You are on page 1of 26

Available online at www.sciencedirect.

com

ScienceDirect
Fuzzy Sets and Systems 263 (2015) 126
www.elsevier.com/locate/fss

Observer-based fuzzy adaptive control for multi-input multi-output


nonlinear systems with a nonsymmetric control gain matrix
and unknown control direction
Wuxi Shi a,b,
a School of Electrical Engineering and Automation, Tianjin Polytechnic University, Tianjin, 300387, China
b Tianjin Key Laboratory of Advanced Technology of Electrical Engineering and Energy, Tianjin, 300387, China

Received 5 December 2013; received in revised form 17 May 2014; accepted 19 May 2014
Available online 24 May 2014

Abstract
In this study, we present an observer-based fuzzy adaptive control scheme for a class of uncertain multi-input multi-output
nonlinear systems with a nonsymmetric control gain matrix, unknown control direction, and unmeasured states. In this scheme,
fuzzy systems are used to approximate the unknown nonlinear functions and an adaptive fuzzy state observer is designed to estimate
the unmeasured states. The parameter adaptive laws are designed by fuzzy basis functions rather than by its filtering and a robust
control term is used to compensate for the lumped errors. The proposed method solves the problem of the nonsymmetric control
gain matrix and it requires no a priori knowledge of the control direction for the states that are unmeasured. We prove that all
of the signals in the resulting closed-loop systems are bounded and that the tracking errors converge asymptotically to zero. Two
simulation examples are used to demonstrate the effectiveness of the proposed scheme.
2014 Elsevier B.V. All rights reserved.
Keywords: Fuzzy adaptive control; MIMO nonlinear system; Nussbaum gain; State observer

1. Introduction
In control engineering, most practical systems are multi-input multi-output (MIMO) systems. In these systems, the
control problem is highly complex due to the coupling between inputs and outputs. It becomes even more difficult
to deal with these systems if they are nonlinear and uncertain. Because of these difficulties, relatively few results are
available for a general class of MIMO systems compared with single-input single-output (SISO) systems.
Thanks to the universal approximation property [1,2], several adaptive fuzzy control schemes have been developed for MIMO nonlinear systems in the past two decades [323]. To handle approximation errors and external
* Correspondence to: School of Electrical Engineering and Automation, Tianjin Polytechnic University, Tianjin, 300387, China.
Tel./fax: +86 22 83955415.
E-mail address: shiwuxi@163.com.

http://dx.doi.org/10.1016/j.fss.2014.05.015
0165-0114/ 2014 Elsevier B.V. All rights reserved.

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

disturbances, the fuzzy adaptive controller was augmented by a robust compensator in some of the proposed schemes
[3,4,7,9,10,12,15]. Furthermore, to obtain a robust compensator and to facilitate performance analyses of these systems, some of the aforementioned results were obtained under the restrictive assumption that the control direction was
known a priori. For example, in [9,10,12,15,16], it was assumed that the control gain matrix was positive definite,
while a restriction was made for the smallest singular of the gain matrix in [11]. There are particular cases where the
control direction can be known a priori [24], but the assumption regarding the control direction does not appear to be
realistic in a general case [25]. If there is no a priori knowledge of the control direction, the adaptive fuzzy control of
these systems becomes much more difficult. The pioneering research in this area involved the development of a class
of first order linear systems [26], where the Nussbaum-type gain was originally proposed, and the Nussbaum-type
function has been used effectively for solving the unknown control direction problem for MIMO nonlinear systems
[13,14,27,28] and for systems in the parameter-strict feedback form [2931]. However, in [13,14], the restrictive assumption required that the control gain matrix had nonzero leading principal minors. In addition, the control direction
was assumed to be positive or negative, and the case with an indefinite direction was not considered. In [27,28], the
two main limitations were that the structure of nonlinear systems comprised a lower triangular and that the boundedness of the control gains were required to be known. For systems in the parameter-strict feedback form, the key
assumption was that the control gains were constant [29,31] or bounded [30].
Another key assumption of some of the schemes mentioned above is that all of the state variables of the systems
are available for measurement. In many control engineering problems, however, the state variables may be partially
unavailable or totally unavailable. In these cases, output feedback control schemes must be employed to obtain the
desired performance. In the last decade, many adaptive fuzzy output feedback control schemes have been developed
for a class of nonlinear systems [3242]. However, these results can only be applied to a relatively simple class of
nonlinear SISO systems. Due to the complexity of MIMO nonlinear systems, most of the schemes developed for
SISO systems cannot be extended directly to MIMO systems. In recent years, several observer-based adaptive fuzzy
control schemes have been developed for a class of MIMO nonlinear systems [3,4,10,17,19,22,23] and for a class of
strict feedback nonlinear systems [18,20,21,23,43]. However, some of the schemes proposed above require strictly
positive real (SPR) conditions on the estimation error dynamics so they allow the filtering of the fuzzy basis functions,
which makes the dynamic order of the system controller/observer very large [32]. To avoid the filtering of fuzzy basis
functions, an observer-based adaptive fuzzy control scheme was developed for a class of uncertain nonlinear systems
[40,41]. However, the use of the assumption u L is not acceptable before the stability analysis [44]; in addition, it
is necessary to assume that the control direction is known a priori. In [17], fuzzy adaptive output feedback control was
developed for a class of uncertain MIMO nonlinear systems with unknown control direction. However, results can
only be obtained in cases where the control gain coefficient was an unknown constant and the filtering of fuzzy basis
functions was used to design the adaptive laws and control law. The schemes in [18,20,21,23,43] do not require SPR
conditions, thus they do not result in the filtering of the fuzzy basis functions. However, it is necessary to assume that
the control direction is known a priori. To the best of our knowledge, however, few previous studies have considered
the class of MIMO uncertain affine nonlinear systems with a nonsymmetric control gain matrix, unknown control
direction, and unmeasured states.
Based on previous results, we developed a fuzzy adaptive output control scheme for a class of MIMO nonlinear
systems with nonsymmetric control gain matrix, unknown control direction, and unmeasured states. In the control
design, fuzzy systems are used to approximate the unknown nonlinear functions and an adaptive fuzzy state observer
is designed to estimate the unmeasured states. A robust control term is used to compensate for the lumped errors.
Although the stability analysis is performed using the SPR conditions, the parameter adaptive laws are designed by
fuzzy basis functions rather than by its filtering. The main original aspects of the proposed control scheme are as
follows: (i) it can solve the problem of a nonsymmetric control gain matrix because a matrix decomposition technique
is used in controller design; and (ii) it does not require a priori knowledge of the control direction for MIMO nonlinear
system when the states are unmeasured due to the incorporation of Nussbaum gain in the robust control term. The
proposed design scheme guarantees that all of the signals in the resulting closed-loop systems are bounded and that
the tracking errors converge asymptotically to zero.
The remainder of the paper is organized as follows. In Section 2, we describe the plant dynamics and control
objective, and we provide a brief description of fuzzy systems. In Section 3, we explain the proposed adaptive fuzzy
output control scheme. In Section 4, simulation results are presented that demonstrate the effectiveness of the method.
Finally, our conclusions are given in Section 5.

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Throughout this paper, . indicates the Euclidean norm.


2. Problem formulation and fuzzy systems
Consider a class of MIMO nonlinear dynamic system represented by the following form [4,10]
xr1 1 = xr1 2
..
.
xr1 (r1 1) = xr1 r1
xr1 r1 = f1 (x) +

p


g1j (x)uj ,

j =1

..
.
xrp 1 = xrp 2
..
.
xrp (rp 1) = xrp rp
xrp rp = fp (x) +
y1 = x r 1 1
..
.
yp = x r p 1

p


gpj (x)uj ,

j =1

(1)

where x = [xr1 1 , , xr1 r1 , , xrp 1 , , xrp rp ]T R L U is the system state vector, which is assumed to be unp
available for measurement, L = i=1 ri , u = [u1 , , up ]T R p and y = [y1 , , yp ]T R p are the system input
vector and output vector, respectively; and fi (x), i = 1, 2, , p and gij (x), i, j = 1, 2, , p are continuous unknown smooth nonlinear functions.
Let us denote

g11 (x) g1p (x)



T

.. ,
..
F (x) = f1 (x), , fp (x) ,
G(x) = ...
.
.

gp1 (x)
A = diag[A1 , , Ap ],
where

0
1
0
0
0
1

Ai =

0
0
0
0
0
0

B = diag[B1 , , Bp ],

0
0

1
0 r r
i

gpp (x)

C T = diag[C1 , , Cp ],


0
0

Bi = .
..
1

Ci = [1, , 0]1ri .

ri 1

Then, Eq. (1) can be written in the following compact form




x = Ax + B F (x) + G(x)u
y = CT x
F (x) R p

(2)
G(x) R pp .

where
and
The controllability-like of (2) requires that G(x) is nonsingular for all x R L , thus we make the following assumptions:
Assumption 1. The matrix G(x) is nonsingular.

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Assumption 2. Let x R p be a nonzero vector and let G1 (x) = (G(x) + GT (x))/2, x T G1 (x)x = 0.
The objective of this study is to design a controller u(t) such that the system output y follows the reference signal
ym = [ym1 , ym2 , , ymp ]T , i.e., the tracking error Ei (t) = ymi (t) yi (t) = 0 (i = 1, , p), while all the signals in
the derived closed-loop system remain bounded.
Assumption 3. The reference signal ymi (t) (i = 1, , p) and its derivatives up to order ri are smooth and bounded.
Remark 1. Some significant results related to the adaptive fuzzy control of MIMO nonlinear systems were reported
in [9,10,12]. However, these results were obtained under the restrictive condition that the control gain matrix was
positive infinite. In [13,14], the control direction was assumed to be unknown, but the case where the direction was
indefinite was not considered. According to the assumptions given above, the control gain matrix of systems (2) may
be nonsymmetric and indefinite. As stated in the Introduction, the Nussbaum-type function has been used effectively
for solving the problem of unknown control direction, thus a Nussbaum-type function is incorporated in the controller
to estimate the sign of the control gain matrix.
A function N( ) is called a Nussbaum-type function if it has the following useful properties [26,29]:
1
lim sup
s+
s

s
N ( )d = +

(3)

N ( )d =

(4)

lim inf

s+

1
s

s
0

Commonly used Nussbaum functions are [26,29]: 2 cos( ), 2 sin( ), and exp( 2 ) cos( 2 ).
Lemma 1. Let V (t) and (t) be smooth functions defined on [0, tf ), with V (t) 0, t [0, tf ), and N ( ) =
exp( 2 ) cos( 2 ) is an even smooth Nussbaum-type function. If the following inequality holds:
t
V (t) c0 +




g( )N ( ) + 1 ( )d,

t [0, tf )

(5)

where c0 represents some suitable constant and g(t) is a time-varying


parameter that takes values in the unknown
t
/ I , then V (t), (t) and 0 (g( )N ( ( )) + 1) ( )d must be bounded on [0, tf ).
closed intervals I =: [l , l + ] with 0
The proof of Lemma 1 is given in Appendix A.
In a special case, the control gain matrix is symmetric in many physical systems, such as robotic systems [9,12,45].
In a general case, however, the control gain matrix is not always symmetric, thus we consider the general case where
the control gain matrix is nonsymmetric. To satisfy the control objective, we need the following useful lemmas.
Lemma 2. (See [46].) Any real matrix can be expressed as the sum of a symmetric matrix and a skew symmetric
matrix.
Lemma 3. (See [46].) Let A be an n n matrix and (A) is the spectral radius of A, then (A) A.
Lemma 4. (See [46].) If A is an n n positive definite matrix and B is an n n skew symmetric matrix, then A + B
is nonsingular.
Lemma 5. (See [46].) Let A be an n n symmetric matrix, let x R n be a nonzero vector, and let = (x T Ax)/(x T x).
Then, there is at least one eigenvalue of A in the interval (, ] and at least one in [, +).

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Using Lemma 2, the matrix G(x) can be decomposed as follows


G(x) = G1 (x) + G2 (x)

(6)

where G1 (x) = (G(x) + GT (x))/2 and G2 (x) = (G(x) GT (x))/2. Using (6), (2) can be rewritten as


 
x = Ax + B F (x) + G1 (x) + G2 (x) u
y = CT x

(7)

Denote
 (r )
(r ) T
(r)
ym
= ym11 , , ympp

(r 1) T
(r 1)
Ym = ym1 , , ym11 , , ymp , , ympp
,

e = Ym x

If the state variables of (1) are available for measurement, and the nonlinear functions fi (x) and gij (x) are known,
then the following control law:

1

(r)
F (x) + ym
u = uc = G1 (x) + G2 (x)
+ KcT e
(8)
where Kc is the feedback gain matrix used to guarantee that the characteristic polynomial of A BKcT is strictly
Hurwitz, can be used to satisfy the control objective. Indeed, by substituting (8) into (7) and using the fact that
(r)
AYm + Bym = Ym [4], we obtain


(9)
e = A BKcT e
Since A BKcT is a strictly Hurwitz matrix, we have limt e = 0 and it follows that Ei (t) = ymi (t) yi (t) 0
(i = 1, , p) as t . In addition, for the given positive definite matrix Q1 , there exists a positive matrix P1 for
the following matrix equation:

T


A BKcT P1 + P1 A BKcT = Q1
(10)
Based on this analysis, we know that the control law (8) is implemented easily in cases where the state variables
of (1) are available for measurement, while F (x) and G(x) are known. However, this control law (8) cannot be
implemented in the present study since the state variables are unavailable for measurement and the nonlinear functions
fi (x) and gij (x) are unknown. Thus, we design an observer to estimate the state variables for these cases and use fuzzy
systems to approximate the unknown nonlinear functions.
The basic configuration of a fuzzy system comprises a fuzzifier, some fuzzy IFTHEN rules, a fuzzy inference
engine, and a defuzzifier. The fuzzy inference engine employs fuzzy IFTHEN rules to perform a mapping from an
input vector x = [x1 , , xn ]T R n to an output variable y R.
The lth fuzzy rule is written in the following form
R (l) : IF x1 is F1l and and xn is Fnl , THEN y is Gl
where Fil and Gl are fuzzy sets, l = 1, 2, , m, and m is the total number of rules.
Using the strategy of singleton fuzzification, product inference, and center-average defuzzification, the output of
the fuzzy system is given as follows
m l n
l=1 y ( i=1 Fil (xi ))
(11)
y(x) = m n
l=1 i=1 F l (xi )
i

yl

where is the point where the membership function of Gl achieves its maximum value and F l (xi ) is the degree of
i
membership of xi . By introducing the concept of the fuzzy basis function vector (x), Eq. (11) can be rewritten as
y(x) = T (x) := f(x| )
where = [y 1 , , y m ]T , (x) = [ 1 (x), , m (x)]T with l (x) =

(12)
n

i=1 F l (xi )
m n i
.
l=1
i=1 F l (xi )
i

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Lemma 6. (See [1,2].) Let f (x) be a continuous function defined on a compact set x . Then, for any constant 0 > 0,
there exists a fuzzy system (12) such that


sup f (x) f(x| ) 0
xx

According to Lemma 6, the fuzzy system (12) can approximate any continuous nonlinear function f (x) on a
compact set x to an arbitrary degree of accuracy. Define the optimal approximation parameter as follows:



= argmin sup f (x) f(x| )

xx

where is the compact set of allowable controller parameters. Define the parameter error as = and o (x) =
f (x) f(x| ) as the minimum approximation error.
Assumption 4. There exists an unknown positive constant such that |0 (x)| .
If we define a compact set z (z = x ) for the fuzzy systems f(z| ), we have the following lemma:
Lemma 7. Let l 1 (s) = 1/(s + )m , is a positive constant, and s is the Laplace variable, then
1.


l 1 (s) f (x) f(z| ) = (z)T +
where







= T (z) l 1 (s) T (z) + l 1 (s) T (x) T (z) + l 1 (s) 0 (x)
2. The positive constants and exist such that
+
| | 
The proof of Lemma 7 is given in Appendix B.
3. Design of the fuzzy observer and adaptive fuzzy controller
From (6), we know that G1 (x) is symmetric and G2 (x) is skew symmetric, and they can be denoted as

g111 (x) g112 (x) g11p (x)


g112 (x) g122 (x) g12p (x)

G1 (x) =

..
..
..
..

.
.
.
.

g11p (x)

g12p (x)

g212 (x)
0
..
.

g212 (x)

G2 (x) =
..

.
g21p (x) g22p (x)

g1pp (x)

..
.

g21p (x)
g22p (x)

..

.
0

Since the nonlinear functions fi (x), g1ij (x), and g2ij (x) are unknown, by using Lemma 6, we assume that fi (x),
g1ij (x), and g2ij (x) can be approximated by the following fuzzy systems
For i = 1, , p
fi (x|f i ) = fTi f i (x)

(13)

T
g1ii (x)
g 1ii (x|g1ii ) = g1ii

(14)

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

For i = 1, , p 1, j = i + 1, , p
T
g 1ij (x|g1ij ) = g1ij
g1ij (x)

(15)

T
g2ij (x)
g 2ij (x|g2ij ) = g2ij

(16)

where f i (x), g1ii (x), g1ij (x), and g2ij (x) are fuzzy basis function vectors, and f i , g1ii , g1ij , and g2ij are the
parameter vectors of each fuzzy system that we describe later.
Since the state variables of (1) are unavailable for measurement, the fuzzy systems (13)(16) cannot be calculated.
Define x x U as an estimate of the state variables x and denote

T
f 1 ), , fp (x|
fp )
F (x|
f ) = f1 (x|

g111 ) g 112 (x|


g112 ) g 11p (x|
g11p )
g 111 (x|
g 112 (x|
g112 ) g 122 (x|
g122 ) g 12p (x|
g12p )

1 (x|
g1 ) =
G

..
..
..
.
.

.
.
.
.
g 11p (x|
g11p ) g 12p (x|
g12p ) g 1pp (x|
g1pp )

g212 )
g 21p (x|
g21p )
0
g 212 (x|
g 212 (x|
g112 )
0
g 22p (x|
g22p )

2 (x|
g2 ) =
G

..
..
..
.
.

.
.
.
.
g 21p (x|
g21p ) g 22p (x|
g22p )

Now, let us consider a control law u = uc , where uc is designed as follows



1

(r)
1 (x|
2 (x|
F (x|
f ) + ym
uc = G
g1 ) + G
g2 )
+ Kc e ,

(17)

where e = Ym x.

1 (x|
2 (x|
Since the matrix G
g1 ) + G
g2 ) is generated online by the estimates of the parameters g1 and g2 , the
1 (x|
2 (x|
control law (17) is not well-defined when the matrix G
g1 ) + G
g2 ) is singular. We can use the symmetric

matrix decomposition technique to decompose the matrix G1(x|


g1 ), as follows [46]
1 (x|
g1 ) = P D P 1
G

(18)

1 (x|
where P is an orthogonal matrix, D = diag[1 , , p ], where i is the characteristic root of matrix G
g1 ). To
avoid the singular problem of the controller, we modify the control law (17) as follows
1


(r)
1 (x|
2 (x|
F (x|
f ) + ym
(19)
uc = G
g1 ) + G
g2 ) + 0 Ip
+ Kc e + us
1 (x|
where 0 = + G
g1 ), and where is a small given positive constant, Ip is the p p identity matrix, and us
is a compensator for the term 0 Ip .
1 (x|
1 (x|
Since G
g1 ) + 0 Ip = P (D + 0 Ip )P 1 , from Lemma 3, we obtain i = 1, , p, |i | (G
g1 ))
1 (x|
x|
1 (x|
G
g1 ), thus the matrix G(
g ) + 0 Ip is positive definite. Using Lemma 4, we known that G
g1 ) +

0 Ip + G2 (x|
g2 ) is nonsingular, therefore the controller (19) is always well defined.
Remark 2. Eq. (18) is introduced mainly for analytical purposes because the values of P and i are not required when
implementing the controller (19).
Design the fuzzy adaptive observer as follows:



 

1 (x|
2 (x|
x = Ax + B F (x|
f ) + G
g1 ) + G
g2 ) uc + K0 y C T x
y = C T x

(20)
1 , , K ri ]T
[K0i
0i

where K0 = diag(K01 , , K0p ), K0i =


is the observer gain vector used to guarantee that the characteristic polynomial of Ai K0i CiT is strictly Hurwitz. Obviously, the characteristic polynomial of A K0 C T is
also strictly Hurwitz.

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

From (19), we have




(r)
1 (x|
2 (x|
F (x|
f ) + G
g1 ) + G
g2 ) uc = ym
+ Kc e + us 0 uc

(21)

Let e0 = x x be an observation error and substituting (21) into (20) yields


 (r)

x = Ax + B ym
+ KcT e + us 0 uc + K0 C T e0
y = C T x

(22)

Since e = Ym x,
using (22) yields


e = A BKcT e + 0 Buc Bus K0 C T e0

(23)

(r)
This considers fact that AYm + Bym = Ym [4] has been used.
We know that approximation errors are unavoidable, thus to compensate for these errors, we append the controller
(19) with a robust control term ur

u = uc + ur

(24)

Substituting (24) into (7) yields






x = Ax + B F (x) + G1 (x) + G2 (x) (uc + ur )
y = CT x

(25)

Since e0 = x x,
using (20) and (25), we obtain





1 (x|
e0 = A K0 C T e0 + B F (x) F (x|
f ) + G1 (x) G
g1 ) uc



2 (x|
+ G2 (x) G
g2 ) uc + G(x)ur

y = C T e0

(26)

where 
y = y y.
Since only the output 
y in (26) is assumed to be measurable, we use the SPR Lyapunov design
approach to analyze the stability of (26). Eq. (26) can be rewritten as



1 (x|

y = H (s) F (x) F (x|
f ) + G1 (x) G
g1 ) uc



2 (x|
(27)
+ G2 (x) G
g2 ) uc + G(x)ur
where H (s) = C T (sI (A K0 C T ))1 B is a known transfer function matrix of (26). To use the SPR-Lyapunov
design approach, (27) can be written as



1 (x|

y = H (s)L(s)L1 (s) F (x) F (x|
f ) + G1 (x) G
g1 ) uc



2 (x|
(28)
+ G2 (x) G
g2 ) uc + G(x)ur
where L(s) = diag(l(s), , l(s)). Let
1
l(s) = s m + Cm
s m1 + + m = (s + )m


m < min ri
i=1,,p

(29)

such that H (s)L(s) is a proper SPR transfer function matrix. To employ the SPR conditions and motivated by the
assumption in [17], we make the following assumption.
i (x|
Assumption 5. The filter L1 (s) is selected such that L1 (s)[G(x)ur ] G(x)ur and L1 (s)[(Gi (x) G
gi ))uc ]
1

L (s)[(Gi (x) Gi (x|


gi ))]uc (i = 1, 2).
Based on Assumption 5, the state space realization of (28) can be written as






1 (x|
e0 = A K0 C T e0 + Bc L1 (s) F (x) F (x|
f ) + Bc L1 (s) G1 (x) G
g1 ) uc


2 (x|
+ Bc L1 (s) G2 (x) G
g2 ) uc + Bc G(x)ur

y = C e0
T

(30)
(31)

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

where
Bc = diag(bc1 , , bcp ),



T
1
bci
= 0, , 0, 1, Cm
, , m 1r (i = 1, , p)
i

Because H (s)L(s) is a proper SPR, then the positive definite matrices P2 and Q2 must exist such that [4,39]

T


A K0 C T P2 + P2 A K0 C T = Q2
(32)
P2 Bc = C

(33)

, , and
Let us define the optimal approximation parameters fi , g1ii
g1ij
g2ij as follows:




f i = argmin sup fi (x) fi (x|f i )
f i f i xU





g1ii
= argmin sup g1ii (x) g 1ii (x|g1ii )
g1ii g1ii xU

g1ij
= argmin



sup g1ij (x) g 1ij (x|g1ij )

g1ij g1ij xU

g2ij
= argmin



sup g2ij (x) g 2ij (x|g2ij )

g2ij g2ij xU

where f i , g1ii , g1ij , and g2ij are the compact set of allowable controller parameters. We define the parameter

errors f i = f i f i , g1ii = g1ii


g1ii , g1ij = g1ij g1ij , g2ij = g2ij g2ij , and


f i (x) = fi (x) fi x|f i

g1ii (x) = g1ii (x) g 1ii x|g1ii

g1ij (x) = g1ij (x) g 1ij x|g1ij

g2ij (x) = g2ij (x) g 2ij x|g2ij
are the minimum approximation errors.
Given the definition above and Lemma 7, (30) can be written as follows
T

f1
f1 (x)
g111 gT11p (x)
g11p
g111 (x)


..
..
..
..
e0 = A K0 C T e0 + Bc
+ Bc
uc
.
.
.
.
T
T
T
fp (x)
fp
g11p (x)
g11p g1pp (x)
g1pp

T
0
g21p (x)
g21p

.
..
..
..
+ Bc
uc + Bc G(x)ur + Bc
.
.
gT21p (x)
g21p
0
where


g111
f 1
.. ..
= . + .

fp

g11p

..
.

g11p
.. u +
. c
g1pp

0
..
.
g21p

..
.

(34)

g21p
.. u
. c
0

is the lumped disturbance term generated by the state estimation error and the approximation errors.







= T (x)
l 1 (s) T (x)
+ l 1 (s) T (x) T (x)
+ l 1 (s) (x)
Using Lemma 7, we know that the unknown constants f i , g1ii , g1ij , g2ij , f i , g1ii , g1ij , and g2ij exist
such that for i = 1, , p,
|f i | f i f i  + f i ,

|g1ii | g1ii g1ii  + g1ii

10

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

and for i = 1, , p 1, j = i + 1, , p,
|g1ij | g1ij g1ij  + g1ij ,

|g2ij | g2ij g2ij  + g2ij

To cancel the lumped disturbance term, a robust control term was designed in [6,9,12] based on the restriction that
the control gain matrix was positive definite. In the present study, there is no a priori knowledge about the control
direction. As stated in the Introduction, however, the Nussbaum-type function is an effective method for solving the
problem of unknown control direction, thus a Nussbaum-type function N ( ) = exp( 2 ) cos( 2 ) is incorporated in
the robust controller. The robust controller ur is designed as
2
ur = N( )
y

y + 2
2
= 
y 2

y + 2

(35)
(36)

where is a time-varying parameter that is designed later and


1
= uc 2 + eT P1 K0  + f Mf + g uc Mg + f + g uc 
(37)
4
p
p
p1 p
In (37), f , g , f , and g are the estimates of
i=1 f i ,
i=1 g1ii + 2
i=1
j =i+1 (g1ij + g2ij ),
p
p1 p
p

i=1 f i , and
i=1 g1ii + 2
i=1
j =i+1 (g1ij + g2ij ), respectively. Mf = maxi=1,,p f i , Mg =

max{maxi=1,,p g1ii , maxi=1,,p1; j =i+1,,p; l=1,2 glij }.


Denote eT P1 B = [M1 , , Mp ], to compensate for the term 0 Ip in (19) and us is designed as follows
us = 02

[|M1 |sgn(M1 ), , |Mp |sgn(Mp )]T



y + 2

(38)

where sgn(.) is a signum function.


We use the following adaptation laws to adjust the unknown parameters f i , g1ii , g1ij , g2ij , f , g , f , g ,
and :
For i = 1, 2, , p
yi
f i = f i f i (x)

ci 
yi
g1ii = g1ii g1ii (x)u

(39)
(40)

For i = 1, 2, , p 1, j = i + 1, , p
yj + ucj 
yi )
g1ij = g1ij g1ij (uci 

yj + ucj 
yi )
g2ij = g2ij g2ij (uci 
f = f M 
y
f

y
g = g uc Mg 
y
f = f 
g = g uc 
y


1
2
= 0 1 + uc 
4

(41)
(42)
(43)
(44)
(45)
(46)
(47)

where f i > 0, g1ii > 0, g1ij > 0, g2ij > 0, f > 0, g > 0, 0 > 0, and (0) = 0.
t
Remark 3. Since (0) = 0, from (47) we obtain = (0) exp(0 0 (1 + 14 uc 2 )d ), which implies that = 0.
Therefore, the robust control term (35) and compensator (38) are well defined.
The property of the designed control scheme is summarized by the following theorem.

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

11

Theorem 1. Given that the plant defined by (1) satisfies Assumptions 15, then the control law (24) and the adaptation
laws (39)(47) will ensure that all of the signals in the closed-loop system are bounded and that the tracking errors
converge asymptotically to zero.
Proof. Now, let us consider the following Lyapunov function candidate:
V = V1 + V2

(48)

where


p 
1
1
1 1 T
1 T
V1 = eT P1 e + e0T P2 e0 +
f i f i +
g1ii g1ii
2
2
2
f i
g1ii
+

1
2

p1


p


i=1

i=1 j =i+1

1
g1ij

T
g1ij
g1ij +

1
g2ij

T
g2ij
g2ij

(49)

1 2
1 2
1 2
1 2
1 2
+
+
+
+

(50)
2f f 2g g 2f f 2g g 20
p
p
p1 p
p
where f = f i=1 f i , g = g [ i=1 g1ii + 2 i=1 j =i+1 (g1ij + g2ij )], f = f i=1 f i ,
p
p1 p
g = g [ i=1 g1ii + 2 i=1 j =i+1 (g1ij + g2ij )].
The time derivative of V1 is

p 
 1
  1 T
1 T
1 T
f i f i +
g1ii g1ii
V1 = e P1 e + eT P1 e + e0T P2 e0 + e0T P2 e0
2
2
f i
g1ii
V2 =

p1


p


i=1 j =i+1

i=1

1
g1ij

T
g1ij g1ij +

1
g2ij

T
g2ij g2ij

(51)

Substituting (23) and (34) into (51) yields


T


1 
V1 = eT A BKcT P1 + P1 A BKcT e + 0 eT P1 Buc eT P1 Bus
2
T


1 
eT P1 K0 C T e0 + e0T A K0 C T P2 + P2 A K0 C T e0
2
T

f1 (x)
f1
g111 (x)
g111 gT11p (x)
g11p

..
..
..
T
..
+ e0T P2 Bc
+ e0 P2 Bc
uc
.
.
.
.
T
T
fTp (x)
fp
g11p (x)
g11p g1pp (x)
g1pp

T
0
g21p (x)
g21p

.
..
T
.
..
..
+ e0 P2 Bc
uc
.
T
g21p (x)
g21p
0


p

1 T
1 T
T
T

+ e0 P2 Bc G(x)ur + e0 P2 Bc
g1ii
f i +

f i f i
g1ii g1ii

p1


p


i=1 j =i+1

i=1

1
g1ij

T
g1ij g1ij +

1
g2ij

T
g2ij g2ij


(52)

From (33), we know that e0T P2 Bc = e0T C = 


y T . By substituting the adaptive laws (39)(42) into (52) and after
some manipulations, we obtain
1
1
V1 = eT Q1 e e0T Q2 e0 + S
2
2

(53)

12

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

where
S = 0 eT P1 Buc eT P1 Bus eT P1 K0
y +
y T G(x)ur + 
yT

(54)

y
If 
y = 0, using Assumption 2 we know that 
y T G1 (x)
y = 0. Denote 1 (t) = 
that two constants and + exist such that




min G1 (x) 1 (t) max G1 (x) +

T G (x)
y
1

yT 
y

and from Lemma 5 we know

(55)

where min (G1 (x)) and max (G1 (x)) are the minimum and maximum eigenvalues of matrix G1 (x), respectively. If
y = 
y 2 for an arbitrary non-zero constant  [ , + ]. Therefore, we can

y = 0, it is obvious that 
y T G1 (x)
conclude that for all 
y,

y T G1 (x)
y = (t)
y 2

(t) if 
y =0
where (t) = 1
.
if 
y =0

Using (6), (35), (36), and (56), and given that 
y T G2 (x)
y = 0 (G2 (x) is a skew symmetric matrix), we have




S 0 eT P1 B uc  eT P1 Bus + eT P1 K0 
y
+

p





f i f i  + f i |
yi | +
g1ii g1ii  + g1ii |uci ||
yi |
p

i=1

(56)

i=1

p1


p




g1ij g1ij  + g1ij + g2ij g2ij  + g2ij |ucj ||
yi | + |uci ||
yj |

i=1 j =i+1


y 2



2
+ 1 + (t)N( )

y + 2

(57)

Using (38) yields






eT P1 B2
0 eT P1 B uc  eT P1 Bus = 0 eT P1 B uc  02

y + 2

1

y  + 2 uc 2
4

(58)

It is obvious that |
yi | 
y  and |uci | uc , thus by substituting (58) into (57) and after some manipulations, we
obtain
p

 T


1
2
2


S 
f i
y  + 
y Mf
y  + uc  + e P1 K0 
4
i=1

 p
p1
p

 
+ uc 
y Mg
g1ii + 2
(g1ij + g2ij )
i=1

+ 
y

p

i=1

f i + uc 
y

i=1 j =i+1

 p

i=1

g1ii + 2

p1


p



(g1ij + g2ij )

i=1 j =i+1



2
+ 1 + (t)N( )
(59)
2

y +
p
p
p1 p
p
We know that i=1 f i = f f , i=1 g1ii + 2 i=1 j =i+1 (g1ij + g2ij ) = g g , i=1 f i =
p
p1 p
f f , i=1 g1ii + 2 i=1 j =i+1 (g1ij + g2ij ) = g g , then inequality (59) can be written as

y 2

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

13



1
S 2 uc 2 f Mf + g uc Mg + f + g uc  
y
4


2
+ 
y  
y 2
+ 1 + (t)N( )

y + 2

(60)

Since
2
2

y + 2
By substituting (61) into (60), we obtain

y  
y 2

(61)

S f Mf + g uc Mg + f + g uc  


y +




1
2
1 + uc  + 1 + (t)N( )
4

(62)

Substituting (62) into (53) yields




1
1
y
V1 eT Q1 e e0T Q2 e0 f Mf + g uc Mg + f + g uc  
2 
2



1
+ 2 1 + uc 2 + 1 + (t)N( )
4
The time derivative of V2 is
1
1
1
1
1
f f + g g +
f f + g g +
V2 =
f
g
f
g
0

(63)

(64)

Using the adaptive laws (43)(47) and the application of some manipulations yields




1
y  2 1 + uc 2
V2 = f Mf + g uc Mg + f + g uc  
4

(65)

Combining (63) and (65) gives




1
1
V eT Q1 e e0T Q2 e0 + 1 + (t)N( )
2
2


1
1
2 min (Q2 )e0 2 + 1 + (t)N( )
min (Q1 )e
2
2
Integrating (66) over [0, t], we have
1
V (t) V (0) min (Q1 )
2


2
1
e(
) d min (Q2 )
2

t
V (0) +



e0 ( )2 d +

(66)
t


1 + ( )N( ) d


1 + ( )N( ) d

(67)

t
According to Lemma 1, V (t), 0 (1 + ( )N( )) d and (t) are bounded in [0, tf ). Similar to the discussion in
[26], we know that the above conclusion is also true for tf = +. Therefore, e,
e0 , f i , g1ii , g1ij , g2ij , f , g ,
f , g , and are all bounded. Therefore, we can conclude the boundedness of u. From (67), we have
1
min (Q1 )
2

t
0


2
1
e(
) d + min (Q2 )
2

t
0



e0 ( )2 d V (0) V (t) +


1 + ( )N( ) d

(68)

This implies that e L2 and e0 L2 . Since F (x) and G(x) are continuous and all the variables on the right side
of (23) and (34) are bounded, then we have e0 L and e L . Since e L2 L and e L , e0 L2 L and
e0 L , by using Babalats lemma [45] we can conclude that limt e = 0 and limt e0 = 0.
Since e0 = x x,
e = Ym x,
and e = Ym x, then we can derive that limt e = 0, thus the tracking error
Ei (t) = ymi (t) yi (t)(i = 1, , p) converges asymptotically to zero.

14

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

4. Simulation
In this section, two examples are used to verify the performance of the proposed observer-based fuzzy adaptive
control scheme. The first is an academic MIMO system and the second is concerned with a two-link rigid robot
manipulator that moves in a horizontal plane.
Example 1. Consider a MIMO nonlinear system with a nonsymmetric control gain matrix, unknown control direction,
and unmeasured states. The dynamic equations of this MIMO system are given by

x11 = x12





2
2

x12 = x21 0.3 sin(x11 x12 ) + x12 + 3 + cos(x11 ) + cos(x12 ) u1 + 1 + x22 u2


x21 = x22





2 + x 2 cos(x ) + cos(x ) u + 1 + x 2 + x

x22 = x11 + x12

12
22
1
22 u2

22
21

y1 = x11 , y2 = x21

Let x = [x11 , x12 , x21 , x22 ]T , u = [u1 , u2 ]T , y = [y1 , y2 ]T


!
"
2
x21 0.3 sin(x11 x12 ) + x12
F (x) =
2 + x2
x11 + x12
22
!
"
2
3 + cos(x11 ) + cos(x12 )
1 + x22
G(x) =
2 +x
(cos(x12 ) + cos(x22 )) 1 + x21
22
A = diag[A1 , A2 ],
We have

B = diag[B1 , B2 ],

3 + cos(x11 ) + cos(x12 )
G1 (x) =
G1

C T = diag[C1 , C2 ].

"
G1
,
2 +x
1 + x21
22

0
G2 (x) =
G2

G2
0

"

where
2 (cos(x ) + cos(x ))
2 + (cos(x ) + cos(x ))
1 + x22
1 + x22
12
22
12
22
G1 =
,
G2 =
2! "
2
!
"
0 1
0
Ai =
, Bi =
, Ci = [1, 0] (i = 1, 2)
0 0
1

Then, the dynamic system can be expressed as follows




 
x = Ax + B F (x) + G1 (x) + G2 (x) u
y = CT x
which is in the form given by (7).
The control objective is to force the system outputs y1 and y2 to track the desired sinusoidal trajectories ym1 (t) =

sin(t)
and ym2 (t) = 30
sin(t), respectively.
30
In this simulation, the nonlinear functions F (x) and G(x) are assumed to be completely unknown, while two fuzzy
systems in the form of (12) are used to approximate the elements of F (x), three are used to approximate the elements
of G1 (x), and one is used to approximate the element of G2 (x). The fuzzy systems have x11 , x12 , x21 and x22 as
inputs. The overall state vector observation is x = [x1 , x2 , x3 , x4 ]T = [x11 , x12 , x21 , x22 ]T . The fuzzy membership
functions are defined for each variable xi (i = 1, , 4) as follows [9]:


 
1 xi + 1.25 2
F 1 (xi ) = exp
i
2
0.6


 
1 xi 2
F 2 (xi ) = exp
i
2 0.6


 
1 xi 1.25 2
F 3 (xi ) = exp
i
2
0.6

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

15

The lth fuzzy rule is constructed as


R (l) : IF x1 is F1k1 and and x4 is F4k4 , THEN y is Gl
where Fiki and Gl are fuzzy sets, ki = 1, 2, 3, i = 1, , 4, l = 1, 2, , m, and m = 34 = 81 is the total number of
rules where the IF parts comprise all of the possible combinations of the Fiki s for i = 1, 2, , 4.
Denote j = 33 (k1 1) + 32 (k2 1) + 3(k3 1) + k4 , then for ki = 1, 2, 3, i = 1, , 4, j {1, 2, , 81}. Collect
j in the ordering for j = 1, 2, , 81, and let
k1 (x1 ) k2 (x2 ) k3 (x3 ) k4 (x4 )
F
F2
F4
F3
j (x)
= 81 1
j =1 k1 (x 1 ) k2 (x 2 ) k3 (x 3 ) k4 (x 4 )
F1

F2

F3

F4

Then, (x)
= [ 1 (x),
, 81 (x)]
T . The fuzzy basic functions f i (x),
g1ii (x),
and g1ij (x)
are selected as (x),
and
we have
f i ) = fTi (x),

fi (x|

T
g 1ii (x|
g1ii ) = g1ii
(x)
(i = 1, 2),

T
T
g 112 (x|
g112 ) = g112
(x),

g 212 (x|
g212 ) = g212
(x)


T
f 1 ), f2 (x|
f 2 )
F (x|
f ) = f1 (x|
!
"
g 111 (x|
g111 ) g 112 (x|
g112 )

g1 ) =
G1 (x|
g 112 (x|
g112 ) g 122 (x|
g122 )
!
"
0
g 212 (x|
g212 )
2 (x|
G
g1 ) =
g212 )
0
g 212 (x|

In this simulation, the initial conditions are x(0)

= [0.1, 0.1, 0.1, 0.1]T , x(0) = [0.02, 0.02, 0.02, 0.02]T , and the

initial values of the estimated parameters f i (0), g1ii (0) (i = 1, 2), g112 (0), and g212 (0) are all set to zero, while
f (0) = 0.01, g (0) = 0.01, f (0) = 0.01, g (0) = 0.01, (0) = 1, and (0) = 0.015.
The feedback and observer gain matrices are selected as

10 0
400
0
7
800
0
0

.
K0 =
Kc =
0 10 ,
0
400
0
7
0
800
For the given Q1 = diag[5, 5, 5, 5], a positive-definite matrix P1 can be solved by (10). Let l(s) = s + 2, then

1 0
2 0

Bc =
0 1
0 2
Let

5
2
P2 =
0
0

2 0
1
0
0
5
0 2

0
0
,
2
1

800
5
Q2 =
0
0

5 0
0
4
0
0

0 800 5
0 5
4

It easy to see that the matrices P2 and Q2 are asymmetric positive definite and that they satisfy (32) and (33).
The design parameters are selected as f i = 10, g1ii = 100, for i = 1, 2, g112 = g212 = 100, f = 0.001,
g = 0.001, f = 0.001, g = 0.001, 0 = 0.001, and = 0.01.
Using the control law (24) to control the nonlinear systems specified above, the simulation results for the trajectories
of the system outputs y1 , y2 and the reference signals ym1 (t), ym2 (t) are shown in Fig. 1 and Fig. 2. These results
illustrate the good tracking performance of the closed-loop system. The trajectories of the system states x11 , x12 , x21 ,
x22 and the observer states x11 , x12 , x21 , x22 are shown in Figs. 36, respectively. It can also be seen that the designed

16

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Fig. 1. The reference signal ym1 (t) (-) and the output y1 (t) (- -).

Fig. 2. The reference signal ym2 (t) (-) and the output y2 (t) (- -).

Fig. 3. The system state x11 (-) and the observer state x11 (- -).

Fig. 4. The system state x12 (-) and the observer state x12 (- -).

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

17

Fig. 5. The system state x21 (-) and the observer state x21 (- -).

Fig. 6. The system state x22 (-) and the observer state x22 (- -).

Fig. 7. The control input u1 (t) (00.5 s).

observer for the unmeasured states is valid. The control signal u1 (t) is shown in Figs. 78. The control signal u2 (t) is
shown in Figs. 910.
In the previous example, the control gain matrix is not symmetric and the control direction is unknown, thus the
control schemes of [10] cannot be applied to the previous example. To compare our control scheme with the scheme
in [10], we consider a two-link rigid robot manipulator that moves in a horizontal plane.
Example 2. The dynamic equations of this two-link rigid robot manipulator are given by [9,10,45]
" ! "$
! " !
"1 #! " !
q1
q1
M11 M12
u1
hq2 h(q1 + q2 )
=

q2
M21 M22
u2
0
q2
hq1
where
M11 = a1 + 2a3 cos(q2 ) + 2a4 sin(q2 ),

M22 = a2 ,

18

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Fig. 8. The control input u1 (t) (0.520 s).

Fig. 9. The control input u2 (t) (00.5 s).

Fig. 10. The control input u2 (t) (0.520 s).

M12 = M21 = a2 + a3 cos(q2 ) + a4 sin(q2 ),

h = a3 sin(q2 ) a4 cos(q2 )

with
2
2
a1 = I1 + m1 lc1
+ Ie + me lce
+ me l12 ,

a3 = me l1 lce cos(e ),

2
a2 = Ie + me lce
,

a4 = me l1 lce sin(e )

The following parameter values are used in the simulation: m1 = 1, me = 2, l1 = 1, lc1 = 0.5, lce = 0.6, I1 = 0.12,
Ie = 0.25, and e = 300 .
Let y = [y1 , y2 ]T = [q1 , q2 ]T , u = [u1 , u2 ]T , x = [q1 , q1 , q2 , q2 ]T and
!
"
"! "
!
hq2 h(q1 + q2 )
q1
f (x)
,
= M 1
F (x) = 1
0
q2
f2 (x)
hq1
!
"
g (x) g12 (x)
G(x) = 11
= M 1
g21 (x) g22 (x)

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

19

Fig. 11. The reference signal ym1 (t) (-) and the output y1 (t) (- -).

Fig. 12. The reference signal ym2 (t) (-) and the output y2 (t) (- -).

Since the matrix M is symmetric, then the matrix G(x) = M 1 is symmetric. The robot system can be expressed as
follows
y = F (x) + G1 (x)u
which is in the form given by (7), where G1 (x) = G(x), G2 (x) = 0.
The control objective is to force the system outputs q1 and q2 to track the sinusoidal reference signals ym1 (t) =

sin(t)
and ym2 (t) = 30
sin(t), respectively.
30
In this simulation, two fuzzy systems in the form of (12) are used to approximate the elements of F (x) and three
are used to approximate the elements of G1 (x). The fuzzy systems have q1 (t), q 1 (t), q2 (t), and q 2 (t) as inputs,
and the overall state vector observation is x = [x1 , x2 , x3 , x4 ]T = [q1 (t), q 1 (t), q2 (t), q 2 (t)]T . The fuzzy membership
functions and the fuzzy basic functions are the same as those used in the previous example. The controller and adaptive
2 (x|g2 ) = 0 in the equivalent controller
laws are also the same as those used in the previous example, except that G
and there is no adaptive law for g212 .
The initial conditions for the robot, the initial values of the estimated parameters, and the design parameters are all
the same as those used in the previous example, except (0) = 0.07.
The simulation results for the trajectories of the links outputs y1 , y2 and the reference signals ym1 (t), ym2 (t) are
shown in Fig. 11 and Fig. 12, respectively. The trajectories of the system states q1 , q1 , q2 , q2 and the observer states
q1 (t), q 1 (t), q2 (t), and q 2 (t) are shown in Figs. 1316, respectively. The control signal u1 (t) and u2 (t) are shown in
Figs. 1720. These results demonstrate the good tracking performance of the closed-loop system.
g )
Remark 4. A comparison with the simulation results in [10] shows that a total of 34 4 parameters for matrix G(x|
4
need to be updated online, whereas the number of parameters in our scheme is only 3 3. Thus, the lower number
of parameters in our scheme helps to make the proposed algorithm more computationally efficient. Moreover, the

20

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Fig. 13. The system state q1 (-) and the observer state q1 (- -).

Fig. 14. The system state q1 (-) and the observer state q 1 (- -).

Fig. 15. The system state q2 (-) and the observer state q2 (- -).

Fig. 16. The system state q2 (-) and the observer state q 2 (- -).

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Fig. 17. The control input u1 (t) (00.5 s).

Fig. 18. The control input u1 (t) (0.520 s).

Fig. 19. The control input u2 (t) (00.5 s).

Fig. 20. The control input u2 (t) (0.520 s).

21

22

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

schemes of [10] makes the assumption that the control direction is known a priori, whereas this assumption is not
needed in our scheme.
5. Conclusion
In this study, we developed a fuzzy adaptive output control scheme for a class of MIMO nonlinear systems with
a nonsymmetric control gain matrix, unknown control direction, and unmeasured states. The fuzzy systems are used
to approximate the unknown nonlinear functions and an adaptive fuzzy state observer is designed to estimate the
unmeasured states. The parameter adaptive laws are designed by fuzzy basis functions rather than by its filtering, and a
robust control term is used to compensate for the lumped errors. The proposed design scheme guarantees that all of the
signals in the resulting closed-loop systems are bounded and that the tracking errors converge asymptotically to zero.
In future, we will develop observer-based fuzzy adaptive control for non-affine MIMO systems and observer-based
fuzzy adaptive terminal mode control for a class of MIMO systems.
Acknowledgements
This research was supported by the National Natural Science Foundation of China under Grant 60974054.
Appendix A. Proof of Lemma 1
The procedure used for this proof is similar to that in [30]. For simplicity, we rewrite (5) as
t
V (t) c0 +



g( )N ( ) ( )d +

( )d,

t [0, tf )

(A.1)

Since g(t) [l , l + ], let us define gmax = max{|l |, |l + |} and gmin = min{|l |, |l + |} for convenience. First, we show
that (t) is bounded on [0, tf ) by seeking a contradiction. Suppose that (t) is unbounded and two cases are considered: (t) has no upper bound and (t) has no lower bound.
For clarity, we define
M2


Ng (M1 , M2 ) = g( )N ( ) d ( )

(A.2)

M1

Noting that g(t) gmax , we have




Ng (M1 , M2 ) gmax (M2 M1 )

sup



N ( ) = gmax (M2 M1 )eM22

[M1 ,M2 ]

(A.3)

for the Nussbaum function N ( ) = e cos( 2 ), which is positive for (4m 1, 4m + 1) and negative for
(4m + 1, 4m + 3) with m an integer.
Let us consider the case g(t) > 0. First, we consider the interval [0, 4m 1] and by applying (A.3), we have


Ng (0, 4m 1) gmax (4m 1)e(4m1)2
(A.4)
Next, let us observe the variation in the interval [4m 1, 4m + 1]. Noting that N ( ) 0, [4m 1, 4m + 1], we
have the following inequality
4m+1/2



g( )N ( ) d ( )

Ng (4m 1, 4m + 1)

(A.5)

4m1/2

By noting that g(t) gmin , we have


Ng (4m 1, 4m + 1) gmin

inf

[4m1/2,4m+1/2]

N ( ) =

2
2/2gmin e(4m1/2)

(A.6)

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

Comparing (A.4) and (A.6) yields


&
2 %
Ng (0, 4m + 1) e(4m1)
2/2gmin e4m3/4 gmax (4m 1) ,

23

(A.7)

which can be written as


2
&
Ng (0, 4m + 1) e(4m1) %
2/2gmin e4m3/4 gmax (4m 1)

4m + 1
4m + 1
From (A.8), we have

(A.8)

Ng (0, 4m + 1)
= +
m+
4m + 1
Let us observe the interval [0, 4m + 1]. Similarly, by applying (A.3), we can obtain


Ng (0, 4m + 1) gmax (4m + 1)e(4m+1)2
lim

(A.9)

(A.10)

Then, let us consider the next immediate interval [4m + 1, 4m + 3]. Noting that N ( ) 0, [4m + 1, 4m + 3], in
the same manner as for [4m 1, 4m + 1], we have the following inequality:


3 2
2
g( )N ( ) d ( )
gmin e(4m+ 2 )
2

4m+5/2

Ng (4m + 1, 4m + 3)

(A.11)

4m+3/2

Accordingly, the inequalities (A.10) and (A.11) lead to


$
#
2
(4m+1)2
(4m+5/4)
Ng (0, 4m + 3) e
gmax (4m + 1) ,
gmin e
2
which can be written as
2

Ng (0, 4m + 3)
e(4m+1)

4m + 3
4m + 3

2
gmin e(4m+5/4) gmax (4m + 1)
2

(A.12)

$
(A.13)

From (A.13), we have


Ng (0, 4m + 3)
=
4m + 3
Therefore, from (A.9) and (A.14), we can conclude that, g(t) > 0
lim

m+

1
lim sup
s+
s

(A.14)



g( )N ( ) d ( ) = +

(A.15)



g( )N ( ) d ( ) =

(A.16)

1
lim inf
s+
s

s
0

In the following, we aim to show that (A.15) and (A.16) also hold for g(t) < 0. Let us observe the intervals
[0, 4m 1], [4m 1, 4m + 1], [0, 4m + 1], and [4m + 1, 4m + 3] for g(x) < 0. In the intervals [0, 4m 1] and
[0, 4m + 1], the inequalities (A.4) and (A.10) remain. In the interval [4m 1, 4m + 1], noting that g(x) < 0 and
N( ) 0, we can also obtain
4m+ 12

Ng (4m 1, 4m + 1)



2
g( )N ( ) d ( ) 2/2gmin e(4m1/2)

4m 12

Combining (A.4) and (A.17) yields

(A.17)

24

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

2
&
Ng (0, 4m + 1)
e(4m1) %
2/2gmin e4m3/4 gmax (4m 1)

4m + 1
4m + 1
From (A.18), we have

lim

m+

Ng (0, 4m + 1)
=
4m + 1

(A.18)

(A.19)

In the interval [4m + 1, 4m + 3], noting that g(x) < 0 and N ( ) 0, we have
4m+5/2



2
g( )N ( ) d ( ) = 2/2gmin e(4m+3/2)

Ng (4m + 1, 4m + 3)

(A.20)

4m+3/2

Combining the inequalities (A.10) and (A.20) on the intervals [0, 4m + 1] and [4m + 1, 4m + 3], respectively, we have
$
2 #
Ng (0, 4m + 3) e(4m+1)
2
(A.21)

gmin e4m+5/4 gmax (4m + 1)


4m + 3
4m + 3 2
From (A.21), we have
Ng (0, 4m + 3)
= +
(A.22)
m+
4m + 3
From (A.19) and (A.22), we can also obtain (A.15) and (A.16). Therefore, we can conclude that (A.15) and (A.16)
hold regardless of whether g(t) > 0 or g(t) < 0.
Next, we show that (t) is bounded on [0, tf ) by seeking a contradiction. Suppose that (t) has no upper bound on
[0, tf ). In this case, limttf (t) = +.
Dividing (A.1) by (t) > 0 yields
lim

V (t)
c0
(0)
0

+1

(t)
(t)
(t)

 (0)
0

g( )N( ( ))d ( )
1
+
(t)
(t)

(t)


g( )N ( ) d ( )

(A.23)

After taking the limit as (t) +, (A.23) becomes


V (t)
1
0 lim
1 + lim
(t)+ (t)
(t)+ (t)

(t)



g x( ) N ( ) d ( ),

(A.24)

which appears to be a contradiction based on (A.16). Therefore, (t) is upper bounded on [0, tf ).
Suppose that (t) has no lower bound on [0, tf ). In this case, limttf ( (t)) = +.
Noting that N(.) is an even function, dividing (A.1) by (ti ) > 0 yields
 (0)
g( )N ( ( ))d( ( ))
V (t)
c0
(0)
0

1+
+ 0
(t) (t)
(t)
(t)
1

(t)

(t)




g( )N ( ) d ( )

(A.25)

After taking the limit as (t) +, (A.25) becomes


V (t)
1
0
lim
1
lim
(t)+ (t)
(t)+ (t)

(t)




g( )N ( ) d ( )

(A.26)

which appears to be a contradiction based on (A.15). Therefore, (t) is lower bounded on [0, tf ).
t
Therefore, (t) must be bounded on [0, tf ). In addition, V (t) and 0 (g( )N ( ) + 1) d are bounded on [0, tf ).

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

25

Appendix B. Proof of Lemma 7


1. Some simple manipulations yield








f (x) f(z| ) = f (x) f x| + f x| f z| + f z| f(z| )


= (z)T + 0 (x) + T (x) T (z)
The optimal approximation parameter is a constant vector, thus we have








l 1 (s) f (x) f(z| ) = (z)T (z)T + l 1 (s) (z)T + l 1 (s) 0 (x) + l 1 (s) T (x) T (z)





= (z)T (z)T + l 1 (s) (z)T




+ l 1 (s) 0 (x) + l 1 (s) T (x) T (z)
= (z)T +
(x) x and |0 (x)| , where z = supt (z), x = supt (x),
2. From the fact that | T (z) | z ,
we can obtain [40,45]
 1  T
 1  T  x
 1 


l (s) (z)  z ,
l (s) (x) 
l (s) 0 (x) 

,
m
m

m
Then, the positive constants and exist such that
+
| | 
where = z (1 +

1
m ),

= ( mx + z )  +

m .

References
[1] L.-X. Wang, Fuzzy systems are universal approximators, in: 1992 IEEE International Conference on Fuzzy Systems, San Diego, CA, USA,
1992, pp. 11631170.
[2] L.-X. Wang, Stable adaptive fuzzy control of nonlinear systems, IEEE Trans. Fuzzy Syst. 1 (2) (1993) 146149.
[3] S. Tong, J. Tang, T. Wang, Fuzzy adaptive control of multivariable nonlinear systems, Fuzzy Sets Syst. 111 (2) (2000) 153167.
[4] H.-X. Li, S. Tong, A hybrid adaptive fuzzy control for a class of nonlinear MIMO systems, IEEE Trans. Fuzzy Syst. 11 (1) (2003) 2434.
[5] S. Tong, H.-X. Li, Fuzzy adaptive sliding-mode control for MIMO nonlinear systems, IEEE Trans. Fuzzy Syst. 11 (3) (2003) 354360.
[6] S. Tong, T. Chai, Direct adaptive control and robust analysis for unknown multivariable nonlinear systems with fuzzy logic systems, Fuzzy
Sets Syst. 106 (3) (1999) 309319.
[7] R. Ordonez, K.M. Passino, Stable multi-input multi-output adaptive fuzzy/neural control, IEEE Trans. Fuzzy Syst. 7 (3) (1999) 345353.
[8] S. Labiod, T.M. Guerra, Direct adaptive fuzzy control for a class of MIMO nonlinear systems, Int. J. Syst. Sci. 38 (8) (2007) 665675.
[9] S. Labiod, M.S. Boucherit, T.M. Guerra, Adaptive fuzzy control of a class of MIMO nonlinear systems, Fuzzy Sets Syst. 151 (1) (2005)
5977.
[10] S. Tong, B. Chen, Y. Wang, Fuzzy adaptive output feedback control for MIMO nonlinear systems, Fuzzy Sets Syst. 156 (2) (2005) 285299.
[11] H. Xu, P.A. Ioannou, Robust adaptive control for a class of MIMO nonlinear systems with guaranteed error bounds, IEEE Trans. Autom.
Control 48 (5) (2003) 728742.
[12] V. Nekoukar, A. Erfanian, Adaptive fuzzy terminal sliding mode control for a class of MIMO uncertain nonlinear systems, Fuzzy Sets Syst.
179 (1) (2011) 3449.
[13] A. Boulkroune, M. Tadjine, M. MSaad, M. Farza, Fuzzy adaptive controller for MIMO nonlinear systems with known and unknown control
direction, Fuzzy Sets Syst. 161 (6) (2010) 797820.
[14] A. Boulkroune, M. MSaad, H. Chekireb, Design of a fuzzy adaptive controller for MIMO nonlinear time-delay systems with unknown
actuator nonlinearities and unknown control direction, Inf. Sci. 180 (24) (2010) 50415059.
[15] W. Shi, M. Zhang, W. Guo, L. Guo, Stable adaptive fuzzy control for MIMO nonlinear systems, Comput. Math. Appl. 62 (7) (2011)
28432853.
[16] A. Boulkroune, M. MSaad, A fuzzy adaptive variable-structure control scheme for uncertain chaotic MIMO systems with sector nonlinearities
and dead-zones, Expert Syst. Appl. 38 (12) (2011) 1474414750.
[17] Y.-J. Liu, S.-C. Tong, T.-S. Li, Observer-based adaptive fuzzy tracking control for a class of uncertain nonlinear MIMO systems, Fuzzy Sets
Syst. 164 (1) (2011) 2544.
[18] S. Tong, Y. Li, P. Shi, Observer-based adaptive fuzzy backstepping output feedback control of uncertain MIMO pure-feedback nonlinear
systems, IEEE Trans. Fuzzy Syst. 20 (4) (2012) 771785.
[19] I.-H. Li, L.-W. Lee, A hierarchical structure of observer-based adaptive fuzzy-neural controller for MIMO systems, Fuzzy Sets Syst. 185 (1)
(2011) 5282.

26

W. Shi / Fuzzy Sets and Systems 263 (2015) 126

[20] S.-C. Tong, Y.-M. Li, G. Feng, T.-S. Li, Observer-based adaptive fuzzy backstepping dynamic surface control for a class of MIMO nonlinear
systems, IEEE Trans. Syst. Man Cybern., Part B, Cybern. 41 (4) (2011) 11241135.
[21] S. Tong, Y. Li, Adaptive fuzzy output feedback control of MIMO nonlinear systems with unknown dead-zone inputs, IEEE Trans. Fuzzy Syst.
21 (1) (2013) 134146.
[22] S. Tong, C. Li, Y. Li, Fuzzy adaptive observer backstepping control for MIMO nonlinear systems, Fuzzy Sets Syst. 160 (19) (2009) 27552775.
[23] S. Tong, T. Wang, Y. Li, B. Chen, A combined backstepping and stochastic small-gain approach to robust adaptive fuzzy output feedback
control, IEEE Trans. Fuzzy Syst. 21 (2) (2013) 314327.
[24] R. Lozano, B. Brogliato, Adaptive control of a simple nonlinear system without a priori information on the plant parameters, IEEE Trans.
Autom. Control 37 (1) (1992) 3037.
[25] X. Ye, Adaptive nonlinear output-feedback control with unknown high-frequency gain sign, IEEE Trans. Autom. Control 46 (1) (2001)
112115.
[26] R. Nussbaum, Some remarks on a conjecture in parameter adaptive control, Syst. Control Lett. 3 (5) (1983) 243246.
[27] T. Zhang, S. Ge, Adaptive neural control of MIMO nonlinear state time-varying delay systems with unknown dead-zones and gain signs,
Automatica 43 (6) (2007) 10211033.
[28] T.-P. Zhang, Y. Yi, Adaptive fuzzy control for a class of MIMO nonlinear systems with unknown dead-zones, Acta Autom. Sin. 33 (1) (2007)
96100.
[29] X. Ye, J. Jiang, Adaptive nonlinear design without a priori knowledge of control directions, IEEE Trans. Autom. Control 43 (11) (1998)
16171621.
[30] S. Ge, F. Hong, T.H. Lee, Adaptive neural control of nonlinear time-delay systems with unknown virtual control coefficients, IEEE Trans.
Syst. Man Cybern., Part B, Cybern. 34 (1) (2004) 499516.
[31] S.S. Ge, J. Wang, Robust adaptive neural control for a class of perturbed strict feedback nonlinear systems, IEEE Trans. Neural Netw. 13 (6)
(2002) 14091419.
[32] A. Boulkroune, M. Tadjine, M. MSaad, M. Farza, How to design a fuzzy adaptive controller based on observers for uncertain affine nonlinear
systems, Fuzzy Sets Syst. 159 (8) (2008) 926948.
[33] W.-Y. Wang, Y.-G. Leu, T.-T. Lee, Output-feedback control of nonlinear systems using direct adaptive fuzzy-neural controller, Fuzzy Sets
Syst. 140 (2) (2003) 341358.
[34] Y.-G. Leu, W.-Y. Wang, T.-T. Lee, Observer-based direct adaptive fuzzy-neural control for nonaffine nonlinear systems, IEEE Trans. Neural
Netw. 16 (4) (2005) 853861.
[35] T.-C. Lin, C.-H. Wang, H.-L. Liu, Observer-based indirect adaptive fuzzy-neural tracking control for nonlinear SISO systems using VSS and
H approaches, Fuzzy Sets Syst. 143 (2) (2004) 211232.
[36] S. Tong, T. Wang, J.T. Tang, Fuzzy adaptive output tracking control of nonlinear systems, Fuzzy Sets Syst. 111 (2) (2000) 169182.
[37] S. Tong, H.-X. Li, W. Wang, Observer-based adaptive fuzzy control for SISO nonlinear systems, Fuzzy Sets Syst. 148 (3) (2004) 355376.
[38] S.-C. Tong, X.-L. He, H.-G. Zhang, A combined backstepping and small-gain approach to robust adaptive fuzzy output feedback control,
IEEE Trans. Fuzzy Syst. 17 (5) (2009) 10591069.
[39] Y.-G. Leu, T.-T. Lee, W.-Y. Wang, Observer-based adaptive fuzzy-neural control for unknown nonlinear dynamical systems, IEEE Trans. Syst.
Man Cybern., Part B, Cybern. 29 (5) (1999) 583591.
[40] J.-H. Park, G.-T. Park, S.-H. Kim, C.-J. Moon, Output-feedback control of uncertain nonlinear systems using a self-structuring adaptive fuzzy
observer, Fuzzy Sets Syst. 151 (1) (2005) 2142.
[41] J.-H. Park, G.-T. Park, Adaptive fuzzy observer with minimal dynamic order for uncertain nonlinear systems, IEE Proc., Control Theory Appl.
150 (2) (2003) 189197.
[42] S. Tong, C. Liu, Y. Li, Fuzzy-adaptive decentralized output-feedback control for large-scale nonlinear systems with dynamical uncertainties,
IEEE Trans. Fuzzy Syst. 18 (5) (2010) 845861.
[43] S. Tong, Y. Li, Y. Li, Y. Liu, Observer-based adaptive fuzzy backstepping control for a class of stochastic nonlinear strict-feedback systems,
IEEE Trans. Syst. Man Cybern., Part B, Cybern. 41 (6) (2011) 16931704.
[44] A. Boulkroune, M. Tadjine, M. MSaad, M. Farza, How to design a fuzzy adaptive controller based on observers for uncertain affine nonlinear
systems, Fuzzy Sets Syst. 159 (8) (2008) 926948.
[45] J. Slotine, W. Li, Applied Nonlinear Control, Prentice-Hall, Englewood Cliffs, NJ, 1991.
[46] R. Horn, C.R. Johnson, Matrix Analysis, Cambridge University Press, 1990.

You might also like