International Journal of Systems Science

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tsys20

Actuator fault estimation based on a proportional-

integral observer with nonquadratic Lyapunov
functions

J. Guzman , F.-R. López-Estrada , V. Estrada-Manzo & G. Valencia-Palomo

To cite this article: J. Guzman , F.-R. López-Estrada , V. Estrada-Manzo & G. Valencia-

Palomo (2021): Actuator fault estimation based on a proportional-integral observer
with nonquadratic Lyapunov functions, International Journal of Systems Science, DOI:
10.1080/00207721.2021.1873451

To link to this article: https://doi.org/10.1080/00207721.2021.1873451

Published online: 20 Jan 2021.

Submit your article to this journal

Article views: 45

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tsys20
INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE
https://doi.org/10.1080/00207721.2021.1873451

Actuator fault estimation based on a proportional-integral observer with

nonquadratic Lyapunov functions
J. Guzmana , F.-R. López-Estrada a , V. Estrada-Manzo b and G. Valencia-Palomo c

a Tecnológico Nacional de México, IT Tuxtla Gutiérrez, TURIX-Dynamics Diagnosis and Control Group, Tuxtla Gutierrez, Mexico; b Dept. of
Mechatronics, Universidad Politécnica de Pachuca, Zempoala, México; c Tecnológico Nacional de México, IT Hermosillo, Hermosillo, Mexico

ABSTRACT ARTICLE HISTORY

This work proposes a proportional-integral fault estimation observer for nonlinear systems by means Received 3 September 2020
of convex structures. Unlike the traditional approaches that consider quadratic Lyapunov functions, Accepted 31 December 2020
this work employs nonquadratic ones. As a result, less conservative conditions in the form of linear KEYWORDS
matrix inequalities are obtained. Furthermore, an H∞ performance is employed in order to ensure Actuator fault-estimation;
robustness against sensor noise and disturbances; thus, the designed observer estimates both the proportional-integral
state and the actuator faults even in the presence of disturbances. The performance of the proposed observer; linear matrix
methodology is illustrated through a numerical example and a three-tank hydraulic system. inequality; convex model;
nonquadratic Lyapunov
function

1. Introduction
(Takagi & Sugeno, 1985) or linear parameter varying
In recent years, safety and performance requirements (LPV) systems (Shamma, 2012). They are called con-
in industrial processes have motivated the develop- vex models because they are a collection of linear sub-
ment of fault diagnosis systems. A fault can be seen models (vertex models) that are combined by convex
as an undesired change in the value of the ele- scalar functions (Tanaka & Wang, 2001), also known as
ments or parameters of a system, above or below the scheduling functions. Moreover, if a TS model is com-
limits considered normal (Huimin & Yafang, 2007). puted via the sector nonlinearity approach (Ohtake
Implementing strategies to detect and correct these et al., 2001), then the resulting convex model is an
abnormal behaviours in time has become pertinent exact representation of the original nonlinear one. For
and necessary to guarantee the process performance. instance, Rodrigues et al. (2014) present an observer
In the literature, there are fault detection meth- for descriptor systems and the work by Aouaouda
ods based on mathematical models (Ahmadizadeh et al. (2015) considers a robust fault detection scheme
et al., 2014; Cheng et al., 2020; Kladis et al., 2016; based-on TS models. Su et al. (2016) propose a
Tong et al., 2011) and data (Pérez-Pérez et al., 2021; fault-based detection for residual generator applied
Santos-Ruiz et al., 2018; Yin et al., 2014). This work is to stochastic systems (Chitraganti et al., 2014). In
dedicated to model-based fault detection methods. Zhong et al. (2017), a robust fractional-order observer
Model-based fault detection systems consist of is designed. A recent survey regarding fault detec-
the creation of analytical redundancy, which is tion methods for convex TS and LPV systems can be
achieved by comparing the system’s nominal oper- consulted in López-Estrada et al. (2019).
ation with the expected performance given by a Nonetheless, most of the works are focused on the
reference model (Puig et al., 2004). A popular stage of detection and isolation employing observer
approach to obtain this redundancy is based on state banks (Castro et al., 2016) or smart classifiers such
observers (Verde et al., 2013). This work adopts as neural networks (Ince et al., 2016) or fuzzy logic
a convex approach for the observer design. Con- (Dehghani et al., 2016). Less work is committed to
vex approaches include Takagi-Sugeno (TS) models estimating the fault magnitude, for example, by mean

CONTACT F.-R. López-Estrada frlopez@ittg.edu.mx, Tecnológico Nacional de México, IT Tuxtla Gutiérrez, TURIX-Dynamics Diagnosis and Control
Group, Carretera Panamericana km 1080 S/N, Tuxtla Gutierrez 29050, Mexico

© 2021 Informa UK Limited, trading as Taylor & Francis Group

2 J. GUZMAN ET AL.

of unknown input estimation for nonlinear descriptor sufficient LMI conditions for calculating the observer
systems (Estrada-Manzo et al., 2015), unknown input gain; also, an observer is presented from an approach
observers (Zhou et al., 2017), proportional-integral that considers disturbances and sensor noise. In
(PI) observers (Nazari & Shafai, 2018), among others. Section 4, the proposed method is applied to an
Regarding PI observers, in Chang (2006), an observer illustrative example taken from the literature and a
was proposed to estimate the charge on lithium-ion three-tank setup. Finally, the conclusions are given in
batteries. A PI observer used for robust fault detection Section 5.
for linear systems is presented in Duan and Wu (2006). Notation: In this paper, AT ∈ Rn×n denotes the
Hassanabadi et al. (2017) proposed a PI observer that transpose of A. The expression A > 0 means that A
includes the detection of faults in both actuators and is positive definite, and A < 0 means that A is nega-
sensors. Moreover, in Tlili (2019), a robust PI observer tive definite. The operator H(A) will be employed to
was proposed to attenuate disturbances and noise by denote A + AT .
considering the H∞ approach.
Notwithstanding, the above solutions are based
2. Preliminaries
on sufficient conditions in the form of linear matrix
inequalities (LMIs), which are obtained by consider- An exact TS model can be obtained from a nonlinear
ing quadratic Lyapunov functions. LMI solutions are system by the well-known nonlinear sector approach
preferred because they can be efficiently solved via proposed in Ohtake et al. (2001). This approach is the
convex optimisation techniques already available in most used because the resulting TS model is an exact
commercial software (Boyd et al., 1994). However, representation of the nonlinear system. For instance,
the main drawback of the quadratic approach is that consider a nonlinear model of the form:
it requires the computation of a single matrix P > 0
for all the vertex models, thus leading to conserva- ẋ(t) = f (z(t))x(t) + g(z(t))u(t), (1)
tive solutions (Bernal & Guerra, 2010). Works such as
Bernal et al. (2011) and Guerra et al. (2012) use non- where x ∈ Rn is the state vector, u ∈ Rm the input vec-
quadratic Lyapunov functions; these families of func- tor, and z ∈ Rp is the premise (scheduling) vector; f (·)
tions have been used for the stabilisation problem in and g(·) are assumed to be smooth and bounded. The
Vafamand and Sadeghi (2015). For instance, in Cherifi scheduling variables are bounded as zj (·) ∈ [zj , zj ],
et al. (2019) a D-stabilisation of TS models involving j ∈ {1, 2, . . . , p}, where zj and zj are the minimum and
average values of membership functions has been pre- maximum values of zj (·) in a compact set of the state
sented. In the case of observers, Wang et al. (2016) space  ⊂ Rn , which includes the origin. Hence, the
employ a matrix decoupling technique together with following membership functions are constructed (also
the nonquadratic approach. known as weighting functions):
This work proposes a methodology for actuator
j zj − zj (·) j j
fault detection and estimation by considering a PI w0 (·) = , w1 (·) = 1 − w0 (·), (2)
observer for TS systems; it is based on a nonquadratic zj − zj
Lyapunov function. In order to guarantee robust-
ness against disturbances and sensor noise, the H∞ to define scheduling functions as
approach is adopted as a performance criterion. If the p
 j
given LMI conditions are feasible, the observer con- hi (z) = wij (zj ), (3)
vergence and the actuator fault estimation are guar- j=1
anteed. It is essential to mention that the proposed
approach provides less conservative solutions with where i ∈ {1, 2, . . . , p}, ij ∈ {0, 1}. The scheduling
respect to a quadratic Lyapunov function. Some simu- functions hi (z) satisfy the convex sum property in

lation examples are proposed to illustrate the method’s : ri=1 hi (·) = 1, hi ≥ 0, and r = 2p is the number
effectiveness. of vertex (linear) models. In this work, the premise
The rest of the document is organised as follows: vector depends only on measurable states (Gómez-
Section 2 presents the TS systems concept. Section 3 Peñate et al., 2019), the general case z(x̂, x) has been
describes the PI structure and the derivation of recently considered by Quintana et al. (2020) and it is
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 3

out of the scope of this research. Finally, an exact TS Thus, (5) can be expressed as:
representation of (1) is obtained: ⎛ ⎞


r    r ⎜     ⎟
ẋ(t) ⎜ Ai Mi x(t) Bi ⎟
ẋ(t) = hi (z(t))(Ai x(t) + Bi u(t)), (4) = ⎜
hi ⎜ + u(t)⎟
⎟
i=1 ḟa (t)
  i=1 ⎝ 0  0  fa(t) 0
 ⎠
x̄˙ Āi x̄ B̄i
where the constant matrices Ai ∈ Rn×n and Bi ∈ ⎛ ⎞
Rn×m are computed by substituting the correspond-  ⎟

r ⎜
ing elements to the weighting functions such that the ⎜ x(t) ⎟
y(t) = hi ⎜[Ci 0] ⎟.
ith-vertex is active. For convenience, in the rest of ⎝  fa (t) ⎠
 i=1  
the paper, ri=1 hi is used as shorthand notation of C̄i
r x̄
i=1 hi (z). (7)
In what follows, the following TS model with actu- Then, in order to estimate both states and actuator
ator faults is considered: faults, the following proportional-integral TS observer
is proposed:

r
ẋ(t) = hi (Ai x(t) + Bi u(t) + Mi fa (t)) ⎛
i=1 ˙ˆ 
r
(5) x̄(t) = ˆ + B̄i u(t)
hi ⎝Āi x̄(t)
r
y(t) = hi Ci x(t), i=1

i=1
⎛ ⎞−1 ⎞
r
+⎝ hj P̄j ⎠ L̄i (y(t) − ŷ(t))⎠ , (8)
where fa (t) ∈ Rnf represents the actuator faults, y(t) ∈
j=1
Ro the measured outputs, and Mi , Ci are constant
matrices of proper size. 
ˆ
where x̄(t) ˆ
and ŷ(t) = ri=1 hi (C̄i x̄(t)) are the esti-
Generally, in the context of TS models, the design
mated state and output vectors, respectively. P̄j ∈
conditions are given in terms of LMIs. To this end, the
R(n+nf )×(n+nf ) and L̄j ∈ R(n+nf )×o , j ∈ {1, 2, . . . , r}
following relaxation lemma will be employed (Tuan
are constant matrices that constitute the nonlinear
et al., 2001):
observer gains.
The estimation error between system (7) and the
Lemma 2.1: Let ϒij = ϒijT , (i, j) ∈ {1, 2, . . . , r}2 be
observer (8) is given by:
matrices of adequate dimensions. Then,
ˆ
ē(t) = x̄(t) − x̄(t), (9)

r 
r
hi hj ϒij < 0
i=1 j=1
whose dynamics can be expressed as:

˙ = x̄(t) ˙ˆ
˙ − x̄(t).
holds if ē(t) (10)
2
ϒii + ϒij + ϒji < 0, (6) Then, by considering (7) and (8), the error dynamics
r−1 can be expressed as:
for all (i, j) ∈ {1, 2, . . . , r}2 .

r
 
˙ =
ē(t) hi Āi x̄(t) + B̄i u(t) ,
3. Proportional-integral observer i=1
⎛ ⎛ ⎞−1
The PI observer allows simultaneous estimation of 
r 
r
− ˆ + B̄i u(t) + ⎝
hi ⎝Āi x̄(t) hj P̄j ⎠
system states and unknown inputs. In this case, the
i=1 j=1
unknown inputs are actuator faults affecting the TS
⎞
model under the assumption that ḟa (t) ≈ 0. This con-
dition is not restrictive, and it can be relaxed in prac- × L̄i (y(t) − ŷ(t))⎠ ,
tice (Chadli et al., 2013; López-Estrada et al., 2019).
4 J. GUZMAN ET AL.

⎛ the system states. Moreover, dealing with the time-


r 
r
= hi ˆ
hj ⎝Āi x̄(t) − Āi x̄(t) derivative could involve the state x̄(t) and the dynam-
i=1 j=1 ics of the system (e.g. Bernal & Guerra, 2010; Cherifi
⎛ ⎞−1 ⎞ et al., 2019; Guerra et al., 2012), thus obtaining a con-

r
vex form (LMI-based conditions) for V̇(ē) is more
−⎝ hj P̄j ⎠ ˆ
L̄j C̄i (x̄(t) − x̄(t))⎠,
difficult. In this work, we adopt the approach proposed
j=1
by Tanaka et al. (2003) with the assumption:
⎛

r 
r
= hi ˆ
hj ⎝Āi (x̄(t) − x̄(t)) |ḣk | ≤ φk , (13)
i=1 j=1
⎛ ⎞−1 ⎞ where φk is a non-negative scalar. In practice, the val-
r ues of φk are proposed from a priori simulations of
−⎝ hj P̄j ⎠ L̄j C̄i ē(t)⎠ . time-derivatives of the membership functions.
j=1 Then, the following result provides LMI conditions
to ensure that the error tends asymptotically to zero,
Finally, the following expression is obtained: i.e. limt→∞ ē(t) ≈ 0.
⎛ ⎛ ⎞−1 ⎞

r 
r 
r Theorem 3.1: The origin ē = 0 of the estimation
˙ =
ē(t) hi hj ⎝Āi − ⎝ hj P̄j ⎠ L̄j C̄i ⎠ ē(t). error (11) is asymptotically stable for given scalars φk ,
i=1 j=1 j=1 k ∈ {1, 2, . . . , r − 1} if there exist matrices P̄j = P̄jT >
(11) 0 and L̄j , ∀i, j ∈ {1, 2, . . . , r} such that P̄k ≥ P̄r , k ∈
The convergence of the observer (8) to system (7) is {1, 2, . . . , r − 1} and the LMIs :
guaranteed if the origin ē = 0 in (11) is asymptotically
stable. This problem has been extensively addressed in 2
ϒii + ϒij + ϒji < 0, (14)
the literature by means of a quadratic Lyapunov func- r−1
tion. Nevertheless, it has been proved that by consid-
ering nonquadratic functions, less conservatism can hold for all (i, j) ∈ {1, 2, . . . , r}2 with
be obtained (Faria et al., 2013; Mozelli et al., 2009).  
H(P1,j Ai − LP ,j Ci ) (∗)
Then, the following nonquadratic Lyapunov function ϒij := T
P2,j Ai − LI ,j Ci + MiT P1,j H(P2,j
T M)
i
candidate is considered:
⎛ ⎞ 
r−1


r + φk (P̄k − P̄r ).
V(ē(t)) = ē (t) ⎝
T
hj P̄j ⎠ ē(t) > 0, P̄j > 0, k=1
j=1
(12) Proof: Consider the time-derivative of the Lyapunov
which clearly holds V(ē) > 0 for all ē = 0; thus it function candidate (12):
is a valid Lyapunov function candidate. The Lya-

r
punov function (12) can be seen as conventional V̇(ē) = hj (ē˙ T (t)P̄j ē(t) + ēT (t)P̄j ē(t))
˙
quadratic function. However, the problem is not triv- j=1
ial because the matrix P̄j is also convex, whose solution

r
also involves the time-derivative of the membership + ḣk ēT (t)P̄k ē(t), (15)
functions. In fact, the structure of the nonquadratic k=1
Lyapunov function is more or less the same since
2001; what changes is the way the authors deal with which after substituting the dynamics (11) gives:
the inherent problem of the time-derivative of the

r 
r
scheduling (membership) functions; this is why they V̇(ē) = hi hj ēT (t)(H(P̄j Āi − L̄j C̄i ))ē(t)
are mainly employed for stability and controller design i=1 j=1
(Guerra et al., 2015). In the case of observer design,

r
a significant problem arises since the Lyapunov func- + ḣk ēT (t)P̄k ē(t), (16)
tion depends on the observation error V(ē) rather than k=1
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 5

⎞
Considering relaxed properties of time-derivatives of
membership functions expressed by Theorem 2 of   ⎟
Ri ⎟
Tanaka et al. (2003): + d(t)⎟
⎟
0 ⎠


r 
r−1
R̄i
ḣk = 0 =⇒ ḣr = − ḣk ; (17) ⎛ ⎞
k=1 k=1
⎜     ⎟

r
⎜ x(t) E ⎟
equation (15) is rewritten as: ȳ(t) = hi ⎜
⎜[C i 0] + i
d(t)⎟
⎟,
⎝   fa (t) 0 ⎠
i=1
C̄i   

r 
r
x̄ Ēi
V̇(ē) = hi hj ēT (t)(H(P̄j Āi − L̄j C̄i ))ē(t)
i=1 j=1
(20)


r−1 where Ri ∈ Rn×nf and Ei ∈ Ro×nf are the perturbation
+ ḣk ēT (t)(P̄k − P̄r )ē(t). (18) and noise matrices with appropriate dimensions.
k=1 The dynamics of the estimation error ē(t) = x̄(t) −
ˆ between system (20) and the observer (8) is given
x̄(t)
Then, considering (13), V̇(ē) < 0 is guaranteed if
by:

r 
r ⎛⎛ ⎛ ⎞−1 ⎞
hi hj ēT (t) r r r
i=1 j=1
˙ =
ē(t) hi hj ⎝⎝Āi − ⎝ hj P̄j ⎠ L̄j C̄i ⎠ ē(t)
 
r−1  i=1 j=1 j=1
⎛ ⎛ ⎞−1 ⎞ ⎞
× H(P̄j Āi − L̄j C̄i ) + φk (P̄k − P̄r ) 
r

  k=1 + ⎝R̄i − ⎝ hj P̄j ⎠ L̄j Ēi ⎠ d(t)⎠ . (21)

ϒij j=1

× ē(t) < 0 (19) In order to ensure disturbance attenuation, the

observer gain is designed such that:
holds. Finally, by means of the relaxation
 Lemma
 2.1
P1,j P2,j re (t) 2
and considering the structure P̄j = PT P3,j and L̄j = sup ≤ γ, γ > 0, (22)
L  2,j
d(t) 2 =
 0 d(t) 2
P ,j
LI ,j , sufficient LMI conditions (14) yield; thus con-
cluding the proof.  where re (t) is the residual vector defined by re (t) =
y(t) − ŷ(t). Hence, the problem is reduced to find
matrices P̄j and L̄j , j ∈ {1, 2, . . . , r} such that condi-
The following section presents the case when the
tion (22) is guaranteed. As a result the following result
system is under disturbances. This problem is faced via
can be formulated:
the H∞ approach.
Theorem 3.2: The estimation error (21) holds the H∞
3.1. H∞ attenuation criterion (22) with a minimum attenuation index γ > 0
and given scalars φk , k ∈ {1, 2, . . . , r − 1} if there exist
The results above can be extended to consider robust- matrices P̄j = P̄jT > 0 and L̄j , ∀i, j ∈ {1, 2, . . . , r} such
ness to sensor noise and disturbances by considering that P̄k ≥ P̄r , k ∈ {1, 2, . . . , r − 1} and the following
the H∞ approach. Thus, under disturbances d(t) ∈ optimisation problem has solution:
Rnd , sensor noise, and actuator faults, the augmented
system (5) becomes: min γ
Pj ,Lj
⎛
   r ⎜     subject to the LMIs
ẋ(t) ⎜ Ai Mi x(t) B
= ⎜
hi ⎜ + i u(t) 2
ḟa (t)
  ⎝ 0  0  fa(t) 0

ϒii + ϒij + ϒji < 0, ∀(i, j) ∈ {1, 2, . . . , r}2 ,
i=1 r−1
x̄˙ Āi x̄ B̄i (23)
6 J. GUZMAN ET AL.

 
with ē(t)
⎡ r−1 × < 0. (27)
H(P̄j Āi − L̄j C̄i ) + d(t)
k=1 φk (P̄k − P̄r )
⎢
ϒij := ⎣ R̄i P̄j − ĒiT L̄Tj
T Finally, (24) can be expressed as
 T  
C̄i  r  r
ē(t)   ē(t)
⎤ Jrd = hi hj ij + Φij <0
P̄j R̄i − L̄j Ēi C̄iT d(t) d(t)
i=1 j=1
⎥
−γ 2 I ĒiT ⎦. (28)
Ēi −I with:
⎡ ⎤
ĀTi P̄j − C̄iT L̄Tj + P̄j Āi
Proof: To guarantee the asymptotic convergence to ⎢ ⎥
⎢ 
r−1
P̄j R̄i − L̄j Ēi ⎥
zero of the estimation error and robustness against ⎢
ij = ⎢ −L̄j C̄i + φk (P̄k − P̄r ) ⎥,
⎥
disturbances d(t), the following criterion H∞ is con- ⎣ k=1 ⎦
sidered (López-Estrada et al., 2016): R̄Ti P̄j − ĒiT L̄j −γ 2 I

Jrd = V̇(ē) + J1 < 0, (24) and

 T 
J1 = reT (t)re (t) − γ 2 dT (t)d(t) < 0, C̄ C̄ C̄iT Ēi
(25) Φij = Ti i .
Ēi C̄i ĒiT Ēi
where Jrd represents the L2 gain of (23) from d(t) to Recall that the Schur complement establishes
re (t) limited by γ and V̇(ē(t)) is a nonquadratic Lya-  
Z11 Z12 T
punov function candidate (15). Replacing the dynam- T −1
Z11 + Z12 Z22 Z12 < 0 ⇐⇒ < 0;
ics of the error (21) in (15), the following is obtained: Z12 −Z22

 r  r  T therefore, with the following definitions:

ē(t)
V̇(ē) = hi hj  T
d(t) C̄
i=1 j=1 T
⎡ ⎤ Z11 = ij , Z12 = Ti ,
Ēi
ĀTi P̄j − C̄iT L̄Tj + P̄j Āi
⎢ ⎥ Z12 = [C̄i Ēi ], −1
and Z22 = I.
⎢ r−1
P̄j R̄i − L̄j Ēi ⎥
⎢
× ⎢−L̄j C̄i + φk (P̄k − P̄r) ⎥
⎥
⎣ k=1 ⎦ The LMI (28) is equivalent to :
R̄Ti P̄j − ĒiT L̄j 0 
r 
r
  hi hj
ē(t)
× . (26) i=1 j=1
d(t) ⎡ ⎤
H(P̄j Āi − L̄j C̄i )
Replacing the value of re (t) = y(t) − ŷ(t), the follow- ⎢  r−1 ⎥
⎢ P̄j R̄i − L̄j Ēi C̄iT ⎥
ing is obtained: ⎢+ φk (P̄k − P̄r ) ⎥
⎢
×⎢ ⎥ < 0,
⎥

r 
r ⎢ k=1 ⎥
J1 = hi hj ((C̄i ē(t) + Ēi d(t))T ⎣ R̄Ti P̄j − ĒiT L̄Tj −γ 2 I ĒiT ⎦
i=1 j=1 C̄i Ēi −I
 
× (C̄i ē(t) + Ēi d(t)) − γ 2 dT (t)d(t)) ϒij


r 
r
such that P̄k ≥ P̄r . Finally, by considering again the
= hi hj (ēT (t)C̄iT C̄i ē(t) + dT (t)ĒiT C̄i ē(t) relaxation Lemma 2.1 over the double convex sum
i=1 j=1 r r
i=1 j=1 hi hj , sufficient LMI conditions are thus
+ dT (t)ĒiT Ēi d(t) − γ 2 dT (t)d(t)), obtained. 

rewriting the previous expression we get:

4. Examples
 r  r  T  T 
ē(t) C̄i C̄i C̄iT Ēi
J1 = hi hj In this section, two examples are presented for fault
d(t) ĒiT C̄i ĒiT Ēi − γ 2 I
i=1 j=1 estimation. The first one is taken from the literature
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 7

and employs conditions of Theorem 3.1 while and the

second one is a three-tank system under disturbances,
so results of Theorem 3.2 are used.

4.1. Example 1
To show the effectiveness of the proposed method-
ology, the following numerical example is borrowed
from Ichalal et al. (2014), that is the TS model (5), with
vertex matrices:
   
1.59 −7.29 −α −4.33
A1 = , A2 = ,
0.01 0 0 0.05
   
1 0 Figure 1. Comparison of the feasibility regions: ‘°’ is given for the
B1 = , B2 = , quadratic solutions and ‘×’ for the nonquadratic solutions.
0 1
   
1 2
M1 = , M2 = ,
1.5 3
   
6 6−β
C1 = , C2 = .
0 −1
The scheduling functions are: h1 (z) = cos x1 and
h2 (z) = 1 − h1 (z).
In line with previous works (Ichalal et al., 2014;
Mozelli et al., 2009; Zhai et al., 2017), a feasibility
region is plotted by evaluating two parameters: α ∈
[6.5, 10] and β ∈ [8.5, 10]. For each combination of
α and β, a different TS model is obtained. In Figure 1,
the feasibility sets for the quadratic ‘°’ (see Appendix)
and non-quadratic ‘×’ approaches are plotted. For Figure 2. States estimation error, β = 9 and α = 7.5.
example, for α = 9 and β = 9, a feasible solution for
Theorem 3.1 is found, but not for the LMIs given in ⎡ ⎤ ⎡ ⎤
Appendix. Thus, it can be seen that a nonquadratic 0.1086 0.1413
Lyapunov function produces a larger feasibility region L1 = ⎣−0.3090⎦ , L2 = ⎣−0.3557⎦ .
than a quadratic one; hence the reduction of conser- 0.0666 −0.0276
vatism is clear.
A comparison is done by adjusting β = 9 and α = To validate the observer performance, simulations are
7.5. In this case, both approaches are feasible, and the carried out by considering initial conditions x(0) =
idea is to illustrate their performance for fault esti- [0 0]T and x̂(0) = [−0.1 0.5]T ; also a perturbation
mation. For the sake of simplicity, it is considered vector is added with zero value mean and magnitude
that φ1 = 0.9 and φ2 = 0.6. By solving the LMIs of 0.1.
Theorem 3.1 within the YALMIP Toolbox and the The state estimation error is shown in Figure 2. As
SEDUMI solver, the resulting matrices are: it can be seen, the observer converges fast and with a
⎡ ⎤ small error despite added noise. Note that in this case,
0.2671 0.0347 0.0403
the example is not considering the H∞ performance.
P1 = ⎣0.0347 0.2571 −0.3194⎦ ,
However, the observer has certain robustness due to
0.0403 −0.3194 1.8345
the reduced conservatism given by the nonquadratic
⎡ ⎤
0.0565 −0.0359 0.0098 Lyapunov function.
P2 = ⎣−0.0359 0.0866 −0.1552⎦ , A fault, as shown in Figure 3, is applied to actuator
0.0098 −0.1552 1.1042 1. The continuous line represents the applied fault; the
8 J. GUZMAN ET AL.

Figure 4. Three-tank hydraulic system scheme.

Torricelli’s law:
Figure 3. Actuator fault fa and its estimates. 
q1 (t) = η1 Sp1 2g|x1 (t) − x3 (t)|;

q2 (t) = η2 Sp2 2gx2 (t);
dashed line represents the fault estimation by consider- 
ing the quadratic Lyapunov function, and the pointed q3 (t) = η3 Sp3 2g|x3 (t) − x2 (t)|;
line represents the estimation of the method proposed
where ηi is the coefficient of discharge, Spi is the cross-
in this work. As displayed, both methods are capable
section of the connection tube i, and g is the gravi-
of estimating the actuator fault. Nevertheless, the non-
tational acceleration. Then, the mathematical model
quadratic approach can estimate the fault with better
that represents the dynamics of the hydraulic system
performance and with less convergence time. This is
is given by:
the result of reducing the conservatism by computing
 1
different matrices Pi that satisfy the LMI conditions. ẋ1 (t) = (−c1 |x1 (t) − x3 (t)| + u1 (t))
S
  1
4.2. Example 2 ẋ2 (t) = (−c2 x2 (t) + c3 |x3 (t) − x2 (t)| + u2 (t)
S
  1
A second example is proposed to illustrate the H∞ ẋ3 (t) = (c1 |x1 (t) − x3 (t)| − c3 |x3 (t) − x2 (t)|) ,
observer performance. In this case, the primary pur- S
pose is not to estimate the states but the actuator faults (29)
with the best possible performance. For such purpose, with g = 9.8 m/s2 , S = 15410 × 10−4 m2 , Sp1 =
consider the three-tank hydraulic system proposed in Sp3 = 0.5 × 10−4 m2 , Sp2 = 0.8 × 10−4 m2 , η1 =
√
Youssef et al. (2017). The system consist in three iden- η3 = 0.456, η2 = 0.652, c1 = η1 Sp1 2g = 1.01 ×
√
tical tanks connected with two identical cylindrical 10−4 , c2 = η2 Sp2 2g = 2.31 × 10−4 , and c3 = η3
√
tubes fed by two pumps that supply water to Tank 1 Sp3 2g = 1.01 × 10−4 . The output of the system is
and 2 respectively, as shown in Figure 4. A directional assumed to be y(t) = [x1 (t) x2 (t) x3 (t)]T . Thus, an
flow is assumed, so the tank levels x1 (t), x2 (t), x3 (t) are observer should be constructed in order to estimate
restricted to x1 (t) > x3 (t) > x2 (t). actuator faults.
Applying the conservation mass law, the following Defining x(t) = [x1 (t), x2 (t), x3 (t)]T and u(t) =
equations can be obtained [u1 (t), u2 (t)]T , the nonlinear model of the three-tank
system can be written in a state-space representation
Sẋ1 (t) = −q1 (t) + u1 (t); as follows:
⎡ ⎤
Sẋ2 (t) = −q2 (t) + q3 (t) + u2 (t); −c z (x) 0 0
1⎣ 1 1
Sẋ3 (t) = q1 (t) − q3 (t); ẋ(t) = 0 −c2 z2 (t) c3 z3 (x) ⎦ x(t)
S
c1 z1 (x) 0 −c1 z3 (x)
⎡ ⎤
where S is the cross-section of the tanks; ui (t) are the 1 0
1
input flows coming from pump i; and qi (t) is the out- + ⎣0 1⎦ u(t) (30)
S
put flow of tank i which can be calculated using the 0 0
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 9

⎡ ⎤
where zi (x) represent the nonlinear terms, which are −13 0 0
given by the following expressions: A7 = ⎣ 0 −62.7 6.5 ⎦ × 10−3 ,
√ √ 13 0 −6.5
|x1 (t) − x3 (t)| x2 (t)
z1 (x) = , z2 (x) = , ⎡ ⎤
x1 (t) x2 (t) −13 0 0
√ A8 = ⎣ 0 −62.7 13 ⎦ × 10−3 .
|x3 (t) − x2 (t)|
z3 (x) = . 13 0 −13
x3 (t)
(31)
The bounds of the states are x1 (t) ∈ [0 0.7], x2 (t) ∈ Finally, the TS representation under faults is given by:
[0 0.18], and x3 (t) ∈ [0 0.4], while the bounds of the
nonlinear terms are z1 (t) ∈ [0.4 2.0], z2 (t) ∈ [2.4 4.2], 
8
and z3 (t) ∈ [1.0 2.0]. The membership functions can ẋ(t) = hi (Ai x(t) + Bu(t) + Mfa (t))
i=1 (34)
be obtained with:
j z̄j − zj y(t) = Cx(t).
w0 = ; (32)
z̄j − z
j j Note that due to the fact that we are considering addi-
w1 = 1 − w 0 . (33) tive actuator faults M = B.
Due to the fact that the system has three nonlinear Then, by implementing the LMI conditions in
terms, the number of local models r that constitute the Theorem 3.2, for the scalars φ1 = 0.2, φ2 = 0.18, φ3 =
TS model is r = 23 , and the possible combinations are: 0.15, φ4 = 0.1, φ5 = 0.09, φ6 = 0.07, φ7 = 0.06, and
φ8 = 0.05, a feasible solution can be found. The fol-
h1 (z) = w01 w02 w03 ; h2 (z) = w01 w02 w13 ; lowing observer matrices can be obtained:
h3 (z) =w01 w12 w03 ; h4 (z) = w01 w12 w13 ;
⎡ ⎤
0.388 0 −0.764 3.482
h5 (z) = w11 w02 w03 ; h6 (z) = w11 w02 w13 ; ⎢ 0 249.17 0.0076 0.0008⎥⎥,
P1 =⎢
h7 (z) = w11 w12 w03 ; h8 (z) = w11 w12 w13 . ⎣−0.764 0.0076 293.79 86.142⎦
The local matrices Ai are calculated by substituting the 3.482 0.0008 86.142 229.35
⎡ ⎤
bounds z̄j and zj in (30): 0.383 0 −0.597 4.797
⎡ ⎤ ⎢ 0 249.18 −0.001 0.0142⎥⎥,
−2.6 0 0 P2 =⎢
⎣−0.597 −0.001 294.64 125.75⎦
A1 = ⎣ 0 −35.8 6.5 ⎦ × 10−3 ,
4.797 0.014 125.75 572.16
2.6 0 −6.5 ⎡ ⎤
⎡ ⎤ 0.404 0 −0.736 3.848
−2.6 0 0 ⎢ 0 249.28 0.007 0.0034⎥⎥,
A2 = ⎣ 0 −35.8 13 ⎦ × 10−3 , P3 =⎢
⎣−0.736 0.007 292.81 89.881⎦
2.6 0 −13
3.848 0.0034 89.881 273.23
⎡ ⎤
−2.6 0 0 ⎡ ⎤
0.420 0 −0.777 3.649
A3 = ⎣ 0 −62.7 6.5 ⎦ × 10−3 , ⎢ 0 249.29 −0.002 0.003 ⎥
⎥,
2.6 0 −6.5 P4 =⎢
⎣−0.777 −0.002 293.19 92.994⎦
⎡ ⎤
−2.6 0 0 3.649 0.003 92.994 283.9
A4 = ⎣ 0 −62.7 13 ⎦ × 10−3 , ⎡ ⎤
0.452 0 −0.947 3.548
2.6 0 −13 ⎢ 0 249.17 0.008 −0.003⎥⎥,
⎡ ⎤ P5 =⎢
−13 0 0 ⎣−0.947 0.008 296.12 97.617 ⎦
A5 = ⎣ 0 −35.8 6.5 ⎦ × 10−3 , 3.548 −0.003 97.617 284.83
13 0 −6.5 ⎡ ⎤
0.467 0 −0.971 3.481
⎡ ⎤
−13 0 0 ⎢ 0 249.18 −0.004 −0.004⎥⎥,
P6 =⎢
A6 = ⎣ 0 −35.8 13 ⎦ × 10−3 , ⎣−0.971 −0.004 296.37 99.519 ⎦
2.6 0 −13 3.481 −0.004 99.519 292.78
10 J. GUZMAN ET AL.

⎡ ⎤
0.478 0 −0.978 3.462 done with initial conditions x(0) = [0.08 0.06 0.07]T
⎢ 0 249.28 0.0076 −0.004⎥⎥, and x̂(0) = [0.089 0.069 0.075]T .
P7 = ⎢
⎣−0.978 0.007 296.2 99.988 ⎦ System inputs are shown in Figure 5. The perturba-
3.462 −0.004 99.988 297.44 tion and noise signal d(t) is considered with zero mean
⎡ ⎤
0.155 0 −0.728 0.916
⎢ 0 83.133 −0.002 −0.001⎥⎥
P8 = ⎢
⎣−0.728 −0.002 129.81 84.958 ⎦ ;
0.916 −0.001 84.958 171.84

and:
⎡ ⎤
1.1389 0.0003 1.571
⎢ 0 105.34 2.167 ⎥⎥,
L1 =⎢
⎣−0.764 −0.494 111.89⎦
3.482 0.005 26.584
⎡ ⎤
1.1336 −0.0002 1.4201
⎢ 0 105.33 3.1601⎥ ⎥,
L2 =⎢
⎣−0.5976 0.0436 110.06⎦
4.797 0.0027 22.365 Figure 5. Dynamics of the system inputs u1 and u2 .
⎡ ⎤
1.154 0.0003 1.544
⎢ 0 99.017 0.6638⎥
⎥,
L3 =⎢
⎣−0.736 1.0176 111.89⎦
3.848 0.004 26.014
⎡ ⎤
1.17 −0.0002 1.5979
⎢ 0 99.025 0.9981⎥ ⎥,
L4 =⎢
⎣−0.777 2.2019 110.08⎦
3.6499 0.0045 26.163
⎡ ⎤
1.2024 −0.0002 4.6946
⎢ 0 105.3 1.2245⎥
⎥,
L5 =⎢
⎣−0.947 0.458 111.84⎦
3.548 0.002 29.993
⎡ ⎤
1.2175 −0.0001 4.723 Figure 6. Measurement noise d(t).
⎢ 0 105.29 2.5681⎥⎥,
L6 =⎢
⎣−0.971 0.6278 110.03⎦
3.481 0.002 29.821
⎡ ⎤
1.228 −0.0002 4.7193
⎢ 0 99.054 −0.1384⎥ ⎥,
L7 =⎢
⎣−0.978 1.8327 111.83 ⎦
3.462 0.0028 30.581
⎡ ⎤
0.9054 0.001 2.9431
⎢ 0 107.91 0.340 ⎥⎥.
L8 =⎢
⎣−0.7284 1.0772 111.96⎦
0.9163 0.0031 24.983

In addition, the value of γ = 1 was obtained. To illus-

trate the observer performance, simulations have been Figure 7. Applied fault and their estimation by the PI observer.
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 11

value and magnitude of 1.5 × 10−5 m as it is shown in MSc degree in Mechatronics Engineering in 2017 from Tec-
Figure 6. First, a ramp type fault starting at t = 200 s nolÃşgico Nacional de MÃľxico, IT Tuxtla GutiÃľrrez. His
research interests are centered on fault diagnosis, fault-tolerant
to t = 400 s is considered, the magnitude of the fault
control systems, TakagiâĂŞSugeno systems, LPV systems, and
remains constant from t = 400 s to t = 550 s reach- their applications.
ing a magnitude of 4.5 × 10−5 , from t = 500 s the fault
begins to decrease until t = 750 s when it disappears. F.-R. López-Estrada received his PhD degree in automatic
control from the University of Lorraine, France, in 2014. He
The performance of the observer can be seen in has been with TecnolÃşgico Nacional de MÃľxico, IT Tuxtla
Figure 7, it is clear that the observer perfectly estimates GutiÃľrrez, as a lecturer since 2008. He is part of the Editorial
the fault even in the presence of disturbances. Board of the Mathematical and Computational Applications and
Int. J. of Applied Mathematics and Computer Science journals.
His research interests are in descriptor systems, TakagiâĂŞ-
5. Conclusions Sugeno systems, fault detection, fault tolerant control, and their
applications.
In this work, a proportional-integral fault estimation
observer has been developed to estimate both states V. Estrada-Manzo was born in Zamora, Mexico, in 1987. He
and actuator faults. The main contribution of the pro- received the PhD degree in Automatic Control from the Uni-
versity of Valenciennes and Hainaut-Cambresis, France, in
posed method was to consider a nonquadratic Lya-
2015, on the subject of nonsingular convex descriptor systems.
punov function, which offers less conservative results, He worked as a Post-Doctoral fellow at the Sonora Institute of
compared to a quadratic Lyapunov function tradition- Technology, Mexico from 2015 to 2018. He currently holds an
ally considered in the literature. Also, an H∞ perfor- Associate Professor position at the Polytechnical University of
mance criteria has been included in order to guaran- Pachuca, Mexico. His research interests are analysis and design
tee robustness against sensor noise, disturbance, and of nonlinear control systems through descriptor quasi-LPV
models and convex optimization techniques.
modelling mismatches. As a result, sufficient condi-
tions have been obtained by a set of linear matrix G. Valencia-Palomo was born in Merida, Mexico, in 1980. He
inequalities that computes the observer gains with received his PhD degree in Automatic Control and Systems
the desired performance. As illustrated for two exam- Engineering from the University of Sheffield, UK, in 2010. He
has been with TecnolÃşgico Nacional de MÃľxico, IT Her-
ples, the proposal shows its effectiveness by atten-
mosillo, as a Tenured Professor since 2010. He is Associate
uating noise and disturbances and estimating with Editor of IEEE Latin America Transactions, and part of the
higher precision the actuator faults in comparison with Editorial Board of Mathematical, and Computational Applica-
the traditional quadratic approach. Future work will tions journals. His research interests include predictive control,
be done to improve the attenuation levels by con- fault detection and isolation, fault-tolerant control, and their
sidering an extended version of the TS systems with applications to different physical systems.
unmeasurable premise variables and their implemen-
tation in fault-tolerant control as in López-Estrada ORCID
et al. (2020). F.-R. López-Estrada http://orcid.org/0000-0002-8724-335X
V. Estrada-Manzo http://orcid.org/0000-0002-2902-8424
G. Valencia-Palomo http://orcid.org/0000-0002-3382-8213
Disclosure statement
No potential conflict of interest was reported by the References
author(s).
Ahmadizadeh, S., Zarei, J., & Karimi, H. R. (2014). Robust
unknown input observer design for linear uncertain time
Funding delay systems with application to fault detection. Asian Jour-
nal of Control, 16(4), 1006–1019. https://doi.org/10.1002/
This work was supported by Consejo Nacional de Ciencia y
asjc.765
Tecnología [grant number 487762].
Aouaouda, S., Chadli, M., Shi, P., & Karimi, H. R. (2015).
Discrete-time H− /H∞ sensor fault detection observer
design for nonlinear systems with parameter uncertainty.
Notes on contributors
International Journal of Robust and Nonlinear Control, 25(3),
J. Guzman is a PhD student at the Department of Elec- 339–361. https://doi.org/10.1002/rnc.v25.3
tronic Engineering at TecnolÃşgico Nacional de MÃľxico, Bernal, M., & Guerra, T. M. (2010). Generalized non-quadratic
Instituto TecnolÃşgico de Tuxtla GutiÃľrrez. He received his stability of continuous-time Takagi-Sugeno models. IEEE
12 J. GUZMAN ET AL.

Transactions on Fuzzy Systems, 18(4), 815–822. https://doi. TS fuzzy systems using fuzzy Lyapunov functions. Inter-
org/10.1109/TFUZZ.2010.2049113 national Journal of Systems Science, 44(10), 1956–1969.
Bernal, M., Soto-Cota, A., Cortez, J., Pitarch, J. L., & Jaadari, https://doi.org/10.1080/00207721.2012.670307
A. (2011). Local non-quadratic H∞ control for continuous- Gómez-Peñate, S., Valencia-Palomo, G., López-Estrada, F. R.,
time Takagi-Sugeno models. In 2011 IEEE international con- C. M. Astorga-Zaragoza, Osornio-Rios, R. A., & Santos-
ference on fuzzy systems (FUZZ-IEEE 2011), Taipei, Taiwan Ruiz, I. (2019). Sensor fault diagnosis based on a sliding
(pp. 1615–1620). mode and unknown input observer for Takagi-Sugeno sys-
Boyd, S., El Ghaoui, L., Feron, E., & Balakrishnan, V. (1994). tems with uncertain premise variables. Asian Journal of
Linear matrix inequalities in system and control theory (Vol. Control, 21(1), 339–353. https://doi.org/10.1002/asjc.v21.1
15). SIAM. Guerra, T. M., Bernal, M., Guelton, K., & Labiod, S. (2012).
Castro, M. L., Escobar, R., Torres, L., Aguilar, J. G., Hernández, Non-quadratic local stabilization for continuous-time Tak-
J., & Olivares-Peregrino, V. (2016). Sensor fault detection and agi–Sugeno models. Fuzzy Sets and Systems, 201, 40–54.
isolation system for a condensation process. ISA Transac- https://doi.org/10.1016/j.fss.2011.12.003
tions, 65, 456–467. https://doi.org/10.1016/j.isatra.2016.08. Guerra, T. M., Sala, A., & Tanaka, K. (2015). Fuzzy control
004 turns 50: 10 years later. Fuzzy Sets and Systems, 281, 168–182.
Chadli, M., Aouaouda, S., Karimi, H. R., & Shi, P. (2013). https://doi.org/10.1016/j.fss.2015.05.005
Robust fault tolerant tracking controller design for a VTOL Hassanabadi, A. H., Shafiee, M., & Puig, V. (2017). Actuator
aircraft. Journal of the Franklin Institute, 350(9), 2627–2645. fault diagnosis of singular delayed LPV systems with inex-
https://doi.org/10.1016/j.jfranklin.2012.09.010 act measured parameters via PI unknown input observer.
Chang, J. L. (2006). Applying discrete-time proportional IET Control Theory & Applications, 11(12), 1894–1903.
integral observers for state and disturbance estimations. https://doi.org/10.1049/iet-cta.2016.1304
IEEE Transactions on Automatic Control, 51(5), 814–818. Huimin, C., & Yafang, W. (2007). Study on the detection system
https://doi.org/10.1109/TAC.2006.875019 of pipeline leakage based on the double parameters method.
Cheng, J., Park, J. H., Cao, J., & Qi, W. (2020). A hidden mode Journal of Shijiazhuang Railway Institute, 2, 013.
observation approach to finite-time SOFC of Markovian Ichalal, D., Marx, B., Ragot, J., & Maquin, D. (2009). Simul-
switching systems with quantization. Nonlinear Dynam- taneous state and unknown inputs estimation with PI and
ics, 100(1), 509–521. https://doi.org/10.1007/s11071-020- PMI observers for Takagi-Sugeno model with unmeasurable
05501-0 premise variables. In 2009 17th Mediterranean conference on
Cherifi, A., Guelton, K., Arcese, L., & Leite, V. J. S. (2019). control and automation (pp. 353–358). IEEE.
Global non-quadratic D-stabilization of Takagi–Sugeno Ichalal, D., Marx, B., Ragot, J., & Maquin, D. (2014). Fault detec-
systems with piecewise continuous membership func- tion, isolation and estimation for Takagi–Sugeno nonlinear
tions. Applied Mathematics and Computation, 351, 23–36. systems. Journal of the Franklin Institute, 351(7), 3651–3676.
https://doi.org/10.1016/j.amc.2019.01.031 https://doi.org/10.1016/j.jfranklin.2013.04.012
Chitraganti, S., Aberkane, S., Aubrun, C., Valencia-Palomo, Ince, T., Kiranyaz, S., Eren, L., Askar, M., & Gabbouj, M. (2016).
G., & Dragan, V. (2014). On control of discrete-time state- Real-time motor fault detection by 1-D convolutional neu-
dependent jump linear systems with probabilistic con- ral networks. IEEE Transactions on Industrial Electron-
straints: A receding horizon approach. Systems & Control ics, 63(11), 7067–7075. https://doi.org/10.1109/TIE.2016.
Letters, 74, 81–89. https://doi.org/10.1016/j.sysconle.2014. 2582729
10.008 Khedher, A., Benothman, K., Benrejeb, M., & Maquin, D.
Dehghani, M., Khooban, M. H., & Niknam, T. (2016). Fast (2010). Adaptive observer for fault estimation in nonlin-
fault detection and classification based on a combination ear systems described by a Takagi-Sugeno model. In 18th
of wavelet singular entropy theory and fuzzy logic in dis- Mediterranean conference on control and automation (pp.
tribution lines in the presence of distributed generations. 261–266). IEEE.
International Journal of Electrical Power & Energy Systems, Kladis, G. P., Menon, P. P., & Edwards, C. (2016). Fuzzy dis-
78, 455–462. https://doi.org/10.1016/j.ijepes.2015.11.048 tributed cooperative tracking for a swarm of unmanned
Duan, G. R., & Wu, A. G. (2006). Robust fault detection in aerial vehicles with heterogeneous goals. International Jour-
linear systems based on PI observers. International Jour- nal of Systems Science, 47(16), 3803–3811. https://doi.org/
nal of Systems Science, 37(12), 809–816. https://doi.org/ 10.1080/00207721.2015.1126380
10.1080/00207720600825511 López-Estrada, F. R., Ponsart, J. C., Theilliol, D., Zhang,
Estrada-Manzo, V., Lendek, Z., & Guerra, T. M. (2015). Y., & Astorga-Zaragoza, C. M. (2016). LPV model-based
Unknown input estimation for nonlinear descriptor systems tracking control and robust sensor fault diagnosis for
via LMIs and Takagi-Sugeno models. In 2015 54th IEEE con- a quadrotor UAV. Journal of Intelligent & Robotic Sys-
ference on decision and control (CDC), Osaka, Japan (pp. tems, 84(1-4), 163–177. https://doi.org/10.1007/s10846-015-
6349–6354). 0295-y
Faria, F. A., Silva, G. N., & Oliveira, V. A. (2013). Reducing López-Estrada, F. R., Rotondo, D., & Valencia-Palomo, G.
the conservatism of LMI-based stabilisation conditions for (2019). A review of convex approaches for control,
 INTERNATIONAL JOURNAL OF SYSTEMS SCIENCE 13

observation and safety of linear parameter varying and Tanaka, K., Hori, T., & Wang, H. O. (2003). A multiple Lya-
Takagi-Sugeno systems. Processes, 7(11), 814. https://doi.org/ punov function approach to stabilization of fuzzy con-
10.3390/pr7110814 trol systems. IEEE Transactions on Fuzzy Systems, 11(4),
López-Estrada, F. R., Santos-Estudillo, O., Valencia-Palomo, 582–589. https://doi.org/10.1109/TFUZZ.2003.814861
G., Gómez-Peñate, S., & Hernandez-Gutiérrez, C. (2020). Tanaka, K., & Wang, H. (2001). Fuzzy control systems design and
Robust qLPV tracking fault-tolerant control of a 3-DOF analysis: A linear matrix inequality approach. Wiley.
mechanical crane. Mathematical and Computational Appli- Tlili, A. S. (2019). Proportional integral observer-based decen-
cations, 25(3), 48. https://doi.org/10.3390/mca25030048 tralized stabilization design scheme for nonlinear complex
Mozelli, L. A., Palhares, R. M., Souza, F. O., & Mendes, E. M. A. interconnected systems. Transactions of the Institute of Mea-
M. (2009). Reducing conservativeness in recent stability con- surement and Control, 41(7), 1811–1823. https://doi.org/10.
ditions of TS fuzzy systems. Automatica, 45(6), 1580–1583. 1177/0142331218788124
https://doi.org/10.1016/j.automatica.2009.02.023 Tong, S., Yang, G., & Zhang, W. (2011). Observer-based fault-
Nazari, S., & Shafai, B (2018). Distributed proportional-integral tolerant control against sensor failures for fuzzy systems with
observers for fault detection and isolation. In 2018 IEEE con- time delays. International Journal of Applied Mathematics
ference on decision and control (CDC), Miami Beach, FL, USA and Computer Science, 21(4), 617–627. https://doi.org/10.
(pp. 6328–6333). 2478/v10006-011-0048-4
Ohtake, H., Tanaka, K., & Wang, H (2001). Fuzzy modeling via Tuan, H. D., Apkarian, P., Narikiyo, T., & Yamamoto, Y. (2001).
sector nonlinearity concept. In Proceedings joint 9th IFSA Parameterized linear matrix inequality techniques in fuzzy
world congress and 20th NAFIPS international conference control system design. IEEE Transactions on Fuzzy Systems,
Vancouver, Canada (Vol. 1, pp. 127–132). 9(2), 324–332. https://doi.org/10.1109/91.919253
Pérez-Pérez, E. J., López-Estrada, F. R., Valencia-Palomo, G., Vafamand, N., & Sadeghi, M. S. (2015). More relaxed
Torres, L., Puig, V., & Mina-Antonio, J. D. (2021). Leak non-quadratic stabilization conditions for TS fuzzy con-
diagnosis in pipelines using a combined artificial neural net- trol systems using LMI and GEVP. International Jour-
work approach. Control Engineering Practice, 107, 104677. nal of Control, Automation and Systems, 13(4), 995–1002.
https://doi.org/10.1016/j.conengprac.2020.104677 https://doi.org/10.1007/s12555-013-0497-7
Puig, V., Quevedo, J., Escobet, T., Morcego, B., & Ocampo, Verde, C., Gentil, S., & Morales-Menéndez, R. (2013). Moni-
C. (2004). Control tolerante a fallos (Parte I): Fun- toreo y diagnóstico automático de fallas en sistemas dinámicos.
damentos y diagnóstico de fallos. Revista Iberoameri- Trillas.
cana de Automática e Informática Industrial, 1(1), 15–31. Wang, L., Zhang, H., & Liu, X. (2016). H∞ observer design for
https://doi.org/10.4995/riai.2004.8021 continuous-time Takagi-Sugeno fuzzy model with unknown
Quintana, D., Estrada-Manzo, V., & Bernal, M (2020). An premise variables via non-quadratic Lyapunov function.
exact handling of the gradient for overcoming persis- IEEE Transactions on Systems, Man, and Cybernetics, Part
tent problems in nonlinear observer design via con- B (Cybernetics), 46(9), 1986–1996. https://doi.org/10.1109/
vex optimisation techniques. Fuzzy Sets and Systems. TCYB.2015.2459016
https://doi.org/10.1016/j.fss.2020.04.012 Yin, S., Luo, H., & Ding, S. X. (2014). Real-time implementa-
Rodrigues, M., Hamdi, H., Braiek, N. B., & Theilliol, D. (2014). tion of fault-tolerant control systems with performance opti-
Observer-based fault tolerant control design for a class of mization. IEEE Transactions on Industrial Electronics, 61(5),
LPV descriptor systems. Journal of the Franklin Institute, 2402–2411. https://doi.org/10.1109/TIE.2013.2273477
351(6), 3104–3125. https://doi.org/10.1016/j.jfranklin.2014. Youssef, T., Chadli, M., Karimi, H. R., & Wang, R. (2017).
02.016 Actuator and sensor faults estimation based on proportional
Santos-Ruiz, I., López-Estrada, F., Puig, V., Pérez-Pérez, E., integral observer for TS fuzzy model. Journal of the Franklin
Mina-Antonio, J., & Valencia-Palomo, G. (2018). Diagno- Institute, 354(6), 2524–2542. https://doi.org/10.1016/j.
sis of fluid leaks in pipelines using dynamic PCA. IFAC- jfranklin.2016.09.020
PapersOnLine, 51(24), 373–380. https://doi.org/10.1016/j. Zhai, D., Lu, A. Y., Dong, J., & Zhang, Q. L. (2017). Stabil-
ifacol.2018.09.604 ity analysis and state feedback control of continuous-time
Shamma, J. S. (2012). An overview of LPV systems. In Control of T–S fuzzy systems via anew switched fuzzy Lyapunov func-
linear parameter varying systems with applications (pp. 3–26). tion approach. Applied Mathematics and Computation, 293,
Springer. 586–599. https://doi.org/10.1016/j.amc.2016.08.053
Su, X., Shi, P., Wu, L., & Song, Y. D. (2016). Fault detec- Zhong, W., Lu, J., & Miao, Y (2017). Fault detection observer
tion filtering for nonlinear switched stochastic systems. design for fractional-order systems. In 2017 29th Chinese
IEEE Transactions on Automatic Control, 61(5), 1310–1315. control and decision conference (CCDC), Chongqing, China
https://doi.org/10.1109/TAC.2015.2465091 (pp. 2796–2801).
Takagi, T., & Sugeno, M. (1985). Fuzzy identification of sys- Zhou, M., Wang, Z., & Shen, Y. (2017). Fault detection and iso-
tems and its applications to modeling and control. IEEE lation method based on H− /H∞ unknown input observer
Transactions on Systems, Man, and Cybernetics, 15(1), design in finite frequency domain. Asian Journal of Control,
116–132. https://doi.org/10.1109/TSMC.1985.6313399 19(5), 1777–1790. https://doi.org/10.1002/asjc.1490
14 J. GUZMAN ET AL.

Appendix. The quadratic approach for observer Then, the following LMI conditions are obtained:
design of TS systems
Lemma A.1: The PI observer (A1) for the system (7) is computed
The stability of TS observers is usually investigated using if there exists matrices P = PT > 0 and Wi , i ∈ {1, 2 . . . , r} such
the direct Lyapunov method. Then, for system (4), the that LMIs
proportional-integral TS observer takes the form:
2
 r ϒii + ϒij + ϒji < 0, ∀(i, j) ∈ {1, 2, . . . , r}2 , (A4)
˙ˆ ˆ + B̄i u(t) + L̄i (y(t) − ŷ(t))) r−1
x̄(t) = hi (Āi x̄(t) (A1)
i=1
hold with:
 
with L̄i = LLPiIi . The estimation error dynamics between (7) ϒij := Āi P + PĀi − Wj C̄i − C̄iT WjT .
and (A1) is expressed as: The observer gains are computed as Li = P−1 Wi , i ∈

r 
r {1, 2, . . . , r}.
ė(t) = hi hj (Āi − L̄j C̄i )ē(t). (A2)
i=1 j=1 The proof is not presented here, but a detailed mathematical
deduction, including the H∞ performance, can be consulted in
Then, in order to guarantee the observer convergence, there
Ichalal et al. (2009), Khedher et al. (2010) and López-Estrada
are plenty of papers that consider the following quadratic Lya-
et al. (2019).
punov function candidate:
V(e() = eT (t)Pe(t), with P = PT > 0. (A3)