International Journal of Control, Automation and Systems 14(1) (2016) 323-330 ISSN:1598-6446 eISSN:2005-4092

http://dx.doi.org/10.1007/s12555-014-0406-8 http://www.springer.com/12555

A Sum of Squares Approach for Polynomial Fuzzy Observer Design for

Polynomial Fuzzy Systems with Unknown Inputs
Ali Chibani, Mohammed Chadli*, and Naceur Benhadj Braiek

Abstract: This paper investigates the problem of designing an unknown inputs observer to estimate states of a
class of nonlinear systems that can be modeled by polynomial fuzzy systems subject to unknown inputs. The
proposed method provides important innovations over the existing unknown inputs observers since the designed
observer has a polynomial fuzzy structure. Based on the Lyapunov approach, the proposed observer is developed
to guarantee the asymptotic stability of the estimation error and sufficient design conditions are given in Sum Of
Squares formulations with equality constraints. Both continuous-time and discrete-time cases are addressed and
relaxed conditions are also introduced by using intermediate variables that can be easily solved via numerical tools.
Finally, an illustrative example is given to provide the effectiveness of the proposed methodology.

Keywords: LMI, Lyapunov method, polynomial fuzzy observer (PFO), polynomial fuzzy system (PFS), unknown
inputs, SOS.

1. INTRODUCTION operator approach. More recently, several research papers

consider fuzzy observer design in finite frequency domain
The T-S fuzzy models, introduced by Takagi and Sugeno to reduce the conservatism generated by those designed in
[1] has received a great deal of attention over the last two the entire frequency domain [18–20].
decades in contol research activities. In this context, sev- These issues of research are motivated by the attractive
eral works in the literature considered stability analysis properties of fuzzy control including robustness to dis-
and controller synthesis of such systems(e.g., [2–9]). Other turbances, uncertainties, and sensors noise. Though T-S
studies addressed the problem of state estimating of T-S fuzzy models and LMI technic have enjoyed great suc-
fuzzy systems. Furthermore, the problem of estimating cess and popularity, there still exists a large number of
state vector of fuzzy systems driven by both known and design problems that either the processes cannot be ex-
unknown inputs has been the subject of a number of re- actly represented by a T-S model, either the results ob-
search studies in nearly past years. These problems are tained through LMIs are too conservative. Therefore, a
of considerable importance since they play an essential new type of fuzzy model had been developed. Indeed,
role in robust control based fault diagnosis. Thus, several in [21], a new structure of fuzzy systems, namely poly-
methods have been proposed in the literature that aims to nomial fuzzy systems (PFS), has been introduced where
design a state observer with unknown inputs for T-S fuzzy each local model is polynomial on state vector [21]. The
systems [10–15]. With the use of the Lyapunov method, polynomial method has been essentially motivated by the
several criterions based on Linear Matrix Inequality (LMI) introduction of the Sum Of Squares (SOS) approach and
formulations have been carried out to state estimation of numerical tools related to SOS programing which permit
T-S fuzzy models [11, 12]. For instance, in [12], sufficient to have polynomial terms in synthesis conditions. Indeed,
LMI conditions are presented to design an observer for T- SOS programming is considered as a powerful technique
S fuzzy models affected by unknown inputs which influ- in automatic and control theory since it provides convex
ence states and outputs simultaneously. On an other point relaxations to solve optimization problem [22, 24]. In ad-
of view, the delta operator approach has been recently ap- dition, SOSTOOLS is a free, third-party MATLAB tool-
plied to fault detection and fault tolerant control problems box that solves SOS problems [23] which requires a SDP
for T-S fuzzy models [16,17]. Hence, a fault detection ob- solver such as SeDuMi [25]. Thus stability analysis and
server for T-S fuzzy models has been designed via delta control synthesis of polynomial fuzzy systems are studied

Manuscript received October 12, 2014; revised January 22, 2015; accepted Febrary 12, 2015. Recommended by Associate Editor Do Wan
Kim under the direction of Editor Euntai Kim.

Ali Chibani and Naceur Benhadj Braiek are with the Advanced Systems Laboratory (Laboratoire des Systèmes Avancés-LSA)
Polytechnic School of Tunisia-EPT, University of Carthage BP. 743, 2078 La Marsa, Tunisia (e-mails: [email protected],
[email protected]). Mohammed Chadli is with University of Picardie Jules Verne, Laboratoire de Modélisation, Information et
Systèmes, MIS(E.A 4290), 33 rue Saint-Leu, 80039 Amiens, France (e-mail: [email protected]).
* Corresponding author.

⃝ICROS,
c KIEE and Springer 2016
324 Ali Chibani, Mohammed Chadli, and Naceur Benhadj Braiek

and effective methods based on the Sum Of Squares (SOS) definite positive (Resp.semi-definite positive) matrix, ()T
technic and Semi-Definite Programming (SDP) are pro- is used to represent the transposed matrix, (∗) denotes the
posed for continuous-time and discrete-time cases (e.g., transposed elements in the symmetric positions, and I is
[21, 26–29]). Hence PFS gives more relaxed criterions in the identity matrix;
stability analysis or control design since, the number of lo-
cal models is generally fewer than the existing T-S fuzzy 2.2. Preliminaries
systems and the SOS design conditions contain the exist- The following polynomial tools are useful for next de-
ing LMI approaches for T-S fuzzy model as special case velopments.
[21].
Definition 1 [22]: A multivariate polynomial p(x) =
On the other hand, the problem of observer design for
p(x1 , ..., xn ) is a sum of squares (SOS) if there exist poly-
such models are also addressed in a number of recent stud-
nomials f1 (x), ..., fm (x) such that
ies. Indeed, the work of [30] examined the case of PFS
such that the consequent local model is independent on m
unmeasurable states, i.e., Ai and Bi only dependent on y, p(x) = ∑ fi2 (x). (1)
and sufficient design conditions for a polynomial fuzzy i=1

observer are given in SOS-based SDP program. Note that,

The existence of an SOS decomposition (1) can be shown
all previous works concerned PFS driven only by known
equivalent to the existence of a positive semi-definite ma-
inputs [30, 31]. In fact, to the best of authors’ knowledge,
trix Q, such that
unknown input observer design for PFS is not provided in
last works. p(x) = Z T (x)QZ(x), (2)
In this context, this study investigates the problem of
state estimation of PFS subject to unknown inputs. This where Z(x) is the vector of monomials of degree less than
work considers a class of PFS in which the polynomial or equal to deg(p)/2. Since every p ∈ ∑n is a sum of
terms depend only on measurable variables and the result squared polynomials, it is clear that p(x) ≥ 0, ∀x ∈ Rn .
is given for both continuous and discrete time cases. However, the converse is not evident.
Thus, a new method to design a polynomial fuzzy ob-
server that reconstitute state vector of such systems is in- Proposition 1 [24]: Let M(x) be an nM × nM symmet-
troduced. Based on Lyapunov method to ensure observ- ric polynomial matrix in x ∈ Rn and v ∈ RnM . Then vT M(x)v
ing error stability, the proposed approach is formulated being an SOS implies M(x) ≥ 0 for x ∈ Rn .
in SOS terms with equality constraints. The given con-
ditions are sufficient and all constraints are presented in 3. POLYNOMIAL FUZZY SYSTEMS
linear form that can be easily solved via numerical tools.
Hence, the proposed method provide important innova- A polynomial fuzzy system is described by if-then rules
tion compared to those related to T-S fuzzy models and [21]. By considering an unknown inputs vector, the ith
the proposed design conditions contain the existing LMI rule of the PFS is generally described by the following
conditions for T-S fuzzy systems as a special case. equation:
The structure of the paper is as follows: next section
provides useful notation and preliminaries. Section III in- i f ζ1 (t) is Mi1 and · · · and ζ p (t) is Mip then
troduces the polynomial fuzzy models. The proposed ob- {
server and the main results are presented in section IV. An δ (x(t)) = Ai (x(t))x(t) + Bi (x(t))u(t) + Ri (x(t))d(t)
illustrative example is introduced in section V and a con- y(t) = Cx(t) i = 1, 2, · · · , r
clusion section closes the paper. (3)

2. PRELIMINARIES AND NOTATIONS where x ∈ Rn is the state vector, u ∈ Rm is the known input
vector, d ∈ Rqd is the unknown input vector, Ai (x) ∈ Rn×n
To lighten the redaction, we introduce, in this section, and Bi (x) ∈ Rn×m are polynomial matrices on x that define
the notation that will be used in the rest of the paper. Fur- the ith local model, C ∈ R p×n is the output matrix, Ri (x) ∈
thermore, we provide a brief review of the SOS decompo- Rn×qd represent the influence matrix of the unknown input
sition that will be useful in the following sections. , ζ j (t)( j = 1, ..., p) are the premise variables assumed to
depend on measurable variables, Mi j (t) is the membership
2.1. Notations function, r denotes the number of model rules and δ (x(t))
Throughout this paper, the notation is rather standard: is employed for simplicity and it is equivalent to ẋ(t) for
Let Rn denotes the set of n-dimensional real vector, ∑n the continuous time case, and x(t + 1) for the discrete time
denotes the subset from Rn that contains all SOS polyno- case.
mial in n variables, P > 0 (Resp. P ≥ 0) means that P is The overall polynomial fuzzy system is given by:
 A Sum of Squares Approach for Polynomial Fuzzy Observer Design for Polynomial Fuzzy Systems with Unknown ... 325

 ( )
 r Ai (x(t))x(t) + Bi (x(t))u(t) and the overall PFO is governed by:
δ (x(t)) = ∑ µi (ζ (t))
+Ri (x(t))d(t)  ( )
 i=1
y(t) = Cx(t),  r Ni (y(t))(z(t)) + Gi (y(t))u(t)
δ (z(t)) = ∑ µi (ζ (t))
(4) +Li (y(t))y(t)
where µi (ζ (t)) are activation functions that verify:  i=1
x̂(t) = z(t) − Ey(t)
r (8)
µi (ζ (t)) ≥ 0, ∑ µi (ζ (t)) = 1. (5)
i=1 in which the same activation functions µi (ζ (t)) as used
Assumption 1: From now, we suppose that the conse- for the PFS (6) are considered and where x̂ ∈ Rn is the es-
quent parts in the PFS depend only on measurable vari- timated vector and Ni (y) ∈ Rn×n , Gi (y) ∈ Rn×m , Li (y) ∈
ables, i.e., Ai (x) = Ai (y), Bi (x) = Bi (y) and Ri (x) = Ri (y). Rn×p and E ∈ Rn×p are the observer gains to be deter-
Under the assumption, we replace the PFS (4) by: mined in order to estimate the state vector of the PFS (6)
such that the estimation error:
 ( )
 r Ai (y(t))x(t) + Bi (y(t))u(t)
δ (x(t)) = ∑ µi (ζ (t)) e(t) = x(t) − x̂(t) (9)
+Ri (y(t))d(t)
 i=1
y(t) = Cx(t) converges towards zero when t → ∞. Note that the esti-
(6) mation error can be rewritten as:
It can be seen that all elements in Ai (y), Bi (y) and Ri (y) e(t) = T x(t) − z(t) (10)
matrices are measurable. This hypothesis is needed later
when designing the observer. Indeed, since the designed with
observer will depend only on the known input u and the
measured state vector y and particulary observer gains are T = I + EC. (11)
polynomial on y, we will see later that equality conditions The dynamic of the observing error, is governed by the
are combined from system matrices and observer gains following equation:
and so they should depend on the same variables.
 
Remark 1: The structures above (4) and (6) with poly- Ni (y)e(t) + T Ri (y)d(t))
(
nomial consequent part, can be obtained by using the sec- r  TAi (y) − Ni (y)T 
δ (e(t)) = ∑ µi (ζ (t)) 
+ x(t) 
.
tor nonlinearity concept and the polytopic transformations i=1
−Li (y)C
for nonlinear functions which appear in the nonlinear sys- +(T Bi (y) − Gi (y))u
tem [21]. In several cases, with fuzzy polynomial struc- (12)
ture (4) we can exactly represent the original nonlinear
system and in general, the number of rules becomes fewer Consequently, if the following constraints hold:
than that of T-S fuzzy model. Moreover, if Ai (x), Bi (x)
and Ri (x) are constant matrices for all i then, (2) reduces TAi (y) − Ni (y)T − Li (y)C = 0 (13)
to classic T-S fuzzy model with unknown input, so the T Bi (y) − Gi (y) = 0 (14)
polynomial structure is a more general representation. T Ri (y) = 0 (15)

4. POLYNOMIAL OBSERVER DESIGN WITH and

UNKNOWN INPUT r
δ (e(t)) = ∑ µi (ζ (t))Ni (y)e(t) (16)
In this section a new method to design an unknown ob- i=1
server will be presented. Indeed, we are concerned by the is a stable system, then we guarantee that the estimated
estimation of the state variable x(t) of unknown inputs states converges to the real ones. Hence, the problem of
PFS (6) using measurable signals, i.e. known input u(t) designing a PFO for the PFS subject to unknown input
and measured output y(t). For this reason, the proposed is reduced to determine the observer gains Ni (y), Gi (y),
observer has a polynomial fuzzy structure in the conse- Li (y) and E such that the error system (16) is stable and
quent parts of the if-then rules. (13)-(15) hold.
The ith rule is described by the following form:
4.1. Synthesis conditions for continuous-time case
i f ζ1 (t) is Mi1 and · · · and ζ p (t) is Mip then
Now we consider the continuous-time case. The error
{ system is given by:
δ (z(t)) = Ni (y(t))(z(t)) + Gi (y(t))u(t) + Li (y(t))y(t)
x̂(t) = z(t) − Ey(t) i = 1, 2, · · · , r r
ė(t) = ∑ µi (ζ (t))Ni (y(t))e(t). (17)
(7) i=1
326 Ali Chibani, Mohammed Chadli, and Naceur Benhadj Braiek

The following theorem, gives sufficient conditions in SOS Using the following variable change:
formulation that permit to design a polynomial observer
in order to estimate states of the unknown inputs system S = PE (33)
(6).
and
Theorem 1: If there exist symmetric matrix P, con-
stant matrix S and polynomial matrices Wi (y) and Ji (y) Wi (y) = PKi (y) (34)
such that the following equations hold ∀i = 1...r:
inequality (32) is equivalent to:
vT1 (P − ε1 In )v1 SOS, (18)
(
−vT2i ((P + SC) Ai (y) −Wi (y)C)T + (P + SC) Ai (y) ((P + SC) Ai (y) −Wi (y)C)T + (P + SC) Ai (y)
(19) (35)
−Wi (y)C + ε2i (y)In ) v2i SOS, −Wi (y)C < 0.
Ji (y) − (P + SC)Bi (y) = 0, (20)
To obtain more relaxed constraints with polynomial struc-
(P + SC)Ri (y) = 0, (21) tures, the sum of squares decomposition is exploited. There-
fore, by using Proposition 1, if (18) hold with ε1 > 0, then
where v1 and v2i are vectors with appropriate dimensions V > 0. Furthermore, if (19) hold with ε2i (y) > 0 for y ̸= 0,
that are independent of y, ε1 > 0 and ε2i (y) > 0 for y ̸= 0. then V̇ < 0 and then the state estimation error is stable.
Then, the estimation error (17) is globally asymptotically Similarly, we use the following variable change:
stable and the estimated states of the PFO converge glob-
ally asymptotically to the states of the PFS. The observer Ji (y) = PGi (y) (36)
gains are given by: to obtain equality (20) which ends the proof. □

E = P−1 S, (22) Remark 2: When Ai (y), Bi (y), Ri (y), Ni (y), Gi (y) and
Li (y) are constant matrices, the polynomial fuzzy system
Gi (y) = P−1 Ji (y), (23)
representation (6) reduces to T-S fuzzy system subject to
−1
Ni (y) = (I + EC)Ai (y) − P Wi (y)C, (24) unknown input and the polynomial fuzzy observer (7) is
the same as the T-S fuzzy observer with unknown inputs
Li (y) = P−1Wi (y) − Ni (y)E. (25)
used in a large body of existing work, (e.g., [12]). Thus,
Proof: To establish the asymptotic stability of the esti- the proposed SOS approach to PFO design for PFS subject
mation error, let us consider the quadratic Lyapunov func- to unknown inputs contains the existing LMI approaches
tion: as a special case. Then, the proposed approach is more
general than the existing LMI conditions.
V (e) = eT Pe, (P > 0). (26) Now we propose to give more relaxed conditions to de-
sign an unknown inputs polynomial observer. Hence, a
Then, the time derivative of V is given by:
transformation based on the introduction of a slack vari-
r ( ) able is used in order to separate the term PNi (y) in equal-
V̇ (e) = ∑ µi eT NiT (y)P + PNi (y) e. (27) ity (28) [32].
i=1
The following result gives more relaxed SOS conditions
Asymptotic stability is guaranteed if: with equality constraints that aims to design a polyno-
mial observer for polynomial fuzzy systems subject to un-
NiT (y)P + PNi (y) < 0. (28) known inputs.
From equations (11) and (13), we get: Theorem 2: If there exist symmetric matrix P, con-
stant matrices S and U, polynomial matrices Wi (y) and
Ni (y) = TAi (y) − Ki (y)C, (29) Ji (y) such that the following equations hold ∀i = 1...r:
where − vTi (Mi + εi (y)InM ) vi SOS (37)
Ki (y) = Li (y) + Ni (y)E. (30) Ji (y) − (U + SC)Bi (y) = 0 (38)
(U + SC)Ri (y) = 0 (39)
Therefore, inequality (28) can be rewritten as:
with
(TAi (y) − Ki (y)C)T P + P(TAi (y) − Ki (y)C) < 0. (31)
 
−U −U T (U + SC)Ai (y) −Wi (y)C + P U
Thus, by exploiting (11), (31) becomes: 
Mi = (∗) −P 0 
((P + PEC)Ai (y) − PKi (y)C)T + (P + PEC) Ai (y) (∗) (∗) −P
(32)
−PKi (y)C < 0. (40)
 A Sum of Squares Approach for Polynomial Fuzzy Observer Design for Polynomial Fuzzy Systems with Unknown ... 327

where εi (y) > 0 for y ̸= 0 and vi are vectors with appro- Theorem 3: If there exist symmetric matrix P, con-
priate dimensions that are independent of y. Then the esti- stant matrix S and polynomial matrices Wi (y) and Ji (y)
mation error (17) is globally asymptotically stable and the such that the following equations hold ∀i = 1...r:
estimated states of the PFO converge globally asymptoti-
cally to the states of the PFS. The observer gains are given vTi (Mi − εi (y)InM ) vi SOS (50)
by: Ji (y) − (P + SC)Bi (y) = 0 (51)
E = U −1 S, (41) (P + SC)Ri (y) = 0 (52)
Gi (y) = U −1 Ji (y), (42)
with
Ni (y) = (I + EC)Ai (y) −U −1Wi (y)C, (43) [ ]
−1 P (∗)
Li (y) = U Wi (y) − Ni (y)E. (44) Mi =
(P + SC)Ai (y) −Wi (y)C P
, (53)

Proof: To ensure the asymptotic stability of the estima- where εi (y) > 0 for y ̸= 0. vi are independent vectors of
tion error, condition (28) is equivalent to: appropriate dimensions. Then, the estimation error (49)
  is globally asymptotically stable and the estimated states
−U −U T UNi (y) + P U converge to the states of the PFS.
 (∗) −P 0  < 0. (45) The observer gains are given by equations (22)-(25).
(∗) (∗) −P
Proof: To prove the asymptotic convergence of the er-
For more detail about the above relaxed condition, the ror system (49), sufficient conditions are derived using the
reader can see for example [12, 32]. quadratic Lyapunov function V (e(t)) = e(t)T Pe(t). The
If we consider the following variable change: variation ∆V = V (t + 1) −V (t) is given by:
r ( )
S = UE (46) ∆V = ∑ µi eT NiT (y)PNi (y)e − eT Pe . (54)
i=1
and
The system (54) is globally stable if:
Wi (y) = UKi (y) (47) NiT (y)PNi (y) − P < 0. (55)

and by using the SOS properties and proposition 1, LMI From equations (11) and (13), we get:
condition (45) is equivalent to the SOS one (37). Further-
more, with the following variable change: Ni (y) = TAi (y) − Ki (y)C (56)

with
Ji (y) = UGi (y) (48)
Ki (y) = Li (y) + Ni (y)E. (57)
we get conditions (38) which ends the proof. □
And hence, inequality (55) becomes:
Remark 3: It is important to have in mind that the pro- (T Ai (y) − Ki (y)C)T P (T Ai (y) − Ki (y)C) − P < 0, (58)
posed formulation in (37)-(39) leads to more of relaxation
by separating the Lyapunov terms PNi (y) and introduc- which is equivalent to:
ing the non symmetric slack variable U in the equality
constraints instead of Lyapunov matrix P which introduce ((P + PEC) Ai (y) − PKi (y)C)T P−1 ((P + PEC) Ai (y)
more degree of freedom in synthesis problem (for more −PKi (y)C) − P < 0.
details we refer the reader to [32]). (59)

4.2. Synthesis conditions for discrete-time case Now, if we consider the following variable change:
For the discrete-time systems, the error system is gov- S = PE (60)
erned by:
and
r
e(t + 1) = ∑ µi (ζ (t))Ni (y(t))e(t). (49) Wi (y) = PKi (y) (61)
i=1
and using the Schur complement, we have:
The following theorem gives sufficient conditions to syn-
[ ]
thesize an unknown input observer for discrete-time fuzzy P (∗)
M= > 0. (62)
systems. ((P + SC)Ai (y) −Wi (y)C)T P
 328 Ali Chibani, Mohammed Chadli, and Naceur Benhadj Braiek

If, the SOS condition (50) hold with εi (y) > 0 for y ̸= 0, 5. SIMULATION EXAMPLE
then, (62) hold and then we guarantee that V > 0 and
∆V < 0. Moreover, by considering the following vari- To illustrate the validity of the proposed method to de-
able’s change: sign an unknown inputs polynomial fuzzy observer, the
following example is considered:
Ji (y) = PGi (y) (63) 
 ẋ1 (t) = cos(x1 (t))x2 (t) + x13 (t)u(t) + d(t)
if condition (51) holds, then, we ensure that (14) is sat- ẋ2 (t) = x1 (t) − x12 (t)x2 (t) (75)

isfied. Finally by satisfying (52), we guarantee that (15) y(t) = x1 (t).
holds which ends the proof. □
The above system can be transformed in a polynomial
Note that, by introducing an intermediate variable, we lead fuzzy structure as follows:
to a more relaxed conditions as the continuous time case. 
 2
Thus, the following theorem gives sufficient conditions ẋ(t) = ∑ µi (ζ (t))(Ai (y)x(t) + Bi (y)u(t) + Ri d(t))
design that offers more degree of freedom in observer syn-  i=1

thesis by using a slack variable U. y(t) = Cx(t),

(76)
Theorem 4: If there exist symmetric matrix P, con-
stant matrices S and U, polynomial matrices Wi (y) and where : ζ (t) = cos x1 (t) ∈ [−1, 1], ∀ x1 (t) ∈( R; µ1 (ζ (t)))=
Ji (y) such that the following equations hold ∀i = 1...r: 1+cos x1 (t) 0 1
; µ2 (ζ (t)) = 1−cos2 x1 (t) ; A1 (y) = ;
2 1 −y2
− vTi (Mi + εi (y)InM ) vi SOS, (64) ( ) ( 3 )
0 −1 y
Ji (y) − (U + SC)Bi (y) = 0, (65) A2 (y) =
1 −y2
; B1 (y) = B2 (y) =
0
; R1 =
( )
(U + SC)Ri (y) = 0, (66) 1
R2 = .
1
where
( ) Note that the considered system can be exactly represented
−P −U + ((U + SC)Ai (y) −Wi (y)C)T
Mi = < 0 (67) by a T-S model with 8 local linear models for x1 (t) ∈
(∗) P −U −U T
[α , β ] with α < β . Hence, one can see that polynomial
and εi (y) > 0 for y ̸= 0. vi are vectors with appropriate fuzzy representation (76) is valid in a wider operating do-
dimensions that are independent of y. Then, the estima- main than the T-S fuzzy one. Furthermore, the number of
tion error (49) is globally asymptotically stable and the local linear models in the T-S fuzzy case was greatly in-
estimated states converge to the states of the PFS. The ob- creased which directly affects the complexity of the syn-
server gains are given by: thesis problem. The SOS conditions described by Theo-
rem 2 are feasible and lead to the following solution:
E = U −1 S, (68) ( ) ( )
Gi (y) = U −1
Ji (y), (69) 2.4367 −0.088984 −0.23728
P= ;S= ;
−0.088984 1.2423 −1.1389
−1
Ni (y) = (I + EC)Ai (y) −U Wi (y)C, (70) ( ) ( )
0.89315y2 + 1.9994 0.51519y3
−1 W1 = ; J1 = ;
Li (y) = U Wi (y) − Ni (y)E. (71) 2
0.46588y + 0.91026 −0.7705y3
( ) ( )
0.89315y2 + 2.5145 0.51519y3
Proof: Identically to the previous section, we use the W2 = ; J2 =
2
0.46588y + 0.14365 −0.7705y3
quadratic Lyapunov function V (e(t)) = e(t)T Pe(t) (P = ( )
0.75247 −0.51519
PT > 0) to ensure the asymptotic convergence of the error and V = .
0.36839 0.7705
system (49). Then, the error system is stable if:
Consequently, the observer parameters are given by:
NiT (y)PNi (y) − P < 0. (72)
( )
−0.67121y2 − 1.5661 1.3076
By introducing the intermediate matrix U, (72) becomes: N1 (y) = ;
( ) −1.0535y2 − 1.2286 −y2 − 1.2724
( )
−P −U + (UNi )T −0.67121y2 − 2.4488 −1.3076
< 0, (73) N2 (y) = ;
(∗) P −U −U T −1.0535y2 − 0.82382 −y2 + 1.2724
( ) ( )
0.88465y3 0.87769y2 + 3.7117
which is equivalent to: G1 = ; L1 = ;
−0.40848y3 0.10508y2 + 0.98741
( ) ( ) ( )
−P −U + ((U + SC)Ai (y) −WiC)T 0.88465y3 0.87769y2 + 1.5382
< 0. (74) G2 = ; L2 = ;
(∗) P −U −U T −0.40848y 3
0.10508y2 + 3.6963
( )
0.30762
Hence, if condition (64) hold, then (74) hold and then we and E = .
−1.2724
guarantee that P > 0 (V > 0) and ∆V < 0. □
 A Sum of Squares Approach for Polynomial Fuzzy Observer Design for Polynomial Fuzzy Systems with Unknown ... 329

[2] A. Sala and C. Arino, “Asymptotically necessary and suffi-

cient conditions for stability and performance in fuzzy con-
trol: Applications of Polya’s theorem,” Fuzzy Sets and Sys-
tems, vol. 158, pp. 2671-2686, 2007. [click]
[3] D. Saifia, M. Chadli, S. Labiod, and T. M. Guerra, “Ro-
bust H∞ static output feedback stabilization of T-S fuzzy
systems subject to actuator saturation," International Jour-
nal of Control, Automation and Systems, vol. 10, no. 3, pp.
613-622, 2012. [click]
[4] H. Yang, P. Shi , J. Zhang, and J. Qiu, “Robust H∞ control
for a class of discrete time fuzzy systems via delta opera-
tor approach,” Information Sciences, vol. 184, pp. 230-245,
Fig. 1. real x1 and estimated x1 2012. [click]
[5] M. Chadli, H. R. Karimi, and P. Shi, “On stability and sta-
bilization of singular uncertain T-S fuzzy systems,” Journal
of Franklin Institute, vol. 351, no. 3, pp. 1453-1463, 2014.
[click]
[6] H. Yang, X. Li, Z. Liu, and L. Zhao, “Robust fuzzy-
scheduling control for nonlinear systems subject to actu-
ator saturation via delta operator approach,” Information
Sciences, vol. 272, pp. 158-172, 2014. [click]
[7] F. Li, P. Shi, L. Wu, and X. Zhang, “Fuzzy-model-dased D-
Stability and non-fragile control for discrete-time descrip-
tor systems with multiple selays,” IEEE Trans. on Fuzzy
Systems, vol. 22, no. 4, pp. 1019-1025, 2014. [click]

Fig. 2. real x2 and estimated x2 [8] S. F. Derakhshan and A. Fatehi, “Non-monotonic Lya-
punov function for stability analysis and stabilization of
discrete time Takagi-Sugeno fuzzy systems,” International
Journal of Innovative Computing, Information and Con-
Figs. 1 and 2 show the asymptotic convergence of the trol, vol. 10, no. 4, pp. 1567-1586, 2014. [click]
estimated states of the PFO to the states of the PFS in [9] E. T. Jeung and K. R. Lee, “Static output feedback control
which we consider x(0) = [0.5 1]T , x̂(0) = [0 0.5]T , for continuous-time T-S fuzzy systems: an LMI approach,"
u(t) = te(−2t) and d(t) = sin(2t)e(−t) . Note that the SOS International Journal of Control, Automation and Systems,
design conditions of Theorem 1 fail to find a feasible solu- vol. 12, no. 3, pp. 703-708, 2014. [click]
tion whereas a solution is found using the SOS conditions [10] S. K. Nguang, P. Shi, and S. Ding, “Fault detection for un-
of Theorem 2. Indeed, the result of Theorem 2 introduces certain fuzzy systems: an LMI approach,” IEEE Trans. on
more relaxation compared to the one of Theorem 1. Fuzzy Systems, vol. 15, no. 6, pp. 1251-1262, 2007. [click]
[11] M. Chadli, A. Akhenak, J. Ragot, and D. Maquin, “State
and unknown input estimation for discrete time multiple
6. CONCLUSIONS
model,” Journal of the Franklin Institute, vol. 346, no. 6,
pp. 593-610, 2009. [click]
This paper proposes a method to design a polynomial
observer for a class of nonlinear system that can be repre- [12] M. Chadli and H. R. Karimi, “Robust observer design for
unknown inputs Takagi-Sugeno models,” IEEE Trans. on
sented by unknown inputs polynomial fuzzy models. The
Fuzzy Systems, vol. 21, no. 1, pp. 158-164, 2013. [click]
convergence of the proposed observer is proved by us-
[13] S. Aouaouda, M. Chadli, V. Cocquempot, and M. T.
ing Lyapunov method and sufficient conditions design are
Khadir, “Multi-objective H− /H∞ faults detection observer
then given in a SOS formulation that can be easily solved
design for Takagi-Sugeno fuzzy systems with unmeasur-
via the existing SOSTOOLS. In this work, both continu- able premise variables: descriptor approach,” Inter. Jour-
ous time and discrete time models are addressed and more nal of Adaptive Control and Signal Processing, vol. 27, no.
of relaxed results are also presented. An illustrative exam- 12, pp. 1031-1047, 2013. [click]
ple is introduced to validate the proposed methodology. [14] T. Youssef, M. Chadli, H. R. Karimi, and M. Zelmat, “De-
sign of unknown inputs proportional integral observers for
REFERENCES TS fuzzy models,” Neurocomputing, vol. 123, no. 10, pp.
156-165, 2014. [click]
[1] T. Takagi and M. Sugeno, “Fuzzy Identification of Systems [15] S. P. He, “Fault estimation for T-S fuzzy Markovian jump-
and Its Applications to Modeling and Contro,” IEEE Trans. ing systems based on the adaptive observer,” International
on Systems, Man and Cybernetics, vol. 15, no. 1, pp. 116- Journal of Control, Automation and Systems, vol. 12, no.
132, 1985. [click] 5, pp. 977-985, 2014. [click]
330 Ali Chibani, Mohammed Chadli, and Naceur Benhadj Braiek

[16] H. Yang, X. Li, Z. Liu, and C. Hua, “Fault detection for [30] K. Tanaka, H. Ohtake, M. Wada, H. O. Wang, and Y.
uncertain fuzzy systems based on the delta operator ap- J. Chen, “Polynomial fuzzy observer design: a sum of
proach,” Circuits, Systems and Signal Processing, vol. 33, squares approach,” Proc. of 48th IEEE Conference on De-
no. 3, pp. 733-759, 2014. [click] cision and Control, pp. 7771-7776, Shanghai, China, De-
[17] H. Yang, P. Shi, X. Li, and Z. Li, “Fault-tolerant control for cember 16-18, 2009. [click]
a class of T-S fuzzy systems via delta operator approach,” [31] K. Tanaka, H. Ohtake, T. Seo, M. Tanaka, and H. O. Wang,
Signal Processing, vol. 98, pp. 166-173, 2014. [click] “Polynomial fuzzy observer designs:a sum-of-squares ap-
[18] T. Iwasaki and S. Hara, “Generalized KYP lemma: Unified proach,” IEEE Trans. On Systems, Man, And Cybernetics,
frequency domain inequalities with design applications,” Part B, vol. 42, no. 5, October 2012. [click]
IEEE Transa. on Automatic Control, vol. 50, no 1, pp. 41- [32] P. Apkarian, H. D. Tuan, and J. Bernussou, “Continuous-
59, 2005. [click] time analysis, eigenstructure assignment, and synthesis
[19] K. Zhang, B. Jiang, P. Shi, and J. Xu, “Analysis and de- with enhanced linear matrix inequalities characterizations,”
sign of robust H∞ fault estimation observer with finite- IEEE Trans. on Automatic Control, vol. 46, no. 12, pp.
frequency specifications for discrete-time fuzzy systems,” 1941-1946, 2001. [click]
IEEE Trans. on Cybernetics, vol. 45, no. 7, pp 1225-1235,
2014. [click] Ali Chibani was born in Gafsa, Tunisia in
1988. He received his B.S and M.S. de-
[20] X. J. Li and G. H. Yang, “Fault detection in finite fre-
grees in electrical engineering, in 2010 and
quency domain for Takagi-Sugeno fuzzy systems with sen-
2012, respectively from the High School
sor faults,” IEEE Trans. on Cybernetics, vol. 44, vol 8, pp.
of Sciences and Techniques of Tunis. Cur-
1446-1458, 2014. [click]
rently, he is a research member of the Ad-
[21] K. Tanaka, H. Yoshida, H. Ohtake, and H. O. Wang, “A vanced Systems Laboratory (LSA) in the
sum of squares approach to modeling and control of non- Polytechnic School of Tunis. His research
linear dynamical systems with polynomial fuzzy systems,” interests include analysis and control of
IEEE Trans. on Fuzzy Systems, vol. 17, no. 4, pp. 911-922, nonlinear systems, fault detection of fuzzy systems.
August 2009. [click]
[22] P. Parrilo, “Stuctured semidefinite programs and semial- Mohammed Chadli received the Master
gebraic geometry methods in robustness and optimiza- degree (DEA) from the Engineering School
tion,” Ph.D. dissertation, California Institute of Technol- INSA-Lyon in 1999, the PhD thesis from
ogy, Pasadena, CA, 2000. [click] the CRAN-Nancy, France in 2002 and his
[23] S. Prajna, A. Papachristodoulou, and P. Parrilo, “Introduc- habilitation in 2011 at the University of Pi-
ing SOSTOOLS: a general purpose sum of squares pro- cardie Jules Verne (UPJV) in Amiens,
gramming solver,” Proc. of IEEE Conference on Decision France. From 1999 to 2004, he was an
and Control, 2002. [click] Assistant Prof CRAN-INPL. Since 2004,
[24] S. Prajna, A. Papachristodoulou, and F. Wu, “Nonlinear he has been an Associate Professor at the
control synthesis by sum of squares optimisation: A Lya- UPJV-MIS France. He was a visiting professorship at the Uni-
punov based approach,” Proc. of Asian Control Confer- versity of Agder-Norway and the University of Shanghai-China.
ence, 2004. [click] Dr Chadli’s research interests include fuzzy/LPV and switched
systems, singular systems, robust control, fault detection and iso-
[25] J. Sturm, “Using SeDuMi 1.02, a Matlab toolbox for op- lation (FDI), fault tolerant control (FTC). On the application side
timization over symmetric cones,” Optimization Methodes he is mainly interested in automotive control and renewable en-
and Software, vol. 11-12, pp. 625-653, 1999. [click] ergy. Dr. Chadli is a senior member of IEEE and he is also
[26] K. Tanaka, H Ohtake, and H. O. Wang, “An SOS-based sta- serving as an Editorial Board member for various peer-reviewed
ble control of polynomial siscrete fuzzy systems,” Amer- journals and was a Guest Editor for Special Issues in interna-
ican Control Conference, Washington, USA June 11-13, tional journals.
2008. [click]
[27] H. K. Lam and H. Li, “Output-feedback tracking control Naceur Benhadj Braiek obtained the Mas-
for polynomial fuzzy-model-based control systems,” IEEE ter degree from both the Ecole Nationale
Trans. on Industrial Electronics, vol. 60, no. 12, pp.5830- d’Ingenieurs de Tunis (ENIT, 1987) and
5840, December 2013. [click] the Institut Industriel du Nord (Ecole Cen-
[28] Y.J. Chen, M. Tanaka, K. Tanaka, H. Ohtake, and H. O. trale de Lille, 1988), the Ph.D. degree in
Wang, “Discrete polynomial fuzzy systems control,” IET Automatic Control from the University of
Control Theory Application, vol. 8, no. 4, pp. 288-296, Sciences and Techniques of Lille, France
2014. [click] (1990). Currently, he is a Professor of Elec-
trical Engineering at the University of Tu-
[29] H. K. Lam and S. H. Tsai, “Stability analysis of
nis and the head of the Automatic Control Laboratory (LECAP)
polynomial-fuzzy-model-based control systems with mis-
at the Ecole Polytechnique de Tunisie. His domain of interest
matched premise membership functions,” IEEE trans. on
is related to the modeling, analysis and control of nonlinear sys-
Fuzzy Systems, vol. 22, no. 1, pp. 223-229, February 2014.
tems with applications on electrical processes.
[click]