Communicated by Dr Hu Jun

Accepted Manuscript

Event-triggered fault detection for discrete-time Lipschitz nonlinear

networked systems in finite-frequency domain

Ying Gu, Guang-Hong Yang

PII: S0925-2312(17)30746-4
DOI: 10.1016/j.neucom.2017.04.037
Reference: NEUCOM 18370

To appear in: Neurocomputing

Received date: 20 January 2017

Revised date: 13 April 2017
Accepted date: 24 April 2017

Please cite this article as: Ying Gu, Guang-Hong Yang, Event-triggered fault detection for discrete-
time Lipschitz nonlinear networked systems in finite-frequency domain, Neurocomputing (2017), doi:
10.1016/j.neucom.2017.04.037

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Event-triggered fault detection for discrete-time

Lipschitz nonlinear networked systems in
finite-frequency domain

T
Ying Gua,b , Guang-Hong Yanga,c,∗

IP
a
College of Information Science and Engineering, Northeastern University, Shenyang,
Liaoning, 110819, P. R. China

CR
b
College of Science, Dalian Jiaotong University, Dalian, Liaoning, 116028, P. R. China
c
State Key Laboratory of Synthetical Automation of Process Industries, Northeastern
University, Shenyang, Liaoning, 110819, P. R. China

Abstract
US
AN
The problem of event-triggered fault detection (FD) filter design for discrete-
time Lipschitz nonlinear networked systems with finite-frequency specifica-
tions is investigated. The event-triggered transmission scheme is introduced
to mitigate the utility of limited network bandwidth. The developed filter
M

combines the H∞ and H− indices. For this class of systems, the general-
ized Kalman-Yakubovic-Popov lemma-based finite-frequency FD filter design
methods are invalid. To solve this problem, the nonlinear error dynamics are
ED

transformed into a linear parameter varying (LPV) system based on the use
of a reformulated Lipschitz property and a new lemma is developed to cap-
ture the system performances in finite-frequency domain. By introducing
PT

slack variable techniques, sufficient conditions for the design of FD filter are
derived in terms of linear matrix inequalities (LMIs). The proposed design
method can significantly reduce the data transmission and achieve a better
CE

FD performance than that in full frequency domain. Finally, two examples

are presented to validate the effectiveness and superiority of the new results.
Keywords: Networked control systems, Lipschitz nonlinear systems, Fault
AC

detection, Event-triggered scheme, Finite frequency, H− /H∞ .

∗
Corresponding author
Email addresses: [email protected] (Ying Gu), [email protected]
(Guang-Hong Yang)

Preprint submitted to Neurocomputing April 29, 2017

 ACCEPTED MANUSCRIPT

1. Introduction
Networked control systems (NCSs) have received a great deal of atten-
tion due to the rapid development of communication technology. The main
advantages of NCSs are low cost, simple installation and maintenance, high

T
efficiency and reliability. As a result, NCSs are successfully applicable to
many fields, ranging from mobile communication, advanced aircraft, remote

IP
surgery and manufacturing processes. However, since the network cable is
of limited capacity, the insertion of network brings about new challenging

CR
issues including transmission delay, signal quantisation, data missing, etc.
Recently, extensive efforts have been made on estimation, analysis, and con-
troller synthesis for NCSs, see, for example [1, 2, 3, 4, 5] and the references
therein.

US
Practically all the control tasks executed today are based on digital sam-
pling measurements. The simplest approach is to transmit measurements
periodically. Although periodic sampling is preferred for system modeling
AN
and analysis, the limited communication band and inadequate computation
resources are the problems that often have to be dealt with in a network
environment. Under typical periodic sampling, a lot of unnecessary sampled
data is transmitted even when the output fluctuation is small compared with
M

the previous one, which may lead to excessive utilisation of the limited bands.
Therefore, how to reduce data transmission while guaranteeing the desired
performances of NCSs is an important topic. Recently, event-triggered data
ED

transmission scheme has been proposed for its capacity of reducing the usage
of communication resources in NCSs. There is an increasing interest in using
event-triggered strategy for control [6, 7, 8, 9] and filtering [10, 11, 12, 13]
PT

areas. Event-triggered method advocates data transmission only when there

is a significant change in the signal. Therefore, the burden of the network
communication is reduced.
CE

Nonlinearities ubiquitously exist in practical systems. As most real phys-

ical processes can be described by Lipschitz nonlinear models, considerable
interest has been shown on designing observers or filters for Lipschitz systems.
AC

Many important results have been reported in the literature, see for example
[14, 15, 16]. For Lipschitz nonlinear networked control systems, fruitful re-
sults have been obtained with the consideration of different network-induced
phenomena [17, 18, 19, 20]. However, these results are all based on the
traditional time-triggered communication scheme, which is not efficient for
saving the network bandwidth. Besides, the provided synthesis conditions

2
 ACCEPTED MANUSCRIPT

in all these methods are generally infeasible for systems with large Lipschitz
constants. A reformulated Lipschitz property proposed in [21] can be used
to cope with this restriction. Due to the introduction of a less conservative
Lipschitz condition, the proposed design method can provide less restrictive
synthesis conditions than those reported in the literature and avoids high

T
gain.

IP
Fault detection (FD), on the other hand, has received much attention
due to the increasing demand for reliability and safety in industrial process.
A lot of detection methods have been achieved, for instance, see [22, 23]

CR
and the references therein. Among these FD methods, robust FD scheme is
accepted as a useful and popular one. The main challenge in robust FD is
to generate a residual signal that should be robust to the disturbances and

US
simultaneously sensitive to the faults. Then compare the residual signal with
a predefined threshold. The robustness and sensitivity should be measured
by a suitable performance index and be optimized, e.g., H∞ optimization
[24], H2 /H∞ optimization [25] or mixed H− /H∞ optimization [26, 27]. The
AN
idea of event-based control can be applicable for FD purposes. In [28], the
event-based FD for the networked systems with communication delay and
nonlinear perturbation has been investigated. Recently, the problem of event-
M

triggered fault detection and isolation filter design for discrete-time linear
systems was addressed in [29]. [30] studied the problems of event-triggered
multi-objective synthesis of feedback controllers and fault diagnosis filters
ED

in a unified framework for discrete-time linear systems. [31] investigated

fuzzy FD problem under event-triggered scheme. [32] studied the problem
of event-triggered FD filter and controller coordinated design for continuous-
time NCSs with biased sensor faults.
PT

Furthermore, it is worth noticing that all the aforementioned FD ap-

proaches were considered in full frequency domain. However, in practice, the
frequency ranges of external disturbances and faults may be known before-
CE

hand. For instance, incipient faults emerge in low frequency domain as the
fault development is slow [22]. Therefore, it will be conservative to design full
frequency filters to detect faults due to the overdesign. To cope with the fre-
AC

quency of the signals, the generalized Kalman-Yakubovich-Popov (GKYP)

lemma [33] based finite-frequency H∞ control and filtering problems have
been studied in [34, 35] for T-S fuzzy systems. More recently, finite fre-
quency FD methods have been presented in [36, 37, 38]. However, these
existing finite-frequency FD results are invalid for nonlinear networked con-
trol systems with Lipschitz nonlinearity under event-triggered scheme. To

3
 ACCEPTED MANUSCRIPT

the best of our knowledge, the FD problem for Lipschitz nonlinear systems
in finite-frequency domain has not been investigated yet, not to mention the
case when event-triggered scheme is considered. Therefore, the main purpose
of this paper is to shorten such a gap.
The main contributions in this paper are as follows: (1) A reformulation of

T
the Lipschitz property is considered to transform the nonlinear error dynam-

IP
ics into a LPV system [21]. (2) An H− /H∞ FD method with finite-frequency
specifications under event-triggered scheme is developed, and a novel lemma
is derived to guarantee the whole system with desired performances in finite-

CR
frequency domain. (3) By using slack variable techniques, the non-convex
analytical conditions are transformed to linear ones. For each performance
index, sufficient conditions are given in terms of LMIs. (4) The FD scheme

US
is proposed based on the obtained conditions which is valid to detect the
fault in finite-frequency domain for systems with large values of Lipschitz
constants. Also, it has a better performance in saving the communication
resources and detecting faults than the existing one.
AN
The rest of this paper is organized as follows: Section 2 gives the sys-
tem description and the problem formulation. In Section 3, the LPV-based
approach for Lipschitz systems is proposed. LMI-based fault detection filter
M

design conditions under event-triggered scheme are then derived in the finite-
frequency domain in Section 4. Two examples are given in Section 5 to show
the advantages of the proposed method. Section 6 concludes the paper.
ED

Notations: In this paper, the notations are standard. For a matrix A,

AT , A⊥ denote its transpose and orthogonal complement, respectively and
A > 0(A < 0) means that it is positive definite (negative definite). The
symbol ∗ will be used in some matrix expressions to denote the transposed
PT

elements in the symmetric positions of a matrix. In is an identity matrix of

size n. The sum of a square matrix A and its transpose AT is denoted by
ith
z}|{
CE

He(A) := A + A . es (i) = (0, · · · , 0, 1 , 0, · · · , 0)T ∈ Rs , s ≥ 1 is a vector

T
| {z }
s components
of the canonical basis of Rs .
AC

4
 ACCEPTED MANUSCRIPT

2. System description and problem statement

2.1. System description
Consider the following nonlinear networked control systems described by


T
x(k + 1) = Ax(k) + Φ(x(k)) + Bf f (k) + Bd d(k)
(1)
y(k) = Cx(k) + Df f (k) + Dd d(k)

IP
where x(k) ∈ Rn is the state vector, y(k) ∈ Rp is the output vector, d(k) ∈ Rl

CR
is the unknown bounded disturbance vector, f (k) ∈ Rq is the fault to be
detected. Matrices A ∈ Rn×n , Bf ∈ Rn×q , Bd ∈ Rn×l , C ∈ Rp×n , Df ∈
Rp×q , Dd ∈ Rp×l are real known constant matrices. The nonlinear function
Φ(x(k)) is assumed to satisfy Φ(0) = 0 and rΦ -Lipschitz, i.e.:

∀x(k), x̂(k) ∈ Rn
US
kΦ(x(k)) − Φ(x̂(k))k ≤ rΦ kx(k) − x̂(k)k,
(2)
AN
where rΦ is a known Lipschitz constant.
Remark 1. Note that globally Lipschitz nonlinear systems are only a limited
class of nonlinear systems. When Φ(x(k)) is locally Lipschitz, all the results
M

derived in this paper are valid in a neighborhood around a nominal point.

Let the event-triggered FD filter be of the following form:

ED

 x̂(k + 1) = Ax̂(k) + Φ(x̂(k)) + L(ỹ(k) − ŷ(k))

ŷ(k) = C x̂(k) (3)

r(k) = V (ỹ(k) − ŷ(k))
PT

where x̂(k) ∈ Rn and ŷ(k) ∈ Rp represent the state and output estimation
vectors, respectively. r(k) ∈ Rs is the residual signal. ỹ(k) ∈ Rp is the last
CE

released data from the sensor to the filter modul. L and V are the filter
matrices to be determined.
For the purpose of saving the limited communication resources, the event-
triggered scheme in [8] is introduced to decide whether or not the sam-
AC

pled data should be transmitted. The event instants are assumed to be

k0 , k1 , k2 , · · · , where k0 = 0 is the initial event-triggered time. Define

ỹ(k) = y(ki ), k ∈ [ki , ki+1 ). (4)

The event-triggered FD scheme is shown in Fig. 1.

5
 ACCEPTED MANUSCRIPT

T
IP
Figure 1: The schematic of event-triggered FD problem.

CR
In order to reduce the data transmission frequency, the current measure-
ment y(k) satisfying

US
δ T (k)Ωδ(k) ≥ σy T (k)Ωy(k) (5)
will be released, where δ(k) = ỹ(k) − y(k) is the difference between the last
released data of the generator and the current output of the plant, σ ∈ [0, 1)
AN
is a scalar, and Ω ∈ Rm×m is a positive weighting matrix.
Remark 2. When σ = 0, {k0 , k1 , k2 , · · · } = {0, 1, 2, · · · }, all the output
signals will be transmitted to the FD module. In this case, the event-triggered
M

scheme reduces to the case of time-triggered scheme.

Let
e(k) = x(k) − x̂(k), η(k) = [xT (k) eT (k)]T , ∆Φ(k) = Φ(x(k)) − Φ(x̂(k)).
ED

(6)
Considering (1), (3) and the event-triggered scheme (5), the filter error sys-
tem can be derived as follows:
PT


 η(k + 1) = Aη(k) + I1 Φ(x(k)) + I2 ∆Φ(k) + Bf f (k) + Bd d(k) + Bδ δ(k)
r(k) = C1 η(k) + Df f (k) + Dd d(k) + Dδ δ(k)

CE

y(k) = C2 η(k) + Df f (k) + Dd d(k)

(7)
where
       
AC

A 0 I 0 Bf
A= , I1 = , I2 = , Bf = ,
0 A − LC 0 I Bf − LDf
   
Bd 0    
Bd = , Bδ = , C1 = 0 V C , C2 = C 0 ,
Bd − LDd −L
Df = V Df , Dd = V Dd , Dδ = V.

6
 ACCEPTED MANUSCRIPT

2.2. Problem statement

In this subsection, the finite-frequency H∞ performance index and finite-
frequency H− index are defined to formulate the fault detection problem.
Consider the following finite frequency interval for frequency θ of the
corresponding signal:

T
Ωθ := {θ ∈ R|ϑ1 ≤ θ ≤ ϑ2 } (8)

IP
where 0 < ϑ2 − ϑ1 ≤ 2π, ϑ1 , ϑ2 are given real scalars and used to describe

CR
the middle-frequency range of the corresponding signal.
Note that, when −ϑ1 = ϑ2 = ϑl , Ωθ denotes the low frequency range
Ωθ := {θ ∈ R||θ| ≤ ϑl } (9)

US
where ϑl is assumed to be a known positive real scalar; when ϑ1 = ϑh , ϑ2 =
2π − ϑh , Ωθ denotes the high frequency range
Ωθ := {θ ∈ R||θ| ≥ ϑh } (10)
AN
where ϑh is assumed to be a known positive real scalar.
Definition 1. The filtering error system (7) has a finite-frequency H∞ index
M

bound γ, if under zero initial condition, the following inequality

kr(k)k2 ≤ γkd(k)k2 (11)
ED

holds for all solutions of (7) with d(k) ∈ l2 in finite-frequency range ϑd1 ≤
θd ≤ ϑd2 , such that
∞
X
jϑd$
(η(k + 1) − ejϑd1 η(k))(η(k + 1) − e−jϑd2 η(k))T ≤ 0
PT

e (12)
k=0

where θd denotes the frequency of d(k), ϑd$ = (ϑd2 − ϑd1 )/2, ϑd1 and ϑd2
CE

reflect the frequency range of d(k). γ is a given real positive scalar which
denotes the worst case criterion for the effect of d(k) on the residual r(k).
Then the parameter γ is called the finite-frequency H∞ index bound of the
AC

control. The smaller the γ is, the smaller the influence of disturbance d(k)
on the residual r(k).
Definition 2. The filtering error system (7) has a finite-frequency H− index
bound β, if under zero initial condition, the following inequality
kr(k)k2 ≥ βkf (k)k2 (13)

7
 ACCEPTED MANUSCRIPT

holds for all solutions of (7) with f (k) in finite-frequency range ϑf 1 ≤ θf ≤

ϑf 2 such that
∞
X
ejϑf $ (η(k + 1) − ejϑf 1 η(k))(η(k + 1) − e−jϑf 2 η(k))T ≤ 0 (14)
k=0

T
where θf denotes the frequency of f (k), ϑf $ = (ϑf 2 − ϑf 1 )/2, ϑf 1 and ϑf 2

IP
reflect the frequency range of f (k). The larger the β is, the greater the
influence of fault f (k) on the residual r(k).

CR
Remark 3. Definitions 1 and 2 are motivated by the work of [37] and
[39], which are given from a time-domain perspective and can be regarded
as extensions of the standard H∞ and H− performances. If ϑl = π, ϑ1 =
−π, ϑ2 = π, or ϑh = 0, the constraints (12) and (14) are automatically

US
satisfied and the finite-frequency range reduces to full frequency range.
Remark 4. Note that the GKYP lemma ([33]) based FD methods are de-
AN
veloped from the viewpoint of transfer function matrix which makes them
cannot be directly applied for Lipschitz nonlinear systems. However, [39]
has provided time domain interpretations of finite-frequency properties. As
stated in the literature, the time domain characterization of these properties
M

are useful for capturing finite-frequency specifications for nonlinear systems

because it is given in terms of the input/output signals rather than the trans-
fer function. Recently, some results in finite frequency domain for nonlinear
ED

systems have been reported, such as [37, 38] for fuzzy systems whose sub-
systems are linear time-invariant system. However, it is hard to use these
existing finite-frequency FD results for this class of Lipschitz nonlinear net-
worked systems subject to the event-triggered scheme.
PT

The FD problem considered in this paper can be formulated as follows:

For a given Lipschitz nonlinear system (1) and two prescribed levels of distur-
bance attenuation and fault sensitivity γ > 0 and β > 0, design a nonlinear
CE

filter in the form of (3) such that under the event-triggered condition (5), the
following three specifications are satisfied:
(S.1) The filtering error system (7) is asymptotically stable.
AC

(S.2) To reduce the effects of disturbances to the residual, the filtering error
system (7) satisfies inequality
kr(k)k2 ≤ γkd(k)k2
under constraint (12).
(S.3) To increase the effects of faults to the residual, the filtering error system

8
 ACCEPTED MANUSCRIPT

(7) satisfies inequality

kr(k)k2 ≥ βkf (k)k2
with f (k) ∈ l2 under constraint (14).

T
3. LPV-based approach for Lipschitz systems

IP
The following lemma will be useful in the later developments.
Lemma 1.[21] Considering the function Φ : Rn → Rn , the two following

CR
items are equivalent:
(a) Lipschitz property: Φ is rΦ -Lipschitz with respect to its argument, i.e.:

US
kΦ(x) − Φ(y)k ≤ rΦ kx − yk, ∀x, y ∈ Rn .
(b) Reformulated Lipschitz property: for all i, j = 1, · · · , n, there exist func-
tions Φij : Rn × Rn → R and constants rΦij and r̄Φij , so that ∀x, y ∈ Rn ,
(15)
AN
n X
X n
Φ(x) − Φ(y) = Φij Hij (x − y) (16)
i=1 j=1
M

∆
and rΦij ≤ Φij ≤ r̄Φij , where Φij = Φij (xyj−1 , xyj ) and Hij = en (i)eTn (j).
Remark 5. More details about the functions Φij in (16) can be found in
ED

[21]. As stated in [21], the standard LMI methods are generally infeasible for
systems with large Lipschitz constants. Lemma 1 takes the detailed structure
of the nonlinearity into account which can reduce the conservatism of the
existing synthesis conditions for Lipschitz systems.
PT

According to Lemma 1, the Lipschitz property on Φ(x(k)) leads to the

existence of bound functions Φij : Rn × Rn → R and constants rΦij and r̄Φij ,
for all i, j = 1, · · · , n, so that
CE

rΦij ≤ Φij ≤ r̄Φij (17)

P P
∆Φ(k) = [ ni=1 nj=1 Φij (xx̂j−1 , xx̂j )Hij ]e(k) (18)
AC

where Hij = en (i)eTn (j).

To simplify the presentation, let us introduce the following notations:
n X
X n n X
X n
Λ := Φij (x0j−1 , x0j )Hij , Υ := Φij (xx̂j−1 , xx̂j )Hij (19)
i=1 j=1 i=1 j=1

9
 ACCEPTED MANUSCRIPT

and
A(Λ) = A + Λ, A(Υ) = A + Υ. (20)
Therefore, the filtering error system (7) can be re-organized into a compact
form


T
 η(k + 1) = A(Λ, Υ)η(k) + Bf f (k) + Bd d(k) + Bδ δ(k)
r(k) = C1 η(k) + Df f (k) + Dd d(k) + Dδ δ(k) (21)

IP

y(k) = C2 η(k) + Df f (k) + Dd d(k)
where

CR
 
A(Λ) 0
A(Λ, Υ) = .
0 A(Υ) − LC
Notice that in view of Lemma 1, the parameters Λ and Υ belong to bounded

US
convex set Hn for which the set of vertices is defined by
VHn = {Γ ∈ Rn×n , Γij ∈ {rΦij , r̄Φij }}. (22)
AN
In the sequel, we denote by (Λ̃, Υ̃) the vertices in VHn × VHn corresponding
to the bounded convex set Hn × Hn .
Remark 6. In this section, the nonlinear filter error system (7) is rewritten
M

under a LPV form by the use of a reformulated Lipschitz property. Then

the considered FD problem is transformed to the problem of H− /H∞ filter
design for a class of LPV systems (21) under event-trigged scheme in finite-
ED

frequency domain.

4. Main results
PT

In this section, the finite frequency H− /H∞ filter design conditions will
be given. Before proceeding further, a new lemma is derived which provides
sufficient conditions for the desired performances.
CE

Lemma 2. Assume that the filtering error system (21) is asymptotically

stable. Letting β > 0 be a given constant, if there exist symmetric matrices
P1 , Q1 > 0 and Ω > 0 such that the following inequality holds, ∀(Λ, Υ) ∈
AC

Hn × H n ,
 T  
A(Λ, Υ) Bf Bδ A(Λ, Υ) Bf Bδ
Ξf (23)
I 0 0 I 0 0
 T    T  
C1 Df Dδ C1 Df Dδ C2 Df 0 C2 Df 0
+ Πf + ‫ג‬ <0
0 I 0 0 I 0 0 0 I 0 0 I

10
 ACCEPTED MANUSCRIPT

   
−I 0 σΩ 0
where Πf = ,‫=ג‬ and
0 β 2I 0 −Ω
1) For the low-frequency range |θf | ≤ ϑf l ,
 
P1 Q1
Ξf = (24)

T
Q1 −P1 − 2cos(ϑf l )Q1
2) For the middle-frequency range ϑf 1 ≤ θf ≤ ϑf 2 ,

IP
 
P1 ejϑf c Q1
Ξf = −jϑf c (25)

CR
e Q1 −P1 − 2cos(ϑf $ )Q1
with ϑf c = (ϑf 2 + ϑf 1 )/2, ϑf $ = (ϑf 2 − ϑf 1 )/2.
3) For the high-frequency range |θf | ≥ ϑf h ,

Ξf =

P1
US −Q1
−Q1 −P1 + 2cos(ϑf h )Q1

(26)
AN
then system (21) satisfies the finite-frequency H− performance kr(k)k2 ≥
βkf (k)k2 .
Proof. We first consider the middle-frequency case for the system (21) with
d(k) = 0. Inspired by [39], multiplying the inequality (23) by [η T (k) f T (k) δ T (k)]
M

from the left and by its transpose from the right, and recalling the fact
uT Q1 v = tr(Q1 vuT ) for arbitrary vector u and v, one can derive
ED

− η T (k)P1 η(k) + η T (k + 1)P1 η(k + 1) − rT (k)r(k) + β 2 f T (k)f (k) (27)

jϑf c T −jϑf c T T
+ tr(Q1 (e η(k)η (k + 1) + e η(k + 1)η (k) − 2cos(ϑf $ )η(k)η (k)))
T T
− δ (k)Ωδ(k) + σy (k)Ωy(k) ≤ 0.
PT

Considering the event-triggered communication scheme (5), for every k ∈

[ki , ki+1 ), δ T (k)Ωδ(k) < σy T (k)Ωy(k), then we have
CE

− η T (k)P1 η(k) + η T (k + 1)P1 η(k + 1) − rT (k)r(k) + β 2 f T (k)f (k) (28)

+ tr(Q1 (ejϑf c η(k)η T (k + 1) + e−jϑf c η(k + 1)η T (k) − 2cos(ϑf $ )η(k)η T (k)))
≤ 0.
AC

Since η(0) = 0 and the system (21) is stable, taking the summation from
k = 0 to ∞, we have
∞
X
(−rT (k)r(k) + β 2 f T (k)f (k)) + tr(Q1 M ) ≤ 0 (29)
k=0

11
 ACCEPTED MANUSCRIPT

where
∞
X
M := (ejϑf c η(k)η T (k + 1) + e−jϑf c η(k + 1)η T (k) − 2cos(ϑf $ )η(k)η T (k))
k=0

T
According to Euler’s formula, it can be verified that −M is equal to the left-
hand side of (14) after some calculations. So the M is positive semidefinite.

IP
Since Q1 > 0, the last term on the P∞ left-hand side ofP
(29) is nonnegative when
(14) is satisfied. Hence we have k=0 r (k)r(k) ≥ ∞
T 2 T
k=0 β f (k)f (k), which

CR
is equivalent to condition (13) for middle-frequency domain in Definition 2.
Similarly, the results follow by choosing ϑf 2 := ϑf l and ϑf 1 = −ϑf l for the
low-frequency case and ϑf 2 := 2π − ϑf h and ϑf 1 = ϑf h for the high-frequency
case. Thus, the proof is completed.
US
Remark 7. The conditions to capture the fault sensitivity performance sub-
ject to event-triggered scheme in finite-frequency domain are given in Lemma
AN
2. However, it contains coupled matrix variables, which leads to a non-convex
problem. In the following, decoupling techniques will be developed to trans-
form the analytical conditions to linear ones.
Lemma 3.(Projection Lemma)[40] Given a symmetric matrix Ψ and two
M

matrices U, V, the problem

Ψ + UXV T + VX T U T < 0 (30)

ED

is solvable with respect to decision matrix X if and only if

T T
U ⊥ ΨU ⊥ < 0, V ⊥ ΨV ⊥ < 0 (31)
PT

where U ⊥ and V ⊥ denote the orthogonal complement of U and V, respectively.

Lemma 4.[41] Let R1 = R1T > 0 and R2 matrices of appropriate size. The
CE

following expression holds:

−R2T R1−1 R2 ≤ R1 − R2 − R2T . (32)

AC

4.1. Fault sensitivity condition

Theorem 1. Assume the filtering error system (21) is asymptotically stable.
For two given constants β > 0, ε1 > 0 and a given matrix S with compatible
dimension, if there exist matrices Ω > 0, V̄ , W11 , W12 , W , det(W ) 6= 0, Y

12
 ACCEPTED MANUSCRIPT

   
P11 P12 Q11 Q12
and symmetric matrices P1 = , Q1 = > 0 such that
∗ P13 ∗ Q13
∀(Λ̃, Υ̃) ∈ VHn × VHn , the following inequality holds:
 
Γ11 Γ12 Γ13 Γ14 Γ15 −ε1 Y

T
 ∗ Γ22 Γ23 Γ24 Γ25 −Y 
 
 ∗ ∗ Γ33 Γ34 Γ35 0 
 <0

IP
 ∗ T  (33)
 ∗ ∗ Γ44 Γ45 −C V̄ 
 ∗ ∗ ∗ ∗ Γ55 −DfT V̄ − S T Y 

CR
∗ ∗ ∗ ∗ ∗ −Ω − V̄
where
T
Γ11 = P11 − W11 − W11 , Γ12 = P12 − W12 − ε1 W T ,
T
Γ13 = µ1 Q11 + W11
T
Γ15 = W11
US
A(Λ̃), Γ14 = µ1 Q12 + ε1 W T A(Υ̃) − ε1 Y C,
Bf − W12 S + ε1 W T Bf − ε1 Y Df , Γ22 = P13 − W − W T ,
AN
Γ23 = µ1 QT12 + W12
T
A(Λ̃), Γ24 = µ1 Q13 + W T A(Υ̃) − Y C,
T
Γ25 = −W S + W12 Bf + W T Bf − Y Df , Γ33 = −P11 − µ2 Q11 + σC T ΩC,
Γ34 = −P12 − µ2 Q12 , Γ35 = σC T ΩDf + A(Λ̃)T W12 S,
M

Γ44 = −P13 − µ2 Q13 − C T V̄ C, Γ45 = −C T V̄ Df + A(Υ̃)T W S − C T Y T S,

Γ55 = β 2 I − DfT V̄ Df + σDfT ΩDf + He(BfT W S − DfT Y T S + BfT W12 S)
ED

and Y = W T L, V̄ = V T V , µ1 = 1, µ2 = 2cos(ϑf l ) for the fault in low-

frequency domain |θf | ≤ ϑf l ; µ1 = ejϑf c , µ2 = 2cos(ϑf $ ) for the fault in
middle-frequency domain ϑf 1 ≤ θf ≤ ϑf 2 ; µ1 = −1, µ2 = −2cos(ϑf h ) for the
PT

fault in high-frequency domain |θf | ≥ ϑf h , then the system (21) satisfies the
finite-frequency H− performance index kr(k)k2 ≥ βkf (k)k2 .
Proof. Based on Lemma 2, the system (21) has a finite frequency H−
CE

performance index β if (23) holds.

We first consider the low-frequency case for the system (21) with d(k) = 0.
By matrix manipulations, the inequality (23) can be rewritten as
AC

 T  
A(Λ, Υ) Bf Bδ A(Λ, Υ) Bf Bδ
 I 0 0  0 0
  Ψf  I  < 0, ∀(Λ, Υ) ∈ Hn × Hn
 0 I 0  0 I 0
0 0 I 0 0 I
(34)

13
 ACCEPTED MANUSCRIPT

where
 
P1 Q1 0 0
 ∗ Ψf 22 −C1T Df + σC2T ΩDf −C1T Dδ 
Ψf = 
∗

∗ 2 T T
β I − Df Df + σDf ΩDf −DfT Dδ 
∗ ∗ ∗ −Ω − DδT Dδ

T
with Ψf 22 = −P1 − 2cos(ϑf l )Q1 − C1T C1 + σC2T ΩC2 .

IP
The following null space bases calculations yield
 ⊥
−I  T 

CR
A (Λ, Υ)
T A (Λ, Υ) I 0 0
  =  BfT 0 I 0 (35)
 BfT  T
Bδ 0 0 I
BδT

∀(Λ, Υ) ∈ Hn × Hn , US
From Lemma 3, the following inequality is a sufficient condition for (34)
Ψf + Uf Xf VfT + Vf XfT UfT < 0 (36)
AN
which has the same form of inequality (30) with
 T
Uf = −I A(Λ, Υ) Bf Bδ , Vf = I. (37)
The matrix Xf is an introduced additional
 variable
 and can be chosen as the
M

form of Xf = W1 Sf , where Sf = I 0 S1 0 . Consequently, the inequality

(36) can be rewritten into
 
P1 − He(W1 ) Q1 + W1T A(Λ, Υ) −W1 S1 + W1T Bf W1T Bδ
ED

 ∗ ϕ22 ϕ23 −C1T Dδ 

 
 ∗ ∗ ϕ33 −Df Dδ + S1 W1 Bδ 
T T T

∗ ∗ ∗ −Ω − DδT Dδ
PT

<0 (38)
where ϕ22 = −P1 − 2cos(ϑf l )Q1 − C1T C1 + σC2T ΩC2 , ϕ23 = −C1T Df + σC2T ΩDf +
AT (Λ, Υ)W1 S1 , ϕ33 = β 2 I − DfT Df + σDfT ΩDf + He(BfT W1 S1 ).
CE

Therefore, by using the convexity principle discussed in [42], we deduce

that the inequality (38) holds if it can be verified for all (Λ̃, Υ̃) ∈ VHn ×
V
 Hn . Partitioning
 the matrices
  W1 and S1 therein into the form of W1 =
AC

W11 W12 0
and S1 = , where W is a nonsingular matrix and denoting
ε1 W W S
Y = W T L, V̄ = V T V , after some matrix manipulation, the condition (33) is
obtained for fault in low-frequency domain.
Further, by using similar techniques, we can obtain the sufficient condi-
tions for middle- and high-frequency cases. The proof is completed.

14
 ACCEPTED MANUSCRIPT

4.2. Disturbance attenuation condition

Theorem 2. Assume the filtering error system (21) is asymptotically stable.
For two given constants γ > 0 and ε2 , if there exist matricesW21 , W22, Ω > 0,
P21 P22
V̄ , W , det(W ) 6= 0, Y and symmetric matrices P2 = , Q2 =

T
  ∗ P23
Q21 Q22

IP
> 0 such that ∀(Λ̃, Υ̃) ∈ VHn × VHn , the following inequality
∗ Q23
holds:

CR
 
Θ11 Θ12 Θ13 Θ14 Θ15 −ε2 Y
 ∗ Θ22 Θ23 Θ24 Θ25 −Y 
 
 ∗ ∗ Θ Θ σC T
ΩD 0 
 33 34 d <0 (39)
 ∗

 ∗
∗
∗
∗
∗
∗
∗
∗
∗
US
∗ Θ44 C V̄ Dd T

Θ55
∗
C V̄ 
T

DdT V̄ 
−Ω + V̄
AN
where
T
Θ11 = P21 − W21 − W21 , Θ12 = P22 − W22 − ε2 W T , Θ13 = µ3 Q21 + W21
T
A(Λ̃),
M

Θ14 = µ3 Q22 + ε2 W T A(Υ̃) − ε2 Y C, Θ15 = W21

T
Bd + ε2 W T Bd − ε2 Y Dd ,
Θ22 = P23 − W − W T , Θ23 = µ3 QT22 + W22
T
A(Λ̃),
Θ24 = µ3 Q23 + W T A(Υ̃) − Y C, Θ25 = W22
T
Bd + W T Bd − Y Dd ,
ED

Θ33 = −P21 − µ4 Q21 + σC T ΩC, Θ34 = −P22 − µ4 Q22 ,

Θ44 = −P23 − µ4 Q23 + C T V̄ C, Θ55 = −γ 2 I + DdT V̄ Dd + σDdT ΩDd
PT

and Y = W T L, V̄ = V T V , µ3 = 1, µ4 = 2cos(ϑdl ) for the disturbance in low-

frequency domain |θd | ≤ ϑdl ; µ3 = ejϑdc , µ4 = 2cos(ϑd$ ) for the disturbance
in middle-frequency domain ϑd1 ≤ θd ≤ ϑd2 ; µ3 = −1, µ4 = −2cos(ϑdh ) for
CE

the disturbance in high-frequency domain |θd | ≥ ϑdh , then the system (21)
satisfies the finite-frequency H∞ performance index kr(k)k2 ≤ γkd(k)k2 .
Proof. We first consider the high-frequency case for the system (21) with
AC

f (k) = 0. Following the process of the proof of Theorem 1, if there exist

symmetric matrices P2 , Q2 > 0, Ω > 0 such that the following inequality

15
 ACCEPTED MANUSCRIPT

holds:
 T  
A(Λ, Υ) Bd Bδ A(Λ, Υ) Bd Bδ
 I 0 0  0 0
  Ψd  I  < 0, ∀(Λ, Υ) ∈ Hn × Hn ,
 0 I 0  0 I 0

T
0 0 I 0 0 I
(40)

IP
where

CR
 
P2 −Q2 0 0
 ∗ Ψd22 C1T Dd + σC2T ΩDd C1T Dδ 
Ψd = 
∗

∗ 2 T T
−γ I + Dd Dd + σDd ΩDd Dd Dδ 
T

∗ ∗ ∗
US −Ω + DδT Dδ

with Ψd22 = −P2 + 2cos(ϑdh )Q2 + C1T C1 + σC2T ΩC2 , then the system (21)
satisfies the finite-frequency H∞ performance index kr(k)k2 ≤ γkd(k)k2 .
AN
Applying Lemma 2, we can obtain that the following inequality is a suf-
ficient condition for (40), ∀(Λ, Υ) ∈ Hn × Hn ,

Ψd + Ud Xd VdT + Vd XdT UdT < 0.

M

(41)

where
ED

 T
Ud = −I A(Λ, Υ) Bd Bδ , Xd = W2 S2 , Vd = I.

Therefore, by using the convexity principle discussed in [42], we deduce that

PT

the inequality (41) holds if it can be verified for all (Λ̃, Υ̃) ∈ VHn× VHn . Par-

W21 W22
titioning the matrices W2 and S2 therein into the form of W2 =
  ε2 W W
CE

and S2 = I 0 0 0 , where W is a nonsingular matrix. Denoting Y =

W T L and V̄ = V T V , after some matrix manipulation, the condition (39) is
obtained for disturbance in high-frequency domain.
Further, by using similar techniques, we can obtain the sufficient condi-
AC

tions for low- and middle-frequency cases. The proof is completed.

Remark 8. The considered FD problem is actually a multi-constrained
design problem which may have no solution due to the increased number
of constraints. Introducing different matrices P1 , P2 and Q1 , Q2 for each
constraint can bring more degrees of freedom. However, the coupling of the

16
 ACCEPTED MANUSCRIPT

matrix variables in (34) and (40) results in a non-convex problem, so (34) and
(40) cannot be handled by the linear optimization procedures. To recover
convexity, it is required that P1 = P2 > 0 and Q1 = Q2 = 0 in full frequency
domain. In finite-frequency domain, the Projection Lemma is exploited to
separate L from P1 , P2 , Q1 and Q2 . Then novel sufficient conditions are

T
obtained in (33) and (39) and matrices P1 , P2 , Q1 and Q2 can be chosen to

IP
be different from one another.

4.3. Stability condition

CR
Then we establish the stability of the filter error system. Note that the
filter design conditions given in the previous subsections cannot ensure a
stable system, so an additional constraint should be added to guarantee the
stability of the system (21).
Theorem 3. For a given constant
W , det(W ) 6= 0, Y and P3 =

P31 P32
US
ε3 , ifthere exist matrices Ω > 0, G11 , G12 ,
> 0 such that ∀(Λ̃, Υ̃) ∈ VHn × VHn ,
AN
∗ P33
the following inequality holds:
 
k11 k12 k13 k14 −Y
M

 ∗ k22 k23 k24 −ε3 Y 

 
 ∗ ∗ k −P 0 <0 (42)
 33 32 
 ∗ ∗ ∗ −P33 0 
ED

∗ ∗ ∗ ∗ −Ω

where
PT

k11 = P31 − G11 − GT11 , k12 = P32 − W T − GT12 , k13 = G11 A(Λ̃),
k14 = W T A(Υ̃) − Y C, k22 = P33 − ε3 W − 3 W T , k23 = G12 A(Λ̃),
k24 = ε3 W T A(Υ̃) − 3 Y C, k33 = −P31 + σC T ΩC,
CE

and Y = W T L, then the system (21) is asymptotically stable.

Proof. Consider the following Lyapunov functional candidate for system
AC

(21)

V (k) = η(k)T P3 η(k) (43)

17
 ACCEPTED MANUSCRIPT

with P3 = P3T > 0. According to δ T (k)Ωδ(k) < σy T (k)Ωy(k), we have

∆V (k) = V (k + 1) − V (k)
≤ η T (k)[AT (Λ, Υ)P3 A(Λ, Υ) − P3 ]η(k) + δ T (k)BδT P3 Bδ δ(k)
+ δ T (k)BδT P3 A(Λ, Υ)η(k) + η T (k)AT (Λ, Υ)P3 Bδ δ(k)

T
+ ση T (k)C2T ΩC2 η(k) − δ(k)T Ωδ(k)

IP
= χ(k)T Ψs χ(k) (44)

CR
where
   T 
η(k) A (Λ, Υ)P3 A(Λ, Υ) − P3 + σC2T ΩC2 AT (Λ, Υ)P3 Bδ
χ(k) = , Ψs = .
δ(k) ∗ BδT P3 Bδ − Ω

US (45)

If Ψs < 0 holds, then system (21) is asymptotically stable. By the Schur

AN
complement, the problem is concluded to satisfy
 
−P3 P3 A(Λ, Υ) P3 Bδ
 ∗ −P3 + σC2T ΩC2 0  < 0, (46)
M

∗ ∗ −Ω

then pre and post-multiplying by diag (GP3−1 , I2n , I2n ) and its transpose,
ED

based on Lemma 4, we can obtain that the following inequality is a sufficient

condition for (46)
 
P3 − G − GT GA(Λ, Υ) GBδ
PT

 ∗ −P3 + σC2T ΩC2 0 <0 (47)

∗ ∗ −Ω
CE

where G is a slack matrix. Therefore, by using the convexity principle we

deduce that the inequality (47) holds if it can be verified for all (Λ̃, Υ̃) ∈ VHn
×
T
G11 W
VHn . Partitioning the matrix G therein into the form of G = ,
G12 ε3 W T
AC

where W is a nonsingular matrix and denoting Y = W T L, after some matrix

manipulation, the condition (42) is obtained. The proof is completed.
Remark 9. Although the LPV method can provide solutions even for large
Lipschitz constants, the drawback is that more demanding LMIs needed to
be solved. In [21], all the LMIs are enforced by a common Lyapunov matrix

18
 ACCEPTED MANUSCRIPT

which generally provides quite conservative results or no solution due to the

increased number of LMIs. By slack variable technique, the LMI feasibility
conditions are derived in which the product of matrix variable P3 and the fil-
ter gain matrix L is not involved. Furthermore, this extra matrix W does not
present any structural constraints, so the conservatism is reduced compared

T
with the one in [21].

IP
4.4. Algorithm
By combining Theorems 1, 2 and 3, specifications 1) − 3) given in Section

CR
2 will be satisfied if LMIs (33), (39) and (42) hold simultaneously. The
following theorem is given to solve the FD problem.
Theorem 4. Let two prescribed performance indices γ, β, three constants

there exist matrices

US
ε1 , ε2 , ε3 and a matrix S be given, system (21) subject to the event-triggered
scheme (5) is asymptotically stable and satisfies conditions (11) and (13), if
 Ω > 0, W11 , W12 , W21 , W 22 , G11 ,G12 , W , det(W ) 6= 0,
AN
P11 P12 P21 P22 P31 P32
Y , V̄ and P1 = , P2 = , P3 = > 0, Q1 =
  ∗ P13
 ∗ P 23 ∗ P33
Q11 Q12 Q21 Q22
> 0, Q2 = > 0 such that (33), (39) and (42) hold.
∗ Q13 ∗ Q23
M

T
Furthermore, the FD filter matrices L = W −1 Y , and V can be obtained
by means of the matrix V̄ , where V is a factorization of V̄ (i.e., V̄ = V T V ).
Finally, the feasible solutions to the FD problem are solved via the following
ED

optimization problem:
max β
s.t.(33), (39), (42) (48)
PT

for the given scalar γ.

Now, for designing the fault detection filter in finite-frequency domain,
CE

one can follow the following design procedures:

step 1: Compute Φij , rΦij , r̄Φij , i, j = 1, · · · , n.
step 2: Give the predefined parameter σ and γ. Construct LMIs (33), (39),
(42) and solve (48) to obtain W , Y and V̄ . If there is a solution, then it means
AC

T
that the FD filter exists. Subsequently, we can compute L by L = W −1 Y
and V by V̄ = V T V . Otherwise, the FD filer dose not exist. Consequently,
the fault detection filter can be constructed as in (3).
Remark 10. It should be claimed that, in the interest of a better fault
sensitivity performance, we solve the optimization problem (48) under an

19
 ACCEPTED MANUSCRIPT

acceptable attenuation level γ rather than the best. To balance these two
performances, we can solve the following optimization problem:
min λ1 γ − λ2 β
(49)
s.t.(33), (39), (42)

T
where λ1 , λ2 > 0 are weighting factors subject to λ1 + λ2 = 1. As pointed

IP
out in [23], a satisfactory FD performance can only be achieved by a proper
trade-off between the robustness and sensitivity. A smaller γ will result in
a smaller β, that is, a better disturbance attenuation performance can be

CR
obtained by sacrificing some level of fault sensitivity performance.
Remark 11. When solving the conditions of Theorem 4, ε1 , ε2 , ε3 and S
need to be given in advance such that (48) becomes a convex optimization

the optimization problem (48). US

problem. The role of the weighting matrix Ω is to enhance the feasibility of

Remark 12. It should be pointed out that considerable computational

AN
burdens may arise when the complexity of the function Φ(x(k)) increases.
The cost of more demanding LMIs does not play an important role on the
feasibility of (48) because the LPV method can provide less restrictive LMI
M

synthesis conditions. Besides, the constant filter gains are computed off line
for the proposed method, so the time taken to solve the LMIs will have no
effect on the real-time applications.
ED

4.5. Detection threshold design

In this section, the threshold for detecting faults is designed and the
detection logic unit is based on the results proposed by [43].
PT

The residual evaluation function Jr (τ ) is chosen as

v
u kτ
u1 X
Jr (τ ) = t rT (k)r(k) (50)
CE

τ k=k
1

where k1 denotes the initial evaluation time instant and kτ stands for the
evaluation time. The threshold Jth is determined by
AC

Jth = sup Jr (51)

f =0,d(k)∈L2

and the faults can be detected using the following logical relationship:
Jr > Jth ⇒ fault alarm, Jr ≤ Jth ⇒ no fault. (52)

20
 ACCEPTED MANUSCRIPT

5. Examples
Two examples are given to illustrate the effectiveness and applicability of
the proposed method.

T
0.2

IP
0.15

0.1

CR
0.05

−0.05

−0.1

−0.15

−0.2
0 2 4
US 6 8 10
AN
Time(s)

Figure 2: Disturbance.
M

5.1. Example 1
Consider system (1) with parameters as follows:
     
ED

−0.1 0.2 0.5 0.2  

A= , Bf = , Bd = , Dd = 0.5, C = −1 1 .
−0.3 0.1 0.1 −0.2
It is assumed that the external disturbance signal is in high-frequency domain
PT

ϑdh = π/1.2, which is shown in Figure 2, and the fault is in low-frequency

domain ϑf l = π/15.
Given γ = 0.8, we solve the optimisation problem (48) in the following
CE

simulation cases:
Case 1 (strictly proper case):
  
0.4, 10 ≤ k ≤ 80 0.1sin(x2 (k))
AC

Df = 0, f (k) = , Φ(x(k)) = .
0, otherwise 0.1sin(x1 (k))
According to Lemma 1, we have
 
0 ±0.1
VHn = .
±0.1 0

21
 ACCEPTED MANUSCRIPT

Different values of σ are considered and the corresponding performance in-

dices β are given in Table 1. It can be seen from Table 1 that the larger
the σ, the worse the fault sensitivity performance. On investigating the

T
Table 1: Comparison of the H− performance β for different σ

IP
σ σ=0.05 σ=0.1 σ=0.2 σ=0.3 σ=0.4
β 0.7768 0.5938 0.4018 0.2792 0.1786

CR
0.8 0.7

US
σ = 0.1 σ = 0.1
σ = 0.4 σ = 0.4
threshold
0.7
0.6

0.6
0.5
Residual signal |r(k)|

0.5
Res.eva.function
AN
0.4
0.4

0.3
0.3

0.2
0.2

0.1
M

0.1

0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time in samples Time in samples
ED

Figure 3: Case 1: Generated residual |r(k)| (left) and residual evaluation function Jr (k)
(right).

performance of the proposed FD filter, the initial conditions are selected as

PT

 T  T
x(0) = 0.01 −0.01 , x̂(0) = 0 0 . When σ = 0.1, the residual response
and residual evaluation response are displayed in Fig. 3. Fig. 4 demonstrates
the release instants and release intervals of the event detector. It is shown
CE

that the fault can be detected timely and the proposed method is effective for
strictly proper system. In the simulation time 10s, we observer that only 69
times are triggered. This result compares favorably with the time-triggered
AC

communication scheme (100 times). For further comparison, we also simu-

late the residual response and residual evaluation function response in the
context of the event-triggered scheme parameter σ = 0.4, which are shown
in Fig. 3. The release instants and release intervals of the event detector
are shown in Fig. 4. In the simulation time, 46 times are triggered which
is less than that in the context of σ = 0.1. The fault can also be detected,

22
 ACCEPTED MANUSCRIPT

8 16

7 14

6 12

5 10
Release interval

Release interval

T
4 8

3 6

IP
2 4

1 2

CR
0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time in samples Time in samples

Figure 4: Case 1: Release interval with σ = 0.1 (left) and release interval with σ = 0.4
(right).

but residual evaluation function response is not as sensitive as the one with
σ = 0.1 because of the less transmission of data.
US
AN
To illustrate the advantage of the proposed approach, we compare it with
the existing full-frequency method ([16, 17]) in the following case since the
method in [16, 17] is not applicable for strictly proper systems, i.e. Case 1.
Case 2 (non-proper case):
M

h r iT 
0.2, 20 ≤ k ≤ 90
Df = 1, Φ(x(k)) = 0 (x1 (k) + |x1 (k)|) , f (k) =
2 0, otherwise.
ED

we have
   
0 0 0 0
VHn = ,
0 0 r 0
PT

where r > 0 is used to boost the comparison. The Lipschitz constant of the
nonlinearity is rΦ = r. When σ = 0.1, the proposed method can tolerate
Lipschitz constants of not more than 1.78 while the existing one can only
CE

tolerate Lipschitz constants of not more than 0.48. Choosing r = 0.4, ε1 = 0,

 T
ε2 = 0, ε3 = 1 and S = −0.36 −0.56 , we can obtain the optimal value
by Theorem 4 for sensitivity performance is β = 0.8962 with the filter gains
AC

 T
L = 0.2188 −0.7911 , V = 0.7075.
The optimal value obtained by the existing full frequency method in [16, 17]
for sensitivity performance is β = 0.6121 with the filter gains
 T
L = 0.4608 0.0496 , V = 0.8582.

23
 ACCEPTED MANUSCRIPT

The residual response and residual evaluation response by these two methods

1.4
by our method 0.9
by existing method by our method
1.2 by existing method
0.8
threshold

T
0.7
1
Residual signal |r(k)|

0.6

Res.eva.function
0.8

IP
0.5

0.6 0.4

0.3

CR
0.4

0.2
0.2
0.1

0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time in samples Time in samples

Figure 5: Case 2: Generated residual |r(k)| (left) and residual evaluation function Jr (k)
(right). US
AN
are displayed in Fig. 5. From which, it is easy to see that fault alarms are
delivered timely and both of these two methods are effective for non-proper
systems, i.e. Df 6= 0. However, the generated residual in this paper is more
sensitive to fault than the one in full frequency domain. Furthermore, only
M

76 times are triggered by our method which can be seen in Fig. 6. These
results indicate that the designed FD filter in finite-frequency domain has
a better performance in saving the communication resources and detecting
ED

fault than the existing one.

5.2. Example 2
PT

Consider the model of an inverted pendulum controlled by a dc motor via

a gear train borrowed from [44], where the states are the angular position,
velocity of the inverted pendulum and the current of the dc motor xT =
CE

(x1 , x2 , x3 ) = (θp , ωp , Ia ), while the output is θp . The state-space model is



 θ̇p = ωp

 g N Km
AC

ω̇p = sin(θp ) + Ia
 l ml2

 KN R 1
 I˙a = − b ωp − a Ia + u
La La La
where Km is the motor torque constant, Kb is the back emf constant, and N
is the gear ratio. The parameters for the plant are g = 9.8 m/s2 , l = 1m,

24
 ACCEPTED MANUSCRIPT

4.5

3.5

T
3
Release interval

2.5

IP
2

1.5

CR
1

0.5

0
0 10 20 30 40 50 60 70 80 90 100
Time in samples

Figure 6: Case 2: Release interval with σ = 0.1.

m = 1kg, N = 10, Km = 0.1 Nm/A, Kb = 0.1 Vs/rad, Ra = 1Ω, La = 100

US
AN
mH. These parameters lead to
        
ẋ1 0 1 0 x1 0 0
ẋ2  = 0 0 1  x2  +  0  u + 9.8sin(x1 )
M

ẋ3 0 −10 −10 x3 10 0

 
  x1
y = 1 0 0 x2 
ED

x3
Using the Euler’s discretization method with a sampling period T = 0.1s, a
PT

discrete-time model of the system is obtained. The control input of this simu-
lation is chosen as u(k) = Kx(k), where K = −19.3239 −7.7777 −0.6908 .
Considering that there exist the disturbance input and the fault, the closed-
CE

loop state-space description is as (1) with system parameters as follows:

     
1 0.1 0 0 0
A=  0 1 0.1  , Bf = −0.01 , Bd = 0.2 ,
  
AC

−19.3239 −8.7777 −0.6908 0.02 0.1

 
  0
C = 1 0 0 , Df = 1, Dd = 0.1, Φ(x(k)) = 0.98sin(x1 (k)) .
0

25
 ACCEPTED MANUSCRIPT

According to Lemma 1, we have

 
 0 0 0 
VHn = ±0.98 0 0 .
 
0 0 0

T
In simulation, the external disturbance signal is assumed to be d(k) =

IP
0.8 0.5
by our method by our method
0.45 threshold
0.7

CR
0.4
0.6
0.35
Residual signal |r(k)|

0.5 Res.eva.function 0.3

0.4 0.25

0.2

US
0.3

0.15
0.2
0.1

0.1
0.05

0 0
AN
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time in samples Time in samples

Figure 7: Generated residual |r(k)| (left) and residual evaluation function Jr (k) (right).

0.2sin(10k) and the fault is set up as

M


0.3sin(0.01k), 10 ≤ k ≤ 80
f (k) =
0, otherwise.
ED

Given γ = 1, σ = 0.3, ϑdh = π/1.5, ϑf l = π/10. By the proposed finite-

 T
frequency FD approach under ε1 = 0, ε2 = 0, ε3 = 1 and S = 0 0 0 ,
one can obtain the filter parameters
PT

 T
L = −0.0439 0.0556 1.8674 , V = 3.4680
and the optimal value on the fault sensitivity performance index is β =
 T
CE

1.3056. Let the initial conditions be x(0) = 0.02 −0.01 0.01 , x̂(0) =
 T
0 0 0 . The generated residual r(k) and its evaluation function are
shown in Fig. 7. The release instants and release intervals are shown in
AC

Fig. 8. We observer that the fault can be easily detected and only 33 times
are triggered. That is the proposed event-triggered FD approach provides
67% reduction in the data transmission. These results further demonstrate
the effectiveness of the proposed method. It is worth noting that we fail to
design a FD filter with the existing method. So the proposed method is less
conservative than the existing one.

26
 ACCEPTED MANUSCRIPT

40

35

30

25

T
Release interval

20

IP
15

10

CR
5

0
0 10 20 30 40 50 60 70 80 90 100
Time in samples

6. Conclusions
US
Figure 8: Release interval with σ = 0.3.
AN
In this paper, we have investigated the problem of finite frequency H− /H∞
fault detection filter design under an event-triggered scheme for discrete-time
Lipschitz nonlinear networked systems. Using a reformulated Lipschitz prop-
M

erty combined with a novel lemma and slack variable techniques, a new filter
design method has been proposed which makes full use of the frequency in-
formation of disturbances and faults to reduce design conservatism. The
ED

proposed method can be applied to systems even with large Lipschitz con-
stants. The H− /H∞ filter design conditions are obtained in the formulation of
LMIs. Finally, simulation results illustrate that the proposed design method
PT

can achieve a better performance than the existing full-frequency method in

saving communication resources and detecting faults.
Although only the inverted pendulum system has been studied in the
CE

simulation, the proposed FD scheme has great potentials to apply to a vari-

ety of engineering systems. It is encouraged to extend the presented results
to more complex systems, such as stochastic dynamic systems [9, 45] and
multi-agent systems [46, 47]. In addition, besides the proposed robust FD
AC

scheme, parameter-adaptation based FD scheme [48] and signal-based di-

agnosis scheme (e.g., spectrum analysis method and phase response curve
method [49, 50]) will also be studied further in our future work.

27
 ACCEPTED MANUSCRIPT

Acknowledgements
This work was supported in part by the Funds of National Science of
China (Grant Nos. 61420106016 and 61621004), and the Research Fund
of State Key Laboratory of Synthetical Automation for Process Industries

T
(Grant no. 2013ZCX01-01).

IP
References

CR
[1] H.J. Gao, T.W. Chen, Network-based H∞ output tracking control, IEEE
Trans. Autom. Control 53 (3) (2008) 655-667.

[2] J. Hu, Z.D. Wang, S. Liu, H.J. Gao, A variance-constrained approach

US
to recursive state estimation for time-varying complex networks with
missing measurements, Automatica 64 (2016) 155-162.

[3] Z.D. Wang, B. Shen, H.S. Shu, G.L. Wei, Quantized H∞ control for non-
AN
linear stochastic time-delay systems with missing measurements, IEEE
Trans. Autom. Control 57 (6) (2012) 1431-1444.

[4] J. Hu, Z.D. Wang, D.Y. Chen, F.E. Alsaadi, Estimation, filtering and
M

fusion for networked systems with network-induced phenomena: New

progress and prospects, Inf. Fusion 31 (2016) 65-75.
ED

[5] X. Zhang, Y.Y. Han, L.G. Wu, Y.T. Wang, State estimation
for delayed genetic regulatory networks with reaction-diffusion
terms, IEEE Trans. Neural Networks Learn. Syst. (2016),
http://dx.doi.org/10.1109/TNNLS.2016.2618899.
PT

[6] C. Peng, Q.L. Han, A novel event-triggered transmission scheme and L2

control co-design for sampled-data control systems, IEEE Trans. Autom.
CE

Control 58 (10) (2013) 2620-2626.

[7] W.P.M.H. Heemels, M.C.F. Donkers, Model-based periodic event-

AC

triggered control for linear systems, Automatica 49 (3) (2013) 698-711.

[8] D. Yue, E. Tian, Q.L. Han, A delay system method for designing event-
triggered controllers of networked control systems, IEEE Trans. Autom.
Control 58 (2) (2013) 475-481.

28
 ACCEPTED MANUSCRIPT

[9] H.L. Dong, Z.D. Wang, B. Shen, D.R. Ding, Variance-constrained H∞

control for a class of nonlinear stochastic discrete time-varying systems:
The event-triggered design, Automatica 72 (2016) 28-36.
[10] L. Zhou, R. Pan, X.Q. Xiao, T.T. Sun, Event-triggered H∞ filtering for

T
discrete-time systems over unreliable networks with package dropouts,
Neurocomputing 218 (19) (2016) 346-353.

IP
[11] J. Hu, Z.D. Wang, F.E. Alsaadi, T. Hayat, Event-based filtering for time-
varying nonlinear systems subject to multiple missing measurements

CR
with uncertain missing probabilities, Inf. Fusion 38 (2017) 74-83.
[12] H.J. Wang, P. Shi, J.H. Zhang, Event-triggered fuzzy filtering for a class

168.
US
of nonlinear networked control systems, Signal Process. 113 (2015) 159-

[13] X.M. Zhang, Q.L. Han, Event-based H∞ filtering for sampled-data sys-
AN
tems, Automatica 51 (2015) 55-69.
[14] Z.W. Gao, S.X. Ding, Fault reconstruction for Lipschitz nonlinear de-
scriptor systems via linear matrix inequality approach, Circuits Syst.
M

Signal Process. 27 (3) (2008) 295-308.

[15] K. Zhang, B. Jiang, P. Shi, Observer-based integrated robust fault es-
timation and accommodation design for discrete-time systems, Int. J.
ED

Control 83 (6) (2010) 1167-1181.

[16] F.M. Guo, X.M. Ren, Z.J. Li, C.W. Hah, Fault detection for discrete-
time Lipschitz nonlinear systems with signal-to-noise ratio constrained
PT

channels, Neurocomputing 194 (19) (2016) 317-325.

[17] H.L. Dong, Z.D. Wang, H.J. Gao, Fault detection for markovian jump
CE

systems with sensor saturations and randomly varying nonlinearities,

IEEE Trans. Circuits Syst. I: Reg. Papers 59 (10) (2012) 253-262.
[18] J.G. Li, J.Q. Yuan, J.G. Lu, Observer-based H∞ control for networked
AC

nonlinear systems with random packet losses, ISA Trans. 49 (1) (2010)
39-46.
[19] M. Liu, Q.L. Wang, H.Y. Li, State estimation and stabilization for
nonlinear networked control systems with limited capacity channel, J.
Franklin Inst. 348 (8) (2011) 1869-1885.

29
 ACCEPTED MANUSCRIPT

[20] J. Hu, D.Y. Chen, J.H. Du, State estimation for a class of discrete non-
linear systems with randomly occurring uncertainties and distributed
sensor delays, Int. J. General Syst. 43 (3-4) (2014) 387-401.

[21] A. Zemouche, M. Boutayeb, On LMI conditions to design observers for

T
Lischitz nonlinear systems, Automatica 49 (2) (2013) 585-591.

IP
[22] J. Chen, R.J. Patton, Robust Model-Based Fault Diagnosis for Dynamic
Systems, Kluwer Academic Publishers, Boston, 1999.

CR
[23] S.X. Ding, Model-Based Fault Diagnosis Techniques: Design Schemes,
Algorithms and Tools, Second Edition. Springer-Verlag, London, 2013.

US
[24] X.D. Zhu, Y.Q. Xia, S. Mab, M.Y. Fu, Actuator fault detection filter
design for discrete-time systems with a descriptor system method, Neu-
rocomputing 142 (22) (2014) 318-325.
AN
[25] M.J. Khosrowjerdi, R. Nikoukhah, N. Safri-shad, Fault detection in a
mixed H2 /H∞ setting, IEEE Trans. Autom. Control 50 (7) (2016) 1063-
1068.
M

[26] X.J. Su, P. Shi, L.G. Wu, Y.D. Song, Fault Detection Filtering for Non-
linear Switched Stochastic Systems, IEEE Trans. Autom. Control 61 (5)
(2016) 1310-1315.
ED

[27] M. Chadli, A. Abdo, S.X. Ding, H− /H∞ fault detection filter design
for discrete-time Takagi-Sugeno fuzzy system Automatica 49 (7) (2013)
1996-2005.
PT

[28] J. Liu, D. Yue, Event-based fault detection for networked systems with
communication delay and nonlinear perturbation, J. Franklin Inst. 350
CE

(9) (2013) 2791-2807.

[29] S. Hajshirmohamadi, M.R. Davoodi, N. Meskin, F. Sheikholeslam,

Event-triggered fault detection and isolation for discrete-time linear sys-
AC

tems, IET Control Theory Appl. 10 (5) (2016) 526-533.

[30] M.R. Davoodi, N. Meskin, K. Khorasani, Event-Triggered multi-

objective control and fault Diagnosis: a unified framework, IEEE Trans.
Ind. Inf. 13 (1) (2017) 298-311.

30
 ACCEPTED MANUSCRIPT

[31] H.Y. Li, Z.R. Chen, L.G. Wu, H.K. Lam, H.P. Du, Event-triggered fault
detection of nonlinear networked systems, IEEE Trans. Cybern. 47 (4)
(2017) 1041-1052.
[32] Y.L. Wang, P. Shi, C. Lim, Y. Liu, Event-triggered fault detection filter

T
design for a continuous-time networked control system, IEEE Trans.
Cybern. 46 (12) (2016) 3414-3426.

IP
[33] T. Iwasaki, S. Hara, Generalized KYP lemma: unified frequency domain
inequalities with design applications, IEEE Trans. Autom. Control 50

CR
(1) (2005) 41-59.
[34] H.M. Wang, G.H. Yang, H∞ controller design for affine fuzzy sys-

US
tems based on piecewise Lyapunov functions in finite-frequency domain,
Fuzzy Sets Syst. 290 (2016) 22-38.
[35] D.W. Ding, G.H. Yang, Fuzzy filter design for nonlinear systems in
AN
finite-frequency domain, IEEE Trans. Fuzzy Syst. 18 (5) (2010) 935-
945.
[36] H. Wang, G.H. Yang, A finite frequency domain approach to fault de-
M

tection for linear discrete-time systems, Int. J. Control 81 (7) (2008)

1162-1171.
[37] H. Yang, Y. Xia, B. Liu, Fault detection for T-S fuzzy discrete systems
ED

in finite-frequency domain, IEEE Trans. Syst, Man Cybern. B, Cybern.

41 (4) (2011) 911-920.
[38] X.J. Li, G.H. Yang, Fault detection in finite frequency domain for
PT

Takagi-Sugeno fuzzy systems with sensor faults, IEEE Trans. Cybern.

44 (8) (2014) 1446-1458.
CE

[39] T. Iwasaki, S. Hara, A.L. Fradkov, Time domain interpretations of fre-

quency domain inequalities on (semi) finite ranges, Syst. Control Lett.
54 (7) (2005) 681-691.
AC

[40] P. Gahinet, P. Apkarian, A linear matrix inequality approach to H∞

control, IEEE IEEE Trans. Autom. Control 40 (1) (1995) 184-190.
[41] Y.Y. Yin, P. Shi, F. Liu, K.L. Teo, Observer-based H∞ control on non-
homogeneous Markov jump systems with nonlinear input, Int. J. Robust
Nonlin. Control 24 (13) (2014) 1903-1924.

31
 ACCEPTED MANUSCRIPT

[42] P. Apkarian, P. Gahinet, G. Becker, Self-scheduled H∞ control of linear

paremeter-varying systems: a design example, Automatica 31 (9) (1995)
1251-1261.

[43] P.M. Frank, Fault diagnosis in dymnamic systems using analytical and

T
knowledge-based redundancy: a survey and some new results, Automat-
ica 26 (1990) 459-474.

IP
[44] J.G. Kuschewski, S. Hui, S.H. Zak, Application of feedforward neural

CR
networks to dynamical system identification and control. IEEE Trans.
Control Syst. Technol. 1 (1) (1993) 37-49.

[45] J. Hu, Z.D. Wang, B. Shen, H.J. Gao, Quantised recursive filtering

US
for a class of nonlinear systems with multiplicative noises and missing
measurements, Int. J. Control 86 (4) (2013) 650-663.

[46] H.J. Ma, G.H. Yang, Adaptive fault tolerant control of cooperative het-
AN
erogeneous systems with actuator faults and unreliable interconnections,
IEEE Trans. Autom. Control 61 (11) (2016) 3240-3255.

[47] X. Wang, G.H. Yang, Distributed H∞ consensus tracking control for

M

multi-agent networks with switching directed topologies, Neurocomput-

ing 207 (26) (2016) 693-699.
ED

[48] H.J. Ma, G.H. Yang, Simultaneous fault diagnosis for robot manipula-
tors with actuator and sensor faults, Inf. Sci. 366 (2016) 10-30.

[49] I. Franović, S. Kostić, M. Perc, V. Klinshov, V. Nekorkin, J. Kurths,

PT

Phase response curves for models of earthquake fault dynamics, Chaos

26 (063105) (2016).
CE

[50] S. Kostić, I. Franović, M. Perc, N. Vasović, K. Todorović, Triggered

dynamics in a model of different fault creep regimes, Sci. Rep. 4 (5401)
(2014).
AC

32
 ACCEPTED MANUSCRIPT

Ying Gu received the B.S. and M.S. degrees in math-

ematics from Northeast Normal University, China, in 2001
and 2005, respectively. Currently, she is pursuing the Ph.D.
degree in control theory and control engineering from North-
eastern University, China. Her research interests include

T
fault detection, nonlinear systems and networked control
systems.

IP
Guang-Hong Yang received the B.S. and M.S. degrees
from Northeast University of Technology, Liaoning, China,

CR
in 1983 and 1986, respectively, and the Ph.D. degree in Con-
trol Engineering from Northeastern University, China (for-
merly, Northeast University of Technology), in 1994. He was

US
a Lecturer/Associate Professor with Northeastern University
from 1986 to 1995. He joined the Nanyang Technological
University in 1996 as a Postdoctoral Fellow. From 2001 to 2005, he was a
AN
Research Scientist/Senior Research Scientist with the National University of
Singapore. He is currently a Professor at the College of Information Science
and Engineering, Northeastern University. His current research interests in-
clude fault-tolerant control, fault detection and isolation, non-fragile control
M

systems design, and robust control. Dr. Yang is an Associate Editor for the
International Journal of Control, Automation, and Systems (IJCAS), the In-
ternational Journal of Systems Science (IJSS), the IET Control Theory &
ED

Applications, and the IEEE Transactions on Fuzzy Systems.

PT
CE
AC

33