Fuzzy Sets and Systems 125 (2002) 122

www.elsevier.com/locate/fss

Design of robust fuzzy-model-based controller with sliding mode

control for SISO nonlinear systems
Wook Changa , Jin Bae Parka; , Young Hoon Joob , Guanrong Chenc
a A432, Department of Electrical and Electronic Engineering, Yonsei University, Shinchondong, Seodaemoonku,
Seoul 120-749, South Korea
b School of Electronics and Information Engineering, Kunsan University, Kunsan, Chonbuk, 573-701, South Korea
c Department of Electronic Engineering, City University of Hong Kong, Kowloon, Hong Kong

Received 14 May 1999; received in revised form 16 June 2000; accepted 18 December 2000

Abstract
In this paper, we present the design of a new type of fuzzy controllers for controlling complex single-inputsingle-output
systems by incorporating sliding mode control theory with fuzzy control technology. First, a fuzzy model of the given
nonlinear system is constructed to represent the local dynamic behaviors of the given nonlinear system. A global controller is
then constructed by combining all local state feedback controllers and a global supervisory sliding mode controller. Improved
robustness and tracking performance are obtained by a seamless integration of fuzzy-model-based control and sliding mode
control. In addition, stability of controlling the original underlying nonlinear systems is guaranteed owing to the robustness
of the sliding mode control method to system parameter uncertainties. Two simulation examples are included to show the
e:ectiveness and feasibility of the developed integrated controller.
c 2002 Elsevier Science B.V. All rights reserved.

Keywords: Fuzzy-model-based control; Sliding mode control; TS fuzzy model

1. Introduction controller. Fuzzy logic control is generally applica-

ble to plants that are mathematically ill-modeled but
Fuzzy logic control is a conspicuous and successful qualitative knowledge of experienced operators is
branch of automation and control theory, which is de- available for design. Fuzzy logic controller is par-
veloped based on the fuzzy set theory of Zadeh [31]. ticularly suitable for those systems with uncertain or
The concept of fuzzy logic control is to utilize the complex dynamics.
qualitative knowledge of a system to design a practical Typical design methodologies for fuzzy logic con-
trollers may be divided into two categories: One is
Corresponding author. Tel.: +82-2-2123-2773; fax: +82-2- model-free and the other is model-based. In general,
362-4539.
E-mail addresses: pegasus@control.yonsei.ac.kr (W. Chang),
a speciGc design of a fuzzy logic controller has dif-
jbpark@control.yonsei.ac.kr (J.B. Park), yhjoo@kunsan.ac.kr Gculties in the acquisition of experts knowledge and
(Y.H. Joo). relies to a great extent on empirical and heuristic

0165-0114/02/$ - see front matter  c 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5 - 0 1 1 4 ( 0 1 ) 0 0 0 3 8 - 0
2 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

knowledge which, in many cases, cannot be justi- mined. The stability of the overall system is then
Ged. Therefore, the performance of fuzzy logic con- determined by Lyapunov stability analysis. This kind
trollers can be degraded in the case of plant parameter of design methods su:er mainly from a few limita-
variations or unpredictable incidents which a designer tions: (1) A common positive deGnite matrix must be
may not be aware or have ignored. Moreover, the pa- found to satisfy a matrix Lyapunov equation, which
rameters of a fuzzy logic controller obtained based can be diJcult especially when the number of fuzzy
on experts experience may not be optimal. To over- rules required to give a good plant model is large
come these diJculties, in recent years many e:orts has so that the dimension of the matrix equation is high.
been devoted to the development of systematic anal- (2) The performance of the closed-loop system is dif-
ysis and design procedure for fuzzy control systems Gcult to predict. (3) The stability is guaranteed only
[14,6 8,12,19 21,15 18,24 30]. for those simpliGed TS fuzzy models although they
Among various kinds of fuzzy control methods, have successfully applied to the original underlying
fuzzy-model-based control is widely used because the nonlinear systems. (4) Tracking problem is not explic-
design and analysis of the overall fuzzy system can be itly addressed. (5) If the original nonlinear system is
systematically performed using the well-established partially known, then it is very diJcult (if not impos-
classical linear systems theory. For example, Takagi sible) to determine the stability of the overall fuzzy
and Sugeno [24] suggested to represent a nonlinear system. The linear matrix inequality (LMI) based
system by a combination of some local linear subsys- approaches have been used to determine the exis-
tems. They have applied this modeling and control tence of a common positive deGnite matrix [28]. And
technique to the control of water cleaning process Cao et al. suggested an alternative method to design
and steel-making process. Lately, Cao et al. [1 3] a fuzzy-model-based controller, where the stability
applied the modern control theory for analysis and de- of the closed-loop controlled system is investigated
sign of some general fuzzy-model-based controllers. with a switching-type fuzzy-model-based controller
Tanaka and Kosaki [25] implemented a fuzzy-model- and piece-wise Lyapunov function [1,3]. Although
based controller for an articulated vehicle with these results solve the problem of determining a
rigorous stability analysis. The idea of using the common positive deGnite matrix, their computation
TakagiSugeno (TS) fuzzy model as the inference can still be expensive and the other problems remain
system in a fuzzy-model-based controller design was unresolved.
Grst proposed in 1985 [24]. In this approach, the TS In the analysis and design of e:ective fuzzy logic
fuzzy model substitutes the consequent fuzzy set in controllers, we believe that it is, nevertheless, desir-
a fuzzy rule by a linear equation of the input vari- able to utilize the well-established modern control
ables. Some variants of the TS fuzzy model have theory and methodology. In this context, the sliding
also appeared in the recent literature aiming at bet- mode control (or variable structure control) is par-
ter analysis and design of fuzzy logic controllers. ticularly useful since it can provide very robust con-
These systems sometimes are called dynamic fuzzy trol performance [11,13,23]. There have been quite a
models [3]. A dynamic fuzzy model represents the lot of research on the combination of sliding mode
local dynamic behaviors of a given nonlinear system control and fuzzy logic control techniques for im-
by linear state space equations in di:erent regions. proving the robustness and performance of uncertain
More precisely, the TS fuzzy model employs lin- systems [7,26,30,27,10]. In [7,30], for example, fuzzy
ear state space equations in the consequent parts in controllers are designed to satisfy the sliding condi-
fuzzy rules. Therefore, it has a convenient analytic tions and applied to nonlinear systems exclusively. In
structure so that some well-established conventional these approaches, one generally focuses on the de-
linear systems theory can be easily applied for theo- sign of sliding surface via the fuzzy logic theory.
retical analysis of the overall closed-loop controlled One approach is proposed by Wang as an alternative
system. to combine the fuzzy control theory with the sliding
Typical fuzzy-model-based controllers use a set mode control [27]. In this approach, the nonlinear dy-
of fuzzy rules to construct suitable local linear state namics of the given system are Grst described by the
models from which local controllers can be deter- Mamdani-type fuzzy inference systems [22] and then
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 3

the additional supervisory control input is designed pendulum system are given to show the advantages
by the adaptive sliding mode control theory. Addi- and e:ectiveness of the proposed approach. Finally,
tionally, in [26], an algorithm is developed to design conclusions are drawn in Section 5.
a supervisory controller with a pre-designed fuzzy
controller by using the sliding mode control theory,
where an application is focused on the Mamdani-type 2. Fuzzy-model-based controller
fuzzy inference systems. Along this line of approach,
Feng et al. [10] suggested a fuzzy-model-based con- Consider a class of uncertain SISO nonlinear dy-
troller which guarantees the stability of the closed- namic systems of the form
loop controlled system. They focused on the regulation
problem and the stability of the closed-loop system x(n) = f(x) + g(x)u; (1)
consisting of the TS fuzzy model and the switching-
where the scalar variable x is the output of interest, u
type fuzzy-model-based controller.
is the control input, and x = [x x x(n1) ]T is the
In this paper, our suggestion is yet another alter-
system state vector. Detailed description of nonlinear
native by means of improving the fuzzy logic control
function f(x) and g(x) are given in the Section 3 be-
method via the sliding mode principle. More precisely,
low. The control problem is to drive the state x to track
we present a systematic design procedure of fuzzy-
a speciGc time-varying state xd = [xd xd xd(n1) ]T
model-based controllers with guaranteed stability and
in the presence of model imprecision on f(x) and
improved tracking performance for the original SISO
g(x). These kinds of systems exist in many me-
nonlinear system. The proposed method utilizes the
chanical systems and are said to be in companion
fuzzy model whose antecedent parts are conventional
form.
fuzzy IF part and the consequent parts are local linear
This SISO nonlinear system can be approxi-
state space models of the given system. The global
mated by the TS fuzzy model, which combines
fuzzy logic controller is constructed by combining
the fuzzy inference rule and the local linear state
such local fuzzy-model-based controllers with a global
model [2,24]. The ith rule of the TS fuzzy model,
sliding mode controller. The proposed method bene-
representing the complex SISO system (1), is the
Gts from both the fuzzy logic control and the sliding
following:
mode control. The design procedure is mainly com-
posed of two parts: One is to design the local compen-
Plant Rule i:
sators for each local linear state model; the other is to
design a supervisory controller whose input is deter- IF x(t) is F1i and and x(n1) is Fni ;
mined by the sliding mode control theory. In doing so,
the overall robustness and tracking ability of the en- THEN x = Ai x(t) + Bi u(t)
tire closed-loop system can be signiGcantly improved. (i = 1; 2; : : : ; r); (2)
Moreover, one can design a global stable fuzzy logic
controller without Gnding a common Lyapunov func- where Rule i denotes the ith fuzzy inference rule,
tion. Furthermore, it guarantees the stability for the Fji ( j = 1; 2; : : : ; n) are fuzzy sets, x(t) Rn is the
original underlying nonlinear system rather than just state vector, u(t) R1 is the input control, Ai Rnn ,
for the local TS models used for approximation. Two Bi Rn1 , r is the number of fuzzy IFTHEN
controller design algorithms are developed for a sys- rules, and the constant matrices are in controllability
tem without uncertainty and a system with uncertainty, canonical form and given by
respectively.
0 1 0 0 0
This paper is organized as follows: Conventional 0 0 1 0 0
fuzzy-model-based controller is Grst brieMy reviewed
.. ; ..
in Section 2. In Section 3, the proposed robust fuzzy- Ai = ... ..
.
.. . .
. . . B i = . :

model-based controller is detailed. In Section 4, 0 0 0 1 0
i i i i
applications to speciGc examples such as the chaotic a1 a2 a3 an bi
DuJng forced-oscillation system and the inverted (3)
4 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

By using the fuzzy inference method with a singleton fuzzy model can still be obtained by some auto-
fuzziGer, product inference, and center average de- matic identiGcation methods such as RLS (recursive
fuzziGer, the fuzzy model (2) can be expressed as the least-squares) or BP (back-propagation) algorithms
following global model: [24,2,16 18,5,14].
r In order to design a global controller for the TS
wi (x(t))(Ai x(t) + Bi u(t)) fuzzy model (2) for the original nonlinear system (1),
x(t) = i=1 r
i=1 wi (x(t)) two types of fuzzy-model-based controllers are avail-
r
 able, i.e., parallel distributed compensation (PDC)
= i (x(t))(Ai x(t) + Bi u(t)) [28] and switching control law [1,3]. The main idea
i=1 of the PDC design is to derive each control rule so
= A((x(t)))x(t) + B((x(t)))u(t); (4) as to compensate each plant rule of the fuzzy sys-
tem and construct a global fuzzy controller by the
where aggregation of the local compensators with fuzzy
n inference system. Unlike PDC technique, switch-

wi (x(t)) = Fji (x(j1) (t)); ing control law is to use only one compensator in
j=1 each sub-space (not fuzzily partitioned sub-space).
In this paper, we concentrate on improving the per-
wi (x(t)) formance of PDC with well-known nonlinear control
i (x(t)) = r ;
i=1 wi (x(t)) theory.
Using the same premise as (2), the ith rule of
(x(t)) = (1 (x(t)); 2 (x(t)); : : : ; n (x(t)))
the fuzzy-model-based controller can be obtained as
and Fji (x( j1) (t)) is the grade of membership of follows:
x( j1) (t) in Fji . It is assumed, as usual, that
Controller Rule i:
r

wi (x(t)) 0 (i = 1; 2; : : : ; r); wi (x(t)) 0 IF x(t) is F1i and and x(n1) (t) is Fni ;
i=1
THEN u = Ki x
for all t. Therefore,
(i = 1; 2; : : : ; r); (5)
r

i (x(t)) 0 (i = 1; 2; : : : ; r); i (x(t)) = 1
where Ki = [k1i kni ] is the feedback gain vector.
i=1
The fuzzy controller (2) is analytically represented by
for all t. For simplicity of notation, let wi = wi (x(t)),
r
i = i (x(t)), and  = (x(t)). i=1 wi (x(t))(Ki x(t))
u(t) = r
i=1 wi (x(t))
Denition 1 (Cao et al. [3]). Model (4) is called
r

the global state-space model of the fuzzy system
= i (x(t))Ki x(t) = K()x(t): (6)
(2). If the pairs (Ai ; Bi ), i = 1; 2; : : : ; r, are control-
i=1
lable, the fuzzy system (2) is said to be locally
controllable. The overall closed-loop fuzzy system obtained by
combining (2) and (6) becomes
The above TS fuzzy model can be obtained by
various identiGcation methods. If the exact dynamic r 
 r
model of the nonlinear system is available, TS fuzzy x = (Ai Bi Kj )x: (7)
model can be identiGed by linearizing the known non- i=1 j=1
linear equations over a number of operating points
[6,15,25,28]. Even if one can only obtain the input A well-known stability condition of fuzzy system (7)
output data pairs from the nonlinear system, the TS is given as follows:
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 5

Theorem 1 (Wang et al. [28]) . The equilibrium of for humans. Additionally, the knowledge of hu-
the fuzzy control system (7) is asymptotically stable man experts can be directly incorporated into the
in the large if there exists a common positive de7nite fuzzy-model-based controller, which makes fuzzy
matrix P such that the following two equations are logic controllers more attractive than other nonlinear
satis7ed: controllers.

(Ai Bi Ki )T P + P(Ai Bi Ki )0
3. Robust fuzzy model-based controller with sliding
(i = 1; 2; : : : ; r) (8) mode control
and
In this section, we describe the new fuzzy-model-
GijT P + PGij 0 (1 6 ij 6 r); (9) based controller which can resolve most of the prob-
lems mentioned in Section 2. The control problem is
where to design a controller u to drive the state vector x to
track a speciGc trajectory, xd = [xd xd xd(n1) ]T . In
Ai Bi Kj + Aj Bj Ki
Gij = : order to control the given nonlinear system (1) based
2
on the TS fuzzy model (2), we combine the conven-
If the common positive deGnite matrix P can be tional fuzzy-model-based controller and sliding mode
found, one can determine the stability of the closed- control theory to achieve synergetic performance.
loop fuzzy system (7). However, in many cases, it The controller design is categorized into two cases:
is diJcult to Gnd this common positive deGnite ma- control of nonlinear system without uncertainty
trix P and, therefore, the closed-loop fuzzy system (Case 1) and control of nonlinear system with uncer-
cannot be guaranteed to be stable (the closed-loop tainty (Case 2). Algorithms are developed for each
system may be unstable although local linear sys- case below.
tems are stable). Conventional fuzzy-model-based
control also has drawbacks that stability analysis is 3.1. Case 1
not conducted for the original underlying nonlin-
ear system but for the simpliGed TS fuzzy model The Grst case is the control of nonlinear system
instead, and that the control objective is mainly fo- without uncertainty. In this case, nonlinear function
cused on the regulation of the given nonlinear sys- f(x) and g(x) in (1) are assumed to be exactly known.
tem (that is, the tracking problem is not explicitly Moreover, the control gain g(x) is assumed to be the
addressed). invertible function of x. One has to provide a sat-
Since the sliding mode control theory [11,13,23] isfactory fuzzy model and then to design the con-
can deal with system parameter uncertainties and can troller based on this fuzzy model. The basic control
achieve good tracking performance, we apply sliding strategy here is the conventional linear control the-
mode control to improve the performance of the fuzzy- ory, since the fuzzy model used in this paper uti-
model-based controller. In the next section, we will lizes a linear state-space model for its consequent
derive a new kind of fuzzy-model-based controller parts.
that guarantees the stability for the original uncer- Most fuzzy-model-based controllers in the literature
tain nonlinear system with improved robustness and are regulators. This is quite restrictive since many ap-
the tracking performance. The approach is to integrate plications require tracking control. In this paper, we
seamless blending of fuzzy-model-based control and use an additional control input to design the intended
sliding mode control together. fuzzy-model-based controller.
Finally, we remark that the fuzzy-model-based Using the same premise as (2), the ith rule of the
control algorithm still e:ective even if the orig- proposed fuzzy-model-based controller is
inal nonlinear system dynamic equations for the Controller Rule i:
plant are available since the resulting fuzzy-model-
based controller is intuitive, easy to understand IF x(t) is F1i and and x(n1) is Fni ;
6 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

THEN u(t) = Ki x(t) + us (t) with

(n1)
(i = 1; 2; : : : ; r): (10) d
s(x; t) = + x
dt
The Grst term Ki x(t) of u(t) is the same as a typi-
cal fuzzy-model-based controller, and the second term = x(n1) + a1 x(n2) + + an1 x; (14)
us (t) is termed a supervisory control. The defuzziGed
output of fuzzy logic controller is where  is a strictly positive constant.
Given an initial condition, the problem of tracking
r xd reduces to that of keeping the scalar function s at
i=1 wi (x(t))(Ki x(t) + us (t))
u(t) = r zero. More precisely, the nth-order tracking problem in
i=1 wi (x(t)) x can be replaced by a Grst-order stabilization problem
r
 in s [23]. The simpliGed, 1st-order problem of keeping
= i (x(t))Ki x(t) + us (t); (11) the scalar s at zero can now be achieved by choosing
i=1 the control law such that
r d 2
where i = wi = i=1 wi . The simplest controller struc- s 6 |s|: (15)
ture for the supervisory control in (11) is to construct dt
the feedforward control input as the multiplication of Di:erentiating s(x; t) with respect to time and using
a feedforward gain and a reference input. However, (12), we obtain
this kind of supervisory controller is quite sensitive
to plant modeling errors and the stability and per- S
s = F(x) + g(x)us (t);
formance of the global fuzzy controller is diJcult to
S
F(x) (16)
= F(x) xd(n) + a1 x(n1) + + an1 x:
predict. Therefore, we use the sliding mode control
theory to design the supervisory control which guar-
Having designed the sliding mode control via the
antees the enhanced robustness and performance.
design of switching functions, the next step is to design
The closed-loop system is obtained from the feed-
the reaching mode and the overall control law. This
back interconnection of the nonlinear system (1) and
gives not only the desired sliding mode but also the
the controller (11), as
desired system dynamics in the reaching mode. The
main requirement in this design is that the control
x(t)(n) = F(x(t)) + g(x(t))us (t); (12)
law should satisfy the reaching condition, which, in
turn, guarantees the existence of the sliding mode on
where
the switching manifold. For this purpose, we adopt
r
 the reaching law method [11]. A general form of the
F(x(t)) = f(x(t)) g(x(t)) i Ki x(t): reaching law is
i=1
s = Q sgn(s) Kh(s); (17)
The scalar supervisory input us (t), which is intro-
duced to guarantee the system stability, is determined where Q and K are strictly positive constants, sh(s)0
based on the well-known sliding mode control theory for all s = 0 and h(0) = 0.
[23] as follows: Having selected the reaching law equation (17), the
Let x = x xd be the tracking error and let control law can then be determined by equating (16)
and (17):
x = x xd = [x x(n1) ]T : (13) S
s = F(x) + g(x)us (t) = Q sgn(s) Kh(s) (18)
In order to incorporate the sliding mode control This equation is solved for the sliding mode control
technique into the fuzzy-model-based control law, we law, yielding
Grst deGne a time-varying surface S(t), in the state
space by imposing the scalar equation s(x; t) = 0, S
us (t) = g(x)1 (F(x) + Q sgn(s) + Kh(s)): (19)
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 7

The sliding condition (15) is satisGed: where

d 2 F(x(t)) = f(x(t)) g(x(t))K()x(t):
s = 2sQ sgn(s) 2sKh(s) 2Q|s|0:
dt
In order to proceed, we have to make the following
Therefore, the closed-loop fuzzy control system (12)
assumption.
is asymptotically stable. The results are summarized
in the following theorem.
Assumption 1. There exist functions fU (x); gU (x),
Theorem 2. If the TS fuzzy model described in (2) is and gL (x) such that |f(x)|6fU (x) and 0gL (x)6
locally controllable; then the closed-loop fuzzy system g(x)6gU (x).
described in (12); with control law (19); is asymptot-
ically stable about the origin. Based on fU (x); gU (x), and gL (x), and using (20),
the upper bound functions of F(x) can be easily ob-
Note that the local controllability condition must be tained as
required for the design of local compensators in each F U (x) = fU + gU |K()x|:
rule; without this condition, tracking may not always
be possible. Let x = x xd be the tracking error, and let

3.2. Case 2 x = x xd = [x x(n1) ]T : (21)

Let us deGne the sliding surface as in (14). Di:er-

In this case, we assume that the nonlinear functions
entiating s(x; t) with respect to time, we obtain
f(x) and g(x) are not exactly known. More speciG-
cally, the nonlinear function f(x) in (1) is not exactly S
s = F(x) + g(x)us ;
known, but the extent of the imprecision of f(x) is
upper-bounded by a known continuous function of x; S
F(x) (22)
= F(x) xd(n) + a1 x(n1) + + an1 x:
similarly, the control gain g(x) in (1) is not exactly
known, but is positive and is bounded by a known Since F(x) and g(x) are unknown, only their
continuous function of x. bounds can be used to construct u. In this case, the
The controller rules and defuzziGed output of the supervisory control law us (t) is chosen to be
controller are the same as Case 1, rewritten below.
us (t) = gL1 {Q sgn(s) + Ks}; (23)
Controller Rule i: where
IF x(t) is F1i and and x (n1)
is Fni ;
Q = [F U + |xd(n) a1 x(n1) an1 x|]:
THEN u(t) = Ki x(t) + us (t)
Substituting (23) into (22), we have
(i = 1; 2; : : : ; r):
s = FS {ggL1 [F U + |xd(n) a1 x(n1)
r
i=1 wi (x(t))(Ki x(t) + us (t)) sgn(s) gg1 Ks;
an1 x|]} (24)
u(t) = r L
i=1 wi (x(t))
r
 ss = sFS s{ggL1 [F U + |xd(n) a1 x(n1)
= i (x(t))Ki x(t) + us (t):
i=1 sgn(s) sgg1 Ks
an1 x|]} L
The closed-loop system is obtained from the feed- S |s|{ggL1 [F U
6 s2 ggL1 K + |s||F|
back interconnection of the nonlinear system (1) and
the controller (11), resulting in the following equation:
+|xd(n) a1 x(n1) an1 x|]}
x(t)(n) = F(x(t)) + g(x(t))us (t); (20) 6 s2 ggL1 K 6 0: (25)
8 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

control theory is seamlessly integrated into the design

of the proposed fuzzy controller, so that the stability
of the overall system is guaranteed without the need
to Gnd a common positive deGnite matrix to satisfy
the Lyapunov equation. Finally, the proposed method
Fig. 1. Controller structure. guarantees the stability of the original underlying non-
linear systems thanks to the sliding mode control al-
gorithm which is very robust to system parameter
Therefore, the closed-loop fuzzy control system uncertainty and nonlinearity. Fig. 1 shows the con-
(20) is asymptotically stable. Fig. 1 shows the con- troller structure.
troller structure. The results are summarized in the The controller design procedure is illustrated as
following theorem. follows:
Theorem 3. If the TS fuzzy model described in (2) 1. Determine the fuzzy model of the given nonlinear
is locally controllable; then the closed-loop control system. If the exact or nominal dynamic equa-
system described in (20); with control law (23); is tions are given then construct the fuzzy model by
asymptotically stable about the origin. linearizing the nonlinear equations over various
operating points of the system [28,25,6,15]. If one
As in the previous theorem, the local controllabil- can only obtain the inputoutput data then the
ity condition must be required for the design of lo- fuzzy model can be also constructed by various
cal compensators in each rule; without this condition, fuzzy modeling methods available in the literature
tracking may not always be possible. [24,2,16 18,5,14]
Note also that although the proposed control scheme 2. Design the state feedback controller for each linear
is able to guarantee the stability of the overall system, state model of the obtained fuzzy model.
it su:ers from the chattering problem because of the 3. Design an appropriate sliding surface.
switching function in the control law. The chattering 4. If the exact nonlinear dynamic equations are known
problem is inherent in sliding mode control but can be then use the control input (19) in Case 1 otherwise
eliminated by replacing the switching function with use the control input (23) in Case 2.
a saturation function. The resulting control law is as
follows:
4. Simulation examples
us = gL1 {Q sat(s=) + Ks};
Two numerical examples are included here to
where  is a boundary constant. Since the allevi-
show the e:ectiveness and feasibility of the proposed
ation of chattering is achieved at the cost of the
method. One is the control of the DuJng-forced os-
tracking precision, the boundary constant  must
cillation system, and the other is the control of the
be carefully chosen. Outside the boundary, the su-
well-known inverted pendulum.
pervisory control law us (t) is identical to (23)
which pushes the trajectory back to inside of the
boundary [23]. 4.1. Example 1: Du8ng forced-oscillation system
Although the exact nonlinear dynamic equations of
the given system are not precisely known, one can ob- In this example, we consider the DuJng oscillation
tain a fuzzy model by determining nominal dynamic system [9]:
equations of the given system. The advantages of the
proposed fuzzy controller are as follows: Firstly, the x1 = x2 ;
proposed fuzzy controller can deal with the tracking x2 = 0:1x2 x13 + 12 cos(t) + u(t): (26)
problem while the other TS-model-based fuzzy con-
trol methods are applied only to the regulation prob- This system is chaotic without control. The trajectory
lem for nonlinear systems. Secondly, the sliding mode of the system with u(t) = 0 is shown in the phase
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 9

Fig. 2. Trajectory of the chaotic DuJng system in the phase plane, with u(t) = 0 and x(0) = [2 2]T .

plane of (x1 ; x2 ) in Fig. 2 for the initial conditions Rule 3: IF x1 is about 4;

x(0) = [2 2]T and the time period from t0 = 0 to THEN x = A3 x + B3 u;
tf = 20. We used the proposed controller to control
the system state x1 to track the reference trajectory where
ym (t) = sin(t). In the phase plane, this reference  
0 1 0
trajectory is the unit circle. A1 = ; B1 = ;
0 0:1 1
The fuzzy model for this chaotic system is obtained
 
by linearizing the nonlinear equations over a number 0 1 0
of operating points in the phase plane. As shown in A2 = ; B2 = ;
12 0:1 1
Fig. 2, x1 is between 4 and 4 for all t (t0 6t6tf ).  
In the manual modeling procedure, x2 is not included 0 1 0
A3 = ; B3 = :
in the fuzzy rule-base since x2 does not appear in the 48 0:1 1
linearized equations. Therefore, we linearize the non-
linear system (26) on x1 = {4; 2; 0; 2; 4} which re- The membership functions for x1 are chosen as shown
sults in Gve fuzzy rules. However, since the fuzzy rules in Fig. 3. The rules of the fuzzy-model-based con-
for x1 = {2; 2}, and x1 = {4; 4} have the same con- troller are:
sequent parts, we Gnally obtain the following fuzzy
rules. Controller rules:
Rule 1: IF x1 is about 0; THEN u = K1 x + us ;
Plant rules: Rule 2: IF x1 is about 2;
Rule 1: IF x1 is about 0; THEN x = A1 x + B1 u; THEN u = K2 x + us ;
Rule 2: IF x1 is about 2; Rule 3: IF x1 is about 4;
THEN x = A2 x + B2 u; THEN u = K3 x + us :
10 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

Fig. 3. Membership functions for state x1 .

Fig. 4. Comparison of output responses.

4.1.1. Case 1 tional fuzzy-model-based controller and the proposed

The desired closed-loop poles for each local model controller are the same. The result is illustrated in
are chosen as 2 and 2. Thus, the feedback control Fig. 4. As can be seen from Fig. 4, the proposed
gains are found as method can stabilize the chaotic system success-
fully, while the conventional fuzzy-model-based
K1 = (4 3:9); K2 = (8 3:9); K3 = (44 3:9): controller fails to stabilize it. That is, the conven-
tional fuzzy-model-based controller may be unstable
The parameters for the sliding mode controller us (t) even though each local linear state space model
are determined by experiments and  = 5; Q = 0:1; is stable. Fig. 5 shows the simulation results for
K = 10, and h(s) = s. For comparison purpose, we the above initial conditions: Fig. 5(a) shows the
applied both the conventional fuzzy-model-based state x1 (t) and its desired trajectory ym (t) = sin(t),
controller and the proposed controller to system Fig. 5(b) shows the state x2 (t) and its desired
(26). In the simulation, parameters of the conven- trajectory ym (t) = cos(t), and Fig. 6 shows the
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 11

Fig. 5. (a) Comparison of state x1 (t) (solid line) and its desired trajectory x1d = sin(t) (dashed line) for the initial conditions x(0) = [2 2]T .
(b) Comparison of state x2 (t) (solid line) and its desired trajectory x2d = cos(t) (dashed line) for the initial conditions x(0) = [2 2]T .
12 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

Fig. 6. Closed-loop controlled system trajectory (x1 (t); x2 (t)) using the proposed controller for the chaotic DuJng system.

Fig. 7. Comparison of output responses.

 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 13

Fig. 8. Closed-loop controlled system trajectories of (a) x1 (t) and (b) x2 (t).

closed-loop trajectory controlled using the proposed 4.1.2. Case 2

fuzzy controller. It can be seen that a quite satisfac- In this case, we assume that the nonlinear equations
tory control result is obtained by using the proposed of the DuJing-forced oscillation system are not ex-
controller. actly known. That is, dynamic equations in (26) are
14 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

Fig. 9. Closed-loop controlled system trajectory (x1 (t); x2 (t)) using the proposed controller for the chaotic DuJng system.

assumed to be nominal system equations of the given the regulation problem is given in Fig. 7. The simula-
DuJng-forced oscillation system. Therefore, for the tion results are shown in Fig. 8, where Fig. 8(a) shows
design of the proposed controller we only use upper the state x1 (t) and its desired trajectory ym (t) = sin(t)
and lower bounds of the nonlinear systems. Nominal and Fig. 8(b) shows the state x2 (t) and its desired tra-
nonlinear equations are used only for the construction jectory ym (t) = cos(t). Fig. 9 shows the phase-plane
of the fuzzy model. The obtained fuzzy model is the trajectory of the controlled DuJng system. The con-
same as Case 1. trol results show chattering phenomenon since the
The desired closed-loop poles for each local model proposed method employs switching function in the
are chosen as 2 and 2. Thus the control gains are control algorithm. As can be seen from the Ggures,
found as the proposed controller can achieve both regulation
and tracking for the given nonlinear system in the
K1 = (4 3:9); K2 = (8 3:9); K3 = (44 3:9): presence of uncertainty.

Finally, we choose K = 10,  = 5, which are same 4.2. Inverted pendulum

as Case 1. In addition, the bounds fU ; gU , and gL are
selected as follows: Consider the problem of balancing an inverted pen-
dulum on a cart, as shown in Fig. 10. The dynamic
|f(x1 ; x2 )| = | 0:1x2 x13 + 12 cos(t)| equations of the pendulum are [29]

6 0:1|x2 | + |x1 |3 + 12 fU (x1 ; x2 ) (27) x1 = x2 ;

and gU = gL = 1 since the control input u(t) is directly g sin(x1 ) amlx22 sin(2x1 )=2 a cos(x1 )u
x2 = ;
applied to system (26). A comparison of the proposed 4l=3 aml cos2 (x1 )
controller and the conventional fuzzy controller for (28)
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 15

Fig. 10. The inverted pendulum system. Fig. 11. Membership functions.

where x1 is the angle in rad of the pendulum 0

from the vertical axis, x2 is the angular velocity B2 = a#
in rad=s1 ; g = 9:8 m=s2 is the acceleration due to 2
4l=3 aml#
gravity, m = 2:0 kg is the mass of the pendulum,
a = (m + M )1 ; M = 8:0 kg is the mass of the cart, with # = cos(88 ). The membership functions for
2l = 1:0 m is the length of the pendulum, and u is the Rules 1 and 2 are shown in Fig. 11.
force applied to the cart.
Since the dynamic equations of the nonlinear sys-
tem is available, a TS fuzzy model used to approxi- 4.2.1. Case 1
mate this system can be manually calculated by simple The rules of the designed fuzzy-model-based con-
linearization method. The resulting TS fuzzy model is troller are:
[20]:
Controller rules:
Plant rules:
Rule 1: IF x1 is about 0; THEN u = K1 x + us ;
Rule 1: IF x1 is about 0; THEN x = A1 x + B1 u; Rule 2: IF x1 is about "=2;
Rule 2: IF x1 is about "=2; THEN u = K2 x + us :
THEN x = A2 x + B2 u;
In order to determine the parameters of the controller,
we simply choose the closed-loop eigenvalues 2 and
where 2 for both A1 B1 K1 and A2 B2 K2 . Then, we
obtain
0 1 0
A1 = g ; B1 = a ; K1 = (120:6667 22:6667);
0
4l=3 aml 4l=3 aml
K2 = (2551:6 764:0):

0 1
The parameters for the sliding mode controller us (t)
A2 = 2g ;
0 are  = 5; Q = 0:1; K = 10; h(s) = s, which are de-
"(4l=3 aml#2 ) termined experimentally. Using these parameters,
16 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

Fig. 12. Comparison of output responses.

we can construct the proposed stable fuzzy logic trajectory x2d = " cot(t)=30 (dashed line), for initial
controller. conditions x(0) = ["=60 0]T . Fig. 14(a) shows the
In order to compare the stabilization performance tracking results, of the state x1 (t) (solid line) and
of the proposed controller to that of the conventional its desired trajectory x1d = " sin(t)=30 (dashed line),
fuzzy-model-based controller, we applied the two con- for initial conditions x(0) = [89"=180 0]T . Fig. 14(b)
trollers to the given system (27). Fig. 12 shows the shows the tracking results, of the state x2 (t) (solid line)
responses of the pendulum system using the conven- and its desired trajectory x2d = " cos(t)=30 (dashed
tional fuzzy-model-based controller and the proposed line), for initial conditions x(0) = [89"=180 0]T .
method, for initial conditions 15 (0:2618 rad). As
can be seen from this Ggure, the response of the pro-
posed method is faster than that of the conventional 4.2.2. Case 2
one. In this case, we assume that the nonlinear equa-
Since the conventional fuzzy-model-based control tions of the inverted pendulum system are not exactly
algorithm only focuses on the regulation problem of known. That is, dynamic equations in (27) are as-
the given nonlinear system, we cannot compare it sumed to be nominal system equations of the given
to the proposed controller. We only show the track- system. Therefore, for the design of proposed fuzzy
ing performance of the proposed controller alone. We controller we only use upper and lower bounds of the
choose the reference signal x1d = " sin(t)=30 in this nonlinear systems as in Example 1. Nominal nonlin-
simulation. Fig. 13(a) shows the tracking results, of ear equations are only used for the construction of the
the state x1 (t) (solid line) and its desired trajectory fuzzy model. The obtained fuzzy model is the same
x1d = " sin(t)=30 (dashed line), for initial conditions as Case 1.
x(0) = ["=60 0]T . Fig. 13(b) shows the tracking re- To apply the proposed control method to this sys-
sults, of the state x2 (t) (solid line) and its desired tem, we Grst need to determine the bounds fU ; gU ,
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 17

Fig. 13. (a) Comparison of state x1 (t) (solid line) and its desired trajectory x1d = " sin(t)=30 (dashed line) for the initial conditions
x(0) = ["=60 0]T . (b) Comparison of state x2 (t) (solid line) and its desired trajectory x2d = " cos(t)=30 (dashed line) for the initial
conditions x(0) = ["=60 0]T .

and gL . For this system, we have 9:8 + 0:005x22

6 = 17:29 + 0:0882x22
  2=3 0:1
 g sin(x1 ) amlx22 sin(2x1 )=2 

|f(x1 ; x2 )| =  
4l=3 aml cos2 (x1 )  fU (x1 ; x2 ) (29)
18 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

Fig. 14. (a) The state x1 (t) (solid line) and its desired trajectory x1d = " sin(t)=30 (dashed line) for the initial condition x(0) = [89"=180 0]T .
(b) The state x2 (t) (solid line) and its desired trajectory x2d = " cos(t)=30 (dashed line) for the initial condition x(0) = [89"=180 0]T .

and If we require that |x1 |6"=6, then

   
 1 
 a cos(x1 )   10 cos "=6 
|g(x1 ; x2 )| =   6 0:1765 |g(x1 ; x2 )|  40:5 
4l=3 aml cos x1 
2  3 10 1
(2)(0:5) cos2 "=6 
gU (x1 ; x2 ): (30) = 0:1464 gL : (31)
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 19

Fig. 15. Output responses of the conventional fuzzy-model-based controller and the proposed controller.

In order to determine the parameters of the conven- the dotted line shows that of the conventional fuzzy-
tional fuzzy-model-based control, we simply choose model-based controller. The proposed method can
the closed-loop eigenvalues 2 and 2 for A1 B1 K1 stabilize the inverted pendulum system for all given
and A2 B2 K2 . Then, we obtain initial conditions, while the linear one fails to balance
the pendulum for x1 = 45 . Compared to the conven-
K1 = (120:6667 22:6667); tional fuzzy-model-based control, the new controller
K2 = (2551:6 764:0): shows faster responses.
Figs. 16 and 17 show the simulation results for ini-
The conventional fuzzy-model-based controller, tial conditions x(0) = ["=60 0]T and x0 = ["=60 0]T ,
however, cannot guarantee the stability of the closed- respectively, where Figs. 16(a) and 17(a) show the
loop system and cannot deal with the tracking prob- state x1 (t) (solid line) and its desired trajectory
lem. In the proposed controller, the parameters for the x1d = " sin(t)=30 (dashed line), Figs. 16(b) and 17(b)
sliding mode control are same as Case 1, i.e.,  = 5 show the state x2 (t) (solid line) and its desired trajec-
and K = 10. In this example, we employ the bound- tory x2d = " cos(t)=30 (dashed line), and Figs. 16(c)
ary concept to eliminate the chattering problem. The and 17(c) show the control u(t). It is clear that the
boundary constant is  = 1. Using these parameters, proposed control method has good tracking perfor-
we can construct the proposed stable fuzzy logic mance under parameter uncertainties within various
controller. operating regions. The simulation results also show
Fig. 15 shows the responses of the pendulum sys- that the proposed controller can stabilize the nonlin-
tem under conventional fuzzy-model-based control, ear system although the design procedure is based on
and the proposed controllers for initial conditions the TS fuzzy model instead of the original nonlinear
45 (0:7854 rad). The solid line indicates the re- plant. Furthermore, unlike the conventional fuzzy-
sponse of the proposed controller, the dashdot line model-based control techniques, it has very good
shows that of the conventional linear controller, and tracking performance.
20 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

Fig. 16. (a) The state x1 (t) (solid line) and its desired trajectory x1d (t) = " sin(t)=30 (dashed line) for the initial condition x(0) = ["=60 0]T .
(b) The state x2 (t) (solid line) and its desired trajectory x2d = " cos(t)=30 (dashed line) for the initial condition x(0) = ["=60 0]T .

5. Conclusion with or without uncertainties. In the design proce-

dure, we represent the fuzzy system by a family of
In this paper, we have developed a robust fuzzy- local state-space models, and construct a global fuzzy
model-based controller for SISO nonlinear systems logic controller by blending all local state feedback
 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122 21

Fig. 17. (a) The state x1 (t) (solid line) and its desired trajectory x1d (t) = " sin(t)=30 (dashed line) for the initial condition
x(0) = [89"=180 0]T . (b) The state x2 (t) (solid line) and its desired trajectory x2d = " cos(t)=30 (dashed line) for the initial condition
x(0) = [89"=60 0]T .

controllers together with a sliding mode controller. mon positive deGnite matrix. (2) Stability is guar-
The proposed method has the following advantages: anteed for the original underlying nonlinear system.
(1) Stability is guaranteed without Gnding a com- (3) Tracking of desired states is easily achieved.
22 W. Chang et al. / Fuzzy Sets and Systems 125 (2002) 122

Finally, simulations have been performed to control [15] Y.H. Joo, G. Chen, L.S. Shieh, Hybrid state-space fuzzy
the uncertain chaotic DuJng-forced oscillator and model-based controller with dual-rate sampling for digital
inverted pendulum, to show the e:ectiveness and control of chaotic systems, IEEE Trans. Fuzzy Systems 7
(1999) 394408.
feasibility of the proposed fuzzy control method. [16] Y.H. Joo, H.S. Hwang, K.B. Kim, K.B. Woo, Linguistic
model identiGcation for fuzzy system, Electron. Lett. 31 (4)
(1995) 330331.
References [17] Y.H. Joo, H.S. Hwang, K.B. Kim, K.B. Woo, Fuzzy system
modeling and its application to mobile robot control, in: Z.
[1] S.G. Cao, N.W. Rees, G. Feng, Stability analysis and design Bien (Ed.), Information Sciences, and Intelligent Systems,
for a class of continuous-time fuzzy control systems, Internat. ch. Fuzzy Logic and its Applications, Kluwer Academic
J. Control 64 (6) (1996) 10691087. Publisher, Dordrecht, 1995, pp. 147156.
[2] S.G. Cao, N.W. Rees, G. Feng, Analysis and design for a [18] Y.H. Joo, H.S. Hwang, K.B. Kim, K.B. Woo, Fuzzy system
class of complex control systems part I: Fuzzy modelling and modeling by fuzzy partition and GA hybrid schemes, Fuzzy
identiGcation, Automatica 33 (6) (1997) 10171028. Sets and Systems 86 (3) (1997) 279288.
[3] S.G. Cao, N.W. Rees, G. Feng, Analysis and design for a [19] S.W. Kim, E.T. Kim, M. Park, A new adaptive fuzzy
class of complex control systems part II: Fuzzy controller controller using the parallel structure of fuzzy controller
design, Automatica 33 (6) (1997) 10291039. and its application, Fuzzy Sets and Systems 81 (2) (1996)
[4] B. Carse, T.C. Fogarty, A. Munro, Evolving fuzzy rule based 205226.
controller using genetic algorithm, Fuzzy Sets and Systems [20] F.H.F. Leung, L.K. Wong, P.K.S. Tam, Fuzzy model based
80 (1996) 273293. controller for an inverted pendulum, Electron. Lett. 32 (8)
[5] C.K. Chak, G. Feng, J. Ma, An adaptive fuzzy-neural network (1996) 16831685.
for MIMO system model approximation in high-dimensional [21] X.J. Ma, Z.Q. Sun, Y.Y. He, Analysis and design of fuzzy
spaces, IEEE Trans. Systems Man Cybernet.Part B: controller and fuzzy observer, IEEE Trans. Fuzzy Systems 6
Cybernet. 28 (3) (1998) 436446. (1) (1998) 4151.
[6] W. Chang, Y.H. Joo, J.B. Park, G. Chen, Design of [22] E.H. Mamdani, Application of fuzzy algorithm for control
robust fuzzy-model-based controller with application to of simple dynamic plant, Proc. IEEE 12 (1221) (1974)
inverted pendulum tracking control, ITC-CSCC98, 1998, 15851888.
pp. 16211624. [23] J.J.E. Slotine, W. Li, Applied Nonlinear Control,
[7] C.L. Chen, M.H. Chang, Optimal design of fuzzy Prentice-Hall, Englewood Cli:s, NJ, 1991.
sliding-mode control: a comparative study, Fuzzy Sets and [24] T. Takagi, M. Sugeno, Fuzzy identiGcation of systems and its
Systems 93 (1998) 3748. applications to modeling and control, IEEE Trans. Systems
[8] C. Li, P.C. Chen, C.K. Chen, Analysis and design of Man Cybernet. SMC-15 (1) (1985) 116132.
fuzzy control system, Fuzzy Sets and Systems 57 (1993) [25] K. Tanaka, T. Kosaki, Design of a stable fuzzy controller for
125140. an articulated vehicle, IEEE Trans. Systems Man Cybernet.
[9] G. Chen, X. Dong, On feedback control of chaotic Part B: Cybernet. 27 (3) (1997) 552558.
continuous-time systems, IEEE Trans. Circuits SystemsI: [26] L.X. Wang, A supervisory controller for fuzzy control systems
Fund. Theory Appl. 40 (9) (1993) 591601. that guarantees stability, IEEE Trans. Automat. Control 39
[10] G. Feng, S.G. Cao, N.W. Rees, C.K. Chak, Design of fuzzy (9) (1994) 18451847.
control systems with guaranteed stability, Fuzzy Sets and [27] L.X. Wang, Adaptive Fuzzy Systems and ControlDesign
Systems 85 (1997) 110. and Stability Analysis, Prentice-Hall, Englewood Cli:s, NJ,
[11] W. Gao, J.C. Hung, Variable structure control of nonlinear 1994.
systems: a new approach, IEEE Trans. Ind. Electron. 40 (1) [28] H.O. Wang, K. Tanaka, M.F. GriJn, An approach to fuzzy
(1993) 4555. control of nonlinear systems: Stability and design issues,
[12] Y.C. Hsu, G. Chen, Fuzzy dynamical modeling and control IEEE Trans. Fuzzy Systems 4 (1) (1996) 1423.
techniques for nonlinear processes with applications to [29] L.K. Wong, F.F. Leung, P.K.S. Tam, Stability design of
multiple-input multiple-output systems, in: C.T. Leondes ts model based fuzzy systems, FUZZ-IEEE1997, Barcelona,
(Ed.), Fuzzy Theory: Systems Technology and Applications, Spain, 1997, pp. 8386.
Academic Press, New York, 1999, Vol. 1, pp. 4768. [30] J.C. Wu, T.S. Liu, A sliding-mode approach to fuzzy control
[13] J.Y. Hung, W. Gao, J.C. Hung, Variable structure control: a design, IEEE Trans. Control Systems Technol. 4 (2) (1996)
survey, IEEE Trans. Ind. Electron. 40 (1) (1993) 222. 141151.
[14] J.S.R. Jang, ANFIS: Adaptive-network-based fuzzy inference [31] L.A. Zadeh, Fuzzy sets, Inform. and Control 8 (1965)
systems, IEEE Trans. Systems Man Cybernet. 23 (3) (1993) 338353.
665686.