Proceedings of the 6th International Conference on SuBD.

4
Systems and Control, University of Batna 2, Batna,
Algeria, May 7-9, 2017

Robust adaptive observer design for fast fault estimation for nonlinear
T-S fuzzy systems using descriptor approach
Salama Makni 1 , Maha Bouattour and Mohamed Chaabane2 and Ahmed El Hajjaji3

Abstract— In this work, we study the detection and es- chosen and conditions must be satisfied.
timation problem of sensor and actuator faults (SAF) for In this work, a robust observer is developed to solve the
nonlinear systems represented by Takagi-Sugeno (T-S) fuzzy simultaneous estimation problem of both SAF for nonlinear
models subject to an Unknown Bounded Disturbance (UBD).
Simultaneously, we study the estimation of the states. To do this, systems. Restrictive method [3] is not needed to solve the
we design an adaptive descriptor observer to jointly estimate LMI. The H∞ performance is considered into the observer
state and SAF vectors. This observer not only lead to estimate to minimize the disturbance effects.
the actuator fault, but also to improve the performance of fault The present paper is structured as follows. In part 2,
estimation. To prove the existence of this observer, two sufficient the fuzzy system description is given. A robust adaptive
conditions are deduced in form of Linear Matrix Inequalities
(LMIs) using Lyapunov theory. A robustness to disturbance descriptor observer design is determined in the third section.
is elaborated by considering the H∞ performance. Finally, a In the fourth part, an example which proves the effectiveness
numerical example and simulation results are afforded to prove of the suggested method is shown. To close, some comments
the mentioned process efficiency. are given in part 5.
Notations. PT and P−1 are respectively the transpose and
I. INTRODUCTION
the inverse of matrix P. Sym(P) = P + PT .
In several researches, the fuzzy modeling approach has
been exploited for control and nonlinear systems diagnosis. II. S YSTEM DESCRIPTION
Recently, the T-S fuzzy representation had attached rising We consider the following T-S fuzzy model :
interest because the synthesis of the observer or the controller
r
becomes less complex for nonlinear systems [10]. In most
ẋ(t) = ∑ µi (ξ (t))(Ai x(t) + Bi u(t))
real control systems, unplanned faults occur at uncertain time i=1 (1)
which may lead to destabilizing and damage the system. y(t) = Cx(t)
Consequently, researches on fault detection and isolation FDI
has taken into consideration these malfunctions [2]. The where x(t) ∈ Rn , y(t) ∈ R p and u(t) ∈ Rm represent state
objective of the robust FDI is to differentiate between the vector, output vector and the input vector respectively.
fault effects and the effect of unknown disturbances. Indeed, Ai , Bi and C are constant real matrices. r is the number of
a fast detection and isolation of system faults is another submodels, ξ is the premise variable assumed measurable
objective of design [5]. and µi are the weighting functions which satisfy :
Many authors like [8] focus on sensor faults and others like r
[7] are obtained for actuator ones only. Whereas, SAF may ∑ µi (ξ (t)) = 1 and µi (ξ (t)) > 0 f or i = 1, . . . , r. (2)
occur at the same time which is more complicated. So, to i=1
be close the reality, it is better to consider SAF under one Now, we take into account this model assumed by SAF
unified framework. For this reason, we are motivated to and UBD :
establish a novel observer to estimate simultaneously SAF
vectors. Different approaches have been suggested for FDI r

of T-S fuzzy faulty systems by involving different techniques. ẋ(t) = ∑ µi (ξ (t))(Ai x(t) + Bi u(t)) + Da fa (t) + Bd d(t)
i=1 (3)
These techniques include fuzzy descriptor approach, sliding
y(t) = Cx(t) + Ds fs (t)
mode observer and linear matrix inequality LMI approach
[9],[8],[4]. where fa (t) ∈ Ra , fs (t) ∈ Rs and d(t) ∈ Rnd represent
Particular attention was considered to the diagnosis [1],[2] of actuator fault, sensor fault and UBD vector respectively.
the fuzzy systems with various performances [6]. To satisfy It is proposed that the derivative of actuator fault is norm
these performances, a suitable Lyapunov function must be bounded with respect to time by a finite value fm . Bd , Da
and Ds are constant real matrices. We note that Ds is a full
*This work was not supported by any organization
1 Salama Makni is with STA Laboratory, University of Sfax, University rank column.
of Gabes, Gabes, Tunisia 6029, maknisalama@gmail.com
2 Maha Bouattour and Mohamed Chaabane are with STA Laboratory,
To estimate state and sensor fault vectors simultaneously,
University of Sfax, 3038 Tunisia maha_bouattour@yahoo.fr, we use the technique of the augmented system and the
chaabane.ucpi@gmail.com
3 Ahmed El Hajjaji is with MIS Laboratory, University of Picardie Jules descriptor approach. Then, we apply an algorithm to
Vernes, Amiens,80000 France hajjaji@u-picardie.fr estimate actuator fault vector. For this, equation (3)

978-1-5090-3960-9/17/$31.00 ©2017 IEEE 249

becomes :
r
Subtracting (8) from (4) yields :
Ē x̄˙(t) = ∑ µi (ξ (t))(Āi x̄(t) + B̄i u(t))
i=1
(4) (Ē + ELC̄)x̄˙(t) − E x̄˙ˆ(t)
+D̄a fa (t) + B̄d d(t) + Nxs (t) r
y(t) = C̄x̄(t) = C0 x̄(t) + xs (t) = ∑ µi (ξ (t)) (Āi + Fi LC0 )x̄(t) − Fi x̄ˆ(t) (9)
i=1

+(N + Fi L)xs (t) + D̄a e fa (t) + B̄d d(t).
where
We define ē(t) = x̄(t) − x̄ˆ(t) as a state error estimation,
e fa (t) = fa (t) − fˆa (t) as an actuator fault error estimation
     
In 0 Ai 0 Bi
Ē = , Āi = , B̄i = ,
0 0 0 −I p 0 and consider that :
   
C̄ = C I p , C0 = C 0 , (5)
      E = Ē + ELC̄ (10)
0 Bd Da
N= , B̄d = , D̄a = .
Ip 0 0

and 
x(t)
 Fi = Āi + Fi LC0 (11)
xs (t) = Ds fs (t), x̄(t) =
xs (t)
xs (t) were supposed as an auxiliary state of (4).
Two Lemmas are given as follows.
N = −Fi L. (12)
Lemma 1 For matrices X and Y and for matrix P = PT > 0,
we have the next inequality :
X T Y +Y T X 6 X T PX +Y T P−1Y Then equation (9) becomes :
Lemma 2 Given matrix X = X T and matrix Π < 0 where
X T ΠX < 0, then there exist a positive scalar ν such that : r

X T ΠX 6 −2νX − ν 2 Π−1 E ē˙(t) = ∑ µi (ξ (t))Fi ē(t) + D̄a e fa (t) + B̄d d(t) (13)
i=1

To estimate conjointly state and SAF vectors, we propose which is the dynamic error.
the next robust observer design. By substituting (12) into (11), we get :
III. ROBUST OBSERVER DESIGN  
Ai 0
In this section, we consider the following fuzzy adaptive Fi = . (14)
−C −I p
descriptor observer to estimate state and SAF vectors :
r Then, equations (14) and (12) give as a result the matrix L:
E ż(t) = ∑ µi (ξ (t))(Fi z(t) + B̄i u(t)) + D̄a fˆa (t)     
i=1 0 −Ai 0 L1
=
x̄ˆ(t) = z(t) + Ly(t) (6) Ip C Ip L2
fˆs (t) = (DTs Ds )−1 DTs 0 I p x̄ˆ(t)
 
−Ai L1
 
=
CL1 + L2
f˙ˆ (t) = αK(ė (t) + e (t))
a y y

By identification, we get :
where z(t) is an auxiliary state vector and x̄ˆ(t) is the
estimated vector of (4). 
0

We denote ey (t) = C̄(x̄(t) − x̄ˆ(t)) and α is a symmetric L= (15)
Ip
positive definite learning rate matrix.
We have introduced the dynamic of the output vector to
Based on [5] and to satisfy condition (10), a solution is
obtain a fast estimation [11].
given as follows :
After that, we have :
 
I 0
z(t) = x̄ˆ(t) − LC̄x̄(t) = x̄ˆ(t) − LC0 x̄(t) − Lxs (t) (7) E= (16)
HC H

By substituting (7) into equation (6), the obtained equation where H ∈ R p is a full rank matrix which is non singular
is given by : and will be calculated.
Then, we have :
r
E x̄˙ˆ(t) − ELC̄x̄˙(t) = ∑ µi (ξ (t)) Fi (x̄ˆ(t)  
i=1 (8) −1 I 0
E = (17)
−LC0 x̄(t) − Lxs (t)) + B̄i u(t) + D̄a fˆa (t) −C H −1

978-1-5090-3960-9/17/$31.00 ©2017 IEEE 250

So, dynamic error (13) can be rewritten as follows :
r
Then, the adaptive observer (6) can realize ē(t) and e fa (t)
¯
ē˙(t) = ∑ µi (ξ (t))Si ē(t) + Ḡe fa (t) + Jd(t) (18) uniformly bounded, α is a symmetric positive definite
i=1 learning rate matrix and Z1 , Z2 and Z3 are a pre-specified
weight matrices.
such that :
  Proof. We take into account the Lyapunov function
Ai 0
Si = E −1 Fi = (19) below :
−CAi − H −1C −H −1
V (t) = ē(t)T Pē(t) + e fa (t)T α −1 e fa (t) such the following
inequality
 
Da
Ḡ = E −1 D̄a = (20) V̇ (t) + v(t)T v(t) − γ 2 d T (t)d(t) < 0 (27)
−CDa

  is satisfied.
Bd
J¯ = E −1 B̄d = (21)
−CBd V̇ (t) + v(t)T v(t) − γ 2 d T (t)d(t)
= sym(ē˙(t)T Pē(t)) + sym(ė fa (t)T α −1 e fa (t))
The dynamic error of actuator fault is considering as follows :
+ v(t)T v(t) − γ 2 d T (t)d(t)
ė fa (t) = f˙a (t) − f˙ˆa (t)
(22) Using equations (18) and (22), we have :
= f˙a (t) − αKC̄ ē˙(t) + ē(t)

V̇ (t) + v(t)T v(t) − γ 2 d T (t)d(t)

Define the next error v as : r
= ∑ µi (ξ (t))sym ē(t)T SiT Pē(t) − e fa (t)T KC̄Si ē(t)


 
ē(t)
v(t) = Z (23) i=1
e fa (t)
+ sym(e fa (t)T ḠT Pē(t)) + sym(d(t)T J¯T Pē(t))
where Z = diag(Z1 , Z2 , Z3 ) is a pre-specified weight matrix − sym(e fa (t)T KC̄Ḡe fa (t)) − sym(e fa (t)T KC̄Jd(t))
¯
with appropriate dimension. − sym(e fa (t)T KC̄ē(t)) + sym(e fa (t)T α −1 f˙a (t))
+ v(t)T v(t) − γ 2 d T (t)d(t)
To fulfil the stabilization condition of system (18), we
formulate the next theorem. Considering (25), we obtain :

Theorem 1. Consider system (3) with observer (6), V̇ (t) + v(t)T v(t) − γ 2 d T (t)d(t)
equation (18) is asymptotically stable and satisfies the r
= ∑ µi (ξ (t))sym ē(t)T SiT Pē(t) − e fa (t)T D̄Ta PSi ē(t)


H∞ performances, for a positive constant σ , if there exist
i=1
symmetric matrices P1 > 0, P3 > 0, R > 0, matrices W3 and
K such that the following conditions are satisfied : + sym(e fa (t)T ḠT Pē(t)) + sym(d(t)T J¯T Pē(t))
− sym(e fa (t)T D̄Ta PḠe fa (t)) − sym(e fa (t)T D̄Ta PJd(t))
¯
sym(ATi P1 ) + Z1T Z1

∗
− sym(e fa (t)T D̄Ta Pē(t)) + sym(e fa (t)T α −1 f˙a (t))
 −P3CAi −W3C sym(−W 3 ) + Z2T Z2
+ v(t)T v(t) − γ 2 d T (t)d(t)
 T
 −Da P1 Ai T
−Da CT P3
T
Bd P1 −BTd CT P3
 From Lemma 1, for R = RT > 0, we obtain the following
∗ ∗ (24)
inequality :
∗ ∗  <0
−sym(DTa P1 Da ) + R + Z3T Z3 ∗ 
T
e fa (t)T α −1 f˙a (t) + e fa (t)T α −1 f˙a (t) 6
BTd P1 Da −σ I
e fa (t)T Re fa (t) + δ
f or i = 1, . . . , r
where δ = fm2 λmax (α −1 R−1 α −1 ).

D̄Ta P = KC̄ (25) Taking into consideration equation (23), we get :

V̇ (t) + v(t)T v(t) − γ 2 d T (t)d(t) 6

where  T  
  r ē(t) ē(t) (28)
P=
P1 0
(26) ∑ µi (ξ (t))  e fa (t)  ∆i  e fa (t)  + δ
0 P3 i=1 d(t) d(t)

978-1-5090-3960-9/17/$31.00 ©2017 IEEE 251

 applying theorem 1, we minimize σ to 0.2817 and we get
the next feasible solutions :
with  
sym(SiT P) + diag(Z1T Z1 , Z2T Z2 ) 1 0 0 0 0


∆i =  −D̄Ta P + ḠT P − D̄Ta PSi 

 0 1 0 0 0 
J¯T P E = 0 0 1 0 0 
  0.2697 0.3294 0.2697 0.2697 0.3294
∗ ∗
−sym(D̄Ta PḠ) + R + Z3T Z3 ∗  −0.1314 0.2565 −0.1314 −0.1314 0.2565
−J¯T PD̄a −γ 2 I
 
−2 1 1 0 0
Replacing Si , Ḡ, J¯ and P by their parameters (19), (20), (21)  −1 −3 0 0 0 
and (26) in ∆i , we obtain LMI (24)  
F1 =  −2 1 −8 0 0
with W3 = P3 H −1 and σ = γ 2 .

 
 −1 0 −1 −1 0 
Applying Lemma 2, if (24) is hold, so there exist a positive
0 −1 0 0 −1

? scalar τ such that

V̇ (t) + v(t)T v(t) − γ 2 d T (t)d(t) < −τkθ (t)k2 + δ , where
ē(t)
θ (t) =  e fa (t) .
d(t)



−3
−1
2
−3
2
0.2 0
0 0
0



F2 =  0.5 2 −5 0 0 
It follows that V̇ (t) < 0 for τkθ (t)k2 > δ , which means
 
 −1 0 −1 −1 0 
that θ (t) converges to a small set Q = {θ (t)|kθ (t)k2 6 δτ } 0 −1 0 0 −1
according to Lyapunov stability theory. Therefore, estimation
errors of both state and faults are uniformly ultimately  
bounded. 0 0
End Proof 

 0 0 

L=
 0 0 

Remark 1. Comparing with [5], in this paper we have
 1 0 
focused on the problem of both SAF estimation. 0 1

The form of the sensor fault is given as follows :

Remark 2. In [3], the resolution of LMIs is realized
on two steps with setting some tuning parameters to 
fs1 (t)

obtain feasible LMIs. Whereas, in our work only one fs (t) =
fs2 (t)
step procedure is used. Moreover, a weighting matrix is
introduced to obtain more relaxed results. We suppose that:

 0, 0 ≤ t < 3.5
IV. N UMERICAL EXAMPLE

0.02(t − 3) + 0.2, 3.5 ≤ t < 6



In this part, we illustrate the validity of the mentioned fs1 (t) = 0, 6 ≤ t < 12
0.2, 12 ≤ t < 17

method. Consider the next theoretical T-S system in the 


0, t ≥ 17

form of (3) with parameters :
   
−2 1 1 −3 2 2 
0, 0≤t<6
A1 =  −1 −3 0  , A2 =  −1 −3 0.2  , 

0.1(−0.25t + 3.5) 6≤t<8


−2 1 −8 0.5 2 −5 

 0, 8 ≤ t < 10


−0.15, 10 ≤ t < 12
   
1 1   fs2 (t) =
1 0 1 0, 12 ≤ t < 14
B1 =  −1  , B2 =  −1  , C =

, 

0 1 0 
0.15, 14 ≤ t < 16
−2

0.5 


t ≥ 16
   
0,

  0.5 0.2
1 0
Ds = , Da =  0  , Bd =  0.5  .
0 1
0 0.4 
The two weighting functions are given below :  0, 0≤t<6
fa (t) = 0.2, 6 ≤ t < 8
0, t≥8

1 − tanh(ξ1 )(t)
µ1 (t) =
2 and input u(t) = 0.1 sin(3t).
1 + tanh(ξ1 )(t) Via applying Theorem 1, we own a feasible solution. De-
µ2 (t) = 1 − µ1 (t) =
2 scriptor observer (6), for α = 400, Z1 = I3 , Z2 = I2 and

978-1-5090-3960-9/17/$31.00 ©2017 IEEE 252

Z3 = 0.1 , allows the estimation of state and both SAF 0.07
x3
vectors. 0.06 x̂3

In Figs 1-3., the state variables and its estimation are shown. 0.05

These figures prove that the proposed observer is able to 0.04

estimate the states accurately, before and after the occurrence 0.03

of any fault. The trajectories of sensor faults fs1 (t), fs2 (t) and 0.02

their estimates are presented in Fig 4. and Fig 5. respectively. 0.01

Fig 6. presents the signal of actuator fault fa (t) and its 0

estimate. These figures show clearly that the mentioned −0.01

−0.02
observer can estimate both SAF successfully. Fig 7. shows
−0.03
the two weighting functions. Despite the presence of the 0 5 10 15 20 25
T ime(s)
white band disturbance, we get a very satisfied estimation
quality. Fig. 3. state signal x3 and its estimate

0.1
x1
x̂1
0.08

0.3
0.06 fs1
0.25 fˆs1
0.04

0.2

0.02
0.15

0
0.1

−0.02 0.05
0 5 10 15 20 25
T ime(s)
0

Fig. 1. state signal x1 and its estimate −0.05

0 5 10 15 20 25
T ime(s)

Fig. 4. sensor fault fs1 and its estimate

0.1 0.25
x2 fs2
x̂2
0.08 0.2 fˆs2
0.15
0.06

0.1
0.04
0.05
0.02
0
0
−0.05

−0.02
−0.1

−0.04 −0.15

−0.06 −0.2
0 5 10 15 20 25 0 5 10 15 20 25
T ime(s) T ime(s)

Fig. 2. state signal x2 and its estimate Fig. 5. sensor fault fs2 and its estimate

978-1-5090-3960-9/17/$31.00 ©2017 IEEE 253

 0.8 [7] D. Kharrat, H. Gassara, M. Chaabane, and A. El-Hajjaji. Fault tolerant
fa control based on adaptive observer for Takagi-Sugeno fuzzy descriptor
0.7 fˆa
systems. In 2015 16th International Conference on Sciences and
0.6 Techniques of Automatic Control and Computer Engineering (STA),
0.5
pages 273–278, Dec 2015.
[8] Y. Lu and Y. Song. Descriptor observer based approach for adaptive
0.4
reconstruction of measurement noises and sensor faults in uncertain
0.3 nonlinear systems. In The 27th Chinese Control and Decision
Conference (2015 CCDC), pages 2764–2769, May 2015.
0.2
[9] K. Tanaka, H. Ohtake, and H. O. Wang. A descriptor system approach
0.1 to fuzzy control system design via fuzzy Lyapunov functions. IEEE
Transactions on Fuzzy Systems, 15(3):333–341, June 2007.
0
[10] K. Tanaka and H. O. Wang. Fuzzy control systems design and analysis,
−0.1
0 5 10 15 20 25
A Linear Matrix Inequality Approach. Wiley Interscience, 2002.
T ime(s) [11] J. Zhang, A. Swain, and S. Nguang. Robust H∞ adaptive descriptor
observer design for fault estimation of uncertain nonlinear systems.
Franklin Institute, 351:5162–5181, 2014.
Fig. 6. actuator fault fa and its estimate

0.56
µ1
µ2
0.54

0.52

0.5

0.48

0.46

0.44
0 5 10 15 20 25
T ime(s)

Fig. 7. weighting functions µ1 and µ2

V. CONCLUSION
To conclude, a robust adaptive observer is designed for
a T-S fuzzy model affected both SAF and an UBD using a
descriptor approach. This technique leads to estimate state
and SAF vectors simultaneously. For the convergence of
the suggested method, sufficient conditions are mentioned
in terms of LMIs by using an appropriate Lyapunov matrix.
Finally, simulation results are presented to show the effec-
tiveness of the nominated method.

R EFERENCES
[1] M. Blanke, M. Kinnaert, J. Lunze, and M. Staroswiecki. Diagnosis
and Fault-Tolerant Control. Springer-Verlag Berlin Heidelberg, 2006.
[2] M. Bouattour, M. Chadli, A. El Hajjaji, and M. Chaabane. State and
faults estimation for T-S models and application to fault diagnosis.
IFAC Proceedings Volumes, 42(8):492 – 497, 2009.
[3] M. Bouattour, M. Chadli, A. El Hajjaji, and M. Chaabane. Estimation
of state, actuator and sensor faults for T-S models. In 49th IEEE
Conference on Decision and Control (CDC), pages 1613–1618, Dec
2010.
[4] S. Boyd, L. El Gaoui, E. Feron, and V. Balakrishnan. Linear Matrix
Inequalities in System and Control Theory. SIAM., 1994.
[5] Z. Gao, X. Shi, and S. X. Ding. Fuzzy state/disturbance observer
design for T-S fuzzy systems with application to sensor fault estima-
tion. IEEE Transactions on Systems, Man, and Cybernetics, Part B
(Cybernetics), 38(3):875–880, June 2008.
[6] J. Han, H. Zhang, Y. Wang, and X. Liu. Robust state/fault estimation
and fault tolerant control for T-S fuzzy systems with sensor and
actuator faults. Franklin Institute, 2016.

978-1-5090-3960-9/17/$31.00 ©2017 IEEE 254