14 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 4, NO.

1, FEBRUARY 1996

n to Fuzzy Control of
s: Stability and Design
Hua 0. Wang, Member, IEEE, Kazuo Tanaka, Member, IEEE, and Michael F. Griffin, Member, IEEE

Abstract-We present a design methodology for stabilization of feedback control is designed. The resulting overall controller,
a class of nonlinear systems. First, we represent a nonlinear plant which is nonlinear in general, is again a fuzzy blending of
with a Takagi-Sugeno fuzzy model. Then a model-based fuzzy each individual linear controller.
controller design utilizing the concept of the so-called parallel
distributed compensation is employed. The main idea of the ThIs paper deals with stability and design issues in the pro-
controller design is to derive each control rule so as to compensate posed fuzzy control approach of nonlinear systems. Stability
each rule of a fuzzy system. The design procedure is conceptually conditions of both fuzzy models and fuzzy control systems
simple and natural. Moreover, the stability analysis and control are given. The design procedure aims at rendering stable
design problems can be reduced to linear matrix inequality fuzzy controllers. More significantly, the stability analysis and
(LMI) problems. Therefore, they can be solved efficiently in
practice by convex programming techniques for LMIs. The control design problems are reduced to linear matrix inequality
design methodology is illustrated by application to the problem (LMI) problems [ 151. Numerically, the LMI problems can be
of balancing and swing-up of an inverted pendulum on a cart. solved very efficiently by means of some of the most powerful
tools available to date in the mathematical programming
literature. Therefore, recasting the stability analysis and control
I. INTRODUCTION
design problems as LMI problems is equivalent to finding
E HAVE witnessed rapidly growing interest in fuzzy solutions to the original problems. The recasting of stability
control in recent years. There have been many success- analysis and design of fuzzy control systems to LMI problems
ful applications. Despite the success, it has become evident was first made in [ l l ] and [12].
that many basic issues remain to be further addressed. Stabil- For illustration the design methodology is applied to the
ity analysis and systematic design are certainly among the problem of balancing and swing-up of an inverted pendulum
most important issues for fuzzy control systems. Recently, on a cart.
there have been significant research efforts on these issues The paper is organized as follows: The main results are
[1]-[6]. This paper attempts to present a systematic design presented in Section 11. Stability analysis of Takagi-Sugeno
methodology for fuzzy control of a class of nonlinear systems. fuzzy models is presented in Section 11-A. In Section 11-B, we
There are several approaches to control of a nonlinear consider the control design problems via parallel distributed
system. A typical approach is the feedback stabilization of compensations. Section 11-C contains an introduction to LMIs
nonlinear systems where a linear feedback control is designed as well as results and discussions on the LMI approach to
for the linearization of the system about a nominal operating the stability analysis and design of fuzzy control systems.
point. This approach, however, generally only renders a local In Section HI, the design methodology is illustrated via a
result. Other approaches [ 141 such as feedback linearization detailed example, namely the balancing and swing-up of an
are rather involved and tend to result in rather complicated inverted pendulum on a car. Concluding remarks are collected
controllers. in Section IV.
In this paper, we consider a nonlocal approach which
is conceptually simple and straightforward. Linear feedback 11. STABILITY ANALYSIS, PARALLEL DISTRIBUTED
control techniques can be utilized as in the case of feedback COMPENSATION AND LINEARMATRIXINEQUALITIES
stabilization. The procedure is as follows: First the nonlinear
In this paper, results for discrete-time systems only are pre-
plant is represented by a Takagi-Sugeno type fuzzy model.
sented for brevitys sake. The results, however, also hold for
In this type of fuzzy model, local dynamics in different state
continuous-time systems subject to some minor modifications.
space regions are represented by linear models. The overall
To illustrate the design procedure, we apply the results to the
model of the system is achieved by fuzzy blending of these
problem of balancing and swing-up of an inverted pendulum
linear models. The control design i s carried out based on the
on a cart, which is a continuous-time system.
fuzzy model via the so-called parallel distributed compensation
In the proposed design procedure, to begin with we represent
scheme. The idea is that for each local linear model, a linear
a given nonlinear plant by the so-called Takagi-Sugeno fuzzy
Manuscript received January 1995; revised April 28, 1995. model [7]. This fuzzy modeling method is simple and natural.
H. 0. Wang and M. F. Griffin are with the United Technologies Research The system dynamics is captured by a set of fuzzy implications
Center, East Hartford, CT 06108 USA. which characterize local relations in the state space. The main
K. Tanaka is with the Department of Mechanical Systems Engineering,
Kanazawa University, 2-40-20 Kodatsuno, Kanazawa 920 Japan. feature of a Takagi-Sugeno fuzzy model is to express the local
Publisher Item Identifier S 1063-6706(96)00639-X. dynamics of each fuzzy implication (rule) by a linear system
1063-6706/96$05.00 0 1996 IEEE
WANG et al.: APPROACH TO FUZZY CONTROL OF NONLINEAR SYSTEMS 15

model. The overall fuzzy model of the system is achieved by

fuzzy blending of the linear system models.
Mi
1
Specifically, the Takagi-Sugeno fuzzy system is described
by fuzzy IF-THEN rules, which locally represent linear input-
output relations of a system. The fuzzy system is of the
following form:
Rule i: IF z l ( b ) is Mal . . and zn(k)is Mzn
+
THEN z ( k 1) = A,z(lc) B,u(k) + 0
where -a a
z T ( k ) = [.l(IC>,.Z(k),...,.n(IC)I Fig. 1. Membership functions of Example 1.

UT@) = [ul(~),uz(IC),...,um(~)]
the system (2) is also said to be quadratically stable and the
i = 1 , 2 , . . . ,r and r is the number of IF-THEN rules. Mij V is called a quadratic Lyapunov function. Theorem 1 thus
+
are fuzzy sets, and z(L 1) = Aiz(k) B;u(k)is the output + presents a sufficient condition for quadratic stability of system
from the ith IF-THEN rule. Given a pair of ( z ( k ) ,u(k)),the (2).
final output of the fuzzy system is inferred as follows: To check the stability of fuzzy system (2), it has long been
r recognized there is a lack of a systematic procedure to find
C w i ( I C ) {Aiz(k)+ IC) 1 a common positive definite matrix P . Most of the time a
trial-and-error type of procedure has been used [2]. In [IO], a
z(k + 1) = i=l r (1)
procedure to construct a common P is given for second-order
fuzzy systems, i.e., the dimension of state n = 2.
a= I
In this paper, we point out that the common P problem
where can be solved efficiently via convex optimization techniques
n for LMIs [15]. To do this a very important observation is
wz(k) = J-J Ma,(.,(k)). that the stability condition of Theorem 1 is expressed in
,=1 LMIs. To check stability we need to find a common P or
is the grade of membership of x 3 ( k ) in Ma,
Ma3(x3(IC)) determine that no such P exists. This is an LMI problem (see
The open-loop system of (1) is Section 11-C for details on LMIs and the related LMI approach
to stability analysis and design of fuzzy control systems).
r
Numerically, the LMI problems can be solved very efficiently
by means of some of the most powerful tools available to date
in the mathematical programming literature. For instance, the
recently developed interior-point methods [ 161 are extremely
i=l efficient in practice.
For system (2) a question naturally arises is whether system
where it is assumed that
r
(2) is stable if all its subsystems are stable, i.e., all Ats
are stable. The answer is no in general as illustrated by the
following example.
Example 1: Consider the fuzzy system:

for all k . Each linear component Aiz(k)is called a subsystem. Rule 1: IF m ( k ) is MI (e.g. Small)
+
THEN ~ ( k1) = Alz(IC),
A. Stability Analysis via Lyapunov Approach Rule 2: IF 22(k) is MZ (e.g. Big)
A sufficient stability condition derived by Tanaka and THEN Z(IC + 1) = A ~ zIC)(
Sugeno [2], for ensuring stability of (2) is given as follows: where %(IC) = \z1(IC) z2(k)IT and
Theorem I, 121: The equilibrium of a fuzzy system (2) is
asymptotically stable in the large if there exists a common
positive definite matrix P such that L J L

ATPA% - P < 0, Fig. 1 shows the membership functions of M I and M2.

i = 1 , 2 , . . . ,r (3)
Since A1 and A2 are stable, the linear subsystems are
i.e., a common P has to exist for all subsystems. stable. However, for some initial conditions the fuzzy system
This theorem reduces to the Lyapunov stability theorem for can be unstable as shown in Fig. 2 for the initial condition
(discrete-time) linear systems when r = 1. = [0.90 -0.70IT. It should be noted that the linearization
The stability condition of Theorem 1 is derived using a of the fuzzy system around 0 is stable (which implies that
quadratic function V(z) = zTPz. If there exists a P > 0 the fuzzy system is locally stable). Obviously there does not
such that the V = z T P z proves the stability of system (2), exist a common P > 0 since the fuzzy system is unstable.
16 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 4, NO. 1, FEBRUARY 1996

61 r 1
I 1

-2.5 -zl-2.5 -2.5 -xz- 2.5 -2.5 -x,- 2.5 -2.5 -x2- 2.5
a=l a = 0.5

(a) (b)

-6
0 2 4 6 8 10 12 14 16 18 20
k

Fig. 2. Response of Example 1 ( a = 1).

This can be shown analytically. Moreover, this can also

be shown numerically by convex optimization algorithms
involving LMIs.
Still an interesting question is for what initial conditions is -2.5 -xl- 2.5 -2.5 -x2- 2.5 -2.5 -x,- 2.5 -2.5 -x2- 2.5

the fuzzy system stable (or unstable). This is determined by a = 0.25 a = 2.0
studying the basin of attraction of the origin. (C) (4
Fig. 3(a) shows the basin of attraction for the case of a = 1. Fig. 3. Basin of attraction for Example 1-(a) a = 1, (b) a = 0.5, (c)
The black area indicates regions of instability (horizontal a = 0.25, and (d) a = 2.0.
axis is 2 1 ) . It is also of interest to consider how the basin
of attraction changes as the membership functions vary. For
system (1). The idea is to design a compensator for each rule
instance, how the basin of attraction would change as a varies
of the fuzzy model. Fig. 4 shows the concept of PDC design.
for this example. Fig. 3(b), (c), and (d) shows the basin of
For each rule, we can use linear control design techniques.
attraction for various values of a. It can be seen that as a
The resulting overall fuzzy controller, which is nonlinear in
decreases (increases) from one the basin of attraction becomes general, is a fuzzy blending of each individual linear controller.
smaller (larger). Therefore, the basin of attraction for the
The fuzzy controller shares the same fuzzy sets with the fuzzy
fuzzy system could be membership function dependent. In the
system (1)
example, when a = 00,the fuzzy system becomes
A I + A2 Rule i: IF ~ ( kis )M I , . . . and ~ ( kis )Mnz
z(k + 1) = ___ 2 4k) THEN = ( k ) = --F,z(k)
which is linear and globally asymptotically stable. where i = 1 , 2 , . - . , T . Hence, the fuzzy controller is
For this specific example, an interesting interpretation can be
Ip

given for the dependence of basin of attraction on membership

functions. As a increases (decreases) the inference process
tends to be fuzzier (crisper). Hence, fuzzier decision
leads to larger basin of attraction while crisper decision leads
to smaller basin of attraction.
As illustrated by the example, we have to take stability
issue into consideration when selecting rules and membership Note that the controller (4) is nonlinear in general.
functions. How to systematically select rules and membership Substituting (4) into (1) we obtain
functions to satisfy prescribed stability properties is an inter- r r
esting topic. In the next subsection, we consider the control
design problems via parallel distributed compensations.

B. Parallel Distributed Compensation

We utilize the concept of parallel distributed compensation
(PDC) [SI and [9] to design fuzzy controllers to stabilize fuzzy When applying Theorem 1, we have the following sufficient
Sugeno mentioned this point in his plenary talk at FUZZ-EEE 92. condition for (quadratic) stability.
WANG et al.: APPROACH TO FUZZY CONTROL OF NONLINEAR SYSTEMS 17

Fuzzy srjtem
Fig. 4. Parallel distributed compensation (PDC) design.
Fuzzy cmtrolier the system (1) is also said to be quadratically stabilizable via
PDC design.
In general, we first design a controller for each rule and
check whether the stability conditions are satisfied. Recall
we can use LMI convex programming techniques to solve
this stability analysis problem. If the stability conditions are
not satisfied, we have to repeat the procedure. Please consult
Section 11-C on how LMIs can be used for directly solving the
control design problem. Here, we note that for some special
cases, the control design problem can be solved analytically.
Assume (A,,B,) are controllable. If B, = B(i =
1 , 2 , . . . , T ) and we choose F, such that

Theorem 2: The equilibrium of a fuzzy control system ( 5 ) where G is a Hunvitz matrix. There exists a P such that
is asymptotically stable in the large if there exists a common
positive definite matrix P such that -P
G~PG 0.

{ A , - B,F,}TP{A,- B,F3}- P < 0, Because of (10) and

for w,(k). w3( k ) # 0, V k , i ,j = 1,2, . . ,T. +

(6)
Gij = G i s j

Note that system ( 5 ) can be also written as
the following theorem follows from Theorem 3.
Theorem 4: In the case of B, = B , i = 1,. . . ,T , the equi-
librium of fuzzy control system ( 5 ) can be made asymptotically
stable in the large by the fuzzy PDC controller (4) where F,
satisfies (10).
We remark that a common G might not always be possible
even if (A,,B , ) are controllable.
Example 2: Consider the fuzzy system:

where
Rule 1: IF 4 k ) is MI
+ +
THEN ~ ( k1) = A l z ( k ) B % ( k ) ,
Rule 2: IF x 2 ( k ) is M2

w= E WZ(k)Wj(k).
+
THEN ~ ( k1) = Azz(k) B u ( ~ ) +
i=l,=l
where A I ,A2 are the same as in Example 1 and
Therefore, we have the following sufficient condition.
Theorem 3: The equilibrium of a fuzzy control system
(5) is asymptotically stable in the large if there exists a
B= [:I.
common positive definite matrix P such that the following Employ the PDC controller (4) and choose the closed-loop
two conditions are satisfied. eigenvalues to be [0.5 0.351, we obtain

{ A , - B,F,}TP{A,- B,Pi} - P < 0 , i = 1 , 2 , . . ,r F1 = [0.15 -0.32501

(8) F2 = [-1.85 -0.32501

and
G ~ P G ,-, P < 0, i<j 5 r. (9)

Remark: The conditions of Theorem 3 are more relaxed

than those of Theorem 2.
A1 - BF1= A2 - BF2 1 G=
0.85 -0.1750
[ I 0 1 .
The control design problem is to select F,(i = 1,2, . . . , r ) The closed loop becomes
such that conditions (8) and (9) in Theorem 3 are satisfied.
Using the notation of quadratic stability, we can also think of z(k + 1) = G z (k )
the control design problem as finding F,s such that the closed-
loop system ( 5 ) is quadratically stable. If there exist such F,s, which is stable since G is stable.
18 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 4, NO. 1, FEBRUARY 1996 ~

Next we consider the more general case. 0.9

Example 3: Consider the fuzzy system:
Rule I: IF ~ c ; ~ is( kM
) I
+
THEN ~ ( k1) = Alic(k) Bla(k), + 0.7

Rule 2: IF z z ( k ) is M z 0.6 -
+
THEN ~ ( k1) = A ~ I G (Baa(k)
~) + 0.5 -
where A I ,Az are the same as in Example 1 and 5 0.4-

B1= [:I; Bz= [-;I 0.3 -

0.2 -
The membership functions of Example 1 ( a = 1) are used in -
0.1
the simulation.
Again choose the closed-loop eigenvalues to be [0.5 0.351, 0-
we have J
6.1
0 2 4 6 8 10 12 14 16 18 20
F1 [0.65 -0.51
F2 = [0.87 -0.111

and
1 NonlinearSystem 1
0.2150
Giz = 0.2400

Note that Gl2 is stable.

-0.9450
0.3050 I

The PDC controller is

Rule 1: IF 1c2(k) is M I
THEN ~ ( k=) -Piic(k),
Rule 2: IF x2(k) is M2

P=
THEN ~ ( k=) - F z ~ ( k ) .

It can be easily shown that if we choose the positive definite

matrix P to be

[ 1.1810
-0.0614
-0.0614
2.3044 1 L
I Parellel-Distributed
Compensation (PDC)

Fuzzy Controller

Fig. 6. Fuzzy model-based fuzzy control design.

the stability conditions (8) and (9) are satisfied. In other words,
the closed-loop fuzzy control system which consists of the
fuzzy model and the PDC controller is globally asymptotically
stable. The P is obtained by utilizing an LMI optimization
algorithm. Fig. 5 illustrates the behavior of the fuzzy control
system for the same initial condition of Fig. 2.
The design procedure represents a systematic framework of
fuzzy modeling and fuzzy control design for a large class of
systems. Fig. 6 illustrates the fuzzy model-based fuzzy control
design methodology. Moreover, the LMI formulation of the
stability analysis and design problems enables very efficient
solutions to these problems in practice. In the next section, we
present an introduction to LMIs as well as the LMI approach
to stability analysis and design of fuzzy control systems.
been found to be extremely efficient in practice. For systems
and control, the importance of LMI optimization stems from
the fact that a wide variety of system and control problems can
be recast as LMI problems [15]. Except for a few special cases
these problems do not have analytical solutions. However, the
main point is that through an LMI framework they can be
efficiently solved numerically in all cases. Therefore recasting
a control problem as an LMI problem is equivalent to finding
a solution to the original problem.
DeJinition 1, (1.51: A linear matrix inequality (LMI) is a
matrix inequality of the form
m

F ( X ) = Fo +C z , F , >0 (11)
C. Stability Analysis and Design Using LMIs 2=1

Recently, a class of numerical optimization problems called where zT = ( 1 c 1 , 2 2 , . . . , z

LMI problems has received significant attention [ 151. These
optimization problems can be solved in polynomial-time and,
hence, are tractable at least in a theoretical sense. The recently
developed interior-point methods [ 161 for these problems have
) is the variable and the symmet-
,
ric matrices P, = FF E W X ni ,= 0, . . . ,m are given. The
inequality symbol > O means that F ( z ) is positive definite.
The LMI (11) is a convex constraint on z,i.e., the set
{zlF(z)> 0} is convex. The LMI (11) can represent a wide
WANG ef al.: APPROACH TO FUZZY CONTROL OF NONLINEAR SYSTEMS 19

variety of convex constraint on 2. In particular, linear inequal- The control design problem is to find a state-feedback
ities, convex quadratic inequalities, matrix norm inequalities, gain F such that the closed-loop system is (quadratically)
and constraints that arise in control theory, such as Lyapunov stable. If such a gain F exists, the system is said to be
and convex quadratic matrix inequalities, can all be cast in the quadratically stabilizable (via linear state feedback). This
form of an LMI. Multiple LMI's F(') > 0, i = 1,. . . , p can be quadratic stabilizability problem can be recast as an LMI
-
expressed as a single LMI diag ( F ( l ) , . , F ( p ) >
+ ) 0. problem.
Very often in the LMI's the variables are matrices, e.g., the The condition (14) is not jointly convex in F and P . Now
Lyapunov inequality multiplying the inequality on the left and right by P e l , and
defining a new variable Q = P - l , we may rewrite (14) as
AP < o
A~P- (12)
where A E Rnxn is given and P = PT is the variable. In
Q { A - B F } T Q - l { A - BF}Q - Q < 0. (15)
this case the LMI will not be written explicitly in the form Define K = FQ so that for Q > 0 we have F = K Q - l .
F ( z )> 0. In addition to saving notation, this may lead to Substituting into (15) yields
more efficient computation [ 151. Of course, the inequality (12)
can be readily put in the form (11): Take Fo = O,F, = Q - {AQ - B K } T Q - l { A Q - B K } > 0. (16)
+
-ATP,A P , where P I , ,P , are a basis for symmetric
n x n matrices. This nonlinear (convex) inequality can now be converted to
LMZprobZems [15]: Given an LMI F ( z )> 0, the LMI LMI form using Schur complements [ 151. The resulting LMI is
Problem is to find zfeassuch that F(zfeas) > 0 or determine
that the LMI is infeasible. This is a convex feasibility problem. >O (17)
As an example, the simultaneous Lyapunov stability condi- [(AQQ B K ) (AQ -Q
tion in Theorem 1 is exactly an LMI problem: Given A, E
in Q and K . Thus, the system (13) is quadratically stabilizable
W n X n , i = 1,. . . , T , we need to find P satisfying the LMI if there exist Q > 0, and K such that the LMI (17) holds. The
P>O, ATPA, - P < o , i = 1,2,...,r state feedback gain is F = K Q - l .
We can easily extend the LMI based control design ap-
or determine that no such P exists. proach to multiple rule ( r > 1) cases of the Takagi-Sugeno
The stability conditions encountered in this paper are ex- fuzzy models. For instance, the quadratic stabilizability of the
pressed in the form of LMI's. This recasting is significant Takagi-Sugeno fuzzy models via a linear state feedback can
in the sense that efficient convex optimization algorithms can be cast as the following LMI problem in Q and K :
be used for stability analysis and control design problems.
The recasting therefore constitutes solutions to the stability
analysis and control design problems in the framework of
Takagi-Sugeno fuzzy model and PDC design. i=l,2;..,r
The design procedure presented in this paper involves
an iterative process. For each rule a controller is designed with the state feedback gain F = K Q - l .
based on consideration of local performance only. Then an The LMI based control design approach has also been
LMI-based stability analysis is carried out to check whether the developed for the control of Takagi-Sugeno fuzzy models via
stability conditions are satisfied. In the case where the stability PDC design.
conditions are not satisfied, the controller for each rule will Some important remarks are in order:
be redesigned. The iterative design procedure has been very Remark 1: The stability conditions presented in this paper,
effective in our experience. However, from the standpoint of not only guarantee stability of fuzzy models and fuzzy control
control design, it is more desirable to be able to directly design systems, they also guarantee stability for related uncertain
a control that ensures the stability of the closed-loop system. linear time-varying systems (LDI's) and nonlinear systems
This is referred to as the control problem in the framework of satisfying some global or local sector conditions. Thus, a
Takagi-Sugeno fuzzy model and PDC design. We claim that controller that works well with the fuzzy model is likely to
the control problem can be recast (hence, solved) using the work well when applied to the real system. This point is
LMI approach. Here, we only briefly state the ideas of the clearly demonstrated by the application in the next section.
LMI approach to the control design problem. Details will be The theoretical details, however, will be discussed elsewhere.
presented elsewhere. Remark 2: The stability analysis and control design results
Consider the case T = 1, i.e., there is only one IF-THEN presented in this paper hold for continuous-time systems as
rule, (1) becomes a linear time invariant system well. Instead of using the Lyapunov inequality for discrete-
z ( k + 1) = h ( k ) + B u ( k ) .
time systems, we should use the Lyapunov inequality for
(13)
continuous-time systems
For a given control gain F , using standard stability theory for
linear time invariant systems or Theorem 2, the system (13) + P A < 0.
is (quadratically) stable if there exist P > 0 such that
In the next section, we apply the PDC approach to a
{ A - B F } T P { A- B F } - P < 0. (14) continuous-time system.
20 LEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 4, NO 1, FEBRUARY 1996

111. APPLICATION:
INVERTED PENDULUM ON A CART Rule 2
To illustrate the PDC approach, consider the problem of
balancing and swing-up of an inverted pendulum on a cart.
The equations of motion for the pendulum are [ 181
31 =x2
g sin (21) - amlx; sin (2x1)/2 - acos ( Q ) U
x2 = (18)
41/3 - am1 cos2 (21)
where q denotes the angle (in radians) of the pendulum from
the vertical, and 52 is the angular velocity. g = 9.8m/s2 is
the gravity constant, m is the mass of the pendulum, M is the
-90 0
mass of the cart, 21 is the length of the pendulum, and U is the Fig. 7. Membership functions of two-rule model.
forced applied to the cart (in Newtons). a = l/(m + M ) . We
choose m = 2.0kg, M = 8.0kg, 21 = 1.0m in the simulations
and
[171.

A. Two-RuleModeling and Control

The control objective of this subsection is to balance the in- Note that G12 is Hunvitz.
verted pendulum for the approximate range 2 1 E (-n/2, ../a). Using an LMI optimization algorithm, we obtain the fol-
To use the PDC approach, we must have a fuzzy model which lowing
represents the dynamics of the nonlinear plant. Therefore,
we first represent the system (18) by a Takagi-Sugeno fuzzy (19)
P = p.6250
0.6250 0.62501
0.2812 '
model. To minimize the design effort and complexity, we try to
use as few rules as possible. Notice that when x1 = f7r/2, the It can be easily shown that the following stability conditions
system is uncontrollable. Hence, we approximate the system are satisfied
by the following two-rule fuzzy model.
Rule 1: IF x1 is about 0
{ A , - B,F,}TP + P { A , - B,F,} < 0, i = 1 , 2 (20)
THEN X = Ala: + B ~ u ,
GT2P + PG12 < 0. (21)
Rule 2: IF 21 is about h / 2 (1x11 < n/2)
THEN X AZX B ~ u + The resulting PDC control law is
where Rule 1: IF x1 is about 0
0 THEN U = - - F I X ,
Rule 2: XF 2 1 is about f7r/2 (1x11 < 7r/2)
THEN U = - F ~ x
41/3 - am1
that is
U = -wlFla: - w ~ F ~ x (22)
1-41/3-aml1 where w1 and w2 are the membership values of Rule 1 and 2,
r 0 11 +
respectively, (w1 w2 = 1). This (nonlinear) control law
guarantees the stability of the fuzzy control system (fuzzy
Ln(41/3 -- amlb2) model + PDC control). To assess the effectiveness of the PDC
controller, we apply the controller to the original system (18).
Simulations indicate the control law can balance the pen-
B2 = dulum for initial conditions 2 1 E [-%" 88"] ( z z = 0). In
41/3 - amlp2 contrast, the linear control alone U = -FIXfails to balance
and p = cos (88"). Membership functions for Rule 1 and 2 the pendulum for initial angles lxll> 45". Fig. 8 shows the
are shown in Fig. 7. response of the pendulum system using linear and fuzzy PDC
Choose the closed-loop eigenvalues [-2, -21 for A1-BIF1 controls for initial conditions x1 = 15", 30", 45O, and 22 = 0.
and A2 - B2F2, we have The solid lines indicate responses with the fuzzy controller.
The dotted lines show those with the linear controller. Fig. 9
F1 = [-120.6667 -22.66671 illustrates the closed-loop behavior of the system with the
8'2 = [-2551.6 -764.01. fuzzy controller for initial conditions x1 = 65", 75", 85", and
x2 = 0.
It follows that
We remark that given the nonlinear plant (18) nonlinear
A1 - BlFl = A2 - B2F2 = G = control laws can be designed to balance the pendulum for
initial angles $1 E (--./Z 7r/2). However, such control laws
WANG et al.: APPROACH TO FUZZY CONTROL OF NONLINEAR SYSTEMS 21

Rule 4

X1
-180 -90 0 90 180 [deg.]
Fig. 10. Membership functions of four-rule model.

[-r,r].In this subsection, we extend the results to the range

of X I E [-r r]except for a thin strip near f r / 2 . Balancing
the pendulum for the angle range of r / 2 < 1x11 5 r is
time (sec) referred to as swing-up control of the pendulum. Recall that
Fig. 8. Angle response using linear and two-rule fuzzy control. for X I = f r / 2 the system is uncontrollable. We add two more
rules (Rule 3 and 4) to the fuzzy model.
Rule 1: IF 2 1 is about 0
THEN X A12 B ~ u , +
Rule 2: IF X I is about f7r-12 (1x11 < r / 2 )
THEN X = A22 B ~ u , +
Rule 3: IF 5 1 is about f r / 2 (1~11> r / 2 )
THEN X = A32 B ~ u , +
Rule 4: IF x1 is about r
THEN X = A42 B ~ u +
where A I ,B1, Az, B2 are the same as above and

time (sec)

Fig. 9. Angle response using two-rule fuzzy control.

often tend to be quite involved. For example, one such control

law is [17] U = k ( x 1 , z ~where
) The membership functions of this four-rule fuzzy model are
4kle2 shown in Fig. 10.
9
k(zl,z2)= - - tan(z1)- -In [sec(z1) tan ( X I ) ]
U 3u
+ Again, choose the closed-loop eigenvalues [-2, -21 for
A3 - B3F3 and A4 - B4F4, we have
(e1 + 4 x 2
+ ele2ml sin ( 5 1 ) + ~- U F3 = [2551.6 764.01
sec ( X I ) - am1 cos ( 2 1 )

and e l , e2 are the specified closed-loop eigenvalues.

I (23)
It follows that
F4 = [22.6667 22.66671.

On the other hand, the PDC design is intuitive and simple. A3 - B3F3 = Aq - B4F4 = G
The resulting controller is simple as well.
and

B. Four-Rule Modeling and Control G34 = -220.5230 --67.4675

l l
Suppose the pendulum on the cart system is built in such
a way that the work space of the pendulum is the full circle Note that GS4is Hunvitz.
22 IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 4, NO.1, FEBRUARY 1996

time (sec)

Fig. 11. Angle response using four-nile fuzzy control Fig. 12. Closer-loop angle response with m changed.

TABLE I
COMPARISONS OF DIFFERENTCONTROL DESIGNS

It can be shown that the P of (19) satisfies the additional

stability conditions

{At - BtFt}TP + P{A, - B,F,}< 0, i = 3 , 4 (24) 0' 0.5 1 1.5 2 2.5

time (sec)
3 3.5 4 4.5 5

+
GT4P PG34< 0. (25) Fig. 13. Closed-loop angle response with M changed.

There is no overlap between membership values w1 and

~ 3 , 2 0 1and w4,w2 and w3, w2 and w4. Hence, only 612 and 180
G34 are needed in stability check.
The PDC controller is 160

Rule 1: IF z1 is about 0 140

THEN U -FIX,
Rule 2: IF 2 1 is about f n / 2 (1x1)< n / 2 ) 120

THEN U = - F ~ x ,
3100
Rule 3: IF 21 is about b r / 2 (1x11 > n/2) W
??.
THEN U = - F ~ x , E 80
Rule 4: IF 51 is about 7r
THEN U = - F ~ x 60

40
that is
2o
U = -wlFlx - w ~ F -~ wx ~ F -~~ x4 8 ' 4 2 . (26)
This control law guarantees stability of the fuzzy control 0' 0.5 I 1.5 2 2.5 3 3.5 4 4.5 5
+
system (four-rule fuzzy model PDC control). This controller time (sec)

is applied to the Original system (18) for evaluation of its Fig. 14. Closed-loop angle response with I changed.
WANG et al.: APPROACH TO FUZZY CONTROL OF NONLINEAR SYSTEMS
performance. Simulation results demonstrate that the controller
(26) is able to balance the pendulum for all initial angles except
when 2 1 is in a thin strip 88 < 1x1I < 94. The size of this thin
strip can be reduced by adding more rules to the model and
controller. Fig. 11 illustrates the response of the closed-loop
system for initial conditions 2 1 = 125,145,165, 180, and
5 2 = 0.
Note that the nonlinear controller (23) does not apply for
T/2 L 1511 I 7r.
Some comparisons between the linear, nonlinear and fuzzy
control designs are summarized loosely in Table I.
To test the robustness of this controller, the following
simulations are conducted: 1) m is changed 2.0 kg to 4.0
kg, 2 ) M is changed from 8.0 kg to 4.0 kg, and 3) 21 is
changed from 1.0 m to 0.5 m. For each case, we simulate
the closed-loop system for the following initial conditions
5 1 = 45, 85, 145, 180, and 22 = 0. The results are shown
in Figs. 12, 13, and 14, respectively, for cases 1, 2, and 3.
IV. CONCLUSION
Stability conditions of fuzzy models and fuzzy control
systems are given. The basin of attraction of a fuzzy system
can be membership function dependent. A design methodology
for stabilization of a class of nonlinear systems based on
Takagi-Sugeno fuzzy model and parallel distributed com-
pensation control design is presented. The design procedure
is conceptually simple and natural. Moreover, the stability
analysis and control design problems are reduced to LMI
problems. Therefore, they can be solved very efficiently in
practice by convex programming techniques for LMIs. The
design methodology is illustrated by application to the problem
of balancing and swing-up of an inverted pendulum on a cart.
ACKNOWLEDGMENT
The authors are grateful to Dr. W. T. Thompkins, Jr. for his
generous support and encouragement.
REFERENCES
K. Tanaka and M. Sugeno, Stability analysis of fuzzy systems using
Lyapunovs direct method, in Proc. NAFIPS 90, 1990, pp. 133-136.
-, Stability analysis and design of fuzzy control systems, Fuzzy
Sets Syst., vol. 45, no. 2, pp. 135-156, 1992.
R. Langari and M. Tomizuka, Analysis and synthesis of fuzzy linguistic
control systems, in Proc. 1990 ASME Winter Ann. Meet., 1990, pp.
3542.
L. X. Wang, Adaptive F u m Systems and Control: Design and Stability
Analysis. Englewood Cliffs, NJ: Prentice-Hall, 1994.
G. Chen and H. Ying, On the stability of fuzzy control systems, in
Proc. 3rd IFIS, Houston, TX, 1993.
S . S . Farinwata and G. Vachtsevanos, Stability analysis of the fuzzy
logic controller, in Proc. IEEE Con8 Decision Contr., San Antonio,
TX, 1993.
23
T. Takagi and M. Sugeno, Fuzzy identification of systems and its
applications to modeling and control, IEEE Trans. Syst., Man, Cybern.,
15, pp 116-132, i w .
K. Tanaka and M. Sano, A robust stabilization problem of fuzzy control
systems and its application to backing up control of a truck-trailer,
IEEE Trans. Fuzzy Syst., vol. 2, no. 2, pp. 119-134, 1994.
K. Tanaka, A Theory of Advanced Fuzzy Control. Tokyo, Japan: Ky-
ouritsu Pub., 1994 (in Japanese).
S . Kawamoto et al., An approach to stability analysis of second order
fuzzy systems, in Proc. Fuzz IEEE 92, pp. 1427-1434, 1992.
H. 0. Wang, K. Tanaka, and M. F. Griffin, Parallel distributed com-
pensation of nonlinear systems by Takagi-Sugeno fuzzy model, Proc.
FUZZIEEELFES 95, pp. 531-538, 1995.
-, An analytical framework of fuzzy modeling and control of
nonlinear systems: Stability and design issues, in Proc. I995 Amer.
Contr. Con$, pp. 2272-2276, Seattle, WA, 1995.
M. Sugeno and G. T. Kang, Fuzzy modeling and control of multilayer
incinerator, Fuzzy Sets Syst., vol. 18, pp. 329-346, 1986.
J.-J. E. Slotine and W. Li, Applied Nonlinear Control. Englewood
Cliffs, New Jersey: hentice-Hall, 1991.
S . Boyd et al., Linear matrix inequalities in systems and control
theory, SIAM. Philadelphia, PA, 1994.
Y. Nesterov and A. Nemirovsky, Interior-point polynomial methods in
convex programming, SIAM. Philadelphia, PA, 1994.
W. T. Baumann and W. J. Rugh, Feedback control of nonlinear systems
by extended linearization, IEEE Trans. Automat. Contr., vol. AC-3 1,
no. I, pp. 40-46, 1986.
R. H. Cannon, Dynamics of Physical Systems. New York McGraw-
Hill, 1967.
H. E. Nusse and J. A. Yorke, Numerical Investigations of Chaotic
Systems: A Handbook for JAYS Dynamics, Draft, Inst. Phys. Sci. Tech.,
Univ. Maryland, College Park, MD, 1992.
Hua 0. Wang (M94), for photograph and biography, see this issue, p. 13.
Kazuo Tanaka (S87-M91), for photograph and biography, see this issue,
p. 13.
Michael F. Griffin (S8O-M84-M87-M89) re-
ceived the BSEET and MSEE degrees in 1979 and
1982, respectively, from the University of Cincin-
nati, Cincinnati, OH. In 1989, he received the Ph.D.
degree in electrical engineering from the University
of Florida, Gainesville.
He is currently a Research Engineer for United
Technologies Research Center in East Hartford, CT,
where he has participated in the analysis of jet
engine controllers, sonadradar adaptive algorithms
development, and fuzzy control theory. Prior to
UTRC, he worked as a biomedical engineerhesearcher for Cardiac Research
at the Childrens Hospital Medical Center in Cincinnati. There, he worked on
system identification of aortic input impedance. In 1992 and 1993, he was a
visiting researcher at the Tokyo Institute of Technology where he participated
in the design of a fuzzy logic controller for an unmanned helicopter. His
current research interests include fuzzy stability issues, nonstationary signal
processing, and modeling of complex nonlinear systems.