An experimental validation of a novel clustering approach

to PWARX identication
Zeineb Lassoued, Kamel Abderrahim
Numerical Control of Industrial Processes, National School of Engineers of Gabes, University of Gabes, St Omar Ibn-Khattab, 6029 Gabes, Tunisia
a r t i c l e i n f o
Article history:
Received 14 March 2013
Received in revised form
2 September 2013
Accepted 16 October 2013
Available online 2 December 2013
Keywords:
Hybrid systems
PWARX models
Clustering
Identication
Chiu's clustering technique
Experimental validation
a b s t r a c t
In this paper, the problem of clustering based procedure for the identication of PieceWise Auto-
Regressive eXogenous (PWARX) models is addressed. This problem involves both the estimation of the
parameters of the afne sub-models and the hyperplanes dening the partitions of the state-input
regression. In fact, we propose the use of the Chiu's clustering algorithm in order to overcome the main
drawbacks of the existing methods which are the poor initialization and the presence of outliers.
In addition, our approach is able to generate automatically the number of sub-models. Simulation results
are presented to illustrate the performance of the proposed method. An application of the developed
approach to an olive oil esterication reactor is also suggested in order to validate simulation results.
& 2013 Elsevier Ltd. All rights reserved.
1. Introduction
Hybrid systems are dynamical systems that involve the interac-
tion of continuous and discrete dynamics. The continuous behavior is
the fact of the natural evolution of the physical process whereas the
discrete behavior can be due to the presence of switches, operating
phases, transitions, computer program codes, etc.
Hybrid systems allow to describe numerous physical processes
such as chemical process control where the continuous evolution
can be changed by the actuators states (valves and pumps),
embedded systems where the discrete part represented by the
program interacts with the physical part represented by the
system, industrial production lines, telecommunication networks,
air trafc management systems, etc. Moreover, the notion of
hybrid system can be used to represent complex nonlinear con-
tinuous systems. In fact, we can exploit the divide and conquer
strategy which consists in decomposing the domain range of the
nonlinear system into a set of operating regions. For each opera-
tion region, a linear or afne model is associated. So, the con-
sidered complex nonlinear system becomes by modeling as an
hybrid system switching between linear or afne sub-models.
The analysis and control of hybrid system require that the
system's dynamic behavior must be dened by a mathematical
model. This model can be developed using the physical laws that
govern the functioning of the system. But, this approach fails when
the physical laws describing the behavior of the system are not
sufciently well known or the system involves not easily measur-
able parameters. In addition, it can lead to very complicated
models that cause problems of exploitation and implementation.
However, for engineering, a mathematical model must provide a
compromise between accuracy and simplicity of implementation.
A solution to this problem consists in using the identication
approach which allows building a mathematical model from
inputoutput data. For this reason, only the problem of identica-
tion of hybrid systems represented by Piecewise AutoRegressive
systems with eXogenous input (PWARX) is considered in this
paper. This problem has received a great attention in the last years,
since the PWARX models can approximate any nonlinear system
with arbitrary accuracy (Lin and Unbehauen, 1992). Moreover, the
properties of equivalence (Heemels et al., 2001) between PWARX
models and other representations of hybrid systems allow to
transfer the results of PWA models to other classes of hybrid
systems, such as Jump Linear models (JL models) (Feng et al.,
1992), Markov Jump Linear models (MJL models) (Doucet et al.,
2001), Mixed Logic Dynamical models (MLD models) (Bemporad
and Morari, 1999; Bemporad et al., 2000), Max-Min-Plus-Scaling
systems (MMPS models) (De Schutter and Van den Boom, 2000),
Linear Complementarity models (LC models) (van der Schaft and
Schumacher, 1998), Extended Linear Complementarity models
(ELC models) (DeSchutter, 1999).
The PWARX systems are obtained by decomposing the state-
input domain into a nite number of nonoverlapping convex
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/engappai
Engineering Applications of Articial Intelligence
0952-1976/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.engappai.2013.10.007
E-mail addresses: zeineb.lassoued1@gmail.com (Z. Lassoued),
kamel.abderrahim@yahoo.fr (K. Abderrahim).
Engineering Applications of Articial Intelligence 28 (2014) 201209
polyhedral regions, and by associating a simple linear or afne model
to each region. Consequently, the PWARX identication problem
involves both the estimation of the parameters of the afne sub-
models and the hyperplanes dening the partitions of the state-input
regression. This is one of the most difcult problems that represents
an area of research where considerable work has been done in
the last decade (Bako et al., 2012; Ferrari-Trecate et al., 2003). The
proposed methods in the literature for the identication of PWARX
systems can be classied in numerous categories of solutions such as
the algebraic solution (Tian et al., 2009), the clustering-based
solution (Ferrari-Trecate et al., 2003), the Bayesian solution (Juloski
et al., 2005), the bounded-error solution (Bemporad et al., 2005), the
sparse optimization solution (Bako, 2011; Bako and Lecoeuche, 2013)
etc. Only the clustering-based procedure is considered in this paper.
This approach is based on three main steps: data classication,
parameter estimation and region reconstruction. It is easy to remark
that the classication of data represents the corner stone of this
approach toward the objective of PWARX system identication
because a successful identication of models' parameters and hyper-
plans depends on the correct data classication. The early approaches
use classical clustering algorithms for the data classication such as
k-means algorithms (Ferrari-Trecate et al., 2003; Ferrari-Trecate and
Muselli, 2003; Nakada et al., 2005; Boukharouba, 2011). These
algorithms may converge to local minima in the case of poor
initializations because they are based on the minimization of non-
linear criterion. Furthermore, their performances degrade in the case
of the presence of outliers in the data to be classied. In addition,
most of them assume that the number of sub-models is a
priori known.
In this paper, we improve the performance of the existing
methods based on clustering procedure thanks to the use of Chiu's
algorithm for data classication (Chiu, 1994). This algorithm allows
us to reduce the effect of outliers. Moreover, it does not need any
initialization. In addition, it is able to automatically generate the
number of sub-models. The performance of this method is
illustrated by a simulation example. An experimental validation
with an olive oil esterication reactor is also considered in order to
show the efciency of the developed approach.
This paper is organized as follows. Section 2 presents the model
and its assumptions. Section 3 recalls the clustering-based procedure
for the identication of PWARX systems. Moreover, it presents the
main drawbacks of the existing methods for the identication of
PWARX based on clustering procedure. In Section 4, we propose a
solution to overcome the main problems of the existing methods.
This solution consists in using the Chiu's algorithm for data classi-
cation. The performance of the proposed approach is evaluated and
compared through simulation results in Section 5. In Section 6, an
application of the developed approach to olive oil esterication
reactor is presented.
2. Model and assumption
In this paper, we address the problem of identifying PieceWise
Auto-Regressive eXogenous (PWARX) systems described by
y(k) =f ((k)) e(k); (1)
where y(k)AR is the system output, e(k) is the noise, k is the now
time index, (k) is the vector of regressors which belongs to a
bounded polyhedron H in R
d
:
(k) = [y(k1); ; y(kn
a
)u(k1); ; u(kn
b
)]
T
; (2)
where u(k) AR
nu
is the system inputs (n
u
is the length of u), n
a
and
n
b
are the system orders and d =n
a
n
u
(n
b
1). f is a piecewise
afne function dened by
f () =

T
1
if AH
1

T
s
if AH
s
_

_
(3)
where = [
T
1]
T
, s is the number of sub-models, {H
i
]
s
i = 1
are
polyhedral partitions of the bounded domain H and
i
AR
d1
is the
parameter vector.
The following assumptions are assumed to be veried:
A1 The number of sub-models s is unknown.
A2 The orders n
a
and n
b
are known.
A3 The noise e(k) is Gaussian sequence independent and identi-
cally distributed with zero mean and nite variance s
2
.
A4 The regions {H
i
]
s
i = 1
are the polyhedral partitions of a bounded
domain H R
d
such as

s
i = 1
H
i
=H
H
i
H
j
=; \i aj
_

_
(4)
Problem statement: Identify the number of sub-models s,
the partitions {H
i
]
s
i = 1
and the parameter vectors {
i
]
s
i = 1
of the
PWARX model using a data set {y(k); (k)]
N
k = 1
where N is the
samples' number.
3. Identication of PWARX models based on clustering
approach
In this section, we recall the main steps of the clustering-based
approach for the identication of PWARX models. These steps can
be summarized as follows: constructing small data set from the
initial data set, estimating a parameter vector for each small data
set, classifying the parameter vectors in s clusters and identifying
the value of s at the same time, classifying the initial data set and
estimating the s sub-models with their partitions.
3.1. Data classication
For every pair of data {(k); y(k)]
N
k = 1
, we construct a local set C
k
collecting {(k); y(k)] and its (n

1) nearest neighbors satisfying:

\(

; y

)AC
k
; J(k)

J
2
rJ(k) ^ J
2
; \( ^ ; ^ y)= cC
k
: (5)
The parameter n

is chosen randomly as n

4d1. This parameter

decisively inuences on the performance of the algorithm.
The optimal value of n

is always a compromise between two

phenomena: accuracy and complexity. The bigger this parameter
is, the better the estimation of the parameters is.
We also dene
k
as follows:
k
= t
1
k
; ; t
n
k
_ _
. It contains, in
ascending order, the indexes of the elements belonging to C
k
.
Among the obtained local sets C
k
, some may contain only data
from the same model. They are called pure local sets. Others can
collect data from multiple sub-models that are called mixed sets.
For each local set C
k
, we identify an afne model using least
square method to determine the local parameters
k
:

k
= (
T
k

k
)
1

T
k
Y
k
; (6)
where

k
= [(t
1
k
)(t
n
k
)]
T
;
and
Y
k
= [y(t
1
k
)y(t
n
k
)]
T
:
Z. Lassoued, K. Abderrahim / Engineering Applications of Articial Intelligence 28 (2014) 201209 202
The obtained parameter vectors {
k
]
N
k = 1
are then classied into
s disjoint clusters using classical clustering algorithms derived
from the k-means approach.
These algorithms are interesting from a theoretical point of
view because they show the possibility of clustering the obtained
parameter vectors. Moreover, they are characterized by the sim-
plicity of implementation. However, these algorithms poorly per-
form in several cases because they do not remove effectively the
outliers. In addition, they can be trapped in local minima in the
case of poor initializations.
3.2. Parameters estimation
As the obtained data are classied, it is possible to determine
the s ARX sub-models dened by the parameter vectors

i
; i =1; ; s using again the least square method. More precisely,
the ith sub-model is obtained from the data of cluster D
i
. Then,
the points in each nal cluster can be used for estimating the
parameter vector of each sub-model according to the rule:
((k); y(k)) AD
i
3
k
AD
i
: (7)
3.3. Regions estimation
The nal step consists in determining the regions H
i
. The
methods of statistical learning such as the Support Vector
Machines (SVM) offer an interesting solution to accomplish this
task (Wang, 2005; Duda et al., 2001). The SVM are a popular
learning method for classication, regression and other learning
tasks. This study is still an ongoing research issue (Hsu and Lin,
2002; Weston and Watkins, 1999).
In our case, it is a matter of nding for every i aj the hyper-
plane that separates points existing in H
i
and in H
j
. Given two sets
H
i
and H
j
, i aj, the linear separation problem is to nd wAR
d
and
bAR such that:
w
T

k
b40 \
k
AH
i
;
w
T

k
bo0 \
k
AH
j
: (8)
This problem can be easily rewritten as a feasibility problem
with linear inequality constraints. The estimated hyperplane
separating H
i
from H
j
is denoted with M
i;j
=m
i;j
where M
i;j
and
m
i;j
are matrices of suitable dimensions. Moreover, we assume that
the points in H
i
belong to the half-space M
i;j
rm
i;j
.
The regions {H
i
]
s
i = 1
are obtained by solving these linear
inequalities (Ferrari-Trecate et al., 2003):
[M

i;1
M

i;s
M]r[m

i;1
m

i;s
m]; (9)
where Mxrm are the linear inequalities describing the bounded
domain H.
4. Chiu's clustering technique
The classication is an important problem to achieve the
objective of PWARX model identication because successful iden-
tication of both sub-models and partitions depends on the
performance of the used clustering approach. In fact, this problem
presents an area of research in which few results have been
devoted in the past (Bemporad et al., 2000; Ferrari-Trecate and
Muselli, 2003; Ferrari-Trecate et al., 2003). Most of the existing
methods for the identication of PWARX models are based on
classical clustering algorithms such as k-means methods (Ferrari-
Trecate et al., 2003). These methods are based on the minimization
of nonlinear criteria. Consequently, they may converge to local
minima in the case of poor initialization. Moreover, these methods
are not able to effectively remove the outliers which degrade their
performance. Also, most of these algorithms assume that the
number of sub-models is a priori known. To overcome these
problems,we propose, in this section, an alternative solution to
these methods based on the Chui's clustering algorithm.
4.1. Principle
Chiu's classication method consists in computing a potential value
for each point of the data set based on its distances to the other data
points and consider each data point as a potential cluster center. The
point having the highest potential value is chosen as the rst cluster
center. The key idea in this method is that once the rst cluster center
is chosen, the potential of all the other points is reduced according to
their distance from the cluster center. All the points which are close to
the rst cluster center will have greatly reduced potentials. The next
cluster center take then the highest remaining potential value. The
procedure for determining a newcenter and updating other potentials
is executed until a predened condition is reached. This condition
depends on the minimum value of the potentials or the required
number of clusters which are reached. This clustering method is based
on three steps: potential computing, outlier processing and clusters
number and centers computing.
Potential computing: We consider the local parameters obtained
by applying the least square method to the grouping obtained by
associating to each its (n

1) nearest neighbors. These N local

parameters (
i
; i =1; ; N) picked out by applying (6) are the
objective of our proposed classication technique. Thus, we
compute a potential value for each local parameter
i
using the
following expression:
P
i
=
N
j = 1
e
(4=r
2
a
) J
i

j
J
2
: (10)
The potential of each local parameter is a function of the distance
from this parameter to all the other local parameters. Thus, a local
parameter with many neighboring local parameters will have the
highest potential value. The constant r
a
is the radius dening the
neighborhood which can be determined by the following expres-
sion:
r
a
=

N

N
i = 1
1
n

n
j = 1
J
i

j
J; (11)
where can be chosen as follows 0oo1.
Outlier processing: From the set of local parameters (
i
; i =
1; ; N), there are some parameters obtained from mixed local
sets, it is clear that we have interest in eliminating them. Eq. (10)
can be exploited to eliminate the misclassied points. As this
equation attribute to the outliers a low potential, we can x a
threshold under which the local parameters are not accepted and
then removed from the data set. This threshold is described by the
following equation:
=min(P)(max(P)min(P)); (12)
where P is the vector containing the potentials P
i
such that
P =[P
1
; ; P
N
] and is a parameter chosen as 0oo1.
Cluster number and centers computing: After this treatment, the
set of local parameters is ltered and reduced to (
i
; i =1; ; N)
(N oN). Then, from this new data set, we select the data point
with the highest potential value as the rst cluster center.
Let
n
1
be this rst center and P
n
1
be its potential. The other
potentials P
j
, (j =1; ; N) are then updated using this expression:
P
i
= P
i
P
n
1
e
(4=r
2
b
) J
i

n
1
J
2
: (13)
Expression (13) allows to associate lower potentials to the local
parameters close to the rst center. Consequently, this choice
guaranties that these parameters are not selected as cluster
centers in the next step. The parameter r
b
is a positive constant
Z. Lassoued, K. Abderrahim / Engineering Applications of Articial Intelligence 28 (2014) 201209 203
that must be chosen larger than r
a
to avoid obtaining cluster
centers which are too close to each other. The constant r
b
is
computed using this formula:
r
b
=

N

N
i = 1
max
j = 1:n
( J
i

j
J): (14)
In general after obtaining the kth cluster center, the potential of
every local parameter is updated by the following formula:
P
i
= P
i
P
n
k
e
(4=r
2
b
) J
i

n
k
J
2
; (15)
where P
n
k
and
n
k
are respectively the potential and the center of
the kth local parameter.
The number of sub-models s is a parameter that we would like
to determine. Therefore, we have developed some criteria for
accepting or rejecting the cluster centers as it is explained in the
algorithm of the next section.
To search the elements belonging to each cluster, we compute
the distance between the estimated output and the real one and
classify (k) within the cluster which has the minimum distance.
arg min(
T
i

k
y
k
); i =1; ; s: (16)
4.2. Algorithm
Algorithm 1. Identication algorithm
Data: Dispose of {
i
]
N
i = 1
from a given data set (
i
; y
i
)
Main steps:
- Compute P
i
for every {
i
]
N
i = 1
according to (10)
- Determine the ltered local parameters {
i
]
N
i = 1
, (N oN)
- Compute the rst cluster center
n
1
from (10)
repeat
Compute the other cluster centers according to the
updated potential formula(15)
If P
n
k
4 then
Compute V(c) such as :
V(c) = J
n
k

n
c
J; c =1; ; k1: (17)
where
n
k
is the current cluster center and
n
c
; c =1; ; k1
are the last selected ones:
if V(c)4; c =1; ; k1 then
[ accept
n
k
as a cluster center and continue
else
[ reject
n
k
and compute a new potential
end

else
[ reject
n
k
and break
end

until V(c)r; c =1; k1

Result:Determination of the number of clusters s and the
parameters {
i
]
s
i = 1
While is a small parameter characterizing the minimum distance
between the new cluster center and the existing ones.
4.3. Properties
The new clustering technique has several interesting properties
which can be summarized as follows:
+
This method automatically generates the number of sub-models.
+
This method does not require the initialization of centers.
Therefore, the problem of convergence towards local minima
does not appear.
+
This method removes the outliers by associating small values to
their potentials. This operation improves the performance of
the clustering step.
5. Simulation example
We now present an example to illustrate the performance of
the proposed method.
Hence, the objective of the simulation is to compare the
efciency of the proposed approach with that of the modied k-
means method (Ferrari-Trecate et al., 2003), the following quality
measures are used (Juloski et al., 2006):
+
The maximum of relative error of parameter vectors is dened
by

= max
i = 1;;s
J
i

i
J
2
J
i
J
2
; (18)
where
i
and
i
are the true and the estimated parameter
vectors for the sub-model i. The identied model is deemed
acceptable if

is small or close to zero.

+
The averaged sum of the squared residual errors is dened by:
s
2
e
=
1
s

s
i = 1
SSR
i
[D
i
[
; (19)
where SSR
i
=
(y(k);(k)) AD
i
(y(k)[(k)1]
i
)
2
and [D
i
[ is the car-
dinality of the cluster D
i
.The identied model is considered
acceptable if s
e
2
is small and/or close to the expected noise
variance of the true system.
+
The percentage of the output variation explained by the model
is dened by:
FIT =100 1
J ^ y yJ
2
Jyy J
2
_ _
; (20)
where ^ y and y are the estimated and the real outputs, and y is
the mean value of y.
The identied model is considered acceptable if the FIT is close
to 100.
Consider the following PWARX model (Boukharouba, 2011):
y(k) =
[0:4 0:5 0:3](k) e(k) if (k)AH
1
;
[ 0:7 0:6 0:5](k) e(k) if (k)AH
2
;
[0:4 0:2 0:2](k) e(k) if (k)AH
3
;
_

_
(21)
where s=3, n
a
=1, n
b
=1 and (k) = [y(k1) u(k1)]
T
is the
regressor vector and the partitions {H
i
]
s
i = 1
are dened by
H
1
={AR
2
: [1 0:3 0]Z0 and [0 0:5 0]40];
H
2
={AR
2
: [1 0:3 0]r0 and [1 0:3 0]o0];
H
3
={AR
2
: [1 0:3 0]Z0 and [0 0:5 0]r0]: (22)
The input signal u(k) and the noise signal e(k) are zero mean,
Gaussian distributed with variances respectively 0.5 and 0.05. The
output y(k) is illustrated in Fig. 1.
Firstly, we determine the local parameters using the least
square method with n

=20. The obtained parameter vectors are

presented in Fig. 2.
Secondly, we use the Chui's clustering algorithm in order to
separate the obtained local parameters. This algorithm consists in
computing the potential of each local parameter.The large poten-
tials represents the cluster centers. However, the lower potentials
represent the outliers which must be removed. It is easy to deduce
that the proposed method is able to remove the outliers. More-
over, it generates automatically the number of submodels. This is
shown in Fig. 3 that illustrates the local parameters separated into
Z. Lassoued, K. Abderrahim / Engineering Applications of Articial Intelligence 28 (2014) 201209 204
s=3 sets where the clusters centers are depicted by the star
symbols.
Table 1 presents the estimated parameter vectors obtained
with the proposed method and the k-means one.
The regions H
i
are determined using the SVM algorithm. For the
proposed method they are dened by the following inequalities:
H
1
= AR
2
:
0:2161 1:9693 0:1850
1:7292 0:4728 0:1822
_ _
Z0
_ _
; (23)
H
2
= AR
2
:
0:2166 1:9704 0:1853
2:1873 0:7128 0:0906
_ _
Z0
_ _
; (24)
H
3
= AR
2
:
1:7292 0:4728 0:1822
2:1873 0:7128 0:0906
_ _
Z0
_ _
: (25)
For the k-means method, the following inequalities are obtained:
H
1
= AR
2
:
0:3255 3:1594 0:1354
3:1094 1:4634 0:1090
_ _
Z0
_ _
; (26)
H
2
= AR
2
:
2:9235 1:4932 0:1469
2:4998 1:6640 0:1243
_ _
Z0
_ _
; (27)
H
3
= AR
2
:
0:3255 3:1594 0:1354
2:4998 1:6640 0:1243
_ _
Z0
_ _
: (28)
Fig. 2. Local parameters.
Fig. 3. Local parameters separated into 3 sets.
Table 1
Estimated parameters.
True values Proposed method k-means

1 0:4
0:5
0:3
_

_
_

_
0:3903
0:4653
0:3075
_

_
_

_
0:4064
0:5464
0:2598
_

_
_

2 0:7
0:6
0:5
_

_
_

_
0:7406
0:6035
0:5027
_

_
_

_
0:6955
0:5903
0:4939
_

_
_

3 0:4
0:2
0:2
_

_
_

_
0:4023
0:1601
0:1847
_

_
_

_
0:4792
0:2101
0:2406
_

_
_

_
Fig. 1. The real output of the system.
1
0.5
0
0.5
1
1.5
y
(
k
)
0 50 100 150 200 250
k
Fig. 4. Real and estimated outputs: proposed method.
1
0.5
0
0.5
1
1.5
y
(
k
)
0 50 100 150 200 250
k
Fig. 5. Real and estimated outputs: k-means method.
Z. Lassoued, K. Abderrahim / Engineering Applications of Articial Intelligence 28 (2014) 201209 205
Now, we can get the output of the system with the two
methods as it is shown in Figs. 4 and 5.
Table 2 presents the quality measures (18), (19) and (20) of the
proposed method and the k-means method.
Based on the results presented in Table 2 we observe that the
proposed method gives the best results compared with those
obtained by the k-means method (

and s
2
e
are close to zero and
FIT is close to 100).
6. Experimental example: a semi-batch reactor
This section presents an application of the proposed method to
an olive oil esterication reactor producing ester with a very high
added value which is used in ne chemical industry such as
cosmetic products.
6.1. The semi batch-reactor
The esterication reaction between vegetable olive oil with free
fatty acid and alcohol, producing ester, is given by the following
equation:
AcideAlcohol2Ester Water: (29)
The ratio of the alcohol to acid represents the main factor of this
reaction because the esterication reaction is an equilibrium
reaction i.e. the reaction products, water and ester, are formed
when equilibrium is reached. In addition, the yield of ester may be
increased if water is removed from the reaction. The removal of
water is achieved by the vaporization technique while avoiding
the boiling of the alcohol. In fact, we have used an alcohol (1-
butanol), characterized by a boiling temperature of 118 1C which is
greater than the boiling temperature of the water (close to 100 1C).
In addition, the boiling temperatures of the fatty acid (oleic acid)
and the ester are close to 300 1C. Therefore, the boiling point of
water may be provided by a temperature slightly greater than
100 1C.
The block diagram of the process is shown in Fig. 6. It is
constituted essentially of:
+
A reactor with double-jackets: It has a cylindrical shape
manufactured in stainless steel. It is equipped with a bottom
valve for emptying the product, an agitator, an orice introdu-
cing the reactants, a sensor of the reaction mixture tempera-
ture, a pressure sensor and an orice for the condenser. The
double-jackets ensure the circulation of a coolant uid which is
intended for heating or for cooling the reactor.
+
A heat exchanger: It allows to heat or to cool the coolant uid
circulating through the reactor jacket. Heating is carried out by
three electrical resistances controlled by a dimmer for varying
the heating power. It is intended to achieve the required
reaction temperature of the esterication. Cooling is provided
by circulating cold water through the heat exchanger. It is used
to cool the reactor when the reaction is completed.
+
A condenser: It allows to condense the steam generated during
the reaction. It plays an important role because it is also used to
indicate the end of the reaction which can be deduced when no
more water is dripping out of the condenser.
+
A data acquisition card between the reactor and the calculator.
The ester production by this reactor is based on three main steps
as illustrated in Fig. 7
+
Heating stage: The reactor's temperature T
r
is increased to
105 1C.
Table 2
Validation results.
Quality measures Proposed method k-means
0.0876 0.1828
s
e
2
0.0023 0.0109
FIT 89.4191 74.195
Fig. 6. Block diagram of the reactor.
0 300 540
100
Time(min)
Reacting stage Cooling stage Heating Stage
Fig. 7. Specic trajectory of the reactor temperature.
Z. Lassoued, K. Abderrahim / Engineering Applications of Articial Intelligence 28 (2014) 201209 206
+
Reaction stage: The temperature should be kept constant at
105 1C until the end of the reaction which can be detected by
the absence of water at the condenser (when no more water is
dripped out of the condenser).
+
Cooling stage: The reactor's temperature is decreased.
6.2. Existing method for modeling the reactor
An experimental study carried out on the reactor showed that the
reactor can be considered as a Single-Input Single-Output (SISO)
plant. The heating power P
e
and the reaction temperature T
r
represent respectively the input and the output of the system.
However, the variation of the quality of the reactants inside the
reactor as well as the external effects can be considered as random
disturbances. In previous work of our laboratory based on some
assumptions, the system is considered as a linear one (Mihoub et al.,
2009). An identication approach is used to estimate the structure
and the parameters of the model of this reactor. In fact, a pseudo-
random binary sequence is applied as input to the real system with
three different amplitude domains in order to excite the three steps
of the operating cycle of the reactor (heating, reaction and cooling).
The obtained measurements, presented in Fig. 8, are then used to
estimate the system structure using the instrumental determinants
ratio method (Mihoub et al., 2009) and to identify the system's
parameters with the least square approach.
The obtained system is described by the following discrete
transfer function (Mihoub et al., 2009):
H(z) =
0:0012
z
2
1:5276:z0:5468
(30)
A validation experiment of this model is then applied. The
obtained results are presented in Fig. 9 which shows that the
obtained model generates an estimated output with large error.
These poor results are due to the nonlinearity that charac-
terizes the dynamic behavior of the reactor as illustrated in
Fig. 10 that depicts the static characteristic of the system. It is
easy to remark that each step of the operating cycle present a
linear behavior. Consequently, this reactor can be modeled by a
PWARX model.
6.3. PWARX model for the semi-batch reactor
The alternative of considering a PWA map is very interesting
because the characteristic of the system can be considered as
piecewise linear in each operating phase: the heating phase,
the reacting phase and the cooling phase. Previous works have
demonstrated that the adequate estimated orders n
a
and n
b
of
each sub-model are equal to two (Talmoudi et al., 2008). Thus, we
can adopt the following structure:
y(k) =
a
1;1
y(k1) a
1;2
y(k2)b
1;1
u(k1)
b
1;2
u(k2) if (k)AH
1

a
s;1
y(k1)a
s;2
y(k2)b
s;1
u(k1)
b
s;2
u(k2) if (k)AH
s
_

_
(31)
where the regressor vector is dened by
(k) =[ y(k1); y(k2); u(k1); u(k2)]
T
and the parameter vectors is denoted by

i
(k) =[a
i;1
; a
i;2
; b
i;1
; b
i;2
]; i =1; ; s:
We have picked out some inputoutput measurements from the
reactor in order to identify a model to this process. We have taken
two measurement les, one for the identication having a length
N=220 and another one of length N=160 for the validation.
The measurement le used in this identication is presented in
Fig. 8. We apply the proposed identication procedure in order
to represent the reactor by a PWARX model with a number of
neighboring n

=70. Our purpose is to estimate the number of

sub-models s, the parameter vectors
i
(k); i =1; ; s and the
hyperplanes dening the partitions {H
i
]
s
i = 1
.
The obtained results are as follows:
+
The number of sub-models is s=3.
+
The parameter vectors
i
(k), i =1; 2 and 3 are illustrated in
Table 3. Fig. 8. The real inputoutput evolution.
0 20 40 60 80 100 120 140 160 180
0
20
40
60
80
100
120
k
Fig. 9. Model validation: real and estimated system outputs.
65
70
75
80
85
90
95
100
105
110
1000 1200 1400 1600 1800 2000 2200
P (W)
Fig. 10. Static characteristic of the reactor.
Z. Lassoued, K. Abderrahim / Engineering Applications of Articial Intelligence 28 (2014) 201209 207
It remains only to determine the partitions {H
i
]
s
i = 1
that are
dened by the following inequalities:
H
i
={AR
4
: M
i
Z0]; (32)
M
1
=
0:3869 0:3130 0:0001 0:0015 4:8740
0:0164 0:0035 0:0018 0:0026 0:7355
_ _
; (33)
M
2
=
0:4548 0:3862 0:0001 0:0014 4:5891
0:2871 0:2519 0:0021 0:0034 0:4359
_ _
;
(34)
M
3
=
0:0163 0:0034 0:0018 0:0026 0:7335
0:1342 0:1041 0:0019 0:0031 0:2969
_ _
: (35)
To validate the obtained models, we have considered a new input
output measurement le having a length N=160. The obtained results
are presented in Fig. 11 which shows that the proposed approach gives
the best results in terms of precision.
7. Conclusions
In this paper, we have considered the problem of clustering
based procedure for the identication of PWARX models. In fact,
we have recalled the main steps of clustering based approach.
This presentation shows that the used clustering algorithms
require a good initial guess in order to converge to global minima.
In addition, these algorithms are sensitive to the presence of
outliers. To overcome this problem, we have suggested the Chiu's
algorithm for data classication. This algorithm presents several
advantages. Firstly, it does not require any initialization and the
problem of convergence towards local minima is overcome.
Secondly, this method is able to remove the outliers by associating
potentials to these points. Finally, our approach generates the
number of sub-models. Numerical simulation results are pre-
sented to demonstrate the performance of the proposed approach.
Also, an experimental validation with an olive oil reactor is
presented to illustrate the efciency of the developed method.
References
Bako, L., 2011. Identication of switched linear systems via sparse optimization.
Automatica 47 (4), 668677.
Bako, L., Boukharouba, K., Lecoeuche, S.T., 2012. Recovering piecewise afne maps
by sparse optimization. In: Proceedings of the IFAC Conference on System
Identication, pp. 356361.
Bako, L., Lecoeuche, S., 2013. A sparse optimazation approach to state observer design
for switched linear systems. In: Systems & Controls Letters, vol. 62, pp. 143151.
Bemporad, A., Morari, M., 1999. Control of systems integrating logic and con-
straints. Automatica 35 (3), 407428.
Bemporad, A., Ferrari-Trecate, G., Morari, M., 2000. Observability and controllability
of piecewise afne and hybrid systems. IEEE Trans. Autom. Control 45 (10),
18641876.
Bemporad, A., Garulli, A., Paoletti, S., Vicino, A., 2005. A bounded-error approach to
piecewise afne system identication. IEEE Trans. Autom. Control 50 (10),
15671580.
Boukharouba, K., 2011. Modlisation et classication de comportements dynami-
ques des systemes hybrides (Ph.D. thesis). Universit de Lille, France.
Chiu, S., 1994. Fuzzy model identication based on cluster estimation. J. Intell.
Fuzzy Syst. 2 (3), 267278.
De Schutter, B., 1999. Optimal control of a class of linear hybrid systems with
saturation. In: Proceedings of the 38th IEEE Conference on Decision and
Control, pp. 39783983.
De Schutter, B., Van den Boom, T., 2000. On Model Predictive Control for Max-Min-
Plus-Scaling Discrete Event Systems. Technical Report bds 00-04, Control
Systems Engineering, Faculty of Information Technology and Systems, Delft
University of Technology, The Netherlands.
Doucet, A., Gordon, N., Krishnamurthy, V., 2001. Particle lters for state estimation
of jump Markov linear systems. IEEE Trans. Signal Process. 49 (3), 613624.
Duda, R.O., Hart, P.E., Stork, D.G., 2001. Pattern Classication, Second edition John
Wiley and Sons.
Feng, X., Loparo, K.A., Ji, Y., Chizeck, H.J., 1992. Stochastic stability properties of
jump linear systems. IEEE Trans. Autom. Control 37 (1), 3853.
Ferrari-Trecate, G., Muselli, M., 2003. Single-linkage clustering for optimal classi-
cation in piecewise afne regression. In: Proceedings Volume from the IFAC
Conference on Analysis and Design of Hybrid Systems (ADHS 03).
Ferrari-Trecate, G., Muselli, M., Liberati, D., Morari, M., 2003. A clustering technique
for the identication of piecewise afne systems. Automatica 39 (2), 205217.
Heemels, W., De Schutter, B., Bemporad, A., 2001. On the equivalence of classes of
hybrid dynamical models. In: Proceedings of the 40th IEEE Conference on
Decision and Control, vol. 1, pp. 364369.
Hsu, C., Lin, C., 2002. A comparison of methods for multiclass support vector
machines. IEEE Trans. Neural Netw. 13 (2), 415425.
Juloski, A., Weiland, S., Heemels, W., 2005. A Bayesian approach to identication of
hybrid systems. IEEE Trans. Autom. Control 50 (10), 15201533.
Juloski, A., Paoletti, S., Roll, J., 2006. Recent Techniques for the Identication of
Piecewise Afne and Hybrid Systems, Current Trends in Nonlinear Systems and
Control. Springer, 7999.
Lin, J., Unbehauen, R., 1992. Canonical piecewise-linear approximations. IEEE Trans.
Circuits Syst. I: Fundam. Theory Appl. 39 (8), 697699.
Mihoub, M., Nouri, A.S., Abdennour, R.Ben., 2009. Real-time application of discrete
second order sliding mode control to a chemical reactor. Control Eng. Pract.
17 (9), 10891095.
Nakada, H., Takaba, K., Katayama, T., 2005. Identication of piecewise afne systems
based on statistical clustering technique. Automatica 41 (5), 905913.
Talmoudi, S., Abderrahim, K., Abdennour, R.Ben., Ksouri, M., 2008. Multimodel
approach using neural networks for complex systems modeling and identica-
tion. Nonlinear Dyn. Syst. Theory 8 (3), 299316.
Table 3
Estimated parameter vectors.
Parameter vectors Estimated parameters

1
1:4256
0:4508
4:9853 10
4
0:0010
_

_
_

2
1:1604
0:2111
0:0015
0:0014
_

_
_

3
1:0847
0:1490
3:9782 10
4
0:0040
_

_
_

_
0 20 40 60 80 100 120 140 160 180
0
20
40
60
80
100
120
k
y
(
k
)
Estimated Output PWARX Model Proposed Method
Real System Output
Estimated Output Validation Existing Method
0 20 40 60 80 100 120 140 160 180
0
500
1000
1500
2000
u
(
k
)
Fig. 11. The input, the real and the estimated validation outputs.
Z. Lassoued, K. Abderrahim / Engineering Applications of Articial Intelligence 28 (2014) 201209 208
Tian, Y., Floquet, T., Belkoura, L., Perruquetti, W., et al., 2011. Algebraic switching
time identication for a class of linear hybrid systems. Nonlinear Anal. Hybrid
Syst. 5 (2), 233241.
van der Schaft, A., Schumacher, J.M., 1998. Complementarity modeling of hybrid
systems. IEEE Trans. Autom. Control 43 (4), 483490.
Wang, L., 2005. Support Vector Machines: Theory and Applications. Springer.
Weston, J., Watkins, C., 1999. Support vector machines for multi-class pattern
recognition. In: Proceedings of the Seventh European Symposium on Articial
Neural Networks, vol. 4, pp. 219224.
Z. Lassoued, K. Abderrahim / Engineering Applications of Articial Intelligence 28 (2014) 201209 209