Fourth International Multi-Conference on Systems, Signals & Devices Volume I : Conference on

March 19-22, 2007 – Hammamet, Tunisia Systems Analysis & Automatic Control

Discrete Fuzzy Adaptive Control for a class of SISO Nonlinear Systems

Hafedh Abid (1-3) Mohamed Chtourou (2) Ahmed Toumi (3)

(1) Institut Supérieur des Etudes Technologiques de Sfax

(2) Unité de Commande Intelligente, design et Optimisation des Systèmes complexes
(ICOS) ENIS, B.P. W, 3038 Sfax, Tunisie
(3) Unité de Commande de Procédés Industriels (UCPI) ENIS, B.P. W, 3038 Sfax, Tunisie

E-mail [email protected] [email protected] [email protected]

Abstract: In this work, we are interested to the Z.D .Yuan and S.H. Hou [18] proposed an indirect
discrete fuzzy adaptive control for a class of SISO adaptive controller based on fuzzy logic systems to
nonlinear systems. . The model of the system is provide tracking of a reference signal for a class of
represented by Takagi-Sugeno type fuzzy system. discrete-time nonlinear systems. Fuzzy logic system
The sub-models are identified on-line based on a is used as the system model of the plant dynamics
recursive algorithm. A Lyapunov function is used to that are unknown and based on the model. The
prove the stability of the adaptive algorithm. The adaptive controller is then constructed.
indirect adaptive control techniques are used to Chung-Shi Tseng [8] proposed a model reference
compute at each time the control law. Inverted fuzzy tracking control design for non linear discrete-
pendulum is used to check performance of the time systems with time-delay. The tracking control
proposed fuzzy adaptive control scheme. design problem is parameterized in terms of a linear
matrix inequality problem.
Key words: Nonlinear system; Takagi-Sugno fuzzy
system; Discrete system; Adaptive control; Tracking In this work, which is dealing with indirect discrete
error; Stability. fuzzy adaptive control law, we use a a recursive
algorithm to identify on-line the sub-system matrix.
1. Introduction A Lyapunov function is used to prove the stability of
system. This work is organized as follows: In section
The adaptive control goes up again to the middle of II, we present the problem statement and Takagi-
the twentieth century. The works of research showed Sugeno (T-S) type fuzzy dynamic model [2]. The
that a controller with the stationary parameters is not third section is reserved to the indirect adaptive
able to assure the wished performances, in the case control laws and the proof of the proposed theorem.
where parameters of the system varied in time. The Before conclude simulation results are presented.
objective of the adaptation law is to adjust the
controller's parameters in order to make the tracking 2. Problem statement and fuzzy system
error toward zero. There exist two families of
adaptive control: the direct control and the indirect Consider a class of SISO nonlinear systems
control. Several theories have been used to reach this described by the following equations:
objective such as adaptive neural control [20], [21],
fuzzy-neural adaptive control [16], [15], [14] or  x(k + 1 ) = f ( x ( k )) + g(x(k))u(k)
fuzzy control [3] [4] [6], [7]. The fuzzy controllers  (1)
have the property of universal approximator for any  y (k ) = C x(k )
continuous function [12]. In the fuzzy adaptive Where, C = [1,0,0, …,0], x(k) ∈ℜn is the state
control literature, Wang and al. [3], and Chan and al. vector that is assumed to be observable

( )
[4] proposed an indirect fuzzy adaptive control law T
for system described in continuous form. They x (k ) = x1 ( k ) ,x2 (k ) ,...,xn ( k ) , u(k)∈ℜ is
proposed adaptive laws based on sliding mode.
Young and al. [6] proposed an indirect fuzzy the control input at time (k).
adaptive control with model of reference (MRAFC), In order for (1) to be controllable, it is required that
they used the fuzzy system of Takagi-Sugeno type. g(x,k)≠0.
Ahmed Rubaai [17] proposed a direct adaptive fuzzy The nominal control law is:
control technique to achieve H∞ tracking for high
performance servo drives with external disturbances.
u(k) = 1
g(x(k))
( x(k + 1 ) − f ( x(k )) ) (2)
The proposed methodology combines the attenuation Our objective is to design a control law which can
technique, fuzzy logic approximation method, and guarantee:
adaptive control algorithm for the robust tracking - The stability of the closed loop system in the sense
control design of the nonlinear servo drives, with that all signals are bounded,
unknown parameter variations and structures.

ISBN 978-9973-959-06-5 / SSD © 2007 1

- The tracking error converges asymptotically to The design objective is to choose the control law
zero. such that all signals in the closed-loop plant are
bounded and the plant state x(k) tracks the state xm(k)
Takagi-Sugeno fuzzy model of the reference model which is described by:

The discrete SISO non-linear system (1) can be  x m (k + 1) = A m x m ( k ) + Bm r(k)

presented by the Takagi-Sugeno fuzzy logic model.  (5)
The fuzzy system is represented by a collection of  y (k) = C x (k) ;
fuzzy IF-THEN rules of the form:
R (i) : IF x 1 is M 1i AND ... AND x n is M ni where,

( )
T
T HEN x ( k + 1) = Ai x ( k ) + Bi u (k ) (3) xm (k ) = xm1 ( k ) ,xm 2 (k ) ,...,xmn ( k ) , is the
where, Ai and Bi are respectively the state matrix and
input vector of the ith sub-system. They will be state vector of reference model, Am and Bm are
written as: respectively a stable matrix and input vector, they
will be written as:
 1 Te 0 ... 0 
 0 1 Te ... 0   1 Te 0 ... 0 
  0
Ai =   ;  1 Te ... 0 
  Am =   ;
 0 0 0 ... Te 
 
 − ai1 − ai 2 − ai 3 ... − a in   0 0 0 ... Te 
 − a m1 − am 2 − am 3 ... − a mn 

Bi = [0 ,0 ,...,b in ]T
B m = [ 0 ,0 ,...,b mn ]T
The output of the fuzzy system can be inferred as:
The reference signal r(k) are chosen such that xm (k)

( )
c represents the desired state response of plant.
 x(k + 1) = ∑ hi ( z (k )) Ai x(k ) + Bi u(k) (4)
 i =1 The state equation of adjustable fuzzy model is
 y(k) = Cx(k) ;
 given by:


( )
For each rule Ri is attributed a weight wi (z(k)) which c
depends on grade of membership function of  x (k + 1) = ∑ hi ( z (k )) Ai x(k ) + Bi u (k )
premise variables zj(k) in fuzzy sets Mij :  i =1
(6)
 y (k ) = Cx(k ) ;
n 
wi ( z(k )) = ∏ ij ( z j ( k)) ;
j =1
where,
µij(zj (k)) is the grade of membership of z j(k) to the x(k ) is the state vector of adjustable model
(x ) ; Ai ( k + 1 )
T
fuzzy set Mij. x(k ) = ( k ), x 2 ( k ), ..., x n ( k )
1
w j (z (k ))
h j (z(k )) = ; 0< hi (z(k)) <1; i = 1, … c ; and Bi ( k + 1 ) are the estimated state matrices and
c
∑ wi (z (k )) vectors respectively for each sub model at the
i =1 discrete time (k+1).
wi (z(k)) >0 ; for i = 1, … c ; We define by E(k) and ε(k) respectively the state
vector prediction error and the state vector error
c c defined by :
∑ w ( z(k)) > 0; ∑ h ( z(k )) = 1
i i
i=1 i =1 E ( k ) = x (k ) − x (k ) (7)

c is the number of fuzzy rules.

It is required that all Bi (k) are different to zero to ε (k) = x(k) − xm (k ) (8)
assure the controllability of (4).
We will drive an adaptive law to make the
3 Indirect adaptive control laws estimation error E(k) equal to zero and with adaptive
law, the estimation model (6) can be the same one as

2
the reference model (5) by adopting the control input Ω(k ) = x m (k + 1) + A m ε (k )
(9). c c

Control law
− ∑∑ h (k )hj (k )(A
i =1 j =1
i i (k ) − Bi (k + 1) K j (k + 1)) x(k )

We proposed the following control law: We use the following recursive algorithm to identify
the parameters, of matrices Ai and vectors Bi , which
u( k ) = u fb ( k ) + u ff ( k ) (9) are susceptible to vary in the time.

Where: ufb represents the state feedback control ∂J ( k )

which is used to stabilise system and uff represents Ai (k + 1) = Ai ( k ) − ηQ (16)
the feed-forward which is used to make the state ∂Ai ( k )
system one to follow the desired state xm(k).
We note hi (z(k)) by hi (k). ∂J ( k )
Bi (k + 1) = Bi ( k ) − ηQ (17)
u fb ( k ) is based on parallel distribute compensation ∂Bi ( k )
(PDC) [19]:
c Where, Q is a specific matrix.
u fb ( k ) = − ∑ h (k ) K
i =1
i i x(k ) (10)
0 . . 0
If we apply the control law (9) to the system (4), we . . . . 
obtain the following equation: Q =  ; J (k ) = E (k )E T (k ) (18)
c 0 . . 0
x (k + 1) = ∑
i =1
h i ( k )A i (k )x(k ) 
1 . . 1

(11)
( )
c
+ ∑
i =1
hi ( k ) B i (k ) (u fb (k ) + u ff ( k )) ~
Ai ( k ) = Ai − Ai (k)
~
The nominal feed-forward control law is given by Ai (k + 1) = Ai − Ai (k + 1)
the following equation (12): ~
Bi ( k ) = Bi − Bi (k) (19)
u ff (k) = (
inv(hi (k)BiT hi (k)Bi )) * (hi (k)BiT ) *
~
Bi (k + 1) = Bi − Bi (k + 1)
 c c  (12)
 x(k +1) −
 ∑∑h (k)hj(k)(A − B K )x(k)
i i i j
 0
 0
0 0 ... 0 
0 
 i=1 j=1   0 0 ...
~
The gain of state feedback is computed such that all Ai = ∆ Ai =  . . . ... .  ( 20 )
eigen values of each sub-system are situated inside  
the unit circle of the complex plan. The gain of state  0 0 0 ... 0 
 ∆ a i1 ∆a i 2 ∆a i 3 ... ∆ a in 
feedback is given by: 
Ki =
1
[(ai1 − a m1 ) . . . (ain − a mn )] (13)
~
bi B i = ∆ B i = [ 0 ,0 ,..., ∆ b i ]T
In fact, Ai, Bi, Kj and x(k+1) are unknown, so we
replace Ai, Bi and Ki by their estimate values.
Theorem

Ki =
1
[(ai1 − am1 ) . . . (ain − amn )] (14)
Let consider a SISO system described by the
equation (1), which can be represented by the T-S
bi
type fuzzy model (4), the adjustable model (6) and
We know that ε (k + 1) − Amε (k ) = 0 so, x(k+1) can be the reference model (5). We assume that the
replaced by xm(k+1) + Am (k) at the sampling time reference input and the state of the reference model
(k+1). are uniformly bounded. If we apply the control law
(9) with the adjustable state matrices and vectors
The feed-forward control law becomes: (16) and (17) of the sub-systems we guarantee that:

(
u ff (k) = inv( hi (k ) BiT (k + 1)hi ( k ) Bi (k + 1)) ) i) E( k ) → 0 ,ε ( k ) → 0 when k → ∞
∂J (k ) ∂J (k )
* (hi (k ) BiT (k + 1)) * Ω(k ) (15) ii) → 0, →0
with: ∂Ai (k ) ∂Bi (k )

3
 ∑ ( ) 
Proof  c ~
∆V ( k + 1 ) =  tr − 2hi ( k )QηE (k ) X T (k − 1) AiT ( k )
We propose the following equation (21) as  i =1 
 
∑ ( )
Lyapunov candidate function: c
+  tr 2η 2 Q 2 hi2 (k ) E 2 (k ) X 2 (k − 1) 
 i =1 
∑ ( )∑ (
1 c 
)
c
~ ~ ~ ~
V( k ) =  tr Ai ( k )AiT ( k ) + tr Bi ( k )BiT ( k )  (21)  c

∑
+  tr (− 2hi ( k )ηQE (k )u (k − 1) )BiT (k )
~
2  i=1 i =1  

 i =1 

∑ ( )
1 c   c 
∑ ( )
2
~ ~
V ( k + 1) =  tr Ai ( k + 1 ) AiT ( k + 1 )  +  tr 2(ηQhi (k ) E ( k )u ( k − 1) 

2  i =1  i =1 
(30)
  
1 c 
∑ ( )
~ ~
+  tr Bi ( k + 1 )BiT ( k + 1 )  (22)

2  i =1 We know that the error equation can be written as:


∑h (k)(A (k)x(k −1) − B (k)u(k −1)) (32)

c
We need to verify that: E(k)= x( k ) − i i i
i=1
∆V( k +1) = V( k +1) −V( k ) p 0 (23)

∑ h ( k )(A (k )x(k − 1) + B (k )u(k − 1)) (33)

c
~ ~
~ ~ ∂J ( k ) E( k ) = i i i
Ai ( k + 1 ) = Ai ( k ) + η Q (24) i =1
∂ Ai ( k )

∂J ( k ) ∆V (k + 1) = −2ηhi (k ) E T (k )QE (k )
~ ~
Bi ( k + 1 ) = Bi ( k ) + ηQ (25)
∂B i ( k ) + 2η 2 hi2 (k ) E T (k ) X T (k − 1) X (k − 1)QQ T E (k )
+ 2η 2 hi2 (k ) E T (k )u (k − 1)u(k − 1)QQ T E (k ) (34)

∑ ( )
1 c
~ ~ 
∆V ( k + 1) =  tr Ai ( k + 1)AiT ( k + 1) 

2  i =1  ∆V (k + 1) = E T (k )φ (k) E ( k ) (35)
1 
∑ ( )
c
~ ~ (26)
+  tr Bi ( k + 1)BiT ( k + 1)  with:
2  i =1 
(
φ ( k ) = − 2ηhi ( k )Q + 2η 2 hi2 (k ) ρ 2 (k − 1)QQ T )
∑( ) ∑(
1 c ~ 
)
c
~ ~ ~
−  tr Ai ( k )AiT ( k ) + tr Bi ( k )BiT ( k ) 

2  i =1 i =1  (
ρ 2( k −1) = X T ( k −1 )X ( k −1 ) + u( k −1 )2 ) (36)

 
1 c  ~ 
2
∂J  ∂J   We know that the bigger eigenvalue of the matrix Q
∆V ( k + 1 ) =  tr  2ηAiT ( k )Q
∑ + ηQ 


  and QQT is the unit "1". So, we can write:
2  i =1 ∂Ai  ∂Ai
     (
∆V (k + 1) p 0 ⇒ −1 + ηhi (k )ρ 2 (k −1) p 0 (37) )
 
1  c  ~T ∂J   if hi (k)= 0, ∆V (k + 1) p 0 ;
2
∂J  
∑
+  tr 2ηBi ( k )Q
2  i =1 
+ ηQ Q
∂Bi  ∂Bi   
(27)
    (−1 +ηh (k)ρ (k −1)) p 0
i
2

∂J (k )
= − 2 hi ( k ) E ( k ) X T ( k − 1) (28) 1
∂ Ai ( k ) if hi (k ) ≠ 0; 0 pη p (38)
hi (k )ρ 2 (k −1)
∂J ( k )
= − 2 hi ( k ) E ( k ) u ( k − 1) (29) So, it leads to ∆V (k + 1) p 0 .
∂ Bi ( k )
So, η can be chosen variable in time, such that the
By using the following properties of trace,
equation (38) is verified at each sampling period.
i) tr(AB) = tr (BA) 4. Illustration:
ii) tr(A + B) = tr (A) + tr(B) for any A,B ∈ℜnxn To illustrate the performance of the presented
approach, we choose inverted pendulum which is
iii) tr(yxT ) = xT y for any x, y ∈ℜnx1 widely used in the control literature of nonlinear

4
system. The equations of system in continuous form
are given by (39)-(40)-(41) [4]:

 x&1(t) = x 2 (t )

 x& 2 (t) = f ( x1 ,x 2 ) + g ( x1 ,x 2 )u (39)
 y(t) = x (t )
 1

Where f(x1, x2) and g(x1, x2) are expressed as :

mlx22 sin x1 cos x1 − (m + M)gsin x1

f(x1,x2 ) = (40)
m l cos2 x1 − 4l (m + M)
3
Figure 1: Evolution of x1 (k ), x m1 (k ) and x1 (k ) .

cos x 1 (41)
g(x 1 ,x 2 ) =
m l cos 2
x 1 − 4 l 3 (m + M)

x1 is the angle in radium of the pendulum from the

vertical axis; x2 is the angular velocity in rad/s; g is
the gravity acceleration; m and 2l are respectively
the mass and the length of the pendulum; M is the
mass of the cart and u is the force applied to the cart.
The nominal values of the parameters are:
g= 9.81m/s2, m = 0.1kg, M = 1kg, 2 l =1m.
x i ∈ ]− π / 2 ; π / 2 [

The membership functions are:

Figure 2: Evolution of x 2 (k ), x m 2 (k ) and x 2 (k ) .

xi (k) x (k )
f1( xi (k)) = 1− ; f2 (xi (k )) = i
π π
2 2

The state matrices and input vectors for sub-systems

are:
 1 0.01  0 
A1 =  B1 = 
0.1729 0   ;
0.0018

 1 0.01  0 
A2 =  B2 = 
0.0936 0   ;
0.000052 
The state matrix and input vector for model Figure 3: Evolution of the error 1
reference are:

 1 0.01   0
Am =   ; Bm =   ;
- 3.275 0.655  0.01
We present in the figures 1, 2, 3 and 4 the simulation
results of the behavior of variables state x1(k), x2(k),
1(k) and 2(k) respectively of inverted pendulum
with parameters slowly varying in time. The
parameter of pendulum varies between the sampling
time 50 seconds and 110 seconds.
The initial conditions are given by:
x(0) = [- π 60 ; 0 ] ,
Figure 4: Evolution of the error 2

5
Commentary Learning Control for Nonlinear Plants With Initial
State Errors" IEEE Transactions on Fuzzy systems
We notice that the state variables of the system
vol 12.N°5, October 2004 pp(724-732).
follow appropriately the reference signal and that the
[11] H. -J.Kang, C. Kwon, H.-J.Lee, M. Park ,“
tracking error converges toward zero. In the first
Robust stability analysis and design method for the
steps the errors are very important but, they decrease
fuzzy feedback linearization regulator" IEEE Trans.
during the time to reach zero.
Fuzzy systems vol 6.N°4 1988 pp(464-472).
5. Conclusion [12] Hao Ying " General SISO Takagi-Sugeno Fuzzy
Systems with Liner rule Consequent Are Universal
In this work, we have used the indirect discrete Approximators" IEEE Transactions on fuzzy
fuzzy adaptive control for the Takagi-Sugeno type systems vol 6.N°4 November 1998 pp(582-587).
fuzzy system. The control law contains two terms, [13] Kazuo Tanaka, Takayuki Ikedaand Hua
the first uses parallel distributed compensation to O.Wang. "Robust stabilization of a class of
stabilise system and the second uses feed forward to uncertain nonlinear systems via fuzzy control:
guarantee the convergence to zero of tracking error.
Quadratic stability, H∞ Control Theory, and linear
The strong point of this approach is that the state
Matrix Inequalities". IEEE Trans on Fuzzy Syst,Vol
matrices of T-S fuzzy model are estimated on-line
4 N°1 Feb 1996.
by an appropriate recursive algorithm. The proposed
[14] Feipeng Da and WenzhongSong "Fuzzy Neural
method guarantees the stability of the closed loop
Networks for Direct Adaptive Control" IEEE
system with parameters slowly varying in time.
Transactions on Fuzzy Systems,Vol 50 N°3 June
2003 pp(507-513).
Bibliographies
[15] Chi-Hsu wang, Han-Leih Liu and Tsung-Chih
[1] Michio Sugeno “On Stability of fuzzy Systems Lin. "Direct Adaptive Fuzzy-Neural Control With
expressed by rules with singleton consequents”. State Observer and Supervisory Controller for
IEEE Trans on Fuzzy Syst,Vol 7 N°2 Feb 1999. Unknown Nonlinear Dynamical Systems". IEEE
[2] T.Takagi and M.Sugeno. "Fuzzy identification of Transactions on Fuzzy Systems, Vol 10 N°1
systems and its applications to modelling and February 2002 pp(39-49).
control", IEEE trans syst. Man,Cybern. Vol15. 116- [16] Hazem N. Nounou and Kevin M. Passino
132 Jan/Feb1985. "Stable Auto-Tuning of Adaptive Fuzzy/Neural
[3] J. Wang, A.B.Rad,P.T.chan "Indirect adaptive Controllers for Nonlinear Discrete-Time Systems"
fuzzy sliding mode control: Part I: Fuzzy switching" IEEE transaction on Fuzzy systems, volume 12, N°
Fuzzy Sets and Systems 12 (2001) pp21-30. February 2004 pp(70-83).
[4] P.T.Chan,A.B.Rad,J.Wang, "Indirect adaptive [17] Ahmed Rubaai, "Direct Adaptive Fuzzy
fuzzy sliding mode control: Part II: Parameter Control Design Achieving H∞ Tracking For high
projection and supervisory control" Fuzzy Sets and Performance Servo Drives". IEEE Transactions on
Systems 122 (2001) pp31-43. Energy Conversion, Vol. 14, No. 4, December 1999.
[5] Shaocheng Ton, Tao wang, Jian Tao Tang [18]Z.D .Yuan and S.H. Hou "Indirect Adaptive
"Fuzzy adaptive output tracking control of nonlinear Fuzzy Control of Discrete-time Nonlinear Systems"
systems" Fuzzy Sets and Systems 111 (2000) pp169- Proceedings of the 3" World Congress on Intelligent
182. Control and Automation June 28-July 2,2000, tlefei,
[6]Young-wan cho,chang-woo Park and Mignon P.R. China pp(1770-1774)
Park. "An Indirect model reference adaptive fuzzy [19]Huao.wang, Kazuo Tanaka and Michael
control for SISO Takagi-Sugeno fuzzy model”. Fuzzy F.Griffin. "An approch to fuzzy control of nonlinear
sets and systems 131 (2002) pp (197-215). systems: stability and design Issues". IEEE
[7] Mehrdad Hojati and Saeed Gazor “Hybrid Transactions on Fuzzy Systems,Vol 4 N°1 February
Adaptive Fuzzy Identification and Control nonlinear 1996.
Systems” IEEE Transactions ON Fuzzy Systems vol [20] Shuzhi Sam Ge and and Jing Wang
10, N°2. April 2002 pp 198-210. "Robust Adaptive Neural Control for a Class of
[8]Chung-Shi Tseng "Model Reference Output Perturbed Strict Feedback Nonlinear Systems".
Feedback Fuzzy Tracking Control Design for IEEE Transactions on neural networks, vol. 13, N° 6,
Nonlinear Discrete-Time Systems With time-Delay" November 2002 pp (1409- 1419)
IEEE transaction on Fuzzy systems, volume 14, N°
February 2006 pp(58-70).
[9]Tsung-Chih lin , Chi-Hsu Wng nd Han-Leih Liu
“Observer-Based Indirect Adaptive Fuzzy Neural
Tracking Control for nonlinear SISO Systems using
VSS and H∞ approaches ” Fuzzy Sets and systems
143 (2004) pp(211-232).
[10] Chiang-Ju Chien,, Chun-Te Hsu, and Chia-Yu
Yao,“ Fuzzy System-Based Adaptive Iterative