Adaptive backstepping control of a class of mimo systems

Jaime Díaz 1,3, Edgar N. Sanchez2

1
Departamento de Automática y Sistemas Computacionales, Facultad de Ingeniería
Eléctrica. Universidad Central “Marta Abreu” de Las Villas. Santa Clara, Cuba.
2
CINVESTAV, Unidad Guadalajara, Apartado Postal 31-430, Plaza La Luna, Guadalajara,
Jalisco C.P. 45091, México.
3
Departamento de Electrónica, Instrumentación y Control, Facultad de Ingeniería
Electrónica y Telecomunicaciones. Universidad del Cauca. Popayán, Colombia.
[email protected].

Abstract

A procedure of adaptive backstepping control design introduced by Choi and Farrel [4] for
a class of SISO nonlinear systems diffeomorphic to output feedback nonlinear systems is
extended to a class of MIMO systems. The proposed scheme provides semiglobal stability
and confines the tracking error to a small region around the origin.

I Introduction

In the last years, there has been an enormous interest in the study of adaptive control of
uncertain nonlinear systems. Particularly, and more recently, many works undertake three
important issues. First, they consider the problem of unmatched uncertainties. A second
problem is that, in general, the bound of the denominated approximation error caused by
function approximator in adaptive systems is unknown. In third place, in many practical
control systems not all the states are available for feedback. To tackle these problems one
of the developed procedures are the adaptive backstepping control design methods for
output feedback systems [1, 2, 3]. In several of the schemes the unknown nonlinearities are
approximated function approxima tors, such that fuzzy systems and neural networks, in
which the bound of the approximation error are assumed finite but unknown [4, 5, 6].

The purpose of this paper is to extend to a class of MIMO systems the adaptive
backstepping control design method introduced by Choi and Farrel [4] for a class of SISO
nonlinear systems diffeomorphic to output feedback nonlinear systems, where only the
output is measured. This scheme provides semiglobal stability and confines the tracking
error to a small region around the origin

II Problem Statement

Consider the MIMO system

x1 = A1 x1 + f10 ( y ) + f1 ( y ) + B1n1 u1
•

Μ
x p = Ap x p + f p0 ( y ) + f p ( y ) + B p p u p
• n

Μ
x s = As xs + fs 0 ( y ) + fs ( y ) + Bsns us
•

y p = Cp x p 1≤ p ≤ s (1)

The system is built up of s subsystems, where C p = [1,0,L ,0] of dimension 1× np ,

[ ]
T
x p = x p1, Λ , x pnp , Ap are n p × n p matrices in the observer's Brunovski form, where n p is
the observability index, 1 ≤ p ≤ s , with n1 ≥ L ≥ np ≥ L ≥ ns . And, where f p0 ( y ) ,
1 ≤ p ≤ s , are known smooth vector fields and f p ( y ) are unknown smooth vector fields,
and Bp p = [0,L ,0,1] . The output is y = [ y1,L , ys ] . Each subsystem of (1) can be written
n T T

as

x p = Ap x p + ∑ Bip ( f pi 0 ( y ) + f pi ( y )) + B p p u p (t )
np
• n

i =1

y p = Cp x p 1≤ p ≤ s (2)

where it is assumed that the unknown functions f pi ( y ) are bounded for all y ∈ R . Each
s

unknown function will be approximate on-line by a linearly parameterized approximation

network, a fuzzy system for example. The inherent approximation error is defined by [4]

ε f i ( y ) = φ f i ( y )θ * f i − f pi ( y ) (3)
p p p

where φ f i ( y ) is a row vector of known functions in y, and θ * f i represents the optimal

p p

parameters values that minimize the approximation error on an approximation region D

over which high control performance is desired. It is worth to note that, under conditions
given, the inherent approximation error is bounded even outside de region D. Then, the
actual estimation error is

φ f i ( y )θˆf i − f pi ( y ) = φ f i ( y )θ f i + ε f i ( y )
~
(4)
p p p p p

where θˆf i are de estimated parameters and θ%f i = θˆf i −θ * f i .

p p p p
III Adaptive observer design

The adaptive observer of subsystem p is given by [4]

( ( ))
np np

xˆ p = Ap xˆ p + ∑ B f ( y ) + fˆpi y,θ f i + B p p u p (t ) − ∑ ζ f i θ f pi − K p (yˆ p − y p )

•
• i n i
p p0 (5)
p p
i =1 i =1

yˆ p = C p xˆ p 1≤ p ≤ s

Here ζ f i , of dimension n p × m f i , is defined by

p p

•
ζ f pi = Ap 0ζ f i − B iφ f i ( y ) (6)
p p

Where Ap 0 = Ap − K pC p . The gain vector K p is designed such that Ap0 is stable. The
parameter update law is given by

•
θ f pi (
= η ζ Tfi C Tp e p0 − σθ f i
p p
) (7)

where ep 0 = yˆ p − y p is the observation error for subsystem p, and σ > 0 , η > 0 are design
parameters. Defining ~ x p = xˆ p − x p , the state error equation can be written as

np • np

x~ p = Ap0 ~x p + ∑ B iφ f i θ f i − ∑ ζ f i θ f pi + d p (t )
• ~
(8)
p p p
i =1 i =1

np

where d p (t ) = ∑ B iε f i .
p
i =1

Define a “augmented estimation error” as

np
x p + ∑ ζ f i θ~f i
ξ~p = ~ (9)
p p
i =1

then, using (6), (7) can be shown that

~•
ξ p = Ap0ξ p + d p (t )
~ (10)

The observation error for subsystem p is given by

 np

ep 0 = yˆ p − y p = C p x~p = C pξ~p − C p ∑ζ f θ~f i

p
i
p
(11)
i =1

Theorem 3.1: The system described by (1), the nonlinear observer given by (5) with the
update laws (7) have the following properties:

•
1. θ~f i , θ f i , θ f pi ∈ L∞ , 1 ≤ i ≤ np , 1 ≤ p ≤ s
p p

2. x~p , ep 0 , ∈ L∞ 1≤ p ≤ s
3. e0 is (d T
d + σ ) -Small in the Mean Square Sense (m.s.s.) [7], where
e0 = [e10,Λ , es 0 ] , d = [d1 ,Λ , d s ] .
T T

Proof: Define the function positive definite in θ%f i

p

n
1 p ~T ~ 1~ ~
V p0 = ∑
2η i =1
θ f i θ f i + ξ pT Ppξ p
p p 2
(12)

where Pp is the positive definite such that ATp 0 Pp + Pp = −2 I . It can be shown [4] that the
time derivative of V p 0 holds the following inequality.

• np
V p0 ≤ −ep2 0 − ξ~pT ξ~p + ξ~p e p0 + λ pm ξ~p d p − σ ∑ θ~fTi θ f i (13)
p p
i =1

where λpm is the maximum eigenvalue of Pp .

After some algebraic manipulations [4] we can arrive to the following expression:

1~ ~ σ ~ ~ σ
• np np 2
1
V p0 ≤ − e2p0 − ξ pT ξ p −
2 4 2
∑
i =1
θ fTi θ f i +
p p 2
∑ θ * f i + λ2pmd 2p
i =1
p
(14)

Choosing 0 < β p < min {ση ,1 2 λ2pm }, (14) becomes

σ
• np 2
V p0 ≤ − β pV p 0 +
2
∑θ *
i =1
f pi
+ λ2pmd 2p (15)

Now, let us define the Lyapunov function for the system (1)

s
V0 = ∑V p 0 (16)
p =1
Then,

σ
• s np 2 s
V 0 ≤ − βV0 +
2
∑∑ θ * f i + ∑ λ2pmd 2p
p =1 i =1
p
p=1
(17)

where β = max { β1 ,L , β s } . Ec. (17) implies that V ∈ L∞ , hence ξ% ~ , θ• i ∈ L . Since

p θ fi
fp ∞
p

φ f i ( y ) is bounded, from (6), ζ f i are bounded since Ap 0 is stable. From (12) ~

x p ∈ L∞ and
p p

also ep 0 = C p x%p . From (16), by integration of (14) we have

t +T
 s np 2  t +T
t +T

∫ e e dτ ≤ 2[V0 (t ) − V0 (t + T )] + λ ∫ d d dτ +  ∑∑
 θ * f i  ∫ σdτ
T 2 T
0 0 (18)
 t
p
t t p =1 i =1

where λ = min {λ1m ,L , λsm } . Then, finally

t +T t +T

∫ e e dτ ≤ c ∫ (d d + σ )dτ + c1
T T
0 0 0 (19)
t t

This implies that e0 is (d T d + σ ) -Small in the m.s.s. sense [7].

IV Backstepping controller design

For the p subsystem the ( j, k ) -th element equation is expressed by

( )
•
ζ f pi ( j ,k ) = −k pjζ f i (1,k ) + 1 − δ n p j ζ f i ( j +1,k ) − δ ijφ f i ( k ) (20)
p p p

for 1 ≤ j ≤ n p , 1 ≤ k ≤ m f i , 1 ≤ p ≤ s , where ζ f i ( j ,k ) is the ( j , k ) th element of ζ f i , φ f i (k ) is

p p p p

the k-th element of φ f i . Where k pj is the j-th element of K p , whereas δ ij = 1 for i = j and
p

δ ij = 0 for i ≠ j . From (5), it is possible to write,

np •

∑ ζ f i θ f pi = v p ζ f 1 , Λ , ζ n p ep 0 + w p  ζ f 1 ,Λ , ζ n p ,θ f 1 , Λ ,θ n p  (21)

i =1
p  p fp   p fp p fp 

T T
where v p = v p1 ,L , v pnp  y w p =  w p1,L , w pn p 
   
  
Define ζ f i [ j ] = ζ f i ,Λ , ζ f i , the j-th row vector of matrix ζ f i . Then, each
p ( j ,1 ) p ( j ,n ) 
p
 p  p

component of v p and wp are expressed by

 m f pi
np 
v pj = η ∑  ∑ ζ f i ζ f i  (22)

i =1 k =1
p ( j ,k ) p (1,k ) 

 

 m f ip
np 
w pj = ση∑  ∑ζ f i θ f i  (23)

i =1 k =1
p ( j ,k ) p (k ) 

 

Using the notation of (21), the observer (5) can be rewritten as

•
( )
xˆ p1 = xˆ p 2 + f p10 ( y ) + fˆp1 y,θ f 1 + w p1 − (k p1 + v p1 )ep 0
p

•
xˆ p 2 = xˆ p3 + f p20 ( y ) + fˆ (y,θ )+ w
p
2
f p2 p2 − (k p2 + v p 2 )e p0
Μ
n n

 fp 
(
xˆ pn p = f p0p ( y ) + fˆp p  y ,θ n p  + wpn p + u p − k pnp + v pnp e p 0
•
) (24)

Assumption 3.1: We assume that the desired output trajectories are given by bounded
(n1 − 1) -time differentiable functions where n1 the major observability index is, and that y (pri )
are known at time t for i = 0,L , n p − 1 .

Based on the procedure developed in [4] a coordinate transformation is designed as

z p1 = y p − y pr = x p1 − y pr
z pi = xˆ pi − y (pri −1) − α pi i = 2, Λ n p (25)

where

(
α p 2 = −γz p1 − f p10 ( y ) − fˆp1 y,θ f 1 − ρz p1λTp1λp1
p
) (26)

where λp1 is a s × 1 vector to be defined later. According to (25), we have

z p1 = x p1 − y (pr1)
• •

= xˆ p2 + f p10 + fˆp1 − ~ (
x p 2 + φ f 1θ f 1 + d p1 − y (pr1)
p
~
p
) (27)
By using f p10 + fˆp1 = −α p 2 − γz p1 − ρz p1λTp1λ p1 ,

•
z p1 = z p2 − γz p1 − ρz p1λTp1λ p1 + λTp1ε~2 (28)

T
where λ p1 = − E p , being E p a unitary vector defined by E p = δ p1,L , δ pk ,L ,δ ps  , and
ε~2 = [ε~21,Λ , ε~2 s ] , ε~2 p = x~p 2 + φ f 1θ~f 1 + d p1 , 1 ≤ p ≤ s .
T
p p

For i=2

•
z p 2 = xˆ p 2 − y (pr2 ) − α p 2
• •
(29)

Noting that α p 2 = α p 2  y , z1 ,θ f 1 , Λ ,θ n p ,ζ f l : 1 ≤ k ≤ m f l ,1 ≤ l ≤ n p 
 p fp p (1,k ) p


m
∂α p 2 • ∂α p 2 • ∂α p 2 • ∂α p 2 •
np np i
• fp

α p2 = z p1 + ∑ θ f p + ∑∑
l ζ + y (30)
∂z p1 l =1 ∂θ f l l =1 k =1 ∂ζ f l
f pl
(1,k ) ∂y
p p (1,k )

Then we have

( )
z p 2 = xˆ p 3 + f p20 ( y ) + fˆp2 y,θ f 2 + wp 2 − (k p 2 + v p 2 )e p 0 − y (pr2)
•

∂α p2
−
∂z p1
(− γz p1 + z p 2 − ρz p1λTp1λp1 + λTp1ε~2 )

∂α p 2
( )
np

−∑ η ζ Tf l C Tp e p0 − σθ f l
l =1 ∂θ f l p p
p

( −k ζ )
m
np
∂α p 2
f pi

− ∑∑ f pl (1,k )
+ ζ f l (2,k ) − δ l 1φ f l ( k )
k =1 ∂ζ f l
p1
p p
l =1
p (1,k )
∂α p 2 •
− y (31)
∂y

Let’ do

α p3 = −γz p 2 − z p1 − f p20 ( y ) − fˆp2 y ,θ f 2 − w p2 ( p

)
∂α p 2
− ρz p 2 (λTp2 λp 2 + ν 2p2 ) − η ∑
np
σθ f l
l =1 ∂θ f l
p
p
 ∂α p 2
+
∂z p1
(
z p 2 − γz p1 − ρz p1λTp1λ p1 )
m
∂α p2
( −k ζ )
np f pi

+ ∑∑ f pl (1,k )
+ ζ f l (2,k ) − δ l 1φ f l ( k )
∂ζ f l
p1
p p
l =1 k =1
p (1,k )
∂α p 2
+ Fˆ (32)
∂y

where Fˆ = [Fˆ1 ,Λ , Fˆs ] , Fˆ p = xˆ p 2 + f p10 + fˆ p1 . We have

T

− γz p 2 − z p1 − ρz p 2 (λTp2 λp 2 + ν 2p2 )
• (2 )
z p 2 = xˆ p 3 − α p 3 − y pr
∂α p2 T
− λp1ε~2
∂z p1
np
∂α p 2
− (k p 2 + v p 2 )e p0 − ∑ ηζ Tf l C Tp e p0
l =1 ∂ θ fl
p
p

∂α p 2   •
+  Fˆ − y  (33)
∂y  

From (27)

•
~
y p = xˆ p2 + f p10 + fˆp1 − x~p 2 + φ f 1θ f 1 + d p1( p p
) (34)

•
hence Fˆ − y = ε~2 . Then

z p 2 = z p 3 − γ z p 2 − z p1 − ρz p 2 (λTp2 λp 2 + ν 2p2 ) + λTp 2ε~2 − ν p 2ep 0

•
(35)

where, we define
∂α p2
ν p2 = (k p 2 + v p2 ) + ∑
np
ηζ Tf l C Tp e p0 (36)
l =1 ∂θ f l p
p

∂α p 2 ∂α p2 T
λTp 2 = − λp1 (37)
∂y ∂z p1

For i=3

•
z p 3 = xˆ p 3 − y (pr3 ) − α p 3
• •
(38)
Noting that

m
∂α p 3 • ∂α p3 • ∂α p 3 • ∂α p 3 • s ∂α p3 •
np np 2 l
• 2 fp

α p3 = ∑ z pj + ∑ θ f p + ∑∑∑
l ζ + y +∑ xˆ p 2 (39)
j =1 ∂z pj l =1 ∂θ f l l =1 j =1 k =1 ∂ζ f l
f pl
( j ,k ) ∂y p =1 ∂x
ˆp2
p p ( j ,k )

We have

( )
z p 3 = xˆ p 4 + f p30 ( y ) + fˆp3 y, θ f 3 + wp 3 − (k p3 + v p 3 )e p 0 − y (pr3)
•

∂α p 3
(z − γz pj − z p , j −1 − ρz pj (λTpj λ pj + ν 2pj ) + λTp2ε~2 − ν pj e p0 )
2
−∑ p , j +1
j =1 ∂ z pj

∂α p 3
( )
np

−∑ η ζ Tf l C Tp e p0 − σθ f l
l =1 ∂θ f l p p
p

( −k ζ )
m
np 2
∂α p 3
f pl

− ∑∑∑ f pl (1,k )
+ ζ f l ( j +1,k ) − δ ljφ f l (k )
j =1 k =1 ∂ζ f l
pj
p p
l =1
p ( j ,k )
∂α p 3
− (Fˆ − ε~2 )
∂y
∂α
− ∑ p 3 (xˆ p 3 + f p20 + fˆp2 + w p2 − (k p 2 + v p2 )e p 0 )
s
(40)
p=1 ∂xˆ p2

Let’ do

α p 4 = −γz p3 − z p 2 − f p30 ( y ) − fˆp3 y ,θ f 3 − wp 3 ( p

)
∂α p 3
− ρz p 3 (λTp 3λ p 3 + ν 2p3 ) − η ∑
np

σθ f l
l =1 ∂θ f l
p
p

∂α p3
(z − γz pj − z p , j −1 − ρz pj (λTpj λ pj + ν 2pj ))
2
+∑ p , j +1
j =1 ∂z pj

( −k ζ )
m
np 2 f lp
∂α p 3
+ ∑ ∑∑ f pl (1,k )
+ ζ f l ( j +1,k ) − δ ljφ f l (k )
∂ζ f l
pj
p p
l =1 j =1 k =1
p ( j ,k )
∂α p 3
+ Fˆ
∂y
∂α p3
(xˆ + f p20 + fˆ p2 + w p2 )
s
+∑ (41)
∂xˆ p 2
p3
p =1

We have
z p 3 = xˆ p 4 − y (pr3) − α p 4 − γz p3 − z p 2 − ρz p 3 (λTp 3λ p 3 + ν 2p3 )
•

∂α p 3 ~ 2
∂α
+ ε 2 − ∑ p 3 λTp 2ε~2
∂y j =1 ∂z pj

∂α p3
− (k p 3 + v p 3 )e p0 + ∑
2
ν pj ep 0
j =1 ∂z pj
np
∂α p 3 ∂α p 3
(k p 2 + v p2 )e p0
s
−∑ ηζ Tf l C Tp e p0 + ∑ (42)
l =1 ∂θ f l p
p =1 ∂ x
ˆ p2
p

Let us define

∂α p 3 ∂α ∂α
ν p3 = (k p 3 + v p 3 ) ν pj + ∑ p 3 ηζ Tf l C Tp − ∑ p 3 (k p2 + v p 2 ) (43)
2 np s
−∑
j =1 ∂ z pj l =1 ∂θ f l p =1 ∂x
ˆ p2
p
p

∂α p 3 2
∂α
λTp 3 = − ∑ p 3 λTp 2 (44)
∂y j =1 ∂z pj

Then we have

z p 3 = z p 4 − γz p3 − z p 2 − ρz p 3 (λTp 3λ p 3 + ν 2p3 ) + λTp 3ε~2 − ν p 3ep 0

•
(45)

For i ≥ 4 ,

•
z pi = xˆ pi − y (pri ) − α pi
• •
(46)

Noting that

m
• ∂α •
i −1 np
∂α •
n p i −1
∂α pi •
f lp
∂α pi • i −1 s ∂α pi •
α pi = ∑ pi z pj + ∑ pi θ f pl + ∑∑∑ ζ + y + ∑∑ xˆ pj
j =1 ∂z pj l =1 ∂ θ f l l =1 j =1 k =1 ∂ζ f l
f pl
( j ,k ) ∂y j = 2 p =1 ∂ x
ˆ pj
p p ( j ,k )
(47)

We have

( )
z pi = xˆ p ,i +1 + f pi 0 ( y ) + fˆpi y ,θ f 3 + w pi − (k pi + v pi )e p0 − y (pri )
•

∂α pi
)
i −1
−∑ (z p , j +1 − γz pj − z p , j −1 − ρz pj (λTpj λpj + ν 2pj ) + λTpjε~2 − ν pj e p0
j =1 ∂ z pj
 ∂α pi
( )
np
−∑ η ζ Tf l C Tp e p0 − σθ f l
l =1 ∂θ f l p p
p

m
∂α pi
( −k ζ )
np i −1 f pl

− ∑∑∑ f pl (1,k )
+ ζ f l ( j +1,k ) − δ ljφ f l (k )
j =1 k =1 ∂ζ f l
pj
p p
l =1
p ( j ,k )
∂α pi
− (Fˆ − ε~2 )
∂y
∂α
− ∑∑ pi (xˆ p , j +1 + f pj0 + fˆpj + wpj − (k pj + v pj )e p 0 )
i −1 s
(48)
j =2 p =1 ∂xˆ pj

Let us do

α p,i +1 = −γz pi − z p ,i −1 − f pi 0 ( y ) − fˆpi y ,θ f 3 − wpi ( p

)
∂α pi
− ρz pi (λTpi λ pi + ν 2pi ) − η ∑
np
σθ f l
l =1 ∂θ f l p
p

∂α
i −1
+ ∑ pi (z p , j +1 − γz pj − z p , j −1 − ρz pj (λTpj λpj + ν 2pj ))
j =1 ∂ z pj
m
∂α pi
( −k ζ )
n p i −1 f lp

+ ∑∑∑ f pl (1,k )
+ ζ f l ( j +1,k ) − δ ljφ f l (k )
j =1 k =1 ∂ζ f l
pj
p p
l =1
p ( j, k )
∂α pi
+ F̂
∂y
i −1 s
∂α
+ ∑∑ pi (xˆ p , j +1 + f pj0 + fˆ pj + w pj ) (49)
j =2 p =1 ∂xˆ pj

we have

z pi = xˆ p,i +1 − y (pri ) − α p ,i +1 − γz pi − z p ,i −1 − ρz pi (λTpi λ pi + ν 2pi )

•

∂α pi ~ i −1 ∂α pi T ~
∂y 2 ∑
+ ε − λpjε2
j =1 ∂ z pj
i −1
∂α pi
− (k pi + v pi )e p0 + ∑ ν pj e p0
j =1 ∂z pj

∂α pi i −1 s
∂α
ηζ Tf l C Tp e p0 + ∑∑ pi (k pj + v pj )e p0
np
−∑ (50)
l =1 ∂θ f l p
j = 2 p =1 ∂ x
ˆ pj
p

Let us define
 ∂α pi ∂α pi ∂α
ν pi = (k pi + v pi ) − ∑ ( )
i −1 np i −1 s
ν pj + ∑ ηζ Tf l C Tp + ∑∑ pi k pj + v pj (51)
j =1 ∂z pj l =1 ∂θ f l p
j =2 p =1 ∂x
ˆ pj
p

∂α pi i −1 ∂α pi T
λTpi = −∑ λpj (52)
∂y j =1 ∂ z pj

We have

z pi = z p ,i +1 − γz pi − z p ,i −1 − ρz pi (λTpi λpi + ν 2pi ) + λTpiε~2 − ν pie p 0

•
(53)

Finally, for i= n p we have

z pn p = f p 0p ( y ) + fˆp p  y, θ n p  + w pn p + u p (t ) − k pn p + v pnp e p0 − y prp

• n n

 fp 
(n )
( )
n p −1
∂α pnp
−∑ ( z p , j +1 − γz pj − z p , j −1 − ρz pj (λTpj λ pj + ν 2pj ) + λTpjε~2 − ν pj e p0 )
j =1 ∂z pj

∂α pnp
( )
np

−∑ η ζ Tfl C Tp e p 0 − σθ f l
l =1 ∂θ f l p p
p

( −k ζ )
m
np n p −1 f lp
∂α pn p
− ∑ ∑∑ f pl (1,k )
+ ζ f l ( j +1,k ) − δ ljφ f l (k )
∂ζ f l
pj
p p
l =1 j =1 k =1
p ( j ,k )
∂α pn p
−
∂y
(Fˆ − ε~2 )
np −1 s
∂α pnp
− ∑∑
∂xˆ pj
(xˆ p , j +1 + f pj0 + fˆ pj + wpj − (k pj + v pj )e p 0 ) (54)
j = 2 p =1

( (n )
)
z pn p = −γz pn p − z p,n p −1 − ρz pn p λTpn p λpn p + ν 2pnp −α p,n p +1 + u p (t ) − y prp + λTpn p ε~2 − ν pnp e p 0
•

(55)

where ν pnp y λpn p are defined in the same manner of (51) and (52), respectively. Choose
the control input by

(n )
u p (t ) = α p ,n p +1 + y prp (56)

then, we have

•
(
z pn p = −γz pn p − z p,n p −1 − ρz pn p λTpn p λpn p + ν 2pnp + λTpnp ε~2 − ν pnp e p 0 ) (57)
Therefore, the transformed closed-loop system is given by

z p = A z p − ϕ (z p , Λ p ,ν p ) + ΛTpε~2 − ν p ep 0
•
(58)

where

−γ 1 0 L 0 0 0
 −1 −γ 1 L 0 0 0
 
A= O  (59)
 
0 0 0 L −1 −γ 1 
 0 0 0 L 0 −1 −γ 

[ ( T 2
) T
(
ϕ (z p , Λ p ,ν p ) = ρz p1 λp1λ p1 + ν p1 ,Λ , ρz pnp λ pnp λ pnp + ν pn p
2
)]
T
(60)

[ ]T
Where Λ p = λp1 ,Λ , λpn p , ν p = ν p1, Λ ,ν pn p
T T T
[ ]
T

The stability of the control system is provided in the following theorem.

Theorem 4.1: The system described by (1) with control law given by (56), and state
observer (5) with parameter update laws given by (7), is stable in the sense that

[
1. z ∈ L∞ , where z = z1T ,Λ , zsT . ]
T

[
2. z is µ 2 -Small in the m.s.s. sense, where µ = e0 + ε~2 . And where e0 = e1 , Λ , es ]
T

Proof: Define the function positive definite

1 T
Vcp = z z (61)
2 p p

The time derivative of Vcp becomes

V cp = zTp (A z p − ϕ (z p , Λ p ,ν p ) − ν p ep 0 + ΛTpε~2 )
•
(62)

Since zTp Az p = −γ zTp z p ,

 1 2  np 2
np
1
V cp = −γzTp z p − ρ ∑  z 2piν 2pi + z piν pi ep 0 +
•
ep 0  + ep0
i =1  ρ 4ρ 2  4ρ
  1 ~T ~  n p ~T ~
np

− ρ ∑  z 2pi λTpi λpi − z pi λTpiε~2 +

1
ε2 ε2  + ε2 ε2 (63)
i =1  ρ 4ρ 2  4ρ

• np 2 n
V cp ≤ −γz Tp z p + ep 0 + p ε~2T ε~2 (64)
4ρ 4ρ

Define now Lyapunov function for system (1)

1 s T
Vc = ∑ zp z p
2 p =1
(65)

It can be shown that

s
n ~T ~
∑n e
• 1
V c ≤ − γz T z + 2
+ ε2 ε2 (66)
4ρ 4ρ
p p0
p=1

•
V c ≤ − γz T z +
4ρ
(
n T
e0 e0 + ε~2T ε~2 ) (67)

µ (t )
• n 2
V c ≤ − γz T z + (68)
4ρ

s
where n = ∑ n p , µ = e0 + ε~2 .
p =1

By Theorem 3.1 µ is bounded. Hence z ∈ L∞ and then Vc ∈ L∞ . From (51), we can show
that

t +T t +T
1 n
∫ zT z dτ ≤ Vc ( t ) − Vc ( t + T ) + ∫ µ dτ
2
(69)
t
γ 4γρ t

This implies that z is (d T d + σ ) -Small in the m.s.s. sense

V Conclusion

In this paper, a procedure of adaptive backstepping control design has been presented for a
class of MIMO nonlinear systems where the state is not directly accessible. The
approxima tors used to estimate the unknown nonlinear functions improve the performance
of the control system allowing achieving small tracking errors. However, outside the
approximation region, the stability is not guaranteed because the inherent approximation
error can become very large. This difficulty can be overcome with a supervisory control
that forces the states to be confined inside a permissible space. Nevertheless, in general, the
determination of this permissible space requires knowledge, even though partial, of the
plant. Future work will be dedicated to this problem.

References

[1] M. Krstic, I. Kanellakopoulos, and P. Kokotovic, Nonlinear and Adaptive Control

Design. New York: Wiley, 1995.

[2] Zhong-Ping Jiang, Guangyu Zhou. “Decentralized Tracking with Disturbance

Attenuation by Nonlinear Output Feedback.” Proceedings of the American Control
Conference, Arlington, VA June 25-27, 2001

[3] Murat Arcak, Petar Kokotovic. “Robust Output-Feedback Design Using a New Class of
Nonlinear Observers.” Proceedings of the 39th IEEE Conference on Decision and Control,
Sydney, Australia, December 2000.

[4] Jin Young Choi and Jay A. Farrell. “Adaptive Observer Backstepping Control Using
Neural Networks.” IEEE Transactions on Neural Networks, Vol. 12, No. 5, pp 1103-1112,
September 2001

[5] M.M. Polycarpou. “Stable adaptive neural control scheme for nonlinear systems.” IEEE
Transactions on Automatic Control, Vol. 41, pp 447-451, 1996.

[6] Julian Stoev , Jin Young Choi, and Jay Farrell. “Adaptive neuro control for output
feedback nonlinear systems.” Proceedings of the American Control Conference, Arlington,
VA June 25-27, 2001

[7] P. A. Ioannou and J. Sun, Robust Adaptive Control. Englewood Cliffs, NJ: Prentice-
Hall, 1996.