Proceedings of the 2011 4th International Symposium on WeA2.

1
Advanced Control of Industrial Processes, Thousand
Islands Lake, Hangzhou, P.R. China, May 23-26, 2011

On observer design for nonlinear Takagi-Sugeno systems with

unmeasurable premise variable
Dalil Ichalal, Benoı̂t Marx, Didier Maquin, José Ragot

Abstract— In this paper, we propose a method for state a dynamic observer. The problem of regularity assumptions
estimation of nonlinear systems represented by Takagi- pointed out in [20] is solved by modifying the H∞ problem.
Sugeno (T-S) models with unmeasurable premise variables. Other classes of nonlinear systems are also studied in the
The main result is established using the differential mean
value theorem which provides a T-S representation of the literature to design observers for nonlinear systems, for
differential equation generating the state estimation error. Linear Parameter Varying systems (LPV) [3] and for bilinear
This allows to extend some results obtained in the case of systems.
measurable premise variables to the unmeasurable one. Using In [22], a new structure for nonlinear representation has
the Lyapunov theory, stability conditions are obtained and been introduced. It is based on the decomposition of the
expressed in term of linear matrix inequalities. Furthermore,
an extension for observer design with disturbance attenuation operating space of the system in several zones. To each zone
performance is proposed. Finally, this approach is illustrated is associated a linear model. Thanks to nonlinear functions
on a DC series motor and compared to the existing approaches. satisfying the convex sum property, the overall behavior of
the nonlinear system can be represented by the following
Index Terms— Takagi-Sugeno systems, state estimation, dif- so-called T-S model
ferential mean value theorem, Lyapunov stability analysis, ⎧
linear matrix inequality. ⎨ ẋ(t) =  r
μi (ξ(t))(Ai x(t) + Bi u(t))
i=1 (1)
⎩
I. I NTRODUCTION AND PROBLEM STATEMENT y(t) = Cx(t)
The control and diagnosis of physical systems often re- where x(t) ∈ Rn is the state vector, u(t) ∈ Rm is the input
quire the knowledge of internal variables of the system (state vector and y(t) ∈ Rp represents the output vector. Ai ∈
variables). These last are often not accessible to measurement Rn×n , Bi ∈ Rn×m , C ∈ Rp×n are known matrices. The
due to economic or technical reasons. In this situation, state functions μi (ξ(t)) are the weighting functions depending on
observers are used to provide an estimation of these variables the variables ξ(t) which can be measurable (as the input or
from input-output data and a mathematical model describing the output of the system) or unmeasurable variables (as the
the behavior of the system. Therefore, the estimation quality state of the system). These functions verify the following
necessarily depends on the precision of the model of nonlin- properties
ear behaviors, leading to complex nonlinear models. r

Early work on the state estimation of nonlinear systems μi (ξ(t)) = 1, 0 ≤ μi (ξ(t)) ≤ 1 ∀i ∈ {1, ..., r} (2)
dates back to the work of Thau [24] when he proposed i=1
an extension of the Luenberger observer [13] to systems Obtaining a T-S model can be performed from different
with Lipschitz nonlinearities. Sufficient conditions are then methods such as linearization of an available nonlinear model
obtained for the convergence of the state estimation error around some operating points and using adequate weighting
toward zero. Thereafter, in [19], an iterative approach is pro- functions. It can also be obtained by black-box approaches
posed for observer gain design. In [20], Rajamani obtained allowing to identify the parameters of the model from input-
necessary and sufficient conditions on the observer matrix output data [7]. Finally, an interesting approach to obtain a T-
that ensure asymptotic stability of the observer and proposed S model is the well-known nonlinear sector transformations
a design procedure, grounded on the use of a gradient based [23], [15]. Indeed, this transformation allows to obtain an
optimization method. In [20] is discussed the equivalence exact T-S representation of a general nonlinear model with
between the stability condition and the H∞ minimization in no information loss, in a compact state space.
the standard form, and pointed out that this design method Takagi-Sugeno model has proved its effectiveness in the
was not solvable since the regularity assumptions are not study of nonlinear systems. Indeed, it gives a simpler formu-
satisfied. Based on the result of Rajamani [20], [18] proposed lation from the mathematical point of view to represent the
behavior of nonlinear systems [22]. Thanks to the convex
D. Ichalal is with the IBISC laboratory, Universit d’Evry-Val-
d’Essonne, 40, rue du Pelvoux, Courcouronnes 91020 Evry Cedex, France
sum property of the weighing functions, it is possible to
[email protected] generalize some tools developed in the linear domain to the
B. Marx, J. Ragot and D. Maquin are with the Centre nonlinear systems. This representation is very interesting in
de Recherche en Automatique de Nancy (CRAN), Nancy- the sense that it simplifies the stability study of nonlinear
Université, 2, avenue de la forêt de Haye 54516 Vandoeuvre-
les-Nancy, France {benoit.marx, didier.maquin, systems and the controller/observer design. In [8], [10], [11],
jose.ragot}@ensem.inpl-nancy.fr the stability and stabilization tools inspired from the study of

353
linear systems are proposed. In [2], [14], the authors worked of DC series motor is proposed to illustrate the performances
on the problem of state estimation and diagnosis of T-S of the proposed methodology.
systems. The proposed approaches in these last papers rely
II. O BSERVER DESIGN
on the generalization of the classical observers (Luenberger
Observer [13] and Unknown Input Observer (UIO) [9]) to the Let us consider the T-S system with unmeasurable premise
nonlinear domain. Recently, some works are dedicated to the variable, given by
relaxation of the conservatism of the stability condition. For r

example, in [21], the Polya’s theorem is used in order to ẋ(t) = μi (x(t)) (Ai x(t) + Bi u(t)) (3)
reduce the conservatism related to the negativity of a sum i=1
matrices inequalities. In [12], the authors proposed a new y(t) = Cx(t) (4)
approach for discrete time T-S systems, it is based on the
The following observer is proposed
evaluation of the variation of the Lyapunov function between
r

two samples taken at times k and k + m with m > 1. ˙
x̂(t) = μi (x̂(t)) (Ai x̂(t) + Bi u(t) + L(y(t) − ŷ(t)))
In this work the considered premise variable ξ(t) depends
i=1
on the state of the system which is not totally measurable.
(5)
The problems of state estimation and diagnosis of nonlinear
systems using T-S model approach have been addressed with ŷ(t) = C x̂(t) (6)
different methods, but most of the published works have Let us remark that the comparison between the state x(t)
considered T-S models with measurable premise variables of the system and the state x̂(t) of the observer seems to be
[1], [17], [14], [2]. It is clear that the choice of measurable difficult. In order to cope with this difficulty, let us introduce
premise variables offers a good simplicity to generalize the the following matrices
methods already developed for linear systems. Contrarily, the
r r
problem becomes harder when the premise variables are not 1 1
A0 = Ai , B0 = Bi (7)
measurable. However, this formalism is very important in r i=1 r i=1
both the exact representation of nonlinear behaviors by T-S
Āi = Ai − A0 , B̄i = Bi − B0 (8)
model and in observer based diagnosis for sensor/actuator
fault detection and isolation. Indeed in this case, the use Then, it is easy to rewrite the system (3) in the following
of measurable premise variables requires to develop two form
different models. The first one uses the input u(t) as a
ẋ(t) = A0 x(t) + B0 u(t)
premise variable and allows to detect and isolate sensor
r
faults. The second one, using the output y(t) of the system as  
+ μi (x(t)) Āi x(t) + B̄i u(t) (9)
a premise variable, is dedicated to the detection and isolation
i=1
of actuator faults. Diagnosis based on a single T-S model
y(t) = Cx(t) (10)
with unmeasurable premise variables allows to detect and
isolate both actuator and sensor faults using observer banks. where it appears that the matrices A0 and B0 play the role of
Furthermore, the T-S models with unmeasurable premise nominal values of the system and Āi and B̄i are variations
variables may represent a larger class of nonlinear systems around these values.
compared to the T-S model with measurable premise vari- The state equation of the observer (5) can also be presented
ables [25]. Only a few works are devoted to the case of in the following form
unmeasurable premise variables, nevertheless, we can cite ˙
[6], [16], [5] where the authors proposed the fuzzy Thau- x̂(t) = A0 x̂(t) + B0 u(t) + L(y(t) − ŷ(t))
 r
Luenberger observer which is an extension of the classical  
+ μi (x̂) Āi x̂(t) + B̄i u(t) (11)
Luenberger observer and, in [25], a filter estimating the state
i=1
and minimizing the effect of disturbances is proposed.
ŷ(t) = C x̂(t) (12)
In this paper a new method is proposed for state estimation
of nonlinear systems. It is based on the use of the Takagi- which allows a simpler comparison with the state equation
Sugeno model representing the behavior of the nonlinear (9)-(10). For that purpose, let us define the state estimation
system. The contribution of this work concerns T-S model error e(t) by e(t) = x(t) − x̂(t). Using (9)-(10) and (11)-
with unmeasurable premise variables (e.g. the state of the (12), the dynamic of the state estimation error is obtained as
system), such a model is commonly encountered when using follows
the sector nonlinearity approach [23]. The main results on
observer design are given in sections II and III. The first ė(t) = (A0 − LC)e(t) + f (x(t), u(t)) − f (x̂(t), u(t)) (13)
result is devoted to the problem of state estimation and the where f (x(t), u(t)) is denoted f (z) with z = [xT (t) uT (t)]T
second one concerns the observer design with disturbance and defined by
attenuation by minimizing the L2 -gain from energy bounded r

unknown exogenous disturbances to the state estimation f (z) = μi (x(t))(Āi x(t) + B̄i u(t)) (14)
error. Finally, in section IV, an application for state estimation i=1

354
Note that the stability analysis of (13) cannot be directly where ãij1 = aij and ãij2 = bij and
achieved with the help of the tools developed for T-S systems ∂fi i
1 ∂xj (z ) − aij
with measurable premise variables. Indeed, the fact that the vij (z i ) = (24)
premise variable is the state of the system leads to a more bij − aij
complex form of the state estimation error (see equation ∂fi i
bij − ∂xj (z )
2
(13)). The key point of the proposed observer design is to vij (z i ) = (25)
obtain a suitable form of the state estimation error in order bij − aij
to re-use the tools proposed for stability and relaxed stability 2

l
analysis of T-S systems with measurable premise variables. vij (z i ) = 1, 0 ≤ vij
l
(z i ) ≤ 1, l = 1, 2 (26)
In conclusion, the objective is to find the gain L of the l=1

observer (11)-(12) that stabilize (13). Using (21) and (23), the dynamic of the state estimation error
The n different entries of the nonlinear vector function is represented by
f (z) : Rn+m → Rn are denoted fi (z). It follows n 
 n 
2
l

T ė(t) = vij (z i )(en (i)eTn (j)ãijl + A0 − LC)e(t)
f (z) = f1 (z) . . . fn (z) (15) i=1 j=1 l=1
where fi (z) : Rn+m → R, i = 1, ..., n + m. (27)
Let us denote es (i) the vector of Rs with all entries being Using the sector nonlinear transformation method proposed
null, except the ith being equal to 1 as given below in ([23], chap. 2), it follows
 T n 
 n 
2 q

es (i) = 01 · · · i−1 0 1 0 ··· 0 l
(16) vij (z i (t))en (i)eTn (j)ãijl = hi (z̃(t))Ai
i i+1 s
i=1 j=1 l=1 i=1
The function f (z) can be written as follows (28)
n
 where z̃(t) depends on the z i (t), Ai depends on the ãijl and
2
f (z) = en (i)fi (z) (17) where q = 2n . Then, the dynamic of the state estimation
i=1 error is written as
q

Theorem 1: Consider fi (z) : Rn+m → R. Let a, b ∈
n+m ė(t) = hi (z̃(t))Ψi e(t) (29)
R . If fi (z) is differentiable on [a, b] then there exists
i=1
a constant vector z i ∈ Rn+m , satisfying z i ∈]a, b[ (i.e.
with Ψi = Ai + A0 − LC.
zji ∈]aj , bj [, for j = 1, . . . , n + m), such that
The stability of this kind of models is largely studied in
∂fi (z i ) the literature. Hence, interesting results exist such as the
fi (a) − fi (b) = (a − b) (18)
∂z quadratic stability established by using a quadratic Lyapunov
Applying the theorem 1 on (17), it is obtained for a, b ∈ function candidate. Relaxed stability conditions are proposed
Rn+m by using the well-known fuzzy Lyapunov functions in the
 n n+m
 ∂fi (z i ) continuous time case and the Lyapunov function proposed
f (a) − f (b) = en (i)eTn+m (j) (a − b) (19) in [11] for the discrete time case.
i=1 j=1
∂zj
In this paper, the stability analysis of the system (29) is
Using (19), the dynamic of the state estimation error (13) studied in order to find the gain L. This analysis is performed
can be then transformed into by using the Lyapunov theorem and a quadratic Lyapunov
n n+m
   function, defined by
T ∂fi (z i ) x(t) x̂(t)
ė(t) = en (i)en+m (j) − V (e(t)) = e(t)T P e(t), P = P T > 0 (30)
∂zj u(t) u(t)
i=1 j=1
+ (A0 − LC)e(t) (20) Theorem 2: The state estimation error asymptotically con-
verges toward zero if there exist a symmetric positive definite
Since zj = xj , for 1 ≤ j ≤ n, then it follows matrix P ∈ Rn×n and a matrix M ∈ Rn×ny such that the
n  n  following linear matrix inequalities hold ∀i = 1, ..., q
∂fi (z i )
ė(t) = en (i)eTn (j) + A0 − LC e(t) AT0 P + P A0 + ATi P + P Ai − M C − C T M T < 0 (31)
i=1 j=1
∂xj
(21) The gain of the observer is derived from
Assumption 1: Assume that f (z i ) is a differentiable func-
L = P −1 M (32)
tion satisfying, for i = 1, ..., n and j = 1, ..., n
Proof: Considering Lyapunov function candidate (30)
∂fi i and definition (29), it is straightforward to obtain
aij ≤ (z ) ≤ bij (22)
∂xj q

∂fi i  
Each nonlinearity ∂xj (z ) can be represented by V̇ (e(t)) = hi (z̃(t))eT (t) ΨTi P + P Ψi e(t) < 0 (33)
i=1
 2
∂fi i l Then, using the definition of Ψi , with properties (2), the
(z ) = vij (z i )ãijl (23)
∂xj proof of the theorem is obvious (see [23] for more details).
l=1

355
 III. E XTENSION TO H∞ PERFORMANCES convex property of the weighting functions lead to the LMIs
The system (3) is now affected by a disturbance input ω(t) (41).
⎧ Remark 1: Note that the observer minimizing the L2
⎨ ẋ(t) = r
μi (x(t)) (Ai x(t) + Bi u(t) + Ei ω(t)) gain of the transfer from the disturbances ω(t) to the state
i=1 (34) estimation error e(t) is obtained by introducing the LMI
⎩
y(t) = Cx(t) + W ω(t) variable γ̄ = γ 2 and by minimizing γ̄ under the LMI
where Ei and W are the incidence matrices defining the constraints (41), where γ̄ is substituted to γ 2 in (41).
influences of ω(t) on the dynamics and the output of the IV. S IMULATION EXAMPLE : DC SERIES MOTOR
system. The disturbance vector is assumed to be energy
In this section, the proposed observer design is applied to
bounded, i.e. ω(t) ∈ L2 . As previously, the introduction of
a series DC motor in order to estimate its current I(t) and
the matrices A0 , B0 , Āi and B̄i , allows to write the state
its angular velocity ω(t). This kind of motors is generally
estimation error as
used in electrical traction due to their high torque and their
r
power autoregulation. The inductor and the armature of this
ė(t) = (A0 − LC)e(t) + μi (x(t))(Ei − LW )ω(t)
type of motor are connected in series, as shown in Figure 1,
i=1
r hence the term ”DC series motor”. The parameteres r and
  
+ (μi (x(t)) − μi (x̂(t))) Āi x(t) + B̄i u(t)
I(t)
i=1
(35) R, L
Cr (t) r
The application of the previously described method leads to
J, F U (t)
write the state estimation error as a T-S system l
q 
 r
ė(t) = hi (z(t))μj (x̂(t)) (Ψi e(t) + (Ej − LW )ω(t))
i=1 j=1
(36) Fig. 1. DC series motor
where
Ψi = A0 + Ai − LC (37) l respectively represent the resistance and the inductance of
the inductor (stator), while R and L respectively represent
Given the system (36), the problem of designing a robust the armature (rotor) resistance and inductance. The voltage
observer (5) is to find the gain L such that U of the motor must be between 0 and 1000 V and the
lim e(t) = 0 if ω(t) = 0 (38) current I is limited to 1000 A. The nonlinear model is given
t→∞
by the following equations
e(t)2 < γ ω(t)2 if ω(t) = 0 (39)
F L Cr (t)
where γ > 0 is a positive scalar representing the attenuation ẋ1 (t) = − x1 (t) + Km x2 (t)2 − (43)
J J J
level of the disturbance. To satisfy the constraints (38) and R L U (t)
(39), it is sufficient to find a Lyapunov function V (e(t)) such ẋ2 (t) = − x2 (t) + Km x1 (t)x2 (t) + (44)
Lt Lt Lt
that
the state vector is given by x(t) = [ω(t) I(t)]T where
V̇ (e(t)) + e(t)T e(t) − γ 2 ω(t)T ω(t) < 0 (40)
ω(t) represents the speed of the motor and I(t) the armature
The following theorem provides sufficient conditions un- current. We define Rt = R + r and Lt = L + l. The motor
der LMI formulation, for the synthesis of an observer robust is powered by a variable voltage which is given by U (t) =
t
to the disturbance ω(t). −70 exp(− 35 ) + 70. The resisting torque Cr (t) is usually
Theorem 3: Given γ > 0, the robust observer (5) for the unknown, but for this example, we consider it as a known
system (34) exists, if there exists a matrix P = P T > 0 ∈ input. Both inputs U and Cr are illustrated in Figure 2. The
Rn×n and a matrix M ∈ Rn×ny such that the following numerical values of the motor parameters are given by R =
LMIs hold 0.001485Ω, r = 0.00989Ω, L = 0.06H, Km = 0.04329,
 J = 30.1N/rad.s−1 , F = 0.1N/rad.s−1 . The inductance
Γi P Ej − M W
<0 (41) of the rotor is very large compared to the stator one, so we
EjT P − W T M T −γ 2 I
have
i = 1, . . . , q /j = 1, . . . , r L >> l ⇒ Lt = L = 0.06H (45)

where The method based on nonlinear sector transformation allows

to exactly transform the system (43)-(44) into the T-S model
Γi = AT0 P + P A0 + ATi P + P Ai − M C − C T M T + I (42) 2

The observer gain is given by the equation L = P M . −1 ẋ(t) = μi (x)(Ai x(t) + Bu(t)) (46)
Proof: The condition (40) expressed with the Lyapunov i=1

function V (e(t)) = eT (t)P e(t) (with P = P T > 0) and the defined by

356
 100
voltage A first simulation is performed without measurement noises
80 to show the convergence of the observer states to the real
U(Volts)
60 states, simulation results are depicted in the figure 3.
40
100
actual ω(t)
20 80 estimated ω(t)

0 60

ω(t)(rad.s )
−1
0 20 40 60 80 100 120 140 160 180 200
t(s) 40

20
350 resisting torque
0

300 −20
0 20 40 60 80 100 120 140 160 180 200
Cr(N.m)

250 t(s)

200
400 actual I
150 estimated I
300
100

I(Ampere)
0 20 40 60 80 100 120 140 160 180 200
t(s)
200

Fig. 2. Inputs of the DC series motor 100

0
0 20 40 60 80 100 120 140 160 180 200
  t(s)

−0.003 0.035
A1 =
−17.317 −0.412 Fig. 3. State of the motor and their estimation
 
−0.003 0 A second simulation is performed by introducing a mea-
A2 =
0 −0.412 surement noise bounded by 20. The goal of this second
  application is to show that even if the premise variable z(t) =
−0.033 0 x2 (t) can be measured, using the estimated state in the
B =
0 16.667
weighting functions gives better state estimation compared
The input vector is u(t) = [Cr (t) U (t)]T . The weighting to that obtained when using the noisy measurement y(t) as
functions are given by the following equations premise variable. In order to compare the two approaches,
⎧ two observers are designed. For the first observer design,
⎪
⎪ Km Lx2 (t)
⎨ μ1 (x2 (t)) = it is supposed that the weighting functions depend on the
1.039
(47) measured output y(t) = x2 (t). This observer is constructed
⎪
⎪ 1.039 − Km Lx2 (t)
⎩ μ2 (x2 (t)) = using the approach proposed in [17] and the estimation
1.039 results are displayed on the figure 4. For the design of the
Suppose that only the current I is measured, which gives the second one, it is considered that the weighting functions
output equation depend on the state variable x2 (t), then the weighting
y(t) = [0 1]x(t) (48) functions of the observer depend on x̂2 (t). The observer is
The state observer is constructed, applying the proposed synthesized by the proposed method using the differential
method using the differential mean value theorem, and the mean value theorem and nonlinear sector transformation and
obtained matrices Ai , i = 1, . . . , 8 are defined as follows the estimation results are displayed on the figure 5. In figure
4 (resp. 5) the estimated states obtained with the first (resp.
    second) observer are represented by the red continuous lines
0 −0.016 0 0.052
A1 = A2 = while the real growths are depicted by green dashed lines. As
−8.661 −3.637 −8.661 −3.637
    a conclusion, the observer using x̂2 (t) as a premise variable
0 −0.016 0 0.052 gives a better state estimation then the second observer using
A3 = A4 =
8.658 −3.637 8.658 −3.637 y(t) as a premise variable. In addition, using the approach
    developed in [4], the LMIs have no solution because of the
0 −0.016 0 0.052
A5 = A6 = high value of the considered Lipschitz constant.
−8.661 0.476 −8.661 0.476
    V. C ONCLUSION
0 −0.016 0 0.052
A7 = A8 =
8.658 0.476 8.658 0.476 In the present article is proposed a new method to de-
The pairs (A0 + Ai ) are observable, then the LMIs sign observers for nonlinear systems described by the well-
conditions in the theorem 2 result in the gain L of the known Takagi-Sugeno systems with unmeasurable premise
observer   variables. The method is based on the writing of the system
−0.3891 generating the state estimation error in the form of a T-S
L= (49)
51.3786 system. To do that, the differential mean value theorem and

357
 150
ω(t) with μi(y) [2] A. Akhenak, M. Chadli, J. Ragot, and D. Maquin. Fault detection and
real ω(t) isolation using sliding mode observer for uncertain Takagi-Sugeno
100
fuzzy model. In 16th Mediterranean Conference on Control and
ω(t)(rad.s−1)
Automation Congress Centre, Ajaccio, France, June 25-27 2008.
50 [3] G. I. Bara, J. Daafouz, F. Kratz, and J. Ragot. Parameter dependent
state observer design for affine LPV systems. International Journal
0 of Control, 74(16):1601–1611, November 2001.
[4] P. Bergsten and R. Palm. Thau-Luenberger observers for TS fuzzy
0 20 40 60 80 100 120 140 160 180 200 systems. In 9th IEEE International Conference on Fuzzy Systems,
t(s) FUZZ IEEE, San Antonio, TX, USA, 2000.
[5] P. Bergsten, R. Palm, and D. Driankov. Fuzzy observers. In IEEE
400 International Fuzzy Systems Conference, Melbourne, Australia, 2001.
I with μi(y)
real I
[6] P. Bergsten, R. Palm, and D. Driankov. Observers for Takagi-Sugeno
300 fuzzy systems. IEEE Transactions on Systems, Man, and Cybernetics
- Part B: Cybernetics, 32(1):114–121, 2002.
I(Ampere)

200 [7] A. Boukhris, G. Mourot, and J. Ragot. Non-linear dynamic system

identification: a multi-model approach. International Journal of
100 Control, 72:591–604, 1999.
[8] M. Chadli, D. Maquin, and J. Ragot. Non quadratic stability analysis
0 of Takagi-Sugeno systems. In IEEE Conference on Decision and
0 20 40 60 80 100 120 140 160 180 200
t(s)
Control, CDC’2002, Las Vegas, Nevada, USA, 2002.
[9] M. Darouach, M. Zasadzinski, and S.J. Xu. Full-order observers for
linear systems with unknown inputs. IEEE Transactions on Automatic
Fig. 4. Observer with measurable premise variable ξ(t) = y(t)
Control, 39(3):606–609, March 1994.
[10] T-M. Guerra, A. Kruszewski, L. Vermeiren, and H. Tirmant. Con-
150 ditions of output stabilization for nonlinear models in the Takagi-
ω(t) with μ (x)
i Sugeno’s form. Fuzzy Sets and Systems, 157(9):1248–1259, May 2006.
real ω(t)
100 [11] A. Kruszewski, R. Wang, and T.M. Guerra. Nonquadratic stabilization
ω(t)(rad.s−1)

conditions for a class of uncertain nonlinear discrete time TS fuzzy

50 models: A new approach. IEEE Transactions on Automatic Control,
53(2):606 – 611, March 2008.
0
[12] A. Kruzewski, R. Wang, and T.M. Guerra. Non-quadratic stabilization
conditions for a class of uncertain non linear discrete-time T-S fuzzy
models: a new approach. IEEE Transactions on Automatic Control,
0 20 40 60 80 100 120 140 160 180 200
t(s)
53(2):606–611, 2008.
[13] D.G. Luenberger. An introduction to observers. IEEE Transactions
on Automatic Control, 16:596–602, 1971.
400
I with μi(x) [14] B. Marx, D. Koenig, and J. Ragot. Design of observers for takagi
real I sugeno descriptor systems with unknown inputs and application to
300
fault diagnosis. IET Control Theory and Application, 1:1487–1495,
I(Ampere)

2007.
200
[15] A.M. Nagy, G. Mourot, B. Marx, G. Schutz, and J. Ragot. Model
structure simplification of a biological reactor. In 15th IFAC Sympo-
100
sium on System Identification, SYSID’09, Saint Malo, France, 2009.
[16] R. Palm and P. Bergsten. Sliding mode observers for Takagi-Sugeno
0
0 20 40 60 80 100 120 140 160 180 200 fuzzy systems. 9th IEEE International Conference on Fuzzy Systems,
t(s) FUZZ IEEE, San Antonio, TX, USA, 2000.
[17] R.J. Patton, J. Chen, and C.J. Lopez-Toribio. Fuzzy observers for non-
Fig. 5. Observer with estimated premise variable ξ(t) = x̂2 (t) linear dynamic systems fault diagnosis. In 37th IEEE Conference on
Decision and Control, Tampa, Florida USA, 1998.
[18] A.M. Pertew, H.J. Marquez, and Q. Zhao. H∞ observer design for
lipschitz nonlinear systems. IEEE Transactions on Automatic Control,
the nonlinear sector transformations are used. After, conver- 51:1211–1216, 2006.
gence conditions are obtained by using the Lyapunov theory [19] S. Raghavan and J. K. Hedrick. Observer design for a class of
nonlinear systems. International Journal of Control, 59(2):515–528,
and a quadratic Lyapunov function. An extension to robust 1994.
observer design with disturbance attenuation is proposed by [20] R. Rajamani. Observers for Lipschitz nonlinear systems. IEEE
minimizing the L2 gain of the transfer from disturbances Transactions on Automatic Control, 43:397–401, March 1998.
[21] Antonio Sala and Carlos Ario. Asymptotically necessary and sufficient
to the state estimation error. The stability conditions are conditions for stability and performance in fuzzy control: Applications
expressed in terms of Linear Matrix Inequalities. In order of polya’s theorem. Fuzzy Sets and Systems, 158(24):2671–2686,
to illustrate the proposed method, an example is devoted to December 2007.
[22] T. Takagi and M. Sugeno. Fuzzy identification of systems and its
the state estimation of a DC series motor. For future works, applications to modeling and control. IEEE Transactions on Systems,
we plan to address the problem of relaxed stability condition Man, and Cybernetics, 15:116–132, 1985.
for observer and controller design for T-S systems with [23] K. Tanaka and H.O. Wang. Fuzzy Control Systems Design and
Analysis: A Linear Matrix Inequality Approach. John Wiley and Sons,
unmeasurable premise variables by using Polya’s theorem 2001.
for example. [24] F.E. Thau. Observing the state of non-linear dynamic systems.
International Journal of Control, 18:471–479, 1973.
R EFERENCES [25] J. Yoneyama. H∞ filtering for fuzzy systems with immeasurable
premise variables: an uncertain system approach. Fuzzy Sets and
[1] A. Akhenak, M. Chadli, J. Ragot, and D. Maquin. Design of sliding Systems, 160(12):1738–1748, 2009.
mode unknown input observer for uncertain Takagi-Sugeno model. In
15th Mediterranean Conference on Control and Automation, MED’07,
Athens, Greece, 2007.

358