Mechanical Systems and Signal Processing 68-69 (2016) 125–137

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Model-based approach for elevator performance estimation

E. Esteban a, O. Salgado a,n, A. Iturrospe b, I. Isasa c
a
Mechanical Engineering, IK4-Ikerlan, J.M. Arizmendiarrieta 2, 20500 Mondragon, Spain
b
Department of Electronics and Computer Sciences, Mondragon Unibertsitatea, Loramendi 4, 20500 Mondragon, Spain
c
Mechanical Engineering, Orona EIC S. Coop., Pol. Lastaola s/n, 20120 Hernani, Spain

a r t i c l e i n f o abstract

Article history: In this paper, a dynamic model for an elevator installation is presented in the state space
Received 7 July 2014 domain. The model comprises both the mechanical and the electrical subsystems,
Received in revised form including the electrical machine and a closed-loop field oriented control. The proposed
29 June 2015
model is employed for monitoring the condition of the elevator installation. The adopted
Accepted 8 July 2015
Available online 30 July 2015
model-based approach for monitoring employs the Kalman filter as an observer. A Kalman
observer estimates the elevator car acceleration, which determines the elevator ride
Keywords: quality, based solely on the machine control signature and the encoder signal. Finally, five
Elevator elevator key performance indicators are calculated based on the estimated car accelera-
Dynamic modeling
tion. The proposed procedure is experimentally evaluated, by comparing the key
State space models
performance indicators calculated based on the estimated car acceleration and the values
Kalman filters
State estimation obtained from actual acceleration measurements in a test bench. Finally, the proposed
procedure is compared with the sliding mode observer.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction

Maintenance and modernization services account for more than a 46% of the total revenue in some elevator leading
companies [1]. Maintenance and modernization services are expected to continue to grow, although with significant
variation between countries and a pricing market environment characterized by very intense competition.
Actually, elevator maintenance services involve periodical in situ inspections in order to evaluate the elevator
performance and its ride quality. In order to reduce variability in the results of elevator ride quality measurements, the
standard ISO 18738-1:2012 [2] encourages industry-wide uniformity in the definition, measurement, processing and
expression of vibration and noise signals defining several key performance indicators [3]. These key performance indicators
are now part of specifications for most elevator installations. Most of the performance indicators established by the standard
ISO 18738-1:2012 are calculated by processing the elevator car acceleration signal in the vertical direction.
Currently, periodical in situ inspections are generally conducted employing commercial equipment designed specifically
to evaluate the performance of an elevator installation. The use of these commercial equipment involves installing
additional sensors in order to measure the required magnitudes. The EVA-625 [4] and the Lift PC [5] system are widely
employed portable systems that evaluate the elevator installation according to the ISO 18738-1:2012. Both evaluation
systems require accelerometers placed on the elevator car floor. However, permanently installing sensors involve additional
costs, and installed sensors are also prone to malfunction. Therefore, it is desirable to employ already existing signals from

n
Corresponding author.
E-mail address: [email protected] (O. Salgado).

http://dx.doi.org/10.1016/j.ymssp.2015.07.005
0888-3270/& 2015 Elsevier Ltd. All rights reserved.
126 E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137

the regulator in order to evaluate the elevator performance. Using regulator signals also opens up the way to provide remote
monitoring services for continuously evaluating the elevator condition.
The Kone Corp. has developed the portable ESiteSurvey™ system [6], for evaluating the condition of elevator
installations. The functionality of the ESiteSurvey™ system goes beyond a mere implementation of the ISO 18738-1:2012
procedures. The ESiteSurvey™ provides a comprehensive set of elevator system parameters, readily after round-trip test
run. The ESiteSurvey™ acquires the electrical power consumed by the machine by means of current probes fixed to the
terminals of the electrical machine. It also measures the acceleration of the elevator car by means of accelerometers. Using
the acquired signals it estimates different parameters, such as car and counterweight masses, guide friction loses, as well as
machine efficiency. These estimations are based on a power balance model approach. The parameters of interest are
estimated by minimizing the deviation between the expected power needed to move the elevator by the model and the
measured consumed power.
Instead of a power balance model, more accurate models accounting coupled electrical dynamics, could explain better
the whole elevator system dynamics, allowing a deeper understanding of the coupling between the mechanical and the
electrical subsystems.
Condition monitoring comprises different monitoring techniques and their selection depend, among other factors, on the
prior knowledge about the monitored system. An excellent survey of condition monitoring techniques can be found in [7]
where model-based and data driven approaches are compared. Comparing the model-based and data-driven modeling
approach, model-based approaches rely on the availability of a theoretical model of the monitored system, which can be
derived using physical modeling principles [8]. The performance of a model-based method depends on a large extent of its
accuracy when describing the system dynamics. Furthermore, a model-based approach should require a sufficiently generic
model to address variations in the elevator configuration. However, model-based approaches have been successfully
employed for both electromechanical systems monitoring and control applications [9,10]. Regarding to data-driven
modeling approaches [8], monitoring systems have to be trained, so enough training data is needed, which is not the
case for our application [11,12], as for other industrial applications [13].
Two of the most employed model based observer algorithms in industry are the Kalman filter (KF) and the sliding mode
observer (SMO) [14–18]. Comparing the KF algorithm and the SMO algorithm, the SMO is simpler to implement but under
noisy measurement conditions the KF algorithm performs better than the SMO algorithm [18–20]. Therefore, the approach
proposed in this paper is based on a state space model and employs the KF as an observer; but it is also compared to
the SMO.
Our approach, as shown schematically in Fig. 1, is based on a dynamic model for a 1:1 elevator installation, comprising
both the mechanical and the electrical subsystems. Then, the Kalman filter (KF) algorithm is applied as an observer,
estimating the elevator car acceleration. The input signal used for the KF algorithm is the regulator signature and the output
signal is the machine encoder. Once the car acceleration is estimated, the five key performance indicators (KPI) are
calculated as described by the ISO 18738–1:2012 in order to estimate the ride quality performance.
The rest the paper is organized as follows. In Section 2, both the dynamics of a 1:1 elevator installation and its model are
described. In Section 3, the Kalman filter is applied for observing elevator car acceleration and five ride quality performance
indicators are calculated. In Section 4 the proposed method is experimentally validated and compared and finally several
conclusions are discussed in the last section.

2. Elevator model

An elevator installation comprehends both mechanical and electrical components as it is shown in Fig. 2. The elevator car
carries the passengers or loads upwards and downwards. The mass of the elevator car is balanced by a counterweight in
order to reduce the torque demanded to the machine. An electrical machine drives through a pulley onto the suspension
ropes which interconnect the elevator car and the counterweight. Both the car and the counterweight move vertically
constrained by a pair of rails each.
The installation modeled in this paper is driven by a permanent magnet synchronous machine (PMSM) [22] which is
controlled using a field oriented control (FOC) regulator [23] and the velocity signature profile is dynamically calculated,
depending on the starting car position and its final destination. The velocity loop obtains its feedback from the motor
encoder.
The mechanical force imbalance between elevator car and counterweight exerts a mechanical torque in the rotor shaft
that is actively balanced with the electromagnetic torque exerted by the machine. The mechanical subsystem and the

Fig. 1. Application of the model-based approach for the estimation of elevator performance indicator. Adapted from [21].
E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137 127

electrical dynamics are coupled by the torque balance equation in the shaft (see Fig. 2) as follows:

I p θ€ m ¼ τe τm ð1Þ

where Ip is the inertia of the driving pulley. The angular acceleration of the rotor shaft is denoted by θ€ m , the electromagnetic
torque exerted by the machine is denoted by τe and τm is the mechanical torque.

2.1. Mechanical subsystem dynamics

The 1:1 rope suspension configuration [11] dynamics, as shown in Fig. 3, are described by a lumped parameter model,
with seven degrees of freedom (DOF) as detailed in Table 1. The DOF represent the absolute position of the inertial elements.
A global frame of reference is located on the ground of the elevator installation for all non-rotary DOFs.

Fig. 2. Simplified schematic description of the 1:1 elevator installation.

Fig. 3. Lumped parameter model of the mechanical subsystem.

128 E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137

Table 1
Description of each DOF.

Angular rotation (rad) Nomenclature

Rotor shaft (or pulley shaft) θm

Vertical position (m) Nomenclature

Machine zm
Rotor shaft zp
Car side rope zrc
Counterweight side rope zrw
Car zc
Counterweight zw

Both the mass and the stiffness of each side of the rope depend on the instantaneous length of the car side rope, lc, and
the counterweight side rope length, lw, as follows [24–28]:
2EA 2EA
mrc ¼ μlc ; mrw ¼ μlw ; kc ¼ ; kw ¼ ð2Þ
lc lw
where E, A and μ are the rope Young modulus, cross-sectional area and the linear mass density respectively.
Based on the force balance in each inertial element, the mechanical subsystem dynamics are stated as follows [29]:

mm z€ m þcm z_ m þkm zm kp zp zm ¼ km ρm kp ρp mm g ð3Þ

mp z€ p þ kp zp zm kc zrc zp kw zrw zp ¼ mp g þ kp ρp þ 2EA ð4Þ

mrc z€ rc þ m
_ rc z_ rc þkc 2zrc zc zp ¼ mrc g ð5Þ

mrw z€ rw þ m
_ rw z_ rw þ kw 2zrw zp zw ¼ mrw g ð6Þ

mc z€ c þkc ðzc zrc Þ ¼ mc g EA ð7Þ

mw z€ w þ kw ðzw zrw Þ ¼ mw g EA ð8Þ

where the lumped masses, stiffness and damping elements of each j-inertial elements are denoted by mj, kj and cj,
respectively and rp denotes the driving pulley radius.
The mechanical torque τm exerted in the rotor shaft from Eq. (1) is then obtained as

τm ¼ r p kw zrw zp rp kc zrc zp ð9Þ
Eqs. (1)–(9) can be rewritten in a matrix form as follows:

g
Mz€ þ Cz_ þKz ¼ F þ fðτe Þ ð10Þ
1
T
z ¼ zm zp zrc zrw zc zw θm ð11Þ
2 3
mm km ρm kp ρp
6 m kp ρp þ 2EA 7
6 p 7
6 7
6 mrc 0 7
6 7
6 7
F ¼ 6 mrw 0 7 ð12Þ
6 7
6 mc EA 7
6 7
6 7
4 mw EA 5
0 0
where M is the mass matrix, C is the damping matrix, K is the stiffness matrix and F is the generalized force matrix. The
position vector of each inertial element is denoted by z and fðτe Þ is a vector that includes the torque exerted by the machine
which is defined in Section 2.4.

2.2. PMSM dynamics

The electrical dynamics of a PMSM can be described by the following equations in the dq0 reference frame [30]:
d
V d ¼ Rid þLd i pθ_ m Lq iq ð13Þ
dt d
E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137 129

d
V q ¼ Riq þ Lq iq þ pθ_ m Ld id þ λ ð14Þ
dt

3
τe ¼ p λiq þ Ld Lq id iq ð15Þ
2
where V, i and L denote the voltages, current and inductance, and the subscripts d and q denote the direct and the
quadrature axis respectively. The resistance of the stator is denoted by R, λ is the magnetic flux linkage and p is the number
of machine pole pairs.
The dynamic model of the PMSM described by Eqs. (13) and (14) is nonlinear due to the product of the currents and the
rotor and the rotor angular velocity. However, Eqs. (13)–(15) can be linearized assuming the magnetic symmetry of the
machine and the same inductances ðLd Lq Þ as described in [31,32]. Hence, the electromagnetic torque is assumed to be
proportional to the direct axis current by a machine torque constant Kt
" # " # " # " # " #
Vd id Ld 0 d id 0
ffiR þ þ ð16Þ
Vq iq 0 Lq dt iq λpθ_ m

3
τe ffi pλiq ¼ K t iq ð17Þ
2

2.3. Integration of the field oriented control in the elevator model

The aim of a FOC strategy is to control the direct and quadrature axis current independently, maximizing the active
power and minimizing the reactive power [33].
As shown in Fig. 4, in our elevator installation the rotor angular velocity is compared to its reference value ωref and
controlled by means of a PI control. In the inner-loop, two PI current controllers are implemented in order to regulate the id
and iq currents. For an ideal FOC control, idref is zero in order to minimize the reactive power consumed by the machine.
The FOC dynamics are described by the following three equations:
Z t
iqref ¼ K ps es þ K is es dt ð18Þ
0

Z t
V d ¼ K pd ed þ K id ed dt ð19Þ
0

Z t
V q ¼ K pq eq þ K iq eq dt ð20Þ
0

The tracking errors are defined as

es ¼ ωref θ_ m ð21Þ

ed ¼ idref id ð22Þ

eq ¼ iqref iq ð23Þ

where the proportional gain and the integral gain of a PI control are denoted with Kp and Ki respectively. The subscript s, d
and q represent the velocity, direct current and quadrature current controller respectively.
The FOC dynamics can be represented as first order differential equations [34] as follows:
2 3 2 3
iqref K ps e_ s þK is es
d6 7 6 _ 7
4 V d 5 ¼ 4 K pd e d þK id ed 5 ð24Þ
dt
Vq _
K pq e q þK iq eq

Fig. 4. Field oriented control scheme.

130 E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137

2.4. State space representation of the elevator model

In this subsection, the electrical equations and the mechanical equations are combined in a matrix form and the whole
system dynamics are represented in the state space domain.
Electrical equations in a matrix form: Eqs. (11), (16), and (21)–(24) are combined and rewritten in a matrix form as follows:
2 3
ω_ ref
6 7
6 ωref 7
6 7
Le_ þRe þ Wz_ þ Nz€ ¼ U6 _ 7 ð25Þ
6 i dref 7
4 5
idref

where the electrical state vector e is expressed as

h iT
e ¼ id iq iqref V d V q ð26Þ

and the matrices are detailed as follows:

2 3
Ld 0 0 0 0
6 0 Lq 0 0 07
6 7
6 7
L¼6
6
0 0 1 0 07
7 ð27Þ
6K 0 0 1 07
4 pd 5
0 K pq K pq 0 1

2 3
R 0 0 1 0
6 0 R 0 0 1 7
6 7
6 7
R¼6
6
0 0 0 0 0 7
7 ð28Þ
6K 0 0 0 0 7
4 id 5
0 K iq K iq 0 0

2 3
0 0 0 0 0 0 0
60 λp 7
6 0 0 0 0 0 7
6 7
W¼6
60 0 0 0 0 0 K is 7
7 ð29Þ
6 7
40 0 0 0 0 0 0 5
0 0 0 0 0 0 0

2 3
0 0 0 0 0 0 0
60 0 7
6 0 0 0 0 0 7
6 7
N¼6
60 0 0 0 0 0 K ps 7
7 ð30Þ
6 7
40 0 0 0 0 0 0 5
0 0 0 0 0 0 0

2 3
0 0 0 0
6 0 0 7
6 0 0 7
6 7
U¼6
6 K ps K is 0 0 7 7 ð31Þ
6 7
4 0 0 K pd K id 5
0 0 0 0

The L and R matrices describe both the PMSM and FOC dynamics. The W matrix couples the PMSM dynamics and the
FOC velocity control loop with the rotor angular velocity. The N matrix couples the velocity control loop with the rotor
angular acceleration. The U matrix relates the electrical dynamics with the regulator inputs.
Mechanical equations in a matrix form: Eqs. (10), (17), and (26) are combined and rewritten in a matrix form as follows:

g
Mz€ þ Cz_ þKz þTe ¼ F ð32Þ
1

where the matrix T represents the electromagnetic torque exerted by the machine into the mechanical subsystem and is
E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137 131

detailed as follows:
2 3
0 0 0 0 0
60 0 0 0 07
6 7
6 7
60 0 0 0 07
6 7
6 07
T¼60 0 0 0 7 ð33Þ
6 7
60 0 0 0 07
6 7
60 07
4 0 0 0 5
0 Kt 0 0 0

By combining both the electrical equation (25) and the mechanical equation (32), the elevator model is represented in
the state space domain [35] as follows:
_ ¼ AxðtÞ þ BuðtÞ
xðtÞ ð34Þ
where the state matrix A and the input matrix B are defined as
2 3 12 3
I7 O O O I7 O
6 7 6 7
A ¼ 4 O M O 5 4 K C T5 ð35Þ
O N L O W R
2 3 12 3
I7 O O O O
6 7 6 7
B¼4O M O5 4O F5 ð36Þ
O N L U O
The identity matrix is denoted with I7 and O is the null matrix with its corresponding size. The state vector x and its time
derivative are
h iT h iT
x ¼ zT z_ T eT ; x_ ¼ z_ T z€ e_
T T
ð37Þ

and the input vector u is

h iT
u¼ ω _ ref ωref i_dref idref g 1 ð38Þ

The elevator model above fits into the framework of Linear Time Varying (LTV) systems because the elevator rope
stiffness changes as a function of its instantaneous length.
Regarding the elevator model above, the mechanical subsystem model only considers the vertical dynamics. Most of the
studies of the literature also focus on the vertical dynamics as a first approximation [36]. Although it is true that the lateral
car-rail forces are coupled with the vertical dynamics, the literature shows that it has little effect on it [27,28]. This
unmodeled lateral dynamics have been considered as additive Gaussian process noise in the KF [37].

3. Kalman observer and the estimation of the ride quality' KPIs

We apply the Kalman observer in order to estimate the elevator car acceleration z€ c by measuring solely the velocity
signature ωref and the encoder angular position θm. The application of the Kalman observer for the state estimation is
explained in Algorithm 1,1[38].

Algorithm 1. Kalman observer.

Tuning of the algorithm: ð39Þ
x^ kjk ¼ x0 ; Pkjk ¼ P0
ð40Þ
Q k ¼ E½ðwk w k Þðwk w k ÞT ¼ Q 0 ; r k ¼ E½ðvk v k Þðvk v k ÞT ¼ r 0
for k ¼ 0; 1; 2; …; N 1
Prediction step:
Compute the predicted state mean: ð41Þ
x^ kjk þ 1 ¼ Ak x^ kjk þ Bk uk
Compute the predicted covariance: ð42Þ
Pkjk þ 1 ¼ Ak Pkjk ATk þ Q k

1
Since in our application there is solely one observation measurement, the correction step of the Kalman observer is written for one measurement.
The general form of the Kalman observer is obtained by replacing Eqs. (43)–(45) as ek ¼ yk ðCk x^ kjk þ 1 Þ; Kk ¼ Pkjk þ 1 CTk ðCk Pkjk þ 1 CTk þ Rk Þ 1 and
x^ k þ 1jk þ 1 ¼ x^ k þ 1jk þ Kk ek respectively.
132 E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137

Correction step:

Compute the innovation:

ð43Þ
ek ¼ yk ðCk x^ kjk þ 1 Þ
Compute the Kalman gain:
ð44Þ
Kk ¼ Pkjk þ 1 CTk ðCk Pkjk þ 1 CTk þ r k Þ 1
Update the predicted state mean:
ð45Þ
x^ k þ 1jk þ 1 ¼ x^ k þ 1jk þ Kk ek
Update the predicted covariance estimate:
ð46Þ
Pk þ 1jk þ 1 ¼ ðI Kk Ck ÞPkjk þ 1

end for

The physical interpretation of the tuning of the Kalman observer is shown in Fig. 5. Each state of the state vector xk is assumed
as a mean value during the elevator ride whereas each variance of the noise vector wk is related with the uncertainty of the
unmodeled dynamics (e.g. lateral car-rail forces) of each state. The Kalman observer assumes the variance as a constant value
during the elevator ride, which is particularly an adequate assumption for the estimation of both LTI and LTV systems.
According to that, we employ the model of the elevator system from Eq. (34) by adding the process noise vector as follows:
_ ¼ AxðtÞ þBuðtÞ þ wðtÞ
xðtÞ ð47Þ
where wðtÞ is the process noise vector which is assumed to be a zero mean Gaussian noise with covariance matrix Q
wðtÞ Nð0; Q Þ ð48Þ
We define the output y(t) as the shaft angular position corrupted by an added measurement noise v(t) as follows:
yðtÞ ¼ CxðtÞ þvðtÞ ð49Þ
where the measurement noise is assumed to be a zero mean Gaussian noise with variance r
vðtÞ Nð0; rÞ ð50Þ
and the output matrix C is detailed as follows:
C ¼ ½0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0: ð51Þ
Finally, the car acceleration estimation is extracted from the 12th row of Eq. (47) as
kc ðx^ 5 x^ 3 Þ EA
x_^ 12 ¼ z€^ c ¼ g þ w12 : ð52Þ
mc

Fig. 5. Physical interpretation of tuning of the Kalman observer for one state.
E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137 133

The estimated car acceleration is then processed according to the ISO 18738:1-2012 standard in order to calculate the
maximum velocity and the maximum jerk reached during the ride, as well as analyzing the peak-to-peak amplitude of the
car vibration in the vertical direction which is of particular relevance to passenger comfort [2]. As shown in Fig. 6, the
peak-to-peak vibration amplitude is calculated by adding the absolute values of two peaks between a zero crossing point.
The peak-to-peak vibration amplitude analysis is based on determining its maximum value, as well as obtaining the A95
indicator which is defined as the 95 percentile of the peak-to-peak vibration amplitude during the ride [2].
The proposed elevator car acceleration estimation procedure allows us to monitor continuously the ride quality for each
ride without installing additional sensor in the elevator car.

4. Experimental results

The proposed procedure has been experimentally tested in an elevator test bench. The elevator test bench emulates a 1:1
rope elevator installation. Two rectangular frames act as the car and the counterweight respectively, as it is shown in Fig. 7.
The elevator parameter values (from Eq. (34)) are detailed in Appendix A.
During a test ride, the angular position of the rotor, θm, and the car acceleration, z€ c , have been simultaneously measured
with a sampling frequency of 20 kHz. The car acceleration has been measured employing a DC accelerometer (B&K model
4575), while the angular position has been measured directly from the SinCos absolute encoder (Heidenhain encoder).
Signals have been sampled by a National Instruments NI-9239 module, fixed to a National Instruments cDAQ-9178
acquisition board.
The velocity signature from the regulator ωref and its time derivative during the test ride are shown in Fig. 8. This velocity
signature can be divided into five zones. The first zone starts in (1) representing the initial floor. Then, from (1) to (2) the

Fig. 6. Illustration of peak-to-peak calculations of the car acceleration in the z-axis.

Fig. 7. 1:1 rope configuration test bench.

134 E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137

Fig. 8. Velocity signature during the test ride.

elevator car accelerates and from (2) to (3) the constant velocity regime is maintained. From (3) to (4) the velocity decreases
and then from (4) to (5) a constant landing velocity is maintained until the brake stops the car.
The estimation results of the elevator car acceleration obtained with Kalman filter are compared to those obtained by the
nonlinear sliding mode observer.
The Kalman filter is tuned by setting the process error covariance matrix P0 ¼ I19 , the process noise covariance matrix
Q 0 ¼ 0:01 I19 and the measurement noise variance r0 ¼0.01. The initial values for the state estimates are statically obtained
(at t¼0) by considering the input vector as u ¼ ½0 0 0 0 g 1T and considering that the elevator car is still, that is x_ 0 ¼ 0. The
estimation of the car acceleration was carried out using Matlab and ReBEL2 toolkit [39].
Likewise, the sliding mode observer is tuned by designing the observer gain as G0 ¼ ½0 0 0 0 0:06 0 0:4 0 0 0 0
0:06 0 0:4 0 0 0 0 0T , using the procedure described in [17] and the same initial values for the state estimates are
employed as for the Kalman filter.
The estimated car acceleration, z€^ c , employing both the Kalman filter and the sliding mode observer algorithm is
compared with the measured car acceleration in Fig. 9. Fig. 9a shows the car acceleration estimation obtained by the Kalman
filter algorithm whereas Fig. 9b shows the car acceleration estimation obtained by the sliding mode observer and compared
to the measured acceleration.
As observer in Fig. 9a, the Kalman filter estimate and the measured acceleration differ mostly when the elevator car is
decelerating. The reason of this deviation can be attributed to the unmodeled rail friction which increases as the car velocity
reduces. Despite this deviation, the standard deviation during the whole ride is less than 0.09 (m/s2).
For the acceleration estimate obtained by the sliding mode observer (see Fig. 9b), the standard deviation is larger than
that for the Kalman filter estimate, more than 0.12 (m/s2) accounting the whole ride. Unlike to the case of Kalman filter
estimation, the larger deviation for the sliding mode observer estimate takes place when the elevator car is accelerating. As
the sliding mode observer gain is fixed for the whole ride, it seems that the algorithm is unable to estimate accurately the
acceleration as it varies during the ride.
Finally, the five ride quality KPIs are calculated based both on the Kalman filter and the sliding mode observer
acceleration estimates, as well as employing the measured acceleration, and the deviations from the KPI obtained based on
the measured acceleration. These KPI values are compared in Table 2. As expected, the deviation between the acceleration
measurement and the Kalman filter estimate, for the estimated KPI, is small (less than 4%). The deviation between the
acceleration measurement and the sliding mode observer estimate, for the estimated KPI, is larger (less than 7%).
As expected, although the sliding mode observer algorithm is simpler to implement, the Kalman filter provides the better
estimates under noisy measurements [19].
Both observers, as they are model based, are not limited to applications with limited number of measurements for
obtaining the empirical model as in the case of data driven techniques. Furthermore, generating data based models that
explain the whole elevator system based only the encoder signal and the control signature seems a priori a difficult task. On
the contrary, model-based approaches rely on the availability of a theoretical model of the elevator system, which is
proposed in this paper using physical modeling principles.

2
ReBEL is a Matlab toolkit for Sequential Bayesian inference ordered from http://www.ohsu.edu/tech-transfer/portal/technology.php?
technology_id=45875.
E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137 135

Fig. 9. Comparison of the car acceleration estimation obtained by Kalman filter and sliding mode observers with the direct car acceleration measurement.
(a) Car acceleration estimation employing KF. (b) Car acceleration estimation employing SMO.

Table 2
Result of the estimated KPIs employing both Kalman filter and sliding mode observer.

KPI name Measured KF estimate KF dev. in % SMO estimate SMO dev. in %

Max. vel. (m/s) 0.592 0.591 0.2 0.556 6

Max. jerk (m/s3) 1.243 1.201 3.4 1.280 3
A95 (m/s2) 0.616 0.590 4 0.575 6.5
A95pk–pk (m/s2) 0.345 0.350 1.5 0.365 6
Max.pk–pk (m/s2) 0.861 0.844 2 0.840 2.4

5. Conclusions

A state space model for a elevator, including the dynamics of a permanent magnet synchronous machine controlled by a field
oriented control regulator, has been proposed. According to the results presented, the model is valid for estimating the ride quality
key performance indicators, with the required accuracy in an industrial application, by employing a Kalman observer. The car
acceleration is estimated by Kalman observer using only the encoder signal. Moreover, the model brings us the possibility to
understand better the behavior of the elevator system and to optimize the elevator installation at a design stage.
The experimental tests show that both the car acceleration and the five ride quality indicators can be estimated correctly
by employing the Kalman filter for this application.
In the future, the extended Kalman filter algorithm would be employed for identifying system parameters; both mechanical and
electrical. The extended Kalman filter algorithm also allows us to estimate both the system state variables and to identify system
parameters simultaneously by augmenting the state vector with the parameters of interest in a joint estimation approach, which is
an extension of the state estimation approach used in this paper. If the lateral dynamics and the car-rail force model are included in
the system model, the acceleration estimation might be improved and the friction forces could be estimated.
Using only the regulator signals as inputs of the observer facilitates the implementation of remote monitoring systems,
which could continuously evaluate the elevator condition. This would also open the door to the implementation of adaptive
control strategies as active vibration control in the future.

Acknowledgments

This study is partially funded by the Basque Government under the Emaitek 2014 Program and by the Ministry of
Economy and Competitiveness of the Spanish Government under the Retos-Colaboración Program (LEMA project, RTC-2014-
1768-4) and the AIRHEM III project. The authors also gratefully acknowledge Orona EIC S. Coop. for supporting this research
line. Any opinions, findings and conclusions expressed in this paper are those of the authors and do not necessarily reflect
the views of funding agencies.

Appendix A

The details of the model parameter values used in the test bench are summarized in Table A1.
136 E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137

Table A1
Model parameter values.

Parameter Nomenclature Value

Motor position in z axis ρm 9m

Rotor position in in z axis ρp 0.125 m
Pulley radius rp 5.75 cm
Motor mass mm 225 kg
Pulley mass mp 15 kg
Car mass mc 150 kg
Counterweight mass mw 140 kg
Pulley and rotor inertia Ip 0.34 kg m2
Elastic force EA 106 N
Motor stiffness km 1.05 107 N/m
Shaft stiffness kp 2 105 N/m
Motor isolation damping cm 0 Ns/m
Gravity g 9.81 m/s2
Proportional gain of velocity loop Kps 0.83 A/rad/s
Integral gain of velocity loop Kis 59.5 A/rad
Proportional gain of current loop Kpd and Kpq 6:1 Ω
Integral gain of current loop Kid and Kiq 404 Ω=s
Motor resistance R 1:212 Ω
Inductance in direct d axis Ld 18.175 mH
Inductance in quadrature q axis Lq 18.175 mH
Motor torque constant Kt 15.936 Wb

Appendix B. Supplementary data

Supplementary data associated with this paper can be found in the online version at http://dx.doi.org/10.1016/j.ymssp.
2015.07.005.

References

[1] Kone Q4 Financial Statement Bulletin 2013, Technical Report, Kone, 2014.
[2] ISO 18738-1:2012 Measurement of Lift Ride Quality—Part 1: Lifts (Elevators), 2012.
[3] G.P. Lorsbach, Analysis of elevator ride quality, vibration, Elevator World 51 (6) (2003) 108.
[4] EVA-625 Elevator Vibration Analysis System, Technical Report, PMT Europe, 2010.
[5] P. Pini, B. Eng, Modules for safe and sure operation, monitoring and documentation in elevators, Lift Rep. 2 (2005) 30–40.
[6] T. Tyni, R. Kontturi, P. Perälä, Electric site survey—on quest of elevator parameters, in: The International Congress on Vertical Transportation
Technologies, Elevcon, no. 19, 2012.
[7] R.J. Patton, F.J. Uppal, C.J. Lopez-Toribio, Soft computing approaches to fault diagnosis for dynamic systems: a survey, in: Fourth IFAC Symposium on
Fault Detection Supervision and Safety for Technical Processes, 2000, pp. 198–211.
[8] R. Isermann, M. Munchhof, Identification of Dynamic Systems: An Introduction with Applications, Springer-Verlag, 2011.
[9] D. Simon, Optimal State Estimation: Kalman, H infinity, and Nonlinear Approaches, John Wiley & Sons, Hoboken, New Jersey, 2006.
[10] G. Mastinu, M. Plöchl, Road and Off-Road Vehicle System Dynamics Handbook, CRC Press, New York, 2014.
[11] L. Janovsky, Elevator Mechanical Design, third ed. Elevator World, Inc., Mobile, AL, 1999.
[12] J.P. Andrew, S. Kaczmarczyk, Systems Engineering of Elevators, Elevator World, 2011.
[13] J. Korbicz, J.M. Koscienly, Z. Kowalczuk, W. Cholewa, Fault Diagnosis: Models Artificial Intelligence, Applications, Springer-Verlag, Berlin, 2004.
[14] V.M. Moreno, A. Pigazo, Kalman Filter: Recent Advances and Applications, InTech, Croatia, 2009.
[15] C. Lascu, I. Boldea, F. Blaabjerg, Comparative study of adaptive and inherently sensorless observers for variable-speed induction-motor drives, IEEE
Trans. Ind. Electron. 53 (February (1)) (2005) 57–65.
[16] SAE-China, FISITA, International Federation of Automobile Engineers', Technicians' Associations, et al., in: Proceedings of the FISITA 2012 World
Automotive Congress: Volume 12: Intelligent Transport System (ITS) & Internet of Vehicles, vol. 12. Springer Science & Business Media, New York,
2012.
[17] H. Alwi, C. Edwards, C.P. Tan, Fault Detection and Fault–tolerant Control Using Sliding Modes, Springer Science & Business Media, London, 2011.
[18] F. Zhang, G. Liu, L. Fang, A battery state of charge estimation method using Sliding mode observer, in: Seventh World Congress on Intelligent Control
and Automation, 2008, WCICA 2008, June 2008, pp. 989–994.
[19] Y. Zhang, Z. Zhao, T. Lu, L. Yuan, Wei Xu, J. Zhu, A comparative study of Luenberger observer, sliding mode observer and Extended Kalman filter for
sensorless vector control of induction motor drives, in: Energy Conversion Congress and Exposition, 2009. ECCE 2009, IEEE, 2009, San Jose, CA,
pp. 2466–2473.
[20] M. Ongkosutjahjo, V.M. Becerra, Integrating the Utkin observer with the Unscented Kalman filter, in: Proceedings of the 17th World Congress The
International Federation of Automatic Control Seoul, Korea, 2008.
[21] R. Isermann, Model-based fault-detection and diagnosis—status and applications, Annu. Rev. Control 29 (1) (2005) 71–85.
[22] A. De Almeida, C. Patrao, J. Fong, R. Araujo, U. Nunes, Project E4: E4 Energy Efficient Elevators and Escalators, Technical Report, ISR-University of
Coimbra (Portugal), ELA (Europe), FhG-SIS (Germany), KAPE (Poland), 2010.
[23] B. Drury, Control Techniques Drives and Controls Handbook, no. 35, The Institution of Engineering and Technology (IET), London, 2001.
[24] N. Noguchi, A. Arakawa, T. Yoshimura, Y. Sekiya, Identification method of elevator rail deformation using operational acceleration data, Jpn. Soc. Mech.
Eng. 78 (786) (2012) 431–445.
[25] R. Roberts, Control of high-rise high-speed elevators, in: Proceedings of the American Control Conference, IEEE, vol. 6, Philadelphia, PA, 1998,
pp. 3440–3444.
[26] A. Arakawa, K. Miyata, A variable-structure control method for the suppression of elevator-cage vibration, in: Proceedings of the Institute of Electrical
and Electronics Engineers Conference—28th IEEE, vol. 3, 2002, pp. 1830–1835.
E. Esteban et al. / Mechanical Systems and Signal Processing 68-69 (2016) 125–137 137

[27] Y. Zhou, Models for an elevator hoistway vertical dynamic system, in: International Congress on Sound and Vibration, no. 5, Adelaida, Australia, 1997,
pp. 2673–2680.
[28] J. Vladic, R. Djokic, M. Kljajin, M. Karakavsic, Modelling and simulations of elevator dynamic behaviour, Tech. Gaz. 18 (3) (2011) 423–434.
[29] J. Angeles, Dynamic Response of Linear Mechanical Systems: Modeling, Analysis and Simulation, Springer-Verlag, New York, 2011.
[30] R.H. Park, Two-reaction theory of synchronous machines generalized method of analysis—part I, Trans. Am. Inst. Electr. Eng. 48 (3) (1929) 716–727.
[31] L.M. Grzesiak, T. Tarczewski, S. Mandra, Permanent magnet synchronous servo-drive with state position controller, in: Proceedings of Power
Electronics and Motion Control Conference—13th EPE-PEMC, 2008, pp. 1071–1076.
[32] W.J. Xu, Permanent magnet synchronous motor with linear quadratic speed controller, Energy Proc. 14 (0) (2012) 364–369. 2011 2nd International
Conference on Advances in Energy Engineering (ICAEE).
[33] N.P. Quang, J.A. Dittrich, Vector Control of Three-Phase AC Machines: System development in the practice, Springer, Verlag, Berlin, 2008.
[34] K. Venu, Ch. Rushikesh, V. Rajasekhar, Design and analysis of DC motor with PID controller—a state space approach, Trans. Electr. Electron. Eng. (ITSI-
TEEE) 1 (2013) 11–14.
[35] K. Ogata, Modern Control Engineering, Prentice-Hall, Upper Saddle River, New Jersey, 2003.
[36] I. Isasa, Model validation applied to locally nonlinear lift structures (Ph.D. thesis), Mondragon Unibertsitatea, 2010.
[37] J.K. Kang, S.K. Sul, Vertical-vibration control of elevator using estimated car acceleration feedback compensation, Trans. Ind. Electron. IEEE 47 (1)
(2000) 91–99.
[38] J.V. Candy, Bayesian Signal Processing: Classical, Modern and Particle Filtering Methods, John Wiley & Sons, Hoboken, New Jersey, 2009.
[39] R. Van Der Merwe, Sigma-point Kalman filters for probabilistic inference in dynamic state-space models (Ph.D. thesis), Oregon Health & Science
University, 2004.