This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 1

Adaptive Fuzzy Observer-Based Active

Fault-Tolerant Dynamic Surface Control for a
Class of Nonlinear Systems with Actuator Faults
Qikun Shen, Bin Jiang*, Senior Member, IEEE, and Vincent Cocquempot

presence of a fault [2, 7-12]. It is valuable to point out that most

Abstract—The problem of fault-tolerant dynamic surface results concerning actuator faults reported in the literature only
control (DSC) for a class of uncertain nonlinear systems with considered bias faults. Gain faults did not attract enough
actuator faults is discussed, and an active fault-tolerant control attention. Investigating both actuator time-varying bias and gain
(FTC) scheme is proposed. Using the DSC technique, a novel fault
diagnostic algorithm is proposed, which removes the classical
faults motivates this paper.
assumption that the time derivative of the output error should be Fuzzy logic system (FLS), as a universal function
known. Further, an accommodation scheme is proposed to approximator, has been widely adopted for nonlinear systems
compensate for both actuator time-varying gain and bias faults, identification and control [13]. Using FLS, various adaptive
and avoids the controller singularity. In addition, the proposed control approaches such as backstepping [4-6, 14-18], have
controller guarantees that all signals of the closed-loop system are been developed for controlling uncertain nonlinear systems. It is
semi-globally uniformly ultimately bounded (SGUUB), and
converge to a small neighborhood of the origin. Finally, the worth noting that the increasing of the system dimension
effectiveness of the proposed FTC approach is demonstrated on a produces a complexity explosion in traditional backstepping
simulated aircraft longitudinal dynamics example. design methods. In order to overcome this problem an original
DSC scheme was proposed in [19-23], where the complexity
Index Terms—Fault-tolerant control, DSC, adaptive control was reduced by introducing the first-order filter in each step of
the backstepping design. In order to deal with the complete
unknown control gains, i.e., both the values and signs of control
I. INTRODUCTION gains are unknown, the Nussbaum-type functions were
introduced [27, 29]. However, if actuator faults occur, then the
I N practical engineering systems, actuator, sensor or
component faults frequently occur, which can cause system
performance deterioration and lead to instability that can further
control schemes in [19-23] do not guarantee the closed-loop
system stability or correct tracking performances. One
produce catastrophic accidents. Thus, many effective FTC motivation of our work is thus to provide an active fault-tolerant
approaches have been proposed to improve system reliability control scheme which guarantees the closed-loop system
and to guarantee system stability in all situations [1-12, 30-38]. stability and maintains satisfactory control performances in all
Basically, the FTC strategies can be categorized into two classes: situations. Another motivation is also to provide a control
the passive approach and the active one. Passive FTC is robust scheme that is applicable in practical applications where both
control technique with respect to an a priori fixed set of faults [1, the values and signs of control gain are not known. In addition,
3-6]. Active FTC consists in online reconfiguring or investigating both actuator time-varying bias and gain faults
reconstructing the controller to recover the stability and system motivates this paper.
performance as soon as a diagnostic algorithm has detected the In this paper, we investigate the problem of tracking control
for a class of nonlinear uncertain systems with complete
unknown control gains and propose an active FTC against
Manuscript received ×, 2012. This work was supported in part by the actuator faults. Compared with existing works, the following
National Natural Science Foundation of China under Grant 61273171, and
Grant 61174046, Doctoral Fund of Ministry of Education of China main contributions are worth to be emphasized. 1) The
(20113218110011) and a Project Funded by the Priority Academic Program proposed FTC scheme considers both gain and bias faults
Development (PAPD) of Jiangsu Higher Education Institutions. simultaneously and does not need the conditions that the bounds
Q. Shen is with the College of Automation Engineering, Nanjing University
of Aeronautics and Astronautics, Nanjing, 210016, China, and with the College
of the varying faults and their time derivatives are known
of Information Engineg, Yangzhou University, Yangzhou 225127, China constants, which thus enlarges the practical application range of
(e-mail: [email protected]). the method. 2) The proposed adaptive fault accommodation
*Corresponding author, B. Jiang is with the College of Automation
algorithm does not need the classical assumption that the time
Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing,
210016, China.(,e-mail,: binjiang@ nuaa.edu.cn) derivative of the output errors must be known. 3) A decision
V. Cocquempot is with the LAGIS UMR CNRS 8219, Lille 1 University: threshold for FDI is defined and applied on an online
Sciences and Technologies, 59655, Villeneuve d'Ascq, France (e-mail: computable fault indicator and not on an asymptotic value of a
vincent.cocquempot @univ-lille1.fr).

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 2

criterion which is not available in practice. The decision objective in healthy and faulty conditions. Under healthy
algorithm is thus more practical than in other works as [9, 12]. 4) condition, control input u is designed, such that the system
In general, the denominator of the fault tolerant control input output y can track asymptotically the reference signal yd .
contains the estimate of the gain fault. If the denominator is Meanwhile, the FDI algorithm is working. As soon as an
equal to zero, a controller singularity occurs. In the proposed actuator fault is detected and isolated, the fault accommodation
FTC scheme, the controller singularity is avoided. 5) In contrast algorithm is activated and a proper fault-tolerant control
with [1-6] where the sign of the control gain must be known, the input u is used such that the tracking performance is still
proposed active FTC scheme does not require the a priori maintained stable.
knowledge of this sign.
The rest of this paper is organized as follows. In Section 2, Assumption 1[3-6]: g j ( x ) ∈ span{g0 ( x )}, g 0 ( x ) ∈ R n , for
the problem formulation, Nussbaum-type function and j = 1,⋯, m , and the nominal system xɺ = f 0 ( x ) + F ( x )θ +
mathematical description of FLS are introduced. Actuator faults
are integrated in such problem and the FTC objective is
g 0 ( x )u0 , y = h ( x ) with u0 ∈ R , is transformable into the
formulated. In Section 3, the main technical results of this paper parametric-strict-feedback (PSF) form with relative degree ρ ,
are given, which include fault detection, isolation, estimation where F ( x ) = [ f1 , f 2 ,⋯, fl ]T , θ = [θ1 ,θ 2 , ⋯,θl ]T .
and fault-tolerant control scheme. An aircraft control
As presented in [3], based on Assumption 1, there exists a
application is presented in Section 4. These simulation results
demonstrate the effectiveness of the proposed technique. diffeomorphism Tr : [ xT ,η T ]T = Tr ( x ), x ∈ R ρ and η ∈ Rγ ,
Finally, Section 5 draws the conclusion. ρ + γ = n such that system (1) can be transformed into the
following PSF form.
II. PROBLEM STATEMENT AND DESCRIPTION OF FLS
 xɺi = xi +1 + ϕ1 ( x[i ] )θ , i = 1, 2,⋯, ρ − 1
A. Problem Statement

 xɺ ρ = ϕ0 ( x,η ) + ϕ ρ ( x,η )θ + β ( x,η ) µ u
T T

Consider the following nonlinear system:  (4)

ηɺ = ψ 1 ( x,η ) + ψ 2 ( x,η )θ
 xɺ = f ( x ) + ∑ l θ f ( x ) + ∑ m µ g ( x )u 
i =1 i i j =1 j j
 y = x1
0 j
 (1)
 y = h( x )
where x[i ] = [ x1 , x2 ,⋯, xi ]T , x = [ x1 , x2 ,⋯, xρ ]T and u = [u1 ,
where x ∈ R is the state, y ∈ R is the output, and u j ∈ R , j =
n
⋯, um ]T denote the measurable state vector and input, µ =
1, 2,⋯, m are the plant control signals, fi (⋅) ∈ R , i = 0,1, ⋯, l ,
n
diag ( µ1 ,⋯, µm ) , µ j , j = 1,⋯, m denote unknown constants,
g j (⋅) ∈ R , j = 1, ⋯, m and h (⋅) ∈ R are smooth functions,
n
β T = [ β1 ,⋯ , β m ]T , βi = Lgi ( x ) Lρf −(1x ) h( x ) , i = 1,⋯ , m , where
θi ∈ R , i = 1,⋯, l and µ j ≠ 0 , j = 1,⋯, m are unknown
0

L f p is the Lie derivative of a scalar function p( x ) along the

constants.
Control objective is to design adaptive controllers for system (1) vector field f ( x) = [ f1 ( x),⋯ , f n ( x)]T , defined as L f p =
to guarantee boundedness of the closed-loop signals and ∂p
∑ i =1 ( ∂x ) fi , ϕi = [ L f1 ( x ) Lif−01( x ) h ( x ),⋯, L fi ( x ) Lif−01( x ) h( x )]T ,
n
asymptotic tracking of a given reference output
1
signal yd ∈ R by y ,
i = 1,⋯, ρ − 1 , ϕ ρ = [ L f1 ( x ) Lρf0−(1x ) h ( x ),⋯ , L fi ( x ) Lρf0−(1x ) h ( x )]T ,
yd ∈ ∏ {( yd , yɺ d , ɺɺyd ) : yd2 + yɺ d2 + ɺɺyd2 ≤ B0 } , B0 > 0 ∈ R
∂Tz ∂T
denotes a known constant. ϕ0 = Lρf0 ( x ) h( x ) , ψ 1 = f 0 ( x ),ψ 2 = z F ( x ) .
Actuator fault model considered can be described as follows: ∂x ∂x
In order to solve the actuator failure compensation problem, the
u jf = ρ j ( x )u j + f ju ( x ) , t ≥ t j , j = 1, ⋯, m (2)
following assumptions are needed:
where unknown functions ρ j ( x ) ∈ [0,1] and f ju ( x ) denote the Assumption 2 [5, 6]: The zero dynamics ηɺ = ψ 1 ( x,η ) +
remaining control rate and a bounded signal, respectively, t j is ψ 2 ( x,η )θ is input-to-state stable with respect to x as the
unknown fault occurrence time. Denote u = [u1 ⋯, um ]T , u f = input, η is measurable and β j ( x,η ) ≠ 0, j = 1, ⋯, m .
Assumption 3: The signs of µ j , j = 1,⋯, m are unknown.
[u1f ⋯, umf ]T , Ρu ( x ) = diag ( ρ1 ( x ),⋯, ρ m ( x )) , F u ( x) =
[ f1u ( x ),⋯, f mu ( x )]T , then, on has Let fi ( x[i ] ) = ϕi ( x[i ] )θ , f ρ ( x,η ) = ϕ0 ( x,η ) + ϕ Tρ ( x,η )θ

u f = Ρu ( x )u + F u ( x ) (3) g T ( x,η ) = β T ( x,η ) µ = [ g1 ( x,η ), ⋯, g m ( x,η )]T , then (4) can

Now, the control objective is re-defined as follows: An active be rewritten as
FTC approach is proposed to obtain the above tracking

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 3

 xɺi = xi +1 + fi ( x[i ] ), i = 1,⋯, ρ − 1 t

V (t ) ≤ c0 + e −c1t ∫ g (τ ) N (ς )ςɺ ec1τ dτ + e− c1t ∫ ςɺ ec1τ dτ
t

 0 0
 xɺ ρ = f ρ ( x,η ) + g ( x,η )u
T
 where constant c1 > 0, g (⋅) is a time-varying parameter which
ηɺ = ψ 1 ( x,η ) + ψ 2 ( x,η )θ
y = x takes values in the unknown closed intervals I := [l −1 , l +1 ] with
 1
0 ∉ I , and c0 represents some suitable constant, then V (t ) ,
Further, one has
t
 xɺ = Ax + Hy + f + Bg T ( x,η )u ς (t ) and ∫ g (τ ) N (ς )ςɺdτ must be bounded on [0, t f ) .
0

ηɺ = ψ 1 ( x,η ) + ψ 2 ( x,η )θ (5)
C. Mathematical Description of FLS
y = x
 1 A FLS consists of four parts: the knowledge base, the
 −h1   h1  0 fuzzifier, the fuzzy inference engine working on fuzzy rules, and
      the defuzzifier. The knowledge base for FLS comprises a
where A =  ⋮ I ρ −1  , H =  ⋮  , B =  ⋮  , f = [ f1 , ⋯, collection of fuzzy if–then rules of the following form:
 − hρ 0 ⋯ 0  hρ  1 
    Rl : if x1 is A1l ⋯and xn is Anl , then y is B l
f ρ ]T , hi ∈ R, i = 1,⋯, ρ are chosen such that A is a strict where Ail , i = 1, 2,⋯ , n , l = 1, 2,⋯, M are fuzzy sets and B l is
Hurwitz matrix. Considering fault model (3), the faulty system the fuzzy singleton for the output in the lth rule, M is the rules
can be described as number. Through singleton fuzzifier, center average
 xɺi = xi +1 + fi ( x[i ] ), i = 1, 2,⋯, ρ − 1 defuzzification and product inference [13], the FLS output can
 be expressed as
 xɺ ρ = f ρ ( x,η ) + g ( x,η )Ρ ( x)u (t ) + g ( x,η ) F ( x)
T u T u
 (6) M  n  M n 
ηɺ = ψ 1 ( x,η ) + ψ 2 ( x,η )θ y ( x) = ∑ y l  ∏ µ Al ( xi )  / ∑  ∏ µ Al ( xi ) 
y = x l =1  i =1 i  l =1  i =1 i 
 1
where µ l ( xi ) is the membership function of the fuzzy set Alj .
Ai
B. Nussbaum Type Gain
Define the fuzzy basis functions as
Any continuous function N ( s ) : R → R is a function of
 n  M  n 
Nussbaum type if it has the following properties: ξ l ( x) = ∏ µ Al ( xi ) / ∑  ∏ µ Al ( xi ) 
1 s  i =1
i i
 l =1  i =1 
s∫0
1) lim sup N (ς )d ς = +∞ ,
s →+∞ Define θ = [ y , y ,⋯, y ] = [θ1 , θ 2 ,⋯, θ M ] and ξ = [ξ 1 , ξ 2 ,
T 1 2 M

1 s ⋯, ξ M ]T , then the FLS output can be rewritten as

2) lim inf ∫ N (ς ) dς = −∞
s → −∞ s 0
y ( x ) = θ T ξ ( x) (7)
For example, the continuous functions ς 2 cos(ς ), ς 2 sin(ς ) ,
The stability results obtained in FLS control literature are
2
and eς cos((π / 2)ς ) verify the above properties and are thus semi-global in the sense that, as long as the input variables x of
Nussbaum-type functions [26].The even Nussbaum the FLS remain within some pre-fixed compact set Ω , where the
2 compact set can be made as large as desired, there exist
function eς cos((π / 2)ς ) is used throughout this paper. controllers with sufficiently large number of FLS rules such that
Lemma 1[27, 29]: Let V (⋅) and ς (⋅) be smooth functions all the signals in the closed-loop remain bounded.
Lemma 3 [13]: Let f ( x) be a continuous function defined on a
defined on [0, t f ) with V (t ) ≥ 0, ∀t ∈ [0, t f ) , and N (⋅) be an
compact set Ω . Then for any constant ε > 0 , there exists a FLS
even smooth Nussbaum-type function. If the following
inequality holds: (7) such as sup | f ( x) − θ T ξ ( x) |≤ ε .
x∈Ω
t
V (t ) ≤ c0 + ∫ ( gN (ς ) + 1)ςɺ dτ , ∀t ∈ [0, t f ) By Lemma 3, the FLS (7) can approximate any smooth function
0
on a compact set to any degree of accuracy. Similar to [17], by
where g ≠ 0 is a constant, and c0 represents a suitable constant,
the FLS (7), fi ( x) , i = 1,⋯, ρ − 1 , f ρ ( x) , g j ( x,η ) , ρ gj ( x,η )
t
then V (t ), ς (t ) and ∫ gN (ς )ςɺdτ must be bounded on [0, t f ) . = g j ( x,η ) ρ j ( x) and ρ fj ( x,η ) = g j ( x,η ) f ju ( x) , j = 1,⋯,
0
Lemma 2[29]: Let V (⋅) and ς (⋅) be smooth functions defined m are approximated as: fˆi ( xˆ , θˆ fi ) = θˆTfi ξ fi ( xˆ[i ] ) , fˆρ ( xˆ , θˆ f ρ ) =
on [0, t f ) with V (t ) ≥ 0, ∀t ∈ [0, t f ) , and N (⋅) be an even
= θˆTf ρ ξ f ρ ( xˆ ,η ) , gˆ j ( xˆ ,η ) = θˆgjT ξ gj ( xˆ ,η ) , ρˆ gj ( xˆ ,η ,θˆgTρ j ) =
smooth Nussbaum-type function. For ∀t ∈ [0, t f ) , if the
θˆgTρ jξ g ρ j ( xˆ,η ) , ρˆ fj ( xˆ,η , θˆTf ρ j ) = θˆTf ρ j ξ f ρ j ( xˆ,η ) , where xˆ,
following inequality holds,

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 4

θˆ fi ,θˆgj ,θˆ f ρ j , θˆg ρ j are the estimates of x, θ *fi ,θ gj* , θ *f ρ j , θ g*ρ j , xˆɺ = Axˆ + Hy + fˆ + Bgˆ T u, yˆ = Cxˆ (8)
respectively. Let us define the optimal parameter vectors where fˆ = [ fˆ1 , ⋯, fˆρ ], gˆ = [ gˆ1 ,⋯, gˆ m ], εˆg = [εˆg1 ,⋯, εˆgm ] ,
T T T

θ *fi , i = 1,⋯, ρ , θ gj* , θ *f ρ j and θ g*ρ j , j = 1,⋯, m as

fˆi , εˆ fi , i = 1,⋯, ρ and gˆ j , εˆgj , j = 1,⋯, m denote the estimates
θ *fi = arg min [ sup | fi ( x[i ] ) − fˆi ( xˆ[i ] ,θˆTfi ) |]
θ fi ∈Ω f x∈U , xˆ∈Uˆ of f i and g j , C = [1 0 ⋯ 0] . Let xˆ = [ xˆ1 , xˆ2 ,⋯, xˆ ρ ]T and

θ *f ρ = arg min [ sup | f ρ ( x) − fˆ ( xˆ,θˆTf ρ ) |] e = x − xˆ , the error dynamics can be written as:
θ f ρ ∈Ω f x∈U , xˆ∈Uˆ
eɺ = Ae + d + Bd g , ey = Ce (9)
θ gj* = arg min [ sup | g j ( x,η ) − gˆ gj ( xˆ ,η ,θˆgjT ) |] ,
where d = [d1 ,⋯, d ρ ]T , di = δ fi = f i − fˆi , d g = ∑ j =1δ gj u j =
m
θ gj ∈Ω g x∈U , xˆ∈Uˆ

θ *f ρ j = arg min [ sup | ρ f ρ j ( x,η ) − ρˆ f ρ j ( xˆ,η , θˆTf ρ j ) |] ,

∑ j =1 ( g j − gˆ j )u j .
m
θ f ρ j ∈Ω f ρ x∈U , xˆ∈Uˆ

θ g*ρ j = arg min [ sup | ρ gj ( x,η ) − ρˆ gj ( xˆ,η , θˆgTρ j ) |] , In the following, based on the previous section, we will
θ g ρ j ∈Ω g ρ x∈U , xˆ∈Uˆ incorporate the DSC technique proposed in [20] into an
adaptive fuzzy control design scheme for the ρ -order system
where Ω f , Ω g , Ω g ρ , Ω f ρ ,U and Û are compact regions for
described by (8). Similar to the traditional backstepping design
θˆ fi ,θˆgj ,θˆ f ρ j , θˆg ρ j , x and x̂ . The FLS minimum approximation method, the recursive design procedure contains ρ steps. From
errors are defined as Step 1 to Step ρ , virtual control laws αi −1 , i = 2,⋯, ρ are
ε fi = f ( x[i ] ) − θ *fiT ξ fi ( xˆ[i ] ) ,ε fρ = f ρ ( x) − θ *f Tρ ξ f ρ ( xˆ ,η ) designed at each step. Finally overall control laws u j ,

ε gj = g j ( x,η ) − θ gj*T ξ gj ( xˆ,η ) , j = 1,⋯, m are constructed at step ρ .

Step 1: Let S1 = xˆ1 − yd . Then, form (8)
ε g ρ j = ρ gj ( x,η ) − θ g*Tρ j ξ g ρ j ( xˆ,η ) ,
Sɺ1 = xˆɺ1 − yɺ d = xˆ2 + θˆTf 1ξ f 1 + θɶTf 1ξ f 1 + ε f 1 − δ f 1 − yɺ d (10)
ε f ρ j = ρ fj ( x,η ) − θ *f Tρ j ξ f ρ j ( xˆ,η )
Choose a virtual control α1 as follows:
δ fi = f ( x[i ] ) − θˆTfi ξ fi ( xˆ[i ] ) , δ f ρ = f ρ ( x) − θˆTf ρ ξ f ρ ( xˆ,η )
α1 = − k1S1 − [ fˆ1 − yɺ d +
(11)
δ gj = g j ( x,η ) − θˆgjT ξ gj ( xˆ ,η ) , ( M ε f 1 + M δ f 1 ) tanh( S1 ( M ε f 1 + M δ f 1 ) / w)]
δ g ρ j = ρ gj ( x,η ) − θˆgTρ j ξ g ρ j ( xˆ,η ) , Here and in the following, ki > 0 ∈ R, i = 1, ⋯, ρ are design
δfρj = ρ fj ( x,η ) − θˆTf ρ j ξ f ρ j ( xˆ,η ) parameters, w > 0 ∈ R is a constant. Introduce a new state
variable z2 and let α1 pass through a first-order filter with time
In order to simplify the notations in the following,
let ξ fi , ξ f ρ , ξ gj , ξ g ρ j and ξ f ρ j denote ξ fi ( xˆ[i ] ) , ξ f ρ ( xˆ ,η ) , constant ε 2 to obtain z2
ε 2 zɺ2 + z2 = α1 , z2 (0) = α1 (0) (12)
ξ gj ( xˆ,η ) , ξ g ρ j ( xˆ,η ) and ξ f ρ j ( xˆ,η ) , respectively. Now, the
Here and in the following, ε i > 0 ∈ R, i = 1,⋯, ρ − 1 are design
following assumptions are made.
Assumption 4 [13, 15, 17]: There exist known positive real parameters.
constants M ε fi , M ε gj , M ε f ρ j , M ε g ρ j , M fi , M gj , M f ρ j , M g ρ j , Step i ( i = 2,⋯, ρ − 1 ): Consider xˆɺi = xˆi +1 + fˆi + hi e1 . Define
the ith error surface Si to be Si = xˆi − zi , then
M δ fi , M δ gj , M δ f ρ j and M δ g ρ j such that | ε fi |≤ M ε fi , | ε gj |≤
Sɺi = xˆɺi − zɺi = xˆi +1 + θˆTfiξ fi + hi e1 − zɺi + θɶTfi ξ fi + ε fi − δ fi (13)
M ε gj , | ε f ρ j |≤ M ε f ρ j ,| ε g ρ j |≤ M ε g ρ j , || θ *fi ||≤ M fi , || θ gj* ||
Choose a virtual control α i as follows:
≤ M gj , || θ g*ρ j ||≤ M g ρ j , || θ *f ρ j ||≤ M f ρ j , | δ fi |≤ M δ fi , | δ gi |≤
α i = −ki Si − [ fˆi − zɺi +
M δ gj , | δ f ρ j |≤ M δ f ρ j and | δ g ρ j |≤ M δ g ρ j , where i = 1,⋯, ρ (14)
( M ε fi + M δ fi ) tanh( Si ( M ε fi + M δ fi ) / w)]
and j = 1,⋯, m .
Introduce a new state variable zi +1 and let α i pass through a
first-order filter with constant ε i +1 to obtain zi +1
III. ADAPTIVE FAULT-TOLERANT DSC DESIGN AND STABILITY
ANALYSIS ε i +1 zɺi +1 + zi +1 = α i , zi +1 (0) = α i (0) (15)

A. Stability Control in Fault-Free Case and Fault Detection Step ρ : Consider xˆɺ = fˆ + h e + gˆ T u .
ρ ρ ρ 1

Since the system states are not all measured, the following Define the ρ th error surface S ρ to be S ρ = xˆ ρ − z ρ , then
observer is constructed to estimate the system states.

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 5

Sɺ ρ = xˆɺ ρ − zɺ ρ = fˆρ + hρ e1 + gˆ T u − zɺ ρ = θˆTf ρ ξ f ρ + where

(16)  S1 ( M ε f 1 + M δ f 1 ) 
hρ e1 + gˆ T u − zɺ ρ + θɶTf ρ ξ f ρ + ε f ρ − δ f ρ + gˆ T u ∂  2 tanh( )( M ε f 1 + M δ f 1 )   xˆ1 
 w 
Finally, let the final control α ρ j , j = 1,⋯, m be as follows: χ1 =   ⋮ ,
∂ ( xˆ1 ,⋯, xˆρ )  
∆  xˆρ 
α ρ j = u j = [ N (ς )(k ρ S ρ + )] / m, ςɺ = −k ρ S ρ2 − ∆ (17)  
Sρ ɺ ∂ξ
B2 = k1Sɺ1 + θˆTf 1ξ f 1 + θˆTf 1
f1
+ χ − ɺɺ
yd , which is a continuous
ρ η fi *T * η gj *T * ∂xˆ1
∆ = ∑ i =1[| Si | ( M ε fi + M δ fi ) + θ fi θ fi + θ gj θ gj ] +
2η1 2η2 function. Similarly, for i = 2,⋯, ρ − 1 ,
ρ −1
µe + ( ρ − 1)σ1 / 2 + S ρ θˆTf ρ ξ f ρ +∑ i =1 Si +1ki +1e1 yɺ i +1 = − yi +1 / ε i +1 + Bi +1 (25)

Just as pointed out in [29], for the above control (17), controller  xˆ1 
ɺ ∂ξ fi  ⋮  + yɺ i + χ is a
singularity may occur since ∆ / S ρ is not well defined at S ρ = 0 . where Bi +1 = ki Sɺi + θˆTfi ξ fi + θˆTfi
∂ ( xˆ1 , ⋯, xˆi )   ε i
i
 xˆi 
Similar to [29], let define Ω cS ⊂ Ω and Ω 0cS s.t.
ρ ρ continuous function,
ΩcS := {Sρ | S ρ |< cSρ } , Ω0cS := Ω − ΩcS ,where cSρ > 0 is  Si ( M ε fi + M δ fi ) 
ρ ρ ρ
∂  2 tanh( )( M ε fi + M δ fi )   xˆ1 
 w 
a constant that can be chosen arbitrarily small and “-” is used to
χi =   ⋮ 
denote the complement of set B in set A ∂ ( xˆ1 ,⋯, xˆ ρ )  
 xˆ ρ 
as A − B := {x | x ∈ A and x ∉ B} . Thus, the final control  
α ρ j = u j , j = 1,⋯, m can be modified as Differentiating Ve = eT Pe with respect to time t and considering
 ∆ (9) and Assumption 4, leads to
 N (ς )[k ρ S ρ + S ] / m, ςɺ = k ρ S ρ + ∆, S ρ ∈ ΩcS ρ
2 0
Vɺe ≤ eT ( PA + AT P + 2λ PP )e + µe (26)
uj =  ρ (18)
0, S ρ ∈ ΩcS where λ > 0 ∈ R is a design parameter, and it is assumed that
 ρ ρ
| u j |≤ u j , u j > 0 ∈ R , µe = (∑ i =1 M δ fi 2 + ∑ j =1 M δ gj u j 2 ) /
m
In the following, we will give the closed-loop system stability
analysis. The closed-loop system in the new coordinates Si , zi (4λ ) . Notice that, this assumption seems to be strict. However,
can be expressed as follows: in many practical systems, such as flight control systems
considered in this paper, control input is bounded. Hence, this
Sɺ1 = xˆ2 + θˆTf 1ξ f 1 + θɶTf 1ξ f 1 + ε f 1 − δ f 1 − yɺ d
assumption is reasonable in some case. In addition, λ can be
Sɺi = xˆi +1 + θˆTfi ξ fi + hi e1 − zɺi + θɶTfi ξ fi + ε fi − δ fi , chosen to be a larger constant such that µe ≤ µe , µe > 0 ∈ R .

Sɺ ρ = θɶTf ρ ξ f ρ + hρ e1 + gˆ T u − zɺ ρ + θɶTf ρ ξ f ρ + ε f ρ − δ f ρ If for a given constant λ matrices P = PT > 0 , Q > 0 are

ε 2 zɺ2 + z2 = α1 , z2 (0) = α1 (0) chosen appropriately such that PA + AT P + 2λ PP ≤ −Q , then,

ε i +1 zɺi +1 + zi +1 = α i , zi +1 (0) = α i (0) , i = 2,⋯, ρ − 1 Vɺe = −eT Qe + µe = − g eVe + µe (27)
Define where g e = λmin (Q ) / λmax ( P) .
y2 = z2 − α1 Consider the following Lyapunov function
(19)
= k1S1 + fˆ1 + ( M ε f 1 + M δ f 1 ) tanh( S1 / w) + z2 − yɺ d 1 ρ 1 1 ρ −1
V1 = Ve + [∑ i =1 ( Si2 + θɶTfi θɶfi + θɶgjT θɶgj ) + ∑ k =1 yk2+1 ]
2 η1 η2
yi +1 = zi +1 − α i = ki Si + fˆi − zɺi −
where η > 0,η > 0 are design constants, θɶ =θ * − θˆ , θɶ =
Si ( M ε fi + M δ fi ) yi (20) 1 2 fi fi fi gj
( M ε fi + M δ fi ) tanh( ) + zi +1 +
w εi θ gj* − θˆgj . Differentiating V1 with respect to time t, leads to
where i = 2,⋯, ρ − 1 . From (11), (14), (18)-(20), one has ρ −1
Vɺ1 ≤ ∑ i =1 ( Si Si +1 − ki Si2 + Si yi +1 ) + S ρθˆTf ρ ξ f ρ + Vɺe +
Sɺ1 = xˆɺ1 − yɺ d = S2 − k1S1 + y2 + θɶTf 1ξ f 1 + ε f 1 − δ f 1 (21)
ρ −1 yk2+1 ρ −1

Sɺi = xˆɺ i − zɺi = Si +1 − ki Si + hi e1 + yi +1 + θɶTfi ξ fi + ε fi − δ fi (22)

∑ k =1 (− ε k +1
+ | yk +1 Bk +1 |) + ∑ i =1 Si +1ki +1e1 +

ρ ρ
Sɺ ρ = θˆTf ρ ξ f ρ + hρ e1 + gˆ T u − zɺ ρ + θɶTf ρ ξ f ρ + ε f ρ − δ f ρ (23) ∑ i =1 SiθɶTfi ξ fi + ∑ i =1 Si (ε fi − δ fi ) +
Since zɺi = (α i − zi +1 ) / ε i = − yi / ε i , i = 2,⋯, ρ − 1 , it gives S ρ ∑ j =1 ( gˆ j + ε gj − δ gj )u j + S ρ ∑ j =1θɶgjT ξ gj u j −
m m

yɺ 2 = zɺ2 − αɺ1 = − y2 / ε 2 + B2 (24)

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 6

1 ɶT ˆɺ
ρ 1 ɶT ˆɺ Si Si +1 ≤ Si2 + Si2+1 / 4, Si yi +1 ≤ Si2 + yi2+1 / 4
∑ i =1 2η
θ fi θ fi − ∑ j =1
m
θ gjθ gj (28) (32)
2η 2
1 | yk +1 Bk +1 |≤ yk2+1 Bk2+1 / σ 1 + σ 1 / 2, ∀σ 1 > 0 ∈ R
ɺ ɺ
Define θˆ fi , i = 1, ⋯, ρ , θˆgj , j = 1, ⋯, m as follows where σ 1 > 0 ∈ R is a design parameter. Substituting (32) to
ɺ (31), yields
θˆ fi =
ρ −1 1 ( ρ − 1)σ 1
2η1Siξ fi − η fiθˆ fi , if || θˆ fi ||< M fi or || θˆ fi ||= M fi Vɺ1 ≤ ∑ i =1 [ Si2+1 + (2 − ki ) Si2 ] + + S ρθˆTf ρ ξ f ρ +
4 2

 and 2η1Siξ fi − η fiθˆ fi ≥ 0; ρ −1 ρ −1 1 1 B2
 (29) ∑ i =1 Si +1ki +1e1 + Vɺe + ∑ k =1 ( − + k +1 ) yk2+1 +
4 ε k +1 2σ 1
 θˆ fiθˆTfi θˆ fiθˆTfi ˆ
η ξ − η θˆ + η ξ − η θ fi ) ρ
 1 i fi ∑ i =1 Si (ε fi − δ fi ) + S ρ ∑ j =1 ( gˆ j + ε gj − δ gj )u j −
2 S (2 S m
fi fi 1 i fi fi
 || θˆ fi ||2 || θˆ fi ||2
 ρη fi ɶT ɶ m η gj ɶT ɶ
 if || θˆ fi ||= M fi and 2η1Siξ fi − η fiθˆ fi < 0 ∑ i =1 2η θ fi θ fi − ∑ j =1 θ gjθ gj +
1 2η 2
ɺ
θˆgj = ρ η fi m η gj

2η2 S ρ ξ gj u j − η gjθˆgj , if || θˆgj ||< M gj or || θˆgj ||= M gj

∑ i =1 2η θ *fiT θ *fi + ∑ j =1 2η θ gj*T θ gj*
1 2
 (33)
 and 2η 2 S ρ ξ gj u j − η gjθˆgj ≥ 0;
 As pointed out in [20], since for any B0 > 0 and p > 0 the sets
 θˆgjθˆgjT
2η2 S ρ ξ gj u j − η gjθˆgj + (2η2 S ρ ξ gj u j − η gj ⋅ Π 0 := {( yd , yɺ d , ɺɺyd ) : yd2 + yɺ d2 + ɺɺyd2 ≤ B0 }
 || θˆgj ||2 ρ
Π i := {∑ l =1 ( Sl2 (0) + θɶTfl (0)θɶfl (0)) + ∑ j =1θɶgjT (0)θɶgj (0) +
m

 θˆgjθˆgj ˆ
T

 ˆ 2 θ gj ), if | | θ gj ||= M gj and2η 2 S ρ ξ gj u j − η gjθ gj < 0

ˆ ˆ
∑ k =1 yk2+1 (0) + 2eT (0)e(0) ≤ 2vp}, i = 1,⋯, ρ
i
||
 gjθ ||
(30) are compact, Π 0 × Π i is also compact. Thus, | Bi +1 | has a
where η fi > 0,η gj > 0 are design constants, u j is a bounded maximum M i +1 on Π 0 × Π i , v = max{1,η1 ,η2 ,1 / λmin ( P)} .
control input which is applied simultaneously to the jth actuator Choose
in the system (5) and the observer (8). Applying Young’s 1
k1 = 2 + β0 , ki = 2 + β 0
Inequality, one has 4
(34)
η fiθɶTfi θˆ fi / (η1 ) ≤ −η fiθɶTfi θɶfi / (2η1 ) + η fiθ *fiT θ *fi / (2η1 ) 1 1 1 M2
k ρ = 1 + β0 , = + k +1 + β0
4 ε k +1 4 2σ 1
ηε fi εɶfiεˆ fi / η1 ≤ −ηε fi εɶ2fi / (2η1 ) + ηε fi (ε *fi )2 / (2η1 ) ,
where k = 1,⋯, ρ − 1, i = 2,⋯, ρ − 1 β 0 > 0 ∈ R is a constant.
η g ,iθɶgjT θˆgj / η2 ≤ −η gjθɶgjT θɶgj / (2η2 ) + η gjθ gj*T θ gj* / (2η2 ) , Thus, from (27) , (33) and (34), one has
ηε gj εɶgj εˆgj / η2 ≤ −ηε gj εɶgj2 / (2η2 ) + ηε gj (ε *gj )2 / (2η 2 ) ρ ρ −1 ρ η fi ɶT ɶ
Vɺ1 ≤ −∑ i =1 β 0 Si2 − ∑ k =1 β0 yk2+1 − g eVe − ∑ i =1 θ fi θ fi +
Substituting the above inequalities into (28), yields 2η1
ρ −1 η gj ɶT ɶ ( ρ − 1)σ 1
Vɺ ≤1 ∑ i =1
( S S − k S 2 + S y ) + S θˆT ξ +
i i +1 i i i i +1 ρ fρ fρ
∑ j =1 2η
m
θ gjθ gj + + S ρ θˆTf ρ ξ f ρ +
2 2
ρ −1 yk2+1
∑ k =1 (− ε + | yk +1 Bk +1 |) + S ρ θˆTf ρ ξ f ρ + ρ −1 ρ
∑ i =1 Si +1ki +1e1 + ∑ i =1| Si | (M ε fi + M δ fi ) +
k +1
ρ −1 ρ η fi η gj *T *
∑ i =1 Si +1ki +1e1 + Vɺe + ∑ i =1 Si (ε fi − δ fi ) + ρ
∑ i =1 2η θ *fiT θ *fi + ∑ j =1
m
θ gj θ gj +
2η 2
ρ η fi ɶT ɶ (31)
1
S ρ ∑ j =1 ( gˆ j + ε gj − δ gj )u j − ∑ i =1
m
θ fi θ fi − µe + S ρ ∑ j =1 ( gˆ j + ε gj − δ gj )u j
m
2η1
ρ η fi ɶT ɶ m η gj ɶT ɶ ( ρ − 1)σ 1 ρ −1
∑ i =1 2η θ fi θ fi − ∑ j =1 θ gjθ gj + Define ∆ = µe + + S ρ θˆTf ρ ξ f ρ +∑ i =1 Si +1ki +1e1 +
1 2η2 2
ρ η fi m η gj ρ η fi η gj
∑ i =1[| Si | ( M ε fi − M δ fi ) + 2η θ *fiT θ *fi ] + ∑ j =1
m
∑ i =1 2η θ *fiT θ *fi + ∑ j =1 2η θ gj*T θ gj* θ gj*T θ gj* ,
1 2η 2
1 2
From Young’s Inequality, one has then, one has

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 7

ρ ρ −1 ρ η fi ɶT ɶ ii)The vectors Si , θɶfi , εɶfi , i = 1,⋯, ρ , and θɶgj , εɶgj , j = 1,⋯, m

Vɺ1 ≤ −∑ i =1 β 0 Si2 − ∑ k =1 β0 yk2+1 − g eVe − ∑ i =1 θ fi θ fi −
2η1 remain in the compact set Ω1 , specified as
m η gj Ω1 :=
∑ j =1 2η θɶgjT θɶgj + ∆ + S ρ ∑ j =1 ( gˆ j + ε gj − δ gj )u j
m

2  ρ θɶTfi θɶfj + εɶ2fi 

(35) ( Si , yi , θɶfi , εɶfi , θɶgj , εɶgj ) ∑ i =1 ( Si2 + )+ 
Substituting control laws (18) into (35), leads to  η1 
 
ɶ T ɶ
m (θ gjθ gj + εɶgj )
2
Vɺ1 ≤ − gV1 + ∑ j =1 (hDj N (ς ) + 1)ςɺ  
m
ρ −1 2
 ∑ j =1 + ∑ k =1 yk +1 + 2λmin ( P )e e ≤ 2µ1 
(36) T

 η 
ηf1 η f ρ η g1 η gm 2
where g = min{β 0 , g e , , ⋯, , ⋯, } and hDj = where µ1 can be adjusted by appropriately choosing the design
2η1 2η1 2η2 2η2
parameter such as η1 ,η2 ,η fi ,ηε fi , η gj , ηε gj and σ 1 , w, β 0 .
( gˆ j + ε gj − δ gj ) . Applying Lemma 2, we can conclude
t
Proof: From the above analysis, it is easy to obtain the
∑ j =1 (hDj N (ς ) + 1)e− gτ ςɺdτ and ς (t ) are SGUUB
m
that, V1 (t ) , ∫ conclusion. The detailed proof is omitted here.
0
From Theorem 1, all signals of the closed-loop system belong
on [0, t f ) . According to Proposition 2 in [28], if the solution of
to the following set Ω1 . Therefore, the detection residual can be
the closed-loop system is bounded, then t f = +∞ . Let µ1 be the defined as J =| yd (t ) − xˆ1 (t ) |=| S1 | . Obviously, it is seen that
t
∑ j =1 (hDj N (ς ) + 1)e− gτ ςɺdτ , we have the the following inequality holds in the healthy case: J ≤ 2µ1 .
m
upper bound of ∫
0
t Then, the fault detection can be performed using the following
∫0 ∑ j =1 e (hDj N (ς ) + 1)e− gτ ςɺ dτ ≤
m − gt
following inequalities: mechanism:
t  J ≤ Td no fault occurred,

∫ ∑ j =1 (hDj N (ς ) + 1)e
m − gτ
ςɺ dτ ≤ µ1 . Thus, (36) becomes  J > Td fault has occurred
0
where threshold Td = 2µ1 .
Vɺ1 ≤ − gV1 + µ1 (37)
Solving inequality (37) gives B. Fault Isolation and Estimation
µ µ µ
0 ≤ V1 (t ) ≤ 1 + [V1 (0) − 1 ]e− gt ≤ 1 + V1 (0) = µ1 (38) Since the system has m actuators and it is assumed that only
g g g one single fault occurs at one time, we have m possible faulty
which means that V1 (t ) is bounded by µ1 . Thus, all signals of the cases in total. When the sth ( 1 ≤ s ≤ m ) actuator is faulty, the
closed-loop system, i.e., S (t ), θɶ ,θɶ , εɶ , εɶ and y are faulty model can be described as: usf = ρ s ( x )us + f su ( x ) . The
i fi gj fi gj i
faulty system can be described as follows:
uniformly ultimately bounded, i.e. for i = 1,⋯, ρ , j = 1,⋯, m ,
 xɺi = xɺi +1 + fi , i = 1, 2,⋯, ρ − 1
2µ1 ,| yi |≤ 2 µ1 , || θɶfi ||≤ 2η1 µ1 , | εɶfi |≤ 2η1 µ1 , || θɶgj ||  ρ
 xɺ ρ = f ρ + ∑ j =1, j ≠ s g j u j + g s ρ s us + g s f s
u

≤ 2η 2 µ1 ,| εɶgj |≤ 2η 2 µ1 ,|| e ||≤ µ1 / λmin ( P) .  (40)

ηɺ = ψ 1 ( x,η ) + ψ 2 ( x,η )θ
Now, the following theorem guarantees the existence of the y = x
observer (8) and the corresponding tracking performance.  1

Theorem 1: Consider system (5) and observer (8) under where fi = fi ( x[i ] ) , i = 1, 2, ⋯, ρ − 1, f ρ = f ρ ( x,η ), gi =
Assumptions 1-4, the virtual control (11), (14) and (18), the
gi ( x,η ) , us is the sth actuator’s desired control input when the
adaptive laws (29) and (30). If matrices H , Q > 0 , P = PT > 0
sth actuator is healthy, ts is the unknown fault occurrence time.
and constant λ > 0 ∈ R are chosen such that After a fault has been detected, the isolation scheme is
PA + AT P + 2λ PP ≤ −Q (39) activated. Inspired by [24], the following m nonlinear adaptive
ρ fuzzy observers are considered:
for all initial conditions satisfying Π i := {∑ l =1 ( Sl2 (0) +
xˆɺs = Axˆs + Hys + fˆ + B(∑ j =1, j ≠s gˆ j u j + gˆ ρ sus + fˆρ s )
m
(41)
θɶTfl (0)θɶfl (0)) + ∑ j =1θɶgjT (0)θɶgj (0) + ∑
m i
y 2 (0) +
k =1 k +1
T
2e (0)
where xˆs = [ xˆs1 ,⋯, xˆs ρ ]T is the observer state; fˆi = θˆTfiξ fi ,
e(0) ≤ 2vp} , i = 1, ⋯, ρ , ki , ε k are chosen as (34), then we
can guarantee the following properties under bounded initial i = 1, ⋯, ρ and gˆ j = θˆgjT ξ gj , j = 1,⋯, m, j ≠ r ,which are the
conditions: estimates of fi and g j ; gˆ ρ s = θˆgTρ sξ g ρ s and fˆρ s = θˆTf ρ sξ f ρ s
i)All signals in the closed-loop system are globally uniformly
ultimately bounded are the estimates of g s ρ s and g s f su . It is assumed that

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 8

r ( 1 ≤ r ≤ m ) is the practical fault pattern where the faulty ρ ρ −1 ρ

V2 = ∑ i =1 Si2 / 2 + ∑ k =1 yk2+1 / 2 + Vse +∑ i =1θɶTfi θɶfi / (2η1 )+
actuators are the r th actuator. Let es (t ) = x − xˆs is the output
∑ j =1θɶgjT θɶgj / (2η2 ) + θɶTf ρ sθɶf ρ s / (2η3 ) + θɶgTρ sθɶg ρ s / (2η4 )
m
error and state error between the faulty plant and the sth
observer, then the error dynamics can be written as follows:
where η3 > 0,η 4 > 0 are design parameters, θɶf ρ s =θ *f ρ s − θˆ f ρ s ,
eɺs = Aes + d + B (d sg + d g ρ + d f ρ ) (42)
θɶg ρ s =θ g* ρ s − θˆg ρ s . Differentiating V2 with respect to time t,
where d g ρ = g ρ s − gˆ s + g r − gˆ ρ r , d f ρ = f ρ s − fˆs + f r − fˆρ r ,
leads to
gˆ ρ r . In the following, similar to the previous subsection, ρ −1 ρ −1
stability analysis will be conducted using DSC method. The
Vɺ ≤
2 ∑ i =1 ( Si Si +1 − ki Si2 + Si yi +1 ) + ∑ i =1 Si +1ki +1e1 + Vɺse +
case ( s = r ) is first considered. ρ −1 y2
S ρ (θˆTf ρ ξ f ρ + θˆTf ρ ξ f ρ ) + ∑ k =1 (− k +1 + | yk +1 Bk +1 |) +
1) s = r : Similar to the previous subsection, we will propose an ε k +1
adaptive DSC scheme for system (41). The recursive design ρ ρ
procedure contains ρ steps. From Step 1 to Step ρ , virtual ∑ S θɶT ξ
i =1 i fi fi
+ ∑ i =1 Si (ε fi − δ fi ) + S ρ (ε f ρ s − δ f ρ s ) +
control laws αi −1 , i = 2,⋯, ρ are designed at each step. Finally S ρ ∑ j =1, j ≠ s ( gˆ j + ε gj − δ gj + θɶgjT ξ gj )u j + S ρ θɶgTρ sξ g ρ s us −
m

overall control laws u j , j = 1,⋯, m are constructed at step ρ .

ρ 1 ɶT ˆɺ
S ρ ( gˆ s + ε g ρ s − δ g ρ s )us − ∑ i =1 θ fi θ fi −
Let us define dynamic surfaces Si , i = 1,⋯ , ρ , virtual control 2η1
laws α i ,first-order filters ε i +1 zɺi +1 + zi +1 = α i and yi +1 , i = 1,⋯ , 1 ɶT ˆɺ 1 ɶT ˆɺ 1 ɶT ˆɺ
∑ j =1, j ≠ s 2η
m
θ gjθ gj − θ f ρ sθ f ρ s − θ g ρ sθ g ρ s
ρ − 1 as in the previous subsection. Note that, the difference 2 2η3 2η4
between the observer (8) and (41) lies in the last equality, i.e., (46)
ɺˆ ɺˆ
xˆɺ = fˆ + h e + ∑ j =1
m
ρ ρ gˆ u ,
ρ 1 j j Adaptive laws θ fi , i = 1, ⋯, ρ , θ gj , j = 1,⋯, m, j ≠ s are

xˆɺ s ρ = fˆs ρ + hs ρ e1 + ∑ defined

m
gˆ u
j =1, j ≠ s j j
+ gˆ ρ s us + fˆρ s
ɺ
as (29) and (30) in the previous subsection. Here, only θˆ f ρ s
Thus, Sɺi , i = 1,⋯ , ρ − 1 have same expressions as in the
ɺ
previous subsection, only Sɺρ is different, which is expressed as and θˆg ρ s are defined as follows:
follows. ɺ
θˆ f ρ s =
Sɺ ρ = xˆɺ s ρ − zɺ ρ = θˆTf ρ ξ f ρ + hρ e1 + ∑ j =1, j ≠ s gˆ j u j +
m

(43) 

gˆ ρ s us + fˆρ s − zɺ ρ + θɶTf ρ ξ f ρ + ε f ρ − δ f ρ 2η3 S ρ ξ f ρ s − η f ρ sθˆ f ρ s ,if || θˆ f ρ s ||< M f ρ s or
Hence, α ρ j , j = 1,⋯, m are defined as follows: 
 | | θˆ f ρ s ||= M f ρ s and 2η3 S ρ ξ f ρ s − η f ρ sθˆ f ρ s ≥ 0;
α ρ j = u j = [ N (ς )(k ρ S ρ + ∆ s / S ρ )] / m, ςɺ = − k ρ S ρ2 − ∆ s (44) 
 ˆ
θˆ f ρ sθˆTf ρ s (47)
( ρ − 1)σ 1 ρ −1
η
 3 ρ f ρs
2 S ξ − η θ
f ρs f ρs + (2 η S
3 ρ ξ f ρs −
where ∆ s = µe + + S ρ ( fˆρ +fˆρ s )+∑ i =1 Si +1ki +1e1 +  || θˆ f ρ s ||2
2 
η fi  θˆ f ρ sθˆTf ρ s ˆ
 f ρ s || θˆ ||2 θ f ρ s ), if || θ f ρ s ||= M f ρ s and
η
ρ ˆ
∑ i =1[| Si | (M ε fi + M δ fi ) + 2η θ *fiT θ *fi ] + M ε f ρ s + M δ f ρ s +  f ρ s
1
η gj η f ρ s *T * η g ρ s *T *  2η3 S ρ ξ f ρ s − η f ρ sθˆ f ρ s < 0

∑ j =1 2η
m
θ gj*T θ gj* + θ f ρ sθ f ρ s + θ g ρ sθ g ρ s .
2 2η3 2η 4
Differentiating Vse = esT Pes with respect to time t and
considering (42) and Assumption 4, leads to
Vɺse ≤ −eT Qe + µe = − geVe + µ se (45)
ρ
where ge = λmin (Q) / λmax ( P) , µ se = [∑ i =1 M δ fi + M δ f ρ s + 2 2

∑ j =1 M δ2gj u j2 + M δ2g ρ s us2 ] / λ .

m

Consider the following Lyapunov function

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 9

ɺ ρ −1 1 ρ −1 1 1
θˆg ρ s = Vɺ2 ≤ ∑ i =1 [ Si2+1 + (2 − ki ) Si2 ] + ∑ k =1 ( − +
4 4 ε k +1

 Bk2+1 2 ( ρ − 1)σ 1
2η4 S ρ ξ g ρ s us − η g ρ sθˆg ρ s ,if || θˆg ρ s ||< M g ρ s or ) yk +1 + + S ρ (θˆTf ρ ξ f ρ + θˆTf ρ ξ f ρ ) +
2σ 1 2

 || θˆg ρ s ||= M g ρ s and 2η 4 S ρ ξ g ρ s us − η g ρ sθˆg ρ s ≥ 0; ρ −1 ρ
∑ i =1 Si +1ki +1e1 + Vɺse + ∑ i =1 Si (ε fi − δ fi ) +

 ˆ
θˆg ρ sθˆgTρ s (48)
S ρ (ε f ρ s − δ f ρ s ) + S ρ ∑ j =1, j ≠ s ( gˆ j + ε gj − δ gj )u j +
m
2η4 S ρ ξ g ρ s us − η g ρ sθ g ρ s + (2η 4 S ρ ˆ ξ g ρ s us −
 || θ g ρ s ||2
 ρ η fi ɶT ɶ
θˆg ρ sθˆgTρ s ˆ S ρ ( gˆ s + ε g ρ s − δ g ρ s )us − ∑ i =1 θ fi θ fi −
 2η1
η g ρ s || θˆ ||2 θ g ρ s ), if || θ g ρ s ||= M g ρ s
ˆ
 gρs η gj ɶT ɶ ρ η fi *T *
∑ j =1, j ≠ s 2η θ gjθ gj + ∑ i =1
m
 and 2η S ξ u − η θˆ < 0 θ fi θ fi +
2η1
 4 ρ gρs s gρs gρs 2

where η f ρ s > 0 , η g ρ s > 0 are design constants. η gj 1 1

∑ j =1, j ≠ s 2η θ gj*T θ gj* − 2η θɶTf ρ sθɶf ρ s − 2η θɶgTρ sθɶg ρ s +
m

Applying Young’s Inequality, one has 2 3 4

η f ρ sθɶTf ρ sθˆg ρ s / η3 ≤ −η f ρ s (θɶTf ρ sθɶf ρ s − θ *f Tρ sθ *f ρ s ) / (2η3 ) η f ρ s *T * η g ρ s *T *

θ f ρ sθ f ρ s + θ g ρ sθ g ρ s
2η3 2η 4
η g ρ sθɶgTρ sθˆg ρ s / η4 ≤ −η g ρ s (θɶgTρ sθɶg ρ s − θ g*Tρ sθ g*ρ s ) / (2η4 )
As pointed out in [20], since for any B0 > 0 and p > 0 the sets
Substituting the above inequalities into (46), yields
ρ −1 Π 0 := {( yd , yɺ d , ɺɺyd ) : yd2 + yɺ d2 + ɺɺyd2 ≤ B0 }
Vɺ2 ≤ ∑ i =1 ( Si Si +1 − ki Si2 + Si yi +1 ) +S ρ (θˆTf ρ ξ f ρ + θˆTf ρ ξ f ρ ) +
ρ
Π i := {∑ l =1 ( Sl2 (0) + θɶTfl (0)θɶfl (0)) + θɶTf ρ s (0)θɶf ρ s (0) +
ρ −1 yk2+1 ρ −1
∑ k =1 (− ε + | yk +1 Bk +1 |) + ∑ i =1 Si +1ki +1e1 + Vɺse +
∑ j =1θɶgjT (0)θɶgj (0) + θɶTf ρ s (0)θɶf ρ s (0) + ∑ k =1 yk2+1 (0) +
m i
k +1
ρ
∑ i =1 Si (ε fi − δ fi ) + Sρ (ε f ρ s − δ f ρ s ) + Sρ ( gˆ s + ε g ρ s − 2eT (0)e(0) ≤ 2vs p}, i = 1,⋯, ρ
δ g ρ s )us +S ρ ∑ j =1, j ≠ s ( gˆ j + ε gj − δ gj )u j −
m are compact, Π 0 × Π i is also compact. Thus, | Bi +1 | has a
maximum M i +1 on Π 0 × Π i , vs = max{1 / λmin ( P ),1,η1 ,η 2 ,η3 ,
ρ η fi 1
∑ i =1 2η (θɶTfiθɶfi − θ *fiT θ *fi ) − 2η θɶTf ρ sθɶf ρ s − 1 1
η 4 } . Choose k1 = 2 + β 0 , ki = 2 + β 0 , k ρ = 1 + β 0 ,
1 3
4 4
η gj ɶT ɶ 1
∑ j =1, j ≠ s 2η (θ gjθ gj − θ gj*T θ gj* ) − 2η θɶgTρ sθɶg ρ s +
m 2
1 1 M
= + k +1 + β 0 , k = 1,⋯, ρ − 1, i = 2,⋯, ρ − 1 , where
2 4 ε k +1 4 2σ1
η f ρ s *T * η g ρ s *T *
θ f ρ sθ f ρ s + θ g ρ sθ g ρ s β 0 > 0 ∈ R is a constant. Thus, from (45), one further has
2η3 2η 4
ρ ρ −1 ρ η fi ɶT ɶ
(49) Vɺ2 ≤ − ∑ i =1 β 0 Si2 − ∑ k =1 β 0 yk2+1 − g eVe − ∑ i =1 θ fi θ fi −
From Young’s Inequality, one has 2η1
Si Si +1 ≤ Si2 + Si2+1 / 4, Si yi +1 ≤ Si2 + yi2+1 / 4, η gj 1 1
∑ j =1, j ≠ s 2η θɶgjT θɶgj − 2η θɶTf ρ sθɶf ρ s − 2η θɶgTρ sθɶg ρ s + ∆ s +
m

| yk +1 Bk +1 |≤ yk2+1 Bk2+1 / σ 1 + σ 1 / 2, ∀σ 1 > 0 ∈ R 2 3 4

S ρ ∑ j =1, j ≠ s ( gˆ j + ε gj − δ gj )u j + S ρ ( gˆ s + ε g ρ s − δ g ρ s )u s
m
where σ 1 > 0 ∈ R is a design parameter. Thus, one has
Substituting (44) into the above inequality, leads to
Vɺ2 ≤ − g sV1 + ( hs N (ς ) + 1)ςɺ (50)
where hs = ∑ j =1, j ≠ s ( gˆ j + ε gj − δ gj ) + gˆ ρ s + ε g ρ s −δ g ρ s , g s
m

ηf1 η f ρ η g1 η gm η f ρ s η g ρ s
= min{β 0 , g se , , ⋯, , , ⋯, , , }.
2η1 2η1 2η2 2η2 2η3 2η4
t − gτ
Applying Lemma 2, we conclude that, ∫0 (hs N (ς ) + 1)e ςɺ dτ ,
V2 (t ) and ς (t ) are SGUUB on [0, t f ) . According to
Proposition 2 in [28], if the solution of the closed-loop system is

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 10

bounded, then t f = +∞ . Let µ2 be the upper bound of 1 1 1 1

follows: k1 = 2 + β 0 , ki = 2 + β0 , kρ = 1 + β0 , =
t 4 4 ε k +1 4
∫0 (hs N (ς ) + 1) ⋅ e− gτ ςɺ dτ , we have the following inequalities:
M k2+1
t − gτ t + + β 0 , k = 1,⋯, ρ − 1, i = 2,⋯, ρ − 1 , then, when the
∫0 e (hs N (ς ) + 1)e− gτ ςɺ dτ ≤ ∫ (hs N (ς ) + 1)e− gτ ςɺ dτ ≤ µ2 2σ 1
0
sth actuator is faulty, for s = r , the closed-loop system is
Thus, (50) becomes Vɺ2 ≤ − g sV2 + µ2 . Further, one has semi-globally uniformly ultimately stable and all signals
µ2 µ2 µ2 involved in the closed-loop systems converge to a small
0 ≤ V2 (t ) ≤ + [V2 (0) − ]e− g st ≤ + V2 (0) = µ2 (51)
gs gs gs neighborhood of the origin Ω2 , and for s ≠ r , all signals
which means that V2 (t ) is bounded by µ2 . Thus, all signals of involved in the closed-loop systems do not converge to the
the closed-loop system are uniformly ultimately bounded, i.e. set Ω2 .
| Si |≤ 2 µ2 ,| yi |≤ 2 µ 2 , || θɶfi ||≤ 2η1 µ2 , || θɶgj ||≤ 2η 2 µ2 , Now, we denote the residuals between the real system and
isolation estimators as follows:
|| θɶf ρ s ||≤ 2η3 µ2 , || θɶg ρ s ||≤ 2η 4 µ2 , || e ||≤ µ2 / λmin ( P). J s (t ) = yˆ s (t ) − y (t ) =|S1|, 1 ≤ s ≤ m
That is to say, Si (t ), θɶfi ,θɶgj ,θɶg ρ s ,θɶf ρ s belong to Ω2 defined According to Theorem 2, when the rth actuator is faulty,
as, i.e., s = r , the residual J s (t ) must tend to Ω2 ,; while for
Ω2 := any s ≠ r , basically, J s (t ) does not belong to Ω 2 .
ρ
( S , y ,θɶ ,θɶ ,θɶ ,θɶ ) ∑ ( S 2 + θɶT θɶ / η ) +  Hence, the isolation law for actuator fault can be designed as
 i i fi gj f ρ s g ρ s i =1 i fi fj 1
  J s (t ) ≤ TI , l = s ⇒ the lth actuator is faulty
  
 ∑ j =1, j ≠ s gj gj 2
m ɶ T ɶ ɶT ɶ ɶ T ɶ
θ θ / η + θ θ
f ρs f ρs / η 3 + θ θ
gρs gρs / η +
4   J s (t ) > TI , l ≠ s
 ρ −1

 ∑ k =1 yk2+1 + 2λmin ( P )eT e ≤ 2µ2  where threshold TI is defined as TI = 2µ2 .
 
2) For s ≠ r , one has d g ρ = ( g ρ s − g s )us + ( g r − g ρ r )ur , d f ρ
ˆ ˆ C. Fault Accommodation
After that the fault information is obtained, we will consider
= g s f − θˆ ξ f ρ r From adaptive laws (29), (30), (47) and (48),
s
u T
f ρr
the fault-tolerant control problem, and design a fault-tolerant
ɺ ɺ ɺ ɺ
it is found out that, θˆg ρ r ≠ θˆg ρ s , θˆ f ρ r ≠ θˆ f ρ s . Thus, both ( g ρ s control law to recover the control system's dynamics
performance when an actuator fault occurs. Firstly, we consider
− gˆ s )us + ( g r − gˆ ρ r )ur and g s f s − θˆ f ρ rξ f ρ r do not converge
u T
the fuzzy control problem for the following nominal system
without actuator faults:
to zero, i.e., limt →∞ [( g ρ s − gˆ s )us + ( g r − gˆ ρ r )ur ] ≠ 0 and
 xɺi = xɺi +1 + fi ( x[i ] ), i = 1, 2,⋯, ρ − 1
limt →∞ ( g s f − θˆ ξ f ρ r ) ≠ 0 . As a result, basically, all signals
u T
f ρr 
 xɺ ρ = f ( x,η ) + g ( x,η )u
s T

of the closed-loop systems such as Si do not remain in Ω 2 using  (52)

ηɺ = ψ 1 ( x,η ) + ψ 2 ( x,η )θ
the above control law and adaptive laws. Therefore, from the y = x
above analysis, we can not obtain (50). Furthermore, we can  1

not obtain (51). Hence, we can draw a conclusion that all From Theorem 1, we can see that, under Assumptions 1-4, if
signals involved in the closed-loop systems do not converge to matrices Q > 0 and P = PT > 0 are chosen such that
the set Ω 2 , i.e., Si (t ), θɶfi ,θɶgj ,θɶg ρ s ,θɶf ρ s do not belong to Ω 2 .
PA + AT P + 2λ PP ≤ −Q , and virtual control law (11), (14)
Now, the control procedures are ended. The above design and (17), and adaptive laws (29) and (30) are adopted, then, the
procedures are summarized in the following theorem. closed-loop system is SGUUB stable, and all signals involved in
Theorem 2: Consider the faulty system (40) and observers (41) the closed-loop systems converge to a small neighborhood of
under Assumptions 1-4, fault model (3) adaptive laws (29), (30), the origin Ω1 , which can be adjusted by appropriately choosing
(47) and (48) and control law (44), If matrices H , Q > 0 and
the design parameter. On the basis of the estimated actuator
P = PT > 0 are such that PA + AT P + 2λ PP ≤ −Q , for all fault, the fault tolerant controller is constructed as
ρ us = ρˆ gs (usN − ρˆ fs ( x)) / ( ρˆ gs
2
+ εu )
initial conditions satisfying Π i := {∑ l =1 ( Sl2 (0) +θɶTfl (0)θɶfl (0)) (53)
where ρˆ gs and ρˆ fs are the estimates of ρ gs = g s ( x,η ) ρ s ( x ) and
∑ k =1 yk2+1 (0) + ∑ j =1θɶgjT (0)θɶgj (0) +
i m
+θɶgTρ s (0)θɶg ρ s (0) +
ρ fs = g s ( x,η ) f su ( x ) , ε u > 0 ∈ R is a design parameter, usN is
2eT (0)e(0) + θɶTfρ s (0)θɶf ρ s (0) ≤ 2vs p} , ki , ε i are chosen as
the sth desired control input under the healthy condition.
Theorem 3: Consider faulty system (52) under Assumptions 1-4,
fault model (3), virtual control law (11), (14) and (17), and

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 11

adaptive laws (29) and (30). If there exists a matrix P = PT > 0 u1 (t ), t<2
u1f (t ) =  , u2f (t ) = u2 (t )
with appropriate dimensions, such that (1 − ρ1 ( x ))u1 (t ) + f1 ( x ), t≥2
u

PA + AT P + 2λ PP ≤ −Q
where ρ1 ( x) = 0.4 cos( x3 ) , f1u ( x ) = 0.2 + sin( x2 ) . Initial
then, system (52) is asymptotically stable under the feedback
values of system state are chosen as x1 (0) = 0.1 , x2 (0) = −0.1 ,
FTC (53) and all signals involved in the closed-loop system are
semi-globally uniformly ultimately bounded, converging x3 (0) = 0.1 , x4 (0) = −0.1 . Firstly, Matlab® LMI control
asymptotically to a small neighborhood of the origin. toolbox is used to solve the matrix inequality (39). Therefore,
Proof: Similar to the proof of Theorem 1, it is easy to obtain the one can design the desired control (18) and further design the
conclusions of Theorem 3. The detailed proof is thus omitted fault-tolerant controller (53). Consequently, the observer-based
here. fault-tolerant control input (53) is used to control the faulty
system. Figure 1 shows that the tracking errors can
IV. APPLICATION TO AIRCRAFT LONGITUDINAL MOTION asymptotically converge to a small neighborhood of the origin.
DYNAMICS When an actuator fault occurs at 2s, keeping the normal
In this subsection, for the purpose of demonstrating the controller, both tracking errors deviate significantly from zero
application of the proposed fault tolerant control scheme, we as shown in Figures 2. However, as shown in Figure 3, when the
apply it to accommodate failure for an aircraft longitudinal proposed FTC (53) is activated at about 2.2s, the better
motion dynamics. The aircraft longitudinal motion dynamics of convergence performance is obtained, which illustrates the
the twin otter [3-6] can be described as follows: effectiveness of the proposed FTC scheme.
Vɺ = [ Fx cos(α ) + Fz sin(α )] / m
 V. CONCLUSION
αɺ = q + [ − Fx sin(α ) + Fz cos(α )] / ( mV ) In this paper, the fault-tolerant control problem for a class of
ɺ (54)
θ = q uncertain nonlinear system in presence of actuator fault is
qɺ = M / I discussed. We first design a bank of observers to detect, isolate
 y
and estimate the fault and a sufficient condition for the existence
where V is the velocity, α is the attack angle, θ is the pitch angle of the observers is derived. Simulation results show that the
and q is the pitch rate, m is the mass, I y is the moment of designed fault detection, isolation and estimation algorithms
inertia, and fault-tolerant control scheme have good dynamic
Fx = qSC x (α , q, δ e1 , δ e 2 ) + T1 cos γ 1 + T2 cos γ 2 − mg sin(θ ) , performances in the presence of actuator fault. In this paper, it is
assumed that only an actuator fault occurs. However, multiple
Fz = qSC z (α , q, δ e1 , δ e 2 ) + T1 sin γ 1 + T2 sin γ 2 − mg cos(θ ) , actuator faults as well as sensor faults may occur simultaneously,
M = qcSCm (α , q, δ e1 , δ e 2 ) . For which q = ρV 2 / 2 is the which will be investigated in our further research.
dynamic pressure, ρ is the air density, S is the wing area, c is the 0.1

mean chord, T1 and T2 are independent thrusts with 0.08

corresponding thrust misalignments γ 1 and γ 2 . C x , C z , C m are 0.06

of the polynomial form defined as in (3-6), δ e1 and δ e 2 are the 0.04

Tracking error

elevator angles of an augmented two-pieces elevator used as 0.02

two actuators for failure compensation study. The notations 0

through the model (54) are illustrated as [3-6]. -0.02
Choosing V , α , θ and q as the states x1 , x 2 , x 3 and x 4 ,
-0.04
and δ e1 , δ e 2 , T1 , T2 as the inputs u1 , u 2 , u 3 , u 4 , as shown in -0.06
[3,4], there exists a diffeomorphism -0.08

[η T , xT ]T = T ( χ ) = [ T1 ( χ ), T2 ( χ ), x3 , x4 ]T such that (55) can -0.1

0 2 4 6 8 10
be transform into the PSF form, i.e. t/s

 xɺ 3 = x4 Fig. 1. Tracking error under normal condition


 xɺ 4 = ϑ φ ( x ) + ∑ i =1 bi x1 ui
T 2 2

and the zero dynamics ξɺ = φ (ξ , χ ) + Φ (ξ , χ )ϑ ,

where ϑ ∈ R 4 is an unknown constant vector. Relative degree
is ο = 2 . The parameters in the simulation study are set based
on the data sheet in [3,4]. The fault case considered in this
example is modeled as

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 12

[9] Q. K. Shen, B., Jiang, V. Cocquempot, Fault-tolerant control for T–S

0.1 fuzzy systems with application to near-space hypersonic vehicle with
0.08 actuator faults, IEEE Transactions on Fuzzy Systems, vol. 20, no 4, pp.
652 - 665, 2012.
0.06 [10] X. He, Z. D. Wang, and D. H. Zhou, “Robust fault detection for
networked systems with communication delay and data missing”,
0.04
Automatica, vol. 45, pp. 2634-2639, 2009.
Tracking error

0.02 [11] Y. Wang, D. Zhou, S. J. Qin, and Wang, H, “Active fault-tolerant control
for a class of nonlinear systems with actuator faults”, International
0
Journal of Control, Automation, and Systems, vol. 6, no.3, pp.339-350,
-0.02 2008.
Fault occurs
[12] B. Jiang, Z. Gao, P. Shi, and Xu, Y. “Adaptive fault-tolerant tracking
-0.04 control of near-space vehicle using Takagi-Sugeno fuzzy models”, IEEE
-0.06
Transactions on Fuzzy Systems, vol.18, no.5, pp. 1000-1007, 2010.
[13] L. X. Wang, “Adaptive fuzzy systems and control: design and stability
-0.08 analysis”, Prentice-Hall, Englewood Cliffs, 1994.
[14] S. C. Tong, X. He, and H. Zhang, “A combined backstepping and
-0.1
0 2 4 6 8 10 small-gain approach to robust adaptive fuzzy output feedback control,”
t/s IEEE Transactions on Fuzzy Systems, vol.17, no. 5, pp.1059-1069, 2009.
Fig. 2. Tracking error under faulty condition without FTC [15] A. Boulkroune, M. Tadjine, M. M. Saad, and M. Farza, “How to design a
fuzzy adaptive controller based on observers for uncertain affine
nonlinear systems”, Fuzzy Sets and Systems, vol.159, pp.926–948, 2008.
[16] S. Tong, Y. Li, “Observer-based fuzzy adaptive control for
0.1 strict-feedback nonlinear systems,” Fuzzy Sets and Systems,vol.160,
no.12, pp.1749-1764, 2009.
0.08
[17] S. C. Tong, C. Y. Li, and Y. M. Li, “Fuzzy adaptive observer
0.06 backstepping control for MIMO nonlinear systems”, Fuzzy Sets and
System, vol.160, pp.2755–2775, 2009.
0.04 Fault occurs
[18] S. C. Tong, C. Liu, and Y. Li, “Fuzzy adaptive decentralized control for
large-scale nonlinear systems with dynamical uncertainties”, IEEE
Tracking error

0.02
Transactions on Fuzzy Systems, vol.18, no. 5, pp.845-861, 2010.
0 [19] D. Swaroop, J. K. Hedrick, P. P. Yip, and J. C. Gerdes, “Dynamic surface
control for a class of nonlinear systems”, IEEE Transactions on
-0.02
Automatic Control, 45(10), 1893–1899, 2000.
-0.04 Fault accommodation
[20] D. Wang, and J. Huang, “Neural network-based adaptive dynamic
is activated surface control for a class of uncertain nonlinear systems in
-0.06
strict-feedback form”, IEEE transactions on Neural Networks, vol.16,
-0.08 no.1, pp.195–202, 2005.
[21] Y. S. Yang and C. Zhou, “Robust adaptive fuzzy tracking control for a
-0.1 class of perturbed strict-feedback nonlinear systems via small-gain
0 2 4 6 8 10
t/s approach”, Information Sciences, vol.170, pp.211–234, 2005.
[22] J. H. Park, S. H. Huh, P. S. Yoon, and G. T. Park, “Robustly stable fuzzy
Fig. 3. Tracking error under faulty condition with FTC controller for uncertain nonlinear systems with unknown input gain sign”,
Proceedings of the 2002 IEEE International Conference on Fuzzy
Systems, Honolulu, 1, pp.639-643, 2002.
[23] S. J. Park, J. Bae, and Y. H. Choi, “Adaptive dynamic surface control
REFERENCES stabilization of parametric strict-feedback nonlinear systems with
[1] D. Efimov, A. Zolghadri and T. Raïssi, “Actuator fault detection and unknown time delays”, IEEE Transactions on Automatic Control, vol. 52,
compensation under feedback control,” Automatica, vol. 47, no. no.12, pp.2360-2365, 2007.
8, pp.1699-1705, 2011. [24] X. Zhang, M. M. Polycarpou, and T. Parisini, Robust detection and
[2] J.H. Richter, W.P.M.H. Heemels, N. van de Wouw, and J. Lunze, isolation scheme for abrupt and incipient faults in nonlinear systems”,
“Reconfigurable control of piecewise affine systems with actuator and IEEE Transactions on Automatic Control, vol. 47, no. 4, pp.576- 593,
sensor faults: Stability and tracking”, Automatica, vol. 47, pp. 678–691, 2002.
2011.. [25] R. H. B. Miller andR. Willian, “The effects of icing on the longitudinal
[3] X. Tang, G. Tao, and S. M. Joshi, “Adaptive actuator failure dynamics of an icing research aircraft”, In Proceedings of the 37th AIAA
compensation for parametric strict feedback systems and an aircraft Aerospace Science Meeting (pp. AIAA-99-0636), Reno, NV, January,
application”, Automatica, vol. 39, pp. 1869-1883, 2003. 1999.
[4] X. Tang, G. Tao, and S. M. Joshi, “Adaptive actuator failure [26] A. Ilchmann, “Non-identifier-based high-gain adaptive control”, London,
compensation for nonlinear MIMO systems with an aircraft application. U.K.: Springer-Verlag, 1993.
Automatica, vol. 43, no. 11, pp.1975-1982, 2007. [27] X. Ye,and J. Jiang, “Adaptive nonlinear design without a priori
[5] W. Wang and C. Wen, “Adaptive actuator failure compensation for knowledge of control directions, IEEE Transactions on Automatic
uncertain nonlinear systems with guaranteed transient performance”, Control, vol. 43, pp.1617-1621, 1998.
Automatica, vol. 46, no. 12, pp. 2082 -2091, 2010. [28] E. P.Ryan, “A universal adaptive stabilizer for a class of nonlinear
[6] W. Wang and C. Wen, “Adaptive compensation for infinite number of systems”, Systems Control Letter, vol. 16, pp. 209-218, 1991.
actuator failures or faults”, Automatica, vol.47, no.12, pp.2197-2210, [29] S. S. Ge, F. Hong, and T. H. Lee, “Adaptive neural control of nonlinear
2011. time-delay systems with unknown virtual control coefficients”, IEEE
[7] W. Chen and M. Saif, “Actuator fault diagnosis for a class of nonlinear Transactions on Systems, Man, and Cybernetics—Part B: Cybernetics,
systems and its application to a laboratory 3D crane”, Automatica, vol. 47, vol.34, no.1, pp.499-516, 2004.
no.7, pp.1435-1442, 2011. [30] T. Lin, S. Chang and C. Hsu, Robust adaptive fuzzy sliding mode control
[8] Y. Zhao, J. Lam, and H. Gao, “Fault detection for fuzzy systems with for a class of uncertain discrete-time nonlinear systems, International
intermittent measurements,” IEEE Transactions on Fuzzy Systems, vol. Journal of Innovative Computing, Information and Control, vol.8,
17, no. 2, pp.398–410, 2009. no.1(A), pp.347-359, 2012.

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.

> IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. ×, NO. ×, ×××× 2012 < 13

[31] Y. Lu, B. Wu and S. Lien, Design of an adaptive fuzzy compensator with Vincent Cocquempot received the Ph.D. degree in
disturbance observer using the sliding-mode technique, International automatic control from the Lille University of
Journal of Innovative Computing, Information and Control, vol.8, no.9, Sciences and Technologies, in 1993. He is
pp.5945-5968, 2012. currently a full Professor in automatic control and
[32] J. Yu, Y. Ma, B. Chen and H. Yu, Adaptive fuzzy backstepping position computer science at Lille 1 University, France. He
tracking control for a permanent magnet synchronous motor, is a researcher of the LAGIS-CNRS UMR 8219:
International Journal of Innovative Computing, Information and Control, Automatic Control, Computer Science and Signal
vol.7, no.4, pp.1589-1602, 2011. Processing Laboratory from Lille 1 University and
Head of the team Fault Tolerant Systems in this
[33] M. Liu, P. Shi, L. Zhang and X. Zhao, Fault tolerant control for nonlinear
laboratory. His research interests include robust
markovian jump systems via proportional and derivative sliding mode
on-line Fault Detection and Isolation (FDI) for uncertain dynamical nonlinear
observer, IEEE Transaction on Circuits and Systems I: Regular Papers,
systems and Fault Tolerant Control (FTC) for Hybrid Dynamical Systems
vol.58, no.11, pp.2755-2764, 2011. (HDS). He is a member of IFAC Technical Committee on Fault Detection,
[34] OBLAK, Simon, ŠKRJANC, Igor, BLAŽIČ, Sašo, Fault detection for Supervision, and Safety of Technical Processes. E-mail:
nonlinear systems with uncertain parameters based on the interval fuzzy [email protected]
model, Engineering Applications of Artificial Intelligence, vol. 20, no. 4,
pp. 503-510, 2007.
[35] OBLAK, Simon, ŠKRJANC, Igor, BLAŽIČ, Sašo, If approximating
nonlinear areas, then consider fuzzy systems, IEEE potentials., vol. 25,
no. 6, pp. 18-23, 2006.
[36] Y. Zhan, H. Gao, J. Lam, Stability and stabilization of delayed T-S fuzzy
systems: a delay partitioning approach, IEEE Transactions on Fuzzy
Systems, vol. 17, no.4, pp.750-762, 2009.
[37] J. Li, K. D. Kumar, Decentralized fault-tolerant control for satellite
attitude synchronization, IEEE Transactions on Fuzzy Systems, vol. 20,
no.3, pp. 572 - 586, 2012.
[38] S. Tong, Y. Li, P. Shi, Observer-based adaptive fuzzy backstepping
output feedback control of uncertain MIMO pure-feedback nonlinear
systems, IEEE Transactions on Fuzzy Systems, vol. 20, no 4, pp. 771-
785, 2012.

Qikun Shen was born in Jiangsu, China, 1971. He

received the B.S. degree in Computer Science and
Applications from Chinese University of Mining
and technology, Xuzhou, China in 1996, and the
M.S. degree in Computer Science and
Applications from Yangzhou University,
Yangzhou, China in 2007. Now, he is currently a
Ph.D. candidate with the College of Automation
Engineering in Nanjing University of Aeronautics
and Astronautics. He is currently an associate
professor in College of Information Engineering, Yangzhou University. His
research interests include fault-tolerant control, adaptive control, fuzzy control
and intelligent control, etc. E-mail: [email protected].

Bin Jiang was born in Jiangxi, China, in 1966. He

obtained the Ph.D. degree in Automatic Control
from Northeastern University, Shenyang, China, in
1995. He had ever been a postdoctoral fellow or a
research fellow in Singapore, France and USA,
respectively. Now he is a Professor and Dean of
College of Automation Engineering in Nanjing
University of Aeronautics and Astronautics, China.
He currently serves as Associate Editor or Editorial
Board Member for a number of journals such as
IEEE Trans. On Control Systems Technology; Int. J.
of Systems Science; Int. J. of Control, Automation and Systems; Int. J. of
Innovative Computing, Information and Control; Int. J. of Applied
Mathematics and Computer Science; Acta Automatica Sinica; Journal of
Astronautics. He is a senior member of IEEE, a member of IFAC Technical
Committee on Fault Detection, Supervision, and Safety of Technical Processes.
His research interests include fault diagnosis and fault tolerant control and their
applications. E-mail: [email protected]

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing [email protected].