See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228458328

A fuzzy logic method for autotuning a PID controller: SISO and MIMO systems

Article

CITATIONS READS

4 1,249

3 authors, including:

Antonio A. R. Coelho
Federal University of Santa Catarina
109 PUBLICATIONS   369 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

sliding mode control based on long range predictive control View project

Model Free Adaptive Controlers View project

All content following this page was uploaded by Antonio A. R. Coelho on 02 June 2014.

The user has requested enhancement of the downloaded file.

Copyright © 2002 IFAC
15th Triennial World Congress, Barcelona, Spain

A FUZZY LOGIC METHOD FOR AUTOTUNING A PID CONTROLLER:

SISO AND MIMO SYSTEMS

Otacílio da M. Almeida1 and Antonio A. R. Coelho2

1
Federal University of Ceará - Department of Electrical Engeneering
2
Federal University of Santa Catarina - Department of Automation and Systems
Box 476 - 88040.900 - Florianópolis, SC, Brazil
e-mail: {otacilio, aarc}@lcmi.ufsc.br)

Abstract: In this paper a new method for autotuning SISO and MIMO PID fuzzy logic
controllers (FLC) is proposed. The fuzzy autotune procedure adjusts on-line the
parameters of a conventional PID controller located in the forward loop of the process.
Fuzzy rules are based on the representation of human expertise on how can be the
behaviour of gain and phase margins of a control system to efficiently compensating the
system errors. Performance and robust stability aspects are assessed by practical and
simulated examples.

Keywords: fuzzy control, PID control, multivariable feedback control, nonlinear systems,
stability.

1. INTRODUCTION
In this paper a method for autotuning SISO and MIMO
For years the fuzzy logic control has proved its broad FLC is proposed. The autotune procedure adjusts on-
potential in industrial applications (Altrock and line the parameters of a conventional PID controller
Gebhardt, 1996; Qin, et al., 1998). The fuzzy control located in the forward loop of the process.
theory has been applied to a number of systems with In order to give an autotuning capacity to SISO and
single-input and single-output (SISO) structures, MIMO cases, a scheme of identification and sequential
mainly to overcome uncertain parameters and multivariable identification are implemented by using
unknown models (Hu et al, 1999). Generally, fuzzy relay feedback (Wang and Shao, 1999; Luyben, 1990;
control shows good performance for controlling Shen and Yu, 1994; Shiu and Hwang, 1998). Since the
nonlinear and uncertain systems that could not be transfer function in each step of the sequential design
controlled satisfactorily by using conventional has a mix of underdamping and overdamping behavior,
controller, for example, a conventional PID controller a second-order plus dead-time structure is adopted as
(Ying, et al., 1990). Also, in applications when there plant model. Also, fuzzy rules, employed to determine
are multiple-input and multiple-output (MIMO) the set of PID gains, are based on the representation of
systems with strong loop interactions, conventional human expertise on how must be the behavior of gain
controllers do not work well and advanced control and phase margins of a control system to efficiently
conceptions are required. Control literature of MIMO compensating the system errors. In both SISO and
fuzzy logic controllers (MIMO FLC) shows limited MIMO cases, gain and phase margins are determined
results and a great effort has been used by researchers by a set of Mandani rules and the membership function
to derive stable control strategies. Usually, a MIMO of the fuzzy sets are based on the system error and its
FLC is tuned by trial-and-error that means a tedious difference.
and time-consuming task, and design techniques for
systematic tuning must be obtained. Also, MIMO FLC Performance and robust stability aspects are assessed
applications are frequently solved by using the by practical and simulated examples of SISO and
conventional decoupling theory and with single FLCs, MIMO systems. Simulation results for other
resulting in high-dimensional rule-bases that may not conventional control algorithms are also included for
be implemented in practical systems, due to required comparison purpose. The proposed control scheme
processing time (Nie, 1997). offers advantages over the conventional fuzzy
controller such as: i) a systematic design is attained in 2π 1
t − jω t
∫ u y r (t )e u dt
both SISO and MIMO cases; ii) it is necessary only ωu = ; G p ( jω u ) = − = − 0t − jω t
one rule base for all loops; iii) the tuning mechanism is Tu Ku ∫0u u r (t )e u dt
simple and control operators can easily understand how (3)
it works; and iv) it is completely autotuned, requiring
only one relay feedback experiment per loop. where, yr(t) and ur(t) are process and relay output,
respectively. Ku and ωu are process critical gain and
frequency, respectively.
2. AUTOTUNING OF FUZZY PID CONTROLLER: Literature shows many methods for tuning PID
SISO CASE - FPID-SISO control. In this paper, the initial PID parameters are
determined by considering the transfer function of the
Since the proposed controller uses a nonlinear form (Wang and Shao, 1999)
fuzzification algorithm and output membership
functions, the controller can be considered as a  As 2 + Bs + C 
nonlinear PID where parameters are tuned on-line Gc ( s ) = k  
 (4)
 s 
based on error e(t) and change of error de(t) about a
setpoint r(t), as shown in Fig. 1. The system error is where A=Kd/k, B=Kc/k, C=Ki/k and (Kc, Ki, Kd) are the
compensated by a set of fuzzy linguistic rules which PID gains.
are derived from the experience and knowledge of a
control designer on how can be the behaviour of gain Zeros of the controller are chosen to cancel the poles
and phase margins for efficiently compensating the of the process model, A=a, B=b and C=c. Then, the
system error. In this sense the FPID-SISO can be following relationship holds
interpreted as a fuzzy gain scheduling PID controller.
ke − sL
G p ( s )Gc ( s ) = (5)
s
where k is obtained by considering the gain (Am) and
phase margin (Φm ) and the gain crossover frequency
(ωg) and phase crossover frequency (ωp). So, the
following relation can be derived

π 1 
Φm = 1 −  (6)

2 Am 
Fig. 1. Fuzzy logic controller system. and PID parameters are

In order to implement the FPID-SISO, it is necessary

Kc  b 
to identify a process model and to design a  K  = π c 
conventional PID control as a starting point. After  i  2A L   (7)
m
that, the fuzzy engine is designed.  K d  a 

A typical value for Am, taking into accounting a

2.1 Tuning phase: identification and initial PID conventional PID controller, is Am=3, so that Φm=60o.
parameters

An approximated model for the process is considered. 2.2 Autotuning fuzzy logic controller engine
A second-order transfer function with time-delay is
usually enough for practical systems and is given by The gain margin Am and the phase margin Φm, Eq.(6)
and (7), are considered linguistic variables which
e − sL values are defined with respect to the same universe
G p (s) = (1) of discourse specified by human expertise about the
as 2 + bs + c
operational knowledge of the process. It is assumed
that the feedback system gain and phase margins are
where a, b, c and Ld are unknown parameters and they
in prescribed ranges [Am,min, , Am,max] and [Φm,min ,
need to be determined by a feedback relay
experiment. Depending on a, b and c, the model may Φm,max], respectively. For convenience, values of Am
have real or complex poles and it is representing both are normalized into a range between zero and one by
monotonic and oscillatory processes. Under relay the following linear transformation
experiments, the parameters a, b and c are given by
the following equations Am' = ( Am − Am, min ) ( Am, max − Am, min ) (8)

cos(ω L ) Φ m' = (Φ m − Φ m,min ) (Φ m,max − Φ m,min ) (9)

1 sin(ω u L) c+ Ku
u
c= ; b= ; a= (2)
Kp ωu Ku ω u2
where Am' and Φ m' are normalized gain and phase
margins, respectively.
Values of Am' are determined by a set of fuzzy rules
of the form

If e(t) is Ai and de(t) is Bi then Am' ,i is Ci; i=1...n

(10) Fig. 3. Implication procedure of a fuzzy rule.

where Am,i is the gain margin for i rule, Ai, Bi and Ci By using the membership functions in Fig.2c, the
are fuzzy sets on the corresponding supporting sets. following condition holds
The membership functions of these fuzzy sets for e(t) n
and de(t) are shown in Fig.2c. Fuzzy rule base sets are ∑ µi = 1 (15)
obtained from operator’s expertise by using the step i =1

response of the process. Figs. 2a and 2b show an

example of a desired time response and a fuzzy rules Then, the defuzzification process, Fig. 3, yields
base.
n
Am' = ∑ µ i Am' ,i (16)
i =1

Once Am' is obtained, Am is calculated from Eq. (8) and

PID parameters are derived from Eq. (7).

2.3 Simulation and experimental results for the FPID-

SISO

(b) The new fuzzy PID controller is now assessed for its
(a) ability to control nonlinear and time-varying plants,
and to evaluate its performance in comparison with
the corresponding PID control tuned without the
fuzzy part of the algorithm. Four experiments are
shown: three simulations and one practical system.
The first system to be tested is a second-order plus
time-delay with the following transfer function
(c)
Fig. 2. (a) Plant response and fuzzy rule base sets; (b)
1
rules base; (c) membership functions. G p (s) = e −2s (17)
(2.5s + 1)(3.75s + 1)
The fuzzy set Ci may be either Big or Small and it is
characterised by logarithmic membership functions. The Since it is known that the PID controller can deal with
'
grade of the membership µ and the variable A has the
m
low-order linear systems with short dead-time and
time-invariant, an error of 50% with variance of 0.01
following relation
is applied to the time-delay estimation for assessing
the performance of the controllers. As shown in Fig.
µ B ( Am' ) = − η1 ln(1 − Am' ) (11) 4, FPID performs better than PID controller.
2.5
−( δ ' |)
µ M ( Am' ) = 1 − e |0.5 − Am
(12)

µ S ( Am' ) = − η1 ln( Am' ) (13)

where, η and δ are adjustable parameters, in Fig. 2c η=4

and δ=0.1.
Fig. 4. Responses for (a) PID (left) ; (b) FPID (right).
The truth value of the ith rule in Eq. (10) µi is obtained by
the product of the membership function values in the Now, considering an error of 50% with variance of 0.01
antecedent part of the rule (Nie, 1997). in the time-delay of the system and a model rupture of
50% in the parameter b, the PID controller was not able
to provide stability while the FPID still tracking the
µ i = µ Ai [e(t )] ⋅ µ Bi [∆e(t )] (14)
setpoint (there is a excessive overshoot in the setpoint
response). Closed-loop results are shown in Fig. 5.
Based on µi, values of Am' for each rule are determined
from their correspondent membership function. The
implication procedure is shown in Fig. 3.
Fig. 5. Responses for (a) PID (left) ; (b) FPID (right).

Next, a nonlinear process given by Eq. (17) is simulated

by using trapezoidal integration. The relay feedback
experiment is used to tune the PID parameters around an
operational point. Output, control and reference signals
of the process when controlled by PID and FPID Fig. 8. Tunnel heating responses for PID and FPID
controllers are shown in Fig. 6. It is evident that the PID controllers.
presents poor performance while a well damped setpoint
response is achieved by FPID controller. According to the set of locus of Fig. 9, Gp(jω) does
not enclose –1/N(A, ω) within [ω1, ω2] so, by the circle
y (t ) = − y (t ) + sin 2 ( y (t ) ) + u (t ) (18) stability criterion (Kin et al., 2000), the system is
asymptotically stable.

Fig. 6. Responses for (a) PID (left) ; (b) FPID (right).

Fig. 9. Circle stability criterion, Gp(jω) – heating tunnel
In order to conclude monovariable performance tests,
PID and FPID control approaches are assessed in a Nyquist diagram; N(A,ω) – describing function
heating tunnel process implemented in the Department of FPID controller; ω1= 0.0305 rd/s; ω2= 0.0528
of Automation and Systems at the University of Santa rd/s.
Catarina. Details of the process are available in
http://www.lcmi.ufsc.br/lcp/. The heating tunnel control
system, shown in Fig. 7, is composed by a fan, a DC 3. AUTOTUNING OF FUZZY PID CONTROLLER:
motor, a 50 cm long air duct with uniform transverse MIMO CASE - FPID-MIMO
area, having on its right extremity an electrical heating
With additional conditions, the proposed FPID-SISO
resistance. The electrical heating part is driven by a
controller, can be generalized for the MIMO case. As in
power actuator circuit whose input is compatible with a
the SISO case, the FPID-MIMO can be divided into two
D/A card. The temperature is measured by two sensors
stages: identification and controller design phases. The
placed on the duct extremity. The hot air inside the duct
identification phase copes with the auto-tuning capacity
is spread by the fan and the control problem is to
of the controller and to do so, a scheme of sequential
regulate the temperature inside the duct (controlled
multivariable identification is implemented. In the
variable) by actuating on the current through the
controller design phase, the main ideas applied to the
electrical resistance (manipulated variable).
FPID-SISO case is extended to the MIMO case. The
reduced dimension of the rule base as well its simplicity
correspond to the most important features of the
proposed fuzzy scheme. In order to asses the FPID-
MIMO performance the Wood-Berry distillation column
(Luyben, 1990) is evaluated. Simulations have shown
that the controller is capable of providing good overall
system performance.

Fig. 7. Experiment Heating tunnel plant.

3.1 Identification phase
Step responses of FPID–SISO and PID controllers for
different setpoints are given in Fig. 8. In this paper a sequential identification is
implemented by utilizing a sequential relay scheme
Although both FPID and PID controllers give good that treat the MIMO system as a series of SISO
control, FPID results is superior with overshoot, rise systems as shown in Fig. 10 for a 2x2 system
time and control variance minimum characteristics. (Luyben, 1990; Shen and Yu, 1994; Shiu and Hwang,
1998).
MIMO fuzzy rules are the representation of human
expertise on how must be the behavior of the gain and
phase margins of a MIMO control system to
efficiently compensating the system errors.

If ei(k) is Av and ∆ei(k) is Bv then Am,j is Cv (23)

v=1...p and i=1...n
where Am,i, is gain margin for i loop, Av, Bv and Cv are
(a) (b) fuzzy sets on the corresponding supporting sets, p is
Fig. 10. Identification under sequential relay. the number of fuzzy sets and n is the number of loops
in the multivariable system. The membership function
In this kind of identification the transfer function gii,CL of the fuzzy sets ei(k) and dei(k) as well the rule base
(i=1,2) generally, has underdamped poles. So, the are generalization of the SISO case. MIMO PID
second-order plus dead-time model given by parameters are determined to ensure adequate gain
and phase margins to the system.
e − sLi Considering that sequential design is addressed, the
G p ,i ( s ) = , i = 1,..., n (19)
a i s 2 + bi s + ci MIMO PID parameters are tuned as in a SISO case.
Considering a two-input/two-output system, as
example of a MIMO transfer function, PID MIMO
where ai, bi, ci and Li are unknown parameters to be
controller can be written as
determined, is used. Depending on ai, bi and ci, the
model may have real or complex poles. Hence, it is
suitable for representing both monotonic and  A s 2 + Bi s + C i 
G c ,i ( s ) = k i  i  i = 1, 2 (24)
oscillatory processes like that in the Eq. (19).  s 
 
The model parameters are determined from sequential
where Ai, Bi, Ci and ki are determined by the same
relay experiment (Shen and Yu, 1994) and given by
SISO case strategy. MIMO PID parameters for a
TITO system is given by
c i = 1 / k p ,i (20)

− jω u ,i Li  K c ,i  bi 
ci − real [e G p ,i ( jω u ,i )]   π c 
ai = (21)  K i ,i  = 2 LA  i i = 1, 2 (25)
ω u2,i  K d ,i  m ,i
  a i 
− jω u ,i Li
imag[e G p ,i ( jω u ,i )]
bi = (22) where the parameters Am,i is a nonlinear variable
ω u ,i determined by the fuzzy engine.
Design steps for the auto-tuning MIMO fuzzy
where the index i is the loop number and ωui is the controller are: i) tune a Ziegler-Nichols PID
crossover frequency controller considering individuals loops; ii) if the
control system is unstable, one step of a sequential
design should be done to tune a PID controller for the
3.2 Autotuning multivariable fuzzy logic controller MIMO process; iii) identify the MIMO system by
engine performing a relay experiment in each loop while the
other loops are under PID controllers designed in step
The multivariable fuzzy logic controller proposed in one. A transfer function like that of Eq. (19) is
this paper utilizes fuzzy rules to determine the set of obtained for each loop; iv) define the discourse
PID parameters. As in the SISO case, control signals universe to the fuzzy variables; v) specify the
in the MIMO case are generated by PID controllers, maximum and minimum values to A’ m. Typical values
Fig. 11. for A’ m ranges from 2 to 5 and is corresponding to
phase margin between 30 to 45; vi) apply the MIMO
fuzzy controller engine.

3.3 Simulation results for the FPID-MIMO

In this section the distillation column of Wood and

Barry, for separating methanol and water is used for
simulation purposes. The Wood and Berry (WB)
column presents strong loop interactions and the
transfer function model is given by
Fig. 11. Fuzzy gain scheduling PID controller: MIMO
case.
 12.8e − s − 18.9e −3s   R( s ) In the SISO case the gain is considered a fuzzy variable
 X t ( s)  16.7 s + 1  and in the MIMO case the sequential design is addressed
21s + 1   
 = (26) in order to be possible the fuzziness of the gain and
 X b ( s)  6.6e
−7 s
− 19.4e −3s   
phase margins. The fuzzy PID controller derived
10.9s + 1 
 S ( s ) 
14.4 + 1  successfully demonstrated better performance than the
conventional PID controller for many case studies,
Manipulated variables are reflux, R(s), and steam flow, particularly for nonlinear plants. The fuzzy PID
S(s). Controlled variables are distillate and bottom controller is also able to tolerate many poor selections or
compositions, Xt(s) and Xb(s), respectively. The inadequate implementation of the controller gains, for
interactor matrix is diagonal with a strong couple example, bad tuning for initial parameters of PID
between the loops. According to step one of the controller, which would make most conventional
proposed algorithm, Ziegler-Nichols PID (ZN-PID) controller unstable.
(Luyben, 1990) parameters must be used as a pre-tuning
and are given in Table 1. Parameters of the identified The main goal of this paper is to provide to the plant
transfer function under relay experiments are shown in operators with easy-to-understand fuzzy PID method
Table 2. for quickly achieving satisfactory control over
unknown monovariable and multivariable systems.
Table 1 – ZN pre-tune parameters. Despite its simplicity, the proposed method yielded
Parameters monovariable and multivariable designs and a
Loops Kc Ki Kd superior behavior to that resulting from empirical
1 0.9450 0.2898 0.5545 method based on trial-and-error procedure. The
2 -0.1960 -0.2177 -0.3175 method has shown be adequated for practical
applications, both in SISO and MIMO cases.
Table 2 – Loop parameters for each transfer function.
Parameters REFERENCES
Loops a b c L
1 0.0231 0.9487 0.1565 1.01 Altrock, C and Gebhardt, J. (1996). Recent successful
2 0.6222 0.4609 0.1032 2.98 fuzzy logic applications in industrial automation,
Proc. of Fifth IEEE International Conference on
Figs. 12 and 13 compare output, setpoint and control Fuzzy Systems, vol. 3, pp. 1845-1851.
signals for WB column under ZN and FPID-MIMO Hu, B.; Mann, G. K. I. and Gosine, R. G. (1999). New
control methods. Simulation results show that the methodology for analytical and optimal design of
FPID-MIMO is better than ZN method. For setpoint fuzzy PID controllers, IEEE Transaction on Fuzzy
responses, the overshoot, the level of interactions and Systems vol. 7, pp. 521-539.
the settling time are improved with more stable Kim, E.; Lee, H. and Park, M. (2000) Limit-cycle
responses. prediction of a fuzzy control system based on
describing function method, IEEE Transaction
on Fuzzy Systems, vol. 8, pp. 11-22.
Qin, Y.; Jia, L. -M. and Zhang, X. -D. (1998).
Automation of combustion process for coke oven
using multivariable fuzzy control technique, IEEE
World Congress on Computational Intelligence,
Proc. of FUZZY-IEEE, Anchorage, Alaska, USA,
(a) (b) pp. 606-610.
Fig. 12. Control, output and setpoint for WB column Luyben, W. L. (1990). Process modeling simulation,
(a) pair R-Xt (b) pair S-Xb. and control for chemical engineers, McGraw-Hill.
Nie, J. (1997). Fuzzy control of multivariable nonlinear
servomechanisms with explicit decoupling
scheme, IEEE Transactions on Fuzzy Systems, vol.
5, pp. 304-311.
Shen, S. –H. and Yu, C. –C. (1994). Use of relay-
feedback test for automatic tuning of multivariable
systems, AIChE Journal, vol. 40, pp. 627-646.
Shiu, S.-J. and Hwang, S., -H. (1998). Sequential design
(a) (b)
method for multivariable decoupling and multiloop
Fig. 13. FPID-MIMO parameters.
PID controllers, Ind. Eng. Chem. Process Des.
Dev., vol. 37, pp. 107-119.
4. CONCLUSION Ying, H.; Siler, W. and Buckley, J. J. (1990). Fuzzy
control theory : a nonlinear case, Automatica, vol.
A systematic method has been developed to design a 23, pp. 513-520.
fuzzy PID controller for SISO and MIMO cases. The Wang, Y. G. and Shao, H. H. (1999). PID tuning for
method is based on gain and phase margin specifications improving performance, IEEE Transactions on
and needs system identification under relay experiments. Control Systems Technology, vol. 7, pp. 457-465.

View publication stats