a methodology for modeling based on fuzzy clustering
TRANSCRIPT
-
7/31/2019 A Methodology for Modeling Based on Fuzzy Clustering...
1/6
5TH INTERNATIONAL SYMPOSIUM ON ROBOTICS AND AUTOMATION 2006, SAN MIGUEL REGLA HIDALGO, MEXICO, AUGUST 25-28, 2006. 1
A methodology for modeling based on fuzzy
clustering - cubic splines for nonlinear systems
Julio Cesar Ramos Fernandez1, Virgilio Lopez Morales2, Gilles Enea3, Jean Duplaix4.
Abstract This paper proposes a novel methodology based onfuzzy clustering and cubic splines for modeling nonlinear systems.The Gustafson-Kessel (G-K) algorithm is used in order to classifyclusters with linear domains. At least in each three different andordered consecutive clusters (with linear trend), there are twopoints of inflexion, one maximum and one minimum. Then, acubic spline can be synthesized, and the intersection with thenext spline is smoothing via fuzzy sub-models. An advantage ofthis approach is a reduction on rules number, and consequentlyan important reduction of Takagi-Sugeno (T-S) models. We obtainmodels with the rule-base, which have trapezoidal membershipfunctions for input and cubic spline for output. An example is
proposed to illustrate our modeling approach.Keywords models reduction, fuzzy modeling, local models,
cubic spline.
I. INTRODUCTION
In control engineering, modeling and identification are
important steps in the design of control laws, supervision and
fault-detection systems. Traditionally modeling is seen as a
conjunction of a thorough understanding of the systems nature
and behavior, and of a suitable mathematical treatment that
leads to an usable model. This approach is usually termed
white box (physical, mechanistic, first-principle) modeling
[Babuska, 1998].
The identification problem consists of estimating the param-
eters in the model. If representative process data is available,
black-box usually models can be developed easily, without
requiring process specific knowledge. A severe drawback of
this approach is that the structure and parameters of these
models usually do not have any physical significance. There
is a range of modeling techniques that attempt to combine the
advantages of the white-box and black-box, such that the know
parts of the systems are modelled using physical knowledge,
and the unknown or less certain parts are approximated in a
black-box approach, using process data and black-box model-
ing structures with suitable approximation properties. These
methods are often denoted as hybrid, semi-mechanistic or
The work of Julio Cesar Ramos Fernandez is partially supported by thePROMEP grant and ANUIES-SEP-CONACyT/ECOS NORD, M02:M03. Thework of Virgilio Lopez Morales is partially supported by the SEP-SESIC-PROMEP research project grant PROMEP - UAEH-PTC-1004 and ANUIES-SEP-CONACyT/ECOS-NORD, M02:M03.
1 Author is at the Universidad Tecnologica de Tula-Tepeji, Col. El33 El Carmen Tula de Allende Hidalgo, Mexico; 2 Author is at theUniversidad Autonoma del Estado de Hidalgo, Centro de Investigacionen Tecnologas de Informacion y Sistemas, Carr. Pachuca-TulancingoKm. 4.5, C.P. 42090, Pachuca, HGO, Mexico; 3,4 Authors are atthe Laboratoire des Sciences de lInformation et des Systemes UMR-CNRS 6168 Equipe COSI, Universite du Sud-Toulon Var, La [email protected], [email protected],[enea,duplaix]@univ-tln.fr
gray modeling [Lafont and Balmat, 2003], [Babuska, 1998].
The modeling approach proposed in this paper is based on
fuzzy models, which describe relationships between variables
by means of If-Then rules, such as:
Ri : If x is region A Then y = f(x) (1)
where Ri is the number of rules or models. Among the differ-
ent fuzzy models methods, the T-S has attracted most attention
[Tanaka and Wang, 2001]. In fact this model consist of rules
like [1], with fuzzy antecedents (region A) and mathematical
functions in the consequent part (y = f(x)). In this paperwe obtain fuzzy models by using G-K algorithm to detectconsecutive linear regions. A cubic spline is calculated by each
three linear regions. An important issue in fuzzy-rule-based
modeling is how to select a set of important fuzzy rules from
the rule base can result in a compact fuzzy model with better
generalizing ability. In [Yen and Wang, 1999] it is introduced
several orthogonal transformation-based methods that provide
new or alternative tools for parameters and rule selection.
Orthogonal Least Squares (OLS) method is used to remove
redundant or less important clusters during the clustering
process. The drawbacks on these approaches is the initializa-
tion of clustering process with an overestimated number of
clusters. The big question is how many number of clustersin the partition are overestimated?. In [Setnes et al., 1998]
it is shown some results to reduce the number of initials
models by using a measure of similarity between the initial
fuzzy sets. Therefore the fuzzy sets with redundant information
are merged to create a common fuzzy set to replace them
in the rule base. In nonlinear systems identification, both
the amplitude and frequency contents of the input signals
are of major importance [Johansen and Murray-Smith, 2000].
The principal components analysis (PCA) allows to reduce
a complex correlation systems into a smaller number of
dimensions, in [Pessel and Balmat, 2005], a reduction in the
matrix of variables input/output is achieved, an application on
a real system (an experimental greenhouse) with a neuronalmodeling. A drawback is that reduction on input variables or
excitations of the modeling, avoid the possibility to design
control laws. Several works with the B-spline basis functions
of different orders are regarded as a class of a membership
functions. Recently, splines have also been proposed for neu-
ronal network modelling [Brown and Harris, 1987] and design
of fuzzy controllers using periodical non uniform B-spline
In [Shimojima et al., 1994] it is proposed a self-tuning fuzzy
inference neuronal networks of membership functions that
are represented by spline function. A key problem is on the
obtention of the region of the spline in order compute the knots
-
7/31/2019 A Methodology for Modeling Based on Fuzzy Clustering...
2/6
5TH INTERNATIONAL SYMPOSIUM ON ROBOTICS AND AUTOMATION 2006, SAN MIGUEL REGLA HIDALGO, MEXICO, AUGUST 25-28, 2006. 2
position of the spline, since the shape of the spline function
depends of the knots.
Several algorithms for ordinary least squares regression
spline fitting are to exhibit in [Thomas, 2002], an important
issue associated with it is the choice of a theoretical best set
of knots. Typically these best knot sets are defined implicitly
as the optimizers of some objective functions.
The methodology proposed in this paper is on the rule-based
structure proposed by T-S, where it is used linear consequents.
The advantages with our approach is a reduction in the rule-
based, and the consequents are nonlinear (cubic) with constant
parameters. We thought that the cubic consequents may char-
acterizer the transients or nonlinearity of the system better
than the linear consequents (T-S modeling). Our methodology
proposed in this paper for modeling nonlinear systems is based
in fuzzy clustering. Because fuzzy clustering classifies objets
according to similarities among them, and the organizing of
data in groups. The term similarity should be understood as
mathematical similarity, measured in some well-defined sense.
In metric spaces, similarity is often defined by means of a
distance norm [Bezdek et al., 1981].This paper has the following structure, in Section 2, we
describe the algorithm of G-K and the knot selection,
and its ability to detect linear regions, by using its
eigenvalues of the covariance matrix on each cluster
obtained. In Section 3, the interpolation cubic splinealgorithm is explained, and we obtain cubic polynomials, are
obtained for the local interpolation, on the interpretation of
local models. This approach is widely used by different
authors, but they was computed linear consequents,
[Babuska, 1998], [Trabelsi et al., 2004], [Yen et al., 1998],
[Foulloy et al., 2003], [Tanaka and Wang, 2001],
[Sam and Rui, 1993], [Bernal Reza, 2005]. Our approach
proposes cubic consequents with the following:
y(x) = a3 x3 + a2 x2 + a1 x + a0. (2)In Section 4 the fuzzy modeling is obtained. The antecedentsare defined by trapezoidal membership function, and the
consequents by cubic equations. Section 5 discusses the mainresults and gives the concluding remarks.
I I . GUSTAFSON-K ESSEL ALGORITHM AND KNOT
SELECTION
The G-K algorithm is a derived from de basic fuzzy c-
means (FCM), by adaptingdistance norm
, in order to de-tect clusters of different geometrical shapes in one data set
[Babuska, 1998], another algorithm that use the adapting dis-
tance norm is based on fuzzy maximum likelihood estimates
clustering (FMLE).
Each cluster has its own norm-inducing matrix Ai, which
yields the following inner-product norm:
D2ikAi = (zk vi)TAi(zk vi) (3)in the FCM algorithm the matrices Ai are identity matrix
(I), in G-K algorithm the matrices Ai employing an adaptivedistance norm. The cost function of the G-K algorithm is
defined by:
J(Z; U , V, A) =c
i=1
Nk=1
(ik)mD2ikA (4)
where U Mfc, V nc and m > 1.Definition 1. (fuzzy partition space) Let
Z = {z1, z2,...,zN} be a finite set and let 2 c N be an in-teger. The fuzzy partitioning space for data matrix Z is the set
Mfc =
U cn ik [0, 1] , i, k; ci=1
ik = 1, k;
0 1 and the termination tolerance > 0. Initializethe partition matrix randomly, such that U0 Mfc Repeat forl = 1, 2,... Step 1: Compute clusters prototypes, 1 i c:
vli = N
k=1 l1ik
mzk
N
k=1 l1ik
m
1
(12)
Step 2: Compute the cluster covariance matrices, 1 i c:
Fi =N
k=1 (l1ik )
m(zk vli)(zk vli)T
Nk=1 (
l1ik )
m1 (13)
Step 3: Compute the distances, 1 i c, 1 k N:D2ikAi = (zk vli)T
i det(Fi)
1/nF1i
(zk vli) (14)
Step 4: Update the partition matrix, if (DikAi > 0), for(1 i c, 1 k N):
lik =
c
j=1
DikAi/DjkAi
2/(m 1)1 (15)
otherwise
lik = 0 if DikAi > 0, and lik [0, 1] with
ci=1
lik = 1.
UntilUl Ul1 <
III. INTERPOLATION CUBIC SPLINE ALGORITHM, AN D
CUBIC CONSEQUENT
In this paper we have integrated two well known technics:
fuzzy clustering, and cubic spline interpolation. The goal is to
-
7/31/2019 A Methodology for Modeling Based on Fuzzy Clustering...
4/6
5TH INTERNATIONAL SYMPOSIUM ON ROBOTICS AND AUTOMATION 2006, SAN MIGUEL REGLA HIDALGO, MEXICO, AUGUST 25-28, 2006. 4
obtained a structure in the rule base with cubic consequents.
The main steps are: for each rule a submodel is obtained, the
consequents are analytical functions, a reduction in the rule
base compared with the T-S models with linear consequents.
The algorithm that we have used for the cubic spline interpola-
tion is shown in [Madar et al., 2003]. The first version appears
in [Horiuchi et al., 1993]. The cubic splines are defined as
a continuous function and fitted to the available measured
data system input/output (SISO). The input/output vector Y =[y1,...,yN]
T, X = [x1,...,xN]T, which are defined by each
pair of knot: [K11 , K12 ], [K
21 , K
22 ], ..., [K
n11 , K
n2 ].
In order to formulate this algorithm, let us define a cubic
spline for a knot sequence of cubic polynomials: x1 = K11