Received: 20 December 2016 Revised: 16 September 2017 Accepted: 27 December 2017

DOI: 10.1002/rnc.4050

RESEARCH ARTICLE

Fault detection and fault-tolerant control over

signal-to-noise ratio constrained channels

Fumin Guo1 Xuemei Ren1 Zhijun Li2 Cunwu Han2

1
School of Automation, Beijing Institute of
Technology, Beijing, China Summary
2
School of Electrical and Control This paper investigates the problems of fault detection (FD) and fault-tolerant
Engineering, North China University of
control for a class of discrete-time systems with signal-to-noise ratio constrained
Technology, Beijing, China
channels. For the purpose of detecting the fault, an observer-based FD scheme is
Correspondence proposed by using a descriptor system approach. In this scheme, an augmented
Xuemei Ren, School of Automation,
robust FD filter is designed such that the residual generation system is admissi-
Beijing Institute of Technology, Beijing
100081, China. ble and the estimations of the system state and fault are both obtained; then, a
Email: [email protected] dynamic threshold is constructed based on stochastic characteristics of the chan-
Funding information nel noise, a constant 𝛼 is introduced to reduce the conservatism of threshold,
National Natural Science Foundation of and false alarm rate of the threshold is given. Subsequently, in order to keep the
China, Grant/Award Number: 61433003,
post-fault system robust asymptotically stable, an output feedback fault-tolerant
61273150, 61321002 and 61573024
control strategy is established on the basis of the fault estimation information.
Finally, two simulation examples are provided to illustrate the effectiveness and
feasibility of the proposed method.

K E Y WO R D S
fault detection, fault-tolerant control, output feedback control, signal-to-noise ratio, threshold

1 I N T RO DU CT ION

Since the early 1970s, fault detection (FD) technique has been gaining tremendous research interest owing to its sig-
nificance for the engineering reliability of modern industrial systems and the well-developed robust control theory.1
After several decades of development, FD has been widely integrated into various engineering areas such as aircraft con-
trol systems,2 electric vehicle drivetrain,3 and power generation systems4 ; meanwhile, many advanced FD approaches
have been proposed, for example, fuzzy logic-based methods,5 observer-based approaches,6 and parameter estimation
approaches.7 Among these FD methods, the observer-based approach is the most common one because it takes full use
of the system mathematical model and observer theory to design a FD filter. Briefly, the core of the observer-based FD
approach is to generate the so-called residual signal to compare with a predefined threshold, and the fault will be detected
if the residual signal surpasses the threshold. Works dedicated to analysis and design the observer-based FD approach
can be found in the works of Zhang et al,8 You et al,9 and Li and Yang.10
Originated around the same time as FD, fault-tolerant control (FTC) technique has been becoming another important
research subject since the 1970s. The main objective of FTC is to construct a controller to compensate for the effects
of faults and maintain the stability and performance of the controlled system in the presence of faults. In general, FTC
systems can be classified into active FTC and passive FTC. The basic idea of active FTC is to reconstruct a controller on
the basis of the estimated fault information to guarantee the stability of the post-fault system.11,12 By contrast, the passive
FTC is such a control scheme that it can tolerate a set of presumed system faults while the structure and parameters of

Int J Robust Nonlinear Control. 2018;1–22. wileyonlinelibrary.com/journal/rnc Copyright © 2018 John Wiley & Sons, Ltd. 1
2 GUO ET AL.

controllers are fixed.13,14 Survey papers of Yu and Jiang15 and Zhang and Jiang16 provided the excellent overviews of recent
research work on FTC.
Descriptor systems, which are also named as singular systems or generalized state-space systems, can be used to describe
many engineering systems, for example, power systems,17 economic model,18 and networked control systems (NCSs).19
Over the past decades, many results for descriptor NCS have been obtained in the literature (see the works of Zhou and
Lu,20 Du et al,21 and Qiu et al22 and the references therein). On the other hand, it is well known that practical communi-
cation channels in NCS are inevitably subject to some constraints, for example, signal-to-noise ratio (SNR) constraints or
power constraints, so one recent research line is to investigate stability of feedback control systems with SNR constrained
channels.23,24 Early significant paper on the topic includes the work of Braslavsky et al,25 which discussed the feedback
stabilization over SNR constrained channels and provided the required minimal SNR for feedback controllers to stabilize
the unstable plant.
Jiang et al26 considered the stabilization problem of NCS with additive white Gaussian noise (AWGN) channel in the
presence of packet dropouts, and the expressions of minimal SNR for stability and the relationship between SNR and
packet dropouts were provided. Then, Rojas27 addressed the step reference tracking problem of feedback control systems
with two different SNR constrained channels, and SNR lower bounds for stability were obtained. For the unstable feed-
back systems with SNR constraints and model uncertainties, Rojas and Freudenberg28 investigated the influences of model
uncertainties on the minimal SNR required for stabilizing the plant. In addition, Freudenberg et al29 and Rojas30 addressed
disturbance rejection issues of discrete-time systems with SNR constrained channels, and admissible SNR expressions for
stabilization were derived when disturbances were present. However, it should be pointed out that the noisy channels con-
sidered in the aforementioned works26-30 are subject to a total power constraint, which may cause an unconstrained power
distribution phenomenon among the channels.31,32 In order to overcome the drawback, a multiple-input–multiple-output
(MIMO) channel consists of multiple independent single-input–single-output (SISO) channels is considered. Besides,
note that most of the existing results on SNR constraints have mainly concerned with stabilization issues, and the FD and
FTC problems for feedback control systems subject to SNR constrained channels have not been well investigated.33 Up
until now, there has been little attention paid to FD for discrete-time systems subject to SNR constrained channels, and
no result has been obtained on FD and FTC for the systems, which motivates this paper.
The main contributions of this paper include the following: (i) The problems of FD and FTC for the discrete-time
systems over SNR constrained channels are first achieved. (ii) A dynamic threshold is designed based on stochastic char-
acteristics of channel noise, a constant 𝛼 is introduced to reduce the conservatism of threshold, and false alarm rate
(FAR) of the threshold is given. (iii) An H∞ FTC strategy is proposed to keep the post-fault system robust asymptotically
stable.
The remainder of this paper is organized as follows. Section 2 presents the preliminaries. Section 3 provides the design
procedure of the observer-based FD scheme that includes an augmented robust FD filter, a dynamic threshold, and
FAR. An output feedback FTC strategy is discussed in Section 4. Two simulation examples are given in Section 5, and a
conclusion is shown in Section 6.

2 PRELIMINARIES

Notations. Rn denotes the n-dimensional Euclidean space, and Rn×m is the set of real n × m matrices. Given a symmetric
matrix P, the notation P > 0 (P < 0) denotes that P is a positive definite (negative definite) matrix. I and 0 represent the
identity matrix and a zero matrix, respectively. In symmetric block matrices or long matrix expressions, the symbol “∗”
denotes a term that is induced by symmetry. diag{· · ·} stands for a block-diagonal matrix. The superscript “T” stands
for matrix transposition. deg(A) and det(A) denote the degree and determinant of the square matrix A, respectively. E{·}
denotes the mathematical expectation. Pb {X} represents the occurrence probability of the event X. |x| denotes Euclidean
norm if x is a vector√and absolute value if x is a scalar. l2 [0, ∞) is the space of square integrable vector-valued function over
∑∞ T
[0, ∞). ||e||E,2 = E{ k=0 e (k)e(k)}. Dr (q, 𝜏) stands for a closed circular region with center q + j0 and radius 𝜏.
Consider the following discrete-time systems over a MIMO AWGN channel, as shown in Figure 1:

{
x (k + 1) = Ax(k) + Bu(k) + Ed d(k) + E𝑓 𝑓 (k)
(1)
𝑦 (k) = Cx(k) + Fd d(k),
GUO ET AL. 3

FIGURE 1 The structure of fault detection (FD) and fault-tolerant control for the feedback control systems with signal-to-noise ratio
constrained channels

where x(k) ∈ Rn and u(k) ∈ Rp are the state vector and system input, respectively; y(k) ∈ Rm is the system output;
d(k) ∈ Rnd is the bounded unknown disturbance; and 𝑓 (k) ∈ Rn𝑓 denotes the fault. A, B, C, Ed , Ef , and Fd are known
constant matrices with the appropriate dimensions.
Throughout this paper, the following general assumptions are considered.
Assumption 1. The MIMO AWGN channel is comprised of m independent SISO channels, and such a chan-
nel is described by w(k) = y(k) + n(k), where w(k) = [w1 (k) w2 (k) · · · wm (k)]T is the channel output, n(k) =
[n1 (k) n2 (k) · · · nm (k)]T is the channel noise, and each SISO channel is subject to a SNR constraint with the definition

P𝑦𝑗
S𝑗 = , 𝑗 = 1, 2, … , m, (2)
𝜎n2𝑗

where Sj and P𝑦𝑗 are the SNR and the channel input power of the jth SISO channel, respectively.

Assumption 2. The channel noise n(k) is a zero mean independent and identically distributed Gaussian white noise
process with a positive diagonal covariance matrix Pnc = diag(𝜎n21 , 𝜎n22 , … , 𝜎n2m ).

Remark 1. In Assumption 1, we restrict the power in each SISO channel, instead of imposing a total power constraint
on all the channels. Thus, each SISO branch of the MIMO channel is assumed to be subject to an individual SNR
constraint. This setting can also be found in the works of Vargas et al.24,32,34

Remark 2. Assumption 2 characterizes the class of channel noise under consideration, and it is a common assumption
in SNR constrained feedback systems as presented in the works of Vargas et al.31,35
This paper is devoted to investigating the FD and FTC problems of system (1) subject to SNR constrained channels
imposed by (2). For the aforementioned purposes, an augmented robust FD filter is designed by a descriptor system
method; then, a dynamic threshold is constructed based on stochastic characteristics of channel noise; finally, on the
basis of the estimated fault information, an output feedback FTC strategy is proposed to keep the post-fault system robust
asymptotically stable.

3 OBSERV ER-BASED FAULT D ETECTION SCHEME

In this section, an observer-based FD scheme including an augmented robust FD filter, a threshold, and FAR will be
designed. In the following, the pair (A, C) is assumed to be detectable.
4 GUO ET AL.

3.1 Robust FD filter design

In this section, we will construct a descriptor system to design the augmented robust FD filter, and estimations of the
system state and the fault will be obtained in the designed filter.
Define x(k) = [x(k); d(k)], for system (1), a descriptor system can be described as
{
E x(k + 1) = Ax(k) + Bu(k) + Ed d(k) + E𝑓 𝑓 (k)
(3)
𝑦(k) = 𝑦(k) = C x(k),
where [ ] [ ] [ ] [ ] [ ]
I 0 A 0 B Ed E𝑓 [ ]
E = 0 0 , A = 0 −I , B = 0 , Ed = I , E𝑓 = 0 , C = C Fd .

For the purpose of generating the residual signal, a FD filter is designed as

⎧̂x(k + 1) = AF ̂
x(k) + BF w(k)
⎪ ̂
⎨r(k) = CF x(k) + DF w(k) (4)
⎪
⎩w(k) = 𝑦(k) + n(k),

where ̂ x(k); ̂
x(k) = [̂ d(k)] ∈ Rn+nd is the state vector of the FD filter; r(k) is the so-called residual signal; AF , BF , CF , and DF
are the free design matrices.
According to the work of Wang et al,36 the estimation of f(k) can be formulated as a weighted fault 𝑓̂w (z) = W𝑓 (z)𝑓 (z)
with Wf (z) being a given stable weighting matrix, and then, a minimal realization of 𝑓̂w (z) can be given by
{
xw (k + 1) = Aw xw (k) + Bw 𝑓 (k)
(5)
𝑓̂w (k) = Cw xw (k) + Dw 𝑓 (k),

where xw ∈ Rnw is the state vector of the weighted fault; 𝑓̂w (k) ∈ Rn𝑓 and 𝑓 (k) ∈ Rn𝑓 are the weighted fault and the
original fault, respectively; Aw , Bw , Cw , and Dw are known matrices.
Remark 3. As shown in the work of Wang et al,36 the introduction of Wf (z) is used to limit the frequency interval, and
the fault can also be identified. On the basis of the obtained fault estimation information, then in Section 4, we will
design a fault-tolerant controller to improve the system performance.
Let re (k) = r(k) − 𝑓̂w (k), the augmented overall system is described as
{
Ee e(k + 1) = Ae e(k) + Be v(k)
(6)
re (k) = Ce e(k) + De v(k),

where
⎡ u(k) ⎤ [ ]
⎡ x(k) ⎤ E 00 ⎡ A 0 0 ⎤
⎢ d(k) ⎥
e(k) = ⎢ ̂x(k) ⎥ , v(k) = ⎢ ⎥ , E e = 0 I 0 , Ae = ⎢B C A 0 ⎥,
⎢ ⎥ ⎢
n(k)
⎥ ⎢ F F ⎥
⎣ xw (k) ⎦ ⎣ 𝑓 (k) ⎦
0 0 I ⎣ 0 0 Aw ⎦
[ ]
B Ed 0 E𝑓
Be = 0 0 BF 0 , Ce = [DF C CF − Cw ], De = [0 0 DF − Dw ] .
0 0 0 Bw

Before designing the robust FD filter, the following definitions should be stated.
Definition 1. (See the work of Dai37 )
The descriptor system Ee e (k + 1) = Ae e(k) is said to be admissible if it is regular, causal, and stable, as follows.
1. The pair (Ee , Ae ) is regular if det (zEe − Ae ) is not identically equal to zero.
2. The pair (Ee , Ae ) is causal if deg(det(zEe − Ae )) = rank(Ee ).
3. The pair (Ee , Ae ) is stable if generalized spectral radius 𝓁 (Ee , Ae )< 1 with 𝓁(Ee , Ae ) = max |𝜆|.
𝜆∈{z| det(z Ee −Ae )=0}
GUO ET AL. 5

Definition 2. (See the work of Hu et al38 )

For any initial condition e0 and v(k) = 0, system (6) is said to be stochastically stable if the following holds:
{N }
∑ T
lim E e (k)e(k) < ∞. (7)
N→∞
k=0

Then, according to Definitions 1 and 2, a sufficient condition for the admissibility of system (6) will be obtained in the
following theorem.
Theorem 1. For a given 𝛾 > 0, if there exist the positive definite matrices P and Q such that
T
Ee PEe ≥ 0 (8)

⎡ −Q−1 0 0 I 0 ⎤
⎢ ∗ −I 0 C e De ⎥
⎢ ⎥
⎢ ∗ ∗ −P−1 Ae Be ⎥ < 0 (9)
⎢ ∗ ∗ ∗ −P 0 ⎥
⎢ ∗ ∗ −𝛾 2 I ⎥⎦
⎣ ∗ ∗
T
hold, where P = Ee PEe + Q, then system (6) is admissible and satisfies the following H∞ performance:

‖re (k)‖E,2 < 𝛾‖v(k)‖E,2 .

Proof. See Appendix A.1.

Remark 4. From inequality (9), it can be seen that it is difficult to handle the problem of robust FD filter design because
of the coupling form. In order to overcome the difficulty, we introduce an auxiliary slack nonsingular matrix Φ.
Premultiplying diag{ΦT , I, ΦT , I, I} and postmultiplying diag {Φ, I, Φ, I, I} to inequality (9), then we obtain
⎡ −ΦT Q−1 Φ 0 0 ΦT 0 ⎤
⎢ ∗ −I 0 Ce De ⎥
⎢ ⎥
⎢ ∗ ∗ −ΦT P−1 Φ ΦT Ae ΦT Be ⎥ < 0, (10)
⎢ ∗ ∗ ∗ −Ee PEe − Q 0 ⎥
T
⎢ ⎥
⎣ ∗ ∗ ∗ ∗ −𝛾 2 I ⎦
Since P is a positive definite matrix and Φ is nonsingular, we have

(P − Φ)T P−1 (P − Φ) ≥ 0,
ie,
( )
−ΦT P−1 Φ ≤ P − ΦT + Φ . (11)
Similarly, we obtain
( )
−ΦT Q−1 Φ ≤ Q − ΦT + Φ . (12)
Substituting inequalities (11) and (12) into (10), then we have the following decoupling inequality:
( T
)
⎡Q− Φ+Φ 0 0 ΦT 0 ⎤
⎢ ∗ −I 0 Ce De ⎥
⎢ ( ) ⎥
⎢ ∗ ∗ P − Φ + ΦT ΦT Ae ΦT Be ⎥ < 0. (13)
⎢ T ⎥
⎢ ∗ ∗ ∗ −Ee PEe − Q 0 ⎥
⎣ ∗ ∗ ∗ ∗ −𝛾 I ⎦
2

After discussing the sufficient condition for the admissibility of system (6), then the FD filter (4) will be obtained
in Theorem 2.

Theorem 2. For a given 𝛾 > 0, if there exist the positive definite symmetric matrices P11 , P22 , P33 , Q11 , Q22 , Q33 , Δ11 , Δ22 ,
and Δ33 and the matrices P12 , P13 , P23 , Q12 , Q13 , Q23 , Δ12 , Δ13 , Δ23 , X, Y, G, Z, A𝑓 , B𝑓 , C𝑓 , and D𝑓 such that the linear
6 GUO ET AL.

matrix inequality (14) holds

⎡ Ψ11 0 0 Ψ14 0 ⎤
⎢ ∗ −I 0 Ψ24 Ψ25 ⎥
⎢ ∗ ∗ Ψ33 Ψ34 Ψ35 ⎥ < 0, (14)
⎢ ∗ ∗ ∗ Ψ44 0 ⎥
⎢ ⎥
⎣ ∗ ∗ ∗ ∗ −𝛾 2 I ⎦
where
[ ]
⎡ Q11 − X Q12 − M Q13 ⎤ X X 0
Ψ11 =⎢ ∗ Q22 − Y Q23 ⎥ , Ψ14 = Y + G Y 0
⎢ ⎥ 0 0 Z
⎣ ∗ ∗ Q33 − Z ⎦
[ ] [ ]
Ψ24 = D𝑓 C + C𝑓 D𝑓 C − Cw , Ψ25 = 0 0 D𝑓 − Dw

⎡ P11 − X P12 − M P13 ⎤ ⎡ XA XA 0 ⎤

Ψ33 = ⎢ ∗ P22 − Y P23 ⎥ , Ψ34 = ⎢ Y A + B𝑓 C + A𝑓 Y A + B𝑓 C 0 ⎥
⎢ ⎥ ⎢ ⎥
⎣ ∗ ∗ P33 − Z ⎦ ⎣ 0 0 ZAw ⎦
[ ]
⎡ XB XEd 0 XE𝑓 ⎤ −Δ11 − Q11 −Δ12 − Q12 −Δ13 − Q13
⎢ ⎥
Ψ35 = Y B Y Ed B𝑓 Y E𝑓 , Ψ44 = ∗ −Δ22 − Q22 −Δ23 − Q23
⎢ ⎥ ∗ ∗ −Δ33 − Q33
⎣ 0 0 0 ZBw ⎦

and X = X + X T , Y = Y + Y T , Z = Z + Z T , M = X + Y T + GT , then the H∞ FD filter (4) can be obtained as AF = A𝑓 G−1 ,

BF = B𝑓 , CF = C𝑓 G−1 , DF = D𝑓 .

Proof. See Appendix A.2.

3.2 Threshold computation and FAR

In this section, we will discuss the threshold computation and FAR. Since channel noise is a stochastic process, we can
design the threshold based on stochastic characteristics of channel noise.
As in most works, the threshold Jth (Jth > 0) is adopted according to the following FD logic:
{
||re𝑓 (k)||2,𝜌 < Jth , fault-free
(15)
||re𝑓 (k)||2,𝜌 ≥ Jth , alarm for fault,

where the residual evaluation function ||r ef (k)||2,𝜌 is determined by

√
√
√ ∑ k
√1
||re𝑓 (k)||2,𝜌 = √ r T (i)r(i) (16)
𝜌 i=k−𝜌+1

and 𝜌 is the length of evaluation window.

Then, recall systems (1) and (4)

⎧x(k + 1) = A x(k) + Bu(k) + E d(k) + E 𝑓 (k)

𝑓
⎪ d
⎪𝑦(k) = Cx(k) + Fd d(k)
⎪ ̂
⎨x (k + 1) = AF x (k) + BF w(k) (17)
⎪ ̂
⎪r(k) = CF x (k) + DF w(k)
⎪w(k) = 𝑦(k) + n(k).
⎩

Define ̃
x (k) = [x (k); ̂
x (k)], we obtain
{
̃
x (k + 1) = Ã̃ ̃u(k) + E
x (k) + B ̃d d(k) + E
̃𝑓 (k) + B
̃F n(k)
̃ (18)
r(k) = C̃ x (k) + F
̃d d(k) + DF n(k),
GUO ET AL. 7

where
[ ] [ ] [ ] [ ]
̃= A 0 ̃ = B ,Ẽd = Ed ̃ 𝑓 = E𝑓 ,
A BF C AF , B 0 BF Fd , E 0
[ ]
̃F = 0 , C
B ̃ = [DF C CF ] , F
̃d = DF Fd .
B F

Then, we have
||r(k)||2,𝜌 = ||ru (k) + rd (k) + r𝑓 (k) + rn (k)||2,𝜌 , (19)
where
ru (k) = r(k)|d=0,𝑓 =0,n=0
rd (k) = r(k)|u=0,𝑓 =0,n=0
r𝑓 (k) = r(k)|u=0,d=0,n=0
rn (k) = r(k)|u=0,d=0,𝑓 =0 .
Besides, the fault-free residual evaluation function is
||ru (k) + rd (k) + rn (k)||2,𝜌 ≤ ||ru (k)||2,𝜌 + ||rd (k)||2,𝜌 + ||rn (k)||2,𝜌 (20)
≤ 𝛼(Jth,u + Jth,d + Jth,n ),
where 𝛼 ∈ (0, 1] is introduced to reduce the conservatism of threshold and
Jth,u = sup ||ru (k)||2,𝜌
u

Jth,d = sup ||rd (k)||2,𝜌

d

Jth,n = sup ||rn (k)||2,𝜌 .

n

Then, the threshold Jth can be chosen as

Jth = 𝛼(Jth,u + Jth,d + Jth,n ) (21)
and Jth,u can be evaluated by
Jth,u = 𝛾ru ||u(k)||2,𝜌 , (22)
||ru (k)||2
where u is assumed to be known online with 𝛾ru = supu ||u(k)|| . According to Theorem 1, 𝛾 ru is calculated by
2

min 𝛾ru
subject to [ ]
A ̃T C
̃ − Pu + C
̃ T Pu A ̃ A ̃
̃ T Pu B
< 0.
∗ BT ̃ − 𝛾ru
̃ Pu B 2
I
In the same way, we obtain
Jth,d = 𝛾rd ed,𝜌 , (23)
||rd (k)||2
where ||d(k)||2,𝜌 ≤ ed,𝜌 and γrd = sup ||d(k)||2
is evaluated by
d

min 𝛾rd
subject to [ ]
A ̃T C
̃ − Pd + C
̃ T Pd A ̃ A ̃T F
̃d + C
̃ T Pd E ̃d
< 0.
∗ E ̃d + F
̃ T Pd E ̃d − 𝛾 2 I
̃T F
d d rd
Similarly, we have
Jth,n = 𝛾rn ||n(k)||2,𝜌 , (24)
||rn (k)||2
where γrn = sup ||n(k)||2
, and γrn is determined by
n
min 𝛾rn
subject to [ ]
̃ T Pn A
A ̃T C
̃ − Pn + C ̃ ̃ T Pn B
A ̃ T DF
̃F + C
̃ ̃ < 0.
∗ BF Pn BF + DF DF − 𝛾rn
T T 2
I
8 GUO ET AL.

Since channel noise n(k) is a stochastic process, it is reasonable to set Jth,n based on stochastic characteristics of n(k), ie,
the mathematical expectation and variance.39 However, we cannot obtain exact values of these parameters, here we use
the upper bounds instead
{ √ { } √ { }}
1
Jth,n = γrn sup E ||n(k)||2,𝜌 + 𝜇 · sup 𝜎 ||n(k)||2,𝜌
2 2
, (25)
𝛼 n n

where 𝜎(·) is standard deviation and 𝜇 > 1∕𝛼 2 is a constant.

Then, for the m-dimensional vector n(k) as shown in Assumption 1, the aforementioned two terms in (25) can be
obtained as (D(·) is variance)
{ }
{ }
1 ∑ T
k
E ||n(k)||2,𝜌 = E
2
n (i)n(i)
𝜌 i=k−𝜌+1

1 ∑
k
{ }
= E nT (i)n(i) (26)
𝜌 i=k−𝜌+1

1 ∑ ( 2
k
)
= 𝜎 (i) + 𝜎n22 (i) + · · · + 𝜎n2m (i)
𝜌 i=k−𝜌+1 n1

{ }
{ }
1 ∑ T
k
D ||n(k)||22,𝜌 =D n (i)n(i)
𝜌 i=k−𝜌+1
{ k }
1 ∑
T
= 2D n (i)n(i) (27)
𝜌 i=k−𝜌+1

1 ∑
k
{ }
= D n21 (i) + n22 (i) + · · · + n2m (i)
𝜌 i=k−𝜌+1
2

2 ∑ { 4
k
}
= 𝜎n1 (i) + 𝜎n42 (i) + · · · + 𝜎n4m (i) .
𝜌 i=k−𝜌+1
2

On the other hand, the FD system cannot get y(k) because n(k) is unknown, but y(k) can be estimated as

𝑦2𝑗 (k) ≤ 𝛾𝑗u

2
||u(k)||22,𝜌 + 𝛾𝑗d
2
||d(k)||22,𝜌 , (28)

Then, γju and γjd are determined by

min 𝛾𝑗u
subject to [ ]
AT P𝑗u A − P𝑗u + C𝑗T C𝑗 AT P𝑗u B
<0
∗ B P𝑗u B − 𝛾𝑗u
T 2
I
and by
min 𝛾𝑗d
subject to [ ]
AT P𝑗d A − P𝑗d + C𝑗T C𝑗 AT P𝑗d Ed + C𝑗T Fd𝑗
< 0,
∗ T T
Ed P𝑗d Ed + Fd𝑗 Fd𝑗 − 𝛾𝑗d
2
I
where Cj and Fdj respectively denote the jth line elements of C and Fd .
Let
P𝑗 (k) = 𝛾𝑗u
2
||u(k)||22,𝜌 + 𝛾𝑗d
2 2
ed,𝜌 , (29)
and based on the SNR constraint (2), we have
P𝑗 (k)
≥ S𝑗 , (30)
𝜎n2𝑗
then we can obtain { } ( P (k) P (k) Pm (k)
)
1 2
E ||n(k)||22,𝜌 ≤ + +···+ (31)
S1 S2 Sm
GUO ET AL. 9

√ {
{ } √√ 2 ( P1 (k) )2 ( P2 (k) )2 ( )2 }
√ Pm (k)
𝜎 ||n(k)||2,𝜌 ≤
2
+ +···+ . (32)
𝜌 S1 S2 Sm

Therefore, Jth,n can be rewritten as

√ √
⎧ √( √ ((
) √ √ )2 ( )2 ( )2 )⎫
⎪1 P1 (k) P2 (k) Pm (k) √ √ √ 2 P1 (k) P2 (k) Pm (k) ⎪
= 𝛾rn ⎨ √
+ 𝜇·
Jth,n + +···+ + +···+ ⎬. (33)
⎪𝛼 S1 S2 Sm 𝜌 S1 S2 Sm ⎪
⎩ ⎭

Consequently, Jth = Jth,u + Jth,d + Jth,n can be obtained from Equations (22), (23), and (33).
Remark 5. In the threshold design process, the conservatism of threshold is reduced by introducing 𝛼, which provides
a reliable estimation of the value range of threshold. Therefore, one may choose a suitable 𝛼 on the basis of experiences
and/or trial and error to design an appropriate threshold in practice.
Next, we will discuss the computation of FAR. Since FAR is an essential performance index for an FD system, it
is reasonable to use FAR to evaluate the designed FD system. In order to compute the FAR, we have the following
theorem.

Theorem 3. For 𝜇 and 𝛼, the FAR of the threshold (21) has a supremum 1∕(𝛼)𝜇2 .

Proof. See Appendix A.3.

4 FTC STRAT EGY

In this section, we will propose an output feedback FTC strategy on the basis of the fault estimation information to keep
the post-fault system robust asymptotically stable.
First, an assumption and a lemma are provided for the FTC strategy.
Assumption 3. rank(B, Ef ) = rank(B)

Remark 6. According to the work of Jiang et al,40 Assumption 3 is equivalent to the existence of matrix B∗ ∈ Rp×n
such that
(I − BB∗ )E𝑓 = 0. (34)

Lemma 1. (See the work of Garcia and Bernussou41 )

For a given matrix A ∈ Rn×n , the eigenvalues of A belong to Dr (q, 𝜏) if and only if there exists a symmetric matrix P ∈ Rn×n
such that
[ ]
−P PA − qP
< 0. (35)
∗ −𝜏 2 P

Then, an output feedback fault-tolerant controller can be constructed as

{
𝜉(k + 1) = Ak 𝜉(k) + Bk w(k)
(36)
u(k) = Ck 𝜉(k) + Dk w(k) − B∗ E𝑓 𝑓̂w (k),

where 𝜉(k) ∈ Rn is the state vector and Ak , Bk , Ck , and Dk are design matrices.
Substitute w(k) into (36), then we have
{
𝜉(k + 1) = Ak 𝜉(k) + Bk Cx(k) + Bk Fd d(k) + Bk n(k)
(37)
u(k) = Ck 𝜉(k) + Dk Cx(k) + Dk Fd d(k) + Dk n(k) − B∗ E𝑓 𝑓̂w (k).
10 GUO ET AL.

Define e𝑓 (k) = 𝑓̂w (k) − 𝑓 (k), based on (1) and (37), we can obtain
( )
x(k + 1) = Ax (k) + B Ck 𝜉(k) + Dk Cx(k) + Dk Fd d(k) + Dk n(k) − B∗ E𝑓 𝑓̂w (k) + Ed d(k) + E𝑓 𝑓 (k)

= (A + BDk C)x(k) + BCk 𝜉(k) + (I − BB∗ )E𝑓 𝑓̂w (k) − E𝑓 𝑓̂w (k) + E𝑓 𝑓 (k) + (Ed + BDk Fd )d(k) + BDk n(k)
= (A + BDk C)x(k) + BCk 𝜉(k) − E𝑓 e𝑓 (k) + (Ed + BDk Fd )d(k) + BDk n(k). (38)

Thus, we have
⎧⌣ ⌣⌣ ⌣
⎪ x (k + 1) = A x (k) + D1 𝜃(k)
⎨ ⌣⌣ ⌣ (39)
⎪w(k) = C x (k) + D2 𝜃(k),
⎩
where
[ ] [ ] [ ]
d(k)
⌣ x(k) ⌣
A + BDk C BCk ⌣
x (k) = 𝜉(k) , 𝜃(k) = e𝑓 (k) ,A = Bk C Ak , C = [C 0]
[ n(k) ]
⌣
BDk Fd + Ed −E𝑓 BDk ⌣
D1 = B F 0 B , D2 = [Fd 0 I] .
k d k
Now, the fault-tolerant controller will be given in Theorem 4.

Theorem 4. For a given disk Dr (q, 𝜏), if there exist the positive definite matrices U, V ∈ Rn×n and the matrices A ∈ Rn×n ,
B ∈ Rn×m , C ∈ Rp×n , D ∈ Rp×m such that
⎡ −U −I AU + BC − qU A + BDC − qI ⎤
⎢ ⎥
⎢ ∗ −V A − qI VA + BC − qV ⎥ < 0 (40)
⎢ ∗ ∗ −𝜏 2 U −𝜏 2 I ⎥
⎢ ⎥
⎣ ∗ ∗ ∗ 2
−𝜏 V ⎦
⎡ −U −I AU + BC A + BDC Ed + BDFd −E𝑓 0 ⎤
BD
⎢ ⎥
⎢ ∗ −V A VA + BC VEd + BFd −VE𝑓 B 0 ⎥
⎢ ∗ ∗ −U −I 0 0 0 UCT ⎥
⎢ ⎥
⎢ ∗ ∗ ∗ −V 0 0 0 CT ⎥ < 0 (41)
T
⎢ ∗ ∗ ∗ ∗ −𝛾𝜃 I 0 0 Fd ⎥
⎢ ∗ ∗ ∗ ∗ ∗ −𝛾𝜃 I 0 0 ⎥⎥
⎢
⎢ ∗ ∗ ∗ ∗ ∗ ∗ −𝛾𝜃 I I ⎥
⎣ ∗ ∗ ∗ ∗ ∗ ∗ ∗ −𝛾𝜃 I ⎦
⌣
hold; then, the eigenvalues of A belong to Dr (q, 𝜏), system (39) is robust asymptotically stable and satisfies the H∞
performance ||w(k)||E,2 < 𝛾 𝜃 ||𝜃(k)||E,2 , and the fault-tolerant controller (36) is
( )
−1
Ak = N A − V (A + BDk C) U M −T − Bk CUM −T − N −1 VBCk
( ) ( )
Bk = N −1 B − VBDk , Ck = C − Dk CU M −T , Dk = D,

where M, N ∈ Rn×n satisfy MNT = I − UV.

Proof. See Appendix A.4.

5 S I M UL ATION R E SU LT S

In this section, two examples including a networked two-tank system and a numerical example are given to show the
effectiveness of the proposed method, and the following parameters are given for the two examples.
Suppose the sampling period h = 0.1 seconds, SNR = 35 dB, and 𝜌 = 20. Set 𝛼 = 0.95 and 𝜇 = 3.9, and then,
the supremum of FAR is 0.0807. The disturbance d(k) is a random signal uniformly distributed between [0, 0.025]. The
weighted matrix of the fault is supposed to be Wf (z) = 0.5z∕(z − 0.5), ie, Aw = 0.5, Bw = 0.25, Cw = 1, and Dw = 0.5, and
the cut-off frequency of the low-pass filter is 0.1461𝜋.
GUO ET AL. 11

FIGURE 2 The physical description of the two-tank system

Example 1. Consider the following networked two-tank system with the physical description as shown in Figure 2.42
In this two-tank system, Lti and Qti (i = 1, 2) are the levels and flow rates of the two tanks, respectively. The physical
equations of the benchmark system are
St L̇t1 = Qt1 − Qt12
St L̇t2 = Qt2 + Qt12 − Qt20 , (42)
√ √
where Qt12 = sgn(Lt1 − Lt2 )Rt |Lt1 − Lt2 |, Qt20 = Rt Lt2 , St is the section of the tank, and Rt is the throttling of the
valve.
Let [ ] [ ] [ ] [ ]
x Lt1 u1 Qt1
x = x1 = L ,u = u = Q ,
2 t2 2 t2

then (42) can be rewritten as

[ ] [ ]
1 −Qt12 1 10
̇ =
x(t) + u(t)
St Qt12 − Qt20 St 0 1
[ ]
10
𝑦(t) = 0 1 x(t). (43)

By using the Taylor series expansion around an operating point with x0 = [x10 x20 ]T , u0 = [u10 u20 ]T , y0 = x0 and by
neglecting the higher order terms, the (43) is linearized as
̇ = Al Δx(t) + Bl Δu(t)
Δx(t)
Δ𝑦(t) = Cl Δx(t) (44)
with
Δx = x − x0 , Δu = u − u0 , Δ𝑦 = 𝑦 − 𝑦0
and [ ] [ ] [ ]
1 −𝜙t1 𝜙t1 1 10 10
Al = , B = , C = 01 ,
St 𝜙t1 −𝜙t1 − 𝜙t2
l l
St 0 1
√ √
Rt 1 R
where 𝜙t1 = 2 |x10 −x20 |
and 𝜙t2 = 2t x1 .
20
Considering the disturbances and faults existed in the system, and the corresponding coefficient matrices are assumed
as [0.7 0.3; 0.3 0.7] and [0; 1], respectively. Then, by taking St = 13 , Rt = 1, x10 = 0.5, x20 = 0.25, u10 = 0.25, u20 = 0, we have
[ ] [ ] [ ] [ ]
−3 3 30 0.7 0.3 0
̇ =
x(t) 3 −6 x(t) + 0 3 u(t) + 0.3 0.7 d(t) + 1 𝑓 (t)
[ ]
10
𝑦(t) = 0 1 x(t). (45)

A discrete-time controller is chosen as [ ]

0.46786 −0.28916
Kc = −0.19708 0.16993 .
12 GUO ET AL.

2.5

Amplitude
1.5

0.5
the fault signal
the weighted fault signal
0
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 3 The fault signal f(k) and the weighted fault signal 𝑓̂w (k) [Colour figure can be viewed at wileyonlinelibrary.com]

The discrete system matrices can be obtained as

[ ] [ ] [ ]
0.7713 0.1949 0.2625 0.0338 0.0646 0.0341
A = 0.1949 0.5764 , B = 0.0338 0.2287 , Ed = 0.0308 0.0568 ,
[ ] [ ]
0.0113 10
E𝑓 = 0.0762 , C = 0 1 .

In the simulation, the fault is considered as

⎧0 0 ≤ k < 20
⎪ ( )
𝑓 (k) = ⎨3 1 − e−0.5(k−20) 20 ≤ k < 80
⎪
⎩0 80 ≤ k < 100.

Then, the fault f(k) and the weighted fault 𝑓̂w (k) are given in Figure 3, in which an accurate estimation of f(k) is
presented.
Then, we have 𝛾 ru = 0.0517, 𝛾 rd = 0.0170, 𝛾 rn = 0.0216, 𝛾 1u = 2.2366, 𝛾 2u = 1.4144, 𝛾 1d = 0.7134, 𝛾 2d = 0.4715. By
using the proposed FD approach, for a given 𝛾 = 0.9986, the FD filter in Theorem 2 is obtained as
⎡ 0.5266 0.1141 0.05340.0290 ⎤ ⎡ −0.3302 −0.0587 ⎤
⎢ 0.1130 0.4128 0.02570.0457 ⎥ ⎢ −0.0583 −0.2699 ⎥
AF = ⎢ −0.0332 −0.0167 −0.0011 ⎥ , BF = ⎢ 0.0066 0.0035 ⎥ ,
−0.9699
⎢ −0.0195 −0.0267 −0.9696 ⎥⎦
−0.0011 ⎢ 0.0042 0.0048 ⎥
⎣ ⎣ ⎦
[ ]
CF = −0.0076 −0.0195 −0.0017 −0.0027 , DF = [0.0018 0.0062] .
According to Theorem 4, the disk is chosen as Dr (0.5, 0.5), and then, we have 𝛾 𝜃 = 1.0230,
[ ] [ ] [ ]
0.0808 0.0050 67.1021 −48.2302 −5.4853 0.3999
U = 0.0050 0.0875 , V = −48.2302 64.1930 , M = 5.8360 0.3758 ,
[ ] [ ] [ ]
0.7106 −0.7036 0.3797 −0.5609 2.1985 −1.9160
N = −0.7036 −0.7106 , Ak = −0.0067 0.2896 , Bk = −0.1453 −0.0666 ,
[ ] [ ]
0.0035 −0.0661 −0.1814 −0.0449
Ck = −0.0003 −0.0508 , Dk = −0.0271 −0.1873 .

The residual signal r(k) and FD result are respectively shown in Figures 4 and 5.
As shown in Figure 5, the residual evaluation function value is smaller than the threshold Jth when there is no fault
and surpasses Jth after the fault f(k) occurs at k = 20, whereas the residual evaluation function value is smaller than Jth
again when f(k) disappears at k = 80, that is, the fault can be detected as long as it occurs.
The simulation results of FTC are shown in Figures 6 and 7. It can be seen from Figure 6 that, for the case of sys-
tem (45) without FTC, f(k) causes the deviations of the system output trajectories after f(k) occurs at k = 20. Compared
GUO ET AL. 13

0.05

0.045

0.04

0.035

0.03

0.025

0.02

0.015

0.01

0.005
residual signal
0
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 4 The residual signal r(k) [Colour figure can be viewed at wileyonlinelibrary.com]

0.05

0.045

0.04

0.035

0.03
Amplitude

0.025

0.02

0.015

0.01

0.005 threshold
residual evaluation function
0
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 5 The residual evaluation function and threshold Jth under SNR = 35 dB [Colour figure can be viewed at wileyonlinelibrary.com]

with Figure 6, the system output performance in Figure 7 can be guaranteed when the proposed FTC strategy is used
in system (45), ie, the designed fault-tolerant controller can not only compensate for the effects of f(k) but also keep the
post-fault system stable.
Example 2. Consider the following continuous-system with the coefficient matrices as in the works of Zhong et al43
and Li et al44 :

⎡ 0 1 0 0 ⎤ ⎡ 0 ⎤ ⎡0 0 0 ⎤ ⎡ 0 ⎤
⎢ 6.52 24.77 33.82 35.56 ⎥ ⎢ 0.2471 ⎥ ⎢0 0 −1 ⎥ ⎢ 0.5 ⎥
̇ =⎢
x(t) 0 0 0 1 ⎥ x(t) + ⎢ 0 ⎥ u(t) + ⎢ 0 0 0 ⎥ d(t) + ⎢ 0 ⎥ 𝑓 (t)
⎢ −12.56 −49.71 −56.25 −68.50 ⎥ ⎢ −0.4758 ⎥ ⎢0 0 0.5 ⎥⎦ ⎢ −0.5 ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎣ ⎦
[ ] [ ]
1000 100
𝑦(t) = 0 1 0 0 x(t) + 0 1 0 d(t). (46)

The initial state is x(0) = [0.5; 0.5; 0.5; 0.5], and a discrete-time controller is

Kc = [0.2665 0.4429 − 0.8298 0.1423] .

14 GUO ET AL.

0.8

0.7

0.6

0.5

Amplitude
0.4

0.3

0.2

y1
0.1
y2
0
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 6 Output responses without fault-tolerant control [Colour figure can be viewed at wileyonlinelibrary.com]

0.6
y1
0.5 y2

0.4

0.3
Amplitude

0.2

0.1

−0.1

−0.2

−0.3
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 7 Output responses with fault-tolerant control [Colour figure can be viewed at wileyonlinelibrary.com]

The discrete system matrices can be obtained as

⎡ 1.0106 0.1356 0.0780 0.0609 ⎤ ⎡ 0.0004 ⎤ [ ]

⎢ 0.1194 1.3429 1.1612 0.7294 ⎥ ⎢ 0.0045 ⎥ 1000
A = ⎢ −0.0216 −0.0826 −0.8870 −0.0223 ⎥ , B = ⎢ −0.0008 ⎥ , C = 0 1 0 0 ,
⎢ −0.2594 −0.9626 −1.5434 −0.5276 ⎥⎦ ⎢ −0.0098 ⎥
⎣ ⎣ ⎦
⎡0 0 −0.0053 ⎤ ⎡ 0.0020 ⎤ [ ]
⎢0 0 −0.1051 ⎥ ⎢ 0.0374 ⎥ 100
Ed = ⎢ 0 0 ,
0.0035 ⎥ 𝑓 ⎢ −0.0017 ⎥ d
E = , F = 010 .
⎢0 0 0.0715 ⎦⎥ ⎢ ⎥
⎣ ⎣ −0.0302 ⎦

The fault is assumed as

{
0 0 ≤ k < 40
𝑓 (k) =
1 − e−0.5(k−40) 40 ≤ k < 100,

and then, the fault f(k) and the weighted fault 𝑓̂w (k) are shown in Figure 8, in which the fault estimation with
satisfactory accuracy is given.
GUO ET AL. 15

1
the fault signal
0.9 the weighted fault signal

0.8

0.7

0.6

Amplitude
0.5

0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 8 The fault signal f(k) and the weighted fault signal 𝑓̂w (k) [Colour figure can be viewed at wileyonlinelibrary.com]

Then, we have 𝛾 ru = 0.0002, 𝛾 rd = 0.0084, 𝛾 rn = 0.0078, 𝛾 1u = 0.0378, 𝛾 2u = 0.0138, 𝛾 1d = 1.2089, 𝛾 2d = 1.1180. Based
on the proposed FD approach, for a given 𝛾 = 0.7826, the FD filter can be obtained

⎡ 0.9213 0.0799 −0.0324 −0.0651 −0.2307 −0.0053 −0.0074 ⎤

⎢ −0.0249 0.7634 0.2442 0.1878 0.0008 −0.1762 −0.3367 ⎥
⎢ −0.0541 −0.3271 0.6441 −0.1520 0.0133 0.0635 0.1698 ⎥
AF = ⎢ −0.1104 −0.3429 −0.6526 0.0109 0.0318 0.0934 0.1139 ⎥
⎢ 0.0195 0.0041 0.0013 −0.0026 −0.8804 0.0004 0 ⎥
⎢ 0.0035 0.0375 0.0094 −0.0159 0.0007 −0.8807 −0.0034 ⎥⎥
⎢
⎣ 0.0249 0.2334 0.0917 −0.0821 −0.0028 −0.0206 −1.0153 ⎦
⎡ −0.2890 −0.0022 ⎤
⎢ 0.0199 −0.1815 ⎥
⎢ 0.0161 0.1221 ⎥
BF = ⎢ 0.0253 0.0404 ⎥
⎢ 0.0690 0.0003 ⎥
⎢ 0.0004 0.0690 ⎥
⎢ ⎥
⎣ −0.0010 −0.0169 ⎦
[ ]
CF = −0.0018 −0.0189 −0.0123 −0.0032 0.0007 0.0038 0.0099
[ ]
DF = 0.0003 0.0038 .

The disk in Theorem 4 is chosen as Dr (0.5, 0.5), and then, we have 𝛾 𝜃 = 1.2224

⎡ 0.0137 −0.0073 −0.0059 0.0075 ⎤

⎢ −0.0073 0.0858 −0.0055 −0.0936 ⎥
U = ⎢ −0.0059 −0.0055 0.0199 −0.0411 ⎥
⎢ 0.0075 −0.0936 −0.0411 0.2610 ⎥⎦
⎣
⎡ 779.5406 722.9633 1666.0379 897.7013 ⎤
⎢ 722.9633 2742.2430 2793.5371 3592.9749 ⎥
V = ⎢ 1666.0379 2793.5371 4700.7867 3602.6659 ⎥
⎢ 897.7013 3592.9749 3602.6659 4776.6477 ⎥⎦
⎣
⎡ −2.0669 1.6867 0.4942 −0.0252 ⎤
⎢ 247.5205 −27.8189 −0.1145 −0.0219 ⎥
M = ⎢ 202.9838 16.8561 −0.2526 −0.0358 ⎥
⎢ −1084.2488 −3.1983 −0.0744 −0.0117 ⎥⎦
⎣
⎡ 0.0973 −0.4556 −0.8627 0.1967 ⎤
⎢ 0.5264 0.2749 0.0963 0.7988 ⎥
N = ⎢ 0.4617 −0.7669 0.4360 −0.0929 ⎥
⎢ 0.7073 0.3587 −0.2375 −0.5609 ⎥⎦
⎣
16 GUO ET AL.

⎡ 0.3382 4.5653 −14.7920 −56.5804 ⎤ ⎡ −16.0172 100.3029 ⎤

⎢ 0.0213 0.6075 1.8525 6.3421 ⎥ ⎢ 2.4234 −9.8217 ⎥
Ak = ⎢ −0.0004 −0.0001 0.8496 −0.3414 ⎥ , Bk = ⎢ 0.3853 0.2993 ⎥
⎢ −0.0001 −0.0038 −0.0047 0.4019 ⎥ ⎢ −0.0686 −0.0946 ⎥
⎣ ⎦ ⎣ ⎦
[ ]
Ck = 0.0071 0.1039 −0.2967 −1.1136 , Dk = [−0.2418 2.1254] .

Figures 9 and 10 show the generated residual signal r(k) and FD result, respectively.
From Figure 10, it can be seen that the residual evaluation function value is always less than the threshold in the
absence of fault and exceeds the threshold immediately after the fault occurs, ie, the fault can be detected effectively by
the proposed FD approach.
The simulation results of FTC are shown in Figures 11 and 12. As shown in Figure 11, the deviations of the system
outputs appear after the fault f(k) occurs at k = 40 for system (46) without FTC. For comparison, the deviations of the
system outputs can be revised in Figure 12 when the proposed FTC scheme is utilized in system (46), which demonstrates
the availability of the proposed FTC method.

−4
x 10
6
residual signal
4

0
Amplitude

−2

−4

−6

−8

−10

−12
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 9 The residual signal r(k) [Colour figure can be viewed at wileyonlinelibrary.com]

−3
x 10
1
threshold
0.9 residual evaluation function

0.8

0.7

0.6
Amplitude

0.5

0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 10 Residual evaluation function and threshold Jth under SNR = 35 dB [Colour figure can be viewed at wileyonlinelibrary.com]
GUO ET AL. 17

2
y1
y2
1.5

Amplitude
0.5

−0.5

−1
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 11 Output responses without fault-tolerant control [Colour figure can be viewed at wileyonlinelibrary.com]
2
y1
y2
1.5

1
Amplitude

0.5

−0.5

−1
0 10 20 30 40 50 60 70 80 90 100
Time step k

FIGURE 12 Output responses with fault-tolerant control [Colour figure can be viewed at wileyonlinelibrary.com]

6 CO N C LU S I O N

This paper has investigated the FD and FTC for the discrete-time systems over a MIMO AWGN channel subject to SNR
constraint. In the FD stage, an augmented robust FD filter has been designed by means of a descriptor system method, and
the estimations of the system state and fault have been obtained; then, a dynamic threshold has been constructed based
on stochastic characteristics of the channel noise, a constant 𝛼 has been used to reduce the conservatism of threshold, and
FAR of the threshold has been given. In the FTC stage, on the basis of the fault estimation information, an active output
feedback fault-tolerant controller has been designed to compensate for the effects of the fault. Two simulation examples
have been used to demonstrate the effectiveness of the proposed method.

ACKNOWLEDGEMENT
This work was supported by the National Natural Science Foundation of China under grants 61433003, 61273150,
61321002, and 61573024.

ORCID

Fumin Guo http://orcid.org/0000-0002-5644-6138

18 GUO ET AL.

REFERENCES
1. Punčochar I, Široký J, Šimandl M. Constrained active fault detection and control. IEEE Trans Autom Control. 2015;60(1):253-258.
2. Thabet REH, Combastel C, Raïssi T, Zolghadri A. Set-membership fault detection under noisy environment with application to the
detection of abnormal aircraft control surface positions. Int J Control. 2015;88(9):1878-1894.
3. Ulatowski A, Bazzi AM. A combinational-logic method for electric vehicle drivetrain fault diagnosis. IEEE Trans Ind Appl.
2016;52(2):1796-1807.
4. Costamagna P, Giorgi DA, Magistri L, Moser G, Pellaco L, Trucco A. A classification approach for model-based fault diagnosis in power
generation systems based on solid oxide fuel cells. IEEE Trans Energy Convers. 2016;31(2):676-687.
5. Wu L, Ho DWC. Fuzzy filter design for Ito stochastic systems with application to sensor fault detection. IEEE Trans Fuzzy Syst.
2009;17(1):233-242.
6. Rios H, Davila J, Fridman L, Edwards C. Fault detection and isolation for nonlinear systems via high-order-sliding-mode
multiple-observer. Int J Robust Nonlinear Control. 2015;25(16):2871-2893.
7. Zhai S, Wang W, Ye H. Fault diagnosis based on parameter estimation in closed-loop systems. IET Control Theory Appl.
2015;9(7):1146-1153.
8. Zhang D, Zhang WA, Yu L, Wang QG. Distributed fault detection for a class of large-scale systems with multiple incomplete measurements.
J Franklin Inst. 2015;352(9):3730-3749.
9. You J, Yin S, Gao HJ. Fault detection for discrete systems with network-induced nonlinearities. IEEE Trans Ind Inform.
2014;10(4):2216-2223.
10. Li XJ, Yang GH. Fault detection and isolation for uncertain closed-loop systems based on adaptive and switching approaches. Int J Robust
Nonlinear Control. 2016;26(13):2916-2937.
11. Jiang B, Gao Z, Shi P, Xu Y. Adaptive fault-tolerant tracking control of near space vehicle using Takagi-Sugeno fuzzy models. IEEE Trans
Fuzzy Syst. 2010;18(5):1000-1007.
12. Tabatabaeipour SM, Stoustrup J, Bak T. Fault-tolerant control of discrete-time LPV systems using virtual actuators and sensors. Int J Robust
Nonlinear Control. 2015;25(5):707-734.
13. Tao Y, Shen D, Fang M, Wang Y. Reliable H∞ control of discrete-time systems against random intermittent faults. Int J Syst Sci.
2016;47(10):2290-2301.
14. Huo Z, Zhang Z, Fang H. Research on fault tolerant control of networked control systems based on information scheduling. J Syst Eng
Electron. 2008;19(5):1024-1028.
15. Yu X, Jiang J. A survey of fault-tolerant controllers based on safety-related issues. Annu Rev Control. 2015;39(1):46-57.
16. Zhang Y, Jiang J. Bibliographical review on reconfigurable fault-tolerant control systems. Annu Rev Control. 2008;32(2):229-252.
17. Aranguiz E, Barbosa KA, Verdejo H. D-stability analysis of descriptor systems and application to power systems. IEEE Lat Am Trans.
2015;13(6):1888-1892.
18. Cantó B, Coll C, Sánchez E. Positive solutions of a discrete-time descriptor system. Int J Syst Sci. 2008;39(1):81-88.
19. Wang J, Wang W, Yang FW, Zhang JX. Fault detection for descriptor networked systems with multiple packet dropouts.
Paper presented at: Joint 48th IEEE Conference on Decision and Control and 28th Chinese Control Conference; 2009;
Shanghai, China.
20. Zhou L, Lu G. Detection and stabilization for discrete-time descriptor systems via a limited capacity communication channel. Automatica.
2009;45(10):2272-2277.
21. Du Z, Yue D, Hu S. H-infinity stabilization for singular networked cascade control systems with state delay and disturbance. IEEE Trans
Ind Inform. 2014;10(2):882-894.
22. Qiu Z, Zhang Q, Zhao Z. Stability of singular networked control systems with control constraint. J Syst Eng Electron.
2007;18(2):290-296.
23. Guo F, Ren X, Li Z, Han C. Robust fault detection for uncertain discrete-time systems subject to signal-to-noise ratio constrained channels.
J Franklin Inst. 2015;352(10):3971-3987.
24. Vargas FJ, Chen J, Silva EI. Stabilization of two-input two-output systems over SNR-constrained channels. Automatica.
2013;49(10):3133-3140.
25. Braslavsky JH, Middleton RH, Freudenberg JS. Feedback stabilization over signal-to-noise ratio constrained channels. IEEE Trans Autom
Control. 2007;52(8):1391-1403.
26. Jiang XW, Hu B, Guan ZH, Zhang XH, Yu L. The minimal signal-to-noise ratio required for stability of control systems over a noisy channel
in the presence of packet dropouts. Inf Sci. 2016;372:579-590.
27. Rojas AJ. Step reference tracking in signal-to-noise ratio constrained feedback control. Int J Control Autom Syst. 2015;13(5):1131-1139.
28. Rojas AJ, Freudenberg JS. Stabilisation over signal-to-noise ratio constrained channels: robust analysis for the discrete-time case. Paper
presented at: 2015 IEEE Conference on Control Applications; 2015; Sydney, Australia.
29. Freudenberg JS, Middleton RH, Solo V. Stabilization and disturbance attenuation over a Gaussian communication channel. IEEE Trans
Autom Control. 2010;55(3):795-799.
30. Rojas AJ. Signal-to-noise ratio performance limitations for input disturbance rejection in output feedback control. Syst Control Let.
2009;58(5):353-358.
31. Vargas FJ, Chen J, Silva EI. Necessary and sufficient conditions for mean square stabilization over MIMO SNR-constrained channels.
Paper presented at: 51st IEEE Conference on Decision and Control; 2012; Maui, HI.
GUO ET AL. 19

32. Vargas FJ, Silva EI, Chen J. Stabilization of TITO systems over parallel SNR-constrained AWN channels. Paper presented at: Proceeding
of the 7th IFAC Symposium over Robust Control Design; 2012; Aalborg, Denmark.
33. Guo F, Ren X, Li Z, Han C. Fault detection for discrete-time Lipschitz nonlinear systems with signal-to-noise ratio constrained channels.
Neurocomputing. 2016;194:317-325.
34. Vargas FJ, Silva EI, Chen J. Explicit conditions for stabilization over noisy channels subject to SNR constraints. Paper presented at: 9th
Asian Control Conference; 2013; Istanbul,Turkey.
35. Vargas FJ, Chen J, Silva EI. On stabilizability of MIMO systems over parallel noisy channels. Paper presented at: 53rd IEEE Conference
on Decision and Control; 2014; Los Angeles, CA.
36. Wang D, Wang W, Shi P. Robust fault detection for switched linear systems with state delays. IEEE Trans Syst Man Cybern B Cybern.
2009;39(3):800-805.
37. Dai L. Singular Control Systems. Lectures Notes in Control and Information Sciences, vol 118. New York, NY: Springer; 1989.
38. Hu S, Yue D, Liu J. H∞ filtering for networked systems with partly known distribution transmission delays. Inf Sci. 2012;194:270-282.
39. Ding SX. Model-Based Fault Diagnosis Techniques: Design Schemes, Algorithms, and Tools. Berlin, Germany: Springer; 2008.
40. Jiang B, Staroswiecki M, Cocqueempot V. Fault accommodation for nonlinear dynamic systems. IEEE Trans Autom Control.
2006;51(9):578-1583.
41. Garcia G, Bernussou J. Pole assignment for uncertain systems in a specified disk for dynamical feedback. IEEE Trans Autom Control.
1995;40(1):184-190.
42. Wang YQ, Ye H, Wang GZ. Fault detection of NCS based on eigendecomposition, adaptive evaluation and adaptive threshold. Int J Control.
2007;80(12):1903-1911.
43. Zhong M, Ding SX, Lam J, Wang H. An LMI approach to design robust fault detection filter for uncertain LTI systems. Automatica.
2003;39(3):543-550.
44. Li T, Wu L, Wei X. Robust fault detection filter design for uncertain LTI systems based on new bounded real lemma. Int J Control, Autom
Syst. 2009;7(4):644-650.
45. He X, Wang Z, Ji YD, Zhou DH. Networked-based fault detection for discrete-time state-delay systems: a new measurement model. Int J
Adapt Control Signal Process. 2008;22(5):510-528.

How to cite this article: Guo F, Ren X, Li Z, Han C. Fault detection and fault-tolerant control over signal-to-noise
ratio constrained channels. Int J Robust Nonlinear Control. 2018;1–22. https://doi.org/10.1002/rnc.4050

APPENDIX A : PROOFS

A.1 Proof of Theorem 1

First, we will prove that system (6) is admissible when v(k) = 0. In this case, the first equation of system (6) is rewritten as
Ee e(k + 1) = Ae e(k).
Then, there exist nonsingular matrices H and K satisfying
[ ] [ ]
I 0 Ae1 Ae2
HEe K = 0 0 , HAe K = ,
Ae3 Ae4
and let [ ]
P P2
H −T PH −1 = PT1 P3 .
2
Considering P > 0 and according to (8) and (9), we have
T T
Ae Ae − Ee PEe < 0. (A1)
Premultiplying and postmultiplying inequality (A1) by KT and K, respectively, and then, we obtain
[ ]
ΛΛ
Λ Γ < 0,
where Λ is the submatrix that is irrelevant to the discussion and Γ is
T T T T
Γ = Ae2 P1 Ae2 + Ae2 P2 Ae4 + Ae4 P2T Ae2 + Ae4 P3 Ae4 < 0. (A2)
20 GUO ET AL.

Based on (A2), we have Ae4 that is nonsingular. Otherwise, we can assume Ae4 is singular. Then, there exists a vector
T
𝜑 ≠ 0 such that Ae4 𝜑 = 0. Premultiplying and postmultiplying inequality (A2) by 𝜑T and 𝜑, we obtain 𝜑T Ae2 P1 Ae2 𝜑 < 0.
T
However, 𝜑T Ae2 P1 Ae2 𝜑 > 0 due to P1 > 0. Thus, the assumption is not satisfied, that is, Ae4 is nonsingular. Then, system (6)
is regular and causal according to Definition 1.
Constructing a Lyapunov function as
T T
V( e(k)) = e (k)Ee PEe e(k) (A3)
and then calculating its difference, we have
ΔV(e(k)) = V( e(k + 1)) − V(e(k))
T T T T
= e (k + 1)Ee PEe e(k + 1) − e (k)Ee PEe e(k) (A4)
( T T
)
T
= e (k) Ae PAe − Ee PEe e(k).
T T
Using the Schur complement, it follows from inequalities (8) and (9) that Ae PAe − Ee PEe < 0. Then, there exists
T
a constant 𝜆 > 0 such that ΔV(e(k))≤ −𝜆e (k)e(k). Thus, we have E{ΔV( e(k))} ≤ −𝜆E{̄eT (k)̄e(k)}. Summing up the
∑∞ T
aforementioned inequality from k = 0 to k = ∞, we can obtain E{ k=0 e (k)e(k)} ≤ 𝜆1 E{V( e0 )} < ∞, where e0 is any
initial condition. This shows that system (6) is stochastically stable. Thus, based on the aforementioned discussion, we
can see that system (6) is admissible under the circumstance of v(k) = 0 .
Next, define the following performance index when v(k) ≠ 0:
{N−1 }
∑( )
J=E reT (k)re (k) − 𝛾 2 vT (k)v(k) , (A5)
k=0

where N is an arbitrary positive integer. Then, under zero initial condition, we have
{N−1 }
∑{ }
J=E re (k)re (k) − 𝛾 v (k)v(k) + ΔV( e(k))
T 2 T
− E{V( e(N))}
k=0
{N−1 }
∑{ }
≤E re (k)re (k) − 𝛾 v (k)v(k) + ΔV( e(k))
T 2 T

k=0
{N−1 }
∑{ T T T T
}
=E reT (k)re (k) − 𝛾 2 vT (k)v(k) + e (k + 1)Ee PEe e(k + 1) − e (k)Ee PEe e(k) (A6)
k=0
{N−1
∑{
=E [ Ce e(k) + De v(k)]T × [ Ce e(k) + De v(k)] − 𝛾 2 vT (k)v(k)
k=0
T
}}
T
+[ Ae e(k) + Be v(k)]T P[ Ae e(k) + Be v(k)] − e (k)Ee PEe e(k)
{N−1 }
∑
=E Π ΥΠ ,
T

k=0

where
[ T T
]
T Ω Ae PBe + Ce De
Π = [e (k) v (k)] , Υ =
T T
T T ,
∗ Be PBe + De De − 𝛾 2 I
T T T
Ω = Ae PAe + Ce Ce − Ee PEe .
Introducing the positive definite matrix Q, and then, Υ can be rewritten as
[ T T
]
Ω+Q−Q Ae PBe + Ce De
Υ= T T . (A7)
∗ Be PBe + De De − 𝛾 2 I
Then, Υ is equivalent to [ ] [ ]
T [ ] T T
Ae −P + Q + Ce Ce C e De
Υ= T P Ae Be + T . (A8)
Be ∗ De De − 𝛾 2 I
GUO ET AL. 21

It follows from (9) and the Schur complement that Υ < 0, which implies J < 0. Thus, the H∞ performance ||re (k)||E,2 <
𝛾||v(k)||E,2 holds for any nonzero v(k) ∈ l2 [0, ∞).
The proof is completed.

A.2 Proof of Theorem 2

According to Theorem 1, system (6) with v(k) = 0 is admissible, and if inequality (9) holds under zero initial condition,
then the H∞ performance ||re (k)||E,2 < 𝛾||v(k)||E,2 can be guaranteed for all the nonzero v(k) ∈ l2 [0, ∞).
Based on inequality (14), we have
[ ]
X + X T X + Y T + GT 0
∗ Y +Y T
0 > 0, (A9)
∗ ∗ Z + ZT

where matrices X, Y, and Z are nonsingular. Premultiplying [I, −I, 0] and postmultiplying [I, −I, 0]T to (A9), we obtain
−G − GT > 0, which implies that G is a nonsingular matrix.
Let
[ ] [ ]
I I 0 X 0 0 T
L = G 0 0 , Φ = Y I 0 L−1 .
0 0 I 0 0Z
By taking a congruence transformation to inequality (13) via diag{L, I, L, L, I} and setting
[ ]
⎡X M 0 ⎤ P11 P12 P13
L (Φ + Φ)L = ⎢ ∗ Y 0 ⎥ , LT PL =
T T
∗ P22 P23 ,
⎢ ⎥ ∗ ∗ P33
⎣∗ ∗ Z⎦
[ ] [ ]
Q11 Q12 Q13 T
Δ11 Δ12 Δ13
LT QL = ∗ Q22 Q23 , LT Ee PEe L = ∗ Δ22 Δ23 ,
∗ ∗ Q33 ∗ ∗ Δ33

A𝑓 = AF G, B𝑓 = BF , C𝑓 = CF G, D𝑓 = DF ,

then (14) is obtained.

The proof is completed.
Remark 7. According to the work of Wang et al,36 the proposed approach can reduce the conservatism reported in the
work of He et al,45 that is, the symmetric Lyapunov matrix is introduced, then by using the matrices L and ΦL, the
problem of the block-diagonal Lyapunov matrices with constraints is solved.

A.3 Proof of Theorem 3

According to (15), a false alarm occurs if and only if the residual evaluation function value is larger than the threshold in
the absence of the fault. Thus, based on the notable Chebyshev's inequality, the FAR of threshold (21) can be obtained by
{ }
Pb ||re𝑓 (k)||2,𝜌 ≥ Jth |𝑓 = 0
{ ( )}
≤ Pb ||ru (k)||2,𝜌 + ‖rd (k)‖2,𝜌 + ‖rn (k)‖2,𝜌 ≥ 𝛼 Jth,u + Jth,d + Jth,n
{ }
≤ Pb ||rn (k)||2,𝜌 ≥ 𝛼 · Jth,n
{ {√ { } √ { }}}
≤ Pb 𝛾rn ||n(k)||2,𝜌 ≥ 𝛾rn sup E ||n(k)||22,𝜌 + 𝛼 𝜇 · sup 𝜎 ||n(k)||22,𝜌 (A10)
n n
{ { } { }}
≤ Pb ||n(k)||22,𝜌 ≥ E ‖n(k)‖22,𝜌 + 𝛼 2 · 𝜇 · 𝜎 ||n(k)||22,𝜌

( )2
≤ 1∕ 𝛼 2 𝜇 .
22 GUO ET AL.

A.4 Proof of Theorem 4

⌣
According to Theorem 1, if there exists the positive definite matrix P such that
⎡ −⌣ ⌣⌣ ⌣⌣
0 ⎤⎥
⎢ P P A P D1
⌣ ⌣T
⎢ ∗ − P 0 C ⎥⎥ < 0
⎢ ⌣T
(A11)
⎢ ∗ ∗ −𝛾𝜃 I D2 ⎥⎥
⎢
⎣ ∗ ∗ ∗ −𝛾𝜃 I ⎦
holds, then system (39) is robust stable and we can obtain ||w(k)||E,2 < 𝛾 𝜃 ||𝜃(k)||E,2 .
Let
⌣
[ ] [ ]
V N ⌣−1 U M
P= , P = ,
NT W MT Ξ
⌣⌣−1
where U, V ∈ Rn×n are the positive definite symmetric matrices. Due to P P = I, we have
⌣
[ ] [ ]
U I I V
P = . (A12)
MT 0 0 NT

Define [ ] [ ]
U I I V
F1 = , F2 = ,
MT 0 0 NT
⌣
we obtain PF1 = F2 . Premultiplying diag{F1T , F1T , I, I} and postmultiplying diag(F1 , F1 , I, I ) to inequality (A11) and
defining

A = V(A + BDk C)U + NBk CU + VBCk M T + NAk M T ,

B = VBDk + NBk , C = Dk CU + Ck M T , D = Dk ,
then using the Schur complement, we have (41).
⌣
Based on Lemma 1, for a given Dr (q, 𝜏), the eigenvalues of A belong to Dr (q, 𝜏) if and only if there exists the positive
⌣
definite symmetric matrix P ∈ Rn×n such that
[ ⌣ ⌣⌣ ⌣]
−P P A − qP
⌣ <0 (A13)
∗ −𝜏 2 P
holds, then premultiplying diag(F1T , F1T ) and postmultiplying diag(F1 , F1 ) to (A13), and according to the expressions of
A, B, C, D, we have (40).
The proof is completed.
⌣⌣−1
Remark 8. Since P P = I, we obtain MNT = I − UV, and full rank matrices M and N can be gained by the singular
value decomposition of I − UV.