Fault Tolerant Control for Takagi-Sugeno systems with unmeasurable
premise variables by trajectory tracking
Dalil Ichalal, Benoit Marx, José Ragot, Didier Maquin
Abstract— This paper presents a new method for fault
tolerant control of nonlinear systems described by Takagi-
Sugeno fuzzy systems with unmeasurable premise variables.
The idea is to use a reference model and design a new
control law to minimize the state deviation between a healthy
reference model and the eventually faulty actual model. This
scheme requires the knowledge of the system states and of
the occurring faults. These signals are estimated from a
Proportional-Integral Observer (PIO) or Proportional-Multi-
Integral Observer (PMIO). The fault tolerant control law is
designed by using the Lyapunov method to obtain conditions
which are given in Linear Matrix Inequality formulation
(LMI). Finally, an example is included.
Index Terms— Takagi-Sugeno fuzzy systems, state and fault
estimation, PI and PMI observers, Lyapunov stability analysis,
linear matrix inequality.
I. INTRODUCTION
Fault tolerant control (FTC) has been recently introduced
in the fault diagnosis framework. It consists to compute a
new control law by taking into account the faults affecting
the system in order to maintain acceptable performances and
preserve stability of the system in the faulty situations.
The existing strategies are classified into two classes.
The first class is called passive fault tolerant control or
robust control. In this approach, the faults are treated as
uncertainties. Therefore, the control is designed to be robust
only to the specified faults. Contrarily to the passive FTC,
active FTC requires a FDI block to detect, isolate and
estimate the faults. The informations issued from the FDI
block are used by the FTC module to reconfigure the control
law in order to compensate the fault and ensure an acceptable
system performances.
The active fault tolerant control has been developed essen-
tially for linear systems [4], [15], [13], [11] and descriptor
linear systems [9]. Clearly, linear models do not often repre-
sent accurately physical systems due to nonlinear behaviors.
It is then interesting to work directly with nonlinear models.
Nevertheless, from the mathematical point of view, working
with nonlinear models is much harder than with linear ones.
A new representation that combines simplicity of linear
models and accuracy of nonlinear behaviors is introduced,
initially, in [16] and known as Takagi-Sugeno (T-S) models.
The idea is to consider a set of system operating points. At
each operating point, the system is represented by a simple
All the authors are with the Centre de Recherche en
Automatique de Nancy (CRAN), Nancy-Université, 2, av-
enue de la forêt de Haye 54516 Vandoeuvre-les-Nancy
{dalil.ichalal, benoit.marx, jose.ragot,
didier.maquin}@ensem.inpl-nancy.fr
linear sub-model. An interpolation of all these sub-models
with nonlinear functions satisfying the sum convex property
allows to obtain the global behavior of the system. One
can cite some works in FTC field for nonlinear systems,
for example, in [5], the authors give a method for actuator
faults by using their estimations, for nonlinear descriptor
systems with Lipschitz nonlinearities. In [14], a method
requiring only the fault isolation is proposed for T-S systems.
It is based on controller based observers bank. A switching
system is then designed, to switch from one controller to an
other, from the residual decision logic.
This paper is dedicated to the design of a fault tolerant
control strategy for nonlinear systems described by Takagi-
Sugeno models with unmeasurable premise variables. The
main idea is to re-use the nominal control input developed
in fault-free case to which two terms, related to the occurred
fault and the tracking error trajectory between the system
and a reference model, are added to be able to compensate
the fault. The reference trajectory is provided by a reference
model representing the system without faults. In addition,
the control law requires the knowledge of the state of the
system and faults affecting it. For that purpose, a PI (or PMI)
observer is used to estimate simultaneously these signals.
The second section is dedicated to a brief presentation
of Takagi-Sugeno models. The third section deals with the
problem of fault tolerant control design with PI and PMI
observers. Finally, an academic example is proposed in order
to illustrate the FTC strategy.
II. TAKAGI -S UGENO STRUCTURE FOR MODELING
Consider a nonlinear system described by

ẋ(t) = f (x(t), u(t))
(1)
y(t) = h(x(t), u(t))
The T-S fuzzy modeling allows to represent the behavior
of a nonlinear system (1) by the interpolation of a set
of linear sub-models. Each sub-model contributes to the
global behavior of the nonlinear system through a weighting
function µi (ξ(t)). The T-S structure is given by

Pr
 ẋ(t) = µi (ξ(t))(Ai x(t) + Bi u(t))


i=1
Pr (2)
 y(t) = µi (ξ(t))(Ci x(t) + Di u(t))


i=1
where x(t) ∈ R is the state vector, u(t) ∈ Rm is the input
n
vector, y(t) ∈ Rp represents the output vector. Ai ∈ Rn×n ,
Bi ∈ Rn×m , Ci ∈ Rp×n and Di ∈ Rp×m are known
matrices. The functions µi (ξ(t)) are the weighting functions
 f (t)
depending on the variables ξ(t) which can be measurable
(as the input or the output of the system) or non measurable
u(t) + uf (t) yf (t)
variables (as the state of the system). These functions verify System
+
the following properties
 r
 P µ (ξ(t)) = 1
i Observer
(3)
 i=1 x̂f (t)
0 ≤ µi (ξ(t)) ≤ 1 ∀i ∈ {1, 2, ..., r} fˆ(t)
−
Obtaining a T-S model (2) from (1) can be performed from Controller
different methods such us linearizing the system (1) around +
some operating points and using adequate weighting func- x(t)
reference
tions. It can also be obtained by black-box approaches which model
allow to identify the parameters of the model from input-
output data. Finally, the most interesting and important way Fig. 1. Fault tolerant control scheme
to obtain a model in the form (2) is the well-known nonlinear
sector transformations [17], [12]. Indeed, this transformation
allows to obtain an exact T-S representation of (1) with no The matrices K1i are determined in order to ensure the
information loss on a compact of state space. stability of the system even if faults occur and to minimize
Thanks to the convex sum property of the weighing the difference between xf (t) and x(t). By analyzing the
functions (3), it is possible to generalize some tools devel- structure of uf (t) given in equation (6), the estimation of
oped in the linear domain to the nonlinear systems. This the state xf (t) and the f (t) faults is required. This task is
representation (2) is very interesting in the sense that it performed via a Proportional-Integral observer which esti-
simplifies the stability studies of nonlinear systems and the mates simultaneously the state and the faults of the system.
design of control laws and observers. In [2], [6], [7], the Let us consider the PI observer
stability and stabilization tools are inspired from the study r
˙x̂f = µi (x̂f )(Ai x̂f + Bi (uf + fˆ)
X
of linear systems. In [1], [10], the authors worked on the
problem of state estimation and diagnosis of T-S fuzzy i=1
systems. The proposed approaches in these last papers rely + H1i (yf − ŷf )) (7)
on the generalization of the classical observers (Luenberger r
˙
fˆ =
X
Observer [8] and Unknown Input Observer (UIO) [3]) to the µi (x̂f ) (H2i (yf − ŷf )) (8)
nonlinear domain. In the remaining of the paper, we use the i=1

following lemmas. ŷf = C x̂f + Rfˆ (9)

Lemma 1: Consider two matrices X and Y with appropri-
which depends on the gains H1i and H2i .
ate dimensions and Ω a positive definite matrix. the following
The output error between the system (5) and the observer
property is verified
(7)-(8) is written by
X T Y + Y T X ≤ X T ΩX + Y T Ω−1 Y Ω > 0 (4)
yf − ŷf = C̃ea (10)
Lemma 2: (Congruence) Let two matrices P and Q, if P
is positive definite and if Q is a full column rank matrix, where
than the matrix QP QT is positive definite.  
  xf
C̃ = C R , xa = ,e a = xa − x̂a (11)
III. FAULT TOLERANT CONTROL OF T-S FUZZY SYSTEMS f
Let us consider the T-S reference model without faults The dynamic of the trajectory tracking error e = x − xf ,
described by (2). The system with faults f is described by obeys to the differential equation
the following T-S model with unmeasurable premise variable r
X
r

 ẋ (t) = P ė = µi (x)(Ai x + Bi u)
f µi (xf (t)) (Ai xf (t) + Bi (uf (t) + f (t)))
i=1
i=1

yf (t) = Cxf (t) + Rf (t) − µi (xf )(Ai xf + Bi (uf − f )) (12)
(5)
Taking into account the definitions (6) and (12) leads to
For sake of simplicity, the time variable is omitted.
r
r X
The goal is to design the control law uf (t) such that the
µi (xf )µj (x̂f )(Ai e − Bi (f − fˆ)
X
system state xf (t) converges toward the reference state x(t) ė =
given by the reference model (2). The control strategy is i=1 j=1

illustrated in the figure 1. The following structure is proposed − Bi K1j (xf − x̂f )) + ∆1 (13)
r r
for the control law XX
r   = µi (xf )µj (x̂f )((Ai − Bi K1j )e − L̃ij ea )
µi (x̂f ) −fˆ + K1i (x − x̂f ) + u
X
uf = (6) i=1 j=1

i=1 + ∆1 (14)
where following theorem 1.


L̃ij = Bi K1j Bi , ea = xa − x̂a (15) Theorem 1: The state tracking error e(t) and the state
r
X and fault estimation errors ea (t) converge asymptotically
∆1 = (µi (x) − µi (xf ))(Ai x + Bi u) (16) toward zero if there exists symmetric and positive definite
i=1 matrices X1 , P2 and matrices H̄1i , H2i and K1i such that γ̄
In order to analyze the evolution of the errors, two cases is minimized under the LMI constraints (27).
are considered : in the first one the faults are supposed to be 
Ψi −B1 K1j −Bi −B1 K1j X1 X1 0

constant ; in the second one they are assumed to be under a  ∗ Θj Ξj 0 0 0 P2 
polynomial form with respect to the time variable. 
 ∗ ∗ Φj 0 0 0 0 

 
 ∗ ∗ ∗ −I 0 0 0 
A. Constant faults  <0
 ∗ ∗ ∗ ∗ −I 0 0 
In this first approach, we have f˙(t) = 0 and, with 
 ∗

∗ ∗ ∗ ∗ −γ̄I 0 
definition (11), the system (5) becomes in augmented form
∗ ∗ ∗ ∗ ∗ ∗ −γ̄I
r
 
 ẋ = P µ (x ) Ã x + B̃ u
 (27)
a i f i a i f
i=1 (17) Ψi = Ai X1 + X1 ATi (28)
yf = C̃xa

Θj = P2 Aj + ATj P2 − H̄1j C − C T H̄1j
T
(29)
where
    Ξj = P2 Bj − H̄1j R − C T H̄2j
T
(30)
Ai Bi Bi T T
Ãi = , B̃i = , (18) Φj = −H̄2j R − R H̄2j (31)
0 0 0

i, j = 1, ..., r
C̃ = C R (19)
The gains are given by K1j and H2j are obtained directly
The state and fault estimation error ea (t) = xa (t) − x̂a (t) from the above optimization problem and H1j are then
between the system (17) and the observer (7)-(8) evolves computed from
according to the following equation H1j = P2−1 H̄1j (32)
r  
the L2 gain from ∆ to ẽ is given by
X
ėa = µi (x̂f ) (Ãi − Hi C̃)ea + Γ∆2 (20)
i=1
√
γ = γ̄ (33)
where
Proof: The gains H1i , H2i and K1i are obtained by a
 
In
Γ= (21) stability analysis of the system described by the differential
0
r
equation (23), using the Lyapunov theory with a quadratic
X function.
∆2 = (µi (xf ) − µi (x̂f ))(Ai xf + Bi (uf + f )) (22)
Let us chose the following quadratic Lyapunov function
i=1

The concatenation of the state tracking trajectory error and V (ẽ) = ẽT P ẽ, P = P T > 0 (34)
the state and faults estimation errors allows to write, from
where P is chosen as follows
(13) and (20), the new augmented system is  
r
r X
P1 0 0
X P =  0 P2 0  (35)
ẽ˙ = µi (xf )µj (x̂f (t))Ãij ẽ + Γ̃∆ (23)
0 0 P3
i=1 j=1

where The time derivative of the function V (ẽ) is given by

 
x − xf  
∆1
ẽ =  xf − x̂f  , ∆= , (24) r
r X  
∆2
X
f − fˆ V̇ (ẽ) = µi (xf )µj (x̂f )ẽT ÃTij P + P Ãij ẽ
  i=1 j=1
In 0 + 2P Γ̃∆
Γ̃ =  0 In  (25) Xr X r
0 0 = µi (xf )µj (x̂f )ẽT Mij ẽ + 2P Γ̃∆ (36)
i=1 j=1
 
Ai − Bi K1j −Bi K1j
−Bi where
Ãij =  0 Aj − H1j C
Bj − H1j R   
Λij −P1 Bi K1j −P1 Bi
0 −H2j C
−H2j R
(26) Mij = S  0 Θj Σj  (37)
0 −P3 H2j C −P3 H2j R
The gains K1i , H1i and H2i are determined by solving a
minimization problem under LMI constraints, given by the Λij = P1 Ai − P1 Bi K1j (38)
 Θj = P2 Aj − P2 H1j C (39) Let us remark that the bloc matrix
 
Σj = P2 Bj − P2 H1j R (40) Ψij −Bi K1j −Bi
 ∗ Zj Υj  (54)
and S is a function that acts on any matrix X as follows ∗ ∗ Tj
S(X) = X T + X (41) of (48) can be written as follows
 
Assume that the input and the faults are bounded and that Ψi −B1 K1j −Bi
the system is stable. As a consequence, ∆ ((24), (16), (22))  ∗ Zj Υj 
is bounded. So, the objective is to minimize the L2 -gain of ∗ ∗ Tj
∆ on the error ẽ(t), this is formulated by −Bi K1j
T 
−Bi K1j T
X1 X1
   
kẽk2 + 0 0 + 0 0 <0
< γ, k∆k2 6= 0 (42) 0 0 0 0
k∆k2 (55)
Then, we are seeking to ensure asymptotic convergence of
where
ẽ(t) toward zero if ∆(t) = 0 and to guarantee a bounded L2
if ∆(t) 6= 0. This problem can be formulated as follows Ψi = Ai X1 + X1 ATi (56)
T 2 T
V̇ (ẽ) + ẽ ẽ − γ ∆ ∆ < 0 (43)
The lemma 1 gives
After some calculation and by using the convex sum property !
Ψi −B1 K1j −Bi
of the weighting functions, the time varying inequality (43) ∗ Zj Υi
is satisfied if the following conditions hold ∗ ∗ Ti
−Bi K1j −Bi K1j T
   
Nij < 0, i, j = 1, . . . , r (44) + 0 Ω −1
0
0 0
where
X1 X1 T
   
 
Λij −P1 Bi K1j −P1 Bi P1 0 + 0 Ω 0 <0 (57)
 0 Θj Σj 0 P2   0 0
 
Nij = S  0 −P3 H2j C −P3 H2j R 0
  0 

 where Ω is a symmetric and positive matrix. After bounding
2
  P1 0 0 −γ I 0  the inequality (48) with (57), and, assuming that
P1 0 0 0 −γ 2 I
(45) H̄1i = P2 H1i (58)
by congruence (lemma 2), for every invertible matrix W , we γ̄ = γ2 (59)
have
Ω = I, P3 = I (60)
Nij < 0 ⇔ W Nij W < 0 (46)
defining W by the LMIs in theorem 1 are obtained.
 
P1−1 0 0 0 B. Time varying faults
 0 I 0 0 
W =
 0
 (47) The assumption that the fault signal is constant over the
0 I 0 
time is restrictive, but in many practical situations where
0 0 0 I
the faults are slowly time-varying signals, the estimation of
the inequality (44) is equivalent to the faults is correct, and the proposed FTC scheme can be
  applied. In the case where the faults are not slowly time-
Ψij −Bi K1j −Bi P1 0
 ∗ varying or constant, the Proportional Integral Observer (PIO)
Z j Υ j 0 P2 

 ∗
 can be replaced by a Proportional Multiple Integral Observer
∗ T j 0 0 <0 (48)
  (PMIO). Such an observer is able to estimate a large class
 ∗ ∗ ∗ −γ 2 I 0 
of time-varying signals satisfying the following assumption
∗ ∗ ∗ ∗ −γ 2 I
where f (q+1) = 0 (61)

Ψij = Ai X1 + X1 ATi − Bi K1j X1 − X1 K1j

T
BiT (49) The principle of this observer is based on the estimation of
all the first q th derivatives of the signal f (t). This observer
Zj = P2 Aj + ATj P2 − P2 H1j C − C T H1j
T
P2 (50)
T T
can also be extended to the case where f (q+1) is bounded.
Υj = P2 Bj − P2 H1j R − C H2j P3 (51) Let consider the system (5) with a fault in the general
Tj = −P3 H2j R − RT H2j
T
P3 (52) polynomial form

X1 = P1−1 (53) f (t) = a0 + a1 t + a2 t2 + ... + aq tq (62)

    
Let consider d0 (t) = f˙(t), d1 (t) = f¨(t), ..., dq−1 (t) = 1 1 1 1
C= , R=
f (q) (t), the system can be transformed into an augmented 1 0 1 0
form  r   The weighting functions depend on the first component of
 x̃˙ = P µ (x ) Ã x̃ + B̃ u the state vector xf ; they are defined by
f i f i f i f
i=1 (63) (

y = C̃ x̃f 1−tanh(x1f )
µ1 (xf ) = 2 (67)
where µ2 (xf ) = 1 − µ1 (xf )
   
xf Ai Bi 0 0 0 The input variation u(t) over the time is depicted in the figure

 d0 


 0 0 I 0 0 
 2 (bottom, continuous blue line). To apply the proposed FTC
x̃f =  .. .. .. .. .. .. strategy, the following reference model is considered
 , Ãi =  ,
   
 .   . . . . .
r
 0 
 dq  0 0 0 I   ẋ = P µi (x) (Ai x + Bi u)
d(q−1) 0 0 0 0 0 i=1 (68)
y = Cx
  
Bi
 0  The fault f (t) is a time varying signal at t = 5. Solving the
 
B̃i =  ... LMIs in theorem 1 results in the following matrices


 , C̃ = C R 0 0 0
 
   
 0  0.91 0.11 0.04
0 X1 =  0.11 0.93 −0.04  ,
0.04 −0.04 0.44
x̃f (t) represents the augmented state vector composed of the
state xf (t) and the q th first successive derivatives of the fault
 
1.53 −0.31 0.50
f (t). The observer simultaneously estimating the state xf (t) P2 =  −0.31 3.04 −0.39  ,
and the faults f (t) with the successive derivatives is given 0.50 −0.39 0.95
in the following form    
−1.93 4.58 −3.39 5.12
r

ˆ˙ f =
 
 x̃ P
ˆf + B̃i uf + H̃i (y − ŷ)
µi (x̂f ) Ãi x̃ H11 =  −3.19 6.27  , H12 =  −3.27 6.67  ,
i=1 (64) −5.35 1.22 −4.47 2.74
 ˆf
ŷ = C̃ x̃    
H21 = 4.885 0.000 , H22 = 3.771 1.114 ,
ˆf (t)
The augmented state estimation error e(t) = x̃f (t) − x̃  
and the error between xf and x are given by K11 = 0.004 0.024 −0.004 ,
 
  X r X r   K12 = 0.003 0.019 −0.004
ė e
= µi (x)µj (x̂f )Ā˜ij + Γ̃∆ (65)
ėa ea The proportional-integral observer provides estimated state
i=1 j=1
and faults. In the figure 2 (top) the real fault and its estimate
where are depicted. The state estimation errors (resp. the state
 
Ai − Bi K1j L̃ij tracking errors) are displayed on the top (resp. bottom) of
Ā˜ij =
0 Ãj − H̃j C̃ figure 3. The figures 4 compares the state variables of the
reference model, of the faulty system without FTC and the
Thus, the structure of the state equations is the same as those
faulty system with FTC. One can see that the state variables
expressed in the case of constant faults. The synthesis of the
of the system affected by fault with FTC is closed to the
gains of the controller and those of the observer are obtained
reference whereas the faulty system with nominal control
by solving the LMIs given in the theorem 1.
deviates.
IV. S IMULATION EXAMPLE
V. CONCLUSION
The proposed fault tolerant control strategy is illustrated
This paper is dedicated to the design of a nonlinear fault
on an academic T-S system described by
tolerant control law. The considered systems are modeled
r

 ẋ = P µi (xf ) (Ai xf + Bi (uf + f )) in the Takagi-Sugeno fuzzy structure with unmeasurable
f
i=1 (66) premise variables. The strategy is based on the use of a
yf = Cxf + Rf

reference model which is the model of the system in the
where fault-free case. The proposed control law is then designed to
    minimize the deviation of the system state compared to the
−2 1 1 −3 2 2
reference state, even in the presence of fault(s). This control
A1 =  1 −3 0  , A2 =  0 −3 0 ,
law uses the nominal control input developed for the system
2 1 −8 5 2 −4
    in fault-free case and two additional terms. The first term is
0 1 related to the estimated fault and the second one corresponds
B1 =  1  , B2 =  1  , to the trajectory tracking error. The stability is studied with
0.25 0 the Lyapunov theory and a quadratic function that allows to
 1.5
derive conditions ensuring the convergence of the state and
1
fault estimation errors and trajectory tracking error toward
0.5 zero. The existence conditions are expressed in terms of
f(t)
0 estimated f(t) LMI that can be solved with classical dedicated softwares.
−0.5 The future works may be oriented, on the one hand, to
−1
the relaxation of these conditions by using polyquadratic
0 1 2 3 4 5 6 7 8 9 10
or non-quadratic Lyapunov functions. On the other hand,
the assumption of open-loop stability (needed for solving
2
u(t)
uf(t)
the LMI problem given in theorem 1) should be relaxed.
1 In addition, it is interesting to develop the FTC control law
0
by taking into account modeling uncertainties, multiplica-
tive faults and some external perturbations, and considering
−1
nonlinear outputs of the system. Real applications will be
−2
developed in future works.
0 1 2 3 4 5 6 7 8 9 10

R EFERENCES
Fig. 2. Fault and its estimate (top) - nominal control input (continuous
blue line) FTC input (dashed red line) (bottom) [1] A. Akhenak, M. Chadli, J. Ragot, and D. Maquin. Fault detection and
isolation using sliding mode observer for uncertain Takagi-Sugeno
state estimation errors
fuzzy model. In 16th Mediterranean Conference on Control and
1 Automation Congress Centre, Ajaccio, France, June 25-27 2008.
0.8 [2] M. Chadli, D. Maquin, and J. Ragot. Non quadratic stability analysis
0.6
of Takagi-Sugeno systems. In IEEE Conference on Decision and
Control, CDC’2002, Las Vegas, Nevada, USA, 2002.
0.4
[3] M. Darouach, M. Zasadzinski, and S.J. Xu. Full-order observers for
0.2
linear systems with unknown inputs. IEEE Transactions on Automatic
0 Control, 39(3):606–609, March 1994.
−0.2 [4] Z. Gao and P.J. Antsaklis. Reconfigurable control system design via
0 1 2 3 4 5 6 7 8 9 10 perfect model following. International Journal of Control, 56(4):783–
798, 1992.
state tracking errors [5] Z. Gao and S. X. Ding. Actuator fault robust estimation and fault-
0.05
tolerant control for a class of nonlinear descriptor systems. Automatica,
0 43(5):912–920, 2007.
[6] T-M. Guerra, A. Kruszewski, L. Vermeiren, and H. Tirmant. Con-
−0.05
ditions of output stabilization for nonlinear models in the Takagi-
−0.1
Sugeno’s form. Fuzzy Sets and Systems, 157(9):1248–1259, May 2006.
[7] A. Kruszewski, R. Wang, and T.M. Guerra. Nonquadratic stabilization
−0.15 conditions for a class of uncertain nonlinear discrete time TS fuzzy
−0.2
models: A new approach. IEEE Transactions on Automatic Control,
0 1 2 3 4 5 6 7 8 9 10 53(2):606 – 611, March 2008.
[8] D.G. Luenberger. An introduction to observers. IEEE Transactions
Fig. 3. State estimation errors (top) - State trajectory tracking (bottom) on Automatic Control, 16:596–602, 1971.
[9] B. Marx, D. Koenig, and D. Georges. Robust fault tolerant control
State of the reference model
for descriptor systems. IEEE Transactions on Automatic Control,
0.4 State of the system without FTC 49:1869–1875, 2004.
State of the system with FTC
0.2 [10] B. Marx, D. Koenig, and J. Ragot. Design of observers for takagi
x1(t)

0 sugeno descriptor systems with unknown inputs and application to

−0.2 fault diagnosis. IET Control Theory and Application, 1:1487–1495,
−0.4 2007.
0 1 2 3 4 5 6 7 8 9 10
[11] M.M. Mufeed, J Jiang, and Z. Zhang. Active Fault Tolerant Control
1 Systems : Stochastic Analysis and Synthesis. Springer, 2003.
[12] A.M. Nagy, G. Mourot, B. Marx, G. Schutz, and J. Ragot. Model
0.5 structure simplification of a biological reactor. In 15th IFAC Sympo-
x2(t)

sium on System Identification, SYSID’09, Saint Malo, France, 2009.

0
[13] H. Noura, D. Sauter, F. Hamelin, and D. Theilliol. Fault-tolerant
−0.5
control in dynamic systems: Application to a winding machine. IEEE
0 1 2 3 4 5 6 7 8 9 10 Control Systems Magazine, 20(1):33–49, 2000.
[14] M. Oudghiri. Commande multimodles tolérante aux défauts : appli-
0.5
cation au contrôle de la dynamique d’un véhicule automobile. PhD
thesis, Université de Picardie Jules Verne, Amiens, France, 2008.
x3(t)

0 [15] M. Staroswiecki. Fault tolerant control : the pseudo-inverse method

revisited. In 16th IFAC World Congress, Seoul, Korea, 2005.
[16] T. Takagi and M. Sugeno. Fuzzy identification of systems and its
−0.5
0 1 2 3 4 5 6 7 8 9 10 applications to modeling and control. IEEE Transactions on Systems,
Man, and Cybernetics, 15:116–132, 1985.
Fig. 4. Comparison between the states of the reference model (no fault), [17] K. Tanaka and H.O. Wang. Fuzzy Control Systems Design and
the states of the system with fault and without FTC and states of the system Analysis: A Linear Matrix Inequality Approach. John Wiley and Sons,
with FTC 2001.