Systems & Control Letters 82 (2015) 18–25

Contents lists available at ScienceDirect

Systems & Control Letters

journal homepage: www.elsevier.com/locate/sysconle

On observer-based fault detection for nonlinear systems

Ying Yang a , Steven X. Ding b,∗ , Linlin Li b
a
State Key Lab for Turbulence and Complex Systems, Department of Mechanics and Engineering Science, Peking University, Beijing 100871, PR China
b
Institute for Automatic Control and Complex Systems (AKS), Faculty of Engineering, University of Duisburg–Essen, 47057 Duisburg, Germany

article info abstract

Article history: This paper addresses analysis and integrated design of observer-based fault detection (FD) for nonlinear
Received 15 May 2014 systems. To gain a deeper insight into the observer-based FD framework, definitions and existence
Received in revised form conditions for nonlinear observer-based FD systems are studied first. Then, a scheme for an integrated
12 May 2015
design of observer-based FD systems for affine nonlinear systems is proposed. Our work is considerably
Accepted 12 May 2015
inspired by the study on input–output stability and stabilization of nonlinear systems. Examples are given
at the end of the paper to illustrate the theoretical results.
Keywords:
© 2015 Elsevier B.V. All rights reserved.
Fault detection
Nonlinear observer-based residual
generation
Input–output stability

1. Introduction studies have been mainly devoted to the application of feedback-

based linearization and differential algebra techniques to residual
Associated with the increasing industrial demands on safety generator design [13–15], and the geometric approach to nonlinear
and reliability, observer-based fault detection and isolation (FDI) FDI [16], the recent research efforts address systems with a special
has attracted intensive attention in the past two decades. One of class of nonlinearities, typically Lipschitz nonlinearity [17,18], sec-
the most challenging topics in the FDI research and application tor bounded nonlinearity [19] or special types of control systems
areas is nonlinear observer-based FDI. A review of the literatures like nonlinear networked control system [20]. Differently, [21–23]
in the last two decades shows that the application of nonlinear have investigated residual evaluation, threshold setting and resid-
observer theory built the main stream in the nonlinear observer- ual generator optimization issues for nonlinear observer-based FD
based FDI study in the 1990s [1]. In recent years, much attention systems. It can be observed that (i) only few of these studies have
has been paid to the application of those techniques to addressing dealt with residual generator and evaluation as well as decision
nonlinear FDI issues. For instance, Fuzzy technique based FDI [2,3], making in an integrated way and (ii) most of efforts have been
adaptive fault diagnosis [4,5], LPV-based FDI [6,7] or sliding mode made on the FD system design but only few on analysis issues.
observer-based fault detection [8,9] as well as their successful Motivated by these observations, in the first part of this paper
applications [10,11] have been reported. we study an essential analysis issue, the existence conditions
In this paper, we deal with analysis and design issues of nonlin- of nonlinear observer-based FD systems. With the aid of input–
ear observer-based fault detection (FD) systems. Roughly speak- output stability theory, the concepts output re-constructability
ing, an observer-based FD system consists of an observer-based and weak output re-constructability are introduced and, in this
residual generator, a residual evaluator and a decision maker with context, existence conditions for different types of nonlinear
a threshold. Reviewing the publications on nonlinear observer- observer-based FD systems are investigated. The objective of this
based FD studies reveals that the major research focus in this area work is to gain a deeper insight into the fundamental properties of
is on the design of observer-based residual generators [1,12]. Serv- nonlinear observer-based FD systems, which may be useful for the
ing as a major methodology, nonlinear observer theory is widely development of nonlinear FD systems using some well established
applied for the investigations in this thematic field. While the first technologies. In the second part of our work, we propose a scheme
for an integrated design of observer-based FD systems for a class
of nonlinear systems, the so-called affine systems. For the purpose
∗
Corresponding author. Tel.: +49 203 379 3386.
of attaining an efficient FD system, a dynamic threshold is first
E-mail addresses: [email protected] (Y. Yang), [email protected] proposed by taking into account of the influence of the input
(S.X. Ding), [email protected] (L. Li). variables on the residual signal with the help of L2 stability
http://dx.doi.org/10.1016/j.sysconle.2015.05.004
0167-6911/© 2015 Elsevier B.V. All rights reserved.
 Y. Yang et al. / Systems & Control Letters 82 (2015) 18–25 19

theory [24]. Moreover, different cases for the application of the (5) is also called fault detection filter (FDF), where L is the observer
observer-based FD system are discussed. gain to be designed such that the FDF is stable and residual signal
The paper is organized as follows. In Section 2, needed pre- r (t ) satisfies
liminaries are introduced and the problems to be addressed are ∀u(t ), x(0), lim r (t ) = 0. (6)
formulated. Section 3 is devoted to the study on definitions and t →∞

existence conditions of observer-based FD systems. The integrated A norm of r (t ), typically L2 - or L∞ -norm, is adopted as residual
design of observer-based FD systems for affine nonlinear systems evaluation function, which is defined by
is proposed in Section 4, and different cases are addressed. In Sec- J = ∥r (t )∥ or J = ∥r (t )∥∞ . (7)
tion 5, examples are provided to illustrate the theoretical results.
Let Jth = supx0 ,w=0 J be the threshold, which is interpreted as the
Notations: Standard notation is adopted in this paper. In addi-
maximum influence of x(0) = x0 on the fault-free (w(t ) = 0)
tion, following definitions known in nonlinear stability theory are
residual vector r (t ). A simple form of detection logic is
used. R+ = [0, ∞). A function γ : R+ → R+ is said to be-
J > Jth =⇒ faulty

long to class K if it is continuous, strictly increasing, and satisfies
γ (0) = 0. A function β : R+ → R+ is said to belong to class L if detection logic: (8)
J ≤ Jth =⇒ fault-free.
it is continuous, strictly decreasing, and satisfies lims→∞ β(s) = 0.
In this paper, two essential nonlinear FD issues will be
A function φ(s, t ) : R+ × R+ → R+ is said to belong to class
addressed. The first one deals with definitions and the existence
KL if for each fixed t the function is of class K and for each fixed
conditions of observer-based FD systems for nonlinear systems (1).
s it is of class L. ∥ · ∥ stands for the Euclidean norm of a vector in
We will apply the well-established input–output stability theory
some Euclidean space; Br := {x ∈ Rn : ∥x∥ ≤ r for some r > 0}.
and the concept of weak detectability for nonlinear systems [26]
L2 (0, ∞) is the space P
 ∞ of functions u : R+ → R which are mea- to our study. It is worth to emphasize that our work is devoted to
surable and satisfy 0 ∥u(t )∥2 dt < ∞. L2,[0,τ ] -norm of u(t ) is de-
 τ 1/2 the overall observer-based FD system with a residual generator,
fined and denoted by ∥uτ ∥2 = 0
∥u(t )∥2 dt , and ∥u∥∞ = an evaluator and a decision maker. The second issue is the design
ess sup {∥u(t )∥, t ≥ 0}. A function f : Rn → R is positive definite of nonlinear observer-based FD systems for affine systems (2). In
if f (x) > 0 for all x > 0, and f (0) = 0. By Vx,x̂ (x, x̂) we denote this work, the L2 -gain technique will serve as the investigation
tool [24]. We will propose a design scheme and apply it to different
Vx,x̂ (x, x̂) = Vx (x, x̂) Vx̂ (x, x̂)
 
design cases.
∂ V (x, x̂) ∂ V (x, x̂)
 
= . 3. On observer-based FD systems
∂x ∂ x̂
In this section, we introduce some basic definitions needed for
constructing a nonlinear observer-based FD system and study the
2. Preliminaries and problem formulation existence conditions.

Consider nonlinear systems described by 3.1. On the construction of observer-based FD systems

Σ : ẋ = f (x, u), y = h(x, u) (1) For our purpose, we first introduce the configuration of the
where x ∈ Rn , u ∈ Rp , y ∈ Rm denote the state, output and observer-based FD systems to be addressed in this paper, which
input vectors, respectively. f (x, u) and h(x, u) are continuously consists of an observer-based residual generator, an evaluation
function and a decision maker.
differentiable nonlinear functions with appropriate dimensions.
The affine form of Σ , Definition 1. Given the nonlinear system (1), a system of the form
Σ : ẋ = a(x) + B(x)u, y = c (x) + D(x)u (2)
x̂˙ = φ(x̂, u, y) (9)
with a(x), B(x), c (x) and D(x) being continuously differentiable r = ϕ(x̂, u, y) (10)
and of appropriate dimensions, is a class of nonlinear systems
which are widely adopted in nonlinear system research. This class is called observer-based residual generator if it delivers a residual
of nonlinear systems can be considered as a natural extension of vector r satisfying that (i) for x̂(0) = x(0), ∀u, r (t ) ≡ 0 (ii) for
linear time-invariant (LTI) systems some w ̸= 0 in the faulty system (4), r (t ) ̸= 0.
In order to avoid loss of information about the faults, the
ẋ = Ax + Bu, y = Cx + Du (3)
residual vector has generally the same dimension like the output
where A ∈ Rn×n , B ∈ Rn×p , C ∈ Rm×n , D ∈ Rm×p . In the FD study vector. For the sake of simplicity, also considering the conditions
on LTI systems, faults are often modelled by extending (3) to (i) and (ii), we suppose that

ẋ = Ax + Bu + Ew w, y = Cx + Du + Fw w r = ϕ(x̂, u, y) = y − ŷ, ŷ = h(x̂, u). (11)

For the residual evaluation purpose, two norm-based evaluation
where w ∈ Rkw represents the fault vector and Ew , Fw are known
functions are considered in this paper: (i) the Euclidean norm-
matrices of appropriate dimensions. Analog to it, an extension of
based instant evaluation
(1) to
JE = α1 (∥r ∥) (12)
Σ : ẋ = f (x, u, w), y = h(x, u, w) (4)
(ii) integral evaluation with an evaluation window [0, τ ]
is adopted for modelling nonlinear faulty systems.  τ
A standard observer-based FD system consists of an observer- J2 = α2 (∥r ∥) dt (13)
based residual generator, a residual evaluator and a decision maker 0

with a threshold [25]. For LTI systems, the most widely applied type where α1 (∥r ∥) , α2 (∥r ∥) are some K -functions. Moreover, the
of residual generators is a full-order observer described by detection logic (8) will be used for decision making. To this end,
threshold setting is a major part of constructing an observer-based
x̂˙ = Ax̂ + Bu + Lr , r = y − ŷ, ŷ = C x̂ + Du. (5) FD system.
20 Y. Yang et al. / Systems & Control Letters 82 (2015) 18–25

Definition 2. Given the nonlinear system (1), a dynamic system is An immediate result of Theorem 1 is the following corollary,
called which can be used for the design of an L∞ observer-based FD
system.
• L∞ observer-based FD system, when it consists of the observer-
based residual generator (9) and (11), residual evaluation func- Corollary 1. Assume that system (1) is output re-constructible,
tion (12) and detection logic (8) with a corresponding threshold. x(0) − x̂(0) ≤ δo and the evaluation window is [t1 , t2 ], i.e. JE =

• L2 observer-based FD system, when it consists of the observer- ϕ1 (∥r (t )∥) , t ∈ [t1 , t2 ]. Then, an L∞ observer-based FD system
based residual generator (9) and (11), residual evaluation func-
can be built by (i) constructing residual generator (9) using φ being
tion (13) and detection logic (8) with a corresponding threshold.
defined in Definition 3 and(ii) setting the threshold Jth = β (δo , t1 ).
In the subsequent two subsections, we are going to study the
The proof follows
 directly from
 Theorem 1 and the definition of
existence conditions of the above two types of FD systems as well
KL-function β x(0) − x̂(0) , t , since
as the construction of the corresponding thresholds.
β x(0) − x̂(0) , t
  
Jth = max
∥x(0)−x̂(0)∥≤δo
3.2. On L∞ observer-based FD systems t ∈[t1 ,t2 ]
 
For our purpose, we first introduce the following definition,
=β x(0) − x̂(0) , min t = β (δo , t1 ) .
 
max
which is motivated by the so-called weak detectability given in [26]. ∥x(0)−x̂(0)∥≤δo t ∈[t1 ,t2 ]
The concept weak detectability is widely used in the study on the
stabilization of nonlinear systems by output feedback [27,28]. Corollary 1 reveals that if system (1) is output re-constructible,
a constant threshold can be found. On the other hand, such a
Definition 3. System (1) is said to be output re-constructible if setting could be too conservative. In order to improve the FD
there exist (i) a function φ : Rn × Rp × Rm → Rn and (ii) functions performance, the influence of the process input variables on the
V (x, x̂) : Rn × Rn → R+ , ϕi (·) ∈ K , i = 1, 2, 3, and positive residual vector should be generally taken into account. It will lead
constants δ, δu such that ∀x, x̂ ∈ Bδ , ∥u∥∞ ≤ δu to a so-called adaptive threshold, which allows a more efficient FD
than a constant one [25]. Moreover, considering that the L2,[0,τ ] -
ϕ1 (∥r ∥) ≤ V (x, x̂) ≤ ϕ2 x − x̂ , r = y − h(x̂, u)
 
(14) norm is often used for the residual evaluation purpose, we are
motivated to investigate the following detection scheme.
Vx (x, x̂)f (x, u) + Vx̂ (x, x̂)φ(x̂, u, y) ≤ −ϕ3 x − x̂ .
 
(15)

3.3. On L2 observer-based FD systems

Remark 1. Substituting ∥r ∥ in ϕ1 (∥r ∥) by x − x̂, Definition 3
 
becomes equivalent with the weak detectability given in [26]. We first introduce the definition of weak output re-construct-
If it is further assumed, as in [29], that ∥h(ζ , u) − h(ς , u)∥ ≤ ability.
γ (∥ζ − ς ∥) for some γ ∈ K , then we have x − x̂ ≥ γ −1 (∥r ∥),
which leads to ϕ1 x − x̂ ≥ ϕ1 γ −1 (∥r ∥) . Since ϕ1 γ −1 (·) ∈ Definition 4. System (1) is said to be weakly output re-con-
     
K , the weak detectability implies the output re-constructability. structible if there exist (i) a function φ : Rn × Rp × Rm → Rn and
(ii) functions V (x, x̂) : Rn × Rn → R+ , ϕ1 (·) ∈ K , ϕ2 (·) ∈ K∞
Remark 2. Consider the (overall) system and a constant δ > 0 such that ∀x, x̂ ∈ Bδ

ẋ = f (x, u), y = h(x, u), x̂˙ = φ(x̂, u, y), Vx (x, x̂)f (x, u) + Vx̂ (x, x̂)φ(x̂, u, h(x, u))
r = y − h(x̂, u) ≤ −ϕ1 (∥r ∥) + ϕ2 (∥u∥). (18)

with u as its input and r as output. Function V (x, x̂) satisfying Comparing Definitions 3 and 4, it becomes evident that
(14)–(15) can be understood as a variant of the IOS-Lyapunov condition (18) is generally weaker than the conditions given in
function [30]. In fact, the residual generation problem can also Definition 3.
be studied in the IOS context. The motivation of introducing The following theorem presents the major result of this sub-
the output re-constructability is that in the model-based FDI section, a sufficient condition for the existence of an L2 observer-
framework, the output estimate ŷ is called analytical redundancy, based FD system.
and residual generation is equivalent with building analytical
redundancy. Theorem 2. Assume that system (1) is weakly output re-constructible.
The following theorem presents a major property of an output Then, an L2 observer-based FD system can be realized using functions
re-constructible system, which provides us with a sufficient ϕ1 , ϕ2 and by (i) constructing residual generator according
τ to (9) and
condition for the existence of an L∞ observer-based FD system and (11), (ii) defining the evaluation function as J = 0 ϕ1 (∥r (t )∥) dt
the threshold setting. and (iii) setting the threshold
 τ
Theorem 1. Assume that system (1) is output re-constructible. Then, Jth = ϕ2 (∥u∥) dt + γ̄o , γ̄o = sup {γ0 } ,
system (9) with (11) as its output delivers a residual vector r (t ), and 0 x(0),x̂(0) (19)
it holds
γo = V x(0), x̂(0) .
 
∥r (t )∥ ≤ β x(0) − x̂(0) , t
  
(16)
Proof. It follows from (18) that
where β x(0) − x̂(0) , t ∈ KL.
  
V̇ (x, x̂) ≤ −ϕ1 (∥r (t )∥) + ϕ2 (∥u∥) .
A similar proof can be found in [31,30] using IOS-Lyapunov
function, as pointed out in Remark 2. Thus, it is omitted here. As a result,
It is evident that for a given initial estimation error x(0) − x̂(0)  τ  τ
ϕ1 (∥r ∥) dt ≤ ϕ2 (∥u∥) dt + V (x(0), x̂(0)). (20)
lim β x(0) − x̂(0) , t = 0.
  
(17) 0 0
t →∞

Property (17) is similar to (6) in the LTI case. The theorem is thus proved. 
 Y. Yang et al. / Systems & Control Letters 82 (2015) 18–25 21

In theoretical study on norm-based residual evaluation, it is Definition 5. Given the nonlinear system (2), the NFDF (25)–(26)
the state of the art that the evaluation window is assumed to be is called L2 -NFDF if it satisfies that for some constant γu ≥ 0
infinitively large, i.e.
 ∞  ∞ ∥rτ ∥22 ≤ γu2 ∥uτ ∥22 + γo (27)
J = ϕ5 (∥r ∥) dt =⇒ Jth = ϕ4 (∥u∥) dt + γ̄o . (21)
0 0 where γo ≥ 0 is a (finite) constant for given x(0), x̂(0).

In practice, this is not realistic, since a large evaluation window It follows from (27) that by an L2 -NFDF the threshold can be,
generally results in a (considerably) delayed fault detection. In on the assumption that γo is bounded for all possible x(0), x̂(0), set
dealing with nonlinear FD, a large evaluation window also means equal to
a high threshold due to the dependence on u. Considering that a
fault may happen after the system is in operation for a long time, Jth = γu2 ∥uτ ∥22 + sup {γo } . (28)
for a large evaluation window the influence of w on J may be much x(0),x̂(0)
weaker than u on Jth . As a result, the FD performance can become
poor. For these reasons, in practice the evaluation function and It allows then the application of the following decision logic,
threshold are often defined by
J = ∥rτ ∥22 > Jth =⇒ faulty

to +τ to +τ (29)
J = ∥rτ ∥22 ≤ Jth =⇒ fault-free
 
J = ϕ5 (∥r ∥) dt =⇒ Jth = ϕ4 (∥u∥) dt + γ̄o (22)
to to
for a successful FD. Note that the L2 -NFDF, the residual evaluation
where γ̄o = supx(t0 ),x̂(t0 ) {γo } represents the maximum γo for all function (29) and threshold setting (28) build a special realization
(bounded) possible x(t0 ), x̂(t0 ). In Section 5, this aspect will be of the FD scheme presented in Section 3. In fact, we have the
illustrated by examples. following relations
In this section, we have derived the existence conditions for  τ
two types of nonlinear observer-based FD systems. Although J = ∥r τ ∥ =2
2 ∥r (t )∥2 dt =⇒ ϕ1 (∥r ∥) = ∥r ∥2
the achieved results do not lead to a direct design of a 0

nonlinear observer-based FD system, they are fundamental for Jth = γu2 ∥uτ ∥ + sup {γo } =⇒ ϕ2 (∥u∥) = (γu ∥u∥)2
2
2
x(0),x̂(0)
the application of some established nonlinear techniques for FD
system design. For instance, inspirited by the T–S-Fuzzy controller where ϕ1 , ϕ2 are defined in Theorem 2. Next, we study the design
design for nonlinear systems [32], we are investigating Fuzzy of L2 -NFDF, which means the determination of gain matrix L(x̂) so
observer-based FD system design based on the conditions given that (27) holds. Let
in Theorems 1–2. A further interesting application aspect is the
   
integrated design of control and diagnosis systems. As well-known, ẋ 0
= ¯ (x, x̂) + G(x, x̂)u +
f
weak detectability is an essential condition for the parametrization x̂˙ L(x̂)r
and design of nonlinear stabilization controllers [28]. It is of
r = y − ŷ, ŷ = c (x̂) + D(x̂)u
reasonable interest to apply the relationship between weak
a(x) B(x)
   
detectability and output re-constructability to the study on the f¯ (x, x̂) = , G(x, x̂) = .
integrated design of fault detection and control as well as fault- a(x̂) B(x̂)
tolerant control systems [25].

4. Design of L2 -FD systems for affine systems Theorem 3. Given the system (2) and the NFDF (25)–(26). Suppose
that (i) there exists a constant γ > 0 so that
In this section, we address FD system design issues. For our T 
γ 2 I − D(x) − D(x̂) D(x) − D(x̂) > 0
 
(30)
purpose, we consider the affine systems described by (2).
and set Θ (x, x̂) given by
4.1. A design method for L2 -NFDF T 
γ 2 I − D(x) − D(x̂) D(x) − D(x̂) = Θ T (x, x̂)Θ (x, x̂)
 
(31)
It follows from (16) that x̂(t ) = x(t ) if x̂(0) = x(0), which means
in turn that for all u with Θ (x, x̂) = Θ T (x, x̂) being a p × p matrix, (ii) there exists
V (x, x̂) ≥ 0 that solves Hamilton–Jacobi inequality (HJI)
φ(x, u, c (x) + D(x)u) = a(x) + B(x)u. (23)
1 T
Vx,x̂ (x, x̂)f¯ (x, x̂) + c (x)c (x) − c T (x̂)c (x̂)

As discussed in [33], for affine systems (2), (9) can then be written 2
into
1 −1 T
+ w(x, x̂) Θ T (x, x̂)Θ (x, x̂) w (x, x̂) ≤ 0

(32)
x̂˙ = φ(x̂, u, y) = a(x̂) + B(x̂)u + l x̂, u, y − c (x̂) − D(x̂)u
 
(24) 2
w(x, x̂) = Vx,x̂ (x, x̂)G(x, x̂) + c T (x) D(x) − D(x̂)
 
(33)
with l x̂, u, y − c (x̂) − D(x̂)u satisfying l (x, u, y − c (x) − D(x)u)
 
= 0. For the sake of simplicity, we restrict our study in the sequel and(iii) there exists L(x̂) solving
to
Vx̂ (x, x̂)L(x̂) = c T (x̂). (34)
x̂˙ = a(x̂) + B(x̂)u + L(x̂) y − c (x̂) − D(x̂)u
 
(25)
Then, it holds
r = ϕ(x̂, u, y) = y − c (x̂) − D(x̂)u. (26)

We call (25)–(26) NFDF, which is similarly like the LTI-FDF (5). ∥rτ ∥22 ≤ γ 2 ∥uτ ∥22 + 2V (x(0), x̂(0)). (35)
22 Y. Yang et al. / Systems & Control Letters 82 (2015) 18–25

Proof. Considering example of the previous study. It is helpful to gain a deeper insight
into the addressed nonlinear FD problems.
V̇ (x, x̂) = Vx,x̂ (x, x̂) f¯ (x, x̂) + G(x, x̂)u
 
Recall that for the L2 -stable system (2) it holds [24] for some
+ Vx̂ (x, x̂)L(x̂) y − c (x̂) − D(x̂)u γ̄ ≥ 0, γu > γ̄ , γ̄o > 0, γo > 0
 

and (34), it holds ∥yτ ∥2 ≤ γ̄ ∥uτ ∥2 + γ̄o ⇐⇒ ∥yτ ∥22 ≤ γu2 ∥uτ ∥22 + γo . (37)
V̇ (x, x̂) = Vx,x̂ (x, x̂) f¯ (x, x̂) + G(x, x̂)u
 
This means that the original system (2) itself can also serve as
T 
+ ŷ − D(x̂)u y − ŷ .
  an NFDF. Using the process input and output variables u, y in the
threshold setting (28) and detection logic (29), an FD system is then
Note that built. On the other hand, since γu2 can be (very) large, this leads to
1 1 1 a high threshold setting. Consequently, the fault detectability may
∥r ∥2 = yT y + ŷT ŷ − ŷT y become poor. In order to improve the FD performance, we now
2 2 2
apply a simple residual generator defined by
1 11
∥y∥2 = ∥c (x)∥2 + ∥D(x)u∥2 + c T (x)D(x)u
2 2 2 Σk : x̂˙ = a(x̂) + B(x̂)u, r = y − c (x̂) − D(x̂)u. (38)
1  2 1 2 1 
ŷ = c (x̂) + D(x̂)u2 + c T (x̂)D(x̂)u

T
Let γ 2 I − D(x) − D(x̂) D(x) − D(x̂) > 0 and set
   
2 2 2
and moreover T 
γ 2 I − D(x) − D(x̂) D(x) − D(x̂) = Θ T (x, x̂)Θ (x, x̂).
 
1
Θ (x, x̂)u − Θ −T (x, x̂)w T (x, x̂)2

If there exists V (x, x̂) ≥ 0 that solves the HJI
2
γ2 1  1
c (x) − c (x̂)2
Vx (x, x̂)a(x) + Vx̂ (x, x̂)a(x̂) +

 D(x) − D(x̂) u2
 
= ∥u∥2 −
2 2 2
1 −1 T 1 −1 T
− w(x, x̂)u + w(x, x̂) Θ T (x, x̂)Θ (x, x̂) w (x, x̂). + w(x, x̂) Θ T (x, x̂)Θ (x, x̂) w (x, x̂) ≤ 0
 
(39)
2 2
It turns out, by HJI (32), w(x, x̂) = Vx (x, x̂)B(x) + Vx̂ (x, x̂)B(x̂)
T 
+ c (x) − c (x̂) D(x) − D(x̂)
 
V̇ (x, x̂) = Vx,x̂ (x, x̂) f¯ (x, x̂) + G(x, x̂)u
 

1  2  then we have
− ∥r ∥2 − ∥y∥2 + ŷ
2 ∥rτ ∥22 ≤ γ 2 ∥uτ ∥22 + 2Vx (x(0), x̂(0)). (40)
T 
− D(x̂)u c (x) − c (x̂) + D(x) − D(x̂) u
   
The proof of (40) is straightforward and can be done as follows.
1 2 Consider
≤ − Θ (x, x̂)u − Θ −T (x, x̂)w T (x, x̂)
2
V̇ (x, x̂) = Vx (x, x̂)a(x) + Vx̂ (x, x̂)a(x̂)
γ2 1
+ Vx (x, x̂)B(x) + Vx̂ (x, x̂)B(x̂) u
 
+ ∥ u∥ 2 − ∥r ∥2
2 2
∥r ∥2 1
c (x) − c (x̂)2 + c (x) − c (x̂) T
γ2
  
1 =
≤ 2
∥ u∥ − ∥r ∥ .
2
(36) 2 2
2 2 1   2
× D(x) − D(x̂) u +  D(x) − D(x̂) u
 
Thus, by adopting the evaluation window [0, τ ], we finally have 2
1
∥rτ ∥22 ≤ γ 2 ∥uτ ∥22 + 2V (x(0), x̂(0)) Θ (x, x̂)u − Θ −T (x, x̂)w T (x, x̂)2

2
which completes the proof.  γ2 1 
 D(x) − D(x̂) u2 − w(x, x̂)u
 
= ∥ u∥ 2 −
Theorem 3 provides us with an algorithm for the design of 2 2
an L2 -NFDF. It consists of (i) solving HJI (32) for V (x, x̂) and 1 −1 T
+ w(x, x̂) Θ T (x, x̂)Θ (x, x̂) w (x, x̂).

(ii) solving (34) for L(x̂). It is worth to notice that the solvability 2
of (32) and (34) leads to (36), which means that the affine system
It turns out
(2) is weakly output re-constructible, as discussed in Section 3.
If they are solvable, the following FD scheme can be applied: ∥r ∥2 γ2 1
c (x) − c (x̂)2

(i) run the residual generator (25)–(26), (ii) set Jth = γ 2 ∥uτ ∥22 + = ∥ u∥ 2 +
2 2 2
2V (x(0), x̂(0)) and (iii) set the decision logic (29). − Vx (x, x̂)B(x) + Vx̂ (x, x̂)B(x̂) u

It is worth to mention that the introduction of the threshold is to 1 −1 T
+ w(x, x̂) Θ T (x, x̂)Θ (x, x̂) w (x, x̂)

reduce false alarms. An alternative way is to apply decoupling tech-
2
nique to make the residual generator robust against uncertainty or
1 2
disturbance [8,25]. If an optimal decoupling can be achieved, the − Θ (x, x̂)u − Θ −T (x, x̂)w T (x, x̂) =⇒ V̇ (x, x̂)
threshold can be set very small so that a high fault detectability 2
becomes achievable. ∥r ∥2 γ2
≤− + ∥ u∥ 2 .
2 2
4.2. On FD schemes for L2 -stable systems
∥rτ ∥2 γ 2 ∥uτ ∥2
It yields 2 2 + V (x, x̂) ≤ 2
2
+ V (x(0), x̂(0)) and finally we
We now consider the NFDF design problem for a special class have (40).
of affine systems. We assume that (2) is L2 -stable. This work is We now compare the FD scheme based on (40) and the one
motivated by practical applications and can be considered as an given in (37). Recall that for the system (2), (37) holds only if
 Y. Yang et al. / Systems & Control Letters 82 (2015) 18–25 23

γu2 I − DT (x)D(x) > 0, and, moreover, the following HJI is solvable

for given γu :
1
Vx (x, x̂)a(x) + Vx̂ (x, x̂)a(x̂) + ∥c (x)∥2
2
1  −1 T
+ w(x, x̂) Θ T (x, x̂)Θ (x, x̂) w (x, x̂) ≤ 0

(41)
2
w(x, x̂) = Vx (x, x̂)B(x) + Vx̂ (x, x̂)B(x̂) + c T (x)D(x)
γu2 I − DT (x)D(x) = Θ T (x, x̂)Θ (x, x̂).
Fig. 1. The comparison of ∥r (t )∥ and β(∥x(0) − x̂(0)∥, t ).
Assume that x̂ is a good estimate of x and ensures that
T 
D(x) − D(x̂) D(x) − D(x̂) < DT (x)D(x), where λ1 , λ2 (λ2 ≥ λ1 ) are eigenvalues of P. Based on Theorem 1,
 
we have V̇ (x, x̂) ≤ − λa V (x, x̂). It follows from Lemma 6.1 in [35]
c (x) − c (x̂)2 < ∥c (x)∥2
 
2
that
and, for a γ 2 smaller than γu2 , the HJI (39) is solvable. Then, the  s
λ2 dr λ2
τ
dependence of the threshold on the process input u, γ 2 0 ∥u∥2 dt, η(s) = − =− ln(s),
τ 1 ar a
is less than γ 0 ∥u∥ dt. On the other hand, since ŷ = c (x̂)+D(x̂)u
2
u
2
− λa t
is independent of any fault, the influence of a fault vector, for γ̃ (s, t ) = η−1 (t + η(s)) = se 2

instance the sensor fault, on the residual vector is identical with s − λa t s

γ (s, t ) = γ̃ (s, t ) + = se 2 +
the one on y. As a result, the FD performance of Σk given in (38) is t +1 t +1
improved in comparison with an FD system based on a direct use 2

a 1

− t
β x(0) − x̂(0) , t = V (x0 , x̂0 ).
 
e λ2 +

of the process input and output variables. In the next section, we
shall demonstrate the possible improvement by an example. λ1 t +1
The above discussion is of practical interest, since in many In our simulation, we choose a = 0.5, b = 1, c = 3.2, d = 3.5 and
automatic control systems, the plant is stable and a parallel input u = 1, which results in l1 = 34.7 and l2 = 18.3. The initial
running model is embedded in the system for monitoring or conditions are chosen as x(0) = (0.15, 0.15) and x̂(0) = (0, 0). A
control purpose (e.g. in an internal model control configuration). comparison of ∥r (t )∥ and β(x(0) − x̂(0) , t ) is depicted in Fig. 1.
By means of the above detection scheme, an L2 -stable NFDF with
satisfactory FD performance can be realized without additional
online computation and engineering costs.
Example 2. This example demonstrates the main results of
5. Examples Theorem 3 for the design of an L2 -stable NFDF. Consider the
following affine system
In this section, we illustrate the main results obtained in the      
ẋ1 x2 − sin3 x1 sin x1
previous sections by means of three examples. = + u, y = sin2 x1
ẋ2 −a sin x1 0
Example 1. This example is taken from [34] and used to illustrate
Theorem 1. Consider the nonlinear system (1) with with a > 12 . Let V (x, x̂) = a(1 − cos x1 )+ 21 x22 + a(1 − cos x̂1 )+ 21 x̂22 ,
   3  the HJI (32) is satisfied with γ ≥  a 1 . By solving Vx̂ (x, x̂)L(x̂) =
a− 2
x1 −x 1 + x 2
x= , f (x, u) = , h(x, u) = x1 . (42) 1 
x2 −x32 + u sin2 x̂1 , we have L(x̂) = sin x̂1
, the FDsystem can be designed as
a
0
Suppose that u ∈ U = [−1, 1]. As can be seen from [34], for every
initial condition x(0) ∈ R2 and u ∈ U = [−1, 1], the solution √ x(t )
 
sin x̂1
˙x̂1 3
(sin x1 − sin x̂1 )
2 2
 
of (42) enters the compact set S = {x ∈ R2 : 12 x21 + 12 x22 ≤ 10/2}. x̂ − sin x̂1 + u sin x̂1 +
= 2 a
Now, design the observer-based residual generator as x̂˙ 2
−a sin x̂1
x̂˙ 1 = −x̂31 + x̂2 + l1 (y − x̂1 ), r = sin2 x1 − sin2 x̂1 .
x̂˙ 2 = −x̂32 + u + l2 (y − x̂1 ), r = y − x̂1 . In order to verify the L2 stability of the NFDF, we choose the input

1 −a
 function shown in Fig. 2. The simulation results in Fig. 3 show that
Let V (x, x̂) = 1
2
(x − x̂)T P (x − x̂), P = −a b
and denote the output signal r (t ) of the NFDF is L2 bounded in the fault-free
e1 = x1 − x̂1 , e2 = x2 − x̂2 . Assume that x21 + x22 ≤ c and x̂21 + x̂22 ≤ d, case with a = 0.6 and x(0) = (1, −0.3), x̂(0) = (−1, −0.2).
it is easy to verify For FD purpose, we choose a constant sensor fault 0.7 occurred
at 60 s. With residual evaluation and threshold computation
Vx (x, x̂)f (x, u) + Vx̂ (x, x̂)φ(x̂, u)
  method provided in Theorem 3, it is evident that the fault can be
9a a detected as shown in Fig. 4.
≤ (c + d) − (l1 − al2 ) e21 − e22 + (1 + al1 − bl2 )e1 e2 .
2
8 2 In the last example, we compare two different L2 -FD schemes
as discussed in the last part of Section 4: (i) FD using y and relation
Select l1 and l2 such that (c + d) − l1 + al2 + 2a = 0, 1 + al1 − bl2 =
9a
8
2
(37) and (ii) FD by means of system (38).
0 hold for arbitrary a, b, d satisfying b > a2 , a > 0 and d > c,
then we have system (42) is output re-constructible for Example 3. Consider a system described by
λ1 λ2 1
r T r ≤ V (x, x̂) ≤ (x − x̂)T (x − x̂) ẋ1 = −x31 + x 1 u, ẋ2 = −x2 − x32 , y = x2 + 2u + w
2 2 2
a
Vx (x, x̂)f (x, u) + Vx̂ (x, x̂)φ(x̂, u) ≤ − (x − x̂) (x − x̂) T
where w represents a sensor fault.
2
24 Y. Yang et al. / Systems & Control Letters 82 (2015) 18–25

Fig. 5. Input signal u(t ).

Fig. 2. Input signal u(t ).

Fig. 6. Fault detection based on y and u.

Fig. 3. Residual signal r (t ).

Fig. 7. Fault detection based on residual r.

Fig. 4. Detection of a sensor fault.
generator, residual evaluation and decision making, we have first
(i) Determining γu for the threshold setting based on in- introduced the concepts of output re-constructability as well as
put/output dynamics and (37). Let V (x) = x21 + x22 . The HJI (41) weak output re-constructability and then proven that if a nonlinear
√ system is output re-constructible, an L∞ observer-based FD
is satisfied for γu > 6, which means ∥yτ ∥22 ≤ 6 ∥uτ ∥22 + γo . As-
sume that ∥x(0)∥ ≤ 0.5. It is then reasonable to set system exists. For constructing an L2 observer-based FD system,
the weakly output re-constructability is sufficient. In the second
 t0 +τ
part of our work, an integrated design scheme for affine nonlinear
Jth = 6 ∥u∥2 dt + 1.
t0
systems with the aid of L2 -stability theory is proposed and
applied to investigating FD issues for L2 stable systems. Finally,
(ii) Determining γ for the threshold setting in FD system
√
(38). Let examples have been given to illustrate the theoretical results. Our
V (x, x̂) = x21 + x22 + x̂21 + x̂22 . The HJI (38) holds for γ >
2
. Assume 2
recent and also future work is dedicated to (i) robustness issues
that ∥x(0)∥ ≤ 0.5, x̂(0) ≤ 0.5. The corresponding threshold is of observer-based FD systems, (ii) design of observer-based FD
 
set to be systems for a general class of nonlinear systems based on the
 t0 +τ existence conditions given in Section 3 and with the aid of e.g.
1
Jth = ∥u∥2 dt + 2. T–S-Fuzzy technique and (iii) study on nonlinear observer-based
2 t0 FD systems in the context of IOS and ROS.
In order to demonstrate the above two FD schemes, a constant
sensor fault w = 1 is considered and inserted at t = 70 s. We Acknowledgement
choose τ = 10 s and the input function u(t ) shown in Fig. 5. It can
be seen from Fig. 6, with the evaluation based on y and threshold This work has been supported by the National Natural Science
computation by means of (37), the fault cannot be detected, while Foundation of China under grants 61433001, 61174052 and
a fault detection based on residual evaluation of FD system (38) 61473004.
can be realized as shown in Fig. 7. The initial conditions are x(0) =
0.5
T T References
, x̂(0) = −0.5 −0.3 .
 
1
[1] E. Alcorta-Garcia, P. Frank, Deterministic nonlinear observer based approaches
to fault diagnosis: A survey, Control Eng. Pract. 5 (5) (1997) 663–670.
6. Conclusions and future works [2] S.K. Nguang, P. Shi, S.X. Ding, Fault detection for uncertain fuzzy systems: An
LMI approach, IEEE Trans. Fuzzy Syst. 15 (6) (2007) 1251–1262.
[3] M. Chadil, A. Abdo, S.X. Ding, H− /H∞ fault detection filter design for discrete-
In this paper, observer-based FD issues for nonlinear systems time Takagi–Sugeno fuzzy system, Automatica 49 (7) (2013) 1996–2005.
have been addressed. For the analysis purpose of an observer- [4] A. Xu, Q. Zhang, Nonlinear system fault diagnosis based on adaptive
based FD system, consisting of an observer-based residual estimation, Automatica 40 (7) (2004) 1181–1193.
 Y. Yang et al. / Systems & Control Letters 82 (2015) 18–25 25

[5] X. Zhang, M.M. Polycarpou, T. Parisini, Adaptive fault diagnosis and fault- [20] Z. Mao, B. Jiang, P. Shi, Protocol and fault detection design for nonlinear
tolerant control of MIMO nonlinear uncertain systems, Internat. J. Control 83 networked control systems, IEEE Trans. Circuits Syst. II: Express briefs 56 (3)
(5) (2010) 1054–1080. (2009) 255–259.
[6] J. Bokor, G. Balas, Detection filter design for LPV systems—a geometric [21] M. Abid, W. Chen, S.X. Ding, A.Q. Khan, Optimal residual evaluation for
approach, Automatica 40 (3) (2004) 511–518. nonlinear systems using post-filter and threshold, Internat. J. Control 84 (3)
[7] S. Armeni, A. Casavola, E. Mosca, Robust fault detection and isolation for (2011) 526–539.
LPV systems under a sensitivity constraint, Internat. J. Adapt. Control Signal [22] A.Q. Khan, S.X. Ding, Threshold computation for fault detection in a class of
Process. 23 (1) (2009) 55–72. discrete-time nonlinear systems, Internat. J. Adapt. Control Signal Process. 25
[8] T. Floquet, J.P. Barbot, W. Perruquetti, M. Djemai, On the robust fault detection (5) (2011) 407–429.
via a sliding mode disturbance observer, Internat. J. Control 77 (7) (2004) [23] A.Q. Khan, M. Abid, S.X. Ding, Fault detection filter design for discrete-time
622–629. nonlinear systems—A mixed H− /H∞ optimization, Systems Control Lett. 67
[9] X.G. Yan, C. Edwards, Robust sliding mode observer-based actuator fault (2014) 46–54.
detection and isolation for a class of nonlinear systems, Int. J. Syst. Sci. 39 (4) [24] A. van der Schaft, L2—Gain and Passivity Techniques in Nonlinear Control,
(2008) 349–359. Springer, 2000.
[10] J. Zarei, M.A. Tajeddini, H.R. Karimi, Vibration analysis for bearing fault [25] S.X. Ding, Model-Based Fault Diagnosis Techniques—Design Schemes, Algo-
detection and classification using an intelligent filter, Mechatronics 24 (2) rithms and Tools, second ed., Springer-Verlag, London, 2013.
(2014) 151–157. [26] M. Vidyasagar, On the stabilization of nonlinear systems using state detection,
[11] S. Yin, G. Wang, H.R. Karimi, Data-driven design of robust fault detection IEEE Trans. Automat. Control 25 (3) (1980) 504–509.
system for wind turbine, Mechatronics 24 (4) (2014) 298–306. [27] D.-J. Pan, Z.-Z. Han, Z.-J. Zhang, Bounded-input-boundd-output stabilization
[12] J. Bokor, Z. Szabo, Fault detection and isolation in nonlinear systems, Annu. of nonlinear systems using state detectors, Systems Control Lett. 21 (3) (1993)
Rev. Control 33 (2) (2009) 113–123. 189–198.
[13] R. Seliger, P. Frank, Fault diagnosis by disturbance decoupled nonlinear [28] W.-M. Lu, A state-space approach to parameterization of stabilizing controllers
observers, in: Proc. of the 30th Conference on Decision and Control, 1991, for nonlinear systems, IEEE Trans. Automat. Control 40 (9) (1995) 1576–1588.
pp. 2248–2253. [29] K. Fujimoto, T. Sugie, State-space characterization of youla parametrization for
[14] H. Hammouri, M. Kinnaert, E.E. Yaagoubi, Observer-based approach to fault nonlinear systems based on input-to-state stability, in: Proc. of the 37th IEEE
detection and isolation for nonlinear systems, IEEE Trans. Automat. Control 44 Conference on Decision and Control, 1998, pp. 2479–2484.
(10) (1999) 1879–1884. [30] E. Sontag, Y. Wang, Lyapunov characterizations of input to output stability,
[15] P. Kabore, H. Wang, Design of fault diagnosis filters and fault-tolerant control SIAM J. Control Optim. 39 (1) (2001) 226–249.
for a class of nonlinear systems, IEEE Trans. Automat. Control 46 (11) (2001) [31] A.-R. Teel, L. Praly, A smooth Lyapunov function from a class-KL estimate
1805–1810. involving two positive semidefinite functions, Control Optim. Calc. Var. 29 (5)
[16] C.D. Persis, A. Isidori, A geometric approach to nonlinear fault detection and (2000) 313–367.
isolation, IEEE Trans. Automat. Control 46 (6) (2001) 853–865. [32] Q. Gao, G. Feng, Y. Wang, J. Qiu, Universal fuzzy controllers based on
[17] A.M. Pertew, H.J. Marquez, Q. Zhao, LMI-based sensor fault diagnosis for generalized T–S fuzzy models, Fuzzy Sets and Systems 201 (2012) 55–70.
nonlinear Lipschitz systems, Automatica 43 (8) (2007) 1464–1469. [33] E. Sontag, Y. Wang, Output-to-state stability and detectability of nonlinear
[18] X. Zhang, M.M. Polycarpou, T. Parisini, Fault diagnosis of a class of nonlinear systems, Systems Control Lett. 29 (5) (1997) 279–290.
uncertain systems with Lipschitz nonlinearities using adaptive estimation, [34] I. Karafyllis, C. Kravaris, Global exponential observers for two classes of
Automatica 46 (2) (2010) 290–299. nonlinear systems, Systems Control Lett. 61 (7) (2012) 797–806.
[19] X. He, Z. Wang, D. Zhou, Robust H∞ filtering for time-delay systems with [35] E. Sontag, Smooth stabilization implies coprime factorization, IEEE Trans.
probabilistic sensor faults, IEEE Signal Process. Lett. 16 (5) (2009) 442–445. Automat. Control 34 (4) (1989) 435–443.