustering techniques are used for
two datasets and more than two datasets, respectively, to group the genes showing
responses in the same conditions.
2.6 Differential Co-expression Analysis
Differentially co-expressed analysis is to identify the patterns of correlated gene

expression in different conditions. It will always give a more informative picture of
the dynamic changes in the gene regulatory networks by comparing the transcrip-
tome of same genome in two conditions. For example, one cluster of genes strongly
correlated in one condition may no longer be strongly correlated in another condition.
Hence, differential co-expression gives high response to potential disease adaptation
in different environments. Differential co-expression analysis can be done in three
ways:
2.6.1 Targeted Differential Co-expression
Differential co-expression analysis starts with targeted approach. In general, it is

completed in three steps.
Firstly, pre-defined clusters are being surveyed with known annotation file to analyze
in different conditions.
Secondly, correlation among genes of individual clusters as well as the correlation
within group of clusters has been derived by using the gene correlation expression.
Finally, comparison is done between gene co-expression values in multiple environ-
mental conditions.
2.6.2 Untargeted Differential Co-expression
Untargeted differential co-expression is the latest approach among all bioinformatics

sequence analyzer. It is also done in three steps. Unlike targeted in the first step,
correlated genes have to be detected which shows different significant behavior in
different conditions. Once clustering is completed, rest two steps are similar with
targeted approach. In 2009, Southworth et al. [5] applied this approach for the first
time which is based on purely untargeted approach for detecting the mice genetic
modules correlation with respect to age.
2.6.3 Semi-targeted Differential Co-expression
It is somewhat in between targeted and untargeted, where pre-defined clusters with

partial annotation files are used. A strong disadvantage of semi-targeted approach is
that it only concerns with those genes which emerge with clusters at least in anyone
different environmental conditions.
2.7 Cluster Detection
Once the analysis is being completed by using any of the analysis methods, viz.
traditional co-expression analysis, differential expression analysis, or differential
co-expression analysis, the affected genes will be discovered. Its efficiency depends
on the effectiveness of the algorithm designed by the researchers. Then, these genes
are clustered as a module so that this cluster can be used in further drug design.
3 Conclusion
In this paper, we have tried to explain a workflow in a sequential manner for detecting
epidemic diseases affected genes using different bioinformatics advancements. In
future, we are going to implement these approaches on different DNA, RNA samples
to detect Cancer and HIV-AIDS affected genes and will try to cluster them separately,
so that our result can help the drug designers at genetic level.
References
1. Tomasz P, Szymon W, Jacek B (2016) Computer representations of bioinformatics models.

Curr Bioinform 11(5):551–560
2. Agbachi CPE (2017) Pathways in bioinformatics: A window in computer science. Int J Comput
Trends Technol 49(2):83–90
3. Sardaraz M, Tahir M, Ikram AA (2016) Advances in high throughput DNA sequence data
compression. J Bioinf Comput Biol 14(3):18
4. Ge SX (2017) Exploratory bioinformatics investigation reveals importance of junk DNA in
early embryo development. BMC Genom 18(1):200
5. Chen S, Liu M, Zhou Y (2018) Bioinformatics analysis for cell-free tumor DNA sequencing
data. In: Computational Systems Biology. Humana Press, New York, NY, USA, pp 67–95
6. Zhang J, Huang K (2017) Pan-cancer analysis of frequent DNA come thylation patterns reveals
consistent epigenetic landscape changes in multiple cancers. BMC Genom 18:1045
7. Van Dam S, Craig T, de Magalhaes JP (2015) Gene friends: a human RNA-seq-based gene and
transcript co-expression database. Nucl Acids Res 43:1124–1132
8. Zeisel A, Munoz-Manchado AB, Codeluppi S et al (2015) Brain structure cell types in the
mouse cortex and hippocampus revealed by single-cell RNA-seq. 347:1138–1142
9. Fiannaca A, La Rosa M, La Paglia L et al (2015) Analysis of miRNA expression profiles in

breast cancer using biclustering. BMC Bioinform 16
10. Xue Z, Huang K, Cai C et al (2013) Genetic programs in human and mouse early embryos
revealed by single-cell RNA sequencing. Nature 500:593–597
11. Sanger F (1980) Google Scholar. http://www.nobelprize.org/nobel_prizes/chemistry/laureates/
1980/sanger-bio.html
12. Hutchison C (2007) DNA sequencing: bench to bedside and beyond Nucleic Acids. Nucl Acids
Res. 35:6227–6237
13. Mardis ER (2008) The impact of next-generation sequencing technology on genetics. Trends
genet. Trends Genet 24:133–141
14. Shendure J, Ji H (2008) Next-generation DNA sequencing. Nat Biotechnol 26:1135–1145
15. Glass K, Huttenhower C, Quackenbush J (2013) Passing messages between biological networks
to refine predicted interactions. PLoS One 8
16. De Smet R, Marchal K (2010) Advantages and limitations of current network inference
methods. Nat Rev Microbiol 8:717–729
17. Yue F, Cheng Y, Breschi A et al (2014) A comparative encyclopedia of DNA elements in the
mouse genome. Nature 515:355–364
18. Amar D, Safer H (2013) Dissection of regulatory networks that are altered in disease via
differential co-expression. PLoS Comput Biol 9
19. Zeisel A, Munoz-Manchado AB, Codeluppi S et al (2015) Brain structure cell types in the
mouse cortex and hippocampus revealed by single-cell RNA-seq. Science 347:1138–1142
20. Bhar A, Haubrock M, Mukhopadhyay A et al (2013) Coexpression and coregulation analysis
of time-series gene expression data in estrogen-induced breast cancer cell. Algor Mol Biol 8
21. Fiannaca A, La Rosa M, La Paglia L (2015) Analysis of miRNA expression profiles in breast
cancer using biclustering. BMC Bioinform 16
22. Kakati T, Bhattacharyya DK, Barah P, Kalita JK (2019) Comparison of methods or differential
co-expression analysis for disease biomarker prediction. Comput Biol Med 10:100–103
23. Madeira SC, Oliveira AL (2004) Biclustering algorithms for biological data analysis: a survey.
IEEE/ACM Trans Comput Biol Bioinform 1:24–45
24. Wang YK, Print CG, Crampin EJ (2013) Biclustering reveals breast cancer tumour subgroups
with common clinical features and improves prediction of disease recurrence. BMC Genom
14:102
25. Oghabian A, Kilpinen S, Hautaniemi S, Czeizler E (2014) Biclustering methods: biological
relevance and application in gene expression analysis. PloS One 9
26. Kakati T, Kashyap H, Bhattacharyya DK (2016) THD-module extractor: an application for
CEN module extraction and interesting gene identification for Alzheimer’s disease. Sci Rep 6
27. Kakati T, Bhattacharyya DK, Barah P, Kalita JK (2019) Comparison of methods for differential
co-expression analysis for disease biomarker rediction. Comput Biol Med 10:113
28. Tan SC, Yiap BC (2009) DNA, RNA, and protein extraction: the past and the present. Hindawi
Publ Corp J Biomed Biotechnol Article ID 574398
29. Zhao Y, Li H, Fang S et al (2016) NONCODE 2016: an informative and valuable data source
of long non-coding RNAs. Nucl Acids Res 44:203–208
30. Guttman M, Donaghey J, Carey BW et al (2011) lincRNAs act in the circuitry controlling
pluripotency and differentiation. Nature 477
31. Ala U, Piro RM, Grassi E et al (2008) Prediction of human disease genes by human-mouse
conserved coexpression analysis. PLoS Comput Biol 4
32. van Someren EP, Vaes BL, Steegenga WT et al (2006) Least absolute regression network
analysis of the murine osteoblast differentiation network. Bioinformatics 22:477–484
33. Friedman N, Linial M, Nachman I et al (2000) Using Bayesian networks to analyze expression
data. J Comput Biol 7:601–620
34. Haeseleer PD (2005) How does gene expression clustering work? Nat Biotechnol 23:1499–
1501
35. Ahmed H, Mahanta P, Bhattacharyya D, Kalita J (2014) Shifting-and-scaling correlation based
biclustering algorithm. IEEE/ACM Trans Computat Biol Bioinform 11:1239–1252
36. Chu S, DeRisi J, Eisen M, Mulholland J, Botstein D, Brown PO, Herskowitz I (1998) The
transcriptional program of sporulation in budding yeast. Science 282:699–705
A Strategic Community Control-Based
Power Flow Between Grid-Integrated PV
Houses
Jatin M. Soni, Deep V. Patel, Rajendra V. Patel, and Hardik P. Modha
Abstract This paper proposes a controlled power flow between interconnected four
photovoltaic (PV) houses and also in grid-connected mode. The voltage and current
schemes are used in PV houses. We have applied Perturb and Observe method to
obtain maximum power. We have connected battery energy storage system (BESS)
at each house by three-phase inverters. When all PV houses are connected, generated
power from each PV is supplied to a load of each house. When a load of any house is
disconnected, this power is utilized in another house. If there is an excess of power,
it is fed to the utility grid. This scheme is validated in MATLAB/SIMULINK in
various load condition. The power quality of this scheme is as per the IEEE standard.
Keywords PV rooftop · BESS · Three-phase inverter · Bidirectional energy

meter · Utility grid
1 Introduction
The consumption of electricity is continuously increasing. But fossil fuels are

decreasing day-by-day. Therefore, the renewable power sources like solar photo-
voltaic (SPV), wind turbine, geothermal are accepted as alternative source. SPV has
gained most attention for research due to its more advantages [1]. SPV is used as
rooftop for low power demand or as large solar farm for high power demand. The
power at output terminal of SPV varies with solar irradiance and temperature. PV is
used in islanded mode or in grid-connected mode [2]. The controller is planned such
J. M. Soni (B) · D. V. Patel · R. V. Patel · H. P. Modha

Department of Electrical Engineering, Ganpat University, Kherva, Gujarat, India
D. V. Patel
R. V. Patel
H. P. Modha
1062 J. M. Soni et al.
that SPV is synchronized with grid to keep stable system and to supply voltage at
desired frequency [3]. Battery energy storage system (BESS) is used to deliver power
in shortage of SPV power. In [4], the control of BESS with non-conventional sources
is proposed with considering its state-of-charge (SOC). To get maximum power from
SPV, maximum power point tracking (MPPT) is used. Perturb and Observe (P&O)
method is used. In [5], the algorithm for MPPT is developed. The three-phase inverter
is used to interconnect PV houses and grid. This inverter requires pulse width modu-
lation (PWM) gate pulses. Inverter gives proper control of voltage and power, system
balancing and quality of power conversion.
The BESS is connected to PV and grid by bidirectional three-phase inverter. It
is used for controlling both charging and discharging state. In [6], reactive power
control method is given. The active power is controlled by controlling disparity
between inverter output voltage and point of common coupling (PCC) voltage [7].
PV houses are interconnected for sharing and utilizing power of other houses. The
excess of power is given to utility grid by connecting it. And if there is less SPV
power generation, the deficit power is supplied by the grid [8]. The bidirectional
energy meter is installed at each PV house terminal. It gives three information: (1)
Status of meter (2) How much power is delivered to other house or to grid (3) How
much power is delivered from other houses or grid [9]. There is also caterpillar line for
indication of direction of power flow. In [10], cluster of PV houses is interconnected.
But it does not give information on excess of power and where to use it. This paper
has implemented that issue by integrating it with grid.
This paper is prepared as along this path: Sect. 2 gives system configuration and
its modeling, Sect. 3 presents control policy for voltage and current control, Sect. 4
gives information about the result attained in different load condition, and Sect. 5
gives conclusion of this paper.
2 System Configuration and Its Modeling
2.1 System Configuration
The four PV houses are interconnected and also connected to utility grid. Each PV
houses has bidirectional energy meter at its terminal and PV rooftop system. The
main architecture of grid-integrated PV houses is shown (see Fig. 1). Community
control panel (CCP) does the power calculation of each house. Each PV house has
loads, BESS, three-phase inverter and bidirectional energy meter. SPV and BESS
are connected to its load and to grid by unidirectional three-phase inverter and bidi-
rectional three-phase inverter, respectively (see Fig. 2). The lead-acid battery is used
which has rating of 5 kW, 24 * 2 = 48 V, 100 Ah. The load of each house is 12 kW.
A Strategic Community Control-Based Power Flow … 1063
Fig. 1 Main architecture of grid-integrated PV houses
Ppv To Grid
DC Rf Lf
Bi-directional
meter
PV array Cf
AC
Three phase Filter
Uni-directional Pbatt
Inverter
AC Load
DC Rb Lb
Battery Energy storage
system (BESS)
AC Cb
Three phase Filter

Bi-directional
Inverter
Fig. 2 System configuration in each PV house
2.2 System Modeling
Modeling of PV system and MPPT The inner structure of PV cell is shown (see
Fig. 3). There is current source with anti-parallel diode [7]. Output voltage is low
Fig. 3 Internal construction

of PV cell
due to voltage drop in series resistance Rs and leakage resistance Rsh . The desired
output voltage and current is done by connecting cell in series and parallel.
Output voltage of PV system is taken from [8] is as follows:
Ipv = Iph ∗ ID −(Vd /Rsh ) (1)

ID = I0 ∗ eVd /VtA − 1 (2)
VD = Vpv + Ipv Rs (3)
Iph = λ ∗ (Isc + A(T − Tref )) (4)

Vpv = Ns (AkT /q) ∗ ln Np Iph − Ipv + Np I0 /Np I0

− Ns /Np Ipv Rs (5)
where Rse —resistance connected in series, I PV —PV output current, q—electron

charge, V PV —obtained output voltage at PV terminal, nse —quantity solar cell series
connected, np —quantity solar cell parallel connected, K—constant (Boltzmann),
A—factor of ideality, T cell —temperature in operation mode, I sat —current when it
becomes saturated, I ph —photocurrent, Rsh —resistance connected in parallel, I sc —
current in short circuit, α—coefficient (temperature), q—charge of electron, T ref —
temperature (reference) and E G —energy gap in PV cell.
MPPT is used to get maximum power. The different MPPT methods are compared
in [10]. Perturb and Observe (P&O) method is used because of easy to implement.
Diagram of P&O method is shown in Fig. 4. We perturb voltage V, then we examine
variation in power P. When we boost voltage in one way, there is raise in power
P. We continue this process until we obtain point where voltage increases, power
decreases. It is named as maximum power point. Vdcref is the reference value of Vdc .
Modeling of BESS system The voltage equations in charging and discharging mode
for battery are given [10]. VBatt, Ch is charging output voltage, VBatt, DCh is discharging
output voltage, V Batt is constant battery voltage, i* is filtered current, Q is Ah capacity,
and Exp(t) is exponential time zone.
Start
Measure Vpv ( k ) and

Ipv ( k )
Ppv ( k )= Vpv ( k )* Ipv ( k )
a = -1 a=1
Vdc _ ref ( k + 1 )= Vpv ( k )+ a
Fig. 4 Diagram of MPPT
VBatt, Ch = VBatt,0 − R1 i Batt − k Q i ∗ /(i Batt + 0.1Q)

− k Q/(Q − i Batt i Batt t + Exp(t) (6)
VBatt, Dch = VBatt,0 − R1 i Batt − k Q i ∗ /(Q − i Batt t)

− k Q i Batt t/(Q − i Batt t) + Exp(t) (7)
3 Control Strategy
3.1 SPV and Its Controller
The block diagram of the PV system coupled to load and integrated with the grid is
presented in Fig. 5. AC voltage (inverter output) is converted by Park transformation
(ABC to DQ transformation). V d , V q , I d and I q , which are obtained from ABC to DQ
(Park transformation) of inverter output voltage, is given to the voltage and current
Pac
Lf
Capacitor 3 Phase Inverter Rf
AC Side
PV Array C (Unidirectional)
Vdc Cf
PLL
Gate Pulse
Filter
PWM Abc to dq
transformation
Maximum power Vdc_Ref
point tracking Voltage and
Vdc
Current controller
Vd, Vq and Id, Iq
Fig. 5 Block diagram of PV system
controller. To get high power and to control active power given to main grid, a voltage
controller is used.
The block diagram of voltage controller is shown in Fig. 6a. According to power
balance theory, DC link voltage is as follows:
0.5 ∗ C ∗ (dV 2 /dt) = PPV − Pdc (8)
Pdc is power at DC side of inverter. If inverter is lossless, Pdc is equal to Ppv .
a
PI Idref
+ +
- Controller +
b Vd Vd
+ PI +
PI Iq_ref + + Division mq
Idref +
Controller
+ Division md
- Controller -
- -
Id Vd/2 wLId Vd/2

Iq
wLId
Fig. 6 aVoltage controller, b current controller

Pac = 1.5 ∗ (VTd i d + VTq Iq ) (9)
VTq should be zero by using PLL. Equation for I dref is achieved by using feedback
linearization technique (FDL).
2Ppv
Idref = u v + (10)
3VTd
It produces the I dref which is given to the current controller as a reference signal.
The control scheme in the current controller is presented in Fig. 6b. Output signals
md and mq are converted into ma , mb and mc signal by inverse Park transformation.
Then, it is compared with a 2 kHz triangular wave to get the SPWM signal.
3.2 BESS and Its Controller
A block diagram of the BESS system coupled to load is presented in Fig. 7. AC voltage
(output from bidirectional three-phase inverter) is converted by Park transformation
(ABC to DQ transformation). When we charge the battery, DC voltage at its terminal
is kept constant. When the battery is in discharging mode, the AC voltage at inverter
output terminal is kept constant. So, value of reference voltage is accepted as per its
mode of operation. The PLL is designed such that V q is zero. The BESS controller
is distributed in two categories: (1) voltage controller (2) current controller as shown
in Fig. 8a, b. DC voltage of battery is taken as reference in battery charging mode.
V d is taken as reference in battery discharging mode. In current controller, I qref = 0.
The SPWM signal is attained by the same process as discussed in the PV system.
Capacitor C
Vdc1
Energy
3 Phase Inverter
Rf Lf Pac
Storage (Bi-directional)
AC Side
Cf
Battery
PLL
Filter
Gate Pulse
PWM
Abc to dq
transformation
Vdc1_Ref
Voltage and Current
Vdc1 controller
Vd, Vq and Id, Iq
Fig. 7 Single line diagram of BESS system

a b
Charging
Iq_ref + PI
Vdc_ref mode mq
- Controller
+
PI
Controller + PI
Controller
md
- -
Vdc Dis-Charging
mode Iq
Vdc or Vd Id
Fig. 8 a Voltage controller, b current controller
4 Results
The fixed irradiance, fixed temperature, PV output power and BESS output power
are 1000 kW/m2 , 25 °C, 10.54 kW and 500 W, respectively, as shown in Fig. 9.
Output voltage after transformer is 400 VRMS voltage as shown in Fig. 10.
Case (1) When the load of one house is disconnected: Load of house 1 is
disconnected at t = 3 s shown in Fig. 11.
Case (2) When the load of two houses is disconnected: Load of house 2 is
disconnected at t = 5 s grid as shown in Fig. 11.
Case (3) When the load of three houses is disconnected: Load of house 3 is
disconnected at t = 7 s as shown in Fig. 11.
Case (4) When a load of all house is disconnected: Load of all houses is disconnected
at t = 9 s as shown Fig. 11.
Fig. 9 Irradiance, temperature and PV output power

Fig. 10 Output voltage after three-phase inverter
Fig. 11 Load power at each house
Fig. 12 Power supplied to grid

Fig. 13 FFT analysis at each house terminal. a Voltage, b current
The power supplied to grid in each case is shown in Fig. 12. The total harmonic
distortion (THD) analysis of the output voltage is also done at each terminal of the
house. And it is as per the IEEE standard as shown in Fig. 13.
5 Conclusion
In this paper, four PV houses are interconnected and also connected to the grid. The
load of each house is abounding by PV generation of that house. If load of house is
varied or disconnected, the PV power of that house is utilized in other houses. The
excess of power, if any, is supplied to grid. The THD of the output voltage and current
is as per IEEE 519 standard. So, the issue regarding excess of power is resolved by
integrating whole system with grid.
References
1. Mahmood H, Michaelson D, Jiang J (2014) A power management strategy for PV/Battery

hybrids system in islanded micro grid
2. Li H, Xu Y, Adhikari S, Rizy DT, Li F, Irminger P (2012) Real and reactive power power
control of a three phase single stage PV system and PV voltage stability. In: Power and energy
society general meeting, 2012, vol 2. IEEE, pp 1–8
3. Blaabjerg F, Teodorescu R, Liserre M et al (2006) Overview of control and grid synchronization
for distributed power generation systems. IEEE Trans Ind Electron 53(5):1398–1409
4. Mastromauro RA, Liserrer M, Dell’Aquila A (2012) Control issues in single stage photovoltaic
system: MPPT, current and voltage control. IEEE Trans Ind Inform 8(2):241–254
5. Hannan MA, Ghani ZA, Mohamed A, Uddin MN (2014) Real-time testing of a fuzzy logic
controller based grid-connected photovoltaic inverter system. In: Industry applications society
annual meeting, 5–9 Oct 2014. IEEE, pp 1–8
6. Delfino F, Denegri GB, Invernizzi M, Procopio R (2010) A control algorithm for the maximum
power point tracking and the reactive power injection from grid-connected PV system. In:
Power and energy society general meeting, 2010, vol 4, no 3, July 2010. IEEE, pp 1–7
7. Eghtedarpour N, Farjah E (2014) Distributed charge/discharge control of energy storages in a
renewable-energy-based DC micro-rid. IET Renew Power Gener 8(1):45–57
8. Jha RR, Srivastava SC (2016) Fuzzy logic and ANFIS controller for grid integration of solar
photo voltaic. In: 2016 IEEE 6th international conference on power systems (ICPS), New Delhi,
pp 1–6
9. Jha RR, Srivastava SC, Kumar M (2017) Development of control schemes for a cluster of
PV-integrated houses in islanded mode. IET Gener Transm Distrib J
10. Soni JM, Pandya MH (2018) Power quality enhancement for PV rooftop and BESS in islanded
mode. In: 2018 IEEE 4th international conference on electrical energy system (ICEES), vol 3,
Feb 2018, pp 242–247
Classification of Buildings and Vehicles
in Google Map Satellite Images Using
Random Forest Classifier
Ritik Gautam, Saravanan Chandran , and Jose Hormese
Abstract The Google Map provides an additional feature for observing various
places of the landscapes as bird’s eye view with the help of the satellite images.
Google Map satellite images are developed with the help of geographical information
systems (GIS) data, aerial images, drone images, and satellite images with the help
of image fusion methods to produce a bird’s eye view of a landscape. This proposed
research work is intended to classify the buildings and vehicles from the Google Map
satellite images that are captured at a zoom level of 20 m. This research work carried
out to survey an area for various applications which require number of buildings and
number of vehicles. The random forest classifier is used for the pixel classification,
and this technique is also referred as pixel segmentation. The random forest classifier
produces a result of 87% accuracy.
Keywords Google map · Satellite images · Random forest classifier · Building

segmentation · Vehicle segmentation
1 Introduction
Segmentation techniques are extensively used in the field of image processing and
computer vision for the applications of object detection and identification [1]. A
digital image is subdivided into a number of regions referred as segmented region.
The segmented region consists of large number of pixels that have similar properties
that represents the image. The segmentation of an image is used in several applica-
tions. The segmentation approach is used to identify both the buildings and vehicles
in a Google Map satellite images that provide the basis for the work. As urbaniza-
tion is on the increase, the significance of such satellite images is increasing which
can be utilized in diverse applications. Such diverse applications include homeland
security, effective resource management, emergency responses, and environmental
monitoring as well. These Google Map satellite images are normally captured through
R. Gautam · S. Chandran (B) · J. Hormese

Department of Computer Science and Engineering, NIT Durgapur, West Bengal, India
1074 R. Gautam et al.
high resolution satellites and have various applications especially in the areas like
agriculture, forestry, geology, surveillance, regional planning, and also in education.
Having such information helps in rural planning which benefits the residents and the
development of various public facilities.
Various approaches have been developed and proposed for applications like
extracting road networks [2–8], water regions, green regions, [9] identification of
buildings, and vehicles. During the unexpected natural disasters, the loss incurred to
assets and people is quite high. If an assessment is prepared in a short time during
the disaster period, it provides the technical details of the disaster, quantity of the
disaster, area affected by the disaster, etc. Thus, prompt decisions are initiated in
short time to control the disaster and start planning for the rescue operation, etc.
The significance of the satellite images originates from the fact that it is of high
resolution, covers a wide area, and is highly efficient in terms of time. Google Map
satellite images are very useful in the context of assessing the damages that are
caused in the event of an unexpected disaster. Identifying the impact of the damage
is possible to quickly initiate various necessary rescue actions. The field of remote
sensing has lot of importance in understanding the extent of land use which helps in
understanding the spatial and temporal features. This helps in providing high data
availability and computing the data which helps in better understanding the changes.
In this research work, classification of buildings and vehicles are carried out and
achieved 87% accuracy using random forest classifier algorithm.
2 Related Works
To know about the structure of terrestrial systems, the land cover details are essential.
The field of remote sensing generates terrestrial information in a global scale. A
methodology is framed for both discrete and continuous land cover that is based on
datasets which is time-based and depends on spatial resolution [10]. A method had
been proposed that detects eroded and non-forested areas that are tropical. They used
image algebra among the colour components that augmented the contrast mainly
between green and brown. This helped to differentiate between the areas that are
forest and the non-forested parts [11, 12]. Brezonik et al. [13] have specified that
images captured from Landsat satellites are useful for estimating various optical
characteristics of lakes and various measurements also considered. Prasad et al.
[14] have identified that the rapid expansion of urbanization has depleted the water
bodies and various infrastructural entities like buildings, roads, parks, and hubs are
spreading in a very quick manner. Goswami et al. [15] have specified the importance
of recognizing images from satellite images and that it comes under the classification
techniques. He proposed artificial neural networks (ANN) for the purpose of object
extraction as ANNs are suitable for extracting objects in an automated manner. The
data associated with land cover are used extensively in the fields like modelling
urban growth, agriculture based on advanced technology, management of coastal
regions, and also climate management [16]. Noticed that land cover altered as an
Classification of Buildings and Vehicles … 1075
effect of natural disasters like landslides, floods, cyclones, and forest fires. Human
interventions like urbanization, deforestation activities, and practices followed in
agriculture also lead to alteration in the land cover [17].
The changes that happen in the ecological and physical systems due to human
activities, understanding the land cover, help to understand both the ecosystem and the
climate models [18]. The conservation of natural resources and its management and
to formulate polices for urban and environmental development is essential. Poullis
and You [19] proposed a road detection method and classified the method into three
categories that are namely region-based, pixel-based, and knowledge-based. Line
detectors [20], [21] and ridge detectors [22], [23] are employed for the identification
and classification of road points. These interconnected road points are producing
road segments that are provided as input to the further processing level stages.
The techniques used in deep learning like convolutional neural networks (CNNs)
are showing better performance than ANNs as they overcome the pitfalls and they
produce better results when compared with machine learning techniques like support
vector machines and random forest [24]. Studies have shown that spectral values are
supplied in a vector format in one-dimensional format to the CNNs that is used for
classification of land cover, but this approach fails to avail full advantage of CNNs
that extracts spatial characteristics from images that are in two-dimensional form
[25].
3 Methodology
Three hundred Google Map satellite images are collected by cropping various loca-
tions. Two hundred images are selected for training, and one hundred images are
reserved for testing. The 200 images are marked with red colour for building and
green colour for vehicle. These marked images are trained using the random classifier
model. Fiji is used for pre-processing the images. The random classifier algorithm
is used for training and testing. The Weka has developed the model for classification
of building and vehicle. Further, the 100 images are tested using the model.
In a Google Map satellite image, segmentation techniques are used to identify
both buildings and vehicles. The image processing tools used for this segmentation
are Fiji and Weka. Random forest classifier is used for classification purpose of the
above-mentioned images. The classifier is being built with around 200 images in the
training set. The classifier is then loaded, and the results are created with a test set
of 100 images.
The practice of subdividing an image into various objects or primary parts is the
main purpose of segmentation. Once this is accomplished, the extracted regions are
applied to different feature extraction algorithms used in both image processing and
computer vision applications. This process depends on the application and grouping
of pixels based on the intensity values is helpful for separating the required and other
objects as well as other parts of the original image. Segmentation is useful for various
areas like detection of vehicles, recognition of faces, in the field of medical diagnosis,
and in various other fields as well. It has immense applications in pre-processing

steps for image compression techniques and identification of object boundaries.
Segmentation algorithms are classified into thresholding, clustering methods, region
growing methods, and histogram-based methods. Partitioning an image into different
regions that is similar and alike satisfies the similarity property while partitioning an
image into several regions that is based on changes in intensity values satisfies the
discontinuity property.
In the field of data science, different classification algorithms are available such
as random forest, support vector machine, decision trees, logistic regression, Naive
Bayes classifier, and few others. One of the frequently used classifiers in this hierarchy
is the random forest classifier. The basic idea is the formulation of decision trees,
and they are combined together to construct a typical random forest. Any random
forest representations are built up using decision trees. Decision trees use features
that help to split the data, and the resultant groups formed will be different in various
respects.
An open source package for image processing that is based on Java named Fiji is
used for this research work. The purpose of Fiji is to provide Java-based plugins that
are bundled. Users are provided with a detailed menu supported with broad docu-
mentation with tutorials and are robust thereby avoiding installation of different
components from various components. Developers use Fiji with version control
systems, numerous development channels, infrastructure for prototyping and support
numerous scripting languages.
The plugin initiates with two classes and hence producing the binary pixel clas-
sification. With the various tools for identifying the region of interest (ROI), users
have the freedom to add traces to these classes.
The training set consists of 200 images that are captured from the Google Map
satellite images taken at a resolution of 20 m zoom level. The random forest classifier
is trained based on this training set, and the classifier is made accordingly. The images
that are used for the training purposes are illustrated as follows. The following two
images displayed in Figs. 1 and 2 are also employed in the training process.
The two classes that are developed here are namely buildings and vehicles, respec-
tively. The ROI tools are used to indicate the regions present in both the classes from
the training images. The pixels that belong to the buildings are marked with red
colour, and the pixels corresponding to vehicles are marked with green colour. These
boundaries are marked for each class for each and every training image that is being
stacked from the training set in the Weka. The resultant classifier images are as
follows Figs. 3 and 4.
After the boundaries for each class are being marked for all the images in the
training set, as shown in the below images Figs. 5 and 6, the images are trained in
Weka. It is noted that the training time depends on the total number of pixels present
in each class. Once the training process is completed, the classifier model is saved
for testing other images.
Fig. 1 Google Map satellite image 1
Fig. 2 Google Map satellite image 2
After the classifier is build and then trained, the next step is to perform the testing
operations on the classifier developed. A sample test dataset of 100 images is consid-
ered for testing, and the results are replicated on these dataset. The classifier is loaded,
which was built previously and then click on the generate results option. The results
obtained are as shown above.
As mentioned earlier, buildings are mapped with red pixels, and the vehicles are
mapped with green pixels. The probability map is shown below where it represents
the probability of a pixel that belongs to a particular class that can either be buildings
or vehicles. It is noted that for classifiers, the performance has to be augmented with
the help of optimization of different parameters (Fig. 7).
Fig. 3 Weka classifier image 1
Fig. 4 Weka classifier image 2
The results obtained are after different stages of optimization techniques. A

commonly employed technique used in machine learning is Bayesian optimization.
When there are different hyper-parameters that deal with machine learning problems,
there arises tuning of the classifier which results in costly evaluations with respect to
time and computational resources, respectively. Bayesian optimization is useful to
overcome this short coming. A probabilistic model is built, and the posterior predic-
tive distribution is computed that integrates the various true functions. This leads to
an optimized proxy function that is cheaper when compared with the true objective.
Fig. 5 Weka test results image 1
Fig. 6 Weka test results image 2
The idea behind this approach is to exploit the amount of randomness to balance
the exploration. This approach may not give the best results with respect to neural
networks [26] because optimization that is based on a manual scheme is applied.
Fig. 7 Weka probability map of the test results
The random forest approach produces the value of accuracy as 0.87 which is
computed as the number of true predictions to the entire number of input images. The
specificity obtained is 0.91 which matches the negative samples wrongly identified
as positive to the whole set of incorrect samples. The precision value obtained is
0.93 which is equal to the proportion of correct positive samples to the whole set of
samples that are positive. The recall results come to 0.91 which is denoted as the
right positive samples that are predicted to the entire set of samples irrespective of
prediction. The F 1 measure that corresponds to the harmonic mean involving both
precision as well as recall is 0.92 which depends on the precision of the classifier.
The geometric mean comes out to be 0.94 which is related to both sensitivity and
specificity.
5 Conclusion
The detection of roads is a critical problem. This is due to complication of road

segments like occlusions, discontinuities, different boundaries, and sharp bends. This
makes the classification of both buildings and vehicles also difficult. This research
article discusses about an approach that classifies buildings and vehicles in Google
Map satellite images based on the random forest approach. This provides an infor-
mative approach for the classification of both buildings and vehicles. The random
forest classifier is employed for this purpose. This research article discusses about
an approach in classifying buildings and vehicles with relatively high accuracies.
The results can be further improved if the images are processed using deep neural
networks that employs deep learning.
References
1. Liu X, Deng Zhidong, Yang Y (2019) Recent progress in semantic image segmentation. Artif
Intell Rev 52:1089–1106
2. Ruiz JJ, Rubio TJ, Urena MA (2011) Automatic extraction of road intersections from images
based on texture characterisation. Surv Rev 43(321):212–225
3. Maboudi M, Amini J, Hahn M, Saati M (2016) Road network extraction from VHR satellite
images using context aware object feature integration and tensor voting. Remote Sens Multidis
Digital Publ Inst 8(8):637
4. Coulibaly I, Spiric N, Sghaier MO, Manzo-Vargas W, Lepage R, St-Jacques M (2014) Road
extraction from high resolution remote sensing image using multiresolution in case of major
disaster. In: Proceedings of 2014 IEEE remote sensing symposium, pp 2712–2715
5. Sghaier MO, Lepage R (2016) Road extraction from very high resolution remote sensing optical
images based on texture analysis and beamlet transform. IEEE J Select Top Appl Earth Observ
Remote Sens 9(5):1946–1958
6. Li M, Stein A, Bijker W, Zhan Qingming (2016) Region-based urban road extraction from
VHR satellite images using Binary Partition Tree. Int J Appl Earth Obs Geoinf 44:217–225
7. Alshehhi R, Marpu PR (2017) Hierarchical graph-based segmentation for extracting road
networks from high-resolution satellite images. ISPRS J Photogram Remote Sens 126:245–260
8. Liu R, Song J, Miao Q, Pengfei X, Xue Q (2016) Road centerlines extraction from high
resolution images based on an improved directional segmentation and road probability.
Neurocomputing 212:88–95
9. Kasu NR, Chandran S, Hormese J (2019) Identification of water and green regions in Open-
StreetMap using modified densitometry 3 channel algorithm. In: IEEE international conference
on advances in computing, communication and control (ICAC3 19), 20–21 Dec 2019, Mumbai,
India
10. Colditz RR, Schmidt M, Conrad C, Hansen MC, Dech S (2011) Land cover classification with
coarse spatial resolution data to derive continuous and discrete maps for complex regions.
Remote Sens Environ 115:3264–3275
11. Cruz HO, Eckert M, Meneses JM, Martinez JF (2017) Precise Real-time detection of
nonforested areas with UAVs. IEEE Trans Geosci Remote Sens 55(2):632–644
12. Mattikalli NM, Richards KS (1996) Estimation of surface water quality changes in response
to land use change: application of the export coefficient model using remote sensing and
geographical information system. J Environ Manage 48(3):263–282
13. Brezonik P, Menken KD, Bauer M (2009) Landsat-based remote sensing of lake water quality
characteristics, including chlorophyll and colored dissolved organic matter (CDOM). Lake
Reservoir Manage 21(4):373–382
14. Prasad PRC, Rajan KS, Bhole V, Dutt C (2009) Is Rapid urbanization leading to loss of water
bodies. J Spat Sci 2(2):43–52
15. Goswami AK, Gakhar S, Kaur H (2014) Automatic object recognition from satellite images
using artificial neural network. Int J Comput Appl 95(10)
16. Adam E, Mutanga O, Odindi J, Abdel-Rahman EM (2014) Land use cover classification in
a heterogeneous coastal landscape using Rapid-Eye imagery: evaluating the performance of
random forest and support vectormachines classifiers. Int J Remote Sens 35(10):3440–3458
17. Carlson TN, Arthur ST (2000) The impact of land use land cover changes due to urbanization on
surface microclimate and hydrology: a satellite perspective. Glob Planet Change 25(1–2):49–65
18. Sterling SM, Ducharne A, Polcher J (2013) The impact of global landcover change on the
terrestrial water cycle. Nature Clim Change 3(4):385
19. Poullis C, You S (2010) Delineation and geometric modeling of road networks. ISPRS J
Photogram Remote Sens 65(2):165–181
20. Zhou YT, Venkateswar V, Chellappa R (1989) Edge detection and linear feature extraction
using a 2-D random field model. IEEE Trans Pattern Anal Mach Intell 11(1):84–95
21. Hedman K, Stilla U, Lisini G, Gamba P (2010) Road network extraction in VHR SAR images
of urban and suburban areas by means of class-aided feature-level fusion. IEEE Trans Geosci
Remote Sens 48(3):1294–1296
22. Hu J, Razdan A, Femiani JC, Cui M, Wonka P (2007) Road network extraction and intersection
detection from aerial images by tracking road footprints. IEEE Trans Geosci Remote Sens
45(12):4144–4157
23. Zhang J, Lin X, Liu Z, Shen J (2011) Semi-automatic road tracking by template matching and
distance transformation in urban areas. Int J Remote Sens 32(23):8331–8347
24. Deng Z, Sun H, Zhou S, Zhao J, Zou H (2017) Toward fast and accurate vehicle detection in
aerial images using coupled region-based convolutional neural networks. IEEE J Sel Top Appl
Earth Observ Remote Sens 10(8):3652–3664
25. Hu F, Xia G-S, Hu J, Zhang L (2015) Transferring deep convolutional neural networks for
the scene classification of high-resolution remote sensing imagery. Remote Sens 7(11):14680–
14707
26. Snoek J, Larochelle H, Adams RP (2012) Practical Bayesian optimization of machine learning
algorithms. Adv Neural Inf Process Syst 25:2951–2959
Addressing the False Positives
in Pedestrian Detection
N. J. Karthika and Saravanan Chandran
Abstract Pedestrian detection is a subfield of object detection that is necessary for

several applications such as person tracking, intelligent surveillance system, abnor-
mal scene detection, and intelligent cars. We prepared a dataset for addressing the
false positives that occur during the person detection process. Some objects have
very similar features to those of a person. If a model is trained using a dataset con-
taining only persons, it leads to several false positives since it cannot differentiate
such objects from that of a person. Our dataset includes person and person-like
objects (PnPLO). Person-like objects that we introduce in our dataset are statues,
mannequins, scarecrows, and robots. We used the SSD model to show that, on train-
ing a model using our dataset, we can significantly reduce the false positives during
detection when compared to models trained on standard person datasets, thereby
improving the precision. The dataset consists of 944 training images, 160 validation
images, and 235 images for testing, with a total of 1626 person and 1368 nonhuman
labelling.
Keywords Pedestrian detection · Nonhuman detection · Deep learning · SSD ·

Computer vision
1 Introduction
Humans can instantly recognise any object in an image. We can also simultaneously
interpret the location of any object, as well as how the objects interact. The human
visual system is very fast as well as accurate, helping us to perform even highly
complex tasks such as driving a vehicle, with little conscious thought. Computer
N. J. Karthika (B) · S. Chandran

Computer Science and Engineering, National Institute of Technology,
Durgapur, West Bengal, India
S. Chandran
1084 N. J. Karthika and S. Chandran
vision is a research area, wherein the researchers try to make the computers work
in the same way as to how the human visual system works. The recent trend in this
field is the use of deep learning models, which is because various researches [1–5]
show that deep learning models have made the computers much faster and accurate
in object detection, classification, recognition, and various other computer vision
problems. Until the year 2016, state-of-the-art object detection systems with the best
accuracy were computationally intensive and too slow to run in real-time (e.g. faster
RCNN [1]). Also, the models which ran on real-time were not accurate enough,
especially for safety-critical applications. With the advent of SSD model [2], there
was a significant improvement in the speed for detection with high accuracy.
Dataset plays a crucial role in problems of object classification, detection, recog-
nition, segmentation, etc. There are many popular datasets widely accepted as bench-
marks for object detection problem. ImageNet [6], PASCAL VOC [7], COCO [8],
and SUN [9] datasets are few examples. Each of these differs in the type of images,
number and type of object labels, and size of datasets.
Pedestrian detection is an object detection problem. It has several real-time appli-
cations such as person tracking, robotics, video surveillance, and driverless cars.
Research works over the years have been using various approaches for the problem
of pedestrian detection, such as part-based detection [10], holistic detection using
features like HOG [11], motion-based detection [12], patch-based detection [13],
and detection using multiple cameras. As in other object detection problems, deep
learning is currently the most used approach in the research related to pedestrian
detection [14].
All the research works related to pedestrian detection which use deep learning
for training their models make use of person datasets as the benchmark [11, 15,
16]. These datasets consist of only the images of persons in various postures and
under different lighting conditions. The popular object detection datasets [6–8] also
have person as a class but no object classes to differentiate the person from objects
having similar features as persons. Since there are objects such as mannequins and
statues that have very similar features as that of a person, a model that is trained
with datasets containing only person images will have higher false positives on
encountering such objects. Considering this problem, we prepared a dataset (PnPLO)
[17] containing persons as well as the objects having features very close to that of
a person such as mannequins, statues, scarecrows, and robots. We train SSD model
with 300*300 image input size (ssd_300) and show the improvements in precision
on testing the newly trained model compared to the model trained on benchmark
datasets, namely, COCO [8], INRIA [11], and PASCAL VOC [7]. We can observe
considerable improvement in precision with the model trained on our dataset, on
testing the models on PnPLO test dataset.
Addressing the False Positives in Pedestrian Detection 1085
2 Recent Works
Computer vision is a field of study where extensive research is going on, especially
with the use of deep learning models. Pedestrian detection or person detection, in
general, is one of the topics of eminence in the field of computer vision, because of its
wide variety of applications in real-world problems. Video surveillance, driverless
cars, and person tracking are some of the applications. Deep learning gained pop-
ularity with the advent of AlexNet [3] in the year 2012, followed by many notable
research works such as [4, 5, 18].
Most of the recent object detection research works use either faster RCNN [1]
or single-shot multibox detector [2] as their backbone network because of their
accuracy and speed of SSD. Before the advent of SSD, faster RCNN was widely
used because of its excellent accuracy. This model is based on a region proposal
network, which is class agnostic. This class agnostic nature of RPN networks leads
to high time consumption, as the network needed two rounds of predictions—first,
to predict the regions which may contain an object, and then to predict the class
of the object present in that region. YOLO [19] considered this disadvantage and
proposed to have only one round of prediction by making the region proposals to be
class-specific so that the network needs to look at an image only once, thus, saving
a great deal of time. As the image passes through the deep convolutional network
only once, the model was speedy and could be run real time. Though YOLO worked
in real time, it compromised the accuracy by a great deal when compared to the
previous state-of-the-art model [1]. Problem with the first version of YOLO was
that it could not capture scale variation and failed to detect very small objects. SSD
provided a solution for this problem by proposing an auxiliary structure that can
perform detections at multiple scales. SSD, therefore, can run in real time with an
accuracy comparable to that of faster RCNN.
2.1 SSD Overview
SSD model was the first model that worked in real time with an accuracy as good as
the previous state-of-the-art models in object detection. Before SSD, models such as
the RCNN series [1, 20, 21] used RPN-based approach, which was time-consuming
because of two stages—region proposal, followed by detecting objects in each pro-
posal. The most significant advantage of SSD is its simplicity, with a single network
encapsulating all the computations, eliminating the need for a proposal generation
as well as the feature resampling stages. Figure 1 shows the architecture of the ssd
model.
SSD takes an image and ground truth boxes as input. The model used VGG-
16 network [5] as the base network. This network forms the first layers, following
which, an auxiliary structure was added to the network to produce detections. For
each location in feature maps of different scales, a small set of default boxes of
38 × 38 × 512 19 × 19 × 1024 10 × 10 × 512 5 × 5 × 256 3 × 3 × 256 1 × 1 × 128
VGG up to
VGG up to
conv4_3
Average
Pooling
Input
layers
layers
Conv
Conv
layers
Conv
fc7
image
DetecƟon & DetecƟon & DetecƟon & DetecƟon & DetecƟon & DetecƟon &
Classifier 1 Classifier 2 Classifier 3 Classifier 4 Classifier 5 Classifier 6
Fast NMS
Final DetecƟon
Fig. 1 SSD architecture
varying aspect ratios are evaluated in a convolutional manner. For each default box,
confidence scores of all object categories, and the shape offsets are predicted. During
training, the default boxes and the ground truth boxes are matched, and the model
loss is calculated, which is the weighted sum of confidence loss and the localisation
loss. The overall loss is given by :
1
L(x, c, l, g) = (L conf (x, c) + αL loc (x, l, g))
N
where L conf and L loc are confidence loss and localisation losses, respectively, and α is
the weight term. Confidence loss is the Softmax loss over multiple class confidences
(c). Localisation loss is the smooth L1 loss [21] between the parameters of ground
truth box and the predicted box (l).
2.2 Datasets
Pedestrian detection is a subject of interest in various researches because of its

widespread real-life applications. Hence, there are multiple standard datasets avail-
able, consisting of person as a class, used for these research works. We have con-
sidered three datasets used as benchmarks viz., COCO, INRIA, and PASCAL VOC
datasets.
2.2.1 COCO
This dataset contains images of complex everyday scenes of common objects in their
natural context. It is a large-scale dataset for object detection. It defines 91 classes,
but only 80 classes are used by the data. Segmentations of 11 other classes were not
collected because of problems like too many instances, ambiguity, and difficulty in
labelling, too few instances, etc. Compared to the previous datasets such as Imagenet
and PASCAL VOC, COCO dataset has more object instances per image and also
gives an additional focus on segmenting individual instances of different objects.
Person is one among the 80 classes considered in the dataset.
2.2.2 INRIA
The INRIA person dataset has a training set constituting 1128 negative images and
614 positive images, and a testing set with 288 images. First created and used by Dalal
and Triggs [11], the static person dataset comprised of people in various positions
and orientations, taken in a variety of backgrounds and lighting conditions.
2.2.3 PASCAL VOC
The PASCAL visual object classes (VOC) challenge had been organised annually
since the year 2005. The dataset associated with this challenge is publicly available
and has been accepted as a benchmark in object detection. It constitutes images as
well as annotations in XML format. The dataset consists of 20 classes which include
‘person’ as one among them.
3 PnPLO Dataset
Data plays a critical role in any deep learning research work, enabling the computers
to work in the same way as to how the humans do. This is especially true in the field
of object detection, where the number and the type of images used for training a deep
learning model play a crucial role while applying the model to real-world problems.
All the datasets used as benchmarks for person detection problem contains only
images labelled with person objects. Training with such a dataset leads to several false
positives while testing, when the images include many objects having features close
to that of a person. If an image contains a statue, then a model that was trained with
only person images tends to identify the statue as a person, leading to a false positive.
To address this problem of false positives, we prepared a dataset containing persons as
well as the objects having features similar to a person—person and person-like objects
(PnPLO) dataset [17]. We have labelled the person-like objects as ‘nonhuman’.
3.1 Image Acquisition
Person images and their corresponding annotation files used for training are consid-
ered from the PASCAL VOC 2012 person training dataset, and images for testing
are taken from PASCAL VOC 2007 person test set. The nonhuman images are taken
from the Internet. These images are completely random, not taken for a specific pur-
pose or a specific event, or from any particular angle. Because of this randomness, we
get an unbiased dataset. Some of the nonhuman images also contain person objects
in them.
3.2 Labelling
The dataset consists of a total of 944 images for training, 160 images in the validation
set, and 235 for testing. In the training set, there are 1106 person and 960 nonhuman
labellings. In the validation set, there are 203 person and 130 nonhuman labellings.
The test set consists of 317 and 278 labellings of person and nonhuman, respectively.
We labelled the nonhuman images using the labelImg tool [22], which is a graphi-
cal tool for generating image annotations. This tool saves the annotations in the form
of XML files in the format of PASCAL VOC dataset. We have labelled the images for
two classes, person and nonhuman class. In the XML annotations of PASCAL VOC
dataset, we have removed the annotations marked as difficult since such objects will
have similar features for both the classes considered and are difficult for even the
human eye to differentiate correctly. We have taken 526 person images for training
from the PASCAL VOC 2012 dataset, and 125 images from PASCAL VOC 2007
test list for testing. The number of person and nonhuman objects in the dataset are
comparable to avoid any over-fitting or under-fitting problems.
4 Experiment
We first tested the ssd_300 model trained on some standard datasets on the test
data of our dataset. The ssd_300 model trained on COCO, INRIA, and PASCAL
VOC datasets, respectively, are considered. Since the PASCAL VOC dataset includes
person as a class, we used the SSD model trained on this dataset as the initial setting
for the model to train on our dataset. This leads to a good initialisation for the model
instead of any random weight initialisation methods. We trained this model on our
train data for 50 epochs. We limited the number of epochs to 50 as further epochs did
not give any considerable improvement in the loss. We have trained the model with
a learning rate of 10−3 for 10 iterations, then with a learning rate of 10−4 , we trained
the model up to 30 iterations, and for the final iterations, we used a learning rate of
10−5 to train the model. We then used the final trained model to evaluate the test data.
Fig. 2 Evaluation using model trained on PASCAL VOC 07+12 dataset: detecting robot, scarecrow,
mannequins, and statues as person
Fig. 3 Evaluation using model trained on PnPLO dataset: robot, scarecrow, mannequins, and statues
are correctly detected as nonhumans
We noticed a significant improvement in the precision of person detection after

training on our dataset when compared to the precisions obtained on training on
datasets containing only persons. This improvement is achieved with the help of
PnPLO dataset [17] that considered the objects with features resembling those of a
person. Figure 2 shows the evaluation on test images of PnPLO dataset using model
trained on PASCAL VOC 07+12 dataset. We can see that the model wrongly detects
the person-like images as persons. Figure 3 shows evaluation on the same four images
using SSD model trained on PnPLO dataset. We can see that the model is able to
differentiate person from other person-like objects.
4.1 Evaluation Metric
Average precision (AP) is the evaluation metric used to compare the performances
of the SSD model trained on different datasets. Following metrics are involved in the
calculation of average precision.
4.1.1 Intersection Over Union (IOU)
IoU is given by the following formula—the area of overlap over the area of the union
of the predicted and ground truth bounding boxes.
area of intersection
IoU =
area of union
IoU is used to measure whether the bounding box predicted by the model is true
positive (TP), false positive (FP), or false negative (FN). If the IoU > 0.5, we consider
the predicted box to be a true positive. The bounding box is considered to be FP either
if IoU < 0.5, or if there are duplicate boxes predicted for the same object in an image.
The predicted bounding box is an FN if IoU > 0.5, but a wrong prediction.
4.1.2 Precision
Precision is nothing but the measure of how accurate our predictions are.
TP
Precision =
TP + FP
4.1.3 Recall
Recall measures how well the model finds all the true positives.
TP
Recall =
TP + FN
Once the above values are obtained, the precision-recall curve (PR curve) is plotted,
as shown in Fig. 4. Average precision is calculated by taking the area under the PR
curve.
Table 1 shows the improvement we achieved after training the SSD model on
PnPLO dataset. We have compared the performance of the model trained on PnPLO
dataset with that trained on three standard datasets, namely, PASCAL VOC 07+12
[7], COCO [8], and INRIA [11] person datasets. Average precision is the metric used
to compare the performances.
On evaluating the model trained on COCO, INRIA, and PASCAL VOC, on
the test set of our dataset, the average precision obtained was 53.6%, 55.3%, and
61.6%, respectively. After training on our training set, the performance significantly
improved to an average precision of 79.8%. Figure 4 shows the precision-recall curve
on evaluating the SSD model on our test data.
Fig. 4 Precision-recall curve for person and nonhuman on SSD model evaluation on our test data
Table 1 Comparing performance of SSD300 trained on different datasets

Dataset Average precision (%)
COCO [8] 53.6
INRIA [11] 55.3
PASCAL VOC 07+12 [7] 61.6
PnPLO (ours) [17] 79.8
Bold represents the average precision of SSD300 model on the PnPLO dataset, which we created
5 Conclusion
Various research works are carried out focusing on detection of persons because of its
widespread applications such as video surveillance, person tracking, and intelligent
cars. These works use datasets comprising of only persons as benchmark dataset.
Many objects have features similar to that of a person. A model trained on only
persons fails to differentiate these objects from a person and person-like objects.
Usage of only person datasets as a benchmark leads to many false positives, detect-
ing person-like objects also to be persons. To overcome this problem, we prepared
a person and person-like object (PnPLO) dataset consisting of persons as well as
person-like objects such as statues, mannequins, scarecrows, and robots. We trained
ssd_300 model on our dataset and tested on PnPLO test data. We show that the perfor-
mance of the model trained on PnPLO dataset is better than performances of models
trained on three standard datasets, namely, COCO, INRIA, and PASCAL VOC. The
model trained on our dataset has an average precision of 79.8% compared to 53.3%,
55.6%, and 61.6% for SSD model trained on COCO, INRIA, and PASCAL VOC
datasets, respectively.
References
1. Ren S, He K, Girshick R, Sun J (2017) Faster R-CNN: towards real-time object detection with
region proposal networks. IEEE Trans Pattern Anal Mach Intell 39(6):1137–1149
2. Liu W et al (2016) SSD: Single shot multibox detector. In: Lecture notes in computer science,
vol 9905. Springer, pp 21–37
3. Krizhevsky A, Sutskevar I, Hinton GE (2012) ImageNet classification with deep convolutional
neural networks. Commun ACM 60:84–90
4. Szegedy C et al (2015) Going deeper with convolutions. In: 2015 IEEE conference on computer
vision and pattern recognition (CVPR), Boston, MA, pp 1–9
5. Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image
recognition. arXiv: 1409.1556
6. Deng J, Dong W, Socher R, Li LJ, Li K, Fei-Fei L (2009) ImageNet: a large-scale hierarchical
image database. in: CVPR
7. Everingham M, Van Gool L, Williams CKI, Winn J, Zisserman A (2010) The PASCAL visual
object classes (VOC) challenge. Int J Comput Vis 88:303–308
8. Lin T-Y, Maire M, Belongie S, Bourdev L, Girshick R, Hays J, Perona P, Ramanan D, Zitnick
CL, Dollar P (2014) Microsoft COCO: common objects in context. In: European conference on
computer vision (ECCV), Zurich
9. Xiao J, Hays J, Ehinger KA, Oliva A, Torralba A (2010) SUN database: large-scale scene
recognition from abbey to zoo. In: CVPR
10. Wang S, Cheng J, Liu H, Wang E, Zhou H (2018) Pedestrian detection via body part semantic
and contextual information With DNN. IEEE Trans Multimedia 20(11):3148–3159
11. Dalal N, Triggs B (2005) Histograms of oriented gradients for human detection. In: Proceedings
of the IEEE computer society conference on computer vision and pattern recognition, pp 886–893
12. Pierard S, Lejeune A, Van Droogenbroeck M (2011) A probabilistic pixel-based approach to
detect humans in video streams. In: IEEE international conference on acoustics, speech and signal
processing (ICASSP), pp 921–924
13. Leibe B, Seemann E, Schiele B (2005) Pedestrian detection in crowded scenes. In: IEEE
computer society conference on computer vision and pattern recognition (CVPR’05), San Diego,
CA, USA, vol 1, pp 878–885
14. Karthika NJ, Chandran S (2019) Recent developments in pedestrian detection using deep
learning. In: 2019 International conference on computing, communication, and intelligent systems
(ICCCIS), Greater Noida, India, pp 353–358
15. Dollr P, Wojek C, Schiele B, Perona P (2012) Pedestrian detection: an evaluation of the state
of the art. IEEE Trans Pattern Anal Mach Intell 34(4):743–761
16. Geiger A, Lenz P, Urtasun R (2012) Are we ready for autonomous driving? The kitti vision
benchmark suite. In: Proceedings of the IEEE conference on computer vision and pattern recog-
nition, pp 3354–3361
17. Karthika NJ, Chandran S. PnPLO Dataset. https://drive.google.com/open?id=1_
HSXRasckZr8-LIZ9ms7LtBxzMwl-3Oa
18. He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In: IEEE
conference on computer vision and pattern recognition (CVPR), Las Vegas, NV, pp 770–778
19. Redmon J, Divvala S, Girshick R, Farhadi A (2016) You only look once: unified, real-time
object detection. In: 2016 IEEE conference on computer vision and pattern recognition (CVPR),
Las Vegas, NV, pp 779–788
20. Girshick R, Donahue J, Darrell T, Malik J (2014) Rich feature hierarchies for accurate object
detection and semantic segmentation. In: 2014 IEEE conference on computer vision and pattern
recognition, Columbus, OH, pp 580–587
21. Girshick R (2015) Fast R-CNN. In: 2015 IEEE international conference on computer vision
(ICCV), Santiago, pp 1440–1448
22. Tzutalin L (2015) https://github.com/tzutalin/labelImg
Hybrid CR Network: An Approach
Based on Interweave- and Underlay-Type
CR Network
Varun Kumar, Mangal Singh, Sarat Kumar Patra, and Poonam Singh
Abstract Increased usage of mobile devices has surged demand for RF spectrum
considerably. Cognitive radios (CR) hold tremendous promise for improving spectral
efficiency in the wireless system. It provides a fair solution for the dynamic allocation
of frequencies between the primary user (PU) and secondary user (SU). Efficient
techniques for spectrum holes’ estimation in available band maximize the system
capacity, but in the limited band, it is tough to provide an unutilized band for SU.
Since the secondary network is established where the licensed band is underutilized
but if the licensed band is over-utilized, SU may still communicate if interference is
lower than I th to the primary user. This paper investigates the capacity of a secondary
user with the variation of the power of SU in interweave and underlay network.
It provides the novel solution for the performance of SU in hybrid network mode,
where SU requires larger bandwidth but the availability of spectrum to the cognitive
user is less.
Keywords Cognitive radio · Interweave network · Underlay network · Hybrid

spectrum access · CFO
V. Kumar
Indian Institute of Information Technology, Surat, India
M. Singh (B)
Institute of Technology, Nirma University, Ahmedabad, India
S. K. Patra
Indian Institute of Information Technology, Vadodara, India
P. Singh
National Institute of Technology, Rourkela, India
1094 V. Kumar et al.
1 Introduction
Increased usage of wireless technology has created a huge demand for high data rate
in the wireless scenario. Additionally, a significant number of users in the limited
available spectrum have created a scarcity in licensed band. It is expected that nearly
20 billion wireless devices will be utilized over the globe for various application by
2023 [1–3]. Demand for spectral efficiency (SE) along with energy efficiency (EE) is
a major challenge in wireless research. Support for high data rate at low power has led
to the wireless network densification and also increases the operational cost. From
Shannon’s channel capacity theorem, system throughput directly depends on SNR or
transmitted power which implies that supporting high data rate in limited power is a
big problem in wireless communication research. In the other side, optimal resource
(time, frequency, power, space, code) allocation in multiple access scenario for uplink
(UL) and downlink (DL) also be one of the emerging challenges. Cognitive radio
(CR) is one of the possible areas in the field of resource allocation which utilizes the
wireless resources opportunistically. CR-based network searches the underutilized
license band support for high user density. CR system provides a fair solution for
dynamic allocation of frequencies between the primary user (PU) and secondary
user (SU) [4]. There are three type of cognitive radio network mode [5]. In the first
mode called as interweave networks where spectrum, holes present in the licensed
band are allocated to the SU for increasing the overall system capacity. Second mode
corresponds, due to unavailability of spectrum in licensed band, the cognitive user
utilizes that band in such a way that it produces interference below the threshold
limit, called as underlay network. Using the same sub-carrier by the different user,
interference could be nullified through intelligent signal processing which is termed
as overlay mode for CR system.
1.1 Prior Research, Motivation, and Contribution
In [6], authors describe the interweave CR network with cooperative sensing, consid-
ering separate PU decoding for the closest PU receiver. The information observed by
secondary base station (SBS) for all closely located SU is highly correlated in coop-
erative sensing environment. Hence, through soft combination, algorithm for user
selection at SBS improvises the overall performance [7]. A solution for an achievable
rate under an average transmit power and interference outage constraints has been
derived for underlay network [8]. After spectrum sensing, a perfect decision requires
based on optimal power. Deterministic and probabilistic model [9] in CR network
helps to the optimal allocation of the power between PU and SU, which maximize
the overall system capacity. Different literature covers the aspect of optimal power
control under various regime. Mung-Chiang et al. worked on the power resource,
where they discussed the optimal power allocation among user and base station (BS)
through geometrical programming [10, 11] to enhance the capacity. Simultaneously
Hybrid CR Network: An Approach … 1095
optimal power allocation among BS and user equipment (UE) solves the problem
of high power requirement in uplink and downlink scenario and also maximizes
the desired SNR at low power condition [12]. The adaptive deactivation of adjacent
sub-carrier provides flexible guard bands between PU and SU and causes mutual
interference cancelation in OFDM-based CR system [13].
Above literature mainly covers (1) spectrum sensing, (2) OFDM-based CR
network, (3) different optimal power control regime, and (4) different modes (inter-
weave, underlay) for CR system. This paper integrates these various aspects and
presents a comprehensive solution for hybrid spectrum access for SU, considering
interweave and underlay scenario. Chu et al. have described the hybrid spectrum
access in cooperative cognitive network [14] but our work is based on OFDM-enabled
single base station (BS) supporting multiple users-based CR system. Interweave and
underlay are the two separate aspects for utilizing the frequency resource. Interweave
mode requires full spectrum slot, and a lot of work were based on the sensing and
allocation of sub-carrier to the cognitive user in this mode. Like interweave, sensing
and sub-carrier allocation for underlay mode have also been addressed in the above
literature. No paper address these two modes jointly for performance analysis of
cognitive user. The numerical result validates the scenario under different real-time
constraints with suitable assumption. Performance analysis in interweave, underlay,
and hybrid mode has been carried out by an adaptive number of sub-carrier allocation
to the SU. This paper investigates the capacity performance enhancement in cognitive
radio through hybrid network approach which includes the benefits of interweave
and underlay mode.
In summary, major contributions of this paper can be stated as follows:
• Based on spectrum sensing and average information about the interference of PU,
we maximize the overall throughput for cognitive user.
• On the basis of two decision threshold, hybrid spectrum access has been
implemented. It depends on the channel condition of licensed sub-carrier in
OFDMA-based user access mechanism.
• Some other effects like carrier frequency offset which degrade the performance
of PU as well as SU have been jointly analyzed.
This paper has been organized as follows, where Sect. 2 covers the related work
about conventional CR mode like interweave and underlay. Section 3 describes the
system model for proposed method along with classical one. Numerical results have
been analyzed in Sect. 4 carried out by the conclusion in Sect. 5.
2 Related Work Based on Interweave and Underlay

Toward Achievable Rate
Based on the intuition framed for interweave and underlay mode [5], its extension
for cognitive network in OFDM system [15] gives the background of this research.
So OFDM-based interweave and underlay performance analysis has been discussed

through two cases.
Case 1: Performance of SU in Interweave Network Generalized frequency

response of received wireless signal for OFDM-based system can be expressed as
Ym,n ( f ) = Hm,n ( f )X m,n ( f ) + Wm,n (1)
where Ym,n ( f ), Hm,n ( f ), X m,n ( f ), Wm,n are the received signal frequency response,
transfer function of wireless channel, transmitted symbol, and additive white Gaus-
sian noise of nth sub-carrier for mth time domain sampling index, respectively. Trans-
mitted symbol of nth sub-carrier for mth time domain sampling index can be expressed
as
j2πmn
X m,n ( f ) = X m,n ( f )e N (2)
where N is the total number of sub-carrier.

In interweave network, PU does not influence the SU. It is mainly affected due to
system noise and carrier frequency offset (loss of orthogonality between consecutive
sub-carrier). The received power of nth sub-carrier can be expressed as
⎛ ⎞2
Bk /2 sin(π f − Tns Ts
P(n) = ρ0 (n)⎝ ⎠ df (3)
−Bk /2
(π f − Ts Ts
n
where P(n) , ρ0 (n), Ts = 1f , Bk are the power allocation of nth sub-carrier, peak
power spectral density of nth sub-carrier, symbol duration, and allocated bandwidth,
respectively. At constant bandwidth, if power level of user terminal increases then
the power spectral density (PSD) also increases. PSD should not exceed from the
maximum limit in such a scenario. In the practical scenario, the upper bound capacity
obtained by Jensen’s inequality gives the maximum achievable observed capacity.
The maximum achievable observed capacity of any random SU in interweave
scenario without (CFO) can be expressed as [16].
⎛ 2 ⎞

pks,n h ks,n
N f ks
αk
Cks = log2 ⎝1 + ⎠
s,n
N f ks − Nsks n=N σn2
sks
αks,n = 1∀Nsks ≤ n ≤ N f ks
αks,n = 0 elsewhere (4)
where Cks is the expected capacity of any ksth cognitive user. Nsks and N f ks are the
starting and last sub-carrier number allotted to respective user. It is also expected that
all these sub-carriers have not been allocated to any PU or other secondary user. N,
B, σn2 are total number of available sub-carriers, total system bandwidth, and noise
variance in license band, respectively. h ks,n is the channel gain of nth sub-carrier

for ksth SU also h ks,n = gks,n βk s where gks,n and βks are the fast fading and slow-
2
fading coefficient and E gk = 1. In interweave network, there is no inter-user
s,n
interference, but the performance of cognitive user is suppressed by only through

system noise and CFO.
CFO in SU Sub-carrier
Inter-carrier interference arises mainly due to incorrect (CFO). It occurs due to
improper sampling. When lth sub-carrier suffers with CFO ε, then the received lth
sub-carrier can be expressed as
N /2
1

Yl = X K HK e− j2π(k−l+ε)n/N (5)
N n K =−N /2
The closed-form equation for observed SINR of SU can be expressed as
2 2
Pks,n h ks,n sin(πε)
(πε)
γ1 = 2 (6)

0.822Pks,n h ks,n sin(π ε)2 + σn2
Case 2: Performance of SU for underlay network When spectrum holes are

unavailable in the licensed band, SU utilizes the licensed band in such a way that
power level of SU could not exceed the threshold limit. Adaptive power controlling
mechanism can limit excess power transmission from the secondary base station
(SBS). Mathematically, the power distribution through the SBS for SU can be
expressed as
Pks,n = min{Ith , Psbs } (7)
where Pks,n , Ith , Psbs are the optimal power for ksth SU, interference threshold to PU
and transmitted power through SBS respectively. Since in underlay scenario PU is
the strong interferer to the secondary, observed achievable capacity in such a scenario
can be expressed as
⎛ N f ks 2 ⎞
1
N f ks −Nsks n=Nsks Pks,n h ks,n
Cks = log2 ⎝1 + N f ks 2 ⎠ (8)
1
N f ks −Nsks n=Nsks P j p,n h j p,n + σn2
where P j p,n is the power of jth PU, whose few sub-carrier starting from n = Nsks to
n = N f ks has been allotted to a random SU where P j p,n|Nsks ...N f ks > Pks,n|Nsks ...N f ks and
power level of SU should not cross the interference threshold limit. In worst case,
scenario achievable capacity in underlay network is jointly influenced by interference
and CFO.
When SU and PU suffer with CFO

Under such scenario, the SU performance significantly degrade and observed SINR
can be expressed as
2 2
Pks,n h ks sin(πε)
(πε)
γ1 = 2 2 (9)
0.822Pks h ks sin(π ε)2 + P h j p + σn2
If CFO arises in PU band then observed SINR can be expressed as

0.822Pp (sin(πε))2
Ps |h s |2 1 + σn
2
γ2 = 2 2 2 (10)
σn2 + 0.822Pp h p (sin(π ε)) + Pp h p sin(πε)
2
(πε)
3 Hybrid Network
Composite network comprises the coexistence of the primary and secondary network
in a certain geographical bound. Primary network consists primary base station (PBS)
and PU, whereas secondary network is framed by SBS or SU. Here, system model
includes the following constraints.
1. Bandwidth allocation for SU should not be greater than Bs Hz for preventing the
unauthorized access of large spectrum from the security point of view.
2. Primary users (PU) are assumed to be not very far from the SBS as well as
secondary users (SU) such that false alarming issues did not come into the picture.
3. Among the SU, there are no spatial correlation.
4. Expected power level across allocated sub-carriers to the PU remain same.
Hybrid spectrum allocation for SU based on interweave/underlay scenario has
been categorized into three steps.
1. Spectrum sensing
2. Decision based on activation threshold
3. Performance analysis for PU and SU.
3.1 Spectrum Sensing Across SBS
In this section, we consider that M p number of PUs are active in the particular
geographical bound and PBS/SBS both receive the signals. PBS remains unaffected,
but SBS acquires the information from the PU. During sensing, no uplink and down-
link data transmission occur by secondary network till sensing time slot. Only SBS
receives the signal and SU remains inactive during sensing. Within sensing time
interval, N sub number of data samples are acquired by SBS. Received data sample
is passed through the FFT block across the SBS. For such a system N sub > N
and f = B/N , where f is the sub-carrier bandwidth. Two scenarios may be
considered for the received signal
⎧
⎪ n 1 (t) A0
⎪
⎨
sl (t) + n l (t) A1
y1 (t) = (11)
⎪
⎪ Nsub
⎩∀ 0<t <
B
From above equation, yl (t) is a normal distributed random variable (RV) with
mean 0 for hypotheses A0 implies the absence of PU, whereas for hypotheses A1 ,
replicate the presence of PU; the received symbol yl (t) is a normal distributed RV
with mean μs . The received power across the SBS can be modeled as
Mp
2
Mp
2 2
Py ( j) = yi j = si j + n j (12)
i=1 i=1
2
where si j is the signal power of ith user for jth sample. From above equation, if i
= 0 (All primary users are inactive), sij = 0; only noise power is considered to be
received power. Statistical analysis gives an idea about the power level of different
PU whose signals are received across the SBS.
4 Numerical Results
In this section, analytical results have been numerically verified. In multiple access
scenario, we have chosen the OFDM-based frequency resource allocation for
different users. These users may be PU or SU. We also consider the availability
of perfect channel state information CSI between PBS to PU and SBS to SU. Cogni-
tive users also know CSI between itself and PBS/PU. The expected channel
variance
2

across each sub-carrier either for SU has been taken unity, i.e., E h ks,n = 1.
In three different scenarios, SU performance variation has been observed due to
nearest reference PU. In LTE-TDD standard, to access multiple users, minimum one
resource block (RB) is transferred to any user, where one RB is 0.5 ms and 180 kHz
wide in time–frequency frame. One-time slot carries 7 OFDM symbol. In normal
cyclic prefix condition, 12 sub-carriers are carried by one resource block. So sub-
carrier spacing _f is equal to 15 kHz. Total number of available sub-carrier, i.e., N
= 512. For simplicity, let Bp = 1.5 MHz bandwidth has been allocated to the PU and
the maximum permissible bandwidth for SU is Bs = 300 kHz. In such a wireless
network, 5 PU are active and for each PU 100, sub-carriers are allocated. In this
assumption, maximum permissible sub-carrier for SU is 20. As per OFDM-based

system constraints, the minimum BW allocation to the cognitive user is 180 kHz and
maximum allocated bandwidth is 300 kHz for preventing the unauthorized access
of larger licensed band. According to dynamic allocation of the frequency band, the
wireless network may allocate the band to SU in three different modes termed as
interweave, underlay, and hybrid mode. Figure 1 shows the capacity performance in
offset scenario. Under such condition, three modes have shown. In interweave mode,
one sub-carrier suffers from 10% CFO due to incorrect sampling, so the achievable
rate degrades. In underlay scenario, two cases have been taken. In the first case, 10%
CFO arises in one sub-carrier and at same sub-carrier PU acts as an interferer, while
in the second case, PU sub-carrier suffers with 10% CFO. Under no CFO and PU as
an interfering agent, the performance degradation in such scenario has been shown in
Fig. 1. Since power control mechanism in underlay is applied in such a way that SBS
did not transmit power more than I th irrespective of maximum power transmit limit by
SBS, power constraints from (9) and taking 5 dB as (I th ) all curve remain to go to the
saturation in underlay scenario. In proposed hybrid spectrum access methodology,
10% CFO has been considered for those sub-carrier which utilizes the PU band and
same amount of CFO in the unused band. If all signals are perfectly sampled, then the
achievable rate for three modes can be observed from Fig. 2. In the joint scenario, two
decision thresholds γl and γavg are used for correct justification about the interweave
and underlay band. Here, γl and γavg have been taken 0 dB and 10 dB, respectively,
for PU. Out of 500 allocated sub-carrier of PU, best 8 sub-carriers are selected from
the PU band. As per assumption, 12 sub-carriers are unused and free from inter-user
interference. Eight sub-carriers of PU band having γl > γavg are aggregated with
twelve unused sub-carriers, which fulfill the BW greed for SU. Based on proportion-
ality average, SNR for hybrid spectrum access methodology is formulated. In such a
Fig. 1 Observed capacity of cognitive user in offset scenario

Fig. 2 Capacity in three modes of CR system
condition, I agg be the transmit power and 5 dB is taken for numerical simulation. The
analytical curve for the upper bound of a closed-form solution to average achievable
rate is obtained from (5), (10) and (13), which are tightly matched to the exact Monte
Carlo simulation. Depending on sub-carrier availability in interweave and underlay
mode, Fig. 3 depicts the achievable rate performance under different proportion-
ality. In this figure, only upper bound close form expression has been used. Due
to increase in the number of unused sub-carrier, the capacity increases with greater
extent. Neglecting the impact on statistics for frequency selective at channel, it is
Fig. 3 Capacity in mixed networks for different value of N und and N int
Fig. 4 Achievable rate of PU vs percentage utilization of sub-carrier in hybrid spectrum access

methodology where Dth = 1:25 dB and I th = 5 dB
2
assumed that all selective channels for PU have unit variance E h ( j p,n ) = 1 for
PU.
Bp NP f
Selective Channel = = (13)
Bc Bc
Figure 4 shows the impact of SU on PU. Multiple cognitive users degrade the
performance of PU in the nonlinear fashion. For numerical computation, detection
and interference threshold have been taken 1.25 and 5 dB, respectively, in underlay
scenario. Since 1.5 MHz and 300 kHz bandwidth have been assigned to PU and
SU, in this practice, the permissible cognitive users are five where all SU consume
maximum bandwidth. Figure 5 depicts two independent scenarios. In the first, when
the cognitive user receives power level up to detection threshold than such situation
causes less interference to the PU gets upper bound capacity. In the second scenario,
when the cognitive user receives the power level up to I th , the PU suffers from more
interference and lower bound capacity can be observed if PSU = I th .
5 Conclusion
In this paper, a detailed analysis has been carried out for SU capacity performance
in different CR network mode. In the case of limited BW availability, hybrid spec-
trum access mode gives moderate solution among another mode like, interweave
and underlay. For maintaining fairness among multiple SU, hybrid spectrum access
methodology is more helpful in comparison to interweave and underlay scenario.

This technique improvises the cognitive user density providing fairness at the cost of
hardware complexity. This paper analyzes the OFDM-based three-mode performance
comparison for a single cognitive user. Multiple users’ achievable rate performance
and their fairness can be an extension of this paper.
References
1. Gohil A, Modi H, Patel SK (2013) 5G technology of mobile communication: a survey’. In:

2013 IEEE international conference on intelligent systems and signal processing (ISSP), pp
288–292
2. Liu Y, Zhang Y, Yu R, Xie S (2015) Integrated energy and spectrum harvesting for 5G wireless
communications’. IEEE Netw 29(3):75–81
3. Wu J, Rangan S, Zhang H (2012) Green communications: theoretical fundamentals, algorithms
and applications. CRC Press
4. Liang YC, Chen K-C, Li GY, Mahonen P (2011) Cognitive radio networking and communi-
cations: an overview. IEEE Trans Veh Technol 60(7):338–340
5. Goldsmith A, Jafar SA, Maric I, Srinivasa S (2009) Breaking spectrum gridlock with cognitive
radios: an information theoretic perspective. Proc IEEE 97(5):894–914 (2009)
6. Kusaladharma S, Tellambura C (2015) Interweave cognitive networks with co-operative
sensing. In: 2015 IEEE global communications conference (GLOBECOM). IEEE, p 16
7. Bhatti DMS, Nam H (2017) Spatial correlation based analysis of soft combination and user
selection algorithm for cooperative spectrum sensing. IET Commun 11(1):3944
8. Sboui L, Rezki Z, Alouini M-S (2013) A unified framework for the ergodic capacity of spectrum
sharing cognitive radio systems. IEEE Trans WirelCommun 12(2):877–887
9. Mallick S, Devarajan R, Rashid MM, Bhargava VK (2012) Robust power allocation designs
for cognitive radio networks with cooperative relays. In: 2012 IEEE international conference
on communications (ICC). IEEE, pp 167–168
10. Chiang M, Tan CW, Palomar DP, Neill DO, Julian D (2007) Power control by geometric
programming. IEEE Trans Wirel Commun 6(7):26402651
11. Boyd S, Kim S-J, Vandenberghe L, Hassibi A (2007) A tutorial on geometric programming.
Optim Eng 8(1):671–674
12. Bansal G, Hossain MJ, Bhargava V (2008) Optimal and suboptimal power allocation schemes
for OFDM based cognitive radio systems. IEEE Trans Wirel Commun 7(11):4710–4718
13. Weiss T, Hillenbrand J, Krohn A, Jondral FK (2004) Mutual interference in OFDM-based
spectrum pooling systems. In: IEEE 59th vehicular technology conference, 2004. VTC 2004-
Spring, vol 4. IEEE, pp 1873–1877
14. Chu TMC, Phan H, Zepernick H-J (2014) Hybrid interweave-underlay spectrum access for
cognitive cooperative radio networks. IEEE Trans Commun 62(7):21832197
15. Zhang Y, Leung C (2009) Resource allocation in an OFDM based cognitive radio system. In:
IEEE Trans Commun 57(7)
16. Shen Z, Andrews JG, Evans BL (2005) Adaptive resource allocation in multiuser OFDM
systems with proportional rate constraints. IEEE Trans Wirel Commun 4(6):2726–2737
Intensive Study on Compact Integrated
Optic Couplers Using Grating Geometry
Bidyut Deka, Aradhana Dutta, and Partha Pratim Sahu
Abstract In the paper, tooth structured grating-assisted (GA) configuration for 2

× 2 compact directional coupler (DC), two-mode interference (TMI) coupler and
multimode interference (MMI) coupler have reported for an intensive study using
a sinusoidal mode simple effective index method (SM-SEIM) centric mathemat-
ical model. It is found that beat length of GA-TMI coupler is ~22.3 μm which is
almost 50% compact in size with comparison to the conventional TMI coupler and is
~25% that for conventional directional coupler. The power imbalance with grating-
assisted structures increases as that of fabrication tolerances which are slightly higher
compared to conventional TMI coupler.
Keywords Integrated optics · Planar waveguide · Grating · Simple effective index

method · Directional coupler · Grating coupler
1 Introduction
The compact planar waveguide-based optical device and its components have become
obligatory for implementation of large-scale integration in photonic integrated device
(PID) [1–6] for accomplish of increasing bandwidth requirements in contemporary
high speed communication. As the fundamental components of integrated circuit,
such as optical couplers (DC/TMI/MMI) and switches with smaller in size have
been growing interests due to the compactness and simple fabrication process. The
B. Deka (B)
Department of Physics, Girijananda Chowdhury Institute of Management and Technology
(GIMT)-Tezpur, Dekargaon, Sonitpur 784501, Assam, India
A. Dutta
Department of Physics, Kaliabor College, Kuwarital, Nagaon 782137, Assam, India
P. P. Sahu
Department of Electronics and Communication Engineering, Tezpur University, Sonitpur 784028,
Assam, India
1106 B. Deka et al.
grating-assisted geometry has appeared highly promising and is thus introduced in the
integrated optical couplers for further compactness that are very much obligatory for
large-scale-integration of PID. Polarization sensitiveness along with higher fabrica-
tion tolerances gives additional advantage. The previous works [7–9] have discussed
the coupling characteristics with a few detail study such as polarization sensitiveness
and fabrication tolerances based on finite difference time domain (FDTD) method
for the tooth structure grating-assisted TMI coupler.
In this paper, a detail intensive analysis of compact directional coupler (DC),
two-mode interference (TMI) coupler and multimode interference (MMI) coupler
with tooth structured grating geometry have been carried out using a sinusoidal
mode centric simple effective index method (SM-SEIM) [2, 10–13] based mathe-
matical model for accurate analysis of modal coupled power. Dependence of beat
length on coupling separation gap between access waveguides with a fixed value
of S bending loss for tooth structured grating-assisted directional coupler (GA-
DC), grating-assisted two-mode interference (GA-TMI) coupler and grating-assisted
multimode interference (GA-MMI) coupler are shown. Coupling behavior for DC,
TMI and MMI couplers with tooth structured grating geometry have discussed and
then compared with conventional structures.
2 Grating-Assisted Structure and the Principle
Figure 1 shows three-dimensional (3D) schematic view of 2 × 2 tooth structured

grating-assisted directional coupler (GA-DC) having tooth-shaped grating-assisted
coupling region with dimensions (length ~ L, coupling separation gap ~ h between
Fig. 1 Schematic 3D view of 2 × 2 directional coupler with tooth structured grating geometry
Intensive Study on Compact Integrated Optic Couplers … 1107
the two-channel waveguides), one pair of input single mode access waveguides
(Waveguide-1 and Waveguide-2) of core size (width ~ a, thickness ~ b) and another
pair of single mode output access waveguides (Waveguide-3 and Waveguide-4) of
similar dimensions, respectively. The coupling region with tooth structured grating
geometry is consisting of two-channel waveguides incorporated with tooth structured
grating placed close to each other. In the coupling region, the guiding layer of width
W m (~2a + h) and grating layer of width W g (~W m + 2W ) are placed alternatively
where W is the width of grating teeth. In this study, rectangular tooth structured
grating is used for higher compactness and simplification of implementation. The
grating structured coupling section is consisting of N total number of grating period,
= lm + l g ; where l m denotes the guiding layer’s length of width (S = m) and lg
gives grating layer’s length of width (S = g), respectively. Refractive index of core
layer and cladding layer are n1 and n2 respectively, whereas n3 is refractive index
of coupling gap cladding section. For input power ~ P1 launched at input lower
most access Waveguide-2, the respective output optical powers obtained through the
Waveguide-3 (as bar state) ~ P3 and Waveguide-4 (as cross state) ~ P4.
Once the mode field with propagation constant β i (λ) is launched as input signal
through single mode input access Waveguide-2, inside the tooth structured grating-
assisted coupling region modes are excited. In coupling section, based on compar-
ative phase difference among the excited modes, light powers are coupled at end
of the section through the output single mode access waveguides (Waveguide-3 and
Waveguide-4). As fundamental and first-order mode are carrying most of optical
power, the beat length which defines the coupling length required for a phase shift ~ π;
of optical coupler with N total number of grating period (~) is found as,
π
L π = (N + 1) lm + Nl g = m g g (1)
β00 − β01
m
+ β00 − β01
where β00 S
and β01S
denote propagation constant for the fundamental and first-order
modes irrespective to guiding layer (S ~ m) and grating layer (S ~ g), respectively. As
(S ~m, ~g), the width, W m = W g and Eq. (1) signifies coupling length for conventional
structures.
For high-index contrast waveguide, mode field penetration in lateral outside direc-
tion of waveguide is negligibly small where input modal field profile of the ith mode,
H i (x) for tooth structured grating-assisted coupling section can be approximated as,

πx
Hi (x) = sin (i + 1) (2)
Wg
Thus, the optical powers at end of tooth structured grating-assisted coupling region
are either coupled toward output access waveguides or diminishes out at end of grating
structured channel waveguide. Since all guided modes traveling through the grating
structured coupling section will contribute to mode field of output access waveguides,
thus, mode fields in Mth access waveguide can be express as
1108 B. Deka et al.

1
HMS (x, L) = S
HM,i (x, L)
i=0
S=m,g

1

= c M,i Hi (x) exp j β0S − βiS L (3)
i=0
S=m,g

where L = [(N + 1)lm + Nl g ] and c M,i ≈ C M,iS
are contribution coefficient of
ith mode at Mth access waveguide, estimated using sinusoidal mode centric simple
effective index method (SM-SEIM) [10–12] based numerical model as,
S
C M,i π2
=
C0 16b2 k 2 n 21 − n 22

1/2
2 1/2
1/2
exp −hk n 2eff − n 22 exp bk n 1 − n 22 − exp −bk n 21 − n 22
π2
+
16b2 k 2 n 21 − n 23

1/2
2 1/2
1/2
exp −hk n 21 − n 23 exp bk n 1 − n 23 − exp −bk n 21 − n 23
(4)
where for TE mode,

0.4 n 21 − n 2eff(TE),S n 2eff(TE),S − n 22
C0 = × (5)
FC
n eff(TE),S n 21 − n 23 W S + √ 2 2
k0 n eff(TE),S −n 22
3(1 + 0.2h)
Fc = (6)
13.5 + 185 β0S − βiS h

λ
n eff(TE),S = βTE(i)
S
; S = m, g (7)
2π
The normalized output coupling power at Mth access waveguide of tooth

structured grating-assisted directional coupler (GA-DC) can be written as,
2
1

H S
(x, L)
M,i
PM,i (x, L) i=0
S=m,g
= 2
P1,i (x, o) 1

H1,i (x, 0)
S

i=0
S=m,g

1
≈ S
C M,i Hi2 (x)
i=0
S=m,g

1
1
+ S
2 C M,i S
C M, j Hi (x)H j (x)
i=0 j=1+i
S=m,g S=m,g
⎧ ⎫⎤
⎪
⎪ ⎪
⎪
⎨ 1
⎬⎥
× cos (N + q S )(βi − β j )l S ⎥
S S
⎦ (8)
⎪
⎪ ⎪
⎪
⎩i=0, j=i+1 ⎭
S=m,g
S 2
where PMi = HM,i (x, L) and i, j = 0, 1 denotes even mode and odd mode such that
j > i, qS = 0, 1 refers to grating layer (S ~ m) and guided layer (S ~ g), respectively,
total number of grating period ~ N and C M,i S
, C M,
S
j are the contribution coefficients
for ith, jth modes that signifies the field contribution into output access waveguides
which can estimated from Eqs. (4), βi , βj = propagation constants for ith and jth
mode that are calculated using dispersive equations [2]. The guiding width length ~
lm and grating width length ~ lg are determined by using the following relation (9)
[8, 9],
λ
lS = ; S = m, g (9)
4n eff( j,S)
2.1 Result and Discussion
Figure 2 shows schematic layout of three-dimensional (3D) tooth structured 2 ×

2 grating-assisted directional coupler (GA-DC) along with the beam propagation
results at the bar coupling (P3 /P1 ) state and cross-coupling (P4 /P1 ) state with W m
= 3.0 μm, h = 0.5 μm, W = 0.25 μm, n = 5%, a = 1.5 μm, b = 1.5 μm, λ ~
1.55 μm obtained by using optiBPM software. It is also show light wave propagation
on half coupling (3-dB) state of GA-DC coupler and cross-coupling point obtained
by optiBPM software that is based on finite difference time domain (FDTD) method
[6, 12]. From the study found that the cross-coupling point is obtained at coupling
length of 45.1 μm which is almost close to that obtained by SEIM based on sinusoidal
modes.
The schematic 3D layout of 2 × 2 tooth structured grating-assisted two-mode
interference (GA-TMI) coupler is shown in Fig. 3 along with the beam propagation
results at bar coupling (P3 /P1 ) point and cross-coupling (P4 /P1 ) point found by using
optiBPM software for W m = 3.0 μm, h = 0 μm, a = 1.5 μm, b = 1.5 μm, W =
1110 B. Deka et al.
Fig. 2 Tooth structured 2 × 2 grating-assisted directional coupler (GA-DC) along with (a) 3D
schematic layout and BPM simulation results for (b) cross-state of beat length ~45.1 μm and
(c) 3-dB coupler of beat length ~23 μm
0.25 μm, n = 5%, λ ~ 1.55 μm, respectively. It is found that cross-coupling beat
length ~22.3 μm which is equivalent to the result obtained by SM-SEIM.
Figure 4 shows 3D device layout of the tooth structured 2 × 2 grating-assisted
multimode interference (GA-MMI) coupler along with beam propagation simulation
results estimated using optiBPM software at the bar coupling (P3 /P1 ) state and cross-
coupling (P4 /P1 ) state with W m = 7.0 μm, n = 5%, W = 0.25 μm, h = 4 μm, a =
1.5 μm, b = 1.5 μm, λ ~ 1.55 μm. The coupling length of GA-MMI coupler obtained
as ~40.1 μm and 3-dB coupler of beat length ~20.2 μm, respectively. Further, a
comparative analysis for beat length (L π ) versus n (%) for tooth structured GA-
MMI, GA-DC and GA-TMI couplers with teeth height W ~ 0.25 μm and that of
conventional couplers (structures with W ~ 0 μm) is shown in the plot Fig. 5. The
figure signifies that as n increases, the beat length reduces. This is obtained that
GA-TMI coupler has the lesser beat length compared to other types of couplers.
In Fig. 6, the relative study of normalized bar and cross-states coupling powers
distribution has shown with respect to grating numbers (~N) which can be estimated
using Eqs. (1)–(9) for tooth structured grating-assisted two-mode interference (GA-
TMI) coupler of coupling separation gap, h ~ 0.0 μm, directional coupler (GA-DC)
for h ~ 0.5 μm and multimode interference (GA-MMI) coupler for h ~ 4.0 μm with
Fig. 3 Tooth structured 2 × 2 grating-assisted TMI (GA-TMI) coupler along with (a) 3D schematic
layout and BPM simulation results for (b) cross-coupling state of beat length ~22.3 μm and (c) 3-dB
coupler of beat length ~11.5 μm
n = 5%, cladding index ~ 1.45, a = b = 1.5 μm, W ~ 0.25 μm, lm = l g =

0.27 μm and wavelength (λ) ~ 1.55 μm, respectively. From Fig. 7, it is observed that
the peak cross-state coupling power (P4 /P1 ) is found at beat lengths corresponding
to the values of N ~ 41, 70, and 85 with respect to the tooth-shaped GA-TMI, GA-
MMI and GA-DC, respectively. Thus, the beat lengths for GA-DC, GA-MMI and
GA-TMI couplers calculated using Eqs. (1) are ~45.1 μm, 40.1 μm and 22.3 μm,
respectively.
Further, these planar waveguide-based conventional DC, TMI coupler and MMI
coupler with waveguide designed parameters are then fabricated and experimentally
tested using waveguide materials, SiON as the core layer along with SiO2 cladding
layer. From the experimental results as shown in Fig. 7, the beat lengths of conven-
tional TMI coupler (h = 0 μm, W = 0 μm) and conventional MMI coupler (with h
= 4 μm, W = 0 μm) are found as ~45 μm and ~80 μm, respectively, whereas for
conventional DC (with h = 0.5 μm, W = 0 μm) is ~91 μm with n = 5%. In the
graph, respective cross and 3-dB coupling points are indicated by the dot, and star
signs show optiBPM simulation results along with experimental results and SEM
photographs of developed DC, TMI coupler and MMI coupler, respectively.
1112 B. Deka et al.
Fig. 4 Tooth-shaped GA-MMI coupler with (a) 3D layout and BPM results for (b) cross-coupling
state of L π ~ 40.1 μm and (c) 3-dB coupler of L π ~ 20.2 μm, respectively
Fig. 5 Beat length (L π ) 300

versus index contrast GA-DC
(n) for tooth structured coupler
grating geometry with W 250
Conventional
~ 0.25 μm and conventional MMI coupler
structures (where W 200
Beat Length ( m)
~ 0 μm) Conventional
TMI coupler GA-MMI
150 coupler
GA-DC
coupler GA-TMI
100
coupler
50
0
0 5 10
n (%)
1
0.9 GA-TMI
Normalized Coupling Power

0.8
0.7
(P3/P1, P4/P1)
P3 /P1
0.6 GA-MMI
0.5
0.4 P4 /P1 GA-DC
0.3
0.2
0.1
0
0 50 100
Grating Number (N)
Fig. 6 Normalized bar and cross-state coupling power distribution versus grating number for tooth
structured GA-TMI coupler with coupling gap, h = 0.0 μm (solid line), multimode interference
(GA-MMI) coupler (dashed lines) for h = 4.0 μm and directional coupler (GA-DC) (dotted lines)
for h = 0.5 μm with cladding index ~1.45, n = 5%, a = 1.5 μm, b = 1.5 μm, W = 0.25 μm
and λ ~ 1.55 μm, respectively
1
P3/P1
0.9
0.8
Normalised Coupled Power
0.7
0.6
0.5
0.4
0.3
P4 /P1
0.2
0.1
0
0 20 40 60 80 100
Beat Length (L )
Fig. 7 Normalized coupling power versus beat length using SM-SEIM-based mathematical model
for conventional TMI coupler, MMI couplers and directional coupler, along with BPM simulation
results and experimental results, respectively
1114 B. Deka et al.
3 Summary
In the paper, a detail comparative study of coupling behavior for tooth structured
grating-assisted two-mode interference (GA-TMI) coupler, multimode interference
(GA-MMI) coupler and directional coupler (GA-DC) have been presented using
a mathematical model based on sinusoidal mode centered simple effective index
method (SM-SEIM). The results are compared to the conventional coupler geom-
etry and verified with beam propagation method (BPM) simulation results obtained
by using commercially available optiBPM software. It is established that GA-TMI
coupler has shorter beat length compared to other couplers.
Acknowledgements The author thankfully acknowledge the help and supports for the fabrication
work, carried out at the CENSE under INUP at Indian Institute of Sciences (IISc.), Bangalore which
have been sponsored by DIT, MCIT, Government of India.
The author dully acknowledges the financial support provided for this work under collabora-
tion research scheme of TEQIP-III from Assam Science and Technology University, Guwahati,
Assam. The author also appreciates fruitful discussions held with Dr. Bharat Kakati.
References
1. Chin MK, Lee CW, Lee SY, Darmawan S (2005) High-index-contrast waveguides and devices.
Appl Opt 44:3077–3086
2. Nishihara H, Haruna M, Suhara T (1989) Optical integrated circuits. McGraw-Hill, New York
3. Das AK, Sahu PP (2006) In: IEEE wireless and optical communication network conference,
Digital No- 01666673, vol 1
4. Sahu PP (2009) Parabolic tapered structure for an ultracompact multimode interference coupler.
Appl Opt 48:206–211
5. Chan HP, Chow CK, Das AK (2003) A wide angle X-junction polymeric thermo optic digital
switch with low crosstalk. IEEE Photon Tech Lett 15:1210–1212
6. Passaro VMN (2000) Optimal design of grating-assisted directional couplers. J Lightwave Tech
18:973–984
7. Hardy A (1984) Exact derivation of the coupling coefficient in corrugated waveguides with
rectangular tooth shape. IEEE J Quantum Electron 20:1132–1139
8. Tsai TY, Lee ZC, Gau CS, Chen FS, Chen JR, Chen CC (2004) A novel wavelength-division-
multiplexer using grating assisted two-mode interference. IEEE Photon Tech Lett 16:2251–
2253
9. Tsai TY, Lee ZC, Chen JR, Chen CC, Fang YC, Cha MH (2005) A novel ultra compact two-
mode-interference wavelength division multiplexer for 1.5 μm operation. IEEE J Quantum
Electron 41:741–746
10. Chiang KS (1991) Effective index method for the analysis of optical waveguide couplers and
arrays: an asymptotic theory. J Lightwave Tech 9:62–72
11. Wang Q, Farrell G, Freir T (2006) Effective index method for planar lightwave circuits
containing directional couplers. J Opt Commun 259:133–136
12. Deka B, Sahu PP (2009) Transformation relationship of directional coupler with multimode
interference coupler and two mode interference coupler. J Opt 38:75–87
13. Deka B, Sahu PP (2011) Tooth-shaped grating-assisted structure for compact multimode
interference (MMI) coupler. Appl Opt 50:E193–E199
Effect of AlN Spacer Layer
on the Proposed MIS-AlGaN/GaN
HEMT
Ravi Ranjan, Nitesh Kashyap, and Ashish Raman
Abstract This paper presents a comparison between AlGaN/GaN high electron

mobility transistor (HEMT) with and without AlN spacer layer between AlGaN and
GaN interface. Two AlN layers are used in the proposed design, one AlN layer is
used as an interfacial passivation layer between Al2 O3 and AlGaN layer. 2DEG is
formed at the interface of AlN and GaN. Spacer layer enhances the mobility of
carrier in the channel. Due to higher mobility, current density and drain current
of the spacer layer HEMT are larger than HEMT without spacer. The results with
spacer-MIS-HEMT are compared with MIS-HEMT without spacer, which shows that
the spacer-MIS-HEMT provides better ON–OFF current, transconductance, cutoff
frequency (11 GHz), and ON to OFF current ratio is (1011 ).
Keywords Gallium nitride (GaN) · Spacer layer · Dopingless · HEMT · MIS
1 Introduction
Silicon has been mostly used in many electronics devices. But due to some limita-
tion of silicon over higher bandgap material, silicon is not preferred for high power
devices. Wider bandgap compound semiconductor is used for power devices and
radio frequency application such as radar, mobile tower and automobiles. III-group
nitrides have many advantages over silicon and other larger bandgap material (GaAs
and SiC) that is saturation velocity and thermal conductivity. Table 1 gives the mate-
rial and electrical properties of GaAs, 4H-SiC, Si and GaN. Due to above advantage,
III-group nitrides are used for power applications [1]. AlGaN/GaN hetro-structure-
based HEMT is better choice for high power and high frequency application [2].
R. Ranjan (B) · N. Kashyap · A. Raman

Dr. B. R. Ambedkar National Institute of Technology Jalandhar, Jalandhar 144011, India
N. Kashyap
A. Raman
1116 R. Ranjan et al.
Table 1 Physical and electronics property of GaN, Si, 4H-SiC and GaAs [1, 3, 5]
Property Si 4H-SiC GaN GaAs
Energy band (eV) 1.12 3.2 3.4 1.4
V s (107 cm/s) 1 2 3 1.3
μ (cm2 /Vs) 1350 800 1300 6000
E 11.8 9.7 9.5 13.1
E cr (MV/cm) 0.25 3 4 0.4
Thermal conductivity, k (W/cmK) 1.5 4.9 1.3 0.5
The drain current is flow through channel, formed through two-dimensional electron
gases (2DEG). Due to different bandgap of AlGaN and GaN layer, 2DEG is formed at
the interface mobility of electrons (carriers) present in 2DEG is higher than mobility
of electron in other three terminal devices [4]. Due to higher mobility of carrier, some
of the carriers move toward Schottky gate side, and due to this, the leakage current
of the device is increased. Metal insulator semiconductor HEMT (MIS-HEMT) is
used to reduce the gate leakage current and enhance the performance of the device.
Different insulators are used in MIS-HEMT are Si3 N4 , Al2 O3 , HfO2 .
Above dielectric is used to reduce leakage current. In the proposed design, Al2 O3
is used as a gate passivation. And AlN layer is used to interfacial layer between
dielectric and AlGaN layer. Lattice of AlGaN and dielectric do not match, so why
AlN layer is used [6].
In AlGaN/GaN HEMT, channel is formed at the interface of AlGaN and GaN. Due
to interface roughness scattering at the interface of AlGaN and GaN, the mobility
of carrier will reduce. A spacer layer (AlN) is used to enhance the carrier mobility,
current density at the (2DEG), transconductance and cutoff frequency [7].
The aim of this paper is to propose and discuss the spacer layer AlGaN/GaN
HEMT and compare it with AlGaN/GaN HEMT without spacer. The analog perfor-
mance of both the devices is discussed, and the results are compared. The spacer
layer AlGaN/GaN HEMT provides better drain current, transconductance, cutoff
frequency and ON-resistance.
2 Device Structure and Simulation Parameters
A larger bandgap material is grown on lower bandgap material for HEMT design.
AlGaN/GaN HEMT is designed by growing AlGaN layer on top of GaN layer [8].
Due to the discontinuities in the conduction band between AlGaN and GaN, the
2DEG is formed at the interface [9, 10]. A potential well is created at the GaN side
of device. Because of the uniform lattice in GaN, spontaneous polarization field is
generated. When larger lattice size material (AlGaN) is grown upon lower lattice size
material (GaN) than strain is induce at the AlGaN side and due to this spontaneous
and piezoelectric polarization field is generated at the AlGaN [11]. A electric field is
Effect of AlN Spacer Layer … 1117
Table 2 Parameters
Parameters Proposed HEMT
AlGaN doping Nil
GaN doping Nil
AlGaN thickness (nm) 20
L G (µm) 1.2
Work function of gate (eV) 5.05
Al2 O3 thickness (nm) 4
Device length (µm) 2
AlN thickness (nm) 2
generated in the AlGaN layer because of piezoelectric polarization. To compensate

the positive charges at the AlGaN, electron is generated at the GaN side. There is not
any requirement of doped AlGaN for generation of 2DEG at the interface (Table 2).
In the proposed HEMT, is shown in Fig. 1. AlGaN layer is grown upon GaN with
the composition (mole fraction) of Al in GaN is 0.25. No any external doping is
required to AlGaN for formation of 2DEG. Si is used as a substrate [12]. GaN is
deposited on the Si substrate with thickness of 0.184 µm. On the top of the GaN,
spacer AlN layer is grown with thickness 1 nm. Due to different lattice of GaN and
AlGaN, the channel electron degradation takes place at the channel. The 2DEG wave
function is mostly present in GaN, but there is a small part of wave function that
penetrates the AlGaN barrier, leading to alloy-disorder scattering. AlN layer reduces
(a)
(b)
Spacer
Fig. 1 a AlGaN/GaN HEMT without spacer b AlGaN/GaN HEMT with spacer

the alloy-disorder scattering. Hence, mobility of electron increases in 2DEG, and

carrier density increases at the interface. The spacing between the source and drain
to gate is 10 nm. The second source and drain are present at the boundary of device
up to upper part of GaN layer for reduction in series resistance. The silicon substrate
is used for the mechanical support of device, and thickness of Si is 0.2 µm. The 2-D
schematic of HEMT with and without spacer layer is shown in Fig. 1.
3 Simulation Results and Discussion
Both the above designs (with spacer and without spacer) are simulated at same
gate voltage and work function. Fig. 2a, b shows the 2-DEG formation in ON-state
condition at the interface of two different bandgap materials such as AlGaN and
GaN interface in case of HEMT without spacer and at the interface of AlN and GaN
with spacer. In both cases of HEMT, 2DEG forms in similar fashion at the interface.
Due to reduction in alloy-disorder scattering, the depth of 2-DEG below Fermi level
is slightly more in case of spacer layer HEMT shown in Fig. 2b. Higher depth of
2-DEG, the current density at the interface is more. The higher current density and
mobility increase the drain current in case of HEMT with spacer.
Electron concentration of both the devices with spacer and without spacer is
shown in Fig. 3a, b, respectively. Because of interface roughness, the concentration
of carrier layer HEMT is slightly larger than HEMT without spacer. From the graph,
the current will flow through drain to source by only one carrier, i.e., electrons.
Figure 4a shows the transfer characteristics of designed HEMT. Due to higher
current density and mobility, the drain current of spacer layer HEMT is larger than
without spacer layer HEMT. Spacer layer HEMT shows higher I ON to I OFF ratio than
conventional HEMT. The OFF-state current of spacer layer HEMT is lower than
HEMT without spacer layer because of lower minority concentration of channel.
Fig. 2 2DEG of HEMT a without spacer b with spacer

Fig. 3 Electron concentration with respect to device length of HEMT a with spacer b without
spacer layer
Fig. 4 a Transfer characteristics of HEMT with and without spacer at V ds = 10 V. b Transconduc-

tance, gm , and drain current, Ids , with respect to V gs at Vds = 10 V
All the simulation is done on V ds = 10 V. At V gs = 5 V, the ON-state currents are

125 mA/mm and 380 mA/mm of HEMT without spacer and spacer layer, respectively.
Figure 4b shows the transconductance and drain current with V gs . Due to higher ON-
state current of the spacer layer HEMT, the transconductance of spacer layer HEMT
is larger than without spacer HEMT.
From Eq. 1, transconductance is directly proportional to the slope of log (drain
current), and due to higher drain current I ds, the transconductance of spacer HEMT
is larger than that of the without spacer HEMT. Transconductance of spacer and
without spacer layer HEMT is 96 mS/mm and 24 mS/mm, respectively.
∂ Ids
gm = (1)
∂ Vgs
Fig. 5 Drain current, I D , as a function of drain voltage, V ds , for GAA and Planar-MIS-HEMT at
V gs = 5 V. b Total capacitance, Cgg with respect to gate voltage, Vgs , at Vds = 10 V
Transconductance (gm ) and drain current (I ds ) graph with gate voltage, V gs , are
shown in Fig. 4b. Dotted line shows the drain current for both the devices that is in
the order of mA scale, and solid line shows the transconductance of both the devices.
Figure 5a shows the graph between drain current with drain voltage. Because of
higher carrier density, the drain current of spacer HEMT is larger than without spacer
layer HEMT.
Figure 5b shows the total capacitance, C gg (C gg = C gd + C gs ), of both devices
at V ds = 10 V where C gd is the gate-to-drain capacitance and C gs is the gate-to-
source capacitance. The value of C gg is higher in the case of the spacer HEMT and
in comparison with HEMT without spacer device. The switching behavior of device
depends upon total capacitance. From Eq. (2), the cutoff frequency, f T , of device
depends on transconductance and total capacitance. Cutoff frequency is directly
proportional to gm and inversely proportional to C gs .
Due to larger value of transconductance of spacer layer AlGaN/GaN HEMT, the
cutoff frequency of spacer layer HEMT is larger than without spacer layer. Figure 6
shows the cutoff frequency, f T , with gate voltage, Vgs , at Vds = 10 V.
gm
fT = (2)
2πCgg
4 Conclusion
AlGaN/GaN HEMT is design for microwave and RF frequency application. Spacer

layer AlGaN/GaN HEMT is compared without spacer layer in terms of their RF
and analog characteristics. AlN is used as a spacer layer between AlGaN and GaN
layer. By using spacer layer, the depth of 2DEG is higher than HEMT without spacer.
Fig. 6 Cutoff frequency, f T ,

with respect to gate voltage,
V gs at V ds = 10 V
Dopingless AlGaN layer is used to facilitate the formation of 2DEG at the interface
of GaN. Carrier density and mobility of carriers are larger in spacer layer HEMT.
Spacer layer HEMT provides better analog performance than HEMT without spacer
layer. ON-state current, transconductance, ON-resistance and cutoff frequency of
GAA-HEMT is higher as compared to HEMT without spacer.
References
1. Wong YY, Chiu YS, Luong TT, Lin TM, Ho YT, Lin YC, Chang EY (2012) 10th IEEE
international conference on semiconductor electronics (ICSE), pp 729–732, IEEE
2. Xiao-Guang H, De-Gang Z, De-Sheng J (2015) Chinese physics B 24(6):067301
3. Hsieh TE, Chang EY, Song YZ, Lin YC, Wang HC, Liu SC, Salahuddin S, Hu CC (2014) IEEE
4. Hwang IH, Eom SK, Choi GH, Kang MJ, Lee JG, Cha HY, Seo KS (2018) Physica status solidi
(a) 215(10):1700650
5. Roccaforte, Fabrizio, Greco G, Fiorenza P, Iucolano F (2019) Materials 12(10)
6. Oka T, Nozawa T (2008) IEEE Electron Device Lett 29(7):668–670
7. Ye PD, Yang B, Ng KK, Bude J, Wilk GD, Halder S, Hwang JCM (2005) Appl Phys Lett
86(6):063501
8. Ohno Y, Kuzuhara M (2001) IEEE Trans Electron Devices 48(3):517–523
9. Palacios T, Chakraborty A, Heikman S, Keller S, DenBaars SP, Mishra UK (2005) IEEE
10. Kumar TV, Balamurugan NB (2018) AEU-Int J Electron Commun 94:19–25
11. Radhakrishnan SK, Subramaniyan B, Anandan M, Nagarajan M (2017) Int J Electron Commun
2018 83:462–469. https://doi.org/10.1016/j.aeue.2017.10.029
12. Ranjan R, Kashyap N, Raman A (February 2020) Appl Phy A
Enhancement of Isolation in MIMO
Antenna System
Neelima Koppala, Sree Lochan Kanthi Kumar Kuppachi,

Rohith Sai Kumar Gajula, Bhanu Ravindranath Kasturi,
and Rupa Sree Jangam
Abstract In this presented document, the analysis of the isolation is improved with
the help of an antenna array which consists of eight radiators divided into four ports.
Each port consists of two radiators for which we are going to increase the isolation.
The number of antennas is greater than one which leads us to the multiple-input
multiple output (MIMO). MIMO is the thing that we are used for achieving high data
rates for the upcoming smart mobiles. This design purely deals with providing the
isolation between any two antenna elements. There are many techniques to improve
isolation, among them this document deals with the defective ground structures,
parasitic elements. This design operates at a frequency of 3 GHz with isolation
parameter range of (−26.0619 to −42.9525) dB.
Keywords MIMO · 5G · Isolation · Dual polarization
1 Introduction
In modern life, high data rates and channel band with are the key parameters of the
communication system. To influence these parameters, we need to study modern
N. Koppala
Department of Electronics and Communication Engineering, Center of VLSI and Embedded
Systems, Sree Vidyanikethan Engineering College, Tirupathi, India
S. L. K. K. Kuppachi (B) · R. S. K. Gajula · B. R. Kasturi · R. S. Jangam
Department of Electronics and Communication Engineering, Sree Vidyanikethan Engineering
College, Tirupathi, India
R. S. K. Gajula
B. R. Kasturi
R. S. Jangam
1124 N. Koppala et al.
communication systems which includes MIMO. This MIMO technology has more
advantages than single input devices, comparatively MIMO has best channel band
width and data rates. The main challenge in MIMO is keeping isolation high between
many elements in antenna array. MIMO was been introduced at BELL laboratories
at period called the BLAST period, which stands for Bell-Layered Space Time.
Because of the mutual coupling between any two elements and high similarity
in channel degrades the performance of MIMO antenna system [1–4]. From past
years, researchers have developed many methods to overcome the mutual coupling.
Among them, we used two methods. They are 1. Defected ground structure and 2.
Parasitic elements [5] isolation [6] and the defected ground structures in turn allows
the antenna to radiate in both sides.
The slotted antenna element contains a square ring as a defective ground structure
and antenna fed with rectangular micro strip lines. In advance, we have a circular-
shaped ring that acts as a parasitic element which is used to reduce mutual coupling
between elements. In the design, we have placed these antennas and ground as a
slots at the ends of square-shaped structure [7]. The single element slot antenna
fundamental properties have been discussed in the following Sections.
2 Single Element Antenna
In this single element antenna, we are going to deal with the single port from four
ports total antenna which consists of eight element antenna.
Figure 1 shows the single slotted antenna with two parasitic elements, two feed
elements and one ground structure. Here, the substrate used is a FR4 dielectric with
a height of 1.6 mm. The ground structure is taken in the form of a rectangular shape
at the bottom of the substrate. Two antennas are placed at the corner of the substrate
with help of design parameters so that it passes from the structure of ground placed
bottom of the substrate. Two circular-shaped rings are taken into consideration that
these are acts as the parasitic elements which can radiate, through which we can
reduce mutual coupling (Table 1).
This table shows the design parameters for single slot antenna L f represents the
length of feed, W f represents the width of the feed, W represents the length of outer
rectangle in ground plane, W 1 represents the length of inner rectangle in the ground
Fig. 1 The proposed single slot antenna

Enhancement of Isolation in MIMO Antenna System 1125
Table 1 The values of the

Parameter Dimension (mm)
design parameter
W 13.4
hs 1.6
Wf 3
Lf 11.75
Ws 30
W1 11.9
W2 13.15
S 0.75
S1 0.75
L1 9
d 5.95
r 3
plane, S is the width of ground plane rectangle, r is the radius of the circular ring
which acts as a parasitic elements.
Figure 2 shows the development of single element antenna in stages; this
development is done in HFSS design.
Substrate contains a material called FR4 dielectric substrate which is used because
it can be easily available and can be designed easily with less cost; its dimensions
mainly consists of 1.6 mm height and a width of (75 × 75) mm2 .
Figure 3 shows the excitation of radiators in single element antenna, (a) shows
excitation of one element where the other is not excited, (b) shows the excitation of
other element where the first one is kept constant.
Main inference from this Figs. 3 and 4 is that the voltage level is increased from
0.7 to 1 V which means by increasing the elements in the design, we can able to
increase voltage levels.
Considering the results of this antenna, we have best return loss at s(1, 1) when
both the radiators are excited and the value is −14.4429 dB where the standard value
goes like this it should be more than −10 dB. Voltage standing wave ratio (VSWR)
is best when both are excited in a frequency of 3 GHz which is 1.4679, usually the
VSWR must be in a range of 1–2.
Considering the isolation parameter which how much one element is different
from other, to obtain the good performance, usually this isolation parameter value
must be more than −15 dB.
We have the isolation parameter value as S(1, 2) = −18.414 dB, which is also the
best one at frequency of 3 GHz range.
Fig. 2 a Ground plane present at the bottom of substrate, b extends by adding feed, c added with
parasitic elements and d final structure
3 Introduction to the Eight Element Antenna
First, we proposed a two element radiating element antenna, for the purpose of the
using this in the smart phone or the upcoming livelihood the need for the antennas
increases drastically. So we designed an eight element antenna [8] to increase the
performance characteristics of the device and as well as to increase the device
capability.
While comparing the values of the parameters like gain, the eight element antenna
meets the required levels.
The metrics that can be used to compare the designs of EDAC Codes for memories
are given as below.
The above design is the eight element antenna design that has been designed
[9] to meet the required levels. The substrate used is the Fr-4 material. The design
challenges that are faced to design the model are:
1. Mutual coupling
2. In situ performance
Fig. 3 a Single element excitation, b second element excitation
Fig. 4 2 elements radiated

3. Antenna size.
1. Mutual coupling:
Mutual coupling is the interference of the radiation pattern between the two
element antenna array.
2. In situ performance
In situ performance is the characteristic representation of an every element in an
array within itself.
3. Antenna size
The size of the antenna is a design challenge because we are designing this design
to an smart phone device; size compatibility is a must in those designs.
4 Design of the Element Antenna
Figure 5 displays the schematic of the designed [10] 8 element antenna. The designed
is arranged on an FR-4 dielectric medium which has a permittivity of 4.4 and the
overall dimension is of 75 × 75 mm2 and an dual-polarized square ring slot antenna
of size 27 × 27 mm2 (Table 2).
Fig. 5 The eight element antenna

Table 2 Design parameters

Parameter Dimension (mm)
W 13.4
hs 1.6
Wf 3
Lf 11.75
Ws 30
W1 11.9
W2 13.15
S 0.75
S1 0.75
L1 9
d 5.95
r 3
5 Performance Characteristics
The output of the design is showed in the 2D and 3D polar plots (Fig. 6).
Figure 7 shows the 3D polar plot of an eight element radiators radiation pattern
when all the antennas are radiated at a time which consists of 2.8 V of voltage level
where as Figure 8 shows the radiation pattern when the single port is radiated from
four port or eight element antenna which exists in the planar mode, when all the ports
are radiated which explains that the design is directional so that this can be viewed
Fig. 6 2D polar plot

Fig. 7 3D polar plot (all ports exited)
Fig. 8 3D polar plot (one port exited)
in MIMO. As we are designing this to the smart phone 5G networks, the MIMO
concept is very useful to give high data rates.
The main parameter of this design is to give the isolation between 2 antenna
elements. Figure 5 shows the eight element antenna design. The isolation should be
more at the nearby elements 1 and 2, 3 and 4, 5 and 6, 7 and 8. The required isolation
should be showed between these pairs; hence, they are the nearby elements; the rest
of the elements are far away & the isolation is quite normal for them. For proving
the isolation between these elements, we are going for the S-parameter due to its
measurement of the power. The next following graphs show the isolation between
those pairs.
The minimum isolation between two antenna elements should be −15 dB. The
above graphs show the isolation between these pairs and the isolation is of the values.
From Fig. 9, it shows S(1, 2): −26.6420 dB, from Fig. 10 it shows S(3, 4): −
25.4492 dB, from Fig. 11 it shows S(5, 6): −26.7471 dB and from Fig. 12 it shows
S(7, 8): −26.0619 dB these values show that more than −15 dB which is best.
Fig. 9 S(1, 2)
Fig. 10 S(3, 4)
Fig. 11 S(5, 6)
Fig. 12 S(7, 8)
According to this designed model, the antenna elements shows a good perfor-
mance in S parameters at 3 GHz.
6 Conclusion
The structure of the proposed MIMO antenna consists of dual-polarized square ring
slot antenna placed at the corners of smart phone PCB which intern reduces mutual
coupling antenna elements, these circular rings act as the parasitic structures at the
top of the dielectric substrate, rectangular ground structures are placed at the bottom
of substrate acts as defective ground structures by using parasitic, defective ground
structures, the mutual coupling is reduced which in turn increases isolation which is
in a range of −26.0619 to 45.9525 dB at operating frequency 3 GHz. This design
antenna offers good performance characteristics at 3 GHz applications.
References
1. Qin Z, Wen GY, Zhang M, Wang J (March, 2016) Printed eight-element MIMO system for
compact and thin 5G mobile handset. Electron Lett 52(6):416–418
2. Hussain R, Alreshaid AT, Podilchak SK, Sharawi MS (2017) Compact 4G MIMO antenna
integrated with a 5G array for current and future mobile handsets. IET Microw Antennas
Propag 11(2):271–279
3. Al-Hadi AA, Ilvonen J, Valkonen R, Viikari V (2014) Eight-element antenna array for diversity
and MIMO mobile terminal in LTE 3500 MHz band. Microw Opt Technol Lett 56(6):1323–
1327
4. Wong KL, Lu JY, Chen LY, Li WY, Ban YL, Li C (December, 2015) 16 antenna array in the
smartphone for the 3.5 GHz MIMO operation. In: 2015 Asia-Pacific microwave conference
(APMC), Nanjing, China, pp 1–3
5. Statement: improving consumer access to mobile services at 3.6–3.8 GHz. Accessed: Oct. 21,
2018
6. Li Y, Zou H, Wang M, Peng M, Yang G (March, 2018) Eight-element MIMO antenna array for
5G/Sub-6 GHz indoor micro wireless access points. In: Proceedings of international workshop
on antenna technology (IWAT), Nanjing, China, pp 1–4
7. Yang HH, Quel YQS (2017) Massive MIMO meet small cell. In: Springer briefs in electrical
and computer engineering. Springer, Cham, Switzerland. https://doi.org/10.1007/978-3-319-
43715-6_2
8. Nadeem Q-U-A, Kammoun A, Debbah M, Alouini M-S Design of 5G full dimension massive
MIMO systems. IEEE Trans Commun 66(2):726–740
9. Osseiran A, Boccardi F, Braun V, Kusume K, Marsch P, Maternia M, Queseth O, Schellmann
M, Schotten H, Taoka H, Tullberg H, Uusitalo MA, Timus B, Fallgren M Scenarios for 5G
mobile and wireless communications: the vision of the METIS project. IEEE Commun Mag
52(5):26–35
10. Sun L, Feng H, Li Y, Zhang Z Compact 5G MIMO mobile phone antennas with tightly arranged
orthogonal-mode
Stochastic Frontier Model Approach
of Technical Efficiency for Paddy
Farming in Odisha
Priyabrata Bhoi, Deepak Kumar Swain, Subhadra Mishra,

Debahuti Mishra, Gour Hari Santra, and Samarjeet Borah
Abstract Paddy production with utilising efficient technology will boost the produc-
tion and economy simultaneously. Technical efficiency approach is adopted for
perfect utilisation of inputs. The study is carried out between the years 2001 and
2016. The secondary data are collected from DES, India, and for Odisha, the data
are generated from scheme “Comprehensive scheme to study cost of cultivation of
principal crops in Odisha” conducted by OUAT, Bhubaneswar. Different parame-
ters are considered, and the efficiency is obtained by stochastic frontier model or
stochastic frontier production function. Total fertiliser and casual labour show statis-
tical significant effect. The significant shifting of mechanisation from 2001 to 2016
can be visible clearly.
P. Bhoi
Department of Agricultural Economics, FAS, Siksha ‘O’ Anusandhan Deemed to be University,
Bhubaneswar, Odisha 751030, India
D. K. Swain
Department of Agricultural Statistics, FAS, Siksha ‘O’ Anusandhan Deemed to be University,
Bhubaneswar, Odisha 751030, India
S. Mishra
Department of Computer Science and Application, CPGS, OUAT, Bhubaneswar, Odisha 751030,
India
D. Mishra (B)
Department of Computer Science and Engineering, ITER, Siksha ‘O’ Anusandhan Deemed to be
University, Bhubaneswar, Odisha 751030, India
G. H. Santra
Department of Soil Science and Agricultural Chemistry, FAS, Siksha ‘O’ Anusandhan Deemed to
be University, Bhubaneswar, Odisha 751030, India
S. Borah
Department of Computer Applications, Sikkim Manipal University, Gangtok, Sikkim, India
1136 P. Bhoi et al.
Keywords Production function · Stochastic frontier model · Cobb–Douglas

function · Technical efficiency · Paddy production
1 Introduction
One of the poor faces of India is Odisha. The livelihood of 70% people depends
on agriculture. Rice is staple crop and food of the people of Odisha. The State
possess 29.14 lakh ha of high land, 17.55 lakh ha of medium land and low land
of 15.11 lakh ha summing up to 61.80 lakh ha of cultivated area. Paddy cropping
covers around 41.24 lakh ha and 3.31 lakh ha area in kharif and rabi, respectively
(DAFE, BBSR). To feed the hunger, the production should be alleviated efficiently
by judicious utilisation of the inputs, due to their scarcity. The knowledge of technical
efficiency of different crops is an important tool for policy formulation. Production
performance of crops can be improved from past efficiency observation. Efficiency
studies have corroborated the possibility of increasing the crop productivity without
incremental in input application which is described by Ali and Chaudhry [1], Umesh
and Bisaliah [2] and Gaddi et al. [3]. The efficiency of production completely depends
on the kind of technology used and the potential of producer explored to exploit it.
The production growth rate can be increased through improved technology, efficient
use of technology, facilitation of technological innovations via institution and rural
infrastructure development through public investment explained by Kalirajan and
Shand [4]. Resource use pattern and production efficiencies of various farms provide
guidelines for national policy and planning as well proportional resource allocation
to different farms. An economy can be benefited in several ways from inefficiency
studies as it is one of essential factor for growth of productivity. The estimation on the
extent of inefficiency can also aid in finding out the potential technology to enhance
the agricultural productivity. Kalirajan et al. [5] confirmed that boosting technical
efficiency is a prime approach to harvest the potential profits of the existing tech-
nology rather than setting for new technology. Shanmugam and Venkataramani [6]
found that without increasing the input application the increment of crop production
is also possible. Determinants of technical efficiency and allocative efficiency can be
accomplished by stochastic frontier production function conferred by Mruthyunjay
et al. [7]. A study was conducted on the productivity of resource and technical effi-
ciency of paddy in Thrissur district of Kerala, where stratified random sampling
technique was used by Suresh and Reddy [8] to collect the farm-level data of 71
farmers.
Stochastic Frontier Model Approach … 1137
2 Materials and Method
2.1 Data Collection
The study used the secondary data compiled from the Department of Economics and
Statistics (DES), Ministry of Agriculture, Govt. of India. Odisha has generated data
from the scheme “Comprehensive scheme to study cost of cultivation of principal
crops in Odisha” being implemented by the Orissa University of Agriculture and
Technology, Bhubaneswar. The current study perused data of input use and output
produced in paddy crop for the year 2001–02 and 2016–17, a gap of 15 years for
technical efficiency estimation and comparison.
2.2 Technical Efficiency
The Cobb–Douglas form of stochastic frontier production function can be employed

to obtain technical efficiency of different crops at farm level. Meeusen et al. [9]
and Aigner et al. [10] firstly published this model. The important feature of this
model is estimation of technical efficiency at farm level for different data types like
cross-sectional and panel data by Battese and Coelli [11]. The disturbance term splits
into a two-sided random error, in which one will check random effects beyond the
control of firm, another is efficiency component. The R Frontier package 1.0 which
is developed by Coelli et al. [12] estimate the ML estimates and individual farm’s
technical efficiencies.
The Cobb–Douglas form of stochastic frontier model is given by,

n
β
Yi = β0 × X i jj × e Dki +(v−u)
j=1
where
Y = Output in q ha−1 or ha −1
X i = Vector of inputs
β j = Estimated coefficient of ith input
Dki = Vector of kth dummy variable
vi = Asymmetrical random term and assumed to be normally

1138 P. Bhoi et al.

distributed N (0, σv2
u i = Farm-specifictechnical inefficiency assumed to follow

.
a half normal distribution
To obtain technical efficiency, the farm will have its own stochastic production
frontier f (X i , β) exp(vi ); a deterministic part f (X i , β) common to all farm and
producer-specific part exp (vi ). The technical efficiency can be obtained as follows,
f (X i , β) exp(vi − u i )
TEi (μi ) = = exp(−u i )
f (X i , β) exp(vi )
where
f = Cobb-Douglas form of the production function
TE(μi ) = Technical efficiency of individual farm.
The technical efficiency in paddy was based on estimated parameters which are given
in Table 1.
In 2001, fertiliser, seed and casual labour were the main inputs which when applied
in increased amount improved the yield significantly. In that period, farmers were
not using mechanisation which is clear from the coefficient which is zero. σ 2 value
showed highly significant which means the model is stochastic rather than determin-
istic. γ value is almost near to 1 (0.97***), dominance of efficiency over random
errors. This means, the factors like fertiliser, seed and casual labour were the major
component of technology in 2001, and their utilisation management was responsible
for 97% variation in the yield of paddy. It also indicates that the technology fron-
tiers must be changed for further yield augmentation as the current technology was
saturated and only 3% possibility was there for other factors to have any effects on
yield.
In 2016, the scenario changed due to shift in the technology level as 15 years is
a period long enough time for technology frontiers to change. From the table, it is
evident that in 2016, mechanisation has already been adopted by Odisha farmers, and
it has highly significant impact on yield; however, seed and fertiliser have now lost
its significance in affecting yield. σ 2 value showed highly significant which means
the model is stochastic rather than deterministic. γ value is almost near to is very
low (0.35) and is non-significant which indicates the dominance of random factors
Stochastic Frontier Model Approach … 1139
Table 1 ML estimates of
Parameters 2001 2016
stochastic frontier production
function of paddy for the (Intercept) 3.92*** 1.38***
years 2001 and 2016, Odisha (0.31) (0.14)
Family labour (Man versus hours) 0.01 0.011
(0.00) (0.009)
Casual labour (Man versus hours) 0.01** −0.002
(0.00) (0.002)
Machine labour (Hours) 0.00 0.009***
(0.00) (0.001)
Total fertiliser (Kg) 0.21*** 0.49***
(0.01) (0.019)
Seed (Kg) 0.28*** −0.005
(0.06) (0.017)
σ2 0.09*** 0.013***
(0.01) (0.003)
γ 0.97*** 0.35
(0.01) (0.28)
Log likelihood value 157.19 39
Mean efficiency 0.82 0.95
* is a significance level of 0.1 indicates a 10% risk of concluding
that a difference exists when there is no actual difference. Similarly
for ** and *** we have 5% and 1% risk of conclusion respectively.
not included in the model over input technical efficiency. This also indicates that
the farmer has the possibility of increasing their current level of input efficiency
to a higher level before any technological change is required. The current level of
technology must be exploited for further augmentation of yield in paddy.
The similar comparative study on paddy crop was carried out in Punjab by Bhoi
et al. [13] in which the technological frontiers, efficiency and maximum likelihood
estimates were used.
4 Conclusion
The present study is based on stochastic frontier production approach for technical
efficiency study in paddy production. MLE estimates of different parameter clarified
the scenario of the study. The visible difference of adoption of mechanisation from
2001 to 2016 has great impact on production of paddy.
1140 P. Bhoi et al.
References
1. Ali M, Chaudhry MA (1990) Inter-regional farm efficiency in Pakistan’s Punjab: a frontier

production function study. J Agric Econ 41:62–74
2. Umesh KB, Bisaliah S (1991) Efficiency of groundnut production in Karnataka: frontier profit
function approach. Ind J Agric Econ 46:20
3. Gaddi GM, Mundinasmani SM, Hiremath GK (2002) Resource use efficiency in groundnut
production in Karnataka-an economic analysis. Agric Sit Ind 58:517–522
4. KalirajanK P, Shand RT (1989) A generalized measure of technical efficiency. Appl Econ
21:25–34
5. Kalirajan KP, Obwona MB, Zhao S (1996) A decomposition of total factor productivity growth:
the case of Chinese agricultural growth before and after reforms. Amer J Agric Econ 78:331–
338
6. Shanmugam KR, Venkataramani A (2006) Technical efficiency in agricultural production and
its determinants: an exploratory study at the district level. Madras Sch Econ 10
7. Mruthyunjaya SK, Rajashekharappa MT, Pandey LM, Ramanarao SV, Narayan P (2005)
Efficiency in Indian edible oilseed sector: analysis and implications. Agric Econ Res Rev
18:153–166
8. Suresh A, Reddy TK (2006) Resource-use efficiency of paddy cultivation in Peechi command
area of Thrissur district of Kerala: an economic analysis. Agric Econ Res Rev 19:159–171
9. Meeusen W, Van den Broeck J (1997) Efficiency estimation from Cobb-Douglas production
functions with composed error. Int Econ Rev 435–444
10. Aigner D, Lovell CK, Schmidt P (1997) Formulation and estimation of stochastic frontier
production function models. J Econ 6:21–37
11. Battese GE, Coelli TJ (1991) Frontier production functions, technical efficiency and panel data:
with application to paddy farmers in India. Department of Econometrics, University of New
England, Australia
12. Coelli T, Henningsen A, Henningsen MA (2013) Package ‘frontier’
13. Bhoi PB, Singh J, Sachdeva J (2017) Farm specific technical efficiency in paddy production
and its determinants in Punjab. Ind J Econ Dev 13(3):493–499
Improving Security Concerns of Mobile
Cloud Computing Environment
with Quantum Cryptographic
Techniques
Sudhanshu Maurya , Kuntal Mukherjee , and R. Badlishah Ahmad
Abstract In the current scenario, mobile cloud computing (MCC) is termed as

speedily developing field of IT. With the development of mobile phones and success
of the Internet, computational resources have become cheaper, easily accessible,
ubiquitous, distributed, and powerful tool for all types of organizations. Mobile
cloud computing is a computational model in which services are offered on mobile
phones in an on-demand fashion. Most of the service provider are concerned with
privacy and security issues. In this endeavor, security concerns of MCC are discussed
and proposed a new quantum cryptography (QC)-based encryption and decryption
method. Proposed technique allows the users to store and process the user’s data
through strengthened cryptographic method.
Keywords Security · Mobile cloud computing · Quantum cryptography
1 Introduction
Amalgamation of Internet, mobile device, and cloud computing (CC) environment

is referred as mobile cloud computing (MCC). Hence, the services offered by cloud
for mobile devices are named as MCC [1]. It uses the technology of CC to deliver
services as well as applications to mobile phones [2]. This amalgamation brings
out advanced computational resources at the fingertips of mobile phone users and
cloud providers [3]. The concept of the cloud computing which is offered by grid
computing, application service provisioning, to makes it easier for its client to utilize
S. Maurya (B)
School of Computing, Graphic Era Hill University, Bhimtal Campus, Dehradun, Uttarakhand,
India
S. Maurya · R. B. Ahmad
School of Computer and Communication Engineering, Universiti Malaysia Perlis (UniMAP),
Arau, Malaysia
K. Mukherjee
Department of Computer Science, BIT Mesra, Ranchi, Jharkhand, India
1142 S. Maurya et al.
the services offered by cloud environment [4, 5]. Google proposed cloud computing
in the year of 2007 and dragged a prodigious attention from idea phase to real-life
utilization phase in a very short span of time [6]. It extended the capacity of storage
and computing, the ability of data service and system trading capacity of mobile
devices. Google beliefs that the CC must provide clients data and data storing in
secure, easily accessible and credible manner [7]. CC provides the facility to update
application and services as per the requirement and concern of the users [8]. In simple
terms, CC is an effervescent environment where user receives the resources as utility
over the Internet [9]. It comes with many perks like efficiency, flexibility, scalability,
etc. It uses “pay-per-use” policy for delivery of user desired services over the cloud
environment [10]. Despite of these positives, it is severely affected by privacy and
security issues [11]. This legacy of limitations is further carried forward by mobile
cloud computing. It is because of scarcity that strong algorithms for resource security
cannot be implemented with mobile phones [12]. There is a need of deploying well-
structured job scheduling algorithms between cloud and mobile nodes to solve this
issue [13]. Unfortunately, most of the users are not aware of these issues associated
with cloud computing and mobile cloud computing [14].
This manuscript presents the security aspects of MCC and proposed novel encryp-
tion and decryption technique based on quantum cryptography. In literature review
section, basics of security, cryptography, security requirements have been discussed.
In next section, we propose a new cryptographic technique for encryption and
decryption. Further, summarizing the paper with result analysis and conclusion.
2 Literature Review
Hidden script communiqué has been proficient for almost four thousand years ago.
In Egypt, a secret script has been outlined around 1900 B.C [15]. Cryptography has
been used ever since there was a need to hide a message. In the ancient times, people
replaced the original words with digits or symbols [16]. Many other classical cryp-
tographic methodologies were implemented in the yester decade. ISO 27001 has
defined cloud computing security as—“safeguarding data’s confidentiality, avail-
ability as well as its integrity.” Authenticity, non-repudiation, accountability and
reliability. Security in cloud computing is a new and emerging research sector as a
subdisciple of network security, computer security and information security.
It comprises of technologies, policies, data protection and such related cloud
infrastructure. In another sense, CC has being associated to large number of busi-
nesses inventively. Productivity of these varied methods leads to privacy and security
risks [17–19]. Security is referred to as an arrangement that decides for the privilege
of accessing and altering information.
Security in cloud is divided into two broad visions: Making of a security strategy
and then checking for the successful accomplishment of the security goal as per
that strategy by code execution. The strategy differentiates sharply that who has the
permission to carry out what operation and on what data set through validation and
Improving Security Concerns of Mobile … 1143
authorizations. After adaptation of cloud cervices client has not confident about the
integrity, recovery of data, privacy and security of data [20]. A protected framework
is a structure where information cannot be seen or changed by a customer without
appropriate approval being explicitly endorsed by the security technique [21].
2.1 User’s Security Requirements
Security in cloud processing has various implications. For instance: having a

protected cloud from a supplier perspective does not imply that it is secure for
clients point. Furthermore, unfortunately, greatest clients deliberate that with client
username and a secret key to get to their record is sufficient protected [22].
Confidentiality: It alludes to the protection information undisclosed. It is upheld by
specialized instruments like encryption and control on access and legitimate backings
too [20]. It gives an answer for some beneath referenced treats:
Unsecured Cloud Service User Access: As we know that most of the services and
resources delivered through a remote connection. There is always the risk of fraud,
phishing and others. Username, password and other credentials are often reused,
which may increase the impact of these attacks [18, 23].
Data leakage: There are two prospects in this, first some organizations might store
their data at servers of another nation. Second, even after deletion by the client, data
may be remains on the provider’s server [24].
Availability: It alludes to utilize the framework at whatever point required. It is
bolstered by acceptable design and limit working by the cloud supplier, notwith-
standing that all around characterized arrangements and terms of the understanding
are likewise required [20]. Approximate all of the cloud service providers’ offer
availability for their servers up to 99.99% [25].
Accountability: It maps the activities in the framework to responsible gatherings
and bolstered by get to control, verification and vigorous personality notwithstanding
that the capacity to record associations and critically review the tracks and record
pernicious ones [20].
2.2 Provider’s Security Requirements
The criteria of security may be minimal distinctive when contrasted with clients: At
one side, it is tied in with picking up the client’s belief, with a decent notoriety in the
cloud advertise. On next side, cloud suppliers dropping their command over the infor-
mation they are developing. Keep information secured on own servers, exchanging
Table 1 Analysis of security preventing methods

Category Solution Users Providers
Security
Confidentiality Availability Accountability
Formal Clear SLA Partially Yes Yes Partially
approaches Data protection Partially Partially Yes Partially
approaches/policies
Managerial Data life cycle Yes Yes Partially Yes
approaches Virtual machine Partially Does not Yes Yes
image management impact
system
Client based privacy Yes Partially Yes Partially
manager
of information to other, reinforcement of their whole architecture or erasing the infor-

mation. Every one of these concerns must be settled on according to the decision of
the suppliers and execution of it must be leveled out. A large portion of the busi-
nesses and analyst use cryptographic procedure for information security. Encryption
is a very notable method for information security, comprehensively named secrecy,
legitimacy and information protection in a cloud domain with the assistance of keys
[26]. At execution part information encryption is finished by the encryption key and
information is decoded by the utilization of decryption key.
In Table 1, examination of few security forestalling methods like Clear SLA [27,
28], data protection approaches/policies [29, 30], data life cycle [31], virtual machine
image management system [30, 32] and client-based privacy manager [33] for all
stake holders of the cloud has been discussed.
Data encryption method is effectively performed when information/data are little
in dimensions, yet it may get quite a while the information is huge. This method in
distributed computing condition can influence its start to finish execution. Some of
the time it is difficult to send and actualize in the cloud [34]. In this manuscript, a new
cryptographic technique is being proposed which is based on quantum mechanics.
Quantum mechanics was originally developed as a theory to describe the behaviors
of microscopic substances like electrons or atoms [35, 36].
Ambiguity in the theory of quantum mechanics has directed toward a new
paradigm referred as quantum cryptography (QC) [37]. It was proposed by Brennet
and Brassaed, based on the no-cloning theorem in the 1970s [38], it means a qubit
cannot be copied or manipulated without changing its actual stage [37]. QC explores
the fundamentals of quantum physics where it is impossible to quantify a state of a
random polarized photon congaing information with zero disturbance [39]. It gives
us perfectly secure data transfer. Various opportunities and market trends of mobile
cloud computing have been discussed in the research work of [40] and also give a
roadmap for security measures.
2.3 Proposed Quantum Encryption and Decryption

Algorithm
The criteria encryption has always been depicted as better resolution to secure data.
As per Sosinsky, symmetric cryptosystem is efficient for cloud environment. Cloud
being handling bulks of data is the reason behind the introduction of symmetric
cryptosystem based on quantum cryptography in this work [41].
A qubit may take binary data simultaneously that is zero and one. In Dirac
representation, it is represented as:
|Ψ ≥ α|0 > +β|1 >

1 0
where: |0 ≥ and |1 ≥ and α, β are composite quantities having the
0 1
credibility as binary numbers that is 0 and 1 discretely
[42]. According to Dirac
1 0
equation as per the free particle solution, the state , while ups to state that
0 1
denotes the roll down of subatomic elements. Finally,
after merging both states, we
10
get a table like structure of the matrix D = .
01

1 −1
We acknowledge zero as minus one and get Dw = .
−1 1
Here, w is served as dimensions of the matrix in 2w form. For higher value of Dw ,
D = D ⊗ D is needed.
The example of the matrix of size 4 × 4 is given in Fig. 1.
Generate a random natural number and replace “−1” by that number. Hence,
obtained matrix Dw is considered as the key matrix for encryption and decryp-
tion operations. Algorithm 1 is responsible for encryption and Algorithm 2 for the
decryption.
Fig. 1 Matrix D4 × 4 , where

−1 is replaceable
3 Experimental Setup and Result Analysis
Overall proposed algorithms are implemented in a cloud environment consisting

of a master–slave node system. Scheduling of jobs to different slaves is the work
of master node. Master Node is Acer Swift 5 Corei7-1065G7 with 16 GB RAM
and SSD is 512 GB and slave nodes are ACER Aspire 3 Core i3 processor with
4 GB of RAM. A fast LAN is used for data communication between the nodes. We
have compared our proposed algorithm with the famous Hill Cipher algorithm. The
comparative observations of time taken, in encrypting and decrypting, are shown in
below figures. Figure 2 depicts the results for message length of 2 while Fig. 3 shows
the results for message length of 4.
Time complexity in encryption using Algorithm 1 is O (n2) whereas talking about
the case decryption using Algorithm 2 is O (n3). In addition to the above results, the
nature of Dirac matrices is combinatorial. On the analysis of all results shown in the
below figures, some observation made are as follows:
9000 7926
8000
7000
6000 4931
Time
5000
4000
3000
2000 1043
613
1000
0
Proposed Algorithm Hill Cipher Algorithm
Message Length = 2
Encryption 613 4931
Decryption 1043 7926
Encryption Decryption
Fig. 2 Encryption and decryption time in millisecond (message length 2)
20000 18531
18000
16000
14000
11167
Time
12000
10000
8000
6000
4000 1724 2356
2000
0
Proposed Algorithm Hill Cipher Algorithm
Message Length = 4
Encryption 1724 11167
Decryption 2356 18531
Encryption Decryption
Fig. 3 Encryption and decryption time in millisecond (message length 4)

1. The time taken in encryption and decryption differ approximately linearly with
respect to the length of message.
2. Time taken in encryption is lower than that taken in decryption.
4 Conclusion
In this endeavor, security and privacy parameters regarding cloud computing have
been discussed. Symmetric key cryptography is used in cloud for securing data. The
techniques presently in use are not sufficient so a new technique has been presented in
this paper. A comparative study of the proposed Algorithm with Hill Cipher algorithm
is done. The proposed algorithm is better at the processing time aspect of encryption
and decryption.
References
1. Sanaei Z, Abolfazli S, Gani A, Khokhar RH (2012) Tripod of requirements in horizontal hetero-

geneous mobile cloud computing. In: International conference on computing, information
systems and communications, Singapore, pp 217–222
2. Liu F, Shu P, Jin H, Ding L, Yu J, Niu D, Li B (2013) Gearing resource-poor mobile devices
with powerful clouds: architecture, challenges, and applications. IEEE Wirel Commun Mag
Special Issue Mobile Cloud Comput 20(3):14–22
3. Khan AR, Othman M, Madani SA, Khan SU (2014) Survey of mobile cloud computing
application models. IEEE Commun Surv Tut 16(1):393–413
4. Ahmed M, Sina AAI, Chowdhury R, Ahmed M, Rafee MMH (2012) An advanced survey on
cloud computing and state-of-the-art research issues. Int J Comput Sci Issues 9(1):201–207
5. Pearson S (2012) Privacy, security and trust in cloud computing. HP Laboratories, HPL-2012-
80R1
6. Li J, Liu JZ (2010) Influence of cloud computing on educational informationization of china
rural areas. In: The proceedings of information science and engineering conference, pp 281–283
7. Meng K (2008) A walk in the cloud: uncovering cloud computing. China Netw World 12–14
8. Mell P, Grance T (2010) The NIST definition of cloud computing. Technical Report. Available
Online at http://www.blogjava.net/zamber/archive
9. Furht B, Escalante A (2010) Handbook of cloud computing. Springer, New York
10. Maurya S, Mukherjee K (2015) A literature survey on mobile cloud computing. Int J Appl Eng
Res 10(79):329–334
11. Kuyoro SO, Ibikunie F, Awodele O (2011) Cloud computing security issues and challenges.
Int J Comput Netw (IJCN) 3(5):247–255
12. Malik S, Chaturvedi MM (2013) Privacy and security in mobile cloud computing: review. Int
J Comput Appl 80(11):20–26
13. Popovic K, Hocenski Z (2010) Cloud computing security issues and challenges. In: Proceedings
of the 33rd international convention MIPRO, pp 24–28
14. Chen D, Zhao H (2012) Data security and privacy protection issues in cloud computing. In:
International conference on computer science and electronic engineering (ICCSEE), pp 23–25
15. Kahn D (1996) The code breakers: the comprehensive history of secret communications from
ancient times to the internet. Enigmas Books, New York
16. Pfleeger CP, Pfleeger SL (2003) Security in computing. Prentice Hall, USA
17. Wang L, Laszewski GV, Younge A, He X, Kunze M, Tao J, Fu C (2010) Cloud computing: a
perspective study. New Gener Comput 28(2):137–146
18. Robinson N, Valeri L, Cave J, Starkey T, Graux H, Creese S, Hopkins P (2010) The cloud:
understanding the security, privacy and trust challenges. TR-933-EC, prepared for unit F.5,
directorate-general information society and media, European Commission
19. Kadam K, Paikrao R, Pawar A (2013) Survey on cloud computing security. Int J Emerg Technol
Adv Eng 3(12):239–249
20. Friedman AA, West DM (2010) Privacy and security in cloud computing. Issues Technol Innov
Center Technol Innov Brookings 3:1–13
21. Maurya S, Mukherjee K (2016) A novel method for secured cloud-health-care-system. Int J
Control Theory Appl 9(19):9291–9303
22. Meryeme A, Hanan EB (2015) Security, privacy and trust in cloud computing: a comparative
study. In: International conference on cloud technologies and applications (CloudTech), pp
23–30
23. Alvi FA, Choudary BS, Jaferry N, Pathan E (2012) A review on cloud computing security
issues and challenges. In: International conference on mobility for life, Chiang Rai, Thailand
24. Kangchan L (2012) Security threats in cloud computing environments. Int J Secur Appl 6(4):25–
32
25. Habib SM, Ries S, Muhlhauser M (2010) Cloud computing landscape and research challenges
regarding trust and reputation. In: IEEE 7th international conference on ubiquitous intelligence
and computing and 7th international conference on autonomic and trusted computing, Xian,
Shaanxi, China, pp 410–415
26. Eloff JHP, Eloff MM, Dlamini MT, Zielinski MP (2009) Internet of people, things and
services—the convergence of security, trust and privacy. In: 3rd companionable workshop—
IoPTS, Novotel Brussels—Brussels, pp 1–8
27. Barron C, Yu H, Zhan J (2013) Cloud computing security case studies and research. World
Congr Eng, London, U.K., II, pp 1287–1291
28. Arora P, Chaudhry R, Satinder WE, Ahuja P (2012) Cloud computing security issues in
infrastructure as a service. Int J Adv Res Comput Sci Softw Eng 2(1):1–7
29. Srinivasamurthy S, Liu DQ (2010) Survey on cloud computing security technical report. Indiana
University—Purdue University, Fort Wayne, Indiana. Available online at: http://salsahpc.org/
CloudCom2010/Poster/cloudcom2010_submission_67.pdf
30. Khan AW, Khan SU, Ilyasand M, Azeem MI (2012) A literature survey on data
privacy/protection issues and challenges in cloud computing. IOSR J Comput Eng 1(3):28–36
31. Pearson S, Shen Y, Mowbray M (2009) A privacy manager for cloud computing. In: IEEE
international conference on cloud computing (CloudCom 2009), lecture notes in computer
science, Springer book series cloud computing, vol 5931, pp 99–106
32. Pearson S, Charlesworth A (2009) Accountability as a way forward for privacy protection in
the cloud. In: IEEE international conference on cloud computing (CloudCom 2009), lecture
notes in computer science, Springer book series Cloud Computing, vol 5931, pp 131–144
33. Miranda M, Pearson S (2009) A client-based privacy manger for cloud computing. In: Proceed-
ings by ACM digital library of fourth international ICST conference on communication system
software and middleware—COMSWARE’09, Dublin, Ireland, Article No. 5
34. Choubey S, Namdeo MK (2015) Study of data security and privacy preserving solutions in cloud
computing. In: International conference on green computing and internet of things (ICGCloT-
2015), pp 1101–1106
35. Dirac PAM (1978) The principles of quantum mechanics, 4th edn. Oxford University Press,
1958. Reprinted-1978, Oxford, UK
36. Landau LD, Lifshitz EM (1987) Quantum mechanics (nonrelativistic theory), 3rd edn.
Pergamon Press, New York, First print-1977. Reprinted-1987
37. Falahati A, Meshgi H (2009) Using quantum cryptography for securing wireless LAN networks.
In: IEEE international conferences on signal processing system, Singapore, pp 698–701
38. Bennett C, Brassard G (2014) Quantum cryptography: public—key distribution and coin
tossing. In: IEEE international conference on computer systems and signal processing,
Bengaluru, India, pp 175–179, 1984. Reprint published by Elsevier J Theor Comput Sci
560(Part 1):7–11
39. Huang X, Wijesekera S, Sharma D (2009) Quantum cryptography for wireless network
communications. In: IEEE 4th international symposium on wireless pervasive computing, pp
01–05
40. Chang RS, Gao J, Gruhn V, He J, Roussos G, Tsai WT (2013) Mobile cloud computing
research—issues, challenges and needs. In: IEEE seventh international symposium on service-
oriented system engineering, pp 442–453
41. Sosinsky B (2011) Cloud computing bible, 1st edn. Wiley India Pvt. Ltd., India, pp 6–42
42. Djellab R, Benmohammed M (2012) Securing encryption key distribution in WLAN via QKD.
In: IEEE international conference on cyber-enabled distributed computing and knowledge
discover, pp 160–165
Author Index
A Bilgundi, Srishail K., 1003

Aarthi, R., 199 Biswas, Saroj Kr., 653, 673, 1039
Adhikary, Tathagata, 11 Borah, Amlan Deep, 919
Afroj, Md. Asif, 143 Borah, Janmoni, 731
Agrawal, Rajeev, 521 Borah, Kunal, 329
Agrawal, Sunil, 465 Borah, Samarjeet, 1135
Agrawal, Vanita, 839 Borah, Shekhar Suman, 285
Ahmad, R. Badlishah, 1141 Bora, Joyatri, 485, 731, 743, 919
Ahmed, Tanvir, 89 Borkakoty, Hsuvas, 387
Al Miraz, Mehedi, 143 Boruah, Arpita Nath, 1039
Angadi, Sachin, 155 Bulo, Yaka, 221
Ansari, Sahil, 831 Buragohain, Arunav, 831
Anwar Hussain, Md., 1017
Arundhati, Priya, 753
C
Chaitan, K., 973
B Chakraborty, Ashis, 253
Babu, Korra Sathya, 177 Chakraborty, Swarnendu K., 697
Baishnab, Krishna Lal, 899, 929 Chandran, Saravanan, 1073, 1083
Bandyopadhyay, Sivaji, 673, 1039 Chanu, Thingbaijam Rajkumari, 21
Banerjee, Abhik, 541, 595, 617 Chatterjee, Saibal, 329
Banik, Santanu, 241 Chhetri, Rupa, 839
Barman, Barnali, 909 Chintesh, I., 575
Barman, Rahul, 653
Baro, Utpal Chandra, 839
Baruah, Bikash, 1051 D
Baruah, Manali, 831 Daimari, Raja Jwbthasa, 853
Baruah, Smriti, 743 Dan, Sanket, 63, 241, 253
Basak, Piyali, 11 Das, Amitesh, 707
Begum, Afruza, 1017 Das, Dolly, 673
Bharti, Manisha, 295, 337 Das, Santos Kumar, 177, 187, 211
Bharti, Sneha, 295, 337 Das, T. D., 965
Bhattacharjee, Utpal, 99, 513 Datta, Brajagopal, 813
Bhattarai, Sagar, 965 Datta, Manishankar, 607
Bhavani, Pilla Devi, 231 Datta, Shrayasi, 885
Bhavani, R., 419 Debela, Tamiru, 863, 875
Bhoi, Priyabrata, 1135 Deka, Aniruddha, 853
in Electrical Engineering 686, https://doi.org/10.1007/978-981-15-7031-5
1152 Author Index
Deka, Bidyut, 1105 Jangam, Rupa Sree, 1123

Deshmukh, Prashant, 211 Jena, Sisir Kumar, 641
Deshmukh, Tanmay, 357 Joshi, Manoj, 133
Dev, Chandana, 387 Jyothsana, A., 307
Devi, Kshetrimayum Linthoinganbi, 731
Devi, Swagata, 899
Dinakar, D., 957 K
Doso, Oying, 265 Kabir, Md. Monirul, 89
Dutta, Aradhana, 1105 Karimulla, Shaik, 347
Dutta, Jiten Chandra, 377, 429 Karthika, N. J., 1083
Dutta, Manash Pratim, 1051 Kashyap, Nitesh, 1115
Dutta, P. K., 565 Kasturi, Bhanu Ravindranath, 1123
Dutta, Subrata, 253 Kaur, Baljit, 295, 337, 367
Kaushik, Rajasve, 585, 607
Keshwani, Gaurav, 377, 429
G Kishore, Putha, 957
Gajula, Rohith Sai Kumar, 1123 Koppala, Neelima, 1123
Ganguly, Amrita, 387 Kri, Rushanti, 983
Gao, Sarsing, 265 Kumari, Pallavi, 407
Gautam, 813 Kumar, Piyush, 485
Gautam, Ritik, 1073 Kumar, Prashant, 585, 607
Ghole, Mukund Subhash, 555 Kumar, Rajesh, 475
Ghosh, Chinmoy, 885 Kumar, Raman, 585, 607
Ghosh, Mourina, 285 Kumar, Rishi Kant, 585, 607
Ghosh, Pritam, 63, 241, 253 Kumar, Varun, 1093
Gogoi, Mriganka, 565 Kumar, Vickey, 367
Goswami, Joyeeta, 77 Kumar, Vinod, 1
Guchhait, Pabitra Kumar, 541, 595 Kunamalla, Sarangam, 395
Guha, Koushik, 899 Kuppachi, Sree Lochan Kanthi Kumar, 1123
Gupta, Lalita, 167
L
H Laskar, Naushad Manzoor, 899
Hajra, Dilip Kumar, 241 Latif, Abdul, 839
Halder, B., 439 Likith Kumar, M. V., 1003
Hampannavar, Santoshkumar, 805
Haripriya, B., 775
Hasibuzzaman, Md., 143 M
Hazarika, Kabyashree, 429 Machavaram, Venkata Raghunadh, 275
Hijam, Somokanta, 21 Mahanta, Pinakeswar, 685
Himabindu, N., 805 Maity, Santanu, 743
Hormese, Jose, 1073 Majumder, Suman, 885
Husainee, Kauser, 357 Malemnganbi, Rajkumari, 37
Hussain, Israfil, 839 Manasani, Swapna, 805
Hussain, Md. Anwar, 919 Mandal, Satyendra Nath, 63, 241, 253, 885
Hussain, Sarfraz, 475 Mandi, Rajashekar P., 805
Marchang, Ningrinla, 627
Maurya, Sudhanshu, 1141
I Maurya, Sudhir, 617
Ieswaria, S., 419 Meghwal, Virendra Kumar, 507
Meher, Preetisudha, 49
Meitei, Moirangthem Goldie, 627
J Meitei, Sorokhaibam Nilakanta, 329
Jamwal, Anbesh, 521 Miah, Md. Aman, 143
Author Index 1153
Mia, Nasim, 89 Prakash, Sai, 495

Mishra, Debahuti, 1135 Preethi, V., 231
Mishra, Madhusudhan, 99, 485 Priyadarsini, Lima, 211
Mishra, Shashidhar, 187 Priya, R., 419
Mishra, Subhadra, 1135 Pudur, Rajen, 407
Mitra, A., 439 Purkayastha, Biswajit, 653
Mittal, Namita, 507
Modha, Hardik P., 1061
Momocha, M. S., 973 R
Mudang, Tagiya, 109 Rajak, Mrinal Kanti, 407
Mukherjee, Kaushal, 707 Raju, A. B., 155
Mukherjee, Kaushik, 63, 241, 253 Raman, Ashish, 1115
Mukherjee, Kuntal, 1141 Raman, Rahul, 317
Murugananthan, V., 453 Ranjan, Ashish, 133
Mustafi, Subhranil, 63, 241, 253 Ranjan, Ravi, 1115
Rashed, Md. Jahidul Islam, 143
Raut, Vaishnavi, 1029
N Ravi, Mancharla, 221
Nandan, Durgesh, 575 Ray, Anjan Kumar, 555
Nandhan, Durgesh, 231 Ray, Kailash Chandra, 939
Naskar, Syamal, 241 Ray, Sangram, 885
Nath, Sourav, 899 Roy, Kunal, 63, 241, 253
Neroula, Sujan, 533
Roy, O. P., 1
Ningthoujam, Sanjit, 697
Roy, Sahadev, 707
Nistala, Bheema Rao, 275, 395, 495, 823
Roy, Subhojit, 241
Nongmeikapam, Kishorjit, 775
Roy, Suvadip, 541
Nyodu, Karter, 663
O S
Obulapathi, L., 957 Saha, Ashim, 77
Okramcha, Maxon, 21 Saha, T., 439
Sahu, Partha Pratim, 1105
Saifullah, Mohammad Khalid, 89
P Sakthivel, G., 453
Padmavathi, Manchineellu, 957 Sakuru, K. L. V., 495
Paily, Roy, 949 Samal, K. Krishna Rani, 177
Panigrahi, Rakhee, 719 Sambyo, Koj, 983
Pani, Santosh Kumar, 753 Santra, Gour Hari, 1135
Panwar, Maj Sunil Kumar, 685 Saring, Yang, 783
Parashar, Shivam, 485 Sarkar, Paramita, 929
Patel, Deep V., 1061 Sarkar, Santanu, 187
Patel, Rajendra V., 1061 Sarkar, Sunita, 653, 673, 1039
Pathak, Uddipta, 831 Sarker, Avijit, 143
Patidar, Akshay, 521 Sasmal, Sajal, 783
Patjoshi, Rajesh Kumar, 719 Satyanarayana, V., 575
Patnaik, Abinash, 719 Sharma, Arpan, 485
Patra, Aswini K., 485 Sharma, Arvind, 965
Patra, Sarat Kumar, 1093 Sharma, Monica, 521
Patra, Swarnajyoti, 909 Sharma, Santanu, 533
Paul, Bikram, 685 Sharma, Suruchi, 295, 337, 367
Phukan, Hillol, 863, 875 Shimray, Benjamin A., 37
Pradeepa, H., 1003 Shivani, 585
Pradhan, Adarsh, 831 Shivu, M., 1003
1154 Author Index
Shopon, Md, 763 T

Shufian, Abu, 143 Thakur, Hiranya Ranjan, 377, 429
Shukla, Swati, 993 Thoudam, Archana, 595
Sindhuja, R., 119 Tilak Babu, S. B. G., 575
Singh, Anil Kumar, 109 Tiwari, Praveen, 685
Singh, Ankur, 285 Tripathy, S. K., 929
Singh, Arvind Kumar, 99, 513 Tripathy, Somanath, 939
Singh, Girdhari, 507 Trivedi, Gaurav, 475, 685, 993
Singh, Gurpartap, 465 Trivedi, Y. N., 27
Singh, Indrasen, 1029 Tumma, Sunil Kumar, 823
Singh, Jiwanjot, 863, 875
Singh, Karam Bharat, 133
Singh, Lukram Dhanachandra, 49
Singh, Mangal, 1093 V
Singh, Poonam, 1093 Valluvan, M., 453
Singh, Renu, 99, 485, 513 Verma, Abhishek, 295, 337
Singh, Vivek, 939 Vijaya, Samudra, 663
Sinha, Khushboo, 27
Sinha, Rohan, 317
Sivakumar, D., 199 Y
Sivaraman, J., 307, 347 Yadav, Menka, 791
Sohi, B. S., 465
Yadav, Vimal Kumar Singh, 949
Soni, Badal, 653
Yaragatti, Udaykumar R., 155
Soni, Jatin M., 1061
Srinivasan, S., 119
Srivastava, Harshit, 187
Suresh, Yellasiri, 155 Z
Swain, Deepak Kumar, 1135 Zimik, Mayanmi, 485

Bok - 978 981 15 7031 5

Uploaded by

Copyright:

Available Formats

Bok - 978 981 15 7031 5

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bok - 978 981 15 7031 5

Uploaded by

Copyright:

Available Formats

Lecture Notes in Electrical Engineering 686

Pradeep Kumar Mallick

More information about this series at http://www.springer.com/series/7818

Alak Majumder Santos Kumar Das

ISSN 1876-1100 ISSN 1876-1119 (electronic)

Prof. Rajshree Srivastava, DIT University, Dehradun, India

It is our pleasure to introduce the proceedings of the First International Conference

Bhubaneswar, India Pradeep Kumar Mallick

• User interfaces and interaction models

• Business information systems

A Reliable and Secure Wireless Network for VoIP Applications . . . . . 1

Optimal Placement of FACTS Devices in the Power System

Modiﬁed Aggressive Packet Combining Scheme with Repetition

Design and Investigation of Tunnel Junction Engineered

Hybrid L + C Edge Detection Algorithm for Harnessing Large

Simulation and Analysis of Hybrid PEMFC and Solar PV Cell Model

ICI Mitigation with Fixed-Number (FN)- and Fixed-Region

Forced Alignment Method for Detection of Bodo

Phrase-Based Machine Translation of Digaru-English . . . . . . . . . . . . . 983

Enhancement of Isolation in MIMO Antenna System . . . . . . . . . . . . . . 1123

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1151

Dr. Pradeep Kumar Mallick is currently working as Associate Professor in the

Dr. Preetisudha Meher is currently working as Assistant Professor in National

Vinod Kumar and O. P. Roy

Abstract In a wireless communication system, the Quality of Service (QoS)

Keywords Wireless network · Voice over internet protocol · Reliability ·

The present day’s society observes a revolutionary change in communication tech-

V. Kumar (B) · O. P. Roy

Table 1 Codec in general

INTRA-SR Domain 2 Domain 3

Fig. 1 Simulation area of (1000 × 1000 m2 )

3.1 Network Design

A network of 1000 × 1000 m2 is created with N number of nodes as shown in Fig. 1.

3.2 Security and Prevention

The problem of routing in VoIP-based communication model is solved by integrating

Fig. 2 Reliable network using GA with (SVM & ANN) approach

Table 2 Node’s properties

In domain 1, during route formation, VoIP message packet is forwarded by node

Malicious Detection of VoIP communication using GA with SVM

where Fs : Current Node Properties and Ft : Threshold properties for all

3 Calculate Length of T for R

4 Results and Analysis

Table 4 Routing overhead analysis

Table 5 Delay analysis

Table 6 Jitter analysis (ms)

Fig. 3 Routing overhead without Attack

Fig. 4 Routing overhead with Attack

Fig. 5 Average delay without attack for VoIP application

Average Delay (ms)

Fig. 6 Average delay with attack for VoIP application

Fig. 7 Average jitter without attack for VoIP application

Fig. 8 Average jitter with attack for VoIP application

Tathagata Adhikary and Piyali Basak

Abstract Computer simulations to address research problems have touched almost

Keywords Bioinformatics · Molecular simulation · Phytochemicals · Bioactivity ·

Pharmacognosy aims to develop databases on natural product-based drugs that

T. Adhikary · P. Basak (B)

extraction/isolation, bioactivity, and toxicity profiling of drugs. Traditional recipes

2 The Diversity of Databases