RBF NN-Based Backstepping Control for Strict

Feedback Block Nonlinear System and Its Application

1,2 2 1
Yunan Hu , Yuqiang Jin , and Pingyuan Cui
1
Department of Astronautics and Mechanics, Harbin Institute of Technology,
Harbin 150001, P.R.China
cui@astro.hit.edu.cn
2
Department of Automatic Control, Naval Aeronautical Engineering Academy,
Yan tai 264001, P.R.China
hya507@sina.com

Abstract. Based on neural networks, a robust control design method is pro-

posed for strict-feedback block nonlinear systems with mismatched uncertain-
ties. Firstly, Radial-Basis-Function (RBF) neural networks are used to identify
the nonlinear parametric uncertainties of the system, and the adaptive tuning
rules for updating all the parameters of the RBF neural networks are derived
using the Lyapunov stability theorem to improve the approximation ability of
RBF neural networks on-line. Considering the known information, neural net-
work and robust control are used to deal with the design problem when control
coefficient matrices are unknown and avoid the possible singularities of the
controller. For every subsystem, a nonlinear tracking differentiator is introduced
to solve the “computer explosion” problem in backstepping design. It is proved
that all the signals of the closed-loop system are uniform ultimate bounded.

1 Introduction

Nonlinear adaptive control has been developed rapidly during last ten years [1~3],and
mostly results limited to the systems in which parametric uncertainties are linear , can
be linearized and the uncertainties satisfy the matched conditions. As a breakthrough
of the difficulties, adaptive backstepping design method can deal with the so-called
mismatched uncertainties [4]. For strict-feedback nonlinear systems, backstepping is a
systematic design method. However, for a class of nonlinear systems in block stan-
dard forms, it is more difficult, fewer results are obtained up to now. Block control
principle is developed on the basis of “block control standard form”. Based on the
block control principle, some design methods are proposed in recent years [5-7].
RBF NN is a typical local approximator and finds wide applications in control sys-
tem design. Generally, only the weight values are updated in most applications, which
is difficult to get ideal approximation when one doesn’t know much about prior
knowledge on the system. In order to solve the problem, reference [8] made some
modifications on the NN updated laws. The author derived the tuning rules for all the
parameters of RBF NN (weights, width and centers of Gaussian functions), but its
implementation of the tuning rules is on the basis of gradient optimization, which
cannot guarantee the stability of the overall system [3]. References [10] and [11] pro-

F. Yin, J. Wang, and C. Guo (Eds.): ISNN 2004, LNCS 3174, pp. 129–137, 2004.
© Springer-Verlag Berlin Heidelberg 2004
130 Y. Hu, Y. Jin, and P. Cui

posed design methods of neural network adaptive control for a class of nonlinear sys-
tems with general uncertainties. But there exist drawbacks in their design method: a)
only can deal with scalar systems at each step of backstepping, cannot be extended to
multi-variable sub system, and controller design of multi-variable system with un-
known control coefficient matrix is a difficult problem; b) highly rely on the ap-
proximation ability of NN and don’t take use of known information efficiently.
In this paper, we discuss the design of Backstepping for a class of nonlinear sys-
tems in block control form and with mismatched uncertainties, solve the problems
mentioned above.

2 Fully Tuned RBF NN

Defination 1. Fully tuned RBF NN: all the parameters of RBF NN (including
weights, widths and centers) are updated.

Assumption 1. Function vector

∆f : Ω 6 R r , Ω belongs to a sub-compact set of

R n . ∀ ε = [ε1 ε 2 " ε r ] ε > 0 , there always exist an optimal Gaussian base

T
, i

function vector
φ* : R n 6 R l and weight matrix W * ∈ R l×r such that

∆f = W *T φ* + ε , ∀x ∈ Ω , (1)

T
  2   2 
where φ* = exp − ς − µ1* / σ 1*2  ... exp − ς − µl* / σ l*2  , µi* , i = 1,2,..., l
    
is the optimal center, l is the number of hidden layer nodes, σ i , i = 1,2,..., l is the
*

optimal affect size, ς ∈ R is the input of RBF NN, and ε is the construction error of
n

NN.
The errors between the optimal values and estimated values will have influence on
the system. The influence is stated in Theorem 1.
~
~ = φˆ − φ* ,
Theorem 1. Define W = Wˆ − W * , φ ~ = µˆ − µ * ,
µ i i i σ~i = σˆ i − σ i* ,
i = 1,2,..., l . The output approximate error of fully tuned RBF NN can be expressed
as:
~
( ) ( )
~ + φˆ ' . ∗ σ~ + d ,
Wˆ T φˆ − W *T φ* = W T φˆ − φˆ 'µ . ∗ µˆ − φˆ 'σ . ∗ σˆ + Wˆ T φˆ 'µ . ∗ µ σ u (2)

where the up bound of d u is:

d u ≤ µ * Wˆ T φˆ 'µ + σ * Wˆ T φˆ 'σ + W *T φˆ 'µ . ∗ µˆ + W *T φˆ 'σ . ∗ σˆ + l W *T . (3)

The meaning of the symbols used in (3) can be founded in reference [12].
 RBF NN-Based Backstepping Control for Strict Feedback Block Nonlinear System 131

3 NN-Based Adaptive Control

3.1 System Description

Consider the following nonlinear system

( ) ( )
X 1 = f1 ( X 1 ) + g1 ( X 1 )X 2 , X 2 = f 2 X 2 + g 2 X 2 X 3 , (4)

"" , X n = f n (X n ) + g n (X n )u , (5)

where X i = X 1T[ " X iT ]

T
( ) ( ) are uncertain functions in the sys-
, f i X iT , g i X iT
[
tem, X i = xi1 " xini ]T
, i = 1,", n , ni is the order of the i th sub-block, and u ∈ R p .
In fact, we can divide the uncertain function into two parts, namely, the nominal part
( ) ( ) ( ) ( ) ( ) ( )
and the uncertain part: f i X i = f i 0 X i + ∆f i X i , gi X i = g i 0 X i + ∆g i X i ,
where gi (X i ) is invertible, f i 0 (X i ) , gi 0 (X i ) are the nominal values, and
∆f i (X i ) , ∆g i (X i ) are uncertain terms.

3.2 Controller Design and Stability Analysis

The goal of the controller design is to cancel the influence of uncertainties on system
performance and track the desired signal X id .
Introduce a new error state z i ∈ R ni , and define

z i = X i − X id , i = 1,", n , (6)

where X id is the desired state trajectory. From (4) - (6), the error state dynamics can
be expressed as

z1 = f1 ( X 1 ) + g1 ( X 1 )X 2 − X 1d . (7)

Step 1: Considering and rearranging the system (7), we have

z1 = f10 ( X 1 ) + g10 ( X 1 )X 2 − X 1d − 1,

(8)

where 1 = −(∆f1 ( X 1 ) + ∆g1 ( X 1 )X 2 ) is introduced by general uncertainties. We ap-

ply a fully tuned RBF NN to compensate the affect. Assume that

1 = W1*T φ1* (Z1 ) + ε1 , (9)

132 Y. Hu, Y. Jin, and P. Cui

where Z1 = X 1T [ X 2T ] is the input of NN, subscript i in W , φ , µ

T
i i ij , σ ij denotes
the NN parameters of the i th sub-system, and ε1 is the construction error of NN.
Taking X 2 as a virtual control of (7), there exists an ideal virtual control
−1
X 2*d = − g10 [
( X 1 ) f10 ( X 1 ) − X 1d + k1 z1 − W1*T φ1* (Z1 ) − ε1 , ] (10)

(
such that z1 = −k1 z1 + g10 ( X 1 ) X 2 − X 2* , where ) k1 > 0 is a design parameter. Be-
cause X 2* is not available, we choose the desired virtual control as
−1
X 2 d = − g10 [
( X 1 ) f10 ( X 1 ) − X 1d + k1 z1 − Wˆ 1T φˆ1 (Z1 ) − v1 , ] (11)

where Ŵ1 , φ̂1 are estimates of W1* and φ1* respectively, and v1 is a robust term, de-
fined by (22). Choose Lyapunov function as

V1 =
1 T
2
1 ~ ~
[ 1 l ~ T −1 ~
z1 z1 + tr W1T Γ W−11W1 + ∑ µ
2 2 i =1
]
1i Γ 1 µ µ1i +
1 l −1 ~ 2
∑ Γ1σ σ 1i ,
2 i =1
( ) (12)

~ ~ = µˆ − µ * , σ~
where W1 = Wˆ 1 − W1* , µ1i 1i 1i 1i = σˆ 1i − σ 1*i , µ̂1i , σ̂ 1i are the estimates of
µ1i* and σ 1i* respectively, i = 1,", l , Γ W1 = Γ W
T
1
> 0 , Γ 1 µ = Γ 1Tµ > 0, Γ1σ > 0 are
design parameters. According to Theorem 1, the derivative of V1 is given by
~
+ z1T g10 ( X 1 )( X 2 − X 2 d ) + z1T d u1 − z1T ε1 + tr W1T Γ W−11W1 
~
V1 = −k1 z1
2
 
~
[ ( ) ~ + φˆ ' . ∗ σ~
+ z1T W1T φˆ1 − φˆ1' µ . ∗ µˆ 1 − φˆ1' σ . ∗ σˆ 1 + Wˆ 1T φˆ1' µ . ∗ µ1 1σ 1( )] (13)

l
( ~
~ T Γ −1 µ
+∑ µ
i =1
T
l
)−1 ~ ~ 
1i 1 µ 1i + z1 v1 + ∑ Γ1σ σ 1i σ 1i
i =1
Let the parameter adaptive laws of NN be

( )
Wˆ 1 = − Γ W1 φˆ1 − φˆ1' µ . ∗ µˆ 1 − φˆ1' σ . ∗ σˆ1 z1T , µˆ 1i = − Γ 1µ φˆ1' µ i Wˆ 1 z1 i , ( ) (14)

( )
σˆ 1i = −Γ1σ φˆ1'σ i Wˆ 1 z1 i , i = 1,", l , (15)

∂ϕ1i '
where δ W1 , δ 1 µ , δ 1σ > 0 are parameter to be designed, φˆ1' µ i = , φˆ1σ i and
∂µ1i
(Wˆ z )
1 1 i denote the i th element of φˆ1' µ , φˆ1' σ and Wˆ 1 z1 . Choose the robust term as

 2 2 2 2
v1 = − z1  Wˆ 1T φˆ1' µ + Wˆ 1T φˆ1' σ + φˆ1' µ . ∗ µˆ 1 + φˆ1' σ . ∗ σˆ1  / η1 , (16)
 
 RBF NN-Based Backstepping Control for Strict Feedback Block Nonlinear System 133

with η1 > 0 . Substituting (14)-(16) into (13) and assuming that ε1 ≤ ε 1H , we can
get that
V1 ≤ −k1 z1 + z1T g10 ( X 1 )z 2 + c1 ,
2
(17)

where c1 =
[(2 + l )η1 ] W *T 2
+
η1
µ1*
2
+
η1
σ1*
2
+
η1
ε12H .
1
4 4 4 4
Step 2: Choosing Lyapunov Function as

Vi = Vi −1 +
1 T
2
1 ~
[~ 1 l ~ T −1 ~
z i z i + tr WiT Γ W−1i Wi + ∑ µ
2 2 j =1
]
ij Γ iµ µij +
1 l −1 ~ 2
(
∑ Γiσ σ ij
2 j =1
) (18)

Similar to the Step 1, we have

+ z iT gi 0 (X i )z i +1 + z i ε iH + z iT v i + z iT d ui + ∑ c j ( )
2 i −1
Vi ≤ −∑ij =1 k j z j
j =1 (19)

≤− ∑ j =1
kj zj ( )
+ z i g i 0 X i z i +1 + ∑ (c ) , j

where ci =
[(2 + l )ηi ] W *T 2
+
ηi
µi*
2
+
ηi
σ i*
2
+
η
j =1
i
ε iH
2
.
i
4 4 4 4
Step n : Choosing Lyapunov function as

Vn = Vn −1 +
1 T
2
1 ~
2
~
[ 1 ~
2
~
z n z n + tr W nT Γ W−1n W n + tr W gT Γ g−1W g ] [ ]
(20)
~ T −1 ~ 
+ tr W gn Γ gnW gn .
 
Taking its time derivative


( ) 
( )
Vn = Vn −1 + z nT  f n 0 X n + g n X n u − Xˆ nd − n
 + tr W

 
~ T −1 ~ 
Γ W
 n W n  n
(21)

= − ∆f n 0 (X n ) + Xˆ nd − X nd  . Denoting gˆ n (X n ) := g n 0 (X n ) + ∆gˆ n (X n )


where n 
[ ]
and assuming that it is invertible. Denoting bij = gˆ n−1 (X n ) , i, j = 1," , nn . For the
simplicity, uncertainties in the sub-system are approximated by RBF NN that only
weight matrix is updated
∆fˆn (X n ) + Xˆ nd − X nd = W n* φn + εWn , ∆g n (X n ) = W g* φ g + ε g ,


[( ) ]
W g = W gij
T
1×l nn ×(nn ×l )
, φ gn = diag (φn ," , φn ) . Choosing control law as

[ ] ( )
u = − g n 0 (X n ) + Wˆ g φg (Z n )  f n 0 (X n ) − Xˆ nd + k n z n + g (Tn−1) 0 X ( n−1) z ( n −1) − Wˆ n φn (Z n ) − v n  , (22)
−1




[
where v n = vn1 ," , vnnn ]T
, v ni = Dεfi sgn (z ni ) + U max sgn (z ni ) ∑ Dεgij , Dεfi ,
nn

j =1
134 Y. Hu, Y. Jin, and P. Cui

Dεgij , i, j = 1," , nn are the up bounds of the NN approximated errors, and

U max > 0 . From (22), we have

( )
ui = ∑ bij  f n 0 (X n ) − Xˆ nd + k n z n + g (Tn −1)0 X ( n−1) z ( n −1) − Wˆ n φn (Z n ) − v n  . (23)

nn

j =1 
  j

Let

( ) ∑ bij  f n0 (X n ) j + Xˆ nd 
nn
ai X n = + kn zn j
( )
+ g (Tn−1) 0 X ( n −1) z ( n −1) 
 j =1 j j

nn

∑  b  ( )
nn nn
+ ij Wˆ nT φn (Z n ) + Dεfi  , hi X n = ∑ bij ∑ Dεgjl , where ⋅ denotes the ab-
j =1
j  j =1 l =1
j

solute of the j th element. Because Dεgjl can be arbitrary small, following the idea of
( )
reference [14], we can assume that 0 ≤ hi X n < 1 . With (23), we can get

ui ≤ ai (X n ) + U max (X n )hi (X n ) ≤ U max (X n ) ,

 a (X ) 
U max (X n ) ≥ max  i n  .
(24)
i =1,",nn 1 − h ( X )
 i n 

Equation (21)can be rewritten as


[ ]
Vn ≤ Vn −1 + z Tn  f n 0 (X n ) + g n 0 (X n ) + ∆gˆ n (X n ) u − Xˆ nd −

n [ ]
+ ∆g n (X n ) − ∆gˆ n (X n ) u  ,

~ ~ (25)
+ tr W nT Γ W−1n W n  + tr W gT Γ g−1W g 
~ ~
   

Because Ŵn , Ŵ gn cannot be guaranteed in the bounded closed set in control process,
we apply the project operator in [15]. Assuming that the feasibility sets of the pa-
rameter space are
  2  2 
≤ β Wn  , ΩWn ∆  Wˆ n Wˆ nT Wˆ n ≤ β Wn + λ.Wn  ,
ΩWn 0 ∆  Wˆ n Wˆ nT Wˆ n
   
 2 
where, β Wn > 0 , λWn > 0 , Ω gn 0 ∆  Wˆ g Wˆ gT Wˆ g ≤ β Wg  ,
 
 ˆ ˆT ˆ 2 
Ω gn ∆  W g W g W g ≤ β Wg + λWg  , with β Wg > 0 , λWg > 0 .
 
The parameter adaptive laws are chosen as:

( 
)
Wˆ n = Γ W n proj Wˆ n , ΦW , Wˆ g = Γ g proj Wˆ g , Φ g ( ) (26)

Defining the project operator as

 RBF NN-Based Backstepping Control for Strict Feedback Block Nonlinear System 135

  ˆ 2  T ˆ
  M − β M ΦM M
Φ −  
( )
2
ˆ , if
M ˆ
M > β M and Φ M M > 0,
T ˆ
ˆ ,Φ
proj M = M 2
M
 δ M Mˆ
Φ , elsewise
 M
where M = W n ,W g , ΦW n = Γ Wn z nφWn
T
and ΦWg = − Γ g z n uT φTg . Using the results
of [15], we have

( )
n n −1
V ≤ − ∑ k n z n
2
+ ∑ cj , (27)
j =1 j =1

{ }
where k = min 2k j . From (27), we can conclude that all the signals in closed-loop
j =1,",n
system are UUB. So we can assume that
1 n
W1 ,",Wn ,Wg 
~ ~
[
1 ~ ~
c0 = sup  ∑ tr W jT Γ W−1j W j + tr W gT Γ g−1W g ] [ ]

 < ∞ . Let c n = kc0 and
ˆ ˆ ˆ 2 j =1 2 
n 1
c = ∑ c j . (27) can be rewritten as V (t ) ≤ V (0)e − kt + c , ∀t ≥ 0 . Consider (20), we
j =1 k
can conclude that
2 2 2
n ~ 2V (t ) n
~ 2V (t ) n 2V (t )
, ∑ σ~ j
∑ Wj
j =1
≤
( )
λmin Γ W−1j
,∑ µ
j =1
j ≤
( )
−1
λmin Γ jµ j =1
≤
( )
λ min Γ −jσ1
, (28)

1 n 2
V (t ) ≥ ∑ zj (29)
2 j =1

k n 2
From (27) and (29), we have V (t ) ≤ − ∑ z j + c . Integrating it yields
2 j =1

∫0 z j (τ ) dτ ≤ 2[V (0) + tc ]/ k , j = 1," , n .

t 2
(30)

From the discussion above, we have the following result,

Theorem 2. Considering the system (4)-(5), if Assumptions 1 holds, using the pro-
posed approach, the following results hold,
(1)The state tracking error of the system z j and the parameter estimated error of NN
are bounded and converge to the neighborhoods of the origins exponentially,
  1 
 2V (0) + c 
  1  2
≤ 
~ ~ ~ n 2 ~ n k 
Ω j =  z j ,W j , µ

j ,σ j ∑ z j
j =1
≤2 V (0) + c , ∑ W j
 k  j =1 ( )
λ min Γ W−1j
,


136 Y. Hu, Y. Jin, and P. Cui

 1   1 
2 V (0) + c  2 V (0) + c  
2 2
k   j = 1," , n
≤ 
k  n ~
≤ 
n
~
∑ µ
j =1
j
( )
λmin Γ −jµ1
,∑ σj
j =1 ( )
−1
λmin Γ jσ 
,

(2)The following inequalities hold,
 1 
lim ∫0t z j (τ ) dτ ≤ 2c / k , ∑ z j
1 2 n 2
≤ 2V (0) + c  .
t →∞ t j =1  k 
Remark 1. From the result (1) of the Theorem 2, we know that adjusting the values of
k i , δ Wi , δ iµ , δ iσ , Γ Wi , Γ iµ , Γiσ , ηi , we can control the convergence rate and the size
of the convergence region.

4 Simulation Example

The missile model with general set of uncertainties considered hereafter is a three-
axes , highly nonlinear model of a BTT missile, The nonlinear dynamic equations are
given as follows:

x1 = f1 ( x1 ) + b1 (x1 )x2 + h1 ( x1 )u , x 2 = f 2 ( x1 , x2 ) + b2u , (31)

x = [α β φ ]T x = [ p q r ]T u = [δ δ δ ]
T
x y z
where 1 , 2 , . The meaning
of the symbols used in this paper can be founded in [12].
In order to validate the effectiveness and validity of the proposed method, nonlinear
6-DOF simulation results for a BTT missile model are presented, The solid lines rep-
resent the command signals, The dotted lines and the dash-dotted lines represent the
simulation results of the adaptive controller under uncertain air parameters with 50
percent (up and down) uncertainty, It is shown that the control system has good sta-
bility, performance and robustness even if the large model uncertainties exist.

Fig. 1. ny tracking curves Fig. 2. β tracking curves Fig. 3. φ tracking curves

 RBF NN-Based Backstepping Control for Strict Feedback Block Nonlinear System 137

5 Conclusions

The main contribution of this paper are: a) relaxed the requirement of SISO in refer-
ences [10] and [11]; b)took the prior knowledge of the system into consideration and
successfully solved the design problem for a class of multi-input multi-output nonlin-
ear systems when the control coefficient matrix is unknown; c)avoided the possible
singularities of the controller; d)Applied Lyapunov stability theory, the parameter up-
dating laws of the fully tuned RBF NN was derived, which guaranteed the stability of
the overall systems; e) by introducing nonlinear tracking differentiators and NNs, the
“computation explosion” is reduced.

References

1. Sastry, S. S., Isidori, A.: Adaptive Control of Linearizable System. IEEE Translations on
Automatic Control, 11 (1989) 1123-1131
2. Seto, D., Annaswamy, A. M., Baillieul, J.: Adaptive Control of Nonlinear Systems with a
Triangular Structure. IEEE Translations on Automatic Control, 7 (1994) 1411-1428
3. Polycarpou, M. M.: Stable Adaptive Neural Control Scheme for Nonlinear Systems. IEEE
Translations on Automatic Control, 3 (1996) 447-451
4. Krstic, M., Kanellakopoulos, I., Kokotovic, P.: Nonlinear and Adptive Control Design.
Wiley–Interscience Publication (1995)
5. Vadim, I. Utkin., De-Shiou, C., Hao-Chi, C.: Block Control Principle for Mechanical Sys-
tems. Journal of Dynamic Systems, Measurement, and Control, 1 (2000) 1-10
6. Loukianov, A., Toledo, B.C., Dodds, S.J.: Nonlinear Sliding Surface Design in the Pres-
ence of Uncertainty. Proceedings of the 14th IFAC, Beijing, P.R.China (1999) 55-60
7. Jagannathan, S., Lewis, F.L.: Robust Backstepping Control of a Class of Nonlinear Sys-
tems Using Fuzzy Logic. Information Sciences, 2 (2000) 223-240
8. Yan, L., Sundararajan, N., Saratchandran, P.: Neuro-controller Design for Nonlinear
Fighter Aircraft Maneuver Using Fully Tuned RBF Networks. Automatica, 8 (2001) 1293-
1301
9. Park, J., Sandberg, I. W.: Universal Approximation Using Radial Basis Function Net-
works. Neural Computation, 2 (1991) 246-257
10. Zhang, T., Ge, S.S., Hang, C.C.: Adaptive Neural Network Control for Strict-Feedback
Nonlinear Systems Using Backstepping Design. Automatica, 12 (2000) 1835-1846
11. Ge, S. S., Wang, C.: Adaptive NN Control of Uncertain Nonlinear Pure-Feedback Sys-
tems. Automatica, 4 (2002) 671-682
12. Jin, Y.Q.: Nonlinear Adaptive Control System Design for Missile. Yantai, P.R.China
(2003)
13. Han, J., Wang, W.: Nonlinear Tracking Differentiator. System Science and Mathematics, 2
(1994) 177-183
14. Ordonez, R., Spooner, J.T.: Stable Multi-input Multi-output Adaptive Fuzzy Control. Pro-
ceedings of the 35th CDC, Japan (1996) 610-615
15. Khalil, H.K.: Adaptive Output Feedback Control of Nonlinear Systems Represented by In-
put-Output Models. IEEE Transactions on Automatic Control, 2 (1996) 177-188