Академический Документы
Профессиональный Документы
Культура Документы
www.elsevier.com/locate/amc
a
BRGA, Aerospace Electronics Systems, Honeywell International Inc., 5353 W. Bell Road Glendale,
AZ 85308, USA
b
The George W. Woodruff School of Mechanical Engineering, Georgia Institute of Technology,
Atlanta, GA 30332-0405, USA
c
Marshall Space Flight Center, NASA, Huntsville, AL 35812, USA
Abstract
This paper presents stability analysis and robust control design for Takagi and
Sugeno (T–S) fuzzy dynamical systems. A systematic approach is proposed to check
the stability of the T–S fuzzy system. We then extend the consideration to the uncer-
tainty case, which can be nonlinear and (possibly fast) time-varying. Only the possible
bound of the uncertainty is needed. If the uncertainty is matched, a robust control
scheme is proposed, which renders the fuzzy system practically stable. If the uncertainty
is mismatched, we show that a mismatched threshold is needed to ensure stability.
2004 Elsevier Inc. All rights reserved.
*
Corresponding author.
E-mail address: yehwa.chen@me.gatech.edu (Y.-H. Chen).
0096-3003/$ - see front matter 2004 Elsevier Inc. All rights reserved.
doi:10.1016/j.amc.2004.06.045
556 T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572
1. Introduction
The basic framework of fuzzy controllers was originally established and ap-
plied to control a steam engine in [1]. The applications of the Mamdani-type
fuzzy control architecture to numerous industrial processes have been generally
recognized as a success. The new control architecture has provided control
engineers an alternative to cope with complicated control problems. A number
of industrial applications by fuzzy control have been developed [2–4].
In addition to this, fuzzy logic theory has been recognized as a means to de-
scribe uncertainty. This is especially important if one wishes to view the uncer-
tainty in terms of the extent of occurrence of an event. The Takagi and Sugeno
(T–S) fuzzy model was proposed to serve this purpose by Takagi and Sugeno
[5].
Stability is generally recognized one vital issue in designing a control system.
It is important for a control system to maintain stability as well as to achieve
good performance. The perspective and advancement of fuzzy control systems
heavily rely on the development of stability analysis. There have been a number
of papers presented to analyze the stability issue of fuzzy control systems based
on the T–S fuzzy model. A stability condition for T–S fuzzy model was pro-
posed in [6]. To meet the stability condition, a common positive definite matrix
is needed for a given T–S fuzzy system.
We propose to study the stability issue for control system analysis and de-
sign when uncertainty appears. The uncertainty is nonlinear and (possibly fast)
time-varying. Only the possible bound of the uncertainty is needed. The result-
ing system performance (e.g., practical stability) can be prescribed and
guaranteed.
The main contributions of the paper are threefold. First, we propose a sys-
tematic approach to find a common positive definite matrix. This in turn helps
to determine the stability of an uncertainty-free T–S fuzzy system. We then
extend to the consideration of an uncertainty case by adopting the means for
quadratic stability. Second, we suggest a robust control design for the T–S fuz-
zy system by utilizing the Lyapunov minimax approach. Under the matching
condition, the controlled T–S fuzzy system is guaranteed to be (globally) prac-
tically stable. This is regardless of the realization of the uncertainty. Third, we
further investigate the mismatched uncertainty case. The practical stability per-
formance is still guaranteed if the uncertainty bound is within a prescribed
threshold.
[x1, x2, . . ., xn]T be the state variable and u = [u1, u2, . . ., um]T be the input vector.
The ith rule of a continuous T–S fuzzy model is expressed as the following:
Ri : If x1 ðtÞ is M ix1 and . . . and xn ðtÞ is M ixn and u1 ðtÞ is M iu1 and . . . and
um ðtÞ is M ium ;
where wi(z(t)) is the overall truth value (i.e., weight) of the premise of the ith
implication. It is calculated as
Y
n Y
m
wi ðzðtÞÞ ¼ M ixp ðtÞ M iuq ðtÞ ; ð2:3Þ
p¼1 q¼1
wi ðzðtÞÞ
ki ðzðtÞÞ ¼ PN : ð2:4Þ
i¼1 wi ðzðtÞÞ
PN
It is reasonable to assume that P wi(z(t)) P 0 and i¼1 wi ðzðtÞÞ > 0. Conse-
N
quently, we have ki(z(t)) P 0 and i¼1 ki ðzðtÞÞ ¼ 1.
It has been shown that a set of fuzzy implications described in (2.1)
can express a highly nonlinear function in spite of a small number of
fuzzy implications [7]. Tanaka pointed out that the dynamics of linear
models and neural network models can be perfectly represented by the T–S
fuzzy models [8]. Furthermore, it has been shown that the standard fuzzy
systems and multilayered feedforward neural networks are special cases
of the T–S fuzzy models. This occurs as the output consequence of each
implication is expressed as a linear combination of Lipschitz continuous func-
tions [9].
Stability is one of the most important and fundamental issues in control sys-
tem design and analysis. We first propose an alternative approach to analyze
the stability of an open-loop (i.e., uncontrolled) T–S fuzzy model.
558 T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572
for i = 1,2, . . ., N.
A stability condition for (3.3) is then the existence of P > 0 such that
ATi PAi P < 0 ð3:4Þ
for i = 1,2, . . ., N.
Theorem 2. For a given positive definite matrix Q, let P1, P2, . . ., PN > 0 be the
unique solutions of the following Lyapunov equations:
AT1 P 1 þ P 1 A1 ¼ Q;
ð3:6Þ
ATi P i þ P i Ai ¼ P i1 ; i ¼ 2; . . . ; N :
If Assumption 1 holds, then the function V(x) = xTPNx is a common Lyapunov
function for each of the individual systems x_ ¼ Ai x, i = 1, 2, . . ., N.
T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572 559
Theorem 3. For a given positive definite matrix Q, let P1, P2, . . ., PN be the
unique solutions of the Lyapunov equations given by (3.6), If Assumption 1 holds,
system (3.1) is asymptotically stable.
We now turn to the system (2.2), which is under the u. Let the control be
T
A i P N þ P N Ai < 0 ð3:12Þ
for i = 1,2, . . ., N.
Substituting (3.11) into (3.1), we obtain
X
N
x_ ðtÞ ¼ ki ðzðtÞÞðAi þ DAi ÞxðtÞ: ð3:13Þ
i¼1
Remark 3 indicates that the stability of system (3.13) can be determined even
if uncertainties exist in the system.
Most of the existing work on control design for T–S fuzzy models is devel-
oped without uncertainty. We shall extend the scope to address the modeling
uncertainty issue. Based on Theorem 3, we propose a robust control, which
renders the uncertain T–S fuzzy model globally practically stable by utilizing
the Lyapunov minimax approach [13].
T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572 561
1
I þ ðEi ðqÞ þ ETi ðqÞÞ P dI; d > 0; ð4:7Þ
2
i = 1, 2, . . ., N.
Remark 5. The choice of Ai and B in (4.3) and (4.4) is not unique. For a given
T–S fuzzy model, one can flexibly choose Ai and B to meet Assumptions 1
and 2.
Define
T
qðxÞ:
lðxÞ :¼ B P N x^ ð4:11Þ
For any > 0, let the control scheme be
uðtÞ ¼ pðxðtÞÞ; ð4:12Þ
where
8
> lðxÞ
>
< ^ðxÞ; if klðxÞk P ;
q
klðxÞk
pðxÞ ¼ ð4:13Þ
>
: lðxÞ q
>
^ðxÞ; if klðxÞk < :
X
N
T T T
¼ ki xT ½Ai P N þ P N Ai þ ðBDi Þ P N þ P N ðBDi Þx
i¼1
^2 T
q T T
x P N Bð2I þ Ei þ Ei ÞB P N x
klk
XN
T T
6 ki ½xT ðAi P N þ P N Ai Þx þ 2dkxT P N Bk^
q 2dklk
i¼1
X
N
T T
¼ ki ½xT ðAi P N þ P N Ai Þx: ð4:17Þ
i¼1
Let
k :¼ min kmin ðRi Þ; ð4:20Þ
i
Consequently, we obtain
2
X
N
2
V_ 6 kkxk ki ¼ kkxk : ð4:22Þ
i¼1
(ii) If klk 6 :
XN
T T T
_
V ¼ ki xT ½Ai P N þ P N Ai þ ðBDi Þ P N þ P N ðBDi Þx
i¼1
q
^ T q
^
lT ðI þ ETi ÞB P N x xT P N BðI þ Ei Þl
XN
T T
¼ ki xT ½Ai P N þ P N Ai þ ðBDi ÞT P N þ P N ðBDi Þx
i¼1
^2 T
q T
x P N Bð2I þ ETi þ Ei ÞB P N x
564 T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572
" #
X
N
T T klk2
T
6 ki x ðAi P N þ P N Ai Þx2dklk 2d
i¼1
XN
T T d 2
¼ ki xT ðAi P N þ P N Ai Þx 2 ðklk klkÞ
i¼1
XN
T T T d 2 1
¼ ki x ðAi P N þ P N Ai Þx 2 klk þ d
i¼1
2 2
2 1
6 kkxk þ d: ð4:23Þ
2
Following the standard argument in [13], the controlled system is globally
practically stable. The uniform boundedness region is with radius
jR; if r 6 R;
dðrÞ ¼ ð4:24Þ
jr; if r > R;
where
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
kmax ðP N Þ
j :¼ ; ð4:25Þ
kmin ðP N Þ
sffiffiffiffiffi
d
R :¼ : ð4:26Þ
2k
The uniform ultimate boundedness ball is with radius dð> jRÞ and the
maximum amount of time it takes to enter this ball (and remains there there-
after) is
8
> 0; if r 6 j1 d;
<
2
T ðd; rÞ ¼ kmax ðP N Þc2 kmin ðP N Þj2 d ð4:27Þ
>
: ; otherwise:
1=2 2 2 1
k j d 2 kd
The uniform stability ball is with radius R. Both d and R can be made arbitrar-
ily small by an appropriate choice of . The proof is thus completed. h
In case the matching conditions (4.5) and (4.6) are not met, we need to inves-
tigate the mismatched case. Let us decompose the uncertainty in the following
way:
~ i ðqÞ;
DAi ðqÞ ¼ BDi ðqÞ þ DA ð5:1Þ
~ i ðqÞ:
DBi ðqÞ ¼ BEi ðqÞ þ DB ð5:2Þ
Let
~ i ðqÞk;
qDA~ :¼ sup max kDA ð5:3Þ
i q2X
~ i ðqÞk:
qDB~ :¼ sup max kDB ð5:4Þ
i q2X
Define
~ i ðqÞx þ DB
~ei ðx; qÞ :¼ DA ~ i ðqÞpðxÞ; ð5:5Þ
^ðxÞ ¼ qkxk=d.
where p(x) is given by (4.13) with q
By applying (5.3) and (5.4), we can obtain
X
N X
N
ki ð~eTi P TN x þ xT P N ~ei Þ ¼ ki ½ðDA ~ i ðqÞpðxÞÞT P T x
~ i ðqÞx þ DB
N
i¼1 i¼1
þ xT P N ðDA~ i ðqÞx þ DB
~ i ðqÞpðxÞÞ
X N
1 2
62 ki kmax ðP N Þ qDA~ þ qDB~ q kxk
i¼1
d
1 2
¼ 2kmax ðP N Þ qDA~ þ qDB~ q kxk
d
¼: ckxk2 : ð5:6Þ
Theorem 5. Consider the T–S fuzzy model (4.1) under the mismatched case ex-
^ðxÞ ¼ qkxk=d
pressed by (5.1) and (5.2). The control scheme given by (4.12) with q
renders system (4.1) globally practically stable if
c < k; ð5:7Þ
where k is given by (4.20) and c by (5.6). Furthermore, the sizes of the uniform
ultimate boundedness region and the uniform stability region can be made arbi-
trarily small by a suitable choice of .
566 T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572
Proof. Let the Lyapunov function candidate V be the same as (4.14). The
derivative of V along the trajectory of the controlled system of (4.1) is
V_ ¼ x_ T P N x þ xT P N x_
X N n T T T
¼ ki ½xT ðAi þ DAi Þ þ uT ðB þ DBTi ÞP N x þ xT P N ½ðAi þ DAi Þx
i¼1
o
þðB þ DBi Þu þ ~eTi P TN x þ xT P N ~ei : ð5:8Þ
Applying (5.6),
1 1
V_ 6 kkxk þ ~kd þ ckxk ¼ ðk cÞkxk þ ~kd:
2 2 2
ð5:10Þ
2 2
Therefore, if (5.7) holds, the controlled system of (4.1) is practically stable
by following the similar argument as in the proof of Theorem 4. The size of
the ultimate boundedness region and the uniform stability ball can be deter-
mined subsequently. The proof is thus completed. h
6. Design example
In this section, we will illustrate the effectiveness of the proposed robust con-
trol design. Consider a mass-spring-damper mechanical system [14],
M€y þ gðy; yÞ
_ þ f ðyÞ ¼ /ð_yÞu; ð6:1Þ
where y is the displacement, y_ is the velocity, €y is the acceleration, M is the
mass, gðy; y_ Þ is the damper force, f(y) is the spring force, and /ð_y Þ is coupled
with the input u. The functions g( Æ ), f( Æ ), and /( Æ ) are nonlinear and possibly
uncertain. We shall first consider that the functions are known.
We adopt the following numerical values and function forms: M = 1.0 kg,
_ ¼ y_ , f(y) = 0.01y + 0.1y3 and /ð_y Þ ¼ 1 þ 0:13_y . The mass-damper-spring
gðy; yÞ
system is proposed to be represented by the following T–S fuzzy model:
Plant Rule 1: If y(t) is F 11 and y_ ðtÞ is F 12 , then
^ 1 xðtÞ þ B
x_ ðtÞ ¼ A ^ 1 uðtÞ;
T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572 567
^ 3 xðtÞ þ B
x_ ðtÞ ¼ A ^ 3 uðtÞ;
^ 4 xðtÞ þ B
x_ ðtÞ ¼ A ^ 4 uðtÞ;
y_ 3 y_ 3
F 12 ðyÞ
_ ¼ 0:5 þ ; F 22 ð_y Þ ¼ 0:5
:
6:75 6:75
^ i and B
At first, we have to decompose A ^ i according to (4.3) and (4.4). Let
0 1
Ai ¼ ; ð6:2Þ
0:01 1:0
0
B¼ ; ð6:3Þ
1
i = 1, . . ., 4. It is noted that Assumption 1 is satisfied. Consequently, we have
0 0 0
DA1 ¼ ; DB1 ¼ ;
0 0 0:4387
0 0 0
DA2 ¼ ; DB2 ¼ ;
0 0 0:4387
568 T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572
0 0 0
DA3 ¼ ; DB3 ¼ ;
0:225 0 0:4387
0 0 0
DA4 ¼ ; DB4 ¼ :
0:225 0 0:4387
It is noted that Ai and Bi, i = 1, . . ., 4, are within the range space of B. The
matching condition is therefore met.
Let also
1 0
Q¼ :
0 1
We can obtain
50:505 50
P4 ¼
50 50:5
by solving the Lyapunov Eq. (3.6).
Based on (4.6) and (4.7), we can choose dnom = 0.5. We choose the nominal
bounding function to be:
^nom ðxÞ ¼ 0:5kxk:
q ð6:4Þ
Here ‘‘nominal’’ stands for that no uncertainty actually occurs. However, we
still have DAi and DBi following the decomposition of Ai and Bi. The bounding
function based on DAi and DBi in the form suggested in Remark 6 is taken. Let
= 0.1. We obtain the nominal control scheme (4.13).
Next, we consider the uncertainty case. Let gðy; yÞ _ ¼ ð1 þ c1 ðtÞÞy,
_
f(y) = (0.01 + c2(t))y + 0.1y3 and /ð_y Þ ¼ 1 þ 0:13_y þ c3 ðtÞ, where jc1(t)j 6 0.5,
jc2(t)j 6 1.0 and jc3(t)j 6 0.25 for all t P 0. Choose the same Ai and B as in
(6.2) and (6.3), we have
0 0 0
DA1 ¼ ; DB1 ¼ ;
c2 ðtÞ c1 ðtÞ 0:4387 þ c3 ðtÞ
0 0 0
DA2 ¼ ; DB2 ¼ ;
c2 ðtÞ c1 ðtÞ 0:4387 þ c3 ðtÞ
0 0 0
DA3 ¼ ; DB3 ¼ ;
0:225 þ c2 ðtÞ c1 ðtÞ 0:4387 þ c3 ðtÞ
0 0 0
DA4 ¼ ; DB4 ¼ :
0:225 þ c2 ðtÞ c1 ðtÞ 0:4387 þ c3 ðtÞ
It is noted that the matching condition is also met. Based on (4.6) and (4.7), we
can choose drob = 0.4. Following Remark 6, we can choose the robust bounding
function,
^rob ðxÞ ¼ 3:375kxk:
q ð6:5Þ
T.-S. Lee et al. / Appl. Math. Comput. 164 (2005) 555–572 569
The difference between q ^rob ðxÞ and q^nom ðxÞ is that the former has to address
the additional bound of uncertainty, besides what has already been addressed
^nom ðxÞ. Certainly, this renders a larger magnitude in q
in q ^rob ðxÞ. For simulation
purpose, we used sinusoidal functions for the uncertainties: c1(t) = 0.5 sin(0.2t),
c2(t) = 0.5 cos(0.5t), c3(t) = 0.25 sin(t). Let = 125. To show the effectiveness of
^rob ðxÞ, as compared with q
q ^nom ðxÞ, in compensating the uncertainty, we use the
control scheme (4.13) when q ^ðxÞ takes either one of the two forms. The control-
led system responses are shown in Figs. 1 and 2. The dotted lines denote the
responses due to the nominal control (i.e., when q ^ðxÞ ¼ q^nom ðxÞ). The full lines
denote the responses due to the robust control (i.e., when q ^ðxÞ ¼ q^rob ðxÞ). It is
noted that significant improvement in performance is obtained while the robust
control is applied. This attests the importance in compensating the uncertainty.
Fig. 3 shows the histories of both controls. It is interesting to note that despite
the use of larger q^rob ðxÞ, a smaller robust control magnitude is observed for
t P 1 s.
7. Conclusions
Theorem A.1 allows one to search only the protruded points for the maxi-
mum of g(Q).
Remark A.1. It has been shown in [11] that the maximum problem is convex
in the sense that any local extreme is also the global extreme. Thus the search
can be carried out by many standard algorithms such as the gradient search.
References
[1] E.H. Mamdani, S. Assilian, An experiment in linguistic synthesis with a fuzzy logic controller,
International Journal of Man Machine Studies 7 (1) (1975) 1–13.
[2] K. Hirota, A. Arai, S. Hachisu, Fuzzy controlled robot arm playing two-dimensional ping-
pong game, Fuzzy Sets and Systems 32 (3) (1989) 140–1159.
[3] M. Sugeno, M. Nishida, Fuzzy control of model car, Fuzzy Sets and Systems 16 (2) (1985)
103–113.
[4] S. Yasunobu, S. Miyamoto, Automatic train operation by predictive fuzzy control, in: M.
Sugeno (Ed.), Industrial Application of Fuzzy control, North-Holland, Amsterdam, pp. 1–18.
[5] T. Takagi, M. Sugeno, Fuzzy identification of systems and its applications to modeling and
control, IEEE Transactions on Systems Man and Cybernetics 15 (1) (1985) 116–132.
[6] K. Tanaka, M. Sugeno, Stability analysis and design of fuzzy control systems, Fuzzy Sets and
Systems 45 (2) (1992) 135–156.
[7] M. Sugeno, K. Tanaka, Successive identification of fuzzy model and its applications to
prediction of complex system, Fuzzy Sets and Systems 42 (4) (1991) 315–344.
[8] K. Tanaka, Stability and stabilizability of fuzzy-neural-linear control systems, IEEE
Transactions on Fuzzy Systems 3 (4) (1995) 438–447.
[9] J.T. Spooner, K.M. Passino, Stable indirect adaptive control using fuzzy systems and neural
networks, in: Proceedings of the 34th Conference on Decision and Control, New Orleans, LA,
December, 1995, pp. 249–254.
[10] K.S. Narendra, J. Balakrishnan, A common Lyapunov function for stable LTI systems with
commuting A-matrices, IEEE Transactions on Automatic Control AC-39 (12) (1994) 2469–
2471.
[11] K. Gu, M.A. Zohdy, N.K. Loh, Necessary and sufficient conditions of quadratic stability of
uncertain linear systems, IEEE Transactions on Automatic Control AC-35 (5) (1990) 601–604.
[12] K. Gu, Quadratic stability bound of discrete-time uncertain systems, in: American Control
Conference, Green Valley, AZ, 1991, pp. 1951–1955.
[13] M.J. Corless, G. Leitmann, Continuous state feedback guaranteeing uniform ultimate
boundedness for uncertain dynamic systems, IEEE Transactions on Automatic Control AC-26
(5) (1981) 1139–1144.
[14] K. Tanaka, T. Ikeda, H.O. Wang, Robust stabilization of a class of uncertain nonlinear
systems via fuzzy control: quadratic stabilizability, H1 control theory, and linear matrix
inequalities, IEEE Transactions on Fuzzy Systems 4 (1) (1995) 1–13.