Automatica 93 (2018) 42–54

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Design of robust fuzzy fault detection filter for polynomial fuzzy

systems with new finite frequency specifications✩
Ali Chibani a , Mohammed Chadli b, *, Steven X. Ding c , Naceur Benhadj Braiek a
a
Advanced System Laboratory, Polytechnic School of Tunisia, University of Carthage BP. 743, 2078 La Marsa, Tunisia
b
University of Picardie Jules Verne, MIS(E.A 4290), 33 rue Saint-Leu, 80039 Amiens, France
c
University of Duisburg–Essen, Bismarkstrasse 81 BB, 47057 Duisburg, Germany

article info a b s t r a c t
Article history: This paper investigates the problem of fault detection filter design for discrete-time polynomial fuzzy
Received 18 December 2016 systems with faults and unknown disturbances. The frequency ranges of the faults and the disturbances
Received in revised form 28 November are assumed to be known beforehand and to reside in low, middle or high frequency ranges. Thus, the
2017
proposed filter is designed in the finite frequency range to overcome the conservatism generated by those
Accepted 2 January 2018
designed in the full frequency domain. Being of polynomial fuzzy structure, the proposed filter combines
the H− /H∞ performances in order to ensure the best robustness to the disturbance and the best sensitivity
Keywords: to the fault. Design conditions are derived in Sum Of Squares formulations that can be easily solved via
T–S fuzzy system available software tools. Two illustrative examples are introduced to demonstrate the effectiveness of the
Polynomial fuzzy system proposed method and a comparative study with LMI method is also provided.
Fault detection filter © 2018 Elsevier Ltd. All rights reserved.
Finite frequency domain
H− /H∞ index
LMI
SOS

1. Introduction to efficient results not only in theoretical developments but also

in industrial applications (Chadli & Karimi, 2013; Zhang & Wang,
Due to the increased requirements for safety as well as for 2016a, b). In its core, this technique consists of generating the
reliability and higher performance in modern systems, the issues residual signals from the difference between the system outputs
of fault detection (FD) have received more attention in the recent and the filter outputs and then a logic algorithm to decide whether
years (Li, Shi, Lim, & Wu, 2016; Li, Zhu, & Ding, 2011; Liu, Huang, a fault has occurred (Liu, Wang, & Yang, 2005; Yang & Wang, 2010;
Shi, & Xu, 2013; Wang, Shi, Lim, & Liu, 2016; Yang, Shi, & Liu, Zhong, Ding, Lam, & Wang, 2003).
2011). In industrial processes and engineering applications such as On the other hand, since most of real processes and industrial
nuclear power stations, chemical or aerospace processes, system systems are in nature of nonlinear dynamics, Takagi–Sugeno (T–S)
components like sensor, actuator etc., may be affected by several fuzzy systems are considered as very useful tool to present the
and serious failures. These problems may drastically change the dynamics of nonlinear processes (Takagi & Sugeno, 1985). The FD
system behavior, which leads to performance degradation or even problem for T–S fuzzy models has been extensively investigated in
system instability. Therefore, FD issues are very essential in order last years and many solutions are formulated as optimization al-
to enhance the system reliability and efficiency. The objective of gorithms with LMI constraints. These optimization algorithms are
fault detection is to detect the fault signal reliably whenever it ap- dedicated to designing the fault detection filter to be as sensitive as
possible to the fault (in order to detect faults in a timely way) and as
pears. One method to do this is to generate residual signals which
robust as possible to the unknown disturbances (in order to avoid
act as an indicator for faulty situation. In the literature, observer-
false alarms). Especially, the disturbance attenuation and the fault
based fault detection filter design is commonly used and leads
sensitivity requirements are expressed in terms of an H∞ index (for
example, Gao, Zhao, Lam, & Chen, 2009; Yoneyama, 2009; Zhang,
✩ The material in this paper was not presented at any conference. This paper was Shi, & Mehr, 2012 and the references therein) or an H− norm or a
recommended for publication in revised form by Associate Editor Michele Basseville mixed H− /H∞ norm (see for example, Chadli, Abdo, & Ding, 2013;
under the direction of Editor Torsten Söderström.
Wang, Shi, & Wang, 2010; Wang, Yang, & Liu, 2007).
*
Corresponding author.
E-mail addresses: [email protected] (A. Chibani), [email protected] In real processes and industrial systems, the frequency ranges of
(M. Chadli), [email protected] (S.X. Ding), [email protected] the disturbances and the faults are usually finite and known in ad-
(N.B. Braiek). vance. In the references previously cited, the frequency properties

https://doi.org/10.1016/j.automatica.2018.03.024
0005-1098/© 2018 Elsevier Ltd. All rights reserved.
 A. Chibani et al. / Automatica 93 (2018) 42–54 43

of such external signals are not considered when designing a fault SDP program and lead to less conservative results than the LMI
detection filter. Incorporating such information into the design can approaches (Chen et al., 2014, 2015; Lam Liu et al., 2015, Lam Wu
definitively improve the detection performance. In this context, et al., 2015; Tanaka et al., 2009). The resolution of the SOS based
the fault detection problem in finite frequency domain has been SDP algorithm can be efficiently performed via numerical tools
addressed in a number of recent works and it has been shown such that SOSTOOLS (Parrilo, 2000; Prajna, Papachristodoulou, &
that filters designed in finite frequency ranges lead to improved Wu, 2004).
FD performance than those designed in the full frequency domain. In the field of state estimation of polynomial fuzzy systems, a
These investigations are essentially motivated by the use of the polynomial fuzzy observer has been proposed in Tanaka, Ohtake,
so-called Generalized Kalman–Yakubovich–Popov (GKYP) lemma, Seo, Tanaka, and Wang (2012). Based on the Lyapunov method,
which converts the various properties of dynamic systems in finite sufficient design condition are formulated in SOS constraints. More
frequency domain into LMI conditions (Iwasaki & Hara, 2005). recently, the state reconstruction problem of PFS with unknown
The finite frequency FD filter design for linear time invariant (LTI) inputs has been discussed in authors’ previous work (Chibani,
systems has been first addressed in Wang and Yang (2008) and Chadli, & Braiek, 2016a). Indeed, a new structure of polynomial
Zhang, Jiang, Shi, and Xu (2015b). The work in Yang, Xia, and Zhang fuzzy observer has been proposed and sufficient design condi-
(2011) proposed a multi-objective H− /H∞ filter in finite frequency tions are derived in SOS terms with equality constraints. Even if
ranges to deal with the FD problem for delta operator systems. this attempts provide important innovations compared with LMI
Recently, the finite frequency H∞ filtering is investigated for state approaches for T–S fuzzy systems, the frequency properties of
delayed systems (Li & Yang, 2015), for switching linear parameter external signals are not incorporated in the design procedure, i.e.,
varying systems (Wang, Ju, & Wang, 2013; Zhong & Yang, 2015) only full frequency domain case is investigated. Moreover, to the
and for uncertain linear discrete-time systems (Zhang, Wang, Ding, best of authors’ knowledge, the fault detection of PFSs is not yet
et al., 2015). Almost all of these papers propose to transform the discussed in the literature.
FD task into optimization algorithms under LMI constraints and The main challenge in this study is how to bring together
demonstrate that the finite frequency conditions are generally less the SOS technique and the finite frequency approach in order to
conservative than their counter-part in the full frequency domain, establish a new criterion that permit to solve the fault detection
and contain the standard ones as a particular case. problem for polynomial fuzzy systems. Hence, this paper investi-
Turning to T–S fuzzy models, the first attempts that discussed gates the problem of fault detection for discrete-time PFSs with
the FD problem in finite frequency ranges for such systems unknown disturbances and faults in finite frequency ranges. A
appeared in Ding and Yang (2010). Indeed, three fuzzy filters, multi-objective H− /H∞ polynomial fuzzy filter will be proposed in
corresponding to disturbance frequency domain are designed in order to guarantee the best disturbances attenuation and the best
order to guarantee the finite-frequency l2 gain which extends the fault sensitivity. Based on the GKYP lemma, new characterizations
standard l2 gain over the whole frequency domain. The design of the H− and the H∞ norms are introduced with finite frequency
conditions are sufficient and derived in LMI terms. Afterward, an specifications. Then, using a fuzzy Lyapunov function, sufficient
H− /H∞ fault detection filter is proposed in Yang, Xia, and Liu design conditions are derived in SOS terms with linear equality
(2011) where two finite frequency performance indices are in- constraints. Moreover, the case of T–S fuzzy systems will be also
troduced to measure fault sensitivity and disturbance robustness. discussed and equivalent LMI design conditions will be presented
In Zhang, Jiang, Shi, and Xu (2015a), the FD problem is investigated instead of the SOS ones. In the simulation study, the effectiveness
with finite frequency specification such that the proposed filter of the proposed method will be shown through two examples
complies with pole-placement constraints. More recently, in the and a comparison study will be introduced to demonstrate the
authors’ previous work (Chibani, Chadli, Shi, & Braiek, 2016b), a superiority of the SOS approach against the LMI one.
multi-objective fuzzy fault detection filter is proposed in the finite The main contributions of this work are as follows: (1) it is the
frequency domain for T–S fuzzy models. The results are given in first attempt to considers the FD problem for PFSs. (2) new char-
LMI formulations. acterizations, including polynomial terms, are proposed to define
Going further, T–S fuzzy models and LMI technique have en- the finite frequency H− and the H∞ norms for PFSs. (3) A new
joyed great success and popularity due to the attractive properties form of polynomial filter is proposed to deal with the FD problem
of fuzzy control including robustness to disturbances, uncertain- of PFSs. (4) By exploiting the SOS technique, a new SOS criterion
ties, and sensors noise. However, there still exists a large number is developed to handle the optimization algorithms with polyno-
of design problems that either the processes cannot be exactly mial terms and less conservative conditions are further proposed.
represented by a T–S model, either the results obtained through (5) Similar results for T–S fussy systems is also treated as a partic-
LMIs are too conservative. ular case.
Recently, a new type of fuzzy model, namely polynomial fuzzy
The reminder of the paper is organized as follows. In Section 2,
system (PFS) has been developed as an extension of T–S fuzzy
the problem is formulated and useful preliminaries are introduced.
models where the consequent parts of the if-then rules are allowed
The main results are derived in Section 3. Simulation examples are
to be polynomial (Tanaka, Yoshida, Ohtake, & Wang, 2009). Indeed,
presented in Section 4 to show the effectiveness and potential of
when the degrees of polynomial terms are all zero, the model
the proposed design techniques. Finally, conclusion remarks are
reduces to the T–S fuzzy model and hence, this last can be regarded
given in Section 5.
as a particular case. The polynomial fuzzy systems are well adopted
in modeling, analysis and different design problems of nonlinear Notation. Throughout this paper, the notation is rather standard:
systems. Indeed, it was been proven that such models give more Let Rn denotes the n-dimensional Euclidean space and l2 [0, ∞)
relaxed criteria in stability analysis or control design since, the denotes the space
∑ of square integrable vector ofn discrete functions
number of local models is generally fewer than the existing T–S over [0, ∞). n denotes the subset from R that contains all
fuzzy systems (Chen, Tanaka, Tanaka, Ohtake, & Wang, 2014; Chen, SOS polynomial in n variables. P > 0 (Resp. P < 0) means
Tanaka, Tanaka, & Wang, 2015; Lam, Liu, Wu, & Zhao, 2015; Lam, that P is definite positive (Resp. definite negative) matrix. Given
Wu, & Lam, 2015; Tanaka et al., 2009). a matrix M, MT and M⊥ are used to represent the transposed and
Note that in this case, the LMI technique can-not be applied due the orthogonal complement (MM⊥ = 0) of M respectively. (∗)
to the polynomial terms. Thus, design criteria in stability analysis denotes the transposed element in the symmetric position, and I
or controller synthesis are derived in Sum Of Squares (SOS) based is the identity matrix with appropriate dimension.
44 A. Chibani et al. / Automatica 93 (2018) 42–54

2. Problem formulation and preliminaries in which the same activation functions µi (ζk ) as used for the
polynomial fuzzy system (2) are considered and where x̂k ∈ Rn is
2.1. Problem formulation the estimated vector and Ni (y) ∈ Rn×n , Gi (y) ∈ Rn×m , Li (y) ∈ Rn×p
and E ∈ Rn×p are the filter gains to be determined.
Consider a discrete-time polynomial fuzzy system affected by Defining the state estimation error to be ek = xk − x̂k and the
faults and unknown disturbances such that the ith local model is residual signal as rk = yk − ŷk , then one can write
described as:
{
ek = Txk − zk + EDdk + EHfk
(6)
if ζ1 k is Mi1 · · · ζp k is Mip then, rk = Cek + Ddk + Hfk

{ with
xk+1 = Ai (yk )xk + Bi (yk )uk + Ri (yk )dk + Fi (yk )fk
yk = Cxk + Ddk + Hfk
(1) T = I + EC . (7)

where xk ∈ Rn is the state vector, uk ∈ Rm is the known input For simplicity, the following notations are used for the rest of
vector, yk is output vector, dk ∈ Rqd and fk ∈ Rqf are the distur- the paper:
r r
bance and the fault signals respectively that belong to l2 and whose ∑ ∑
frequency ranges are finite. Let θd and θf be the frequency ranges Aµ (y) = µi (ζk )Ai (y), Bµ (y) = µi (ζk )Bi (y) (8a)
of the disturbance and the fault signals respectively. Ai (yk ) ∈ Rn×n , i=1 i=1
Bi (yk ) ∈ Rn×m , Ri (yk ) ∈ Rn×qd and Fi (yk ) ∈ Rn×qf are polynomial ∑r
∑r

matrices of yk . C , D and H are constant matrices of appropriate Rµ (y) = µi (ζk )Ri (y), Fµ (y) = µi (ζk )Fi (y) (8b)
dimensions. ζk = [ζ1 k , . . . , ζp k ] are the premise variables assumed i=1 i=1
to depend on measurable variables, Mij is the membership function r r
∑ ∑
and r denotes the number of polynomial local models. To lighten Nµ (y) = µi (ζk )Ni (y), Gµ (y) = µi (ζk )Gi (y) (8c)
the notation, in all the rest of the paper, we drop the notation with i=1 i=1
respect to time samples k for all polynomial matrices on yk . For ∑ r

instance, we will use Ai (y), Bi (y), Ri (y) and Fi (y) instead of Ai (yk ), Lµ (y) = µi (ζk )Li (y). (8d)
Bi (yk ), Ri (yk ) and Fi (yk ), respectively. i=1
The overall polynomial fuzzy system is given by: The dynamic of the filtering error system (6) can be rewritten as:
⎧ r ( ⎧ ( )
∑
x = µ ( ζ ) Ai (y)xk + Bi (y)uk + Ri (y)dk + e = N (y)e + TA (y) − N (y)T − L (y)C xk +
⎪
µ µ µ µ
⎪
⎪ k+1 i k
⎪
⎪ k+1
⎪
⎨ ⎪
⎪
i=1
⎪
⎪ ( ) (
(2)
) ⎪
⎪
⎪
⎪ Fi (y)fk ⎨ TBµ (y) − Gµ (y) uk + TRµ (y) − Nµ (y)ED −
⎪
(9)
⎪
⎪ ) ( )
⎩
yk = Cxk + Ddk + Hfk ⎪
⎪Lµ (y)D dk + TFµ (y) − Nµ (y)EH − Lµ (y)H f +
⎪
⎪
with µi (ζk ) are the activation functions that verify:
⎪
⎪
EDdk+1 + EHfk+1
⎪
⎪
⎪
rk = Cek + Ddk + Hfk .
⎩
r
∑
µi (ζk ) ≥ 0, µi (ζk ) = 1. (3)
Consequently, if the following constraints hold:
i=1

TAµ (y) − Nµ (y)T − Lµ (y)C = 0 (10a)

Remark 1. Fuzzy representation (2) with polynomial consequent
part, can be obtained by using the sector nonlinearity concept. TBµ (y) − Gµ (y) = 0 (10b)
In several cases, fuzzy polynomial structure (2) exactly repre-
sents the original nonlinear system and in general, it leads to a E [D H ] = 0 (10c)
fewer number of local models than that of a T–S fuzzy system. then, we obtain:
Moreover, if all matrices in (2) are independent of y (i.e., Ai , Bi , Ri
ek+1 = Aµ (y)ek + B1µ (y)dk + B2µ (y)fk
{
and Fi are constant matrices for all i) then, (2) reduces to a T–S fuzzy
model. Hence, polynomial fuzzy systems are more general repre- (11)
sentation and contain T–S fuzzy models as a particular case (Tanaka rk = C ek + D1 dk + D2 fk
et al., 2009). where
In order to reconstitute the state vector and to detect the fault, Aµ (y) = Nµ (y) (12a)
a polynomial fuzzy filter is proposed such that the ith local model
is described as: B1µ (y) = TRµ (y) − Kµ (y)D (12b)

if ζ1 k is Mi1 and · · · and ζp k is Mip then, B2µ (y) = TFµ (y) − Kµ (y)H (12c)

zk+1 = Ni (y)zk + Gi (y)uk + Li (y)yk C=C (12d)

{
x̂k = zk − Eyk (4)
D1 = D (12e)
ŷk = C x̂k
and the overall fault detection polynomial fuzzy filter has the D2 = H (12f)
following form: with
⎧ r ( )
Kµ (y) = Nµ (y)E + Lµ (y). (13)
∑
µ ζ
⎪
z = ( ) N (y)z + G (y)u + L (y)y
⎪
⎨ k+1
⎪ i k i k i k i k
i=1 (5) Note that from (10a), we get:
⎪x̂ = zk − Eyk
⎩ k
⎪
Nµ (y) = TAµ (y) − Kµ (y)C .
⎪
ŷk = C x̂k (14)
 A. Chibani et al. / Automatica 93 (2018) 42–54 45

Now, the problem can be formulated as follows: Given a poly- (3) For the high frequency range |θd | ≥ θdh :
nomial fuzzy model (2), find a finite frequency filter of the form (5) ∞ ∞
)2 ∑
θdh
(
in order to meet the following requirements:
∑
(ek+1 − ek )(ek+1 − ek )T ≥ 2 sin ek eTk . (22)
2
k=0 k=0
(i) The asymptotic stability of system (11);
(ii) The best robustness to disturbances, i.e. for d ̸ = 0, the H∞ Proof. The proof comes from Lemma 1, so omitted here.
performance from d to r is less than a given positive scalar
γ in a finite frequency domain; Definition 1 (Parrilo, 2000). A multivariate polynomial p(x) =
(iii) The best sensitivity to faults, i.e. for f ̸ = 0, the H− perfor- p(x1 , . . . , xn ) is a sum of squares (SOS) if there exist polynomials
mance from f to r is greater than a given positive scalar β in f1 (x), . . . , fm (x) such that
a finite frequency domain. m
∑
p(x) = fi2 (x). (23)
In the next section, we will recall some useful tools of the finite
i=1
frequency approach and we will provide a brief review of the SOS
decomposition that will be needed in the sequel. The existence of an SOS decomposition (23) can be shown
equivalent to the existence of a positive semi-definite matrix Q ,
such that
2.2. Preliminaries
p(x) = Z T (x)QZ (x) (24)
Lemma 1 (Ding & Yang, 2010). System (11) has a finite frequency where Z (x) is the vector of monomials of degree less than or equal
H∞ index bound γ from dk to rk , if under zero initial condition, the to deg(p)/2. Since every p ∈
∑
n is a sum of squared polynomials,
following inequality it is clear that p(x) ≥ 0, ∀x ∈ Rn . However, the converse is not
∞
∑ ∞
∑ evident.
rkT rk ≤ γ 2 dTk dk (15)
k=0 k=0 Proposition 1 (Prajna et al., 2004). Let M(x) be an nM × nM
symmetric polynomial matrix in x ∈ Rn and v ∈ RnM . Then v T M(x)v
holds for all solutions of (11) with dk ∈ l2 such that
being an SOS implies M(x) ≥ 0 for x ∈ Rn .
(1) For the low frequency range |θd | ≤ θdl :
Lemma 3. Projection lemma (Gahinet & Apkarian, 1994). Let Γ , Λ,
∞ ∞
)2 ∑
θdl and Θ be given. There exists a matrix Σ satisfying Θ + Γ ΣΛT +
∑ (
(ek+1 − ek )(ek+1 − ek ) ≤ T
2 sin ek eTk . (16) ΛΣ T Γ T < 0 if and only if:
2
k=0 k=0
T T

(2) For the middle frequency range θd1 ≤ θd ≤ θd2 :

Γ ⊥ ΘΓ ⊥ < 0 and Λ⊥ ΘΛ⊥ < 0.
∞ ( )( The following section gives the main results of this paper.
exp(jθdm )
(jθd ) Indeed, an H− /H∞ fuzzy fault detection filter will be designed in
∑
ek+1 − exp 1 ek ek+1
finite frequency domain to achieve the best robustness to distur-
k=0 (17)
)∗ bances and the best sensitivity to faults.
(jθd )
− exp 2 ek ≤0
3. Main results
where θdm = (θd2 − θd1 )/2
(3) For the high frequency range |θd | ≥ θdh : We start with the first requirement aiming at establishing the
stability conditions for the detection filter to be designed.
∞ ∞
)2 ∑
θdh
∑ (
(ek+1 − ek )(ek+1 − ek )T ≥ 2 sin ek eTk . (18) 3.1. Stability condition
2
k=0 k=0

When we set d = 0 and f = 0, the filtering error system is given

Lemma 2. System (11) has a finite frequency H− index bound β from
by:
fk to rk , if under zero initial condition, the following inequality
∞
∑ ∞
∑ ek+1 = Aµ (y)ek . (25)
rkT rk ≥ β 2 fkT fk (19)
The following theorem gives a sufficient SOS condition to guaran-
k=0 k=0
tee the asymptotic stability of system (25).
holds for all solutions of (11) with dk ∈ l2 such that
Theorem 1. System (25) is asymptotically stable if there exist con-
(1) For the low frequency range |θd | ≤ θdl :
stant matrices S, Ui = UiT > 0, a non singular matrix Σ and
∞
∑ (
θdl
∞
)2 ∑ polynomial matrices Wi (y), such that the following conditions hold
(ek+1 − ek )(ek+1 − ek )T ≤ 2 sin ek eTk . (20) ∀i = 1, . . . , r and ℓ = 1, . . . , r:
2
k=0 k=0
− ViT Mi,ℓ + Ei (y)IM Vi
( )
SOS (26)
(2) For the middle frequency range θd1 ≤ θd ≤ θd2 :
∞ ( )( where
(jθdm ) (jθd ) ∆
∑ ( )
exp ek+1 − exp 1 ek ek+1 (∗)
Mi,ℓ = (27)
k=0 )∗ (21) −Σ T + (Σ + SC )Ai (y) − Wi (y)C Uℓ − Σ − Σ T

− exp(jθd2 ) ek ≤0 with ∆ = ((Σ + SC )Ai (y) − Wi (y)C ) + ((Σ + SC )Ai (y) − Wi (y)C )T − Ui ,
Ei (y) > 0 for y ̸ = 0, Vi are vectors with appropriate dimensions that
where θdm = (θd2 − θd1 )/2 are independent of y and IM is the identity matrix.
46 A. Chibani et al. / Automatica 93 (2018) 42–54

Proof. System (25) is stable if case by choosing θd1 = −θdl and θd2 = θdl and for the high
frequency case by choosing θd1 = θdh and θd2 = 2Π − θdh . Thus,
−Uµ+ + Σ Aµ (y) + ATµ (y)Σ T
( )
(∗)
< 0. (28) the proof is complete.
−Σ T + Σ Aµ (y) Uµ+ − Σ − Σ T
Remark 2. Note that when all matrices in (29) and (30) are
[ ]
Indeed, pre-multiplying Eq. (28) by I Aµ (y) and post-multiplying
]T independent of y, system (29) reduces to a T–S fuzzy system and
it by I Aµ (y) , we obtain ATµ (y)(Uµ+ )Aµ (y) − Uµ < 0. This
[
Theorem 2 is not other than Lemma 3 in Ding and Yang (2010)
inequality corresponds to ∆V (ek ) = V (ek+1 ) − V (ek ) = ek+1 Uµ+ which introduces a finite frequency l2 gain conditions for T–S fuzzy
ek+1 − ek Uµ ek < 0 with V (ek ) = eTk Uµ ek , Uµ > 0. Finally using models. Hence the proposed conditions are more general than
Proposition 1, results in (26), which ends the proof. those given in Ding and Yang (2010). Furthermore, if Q = 0 in Ξd ,
inequality (30) becomes:
3.2. Finite frequency robustness condition
( )
ATµ Pµ+ Aµ − Pµ + C T C ATµ Pµ+ B1µ + C T D1
< 0 (33)
In this sub-section, we address requirement (ii). To this end, we B1Tµ Pµ+ Aµ + D1T C B1Tµ Pµ+ B1µ + D1T D1 − γ 2 I
consider f = 0 (d ̸ = 0). Then, system (11) becomes
{ which is none other than the standard H∞ norm condition de-
ek+1 = Aµ (y)ek + B1µ (y)dk signed in the full frequency range for discrete-time T–S fuzzy
(29)
rk = C ek + D1 dk . systems where the fuzzy Lyapunov function is chosen as V (ek ) =
eTk Pµ ek with Pµ = PµT > 0. Thus, the finite frequency conditions are
Theorem 2. System (29) has a finite frequency H∞ index bound γ , more general and contain the full frequency domain conditions as
if there exists a polynomial fuzzy-basis-dependent symmetric matrix particular cases.
Pµ (y) and a symmetric positive definite matrix Q such that the follow-
ing conditions hold: Remark 3. Note that for a linear system, i.e., when system (29)
is independent of µ and the matrices in (30) are independent of µ
Φ T Ξd Φ + Ψ T Πd Ψ < 0 (30) and µ+ , the sufficient condition is also a necessary one. In this case,
(
Aµ (y) B1µ (y)
) (
C D1
) (
I 0
) Theorem 2 is none other than the generalized KYP lemma (Iwasaki
where Φ = I 0
,Ψ = 0 I
, Πd = 0 −γ 2 I
and : & Hara, 2005).
(1) For the low frequency range |θd | ≤ θdl : Based on Theorem 2, the following result gives sufficient design
( +
) conditions on the existence of a finite frequency fault detection
−Pµ+ (y ) Q
Ξd = . filter (5) such that system (29) satisfies the H∞ performance re-
Q Pµ (y) − 2cos(θdl )Q quirement ∥rk ∥2 < γ ∥dk ∥2 . Since we consider a polynomial fuzzy
(2) For the middle frequency range θd1 ≤ θd ≤ θd2 : representation, design conditions are derived in SOS formulations.

−Pµ+ (y+ ) exp(jθdc ) Q

( )
Theorem 3. System (29) has a finite frequency H∞ performance index
Ξd =
exp(−jθdc ) Q Pµ (y) − 2cos(θdm )Q γ in a finite frequency domain if there exist constant matrices S and
Q = Q T > 0, a non singular matrix Σ and polynomial matrices
where θdc = (θd1 + θd2 )/2 and θdm = (θd2 − θd1 )/2
Wi (y), and Pi (y) = PiT (y) such that the following conditions hold
(3) For the high frequency range |θd | ≥ θdh :
( ) ∀i = 1, . . . , r and ℓ = 1, . . . , r:
−Pµ+ (y+ ) −Q T
Ξd = . − Vid Mdi,ℓ (y) + Eid (y)IMd Vid
( )
−Q Pµ (y) + 2cos(θdh )Q SOS (34)
where
Here µ+ is used to denote the fuzzy basis functions at the time instant
k + 1 and can be obtained by (3), where ξk are replaced by ξk+1 . (1) For d in the low frequency range |θd | ≤ θdl :
Moreover, y+ is used to denote y(k + 1).
−Pℓ (y+ ) (∗) (∗) (∗)
⎛ ⎞

Proof. Firstly, we consider the middle frequency range. Suppose ⎜Q −Σ Θi (y) (∗) (∗)⎟
Mdi,ℓ (y) = ⎝ (35)
(30) holds. Then, multiplying it by [eTk dTk ] from the left and by its 0 Ω −γ 2 I (∗)
⎠
conjugate transpose from the right, we have 0 C D −I
where Θi (y) = (Σ + SC )Ai (y) − Wi (y)C + (Σ + SC )Ai (y) −
(k − ek+1 Pµ+ (y )ek+1 + rk rk − γ dk dk
( ) (
eTk Pµ (y)e T + T 2 T
)T
Wi (y)C + Pi (y) − 2 cos(θdl )Q .
+ trace Q exp(−jθdc ) ek+1 eTk + exp(jθdc ) ek eTk+1 −
(
) (31) (2) For d in the middle frequency range θd1 ≤ θd ≤ θd2 :

2 cos(θdm )ek ekT

< 0. −Pℓ (y+ )
) ⎛ ⎞
(∗) (∗) (∗)
(−jθdc )
⎜exp Q −Σ Θi (y) (∗) (∗)⎟
Mdi,ℓ (y) = ⎜ ⎟ (36)
Note that e(0) = 0 and limk→∞ ek = 0 (due to the stability of ⎝ 0 Ω −γ I 2
(∗)⎠
(29) which has been established in Section 3.1), then taking the 0 C D −I
summation from k = 0 to ∞, we obtain
where Θi (y) = (Σ + SC )Ai (y) − Wi (y)C + (Σ + SC )Ai (y) −
( ) (
∞
∑ )T
rkT rk − γ 2 dTk dk < −trace(Q E ) (32) Wi (y)C + Pi (y) − 2cos(θdm )Q , θdc = (θd1 + θd2 )/2 and
k=0 θdm = (θd2 − θd1 )/2.
(−jθdc ) (3) For d in the high frequency range |θd | ≥ θdh :
ek+1 eTk + exp(jθdc ) ek eTk+1 − 2 cos(θdm )ek eTk .
∑∞
with E = k=0 exp
Then, we can easily verify that −E is equal to the left-hand side −Pℓ (y+ ) (∗) (∗) (∗)
⎛ ⎞
of (17). Then, we conclude that E > 0 and trace(Q E ) > 0 (since ⎜−Q − Σ Θi (y) (∗) (∗)⎟
Q > 0). Hence, from (32), condition (15) is satisfied for the middle Mdi,ℓ (y) = ⎝ (37)
0 Ω −γ 2 I (∗)
⎠
frequency ranges. Similarly, the result comes for the low frequency 0 C D −I
 A. Chibani et al. / Automatica 93 (2018) 42–54 47

where Θi (y) = (Σ + SC )Ai (y) − Wi (y)C + (Σ + SC )Ai (y) −

( ) (
The following theorem gives an SOS condition to satisfy the H−
performance ∥rk ∥2 > β∥fk ∥2 .
)T
Wi (y)C + Pi (y) + 2cos(θdh )Q ,
)T
with Ω = (Σ + SC )Ri (y) − Wi (y)D for low, middle and high Theorem 4. System (44) has a finite frequency H− performance index
(
frequency cases, Vid are vectors with appropriate dimensions that are β in a finite frequency domain if there exist constant matrices S and
Q = Q T > 0, a non singular matrix Σ and polynomial matrices
independent of y, Eid (y) > 0 for y ̸ = 0 and IMd is the identity matrix.
Wi (y), and Pi (y) = PiT (y) such that the following conditions hold
∀i = 1, . . . , r and ℓ = 1, . . . , r:
Proof. To prove the result, we start with the low frequency case.
T
( )
From Theorem 2, we have the finite frequency H∞ inequality − Vif f f
Mi,ℓ (y) + Ei (y)IMf
f
Vi SOS (45)
Φ Ξd Φ + Ψ Πd Ψ < 0
T T
(38) where
( ) ( ) ( )
I 0
where Φ =
Aµ (y) B1µ (y)
, Ψ =
C D1
, Πd = and (1) For f in the low frequency range |θf | ≤ θfl :
−γ 2 I
+
I 0
) 0 I 0
−Pℓ (y+ )
(
−Pµ+ (y ) Q
⎛
(∗) ( ∗) (∗)
⎞
Ξd = P (y) − 2cos(θ )Q
.
Q µ dl
f ⎜Q −Σ Υi (y) (∗) (∗)⎟
Eq. (38) can be rewritten as follows: Mi,ℓ (y) = ⎝ (46)
0 Λ −H T H (∗)
⎠

Γ ⊥
(
M T Ξd M + N T Πd N Γ
) ⊥T
<0 (39) 0 0 βI −I
where Υi (y) = ((Σ + SC )Ai (y) − Wi (y)C ) + ((Σ + SC )Ai (y) −
ATµ (y)
( ) ( ) ( )
I 0 I 0 0 0 C D1
where Γ ⊥ = ,M = 0 I 0
,N = 0 0 I
. Wi (y)C )T − C T C + Pi (y) − 2 cos(θfl )Q .
B1µ (y) 0 I
According to Lemma 3, if there exists a non singular matrix Σ such (2) For f in the middle frequency range θf1 ≤ θf ≤ θf2 :
that: ⎛
−Pℓ (y+ ) (∗) (∗) (∗)
⎞
(−jθfc )
exp Q −Σ Υi (y) (∗) (∗)⎟
Γ Σ T ΛT + ΛΣΓ T + M T Ξd M + N T Πd N < 0 (40) f
Mi,ℓ (y) = ⎝
⎜
(47)
0 Λ −H T H (∗)
⎠
then condition (39) holds and 0 0 βI −I
) T where Υi (y) = ((Σ + SC )Ai (y) − Wi (y)C ) + ((Σ + SC )Ai (y) −
Λ⊥ M T Ξd M + N T Πd N Λ⊥ < 0
(
(41)
Wi (y)C )T − C T C + Pi (y) − 2cos(θfm )Q and θfc = (θf1 + θf2 )/2
for given vectors Γ and Λ where Γ ⊥ and Λ⊥ designate the or- and θfm = (θf2 − θf1 )/2.
thogonal)complements of Γ and Λ respectively. Note that for Γ =
( (3) For f in the high frequency range |θf | ≥ θfh :
−I (0)
−Pℓ (y+ ) (∗) (∗) (∗)
⎛ ⎞
Aµ (y)T it is easy to verify that Γ ⊥ Γ = 0. Now, let Λ = I . By
B1µ (y)T 0 f ⎜−Q − Σ Υi (y) (∗) (∗)⎟
Mi,ℓ (y) = ⎝ (48)
some matrix manipulations, condition (40) becomes 0 Λ −H T H (∗)
⎠
0 0 βI −I
−Pµ+ (y+ ) (∗) (∗) (∗)
⎛ ⎞
⎜ Q −Σ ∆d (y) (∗) (∗)⎟ where Υi (y) = ((Σ + SC )Ai (y) − Wi (y)C ) + ((Σ + SC )Ai (y) −
⎝ 0 B1Tµ (y)Σ T −γ 2 I (∗)
⎠<0 (42) Wi (y)C )T − C T C + Pi (y) + 2cos(θfh )Q .
0 C D1 −I with Λ = ((Σ + SC )Fi (y) − Wi (y)H)T − H T C for low, middle and high
f
where ∆d (y) = Σ Aµ (y) + Aµ (y)Σ + Pµ (y) − 2 cos(θdl )Q . Now,
T T frequency cases, Vi are vectors with appropriate dimensions that are
f
taking account definitions (12) and (14) and letting S = Σ E and independent of y, Ei (y) > 0 for y ̸ = 0 and IMf is the identity matrix.
Wi (y) = Σ Ki (y), then Eq. (42) leads to
Proof. The main idea to prove this theorem is the same as for
−Pℓ (y+ ) (∗) (∗) (∗)
⎛ ⎞
Theorem 3. However, notes
( are included here for completeness.
⎜Q −Σ Θi (y) (∗) (∗)⎟
)
−I 0
⎠<0 (43) Indeed, choosing Πf = for a given β > 0 instead
Ω −γ 2 I 0 β2I
⎝ 0 (∗)
0 C D −I of Πd in the finite frequency condition (30), we have the finite
frequency H− index for system (44). The rest is similar as the proof
where Θi (y) = (Σ + SC )Ai (y) − Wi (y)C + (Σ + SC )Ai (y) −
( ) (
of Theorem 3.
)T )T
Wi (y)C + Pi (y) − 2 cos(θdl )Q and Ω = (Σ + SC )Ri (y) − Wi (y)D .
(
In the sequel, we state the global result. Indeed combining the
Thus, By using (Proposition 1, if condition
) (34) holds, then we previous theorems we present sufficient SOS conditions on the
guarantee that Mdi,ℓ (y) + Eid (y)IMd ≤ 0 and consequently (43) existence of a fuzzy filter (5) for polynomial fuzzy system (2).
holds.
The same lines are followed for the middle and the high fre- 3.4. Finite frequency filter design
quency ranges, which lead to Eqs. (36) and (37), and thus ends the
proof. Theorem 5. Consider a discrete-time polynomial fuzzy system (2).
There exists a polynomial fuzzy fault detection filter (5) (in a finite
frequency domain) such that system (11) is stable and has a finite
3.3. Finite frequency fault sensitivity condition frequency H∞ performances index γ and a finite frequency H− per-
formances index β if there exist constant matrices S, Ui = UiT > 0
For a successful fault detection, we propose to enhance the and Q = Q T > 0, polynomial matrices Wi (y), Ji (y) and Pi (y) = Pi (y)T
sensitivity of the residual output on faults by means of the max- and a non singular matrix Σ such that the following conditions hold
imization of an H− index from f to r (i.e. to satisfy the requirement ∀i = 1, . . . , r and ℓ = 1, . . . , r:
(iii)). The dynamic of the filtering error system, when d = 0 (f ̸ = 0)
is given by SOS condition (26)
{ SOS condition (34)
ek+1 = Aµ (y)ek + B2µ (y)fk
(44) SOS condition (45)
rk = C ek + D2 fk .
48 A. Chibani et al. / Automatica 93 (2018) 42–54

Ji (y) − (Σ + SC )Bi (y) = 0 (49)

( )
− Vi T Mi,i (y) + Ei (y)IM Vi SOS (56)
S [D H ] = 0. (50)
Then, the filter’s parameters are obtained by: ( 2
−Vi T Mi,i (y) + Mi,ℓ (y) + Mℓ,i (y)
r −1 (57)
E = Σ −1 S (51a) )
+ Ei (y)IM Vi SOS
Ni (y) = TAi (y) − Σ −1 Wi (y)C (51b)
Ji (y) − (Σ + SC )Bi (y) = 0 (58)
Li (y) = Σ −1
Wi (y) − Ni (y)E (51c)
S [D H ] = 0 (59)
Gi (y) = Σ −1 Ji (y). (51d)
where, Mdi,ℓ (y) is defined in Eqs. (35), (36) and (37) according to the
Proof. Equality condition (58) is introduced to verify (10b). Indeed, f
frequency ranges of the disturbances, Mi,ℓ (y) is defined in Eqs. (46),
post-multiplying (10b) by Σ and making Ji (y) = Σ Gi (y), we (47) and (48) according to the frequency ranges of the fault and
obtain Eq. (58). Similarly, condition (59) is equivalent to (10c). The f
Mi,ℓ (y) is defined in Eq. (27). Vid , Vi and Vi are vectors with appropri-
rest comes from Theorems 1, 3 and 4. f
ate dimensions that are independent of y, Eid (y) > 0, Ei (y) > 0 and
Remark 4. Theorem 5 proposes the sufficient design conditions Ei (y) > 0 for y ̸ = 0 and IMd , IMf and IM represent the identity
for the existence of a polynomial fuzzy fault detection filter for matrices with appropriate dimensions.
polynomial fuzzy systems with new finite frequency specifications.
The proposed approach has the advantage to define new finite Proof. The proof follows immediately from Theorem 5 by exploit-
frequency characterizations for the H∞ and H− norms with polyno- ing the work in Tuan et al. (2001).
mial terms. This facts makes the proposed filter useful for a large
class of nonlinear processes that can be modeled by polynomial
systems. Hence, the results are given in SOS formulation which 3.6. Case of T–S fuzzy models
provides more of relaxations than the LMI technique. Moreover,
the stability of the filter is ensured by a fuzzy Lyapunov function In the previous section, the finite frequency fault detection
which leads to less conservative conditions. problem is investigated for polynomial fuzzy systems. As dis-
cussed in Remark 1, T–S fuzzy models present a particular case of
Remark 5. Exploiting the previous results, the fault detection polynomial structure. Indeed, when all system’s matrices become
problem of polynomial fuzzy systems can be formulated by an constant (i.e., independent of y), it becomes the classical T–S fuzzy
optimization algorithm as follows: models. In this section, the fault detection filter design problem for
max β, minγ T–S fuzzy models is addressed in finite frequency ranges. To this
s.t . (26), (34), (45), (58), (59). end, polynomial fuzzy system (1) is reduced to the following T–S
fuzzy model
Then, the filter’s parameters can be calculated by (51).
r
⎧
∑
µi (ζk ) Ai xk + Bi uk + Ri dk + Fi fk
⎪
⎨x ( )
Remark 6. Note that we have used a multiple Lyapunov function k+1 =
instead of a single one. Hence, more relaxed conditions can be (60)
⎪ i=1
given by exploiting the work of Tuan, Apkarian, Narikiyo, and
⎩
yk = Cxk + Ddk + Hfk
Yamamoto (2001). The following result provides less conservative
conditions. and the proposed filter is given by
⎧ r
∑
µi (ζk ) Ni zk + Gi uk + Li yk
⎪ ( )
z =
⎪
3.5. SOS relaxations for the design of a fault detection filter ⎨ k+1
⎪
i=1 (61)
Lemma 4. Consider a discrete-time polynomial fuzzy system (2).
⎪ x̂ = zk − Eyk
⎩ k
⎪
ŷk = C x̂k .
⎪
There exists a polynomial fuzzy fault detection filter (5) (in a finite
frequency domain) such that system (11) is stable and has a finite
To deal with the fault detection problem for T–S fussy systems,
frequency H∞ performance gain γ and a finite frequency H− perfor-
similar theoretical developments are followed and the same def-
mance gain β if there exist constant matrices S, Ui = UiT > 0 and
initions as for the polynomial case are considered such that all
Q = Q T > 0, polynomial matrices Wi (y), Ji (y) and Pi (y) = Pi (y)T
polynomial matrices are restricted to be constant (i.e., independent
and a non singular matrix Σ such that the following conditions hold
of yk ). Consequently, the error system (11) becomes
∀i = 1, . . . , r and ℓ = 1, . . . , r and i ̸= ℓ: {
T
( ) ek+1 = Aµ ek + B1µ dk + B2µ fk
− Vid Mdi,i (y) + Eid (y)IMd Vid SOS (52) rk = C ek + D1 dk + D2 fk .
(62)

T
( 2 Thus, we propose the following result, derived from Theorem 5, to
−Vid Md (y) + Mdi,ℓ (y) + Mdℓ,i (y) achieve the fault detection goal for T–S fuzzy models by means of a
r − 1 i,i (53)
) T–S fuzzy multi-objective H− /H∞ filter. The design conditions are
+ Eid (y)IMd Vid SOS presented in LMI formulation.

T
( )
− Vif f f
Mi,i (y) + Ei (y)IMf Vi
f
SOS (54) Corollary 1. Consider a discrete-time T–S fuzzy model (60). There
exists a fuzzy fault detection filter (61) (in a finite frequency domain)
( 2 such that system (62) is stable and has a finite frequency H∞ perfor-
T
−Vif f f f
M (y) + Mi,ℓ (y) + Mℓ,i (y) mances index γ and a finite frequency H− performances index β if
r − 1 i ,i ) (55) there exist matrices S, Wi , Ji , Hermitian matrices Pi , Ui > 0 and Q > 0
+ Eif (y)IMf Vif SOS and a non singular matrix Σ such that the following conditions hold
 A. Chibani et al. / Automatica 93 (2018) 42–54 49

∀i = 1, . . . , r and ℓ = 1, . . . , r: The filter’s matrices are given by:

Ji − (Σ + SC )Bi = 0 (63) E = Σ −1 S (72a)

S([D H ] = 0 (64) Ni = (I + EC )Ai − Σ −1
Wi C (72b)
Uℓ − Σ − Σ T
)
(∗)
<0 (65) Li = Σ −1
Wi − Ni E (72c)
((Σ + SC )Ai − Wi C )T − Σ
∆
)T Gi = Σ −1
Ji . (72d)
where ∆ = (Σ + SC )Ai − Wi C + (Σ + SC )Ai − Wi C − Ui .
( ) (
and Proof. The corollary is derived from Theorem 5 and the proof is
according to θd immediate and omitted here.

(1) For d(t) in the low frequency range |θd | ≤ θdl : Remark 7. The T–S fuzzy model is regarded as a particular case
⎛
−Pℓ (∗) (∗) (∗)
⎞ of polynomial fuzzy systems. The use of the polynomial fuzzy
⎜Q − Σ Θi (∗) (∗)⎟ representation is motivated by the invention of the SOS technique
⎝ 0 Ω −γ 2 I (∗)
⎠<0 (66) to deal with the finite set of polynomial constraints (Parrilo, 2000).
0 C D −I The advantages of the polynomial structure and the SOS technique
are widely shown in a lot of recent works (Prajna et al., 2004;
where Θi = (Σ + SC )Ai − Wi C + (Σ + SC )Ai − Wi C
( ) ( )T
+ Tanaka et al., 2009). A comparison study will be made in the
Pi − 2 cos(θdl )Q . simulation section to show the benefits of the polynomial approach
(2) For d(t) in the middle frequency range θd1 ≤ θd ≤ θd2 : against the classical T–S method.
−Pℓ (∗) (∗) (∗)
⎛ ⎞
⎜e−jθdc Q − Σ Θi (∗) (∗)⎟
⎠<0 (67) 4. Simulation study
0 Ω −γ 2 I (∗)
⎝
0 C D −I This section demonstrates the validity of our proposed ap-
where Θi = (Σ + SC )Ai − Wi C + (Σ + SC )Ai − Wi C T + proach through two examples.
( ) ( )
Pi − 2cos(θdm )Q , θdc = (θd1 + θd2 )/2 and θdm = (θd2 − θd1 )/2
(3) For d(t) in the high frequency range |θd | ≥ θdh : 4.1. Example 1: polynomial fuzzy model
− Pℓ (∗) (∗) (∗)
⎛ ⎞
Let consider a two-rules polynomial fuzzy system (2) given by
⎜−Q − Σ Θi (∗) (∗)⎟
⎝ 0 Ω −γ I 2
(∗)
⎠<0 (68) the following data:
0.5 0.5
( ) ( )
0 C D −I 0 0
A1 (x2 ) = , A2 (x2 ) =
− x 2
1 − x2
− 0. 2172
where Θi = (Σ + SC )Ai − Wi C + (Σ + SC )Ai − Wi C
( ) ( )T
+ ( 2 ) (2 )
0.1 0.1
( )
Pi + 2cos(θ(dh )Q . 0
B1 = B2 = , R1 = , R2 =
0.1 0 0.1
with Ω = (Σ + SC )Ri − Wi D T .
)
0.001
( ) ( )
1 0
according to θf C = , D1 = D2 =
0 1 0.001
( )
(1) For f (t) in the low frequency range |θf | ≤ θfl : 1
F1 = F2 = H1 = H2 =
1
− Pℓ (∗) (∗) (∗)
⎛ ⎞
sin(x2 ) + 0.2172x2 x2 − sin(x2 )
⎜Q − Σ Υi (∗) (∗)⎟ µ1 (x2 ) = , µ2 (x2 ) = .
⎝ 0 Λ −H T H (∗)
⎠<0 (69) 1.2172x2 1.2172x2
0 0 βI −I Our objective is to design a polynomial fuzzy filter of the form (4).
As noted above, the designed filter should be as robust as possible
where Υi = (Σ + SC )Ai − Wi C + (Σ + SC )Ai − Wi C
( ) ( )T
−
against the disturbances and as sensitive as possible to the fault.
C T C + Pi − 2 cos(θfl )Q .
For this reason, assume that d and f belong to the low frequency
(2) For f (t) in the middle frequency range θf1 ≤ θf ≤ θf2 :
ranges and let θdl = θfl = 0.1. Then, performing the SOS algorithm
− Pℓ (∗) (∗) (∗) described by Theorem 5, we get γ = 9.9727 × 10−5 , β = 5.572
⎛ ⎞
−jθfc
⎜e Q −Σ Υi (∗) (∗)⎟ and the filter matrices are given by:
⎠<0 (70)
0 Λ −H T H (∗)
⎝
0.46877 −0.46877 0.1
( ) ( )
0 0 βI −I E= , G1 = G2 = ,
0.4819 −0.4819 0.1
where Υi = (Σ + SC )Ai − Wi C + (Σ + SC )Ai − Wi C T − C T C +
( ) ( )
0.46877x22 + 0.35717 −0.84598
( )
Pi − 2cos(θfm )Q , θfc = (θf1 + θf2 )/2 and θfm = (θf2 − θf1 )/2. N1 (x2 ) = ,
(3) For f (t) in the high frequency range |θf | ≥ θfh :
−0.5181x22 − 0.14683 0.13032
0.46877x22 + 0.35667 −0.2759
( )
−Pℓ (∗) (∗) (∗) ,
⎛ ⎞
N2 (x2 ) =
⎜−Q − Σ Υi (∗) (∗)⎟ −0.5181x22 − 0.14734 −0.50083
⎝ 0 Λ −H T H (∗)
⎠<0 (71) ( )
0 0 βI −I −0.21974x22 + 0.61746 0.21974x22 + 0.13696
L1 (x2 ) =
0.24287x22 + 0.39381 −0.24287x2 + 0.38175
2

where Υi = (Σ + SC )Ai − Wi C + (Σ + SC )Ai − Wi C

( ) ( )T
−
C T C + Pi + 2cos(θfh )Q .
( )
−0.21974x22 + 0.34347 0.21974x22 + 0.41195
L2 (x2 ) = .
0.24287x22 + 0.69871 −0.24287x2 + 0.077877
2
with Λ = (Σ + SC )Fi − Wi H
( )T
− HT C.
50 A. Chibani et al. / Automatica 93 (2018) 42–54

Fig. 1. Unknown disturbance. Fig. 3. State estimation errors x − x̂.

Fig. 2. Fault signal. Fig. 4. Residual output r1 .

To analyze the effectiveness of the designed filter, the unknown 4.2.1. Polynomial approach
disturbance and the fault signal are given in Figs. 1 and 2. The dynamic of the Tunnel Diode Circuit is characterized by
Fig. 3 shows the dynamic of the estimation error ek = xk − x̂k .
It can be concluded that the state estimation for the system with iD (t) = 0.002vD (t) + 0.01vD3 (t). (73)
unknown disturbance is performed as well. On the other hand, the Letting x1 = vC and x2 = iL to be the system’s states, then the
generated residual outputs are depicted in Figs. 4 and 5 and show dynamic of the system is governed by Ding and Yang (2010):
that the fault detection aim has been successfully achieved by the
⎨C ẋ1 (t) = −0.002x1 (t) − 0.01x1 (t) + x2 (t)
3
⎧
designed polynomial filter. Indeed, the occurrence of faults can be ⎪
easily detected and the disturbance has been well discriminated Lẋ2 (t) = −x1 (t) − Rx2 (t)
(74)
from the fault. ⎩y1 (t) = x1 (t)
⎪
y2 (t) = x2 (t).
Taking the parameters C = 20 mF, L = 1000 mH and R = 1 , the
4.2. Example 2: Tunnel Diode Circuit
circuit is modeled by

⎨ẋ1 (t) = −0.1x1 (t) − 0.5x1 (t) + 50x2 (t)

3
⎧
In this example, the proposed method will be validated through ⎪
a real process, namely Tunnel Diode Circuit and given by Fig. 6. ẋ2 (t) = −x1 (t) − x2 (t)
(75)
Firstly, a discrete-time polynomial model will be established for ⎩y1 (t) = x1 (t)
⎪
y2 (t) = x2 (t).
the addressed system and a polynomial filter will be designed on
finite frequency domain via the SOS conditions. Then, the dynamic Fixing T = 0.02 to be the step of discretization, a discrete-time
of the studied system will be presented by a T–S fuzzy system. In model for (75) can be obtained by the Euler’s discretization method
this case, a T–S fuzzy filter will be constructed to deal with the as follow
fault detection problem in finite frequency domain using the LMI = x1k + T (−0.1x1k − 0.5x31k + 50x2k )
⎧
x
technique. A comparison study will be made to show the advantage ⎨ 1k+1
⎪
x2k+1 = x2k + T (−x1k − x2k )
of the polynomial approach and the SOS method against the T–S (76)
⎩y1k = x1k
⎪
fuzzy approach and the LMI technique. y2k = x2k .
 A. Chibani et al. / Automatica 93 (2018) 42–54 51

As described in Remark 5, our objective is to minimize the H∞

index and to maximize the H− index. To make comparison, the
optimal value obtained for γ and β are summarized in Table 1.

4.2.2. T–S fuzzy approach

In this case, a T–S fuzzy representation for the Tunnel Diode
Circuit will be given. Indeed, assuming that |x1k | ≤ 3, the following
T–S fuzzy model can be derived from polynomial system (77)
⎧ 2
∑
µi (x1k )(Ai xk + Bi uk + Ri dk + Fi fk )
⎪
xk+1 =
⎨
(78)
⎪ i=1
⎩
yk = Cxk + Ddk + Hfk
where (
0.998 0.908
) ( ) ()
1 1 0
A1 = −0.02 0.98
, A2 = −0.02 0.98
, B1 = B2 = 0 ,
0.2 0.1 0.01
( ) ( ) ( ) ( )
1 0
R1 = R2 = 0.02
, F1 = F2 = 0.1
,C = 0 1
, D = 0.01 ,
()
Fig. 5. Residual output r2 . 1
H= 1
,
with
⎧
x1k + 3
⎨ 3 , for − 3 ≤ x1k ≤ 0
⎪
⎪
⎪
µ1 (x1k ) = 3 − x1k
⎪ , for 0 ≤ x1k ≤ 3
⎩ 3
⎪
⎪
0, for x1k < −3 and x1k > 3
and µ2 (x1k ) = 1 − µ1 (x1k ).
In this case, a multi-objective H− /H∞ filter is designed by
solving the LMI problem according to Corollary 1 and)the fil-
0.6481 0.2085
(
ter’s gain matrices are given by: N1 = 0.2113 0.6472
, N2 =
0.6541 0.2146 −0.0419 0.0046
( ) () ( )
0
0.2174 0.6534
, G1 = G2 = , L1 = −0.0415 0.0025
,
(0
−0.0281 −0.0092 −1.1356 1.1356
( ) )
L2 = −0.0277 −0.0113 and E = −0.1353 0.1353
.
Fig. 6. Tunnel-Diode circuit.
4.2.3. Results and discussions
We compare between the T–S fuzzy approach and the polyno-
mial approach. The obtained results for the two cases are included
Note that (76) can be regarded as a polynomial fuzzy representa-
in Table 1.
tion (2) with r = 1. Indeed, adding unknown disturbances and
Firstly, it can be easily seen that the polynomial representation
faults, system (76) can be rewritten as:
{ is able to present the real nonlinear process in a wider domain
xk+1 = A(y1 )xk + Buk + Rdk + Ffk than the T–S fuzzy modeling. This fact shows the superiority of
(77) PFS than the T–S fuzzy one. In addition, the obtained results for
yk = Cxk + Ddk + Hfk
the optimal values of γ and β are significantly enhanced via the
−0.01y21 + 0.998 0.02
( ) () ( )
1 0 polynomial method than the T–S fuzzy method. Indeed, since our
with: A(y1 ) = −)0.02
, B = , R = 0.2
,
)0.98 0
objective in the FD task was to simultaneously minimize γ and
0.1 0.01
( ) ( ( ()
1 0 1
F = 0.1
,C = 0 1
,D= 0.01
,H = 1
. maximize β , one can see that the value of γ obtained by the
According to Theorem 5, the fault detection filter design prob- SOS approach is more smaller than that obtained by the LMI ap-
lem has been converted into an optimization algorithm under proach. Furthermore, the value of β is clearly increased by the SOS
SOS conditions with equality constraints in which the external method. Hence, one can note that the proposed polynomial filter
disturbance d and the fault signal f are assumed to belong to the significantly improves the fault detection procedure for nonlinear
low frequency range (i.e. |θd | < θdl = 0.1 and |θf | < θfl = systems.
0.5). Solving the SOS algorithm described in Remark 5, we get the In order to check the efficiency of our proposed fault detection
following filter’s parameters: filter, unknown disturbance and fault signals are given by Figs. 7
and 8 respectively. System’s states and their estimates are shown
−0.014893y21 + 0.45728 −0.029069 in Figs. 9 and 10 with initial conditions x(0) = [0.01 0.01]T ,
( )
N(y1 ) = , x̂(0) = [0 0]T . It can be seen that the system’s states are well
−4.2619 × 10−5 y21 − 0.027352 0.9684
reconstituted by the proposed polynomial filter despite that the
0.48933 −0.48933
( ) ( )
0
G(y1 ) = ,E= , unknown disturbance exists. Figs. 11 and 12 show the residual
0 0.0042619 −0.0042619 outputs r1 (t) and r2 (t) generated for T–S fuzzy model (in blue
( ) lines) and for polynomial model (in red lines). They show that
L1 L2 both T–S fuzzy and polynomial filters successfully achieve the fault
L(y1 ) = , detection aim. Indeed, the residual outputs are firstly almost zero
L3 L4
which means that they are robust against disturbance. Once the
with L1 = 0.0072877y21 + 0.81522, L2 = −0.0072877y21 + 1.2625, fault appears, the residual signals become so important so they
L2 = 2.0854 × 10−5 y21 + 0.020947, L2 = −2.0854 × 10−5 y21 + easily detect the fault and the disturbances are well discriminated
0.0024334. from the fault. In the faulty case, one can see that the generated
52 A. Chibani et al. / Automatica 93 (2018) 42–54

Table 1
Obtained results for polynomial and T–S approaches.
Domain of validity of the model γ β
∞ ≤ x1 ≤ ∞
Polynomial approach 3.0482 × 10 −6
8.0715
∞ ≤ x2 ≤ ∞
−3 ≤ x1 ≤ 3
T–S fuzzy approach 0.0143 1.4142
∞ ≤ x2 ≤ ∞

Fig. 7. Unknown disturbance. Fig. 9. Real x1 and estimated x1 .

Fig. 8. Fault signal. Fig. 10. Real x2 and estimated x2 .

residuals by the polynomial filter are more sensitive than those

generated by the T–S fuzzy filter. Hence, we can conclude that
the proposed polynomial approach significantly improves the fault
detection task for nonlinear processes.

5. Conclusions

In this study, a polynomial fuzzy filter is designed in finite

frequency domain to deal with the fault detection problem for
polynomial fuzzy systems. Firstly, new conditions for the H− and
the H∞ performances have been introduced for polynomial struc-
ture using the GKYP lemma. Then, SOS design conditions for the
existence of a polynomial fuzzy filter have been presented. It has
been shown that the finite frequency conditions contain the full
frequency ones as a particular case. Moreover, it has been illus-
trated that the proposed SOS conditions are more general than the
LMI conditions for T–S fuzzy systems. Two examples have been Fig. 11. Residual output r1 . (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)
introduced to validate the developed method. Finally, we note that
 A. Chibani et al. / Automatica 93 (2018) 42–54 53

Liu, Fuqiang, Huang, Ji, Shi, Yang, & Xu, Demin (2013). Fault detection for discrete-
time systems with randomly occurring nonlinearity and data missing: a quadro-
tor vehicle example. Journal of the Franklin Institute, 350(9), 2474–2493.
Liu, Jian, Wang, Jian Liang, & Yang, Guang-Hong (2005). An LMI approach to mini-
mum sensitivity analysis with application to fault detection. Automatica, 41(11),
1995–2004.
Parrilo, Pablo A. (2000). Structured semidefinite programs and semialgebraic geometry
methods in robustness and optimization. (Ph.D. thesis), Citeseer.
Prajna, Stephen, Papachristodoulou, Antonis, & Wu, Fen (2004). Nonlinear con-
trol synthesis by sum of squares optimization: A lyapunov-based approach.
In Control conference, 2004 5th Asian, Vol. 1 (pp. 157–165). IEEE.
Takagi, Tomohiro, & Sugeno, Michio (1985). Fuzzy identification of systems and its
applications to modeling and control. IEEE Transactions on Systems, Man, and
Cybernetics,(1), 116–132.
Tanaka, Kazuo, Ohtake, Hiroshi, Seo, Toshiaki, Tanaka, Motoyasu, & Wang, Hua O.
(2012). Polynomial fuzzy observer designs: a sum-of-squares approach. IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 42(5),
1330–1342.
Tanaka, Kazuo, Yoshida, Hiroto, Ohtake, Hiroshi, & Wang, Hua O. (2009). A sum-
of-squares approach to modeling and control of nonlinear dynamical systems
with polynomial fuzzy systems. IEEE Transactions on Fuzzy Systems, 17(4),
Fig. 12. Residual output r2 . (For interpretation of the references to color in this 911–922.
figure legend, the reader is referred to the web version of this article.) Tuan, Hoang Duong, Apkarian, Pierre, Narikiyo, Tatsuo, & Yamamoto, Yasuhiro
(2001). Parameterized linear matrix inequality techniques in fuzzy control
system design. IEEE Transactions on Fuzzy Systems, 9(2), 324–332.
Wang, Heng, Ju, He-Hua, & Wang, Yu-Long (2013). H∞ switching filter design for lpv
this work is still open to many possible extensions in the future. systems in finite frequency domain. International Journal of Control, Automation
In particular, we intend to consider the case of polynomial fuzzy and Systems, 11(3), 503–510.
systems with unmeasurable premise variables and to investigate Wang, Yu-Long, Shi, Peng, Lim, Cheng-Chew, & Liu, Yuan (2016). Event-triggered
fault detection filter design for a continuous-time networked control system.
the problem of fault tolerant control of polynomial fuzzy systems
IEEE Transactions on Cybernetics, 46(12), 3414–3426.
in finite frequency domain. Wang, Dong, Shi, Peng, & Wang, Wei (2010). Robust fault detection for continuous-
time switched delay systems: an linear matrix inequality approach. IET Control
References Theory & Applications, 4(1), 100–108.
Wang, Heng, & Yang, Guang-Hong (2008). A finite frequency domain approach to
Chadli, Mohammed, Abdo, Ali, & Ding, Steven X. (2013). H− /H∞ fault detection fault detection for linear discrete-time systems. International Journal of Control,
filter design for discrete-time Takagi–Sugeno fuzzy system. Automatica, 49(7), 81(7), 1162–1171.
1996–2005. Wang, Jian Liang, Yang, Guang-Hong, & Liu, Jian (2007). An LMI approach to H−
Chadli, Mohammed, & Karimi, Hamid Reza (2013). Robust observer design for index and mixed H− /H∞ fault detection observer design. Automatica, 43(9),
unknown inputs Takagi-Sugeno models. IEEE Transactions on Fuzzy Systems, 1656–1665.
21(1), 158–164. Yang, Rongni, Shi, Peng, & Liu, Guo-Ping (2011). Filtering for discrete-time net-
Chen, Ying-Jen, Tanaka, Motoyasu, Tanaka, Kazuo, Ohtake, Hiroshi, & Wang, Hua O. worked nonlinear systems with mixed random delays and packet dropouts. IEEE
(2014). Brief paper-discrete polynomial fuzzy systems control. IET Control The- Transactions on Automatic Control, 56(11), 2655–2660.
ory & Applications, 8(4), 288–296. Yang, Guang-Hong, & Wang, Heng (2010). Fault detection for a class of uncertain
Chen, Ying-Jen, Tanaka, Motoyasu, Tanaka, Kazuo, & Wang, Hua O. (2015). Stability state-feedback control systems. IEEE Transactions on Control Systems Technology,
analysis and region-of-attraction estimation using piecewise polynomial lya- 18(1), 201–212.
punov functions: polynomial fuzzy model approach. IEEE Transactions on Fuzzy Yang, Hongjiu, Xia, Yuanqing, & Liu, Bo (2011). Fault detection for T–S fuzzy discrete
Systems, 23(4), 1314–1322. systems in finite-frequency domain. IEEE Transactions on Systems, Man, and
Chibani, Ali, Chadli, Mohammed, & Braiek, Naceur Benhadj (2016a). A sum of Cybernetics, Part B (Cybernetics), 41(4), 911–920.
squares approach for polynomial fuzzy observer design for polynomial fuzzy Yang, Hongjiu, Xia, Yuanqing, & Zhang, Jinhui (2011). Generalised finite-frequency
systems with unknown inputs. International Journal of Control. Automation and KYP lemma in delta domain and applications to fault detection. International
Systems, 14(1), 323–330. Journal of Control, 84(3), 511–525.
Chibani, Ali, Chadli, Mohammed, Shi, Peng, & Braiek, Naceur Benhadj (2016b). Fuzzy Yoneyama, Jun (2009). H∞ filtering for fuzzy systems with immeasurable premise
fault detection filter design for ts fuzzy systems in finite frequency domain. IEEE variables: an uncertain system approach. Fuzzy Sets and Systems, 160(12),
Transactions on Fuzzy Systems. 1738–1748.
Ding, Da-Wei, & Yang, Guang-Hong (2010). Fuzzy filter design for nonlinear systems Zhang, Ke, Jiang, Bin, Shi, Peng, & Xu, Jinfa (2015a). Analysis and design of robust H∞
in finite-frequency domain. IEEE Transactions on Fuzzy Systems, 18(5), 935–945. fault estimation observer with finite-frequency specifications for discrete-time
Gahinet, Pascal, & Apkarian, Pierre (1994). A linear matrix inequality approach to fuzzy systems. IEEE Transactions on Cybernetics, 45(7), 1225–1235.
H∞ control. International Journal of Robust and Nonlinear Control, 4(4), 421–448. Zhang, Ke, Jiang, Bin, Shi, Peng, & Xu, Jingfa (2015b). Fault estimation observer de-
Gao, Huijun, Zhao, Yan, Lam, James, & Chen, Ke (2009). Fuzzy filtering of nonlinear sign for discrete-time systems in finite-frequency domain. International Journal
systems with intermittent measurements. IEEE Transactions on Fuzzy Systems, of Robust and Nonlinear Control, 25(9), 1379–1398.
17(2), 291–300. Zhang, Hui, Shi, Yang, & Mehr, Aryan Saadat (2012). On H∞ filtering for discrete-
Iwasaki, Tetsuya, & Hara, Shinji (2005). Generalized kyp lemma: unified frequency time Takagi–Sugeno fuzzy systems. IEEE Transactions on Fuzzy Systems, 20(2),
domain inequalities with design applications. IEEE Transactions on Automatic 396–401.
Control, 50(1), 41–59. Zhang, Hui, & Wang, Junmin (2016a). Active steering actuator fault detection for
Lam, Hak-Keung, Liu, Chuang, Wu, Ligang, & Zhao, Xudong (2015). Polynomial an automatically-steered electric ground vehicle. IEEE Transactions on Vehicular
fuzzy-model-based control systems: Stability analysis via approximated mem- Technology.
bership functions considering sector nonlinearity of control input. IEEE Trans- Zhang, Hui, & Wang, Junmin (2016b). Adaptive sliding-mode observer design
actions on Fuzzy Systems, 23(6), 2202–2214. for a selective catalytic reduction system of ground-vehicle diesel engines.
Lam, Hak-Keung, Wu, Ligang, & Lam, James (2015). Two-step stability analysis for IEEE/ASME Transactions on Mechatronics, 21(4), 2027–2038.
general polynomial-fuzzy-model-based control systems. IEEE Transactions on Zhang, Sunjie, Wang, Zidong, Ding, Derui, Shu, Huisheng, Hayat, Tasawar, & Dobaie,
Fuzzy Systems, 23(3), 511–524. Abdullah M. (2015). On design of robust fault detection filter in finite-frequency
Li, Fanbiao, Shi, Peng, Lim, Cheng-Chew, & Wu, Ligang (2016). Fault detection domain with regional pole assignment. IEEE Transactions on Circuits and Systems
filtering for nonhomogeneous markovian jump systems via fuzzy approach. II: Express Briefs, 62(4), 382–386.
IEEE Transactions on Fuzzy Systems. Zhong, Maiying, Ding, Steven X., Lam, James, & Wang, Haibo (2003). An LMI
Li, Xiao-Jian, & Yang, Guang-Hong (2015). Fault detection in finite frequency do- approach to design robust fault detection filter for uncertain lti systems.
mains for multi-delay uncertain systems with application to ground vehicle. Automatica, 39(3), 543–550.
International Journal of Robust and Nonlinear Control, 25(18), 3780–3798. Zhong, Guang-Xin, & Yang, Guang-Hong (2015). Fault detection for discrete-time
Li, Wei, Zhu, Zhencai, & Ding, Steven X. (2011). Fault detection design of networked switched systems in finite-frequency domain. Circuits, Systems, and Signal Pro-
control systems. IET Control Theory & Applications, 5(12), 1439–1449. cessing, 34(4), 1305–1324.
54 A. Chibani et al. / Automatica 93 (2018) 42–54

Ali Chibani was born in Gafsa, Tunisia in 1988. He received Steven X. Ding received Dr.-Ing. degree in electrical en-
his B.S. and M.S. degree in electrical engineering in 2010 gineering from the Gerhard-Mercator University of Duis-
and 2012, respectively from the High School of Sciences burg, Germany, in 1992. From 1992 to 1994, he was a
and Techniques of Tunis. He obtained his Ph.D. Thesis in R&D engineer at Rheinmetall GmbH, Germany. From 1995
electrical engineering in 2016. Currently, he is a research to 2001, he was a full-professor of control engineering at
member of the Advanced Systems Laboratory (LSA) in the the University of Applied Science Lausitz in Senftenberg,
Polytechnic School of Tunisia. His research interests in- Germany, and served as a vice president of this university
clude robust control, state estimation and FDI of nonlinear during 1998–2000. Since 2001, he has been a full-
systems, LPV and fuzzy systems. professor of control engineering and the head of the In-
stitute for Automatic Control and Complex Systems (AKS)
at the University of Duisburg–Essen. His research interests
are model-based and data-driven fault diagnosis, fault tolerant systems and their
applications in industry with a focus on automotive systems, chemical processes
Mohammed Chadli received the M.Sc. from ‘‘Ecole Nor- and renewable energy systems.
male Sup’’ Mohammedia-Morocco (1993) and (DEA) from
Engineering School INSA-Lyon (1999), the Ph.D. thesis
from the CRAN-Nancy France in 2002 and his Habilitation
in 2011 at the University of Picardie Jules Verne (UPJV) Naceur Benhadj Braiek was born in Mahdia Tunisia. He
in Amiens, France. From 1999 to 2004, he was an As- obtained the engineer degree in Electrical Engineering,
sistant Prof at University of Lorraine (CRAN-Institut Na- and the master degree in Systems Analysis and Signal
tional Polytechnique de Lorraine). Since 2004, he has been Processing, both from National High School of Engineers
Associate Professor at the UPJV-MIS France. He was a vis- of Tunis in 1987, and the Doctorate degree (Ph.D.) in
iting professorship at the Technical University of Ostrava- Automatic Control from the University of Sciences and
Czech Rep., University of Agder-Norway and University of Technologies of Lille-France in 1990, and the State Doc-
Shanghai-Chinn. torate of Sciences (Doctorat d’Etat ès Sciences) in Electrical
His research interests include fuzzy/LPV systems, singular systems, robust con- Engineering from the National High School of Engineers of
trol, fault detection and isolation (FDI), fault tolerant control (FTC) via LMI and SOS Tunis – University of Tunis – in 1995.
optimization techniques, and their applications (automotive control, renewable He is Professor of Automatic Control at the University
energy, . . . ). He is author/co-author of 4 books, book chapters and numerous articles of Tunis and the Head of research Laboratory ‘‘Laboratory of Advanced Systems’’ at
published in international journals and conferences. the Tunisia Polytechnic High School-University of Carthage. His domain of interest
Dr. Chadli is a senior member of IEEE, he has actively served in the editorial includes nonlinear systems modeling, control and optimization in both theoretical
board of a number of journals, including IEEE Transactions on Fuzzy Systems, developments and applications. He has supervised more than 80 Master thesis and
IEEE/CAA Automatica Sinica, Journal of Franklin Institute, IET Control Theory and 30 Ph.D. thesis which have been successfully defended. His main contributions in
Applications, Asian Journal of Control, . . . and was a Guest Editor for Special Issues the field of nonlinear systems concern the development of algebraic and analytic
in international journals. tools for the modeling, analysis and control of polynomial systems.