8th International IFAC Symposium on Dynamics and Control of Process Systems

Preprints Vol.2, June 6-8, 2007, Cancn, Mexico

STABLE INDIRECT ADAPTIVE FUZZY CONTROL BASED ON TAKAGI-SUGENO MODEL Ruiyun Qia,*, Mietek A. Brdysa
a

Department of Electronic Electrical & Computer Engineering, University of Birmingham, Edgbaston, Birmingham B15 2TT, UK

Abstract: This paper presents an indirect adaptive fuzzy control scheme for nonlinear uncertain stable plants with unmeasurable states. A discrete-time T-S fuzzy model is employed as a dynamic model of an unknown plant. Based on this model, a feedback linearization controller is designed and applied to both the model and the plant. Parameters of the model are updated on-line to allow for partially unknown and timevarying plants. Stability analysis shows that the adaptive controller guarantees the boundedness of all the closed-loop signals and achieves bounded tracking error. In the ideal case where there is no modelling error and the signal for parameter learning is persistently exciting, perfect tracking is ensured. The effectiveness of the method is verified by simulation examples. Copyright 2007 IFAC

Keywords: nonlinear control, adaptive control, fuzzy control,.

1. INTRODUCTION Adaptive control has become very popular in many fields of control engineering and attracted much attention in developing advanced applications. Adaptive control is based on feedback of signals in a controlled system for control adaptation to effectively handle system uncertainties (Tao, 2003). The adaptive control can be divided into two classes: direct adaptive control and indirect adaptive control. The former tunes the parameters of the controller while the latter updates the parameters of the model. This paper focuses on indirect adaptive control, which means the T-S fuzzy model is adapted on-line. Design of provably stable indirect adaptive fuzzy controller posed more difficult theoretical questions. Some of the early adaptive control schemes using continuous-time T-S fuzzy model (Takagi &Sugeno, 1985) were presented in the papers (Wang, 1994;
*

Corresponding author. Tel: 44-1214143150; fax: 44-121 4144291. E-mail address: [email protected] [email protected] (M.A. Brdys) (Ruiyun Qi),

Spooner 1996; Yoo, 1998) with stability results. Wangs control law is composed of a certainty equivalence control and a supervisory controller which is used to force the tracking error not to exceed certain error ball. Spooner and Passinos approach consists of a certainty equivalence control term, a bounding control term used to restrict the plant output trajectory, and a sliding mode control term to compensate the approximation errors. Their parameter adaptation algorithms are chosen to make the derivative of a certain Lyapunov equation nonpositive. Recent adaptive fuzzy control schemes for continuous-time system can be found in the papers (Golea2003; Park & Cho, 2004). In Park and Cho (2004), an adaptive law for updating parameters is designed based the Lyapunov theory and with the on-line parameter estimator, any type of fuzzy controllers works adaptively to the parameter perturbations. All the above adaptive control schemes assume that full states of the controlled plant are available for measurement. Adaptive output feedback control for nonlinear systems when full states are not measurable remains one of the outstanding problems. Kulawski and Brdys (2000) present an adaptive

385

control scheme for uncertain nonlinear plants with unmeasurable states based on dynamic neural networks. In their approach, a recurrent neural network is employed as a dynamical model of the plant and the control law is designed by using the states of the neural network instead of the plant. Our proposal in this paper is inspired by the above consideration and applies it to discrete-time adaptive fuzzy control system. A discrete-time T-S fuzzy model is trained to approximate the input-output behaviour of an unknown continuous-time plant by using I/O data from the plant. Then an appropriate control law is calculated completely based on the model, using the model structure, parameters and states. The control input is applied to both the plant and the model. The parameters of the model are updated on-line by recursive least square estimation (RLSE) method, which is fast and is able to ensure the boundedness of parameters (Tao, 2003). With this approach, at least the model can be stabilized and provides bounded control input without extra supervisory controller. The sufficient conditions for the convergence of tracking error are derived for stable plant. The paper is organized as follows. The control algorithm including parameter adaptation law is presented in Section 2. Stability of the closed-loop system and convergence of the tracking error are proven in Section 3. Section 4 shows the simulation results on control of a single-link robot manipulator with parameter disturbance. Finally, Section 5 concludes the paper. 2. CONTROL ALGORITHM 2.1 Modelling for adaptive control It has been shown that T-S fuzzy systems with piecewise linear rule consequents are universal approximators to approximate any continuous nonlinear system to an arbitrary accuracy (Ying, 1998 & 1999) if all fuzzy membership functions are Gaussian and the t-norm is the multiplication. In this paper, the T-S fuzzy model is used to express a real nonlinear plant. That is, the nonlinear plant can be represented by the following fuzzy rules: i R i : If x1 (k ) is M1i and L and xn (k ) is M n , then (1) i i xn (k + 1) = q 0i + q1i x1 (k ) + L + q n xn (k ) + q ni +1u (k )
i

y (k - 1), y (k )]T , then the overall T-S fuzzy model is calculated as

output relation in the local operating region. Let x (k ) = [ x1 (k ), x2 (k ),L , xn (k )]T = [ y (k - n + 1),L,

= Y T ( x ( k ) | c , )Q y ( k ) = xn ( k ) where

(3)

w i ( x | c i , i ) = M ij ( x j | c ij , s ij ),
j =1

w ( x | c, ) = w ( x | c i , i )
i i i j i j

w ( x | c, ), c = {c | i = 1, L, N ; j = 1,L, n} = {s | i = 1, L , N ; j = 1,L , n}
i i =1

(4) ,
T

T i i ( 1 )T i = q0 L q n +1 ,Q = T xe (k ) = 1 x (k ) u (k ) , 1 T YT ( x | c , ) = w xe T w 2 xe T

( )
T

N T

T L w N xe .

An assumption is made here. Assumption 1. The modelling error is negligible and there exist parameter values (c , ) and Q that make the model (3) to become a perfect representation of the real plant. From now on, the nonlinear plant is completely represented by (3). Since all the parameters in (3) are unknown, we define an estimation model as
T T i 1 x |c i , n (k + 1) = w i ( x i ) x (k ) u (k ) i =1 N

T (x (k ) (k ) | c (k ), (k ))Q =Y (k ) = x n (k ) y

(5)

(k ) are the estimates of (k ), (k )) and Q where (c (k ) n is the model state (c , ) and Q at time k. x vector. 2.2 Control synthesis (k + 1) e
z -1 M

where R ( i = 1, 2,L , N ) denotes the ith fuzzy rule and N is the number of rules. x1 (k ), L, xn (k ) are the premise variables which are the past plant output measurements and u (k ) is the control input. M ij ( i = 1,L , n ) are fuzzy sets which are described by Gaussian functions ( x - c i )2 M ij ( x j | c ij , s ij ) = exp - j i j2 (2) 2(s j ) The model parameter values q ij , c i = 1,L , N , j = 1,L , n are unknown. The T-S fuzzy model approximates a nonlinear system by combining a group of local affine models. The premise of a fuzzy rule defines a local operating region and each consequent describes a local input386
i j

(k ) e
kT

z -1
(k ) x u(k )

yr (k + 1)

v (k )

( k + 1) y

x (k )

z -1
y ( k + 1)

Fig. 1 Adaptive control based on T-S fuzzy model The control objective is to make the output of the nonlinear plant (3) tracking a specified reference trajectory yr . Assumption 2. The reference yr (k ) satisfies y r (k ) U (6)

and s ij ,

where y r ( k ) = [ yr ( k - n + 1) L yr ( k ) ] and U is a known bound. The overall control structure is shown in Fig. 1. It consists of three parts: a nonlinear uncertain plant, an adaptive T-S fuzzy model and a model based feedback linearization controller.
T

The control law is designed based on the model (5) whose evolution of output can be expressed as
T T i (k ) 1 x (k ) | c i , (k + 1) = w i ( x i ) y (k ) u(k ) i =1 N T T i (k ) 1 x (k ) | c i , i ) = wi ( x f (k ) i =1 N

2005). The initial model is good enough for the model based controller to produce proper control signal in the beginning of closed-loop control. However, the plant may have external disturbances and parameters disturbances so that the initial model needs to be updated. Therefore, the model should be adaptive when plant changes. The parameters of the model can be adjusted on-line to provide an adaptive model. (k ), (k )) , Q can be For fixed parameters (c estimated by recursive least square estimation (RLSE). (k + 1) = Q (k ) - S (k )Y (k )e (k + 1) Q (13) 1 + YT ( k ) S ( k ) Y ( k ) S (k )Y (k )Y T (k )S (k ) (14) 1 + Y T (k ) S (k )Y (k ) (k ), (k )) and where Y (k ) denotes Y ( x | c T (k ) - y (k + 1) . (k + 1) - y (k + 1) = Y (k )Q e (k + 1) = y S (k + 1) = S (k ) The initial conditions are chosen as T T (0) q T (0) L q T (0) , S (0) = WI (0) = q Q 1 2 N T (0) q T (0) L q T (0) are the parameters where q 1 2 N of the initial T-S model and W is a large positive constant. Large W can speed up parameter convergence, but too large may cause instability. It needs to be chosen properly during on-line adaptation. 3. STABILITY ANALYSIS 3.1 Convergence of model tracking error Applying (9) to (7) , we get (k | p (k )) (k + 1) = yr (k + 1) + k T e y (15)

i ( k )u ( k ) (k ) | c i , i ) + w i ( x g
i =1

(x (k ) | p (k )) + g (k ) | p (k ))u (k ) (x = f (7) where i (k ) = q i i i f 0 (k ) q1 (k ) L qn (k ) , i (k ) = q i (k ), g n +1

(k ) (k ) = c (k ), (k ), Q p
N

(8) T T (x i (k ) 1 x (k ) | p (k )) = w i ( x (k )) f f (k ) ,
i =1 N

i (k ). (k ) | p (k )) = w i ( x (k )) (x g g
i =1

To apply feedback linearization technique to design (k ) | p (k )) is required to be (x the control law g invertible. (k ) | p (k )) > e where e is a (x Assumption 3. g small real positive number, which implies the relative degree of the T-S fuzzy model is equal to one. With the assumption 3, the control law can be calculated as 1 (x |p (k )) + v (k )] (9) u (k ) = [- f |p (k )) (x g where v( k ) is a new input chosen as where error (k )) = yr (k ) - y (k )) ( k | p (k | p e (11) and k = (k n ,L , k1 )T be such that the zeros of the (k | p (k )) v (k ) = yr (k + 1) + k T e (10) (k | p (k )) = [e (k - n + 1)) L (k - n + 1| p e
T

(k )) = yr (k + 1) - y (k + 1)) , (k + 1| p (k + 1| p Since e we have the following error dynamics (k + 1| p (k + 1)) = -kn e (k - n + 1| p (k - n + 1)) e (16) (k - 1)) - k1e (k )) (k - 1| p (k | p - L - k2 e which can be formulated in matrix form (k + 1| p (k + 1)) = Le (k | p (k )) e where 1 0 0 L 0 0 0 0 0 1 0 L 0 0 L= M M M M O M M , (18) 0 0 0 L 0 1 0 -k n -k n -1 L L L L -k1 Since L is a stable matrix ( zI - L = z n + k1 z n -1 + L + k n is stable), the system (17) is asymptotically stable, which implies (k | p (k )) 0 when k . The model tracking e error will converge to zero. 3.2 Boundedness of parameters and control signal Assumption 4. In order to deal with the T-S model bring linear in parameters when performing tracking (17)

(k ))] and e (k )) is the model tracking ( k | p ( k | p e

polynomial h( z ) = z n + k1 z n -1 + L + kn lie within the unit circle centered at the origin of the z plane. The control law (9) applied to (7) results in it being decoupled and linear with respect to the new input v (k ) (k + 1) = v (k ) y (12) 2.3 Online parameter adaptation The initial T-S fuzzy model can be obtained by either off-line or on-line identification methods. In this paper, the initial model is obtained by the on-line identification method proposed in ((Qi and Brdys,

387

error boundedness analysis, we shall assume that the , ) are the same premise parameters of the model (c as the plant parameters (c , ) . However, updates of all paramters will be carried during the control simulation studies. Lemma 1. The adaptive algorithm (13)-(14) % (k ) and Q (k ) are bounded. guarantees that Q Lemma 2. If the signal Y (k ) is persistently exciting, then the least-squares algorithm (13)-(14) ensures % (k ) = 0 . that lim Q k These follow from standard properties of the leastsquare estimation methods applied to linearly parameterized models. Applying (9) to (7) yields (k ) (k + 1) = yr (k + 1) + k T e y (19) From assumption 2, yr (k + 1) is bounded and since (k ) is bounded. Thus, equation (19) implies the e (k ) is a vector (k + 1) . Since x boundedness of y (k ) is bounded. consisting of past outputs, x (k ) , bounded parameters Q(k ) and Bounded x (x (k )) imply bounded f (k )) and bounded w i ( x (k )) in (8). Thus, from (9) we obtain (x bounded g the boundedness of u (k ) . 3.3 Boundedness of plant tracking error From (3), we have the plant output evolution.
T y (k + 1) = w i ( x (k )) i 1 x (k ) u (k ) i =1 N T N N T

) - w i ( x ) ) iT x e (k ) = (w i ( x
i =1 iT iT (k ) + w i ( x )x e - w i ( x )x e (k ) i =1 i 1 i =1 i x i 2 i n T N N

(23)

where = q q L q . Substituting (23) into (22), the plant tracking error dynamics can be obtained as follows T (k )Q % ( k ) + (w i ( x e (k ) e(k + 1) = Y ) - w i ( x) ) iT x
i =1 iT (k ) + w i ( x )x (k ) - w ( x ) e e (k ) - k T e i N i =1 N iT x N i =1 N

= w ( x ) e (k ) + G(k )
i i =1 iT x

(24) where % (k ) - w i ( x ) iT e T (k )Q (k ) - k T e (k ) G( k ) = Y x
i =1 N

) - w i ( x) ) iT x e (k ) + (w ( x
i i =1

(25)

The error dynamics (24) can also be written into a matrix form e (k + 1) = w i ( x)Ai e (k ) + G(k )
i =1 N

(26)

where
0 1 0 0 0 0 1 0 i A =M M M M 0 0 0 0 i i q1 q 2 L L

L L

0 0 0 O M M , G (k ) = . G( k ) L 0 1 L L q ni 0 0

T i i = w i ( x ) if 1 x (k ) + w ( x )g (k )u(k ) i =1 i =1

(27) Assumption 5. There exists a positive constant a and a common positive definite symmetric matrix P for all the matrices Ai ( i = 1,L, N ) such that AiT PAi - P -aI < 0 for i = 1,L , N (28)

= f ( x (k )) + g ( x (k ))u (k ) (20) Applying (9) to (20) and after some manipulations, we have y (k + 1) = f ( x(k )) + g ( x (k ))u (k ) (x ) + g )u (k )) (21) (x = f ( x ) + g ( x )u ( k ) - ( f (k ) + yr (k + 1) + k T e Moving yr (k + 1) to the left side of (21) and after some manipulation, we get T ( k )Q ( k ) - YT ( k ) Q - k T e (k ) e(k + 1) = Y (22) T ( k )Q % (k ) + Y % T (k )Q - k T e (k ) =Y (k ) where Y (k ) denotes Y ( x (k ) | c , ) and Y (x (k ) | c (k ), (k )) . denotes Y Considering % T ( k )Q = Y T (k )Q - YT (k )Q Y ) x e (k ) - w ( x ) xe (k ) = w (x
i iT i iT i =1 i =1 N N

Theorem 1. Consider the closed-loop system consisting of plant (3), controller (9) and parameter adaptation law (13)-(14). If the assumptions 1-5 hold, then (i) the plant tracking error e (k ) is bounded; (ii) if in addition, the signal Y (k ) is persistently exciting, e (k ) converges to zero. Proof. Consider the following possible Lyapunov function V (e (k )) = e (k )T Pe (k ) (29) Then the difference of the Lyapunov function along the plant tracking error trajectory (26) is calculated as DV = e (k + 1)T Pe (k + 1) - e (k )T Pe (k ) = (w i ( x ))2 e (k )T ( AiT PAi - P)e (k ) + +
i =1 N i =1 j = i +1 N

w w e (k )
i j N i =1

( AiT PA j + A jT PAi - 2 P )e (k )

+ 2G(k )T P w i ( x )Ai e (k ) + G(k )T PG(k ) (30)

388

With assumption 5, we have AiT PA j + A jT PAi - 2 P -2aI for all i and j (31) Applying (31) in (30), the following inequality can be obtained N N DV -e (k )T w i ( x ) w j ( x )a e (k ) j =1 i =1 + 2G(k )T P w i ( x )Ai e (k ) + G(k )T PG(k ) 2M 2 1 2 - a e (k ) + + lmax ( P) G(k ) 2 a where
i i =1 2 N

(32)

M = lmax ( P) max Ai , i = 1,L N P ( w i ( x ) Ai )

where y (t ) = q1 (t ) is the arm position which is the measured output, q2 (t ) is the angular velocity which is unmeasurable and u (t ) is the input torque. The parameters a1 and a 2 depend on the mass and length of the arm. The values of a1 , a 2 and b are chosen as a1 = a 2 = b = 1 . The premise variables are selected as x1 (kT ) = y[(k - 1)T )] and x2 (kT ) = y (kT ) . The control The sampling time T = 0.01sec . objective is to make q1 track a bounded reference signal. The initial T-S fuzzy model for closed-loop control is an input-output model obtained by the online identification method proposed in Qi and Brdys (2005) which follows the structure of (38).
2 1 Outputs 0 -1 0 0.2 Plant Tracking Error 0.15 0.1 0.05 0 -0.05 0 10 20 30 Time (sec) 40 50 60 10 20 30 Time (sec) 40 50 60 plant output reference

Consider G( k ) = G( k )
N T (k ) Q % (k ) + w i ( x ) iT + k T e Y x (k ) (33) i =1

+
i =1

) - w i ( x) i x e (k ) w (x
i

Since all the elements in (33) are bounded, G(k ) is bounded. Hence there exists positive constant rG such that G( k ) r G (34) We can conclude outside the ball 4M 2 + 2almin ( P ) e ( k ) : e ( k ) < e = rG (35) a the Lyapunov function difference can be bounded by DV (e (k )) 0 " e (k ) e (36) The hypothesis (i) has now been proven. With Lemma 2, if the signal Y ( k ) is persistently exciting, then the least-squares algorithm (13)-(14) % ( k ) 0 . With the assumption 4, ensures that lim Q
% ( k ) 0 means the model become a perfect lim Q
k k

Fig. 2. Plant tracking response with parameter disturbance ( a 2 changes from 1 to 1.5 at t = 5s )
5 0 -5 -10 -15 0 x 10
-3

a30

2.01 2.005 2 1.995

a31

20 40 Time (sec)

60

1.99 0 x 10
-3

20 40 Time (sec)

60

-0.99 -0.992 -0.994 -0.996 -0.998 -1 0 20 40 Time (sec) 60 a32

3 2 1 0 -1 0

a33

representation of the plant, which will make the model state x (k ) become the same as the plant state
) - w i ( x ) = 0 . With x (k ) . Thus we have lim w i ( x
k

(k ) 0 . the control law (9), we have proven lim e

20 40 Time (sec)

60

Therefore, when k , G(k ) in (33) converges to zero and (32) becomes 1 2 DV (e (k )) - a e (k ) k (37) 2 we obtain e (k ) 0 as k . The hypothesis (ii) and the Theorem 1 have now been proven. 5. SIMULATION RESULTS In this example, the validity and effectiveness of the proposed control scheme are illustrated through the tracking control of a single-link robot arm described by &1 (t ) = q2 (t ) q (38) &2 (t ) = -a1 sin[q1 (t )] - a 2 q2 (t ) + bu(t ) q

Fig. 3. Consequent parameter adaptation of rule 3

0.5 0 -0.5 -1 -1.5 -2 0 5 10 15 20

25

30

35

40

Fig. 4. Control signal Fig. 2 illustrates the plant tracking response with disturbance and tracking error and Fig. 4 shows the control signal. The plant parameter a 2 changes from 1 to 1.5 at t = 5s , which causes the tracking error to
389

increase after t = 5s . However, the controller is able to adapt to the change and regain good tracking performance. The tracking error decreases gradually as the model is learning to become a good approximation of the plant. Parameter adaptation of rule 3 is shown in Fig. 3. The learning rate W is chosen as W = 1.0 106 . The reference trajectory in Fig. 5 is obtained by passing a piecewise constant through a first-order stable linear filter. A disturbance in a1 was introduced at t = 44s . As a result, it can be seen that the tracking error increases after 44s . The parameter adaptations are shown in Fig. 6.The control law can produce control signal that adapts to the change and regain good tracking performance.
2 1 Outputs 0 -1 0 0.8 Plant Tracking Error 0.6 0.4 0.2 0 -0.2 0 10 20 30 40 50 60 Time (sec) 70 80 90 100 10 20 30 40 50 60 Time (sec) 70 80 90 100 plant output reference

convergence are derived for stable nonlinear plant. It should be noted here that although the error boundedness or the closed-loop global stability are proven under the assumption that the premise parameters of the model are the same as in the plant, a local stability or the error boundedness are achieved if the initial parameter values are close enough to the plant parameter values. Simulation examples verify the effectiveness, adaptation and tracking performance of the proposed control scheme. ACKNOWLEDGEMENT This work has been partly supported by Universities UK on behalf of the Department for Education and Skills (DfES) through the UK funding councils and by the Polish State Committee for Scientific Research under grant No. 4 T11A 008 25. REFERENCES Golea, N., Golea, A. and Benmahammed, K. (2003), Stabel indirect fuzzy adaptive control, Fuzzy Sets and Syst., vol. 137, pp. 353-366. Kulawski, G. J. and Brdys, M. A. (2000), Stable adaptive control with recurrent networks, Automatica, vol. 36, pp. 5-22. Park, C.-W. and Cho, Y.-W. (2004), T-S Model Based Indirect Adaptive Fuzzy Control Using Online Parameter Estimation, IEEE Trans. Syst. Man. And Cyber., Part B, Vol. 34, pp. 2293-2302. Qi, R. and Brdys, M.A. (2005), Adaptive fuzzy modelling and control for discrete-time nonlinear uncertain systems, in Proc. Of the 2005 American Control Conference, pp. 11081113. Spooner, J. T. and Passino, K. M., Stable adaptive control using fuzzy systems and neural networks, IEEE Trans. Fuzzy Syst., vol. 4, pp. 339-359, Aug. 1996. Takagi, T. and Sugeno, M. (1985), Fuzzy idenfication of systems and its applications to modeling and control, IEEE Trans. Syst., Man, Cybern., vol. SMC-15, pp. 116-132. Tao, G. (2003), Adaptive design and analysis. Hoboken, NJ: John Wiley & Sons, Inc.. Wang, L.-X. (1994), Adaptive Fuzzy Systems and Control: Design and Stability Analysis. Englewood Cliffs, NJ: Prentice-Hall. Ying, H. (1998), General MISO Takagi-Sugeno fuzzy systems with simplified linear rule consequent as universal approximators for control and modeling applications, IEEE Trans. Fuzzy Syst, vol. 6, pp. 582-587. Ying, H. (1999), Comparison of necessary conditions for typical Takagi-Sugeno and Mamdani fuzzy systems as universal approximators, IEEE Trans. Syst., Man, Cybern., vol. 29, pp. 508-514. Yoo, B. and Ham, W. (1998), Adaptive fuzzy sliding mode control of nonlinear systems, IEEE Trans. Fuzzy Syst., vol. 6, pp. 315-321.

Fig. 5. Plant tracking response with parameter disturbance ( a1 changes from 1 to 1.5 at t = 44s )
0.1 0.05 0 -0.05 -0.1 0 -0.94 ^ 3 1 -0.96 1 -0.98 1.995 1.99 0 x 10
-4

2.005 ^ 3 0 2 ^ 3 2

50 Time (sec)

100

50 Time (sec)

100

^ 3 3

-1 0

50 Time (sec)

100

0 0

50 Time (sec)

100

Fig. 6. Consequent parameter adaptation of rule 2 6. CONCLUSION The indirect adaptive fuzzy control algorithm for uncertain nonlinear stable plants is presented in this paper. A discrete-time T-S fuzzy model is employed as a dynamical model of an unknown continuoustime plant with unmeasurable states. By proper training, the T-S fuzzy model is able to achieve input-output behaviour close to the plant and provide useful information about the states of the plant. A feedback linearization control is calculated entirely based on the model, using the model structure, parameters and states, while the parameters of the model are updated on-line by RLSE method. Sufficient conditions stability and tracking error
390