Nonaffine Nonlinear

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Second International Symposium on Intelligent Information Technology Application

Adaptive Backstepping Control for a Class of Nonaffine Nonlinear Systems Based


Neural Networks

Jianqing Min, Zibin Xu, Yingguo Fang


College of Biology and Environment Engineering,
Zhejiang Shuren University, Hangzhou, ZheJiang, 310015, China
minjq@sina.com, hzxuzibin@gmail.com

Abstract In this paper, the problem of designing an adaptive


neural networks controller is studied aiming at a class of
Aiming at a class of nonaffine nonlinear system with nonaffine nonlinear system with uncertainties, and a
uncertainties, an adaptive backstepping neural controller simulation example is presented to demonstrate the
design is presented. By applying backstepping design effectiveness of the proposed control design.
strategy and online approaching nonlinearity with fully
tuned radial basis function (RBF) neural networks, the 2. Problem formulation
adaptive tuning rules are derived from the Lyapunov
stability theory. A nonlinear tracking differentiator is Consider the uncertain nonaffine nonlinear system in
introduced to deal with the problem of extremely the form of
expanded operation quantity of backstepping method. The
developed control scheme guarantees that all the signals ⎧ x& i = f i ( X i ) + g i ( X i ) xi +1 (1 ≤ i < n) (1)

of the closed-loop system are uniformly ultimately ⎩ x& n = b( X , u )
bounded. The effectiveness of the proposed controller is
Where, X = [ x1 , x 2 L , x n ]T ∈ R n is the state vector,
illustrated through a simulation example.
X i = [ x1 , x 2 L , xi ] ∈ R ; u ∈ R is the control input; fi(Xi)
T i

1. Introduction and gi(Xi) are unknown smooth functions, and cannot be


expressed as linearization form.
The study on uncertain nonlinear systems adaptive The aim is to design a controller that can eliminate the
control has attracted wide attention and some important effect of unexpected factors, so that the system output can
achievements were obtained during the recent years [1-6], track the desired control output and it can be ensured all
in particular, the nonlinear systems control based on the signals of the closed-loop system are uniformly
neural networks is one of the active research areas. ultimately bounded.
However, the majority of research results focus on affine Before the main results are given, the assumptions and
systems rather than nonaffine, this is because the lemmas are introduced below.
nonlinear function of its state equation implies control Assumption 1: Suppose g u ( X , u ) := ∂b[( X , u )]/∂u ≠ 0 , there
input, which leads to difficulty studying. In fact, many of always exist normal number b1, b2 and b0 which satisfy
the actual systems can not be described by the affine, such
b1 ≤ g u ( X , u ) ≤ b2 and g& u ( X , u ) ≤ 2b0 .
as flight control systems, chemical control systems etc. In
[9], a direct adaptive control design method is proposed Assumption 2: gi(Xi) is positive or negative, and the sign
for a class of uncertain nonaffine nonlinear system using is invariable, let’s set gi(Xi)>0. There exists normal
multilayer neural networks, but the discussion of higher number gi0 and smooth bounded function Gi(Xi), which
order systems isn’t given; and the uncertain systems are satisfy gi0 ≤ gi (Xi ) ≤ Gi (Xi ) and β&i ( X i ) ≤ 2α i . Where α i ≥ 0 ,
not studied in detail in [10]. In addition, the neural βi (Xi ) = Gi (Xi )/ gi (Xi ).
networks only provide a limited usage of adjusting the
Assumption 3: For a real continuous function
weights in most of the applications, and the desired
approximating effect is hard to obtain when a sufficient Δf : Ω a R , Ω is a compact subset of R. For any ε>0,
number of system parameters are not available. To solve there always exists a Gauss function vector h ∈ R l ×1 and
this problem, the fully tuned RBF neural networks is an optimum weight vector w ∈ R l ×1 , which fulfils the
proposed in [11], unfortunately, the stability of whole following equation
closed-loop system cannot be proved.
Δf = wT h(ξ, c, b) + ε (2)

978-0-7695-3497-8/08 $25.00 © 2008 IEEE 716


DOI 10.1109/IITA.2008.167
2
Where, the Gauss function h j = exp( −0.5 X − c ij 2
/ bj ), 1 1 ~T −1 ~
V1 ( Z 1 ) = β 1 ( X 1 ) z1 2 + tr [ w 1 Γw1 w 1 ] (7)
2 2
(i=1,2,…,n; j=1,2,…,l), ξ∈ R n×1 is reference input vector; 1 1 ~ T −1 ~
+ tr [ c~1 Γc1 c~1 ] + tr [b1 Γb1 b1 ]
T −1

l is the number of hidden nodes, c ∈ R m×l is the center 2 2


matrix, b ∈ R l ×1 is the width vector, ε is the Where Γw1 = Γw1T > 0 , Γc1 = Γc1T > 0 and Γb1 = Γb1T > 0
approximation error. are constant matrixes.
Compared to the RBF neural networks, which is a The time derivative of V1 is given by
typical local-approached network, the fully tuned RBF
can adjust not only to the weight but also to the center and 1 ~ T Γ −1 w
~& )
V&1 ( Z 1 ) = β&1 ( X 1 ) z1 + β 1 (X 1 ) z1 z&1 + tr ( w
2
1 w1 1
the width of hidden node, that is to say, the good 2
~ −1 ~ ~ ~
&
+ tr (c1 Γc1 c&1 ) + tr (b1 Γb1 b1 )
T T −1
approximate ability of the networks on-line.
We define ŵ , cˆ and b̂ as estimate value of w, c and b
~ = wˆ - w , From assumption 2 and 3 we obtain
respectively, and then the estimate errors are w
~ ~ ˆ
c = cˆ - c and b = b - b . For the convenience, we denote β1 ( X 1 )[ f1 ( X 1 ) − x&1d ] + α1 z1 = w1T h1 (ξ1 , c1 , b1 ) + ε 1 (8)
h := h(ξ , c, b) and hˆ := h(ξ , cˆ, bˆ) . Then the following lemma
We have
can be obtained [12].
Lemma 1: the approximation error of fully tuned RBF ~ T Γ −1 w
V&1 ( Z 1 ) ≤ z1 [ w1 h1 + ε 1 + G1 ( X 1 ) x 2 d ] + tr ( w
T ~& )
1 w1 1
neural networks is as follows ~ T −1 ~&
+ tr (c~ Γ c~& ) + tr (b Γ b )
T −1
1 c1 1 1 b1 1
~ T ( hˆ - hˆ ′ cˆ - hˆ ′ bˆ) + wˆ T (hˆ ′ c~ + hˆ ′ b~ ) + γ (3)
wˆ T hˆ − w T h = w c b c b
Where ξ1 = [ X 1 , x1d , x&1d ]
Where hˆ c′ = ∂hˆ /∂c , hˆ b′ = ∂hˆ /∂b , the upper bound of Choose the desired virtual input as follows
interference term γ satisfies x 2 d = G1 ( X 1 )[−k1 z1 - wˆ 1 hˆ1 ]
−1 T
(9)
γ ≤ w ( hˆ c′ cˆ + hˆ b′ bˆ ) + wˆ T hˆ c′ c F
+ wˆ hˆ b′
T
b + 2 w 1 (4) Where k1 is design parameter which will be given in the
F F F F

following.
Where ⋅ , ⋅ and ⋅ denote 1, 2 and Frobenius norms The adaptive law of each neural networks parameter is
1 F
respectively. chosen as

3. Adaptive backstepping controller design ⎧w&ˆ = Γ (hˆ - hˆ ′ cˆ - hˆ ′ bˆ ) z − Γ σ wˆ


⎪ 1 w1 1 1c1 1 1b1 1 1 w1 w1 1
⎪& (10)
⎨cˆ1 = Γc1 hˆ1′c1 wˆ 1 z1 − Γc1σ c1cˆ1
T
The Backstepping design, aiming at the mismatched
⎪&
uncertainties system, is a kind of systematic synthetic ⎪⎩bˆ1 = Γb1 hˆ1′b1 wˆ 1 z1 − Γb1σ b1bˆ1
T

technique to control. It is a recursive procedure that


combines the choice of a Lyapunov function with the Where σ w 1 , σ c1 and σ b1 are positive parameters.
design of feedback control. The Backstepping design
technique starts from the differential equation in its From lemma 1 we also know
smallest order of the system, introduces the concept of V&1 (Z1 ) = −k1 z1 + z1 (ε1 − γ 1 ) −σ w1tr(w
2 ~ T wˆ )
1 1
virtual control, designs the satisfactory virtual control step ~
−σ c1tr(c~1 cˆ1 ) −σ b1tr(b1 bˆ1 )
T T
by step, and the real controller can be obtained in the end.
Define the state variable Zi Taking into account that
Z i = [ z1, L , z i ] , z i = xi − xid , i = 1, L, n (5) ⎧ ~T 1 2 1 ~ 2
⎪- σ w1tr ( w1 wˆ 1 ) ≤ 2 σ w1 w1 − 2 σ w1 w1
Where xid is the desired fictitious control value of each ⎪
order. ⎪ ~T 1 2 1 ~ 2
⎨- σ c1tr (c1 cˆ1 ) ≤ σ c1 c1 F − σ c1 c1 F
Step 1 (i=1): x2d is the desired virtual input, the first-order ⎪ 2 2
subsystem of system (1) can be expressed as ⎪ ~T ˆ 1 2 1 ~ 2
⎪- σ b1tr (b1 b1 ) ≤ 2 σ b1 b1 − 2 σ b1 b1

x&1 = f1 ( X 1 ) + g1 ( X 1 ) x2 d (6)
2
We consider the stabilization of subsystem (6). Define 2 z1
the state variable as z1 = x1 − x1d . In order to disposal z1 ε 1 ≤ η1 ε 1 +
4η1
effectively the issue of the singularity-free adaptive
control we choose the Lyapunov function candidate as Let

717
2 2 2 The adaptive law expression, ki and Ci (including the
k1 = [ wˆ 1 hˆ1′c1 + wˆ 1 hˆ1′b1 + hˆ1′c1 cˆ1
T T
F F F (11) following Cn ) are similar to (10), (11) and (12)
2
+ hˆ1′b1 bˆ1
2
+ 5 / 4 + G1 ( X 1 )] / η1 η1 > 0 respectively except that the each variable subscript
F changes from 1 to i (n).
Combine (4), the following inequality can be gotten Step n (i=n): Define z1 = x1 − x1d ,…, z n = x n − xnd ,

1 z n +1 = u − x ( n +1) d = 0 , then Eq. (13) can be expressed as


~ 2 − 1 σ c~
V&1 ( Z1 ) ≤ −G1 ( X1 ) z1 /η1 − σ w1 w
2 2 2 1 ~ 2
− σ b1 b1 + C1
1 c1 1 F
2 2 2
⎧ z& i = f i ( X i ) + g i ( X i ) x i +1 − x& id (1 ≤ i < n) (16)
Then ⎨
⎩ z& n = b( X , u ) − x& nd
2 2 2
c1 b1 w1 2 2 We use differential intermediate value theorem as the
C1 = η1 [ F
+ + + w1 1
+ ε1 ] (12) system is nonaffine, that is
4 4 2
1 2 1 2 1 2
+ σ w1 w1 + σ c1 c1 F
+ σ b1 b1 b( X , u ) = b( X , u ∗ ) + g uλ (u - u ∗ ) , g uλ = g u ( X , u ) u =uλ
2 2 2
Step i (2 ≤ i ≤ n −1) : A similar procedure is employed uλ = λ ⋅ u + (1 − λ )u ∗ , 0 < λ < 1 .
recursively for each step i. we assume that x(i+1)d is virtual Suppose μ = z n −1G n −1 ( X n -1 ) + k n z n − x& nd , where kn>0.
input, consider the stabilization of the ith-order subsystem
described as Then
∂μ / ∂u = 0 , ∂[b( X , u ) + μ ] /∂u ≠ 0
⎧ x& i -1 = f i -1 ( X i -1 ) + g i -1 ( X i -1 ) xi (13)
⎨&
⎩ xi = f i ( X i ) + g i ( X i ) x(i +1) d By means of the implicit function theorem, we have

The Lyapunov function candidate is chosen as b( X , u ∗ ( X , μ )) + μ = 0

1 1 ~ T −1 ~ Which yields
β i ( X i ) z i 2 + tr[ w
Vi ( Z i ) = Vi −1 ( Z i −1 ) + i Γwi w i ]
2 2 (14) (17)
z& n = g uλ (u - u ∗ ) − z n −1Gn−1 ( X n-1 ) − k n z n
1 ~ T −1 ~ 1 ~ T −1 ~
+ tr[c i Γci c i ] + tr[bi Γbi bi ]
2 2 Let
Where z1 = x1 − x1d , …, z i = xi − xid . u ∗ = w n hn (ξ n , c n , bn ) + ε n
T
(18)
The time derivation is given by
Where ξ n = [ X , x ( n −1) d , x nd , xˆ& nd ]
1
V&i ( Z i ) = V&i −1 ( Z i −1 ) + z i −1Gi −1 ( X i −1 ) z i + β& i ( X i ) z i
2
The final control law is chosen as:
2
+ β i (X i ) z i [ f i ( X i ) + g i ( X i ) x ( i +1) d − x& i d ]
u = −(c1 + c 2 ) z n + wˆ n hˆ n + rn
T
(19)
+ tr ( w ~ T Γ −1 w ~& ) + tr (c~ T Γ −1 c~& ) + tr (b~ T Γ −1 b~& )
i wi i i ci i i bi i
Where c1>0, c2>0 are design parameters, rn is an adopted
Let robust term. Then we have
β i ( X i )[ f i ( X i ) − x& i d ] + α i z i = w i T hi ( ξ i , c i , bi ) + ε i ~ T (hˆ - hˆ ′ cˆ - hˆ ′ bˆ ) + wˆ T (hˆ ′ c~ + hˆ ′ b~ )
z& n = g uλ [w n n ncn n nbn n n ncn n nbn n

Where ξ i = [ X i , xid , xˆ& id ] , x&ˆ id is the nonlinear tracking + γ n − ε n + rn ] − z n−1Gn−1 ( X n-1 ) − k n z n − g uλ (c1 + c2 ) z n
differentiator output which xid is the input. The purpose of Choose the nth-order Lyapunov function as
introducing differentiator is to deal with the problem of
1 2 1 ~ T −1 ~
extremely expanded operation quantity of backstepping Vn = Vn −1 + z n + tr[ w n Γwn w n ]
method. The related properties of differentiator can refer 2 g uλ 2 (20)
to [13] and [14] 1 ~ T −1 ~ 1 ~ T −1 ~
+ tr[c n Γcn c n ] + tr[bn Γbn bn ]
Then choose the virtual control input as 2 2

x ( i +1) d = G i ( X i )[ − z i −1G i −1 ( X i −1 ) − k i z i - wˆ i hˆi ]


−1 T
(15) and
g& z z&
V&n = V&n−1 + z n −1Gn −1 ( X n −1 ) z n − uλ2 z n + n n
2
Then
1 2g uλ g uλ
2 ~
V&i (Z i ) ≤ V&i −1 (Zi −1 ) − Gi ( X i )zi / ηi − σ wi w
2 2
~ −1 ~
&
i ~ T −1 ~ ~ T − 1 ~
+ tr[wn Γwn w& n ] + tr[cn Γcn c&n ] + tr[bn Γbn bn ]
T
2
1 1 ~ 2
− σ ci c~i F − σ bi bi + Ci
2

2 2 The adaptive law is designed as follows

718
⎧w)& = −Γ (hˆ - hˆ ′ cˆ - hˆ ′ bˆ ) z − Γ σ wˆ Vn (t ) ≤ [Vn (0) − C / ρ ]e − ρt + C / ρ (23)
⎪ n wn n ncn n nbn n n w n wn n
⎪& ˆ (21) From (23), we know that Vn(t) is bounded in C/ρ.
⎨cˆ n = −Γcn hnc′ n wˆ n z n − Γcnσ cn cˆ n
T

⎪& Consequently, combining (20) and Lemma 1, we can infer


⎪⎩bˆn = −Γbn hˆnb
′ n T wˆ n z n − Γbnσ bn bˆn that all signals in the closed-loop system are globally
uniformly bounded, and the tracking error satisfies
Which yields n −1

∑z + z n / b2 ≤ 2[Vn (0) − C / ρ ]e − ρt + 2C / ρ
2 2
j
1
V&n = V&n−1 + (1 − ) z n Gn ( X n −1 ) z n − c2 z n2 j =1
g uλ −1 −1
g& k
− (c1 + uλ2 ) z n − n z n + z n (γ n − ε n + rn )
2 2

2g uλ g uλ 4. Simulations
− σ wn tr(w ~ T wˆ ) − σ tr(c~ T cˆ ) − σ tr(b~ T bˆ )
n n cn n n bn n n
To prove the validity of our results, the following
nonaffine nonlinear system was simulated
Let c > b0 > g& uλ , we have
1
b1 2 g u2λ ⎧ x&1 = x1 sin( x1 ) + (sin x1 + 5) x 2 (24)
⎨ 2
1 (1 −1 / b ) 2 2
2 2 ⎩ x& 2 = (1 + x 2 )u + sin(0.2u )
(1 − ) z n Gn ( X n −1 ) z n − c2 z n2 ≤ z n −1 Gn −1 ( X n −1 )
g uλ −1 −1 4c2
To ensure the smooth tracking system signals, the
The robust term is selected as desired trajectory x1d is chosen as the following
2 2 2 reference model
rn = − z n [ wˆ n hˆ nc
′n + wˆ n hˆ nb
′n + hˆ nc
′ n cˆ n
T T

2
F F F
x1d ω n2
+ hˆ nb
′ n bˆn + 5 / 4] / η n ηn > 0 = 2
F r s + 2ξ nω n s + ω n2
We get Where ω n = 20 , ξ n = 0.5 , r is a square wave signal which
n −1
2
G j ( X j )z j
2
kn zn
2
(1 − 1 / b ) 2 2 its frequency is 1 Hz and the amplitude is 1.
V&n ≤ −∑
2 2
− + z n −1 G n −1 ( X n −1 ) The initial state is [−1,0]T , and the parameters of
j =1 ηi g uλ 4c 2
1 n ~ 2 ~ 2 n controller are given as follows
− ∑ [σ wj w + σ cj c~ j + σ bj b j ] + ∑ C j
2

2 j =1
j F
j =1
The width vector and the center vector of RBF neural
n−2 G
2
( X ) z
2 2 networks are both 0.5, the initial weight is 0,
k z
≤ −∑
j j j 2 2
− n n − c 0 z n −1 G n −1 ( X n −1 ) Γw1 = Γw2 = diag{2} , Γc 2 = Γc 3 = diag{5} ,
j =1 ηj g uλ
1 n n Γb 2 = Γb 3 = diag{5} , σ w1 = σ w 2 = 0.1 , η1 = 0.5 , η 2 = 0.1 ,
− ∑ [σ wj w ~ 2 + σ c~ 2 + σ b~ 2 ] + C
2 j =1
j cj j F bj j ∑
j =1
j
c 0 = 0.2 , σ c 2 = σ c 3 = σ b 2 = σ b3 = 0.02 .
The simulation results are shown in Fig.1 and 2, it is
Where c = 1 (1 − 1 / b 2 ) 2 clear that the adaptive neural networks backstepping
− >0 controller has good control characteristics with robustness.
η n −1
0
4c 2
2
Define x 1d
1.5
x1
ρ = min {2g2j 0 / η j ,2c0 g(2n−1)0 ,2kn , σwj / λmax(Γwj−1),σcj / λmax(Γcj−1),
j =1,L,n−2
1
σbj / λmax(Γbj−1),σw(n−1) / λmax(Γw−(1n−1) ),σc(n−1) / λmax(Γc−(1n−1) ),
σb(n−1) / λmax(Γb−(1n−1) ),σwn / λmax(Γwn
−1
),σcn / λmax(Γcn−1),σbn / λmax(Γbn−1)} 0.5
x1, x1d

n 0
C = ∑C j
j =1
-0.5

Then, we obtain
-1
V&n ≤ − ρV n + C (22)
-1.5
0 0.5 1 1.5 2 2.5 3
Multiply e ρt to the two sides of (22), and integrate in t (s)
the space [0, t], we yield Figure 1 Result of system state x1

719
[6] Peng, Z P Jiang (2000). “Stable neural controller design for
8
x 2d unknown nonlinear systems using backstepping”. IEEE
x2 Transaction on neural networks, Vol.11, No.6, pp. 1347-
6 1360
[7] Mou Chen , Changsheng Jiang, Qingxian Wu (2007).
4 “Backstepping control for a class of uncertain nonlinear
systems with neural network”. International Journal of
2 Nonlinear Science, Vol.3, No.2, pp. 137-143.
[8] Niu Y, Lam J, Wang X, et al (2003). “Sliding mode control
for nonlinear state-delayed systems using neural network
x2, x2d

0
approximation”. IEE Proceedings-Control Theory and
-2 Applications, Vol.150, No.3, pp. 233-239.
[9] Jin Yuqiang, Liang Ruisheng, Shen Hao (2007). “Neural
-4 adaptive control for a class of nonlinear systems”. Journal
of Naval Aeronautical Engineering Institute, Vol.22, No.2,
0 0.5 1 1.5 2 2.5
pp. 222-226.
t (s) [10] Ge S S , Hang C C(1998). “Nonlinear adaptive control
Figure 2 Result of system state x2 using neural networks and its application to CSTR
systems”. Journal of Process Control, Vol.9, No.4, pp.
5. Conclusions 313-323.
[11] Li Y, Sundararajun N, Saratchandran P (2001). “Neuro-
In this paper, aiming at a class of nonaffine nonlinear controller design for nonlinear fighter aircraft maneuver
system with uncertainties, an adaptive backstepping using fully tuned RBF networks”. Automatic, Vol.37, No.8,
pp. 1293-1301.
neural networks control method is present. Based on
[12] Zhao Tong, Zhang Huaxiang; Cao Menglong, et al (2005).
Lyapunov stability theory, the parameter adaptive rules of “The adaptive control for hysteresis nonlinear systems via
fully tuned RBF neural networks are derived, and the neural networks”. Journal of Shanghai Jiaotong University,
stability of closed-loop system is guaranteed. Meanwhile Vol.39, No.12, pp. 2002-2005. (in Chinese)
the nonlinear tracking differentiator is introduced to deal [13] Han Jingqing, Wang Wei (1994). “Nonlinear tracking-
with the problem of extremely expanded operation differentiator”. Journal of Systems Science and
quantity of backstepping method. Simulation results prove Mathematical Sciences, Vol.12, No.2, pp. 177-183. (in
the effectiveness of the designed controller. Chinese)
[14] Wang Xinhua, Chen Zeng-qiang, Yuan Zhuzhi (2003).
“Nonlinear tracking-differentiator with high speed in whole
Acknowledgement course”. Control Theory & Applications, Vol.20, No.6, pp.
875-878. (in Chinese)
This work is supported by the Natural Science
Foundation of Zhejiang Province of China No. Y107605.

References

[1] M. O. Efe, C. nsal, O. Kaynak, X. Yu (2004). “Variable


structure control of a class of uncertain systems”.
Automatica, Vol.40, No. 1, pp. 59-64.
[2] M A Rahman, D M Vilathgamuwa, M N Uddin, K J Tseng
(2003). “Nonlinear control of interior permanent-magnet
synchronous motor”. IEEE Trans. on Industry Applications,
Vol.39, No.2, pp. 408-416.
[3] Sastry S S, Isidori A (1989). “Adaptive control of
linearizable system”. IEEE Trans on Automatic Control,
Vol.34, No.11, pp. 1123-1131
[4] Jia Qingzhong, Liu Yongshan (2007). “Backstepping
design for nonlinear control system with uncertainties”.
Journal of System Simulation, Vol.19, No.21, pp. 5002-
5007. (in Chinese)
[5] Ki-SeokKim, YoudanKim (2003). “Robust backstepping
control for slew maneuver using nonlinear tracking
function”. IEEE Trans on Automatic Control, Vol.11, No.6,
pp. 822-829.

720

You might also like