Vous êtes sur la page 1sur 25

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/2804791

A Fuzzy Logic Approach to Optimal Control of Nonlinear Systems

Article · October 2000


Source: CiteSeer

CITATIONS READS

29 48

4 authors, including:

Owen Wang Linda Bushnell


California Polytechnic State University, San Luis Obispo University of Washington Seattle
132 PUBLICATIONS   4,212 CITATIONS    138 PUBLICATIONS   5,590 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

CPS: Synergy: Certifiable, Scalable, and Attack-resilient Submodular Control Framework for Smart Grid Stability View project

All content following this page was uploaded by Linda Bushnell on 21 February 2015.

The user has requested enhancement of the downloaded file.


A Fuzzy Logic Approach to Optimal Control of
Nonlinear Systems
Jing Li Hua O. Wang
 Linda Bushnell
y Yiguang Hong

Laboratory for Intelligent and Nonlinear Control (LINC)


Department of Electrical and Computer Engineering
Duke University, Durham, NC 27708, U.S.A
Kazuo Tanaka

Department of Mechanical and Control Engineering


University of Electro-Communications
1-5-1 Chofugaoka, Chofu, Tokyo 182, Japan

Abstract
In this paper, a new scheme of fuzzy optimal control for nonlinear systems is proposed based
on the framework of T-S (Takagi-Sugeno) fuzzy model and PDC (parallel distributed compen-
sation) controller. Both the discrete system case and the continuous system case are discussed
in this paper. The controller design procedure is to solve a set of LMI (linear matrix inequality)
conditions. With the state-of-art LMI solving algorithms, this can be done quite eciently. It is
also shown in this paper that various control performance speci cations can be incorporated into
this framework as well. Finally, this fuzzy optimal control approach is applied to a nonlinear
control benchmark problem: the TORA system. The results demonstrate the e ectiveness of
the approach towards achieving the optimal control objective.

 Corresponding author. E-mail: hua@ee.duke.edu; Tel: (919) 660-5273; Fax: (919) 660-5293. Dr. Wang is
also with the Center for Nonlinear and Complex Systems and the Department of Control Science and Engineering,
Huazhong University of Science and Technology, Wuhan, China. This research was supported in part by the Army
Research Oce grant no. DAA G55-98-D-0002 and by Otis Elevator Company.
y Dr. Bushnell is also with the U.S. Army Research Oce, PO Box 12211, RTP, NC 27709.

1
1 Introduction
In many physical and engineering systems, engineers are hindered by strong nonlinearity from
successful application of linear control theory. One of the dicult problems encountered is optimal
control for nonlinear systems. That is, in many practical situations, an optimal controller is desired
that can minimize (or maximize) certain performance criterion and satisfy some physical constraints
at the same time. For linear systems, the problem of designing optimal controllers reduces to
solving algebraic Riccati equations (AREs). AREs are usually easy to solve and detailed discussion
of their solutions can be found in many textbooks [29]. However, for a general nonlinear system, the
optimization problem will reduce to the so-called Hamilton-Jacobi (HJ) equations that are nonlinear
PDEs [17]. Di erent from their counterparts for linear systems, HJ equations are usually hard to
solve both numerically and analytically. Results have been given on the relationship between the
solution of HJ equation and the invariant manifold for the Hamiltonian vector eld [16]. Progresses
have also been made on the numerical computation of the approximated solution of HJ equations
[15]. But few results so far can provide an e ective way of designing optimal controllers for general
nonlinear systems. In this paper, we try to propose an alternative approach to nonlinear optimal
control based on fuzzy logic, which has emerged as one of the important techniques to contend with
nonlinearities.
In the past few decades, as the interest in fuzzy systems has increased, researchers have consid-
ered the stability analysis of these fuzzy systems using a variety of modeling and control frameworks.
One of the modeling techniques that has attracted a great deal of attention is the so-called fuzzy
Takagi-Sugeno (T-S) model originating from fuzzy identi cation [18]. The popularity of T-S model
arises not only from its simplicity, but also from the idea that local dynamics of a nonlinear plant
can be represented by di erent fuzzy rules (linear models) in T-S model. Recent results have also
shown that T-S model can be a universal approximator of any smooth nonlinear dynamic system
[4] [11] [25]. These results di er from the commonly held view that a T-S model has only limited
capability in representing a nonlinear system. Moreover, it has been found out that many nonlinear
systems can be represented exactly through sectorization by T-S models with only a few number of
rules. From the point of view of system analysis, T-S model are appealing as well since the stability
and performance characteristics of the system can be analyzed using a Lyapunov approach [2] [3] [8]
[21] [23] [26]. In this approach, sucient conditions for the stability and performance of a system
are stated in terms of the feasibility of a set of linear matrix inequalities (LMIs). Very ecient
numerical algorithms exist for determining the feasibility of LMIs, so even problems involving large

2
scale systems are computationally tractable [27] [28]. This LMI based technique is not only an anal-
ysis tool; it can also be utilized to automate the controller design process. In the paper [24], the
authors introduced a controller design called parallel distributed compensation (PDC). The PDC
design utilizes a rule-based state feedback which mirrors the structure of the T-S model. Using the
LMI approach, a set of stabilizing gains for the PDC controller can be numerically computed [10]
[19] [20] [22]. Moreover, since many linear control results can be expressed in LMI conditions, these
results can be extended to T-S models [5]. For example, multi-objective controller design can be
done through solving an augmented LMI problem [9].
Our previous successes of applying PDC control lead us to believe that T-S model may be a good
way to handle nonlinear control problems, or at least partially. Therefore, in this paper, we are going
to investigate the problem of optimal controller design for nonlinear systems in the framework of
fuzzy T-S models. In the next few sections, a new scheme of fuzzy optimal control will be proposed.
The idea is to set up an upper bound for the performance measure that needs to be optimized.
Then, a controller is designed to minimize this upper bound. For this scheme, PDC controller is
adopted as our fuzzy controller and the design procedure is converted into an LMI problem. From
the feasible solution of LMIs, the parameters of the fuzzy controller can be calculated. Compared
to classical nonlinear optimal results, this new approach is both straightforward and easy to solve.
In this paper, both the discrete system case and the continuous system case are discussed. It will
also be shown that performance speci cations, such as decay rate and input (output) constraint,
can be incorporated into this design framework as well. Finally, this new fuzzy optimal control
approach is applied to the TORA (translational oscillation by a rotational actuator) system. TORA
system is an ACC nonlinear benchmark problem [1] [6] [7]. The problem considers controlling the
translational oscillations by a rotational actuator. Many di erent approaches have been proposed
to control the system [1] [6] [7]. Most of these approaches are rather complicated and some needs
speci c information about the system. In this paper, we will rst model the TORA system by
a T-S fuzzy model. Then we will apply the new fuzzy optimal control techniques to the TORA
system. The nal control performance demonstrates the e ectiveness of our approach.
The rest of the paper is organized as follows. In section 2, the concepts of T-S model and
PDC controller are introduced. In section 3, the fuzzy optimal controller design is introduced and
discussed. Controller design for TORA system is given in section 5 and concluding remarks are
collected in section 6.
In this paper, the Rnm is used to denote the n  m matrices whose components are real

3
numbers. The notation M > 0 stands for positive de nite symmetric matrix and M  0 stands for
nonnegative de nite symmetric matrix.

2 Takagi Sugeno Models and Parallel Distributed Compensation


2.1 Takagi-Sugeno Fuzzy Model
The main feature of a Takagi-Sugeno (T-S) fuzzy model is the expression of each local dynamics by
a fuzzy implication (rule). The overall fuzzy model of the system is achieved by fuzzy interpolation
of these linear system models. Speci cally, the Takagi-Sugeno fuzzy system is of the following form:
Discrete System Case:
Rule i: IF z (k) is Mi    and zp (k) is Mip
1 1

THEN x(k + 1) = Ai x(k) + Bi u(k), and y(k) = Ci x(k)


where
xT (k) = [x1 (k); x2 (k);    ; xn (k)];
uT (k) = [u1 (k); u2 (k);    ; um (k)];
i = 1; 2;    ; r and r is the number of IF-THEN rules. zi(k) are some fuzzy variables, Mij are fuzzy
sets, and x(k + 1) = Ai x(k) + Bi u(k); y(k) = Ci x(k) is the output of the i-th IF-THEN rule.
By using the center of gravity for the defuzzi cation, we can simplify the expressions of the T-S
model as:

X
r
x(k + 1) = hi (z)(Ai x(k) + Biu(k)) (1)
i=1
Xr
y(k + 1) = hi (z)Ci x(k) (2)
i=1
where hi (z ) is the possibility for the ith rule to re ,i.e ,
h (z) = wi (k)
i X
r
wi (k)
i=1
and
Yp
wi (k) = Mij (zj (k)):
j =1

4
Mij (zj (k)) is the grade of membership of zj (k) in Mij .
Xr
It is assumed that hi (z ) have already been normalized, i.e. hi (z )  0 and hi (z ) = 1. In
i=1
the following, we will just write hi (z ) as hi . But it should be kept in mind that hi are nonlinear
functions of system states.
Continuous System Case:
Rule i: IF z (t) is Mi    and zp (t) is Mip
1 1

THEN x_ (t) = Ai x(t) + Bi u(t) and y(t) = Ci x(t),


where
xT (t) = [x1(t); x2 (t);    ; xn(t)];
uT (t) = [u1(t); u2 (t);    ; um (t)];
i = 1; 2;    ; r and r is the number of IF-THEN rules. zi(t) are some fuzzy variables, Mij are fuzzy
sets, and x_ (t) = Ai x(t) + Bi u(t) , y(t) = Ci x(t), is the output from the i-th IF-THEN rule.
Similarly, we can simplify the expression of the T-S model as:

X
r
x_ (t) = hi(z)(Ai x(t) + Bi u(t)) (3)
i=1
Xr
y(t) = hi(z)Ci x(t) (4)
i=1
where
hi (z) = Xwi(t)
r
wi(t)
i=1
and
Yp
wi (t) = Mij (zj (t)):
j =1

2.2 Parallel Distributed Compensation


The idea of parallel distributed compensation (PDC) is to associate a compensator for each rule
of the fuzzy model. The resulting overall controller, which is nonlinear in general, is a fuzzy
interpolation of each individual linear controller. A PDC controller usually takes the following
form:

5
Discrete System Case:
Rule i: IF z (k) is M i    and zp (k) is Mip
1 1

THEN u(k) = Fi x(k), where i = 1; 2;    ; r.


Hence the fuzzy controller is given by:
X
r
u(k) = hi (z)Fi x(k) (5)
i=1
In [24], the following result is given on the stabilization of the discrete T-S model:
Theorem 1 The fuzzy control system of T-S model (1) (2) using PDC control (5) is stabilizable
in the large if there exist a Q > 0 and Yi , i = 1; 2; : : : ; r such that the following LMI conditions
hold:
2 3
4 Q QATi + YiT BiT 5
> 0 i = 1; : : : ; r (6)
Ai Q + Bi Yi Q
2 3
4 2Q QATi + QATj + YjT BiT + YiT BjT 5 > 0; i < j  r: (7)
Ai Q + A j Q + B i Yj + B j Yi 2Q
And the PDC controller is given by:
Fi = Yi Q?1 (8)
The Lyapunov function is given by
V = xT Q?1 x (9)
Continuous System Case:
Rule i: IF z (t) is M i    and zp (t) is Mip
1 1

THEN u(t) = Fi x(t), where i = 1; 2;    ; r.


Hence the fuzzy controller is
X r
u(t) = hi (z)Fi x(t) (10)
i=1
Similarly, the following result holds for the stabilization of the continuous T-S model.
Theorem 2 The fuzzy control system of T-S model (3) (4) using PDC control (10) is stabilizable
in the large if there exist a Q > 0 and Yi , i = 1; 2; : : : ; r such that the following LMI conditions
hold:

6
QATi + AiQ + BiYi + YiT BiT < 0; i = 1; : : : ; r (11)
QATi + Ai Q + QATj + Aj Q + Bi Yj + YjT BiT + Bj Yi + YiT BjT < 0; i < j  r: (12)
And the PDC controller is given by:
Fi = Yi Q?1 (13)
The Lyapunov function is given by
V = xT Q?1 x (14)
Remark: The above result guarantees quadratic stability for the region where a T-S model repre-
sentation is valid. Within this region, it is a semi-global stabilization result for the original system.
If the T-S model is valid globally, the above result represents a global stabilization result.

3 Fuzzy Optimal Control


In this section, a new scheme of fuzzy optimal control will be presented. The control objective
is to minimize certain performance measure. It will also be shown in this section that design
speci cations, such as decay rate and input (output) constraint, can be incorporated into this
design framework as well. We begin with the discrete system case. Before the discussion, two
lemmas used in the later proof are given.
Lemma 3
YjT RYi + YiT RYj  YiT RYi + YjT RYj (15)
where R = RT  0.

Lemma 4 (S-procedure) [27]


2 3
4 QT S 5 < 0
S R
is equivalent to R < 0 and Q ? SR?1 S T < 0

7
3.1 Discrete Optimal Controller
Suppose the T-S model is given as:
X
r X
r
x(k + 1) = hi Aix(k) + hiBi u(k) (16)
i=1 i=1
Xr
y = Cix(k) (17)
i=1
The control objective is to minimize
X
1
J= (yT (k)Wy(k) + uT (k)Ru(k)) (18)
k=0
where W = W T > 0 and R = RT > 0. The initial condition is given as x(0) = x0
Consider the Lyapunov function V (x) = xT (k)Px(k). We rst assume that the closed-loop
system is asymptotically stable which will be justi ed later in the design conditions. Then if the
condition
V (k) , V (x(k + 1)) ? V (x(k)) < ?(yT (k)Wy(k) + uT (k)Ru(k)) (19)
is satis ed, the following inequality can be obtained by just summing up (19) from 0 to 1.
J < V (x(0)) (20)
Therefore, V (x(0)) gives an upper bound of J under the above assumptions. The controller to be
introduced is in the sense a \sub-optimal" controller since V (x(0)) will be optimized instead of J
in the controller design procedure. By doing so, the controller design procedure is converted into
the following problem:
Problem: Find a controller to minimize
subject to: J < xT (0)Px(0) 
If we de ne Q = P ?1 , the problem can be rewritten in the following form using S-procedure:
Problem: Find a controller to minimize
subject to: 2 3
41 xT (0) 50 (21)
x(0) Q
To solve this problem, the following PDC controller is adopted as our fuzzy optimal controller:
X
r
u(k) = Fi x(k) (22)
i=1

8
With this controller, the closed-loop system becomes
r X
X r
x(k + 1) = hi hj (Ai + Bi Fj )x(k)
i=1 j =1
Xr Xr X r
= h2i Giix(k) + 2 hihj ( Gij +2 Gji )x(k)
i=1 i=1 j =i+1
where Gij = Ai + Bi Fj
Then we can write V (x(k)) as
V (x(k)) = V (x(k + 1)) ? V (x(k))
r X
X r X
r X r
= hi hj hk hl xT (k)(Gij PGkl ? P )x(k)
i=1 j =1 k=1 l=1
Xr X r X r X r  
= 41 hi hj hk hl xT (k) (Gij + Gji)T P (Gkl + Glk ) ? 4P x(k)
i=1 j =1 k=1 l=1
Applying Lemma 3, we can get the following inequality:
1 Xr X r
V (x(k))  4 hi hj xT (k)[(Gij + Gji)T P (Gij + Gji) ? 4P ]x(k)
i=1 j =1
r X
X r  (Gij + Gji)T (G + G ) 
= hi hj xT (k) 2 P ij ji ? P x(k)
2
i=1 j =1
Xr
= h2i xT (k)(GTii PGii ? P )x(k)
i=1
X L X L  T 
+2 hihj xT (k) (Gij +2 Gji) P (Gij +2 Gji ) ? P x(k)
i=1 j =i+1
So if
GTii PGii ? P + CiT WCi + FiT RFi < 0 (23)
(Gij + Gji)T P (Gij + Gji) ? P + CiT WCj + CjT WCi + FiT RFj + FjT RFi < 0 (24)
2 2 2 2
Then
X
r
V (x(k)) < h2i xT (k)(?CiT WCi ? Fi RFi )x(k)
i=1
X
r X
r " #
CiT WCj + CjT WCi FiT RFj + FjT RFi
+2 hi hj xT (k) ? 2 ? 2 x(k)
i=1 j =i+1

9
Since xT (k)Wx(k) + uT (k)Ru(k) can be written as
2r 3
X r X
X r C T WCj + C T WCi F T RFj + F T RFi
xT (k) 4 hi (CiT WCi + FiT RFi) + 2
2
hi hj ( i 2
j + i 2
j )5 x(k)
i=1 i=1 j =i+1
therefore
V (x(k)) < ?yT (k)Wy(k) ? uT (k)Ru(k)
which is exactly (19).
Now we need to convert the condition (23), (24) to LMIs. From (23) we get:
(Ai + Bi Fi )T P (Ai + Bi Fi ) ? P + CiT WCi + FiT RFi < 0
or equivalently
(Ai + Bi Fi )T Q?1 (Ai + Bi Fi ) ? Q?1 + CiT WCi + FiT RFi < 0

If we de ne Yi = Fi Q, the above equation can be written as:


(Ai Q + Bi Yi )T Q?1 (Ai Q + Bi Yi ) ? Q + QCiT WCi Q + YiT RYi < 0
which is equivalent to (25) using S-procedure.
2 3
(Ai Q + Bi Yi )T QCiT W = YiT R =
66 Q
1 2 1 2
77
66 AiQ + BiYi Q 0 0 77 > 0 (25)
64 W = CiQ
1 2
0 I 0 75
R1=2 Y 0 0 I
From (24), we get
(Ai + Bi Fj + Aj + Bj Fi )T Q?1 (Ai + Bi Fj + Aj + Bj Fi )
?4 Q?1 + 2CiT WCj + 2Cj WCi + 2FiT RFj + 2FjT RFi < 0
or equivalently
(Ai Q + Bi Yj + Aj Q + Bj Yi )T Q?1 (Ai Q + Bi Yj + Aj Q + Bj Yi ) ? 4Q
+2 ?1 Q(CiT WCj + Cj WCi)Q + 2 ?1 YiT RYj + 2 ?1 YjT RYi < 0

10
Applying Lemma 3, we can get CjT WCi + CiT WCj  CiT WCi + CjT WCj and YjT RYi + YiT RYj 
YiT RYi + YjT RYj , then the above inequality can be rewritten as:
4Q ? (Ai Q + Bi Yj + Aj Q + Bj Yi )Q?1 (Ai Q + Bi Yj + Aj Q + Bj Yi )
?2 ?1 Q(CiT WCi + CjT WCj )Q ? 2 ?1 YiT RYi ? 2 ?1 YjT RYj > 0
which is equivalent to (26) using S-procedure
2 T p2QC T W = p2QC T W = p2Y T R = p2Y T R =
3
4Q S
66 77
1 2 1 2 1 2 1 2
i j i j
66 p S Q 0 0 0 0 77
66 2W = Ci Q 0
1 2
I 0 0 0 77
66 p = 77 > 0 (26)
66 p2W Cj Q 0
1 2
0 I 0 0 77
64 2R = Yi 0
1 2
0 0 I 0 75
p
2R1=2 Yj 0 0 0 0 I
where S = Ai Q + Bi Yj + Aj Q + Bj Yi .
Finally, we have the following theorem:
Theorem 5 The fuzzy T-S model (16) (17) is stabilizable with PDC control (22) if there exist a
Q > 0 and Yi ; i = 1; 2; : : : ; r such that the LMI conditions (21)(25)(26) are satis ed. And the
performance measure J will be less than . The parameters of the fuzzy optimal controller are given
by Fi = Yi Q?1 and the Lyapunov function is V (x) = xT Q?1 x
Remark: To arrive at (19), it was assumed that the closed-loop system is asymptotically stable.
This assumption would be automatically satis ed if the LMI design conditions in the above theorem
hold.
From the above proof, the following corollary can also be easily derived:
Corollary 6 Under the assumptions in Theorem 5 , the region xT (k)Q? x(k)  1 is an invariant
1

set of the system under the PDC control (22). That is, any system starts in this region will remain
in it afterwards.
Sometimes, the relationship between y and x is linear, i.e. y(k) = Cx(k). In these cases, the
following corollary holds true:
Corollary 7 The fuzzy T-S model (16) and y(k) = Cx(k) is stabilizable with PDC control (22)
if there exist a Q > 0 and Yi ; i = 1; 2; : : : ; r such that in addition to (21) the following LMI

11
conditions are satis ed.
2 3
(Ai Q + Bi Yi )T QC T W = YiT R =
66 Q
1 2 1 2
77
66 Ai Q + BiYi Q 0 0 77 > 0 (27)
64 W = CQ
1 2
0 I 0 75
R1=2 Y 0 0 I
2 p p T = 3
4Q S T 2QC T W = 2YiT R = 2 Yj R
66 77
1 2 1 2 1 2

66 S Q 0 0 0 77
66 2W = CQ 0
1 2
I 0 0 77 > 0 (28)
66 p = 77
4 p 2 R Yi 0
1 2
0 I 0 5
2R1=2 Yj 0 0 0 I
where S = Ai Q + Bi Yj + Aj Q + Bj Yi. And the performance measure J will be less than . The
parameters of the fuzzy optimal controller are given by Fi = Yi Q?1 and the Lyapunov function is
V (x) = xT Q?1 x
Remark : Multi-objective optimal controller design can be easily incorporated into this design
framework by combining the above LMIs with LMI conditions for other design requirements [27].
For example, if we need to put constraint on the decay rate of the closed-loop system Lyapunov
function, the following LMIs should be included:
 Decay Rate: The condition that V (x(k))  ( ? 1)V (x(k)) is equivalent to
2 3
4 Q QATi + YiT BiT
5  0 i = 1; : : : ; r (29)
Ai Q + Bi Yi Q
2 3
4 2 Q QATi + QATj + YjT BiT + YiT BjT 5
 0; i < j  r: (30)
Ai Q + A j Q + B i Yj + B j Yi 2Q
LMI conditions for other performance speci cations can be similarly derived:
 Constraint on Input:
X
r
As we know maxk0 ku(k)k = maxk0 k2
2 hi Yi Q?1 x(k)k22 . From Corollary 6, we also know
i=1
that xT (k)Q?1 x(k) is an invariant set. Therefore,
X
r
max
k0
ku(k)k 
2
max k hi Yi Q?1 zk22
2
zT Q?1 z<1 i=1
 max max kY Q? z k
i zT Q?1 z< i
1 2
2
1

 max (Q? = YiT YiQ? = )


1 2 1 2

12
So the constraint ku(t)k2  umax is enforced at all time t  0 if:
2 3
Q Y T
4 i 50 (31)
Yi u2max I
 Constraint on output:
2 3
X
r X
L X
r
max
k1
ky(k)k22 = max
k
k hl Cl hi 4Aix(k) + Bi hj Fj x(k)5 k
0
2
2
l=1 i=1 j =1
= max max kC (Ai + Bi Fj + Aj + Bj Fi ) Q1=2 k2
i;j;l zT Q?1 z<1 l 2 2

Therefore, the constraint ky(t)k2  ymax is enforced at all time t  0 if the following inequality
holds true.
2 QATi +YjT BiT +QAj +YiT BjT ) T
3
4 Q (
Cl 5  0; 8i; j; l
2
(32)
Cl (Ai Q+Bi Yj +2 Aj Q+Bj Yi) ymax
2
I

4 Continuous System Case


In this section, we consider the continuous system case
X
r X
r
x_ (t) = Ai x(t) + Biu(t) (33)
i=1 i=1
Xr
y(t) = Ci x(t) (34)
i=1
The control objective is to minimize
Z1
J= yT (t)Wy(t) + uT (t)Ru(t) dt (35)
0

Consider the Lyapunov function V (x) = xT Px. First, we assume that the closed-loop system is
asymptotically stable which would be justi ed later in the design conditions. Then if the condition
dV < ?(yT (t)Wy(t) + uT (t)Ru(t)) (36)
dt
is satis ed, the following inequality can be obtained by just integrating both side of (36) from 0
and 1.
J < V (x(0)) (37)

13
Therefore, V (x(0)) gives an upper bound of J under the above assumptions. Similar to the discrete
system case, the controller design to be proposed is to optimize V (x(0)) instead of J . By doing
that, the design problem is converted to the following problem:
Problem: Find a controller to minimize
subject to: J < xT (0)Px(0) 
If we de ne Q = P ?1 , the problem can be rewritten in the following form using S-procedure :
Problem: Find a controller to minimize
subject to: 2 3
1 x T (0)
4 50 (38)
x(0) Q
The following PDC controller will be applied to the original system
X
r
u(t) = Fi x(t) (39)
i=1
With this controller, the closed-loop system becomes
r X
X r
x_ (t) = hi hj (Ai + Bi Fj )x(t)
i=1 j =1
X
r r X
X r
= hi Gii x(t) + 2
2
hi hj ( Gij +2 Gji )x(t)
i=1 i=1 j =i+1
where Gij = Ai + Bi Fj
Then, we can write dV (dtx(t)) as:
dV (x(t)) = X
r X
r X
r
dt h2i xT (t)(GTii P + PGii)x(t) + 2 hi hj
i=1 i=1 j =i+1
 G +G G + G 
x (t) ( ij ji )T P
T
2 + P( ij
2
ji ) x(t)
So if
GiiP + PGii + CiT WCi + FiT RFi < 0 (40)
C T WCj + CjT WCi FiT RFj FjT RFi
( Gij +2 Gji )T P + P ( Gij +2 Gji ) + i 2 + 2 + 2 < 0 (41)
Then " #
dV (x(t)) < ?xT (t) Xr
h2i (CiT WCi + FiT RFi) x(t) ?
dt i=1 2 r r T 3
X X T T T
C WCj + Cj WCi Fi RFj + Fj RFi 5
xT (t) 42 ( i + 2 ) x(t) 2
i=1 j =i+1

14
Since xT (t)Wx(t) + uT (t)Ru(t) can be written as:
2r 3
X r X
X r C T WCj + C T WCi
xT (t) 4 hi CiT WCi + 2
2
hi hj ( i 2
j )5 x(t)
i=1 i=1 j =i+1
therefore
dV (x(t)) < ?yT (t)Wy(t) ? uT (t)Ru(t)
dt
which is exactly (36). Now we need to convert (40) (41) into LMIs. If we de ne Yi = Fi Q, (40) can
be written as:
(Ai + Bi Fi )T P + P (Ai + Bi Fi ) + CiT WCi + Fi RFi < 0
or equivalently
AQ + QAT + BiYi + YiT BiT + ?1 QCiT WCiQ + ?1 YiT RYi < 0
The above inequality can be shown to be equivalent to equivalent to the following LMI using
S-procedure 0 1
BB ?AiQ ? QAi =? BiYi ? Yi Bi QCi W Yi R C
T T T T = T = 1 2 1 2

@ W Ci Q
1 2
I 0 CA>0 (42)
R1=2 Yi 0 I
From Lemma 3, we know that CjT WCi + CiT WCj  CiT WCi + CjT WCj and YjT RYi + YiT RYj 
YiT RYi + YjT RYj . Application of these inequality to (41) will result
(Ai + Bi Fj + Aj + Bj Fi )T P + P (Ai + Bi Fj + Aj + Bj Fi )P
+CiT WCi + CjT WCj + FiT RFi + FjT RFj < 0
or equivalently
AiQ + QATi + BiYj + YjT BiT + Aj Q + QAj + Bj Yi + YjT BjT
+ ?1 QCiT WCiQ + ?1 QCjT WCj Q + ?1 YiT RYi + ?1 YjT RYj < 0
This inequality can be converted into the following LMI using S-procedure.
0 1
QCiW = QCj W = YiT R = YjT R =
BB =T C
1 2 1 2 1 2 1 2

BB W CiQ I
1 2
0 0 0 C
C
BB W = Cj Q C
C >0 (43)
1 2

BB = 0 I 0 0 C
C
@ R Yi
1 2
0 0 I 0 C
A
R1=2 Yj 0 0 0 I
15
where T = ?Ai Q ? Aj Q ? QATi ? QATj ? Bi Yj ? YjT BiT ? Bj Yi ? YiT BjT Finally we have the following
theorem:
Theorem 8 The fuzzy T-S model (33) (34) is stabilizable with PDC control (39) if there exist a
Q > 0 and Yi ; i = 1; 2; : : : ; r such that the LMI conditions (38)(42)(43) are satis ed. And the
performance measure J will be less than . The parameters of the fuzzy optimal controller are given
by Fi = Yi Q?1 and the Lyapunov function is V (x) = xT Q?1 x
Remark: To arrive at (36), it was assumed that the closed-loop system is asymptotically stable.
This assumption would be automatically satis ed if the LMI design conditions in the above theorem
hold.
From the above proof, the following corollary can be easily derived:
Corollary 9 Under the assumptions in Theorem 8 , the region xT (t)Q? x(t)  1 is an invariant
1

set of the system under the PDC control (39).


Similar to the discrete system case, we have the following corollary based on Theorem 8 for the
cases where the relationship between y and x is linear, i.e. y(t) = Cx(t).
Corollary 10 The fuzzy T-S model (33) and y(t) = Cx(t) is stabilizable with PDC control (39)
if there exist a Q > 0 and Yi ; i = 1; 2; : : : ; r such that in addition to (38) the following LMI
conditions are satis ed.
0 1
BB ?AiQ ? QAi =? BiYi ? Yi Bi QC W Yi R
T T T = T = 1 2 1 2
C
C
@ W CQ 1 2
I 0 A>0 (44)
R1=2 Yi 0 I
0 p 1
T 2QCW 1=2 YiT R1=2 YjT R1=2
BB p 1=2 C
C
BB 2W CQ I 0 0 C
C >0 (45)
B@ R1=2 Yi 0 I 0 C
A
R1=2 Yj 0 0 I
where T = ?Ai Q ? Aj Q ? QATi ? QATj ? Bi Yj ? YjT Bi ? Bj Yi ? Yi Bj
And the performance measure J will be less than . The parameters of the fuzzy optimal controller
are given by Fi = Yi Q?1 and the Lyapunov function is V (x) = xT Q?1 x
Remark: For the linear system case, the above LMIs become
AT P + PA + F T B T P + PBF + C T WC + F T RF < 0
16
It can be proven that the feasibility of this inequality is equivalent the solvability of the ARE
(Algebraic Riccati Equation) [12]
PA + AT P + PBR?1B T P + C T WC = 0
That is, our approach and the classical optimal control coincide for the linear system case .
Remark: Other performance speci cations such as decay rate, constraint on input, constraint on
output can also be incorporated in the design framework just as in the discrete system case.
 Decay Rate: The condition V_ (x(t))  ? V (x(t)) is equivalent to:
QATi + AiQ + BiYi + YiT BiT + Q  0; i = 1; : : : ; r (46)
QATi + Ai Q + QATj + Aj Q + Bi Yj + YjT BiT + Bj Yi + YiT BjT + 2 Q  0; i < j  r: (47)
 Input Constraint: Assume that the initial condition x(0) is known, the constraint ku(t)k  2

umax is enforced at all time t  0 if (31) holds true.


 Output Constraint: The constraint ky(t)k  ymax is enforced at all time t  0 if (48) holds
2

true. 2 3
Q C T
4 l 5  0; 8l (48)
Cl ymax
2
I

5 TORA Example
TORA (translational oscillation by a rotational actuator) is a nonlinear benchmark problem corre-
sponding to the physical system which has been built by Dr. Bernstein and his co-workers at the
University of Michigan [1] [6] [7]. It is illustrated in Fig. 1. The problem considers a translational
oscillator with an attached eccentric rotational proof mass actuator, where the nonlinear coupling
between the rotational motion of the actuator and the translational motion of the oscillator pro-
vides the control mechanism. The behavior of TORA is similar to that of a dual-spin spacecraft
exhibiting the resonance phenomenon.
We let x1 be the normalized displacement of the platform from the equilibrium position, x2 = x_1 ,
x3 =  be the angle of the rotor, and x4 = x_3 . Then, the system dynamics can be expressed as the
following equation:

x_ = f (x) + g(x)u (49)


17
M
Ν
e
θ

Translational Movement

Figure 1: Nonlinear Benchmark Problem

where u is the torque applied to the eccentric mass.


2 3
x2
66 ?x1 +x24 sin x3 77
f (x) = 666 1?2 cos2 x3 77
75
4 x4
 cos x3 (x1 ?x24 sin x3 )
1?2 cos2 x3

2 3
0
66 ? cos x3 77
g(x) = 666 1 ?2 cos2 x3 77
75
4 0
1
1 ?2 cos2 x3
The equilibrium point of this system could be any point [0; 0; x3 ; 0] among which only the
point [0; 0; 0; 0] is the desired equilibrium point. The linearization around the point [0; 0; 0; 0] has
two eigenvalues i, which means that the TORA system is a critical nonlinear system. Many
approaches have been proposed to control this system. For example, the following controller can
be obtained based in feedback passivation [1] [6]:
u = 2 y22 sin y1 cosy1 ? 3 cos2 y1(z1 ?  sin y1 ) ? (1 ? 2 cos2 y1 )(k1 y1 + k2y2 ) (50)
In this paper, the proposed fuzzy optimal control will be applied to the TORA system. It will
be shown that this approach is simpler and more straightforward than other approaches. First, a
fuzzy T-S model is constructed for the TORA system. The resulting T-S model consists of four
fuzzy rules:
Rule 1: IF jx3 (t)j is near 0,

18
Then 2 3 2 3
66 0 1 0 0
7
7 6
6
0
2 7
7
?
A = 666 ?
1
2 0  0 77 ; B = 66 1? 77?
0 0 1 75 64 0 75
1

4 0
 0 0 0 1
1?2 1?2

Rule 2: IF jx3 (t)j is near 2 and jx4 (t)j is small,


Then 2 3 2 3
0 1 0 0 0
66 7
7 6
6 77
A = 666
? 1 0 0: 01 2
 0 77 ; B = 66 77 0
4 0 0 0 1 75 64 0 75
0 0 0 0 1
Rule 3: IF jx3 (t)j is near 2 and jx4 (t)j is big,
Then 2 3 2 3
0 1 0 0 0
66 7 6
66 0 777
6 ? 1 0 2 0 77
A = 66 7 ; B = 66 77
4 0 0 0 1 75 405
0 0 0 0 1
Rule 4: IF jx3 (t)j is near ,
Then 2 3 2 3
0 1 0 0 0
66 1 77 66  77
?
6
A = 66 1? 2 0 0 0
77 ; B = 666 ? 1?2 777
7
4 0 0 0 15 4 0 5
 0 0 0 1
1?2 1?2

De ne z1 as jx3 (t)j and z2 as jx4 (t)j, then the membership function can be chosen as following:
8 ? x
>
< x 0x  1
sin 3
3
M (jx j) = M (jx j) > ?  1
2 3 2
21 3 31
:  xx   x  
3

2
sin 3
3 2 3

8
< 1?M 0 x  
M (jx j) = :
11 3
21
 x 
3 2

0 2 3

and 8
<0 0x  
M (jx j) = : 3 2

1?M  x 
41 3
21 2 3

19
and
M22 (jx4 j) = x24 0  jx4 j  1
M32 (jx4 j) = 1 ? x24 0  jx4 j  1
Remark: The T-S model above is constructed through the sector bounds on the nonlinearities.
It is known that many nonlinear physical systems can be represented as a feedback connection of
a linear dynamical system and a nonlinear element satisfying a sector condition. Therefore, the
above construction procedure for T-S model has a wide area of application.
In this paper, the performance index is chosen as
Z1
J= 100x21 (t) + u2 (t) dt (51)
0

Now the fuzzy optimal controller will be designed for the above T-S model. After solving the LMIs
presented in the previous section, we obtain the following set of feedback gains:
F1 = [ 0:8910 ?0:3049 ?1:2311 ?0:8615 ]
F2 = [ 0:9979 ?0:3266 ?1:2508 ?0:9139 ]
F3 = [ 0:9914 ?0:3066 ?1:2360 ?0:9073 ]
F4 = [ 0:9847 ?0:2873 ?1:2177 ?0:8553 ]
The resulting fuzzy optimal controller is now applied to the TORA system and simulation results
are shown in Figure 2. Simulation results using passivation control [6] and classical PDC control
[13] are also shown in Figures 3 and 4, respectively, for comparison.
The performance index for these three control approaches are also calculated and shown in the
table below. The results in this table demonstrate the e ectiveness of the proposed fuzzy optimal
controller towards achieving the optimal control objective.

Passivation Control PDC Control Fuzzy Optimal Control


Performance Index 2.7957 3.5412 1.6343

6 Conclusions
In this paper, a new scheme of fuzzy optimal control of nonlinear systems is proposed based on the
fuzzy T-S model together with the PDC controller structure. The controller is designed by solving
the optimization problem that minimizes the upper bound of a given quadratic performance index.

20
0.03 0.08

0.06
0.02
0.04
0.01

x1

2
0.02

x
0
0
−0.01
−0.02

−0.02 −0.04
0 20 40 60 0 20 40 60
t t

1.2 0.5

1 0
0.8
−0.5
0.6
3

4
−1
x

x
0.4
−1.5
0.2

0 −2

−0.2 −2.5
0 20 40 60 0 20 40 60
t t

Figure 2: Fuzzy Optimal Control of the TORA System

This technique has been applied to a nonlinear control benchmark problem. The control result
shows that this approach is both simple and e ective.

References
[1] R. T. Bupp, D. S. Bernstein and V. T. Coppola, \ A benchmark problem for nonlinear control
design: problem statement,experiment testbed, and passive nonlinear compensation," in Proc.
1995 American Control Conference, 1995, pp. 4363-4368.
[2] S.G. Cao and N.W. Rees, \Identi cation of dynamic fuzzy models," Fuzzy Sets and Systems,
vol. 74, no. 2, pp. 307-320, 1995.
[3] S.G. Cao, N.W. Rees and G. Feng, \Fuzzy control of nonlinear continuous-time systems," in
Proc. 35th IEEE Conf. Decision and Control, New York, NY, 1996, pp. 592-597.
[4] C. Fantuzzi and R. Rovatti, \On the approximation capabilities of the homogeneous Takagi-
Sugeno model," in Proc. FUZZ-IEEE'96, 1996, pp. 1067-1072.
[5] S. K. Hong and R. Langari, \Synthesis of an LMI-based fuzzy control system with guaranteed
optimal H1 performance," Proc. FUZZ-IEEE'98, 1998, pp. 422-427.
[6] M. Jankovic, D. Fontaine, P. Kokotovic, \TORA example: cascade and passivity control
design," IEEE Trans. on Control Systems Technology, vol. 4, no. 3, 1996

21
0.03 0.02

0.02
0.01
0.01
x1

2
0 0

x
−0.01
−0.01
−0.02

−0.03 −0.02
0 20 40 60 0 20 40 60
t t

1 0.4

0.2
0.5

0
3

4
0
x

x −0.2

−0.5
−0.4

−1 −0.6
0 20 40 60 0 20 40 60
t t

Figure 3: Passivation Control of the TORA System

0.06 0.06

0.04 0.04

0.02 0.02
x1

0 0
x

−0.02 −0.02

−0.04 −0.04

−0.06 −0.06
0 20 40 60 0 20 40 60
t t

1 0.2

0.8
0
0.6

0.4 −0.2
3

4
x

0.2 −0.4
0
−0.6
−0.2

−0.4 −0.8
0 20 40 60 0 20 40 60
t t

Figure 4: Classical PDC Control of the TORA System

22
[7] I. Kanellakopoulos and J. X. Zhao, \Tracking and disturbance rejection for the benchmark
nonlinear control problem," in Proc. 1995 American Control Conference, 1995, pp. 4360-4362.
[8] G. Kang, W. Lee and M. Sugeno, \Design of TSK fuzzy controller based on TSK fuzzy model
using pole placement," in Proc. FUZZ-IEEE'98, 1998, pp. 246-251.
[9] J. Li, D. Niemann, H.O. Wang and K. Tanaka, \Parallel distributed compensation for Takagi-
Sugeno fuzzy models: multi-objective controller design," in Proc. American Control Confer-
ence, San Diego, CA, June 1999, pp. 1832-1836.
[10] J. Li, H.O. Wang, D. Niemann and K. Tanaka, \Dynamic parallel distributed compensation
for Takagi-Sugeno fuzzy models," Information Sciences, no. 123, pp. 201-221, 2000.
[11] J. Li, H.O. Wang, D. Niemann and K. Tanaka, \T-S Fuzzy model with linear rule consequence
and PDC controller: a universal framework for nonlinear control systems," in Proc. 9th IEEE
Int. Conf. on Fuzzy Systems, May 2000.
[12] J. Li, H.O. Wang and L. Bushnell, \On the Relationship Between LMIs and AREs: Applica-
tions to Absolute Stability Criteria, Robustness Analysis and Optimal Control," submitted to
CDC 2000.
[13] J. Li, H.O. Wang and K. Tanaka, \Stable fuzzy control of the benchmark nonlinear control
problem: a system-theoretic approach," in Joint Conf. of Information Science, 1997, pp. 263-
266.
[14] C. J. Lin and C. T. Lin, \An ART-based fuzzy adaptive learning control network," IEEE
Trans. on Fuzzy Systems, vol. 5, no. 4, pp. 477-496, 1997.
[15] W.M. Lu and J. C. Doyle, \H1 control of nonlinear systems: a convex characterization,"
IEEE Trans. on Automatic Control, vol. 40, no. 9, pp. 1668-1675, 1995.
[16] D. L. Lukes, \Optimal regulation of nonlinear dynamical systems," SIAM J. Control, vol. 7,
pp. 75-100, 1969.
[17] A. J. van der Schaft, \On a state space approach to nonlinear H1 control," Systems and
Control Letters, vol. 16, pp.1-8, 1991
[18] T. Takagi and M. Sugeno, \Fuzzy identi cation of systems and its applications to modeling
and control," IEEE Trans. on SMC, vol. 15, no. 1, pp. 116-132, 1985
23
[19] K. Tanaka, T. Ikeda and H.O. Wang, \Robust stabilization of a class of uncertain nonlinear
system via fuzzy control," IEEE Trans. on Fuzzy Systems, vol. 4, no. 1,pp. 1-13,1996.
[20] K. Tanaka, T. Ikeda and H.O. Wang, \Fuzzy regulators and fuzzy observers: relaxed stability
conditions and LMI-based designs," IEEE Trans. on Fuzzy Systems, vol. 6, no. 2, pp. 250-265,
1998.
[21] K. Tanaka and M. Sano, \Fuzzy stability criterion of a class of nonlinear systems," Information
Sciences, vol. 71, no. 1,2, pp. 3-26, 1993.
[22] 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. on Fuzzy Systems, vol. 2,
no. 2, pp. 119-134, 1994.
[23] K. Tanaka and M. Sugeno, \Stability analysis and design of fuzzy control systems," Fuzzy Sets
and Systems, vol. 45, no. 2, pp. 135-156, 1992
[24] H.O. Wang, K. Tanaka and M. Grin, \An approach to fuzzy control of nonlinear systems:
stability and design issues," IEEE Trans. on Fuzzy Systems, vol. 4, no. 1, pp. 14-23, 1996.
[25] H. Ying, \Sucient conditions on uniform approximation of multivariate functions by general
Takagi-Sugeno fuzzy systems with linear rule consequence," IEEE Trans. SMC, vol. 28, no. 4,
pp. 515-521, 1998.
[26] J. Zhao, V. Wertz and R. Gorez, \Fuzzy gain scheduling controllers based on fuzzy models,"
in Proc. of the FUZZ-IEEE'96, New Orleans, LO, 1996, pp. 1670-1676.
[27] S. Boyd, L. E. Ghaoui, E. Feron and V. Balakrishnan, Linear Matrix Inequalities in Systems
and Control Theory, SIAM, Philadelphia, PA, 1994.
[28] P. Gahinet, A. Nemirovski, A. J. Laub and M. Chilali, LMI Control Toolbox, The Math Works,
1995.
[29] D. E. Kirk, Optimal Control Theory: An introduction, Prentice Hall, 1970.

24

View publication stats

Vous aimerez peut-être aussi