Automatica 38 (2002) 1815 – 1821

www.elsevier.com/locate/automatica

Brief Paper
Sliding mode observers for detection and reconstruction
of sensor faults

Chee Pin Tan1 , Christopher Edwards ∗
Control and Instrumentation Group, Department of Engineering, University of Leicester, Leicester, LE1 7RH, UK
Received 16 February 2001; received in revised form 9 October 2001; accepted 24 May 2002

Abstract

This paper proposes two methods for detecting and reconstructing sensor faults using sliding mode observers. In both methods, 2ctitious
systems are introduced in which the original sensor fault appears as an actuator fault. The original sensor faults are then reconstructed using
a ‘secondary’ sliding mode observer. For both methods, there are certain conditions which must be satis2ed for successful fault detection
and reconstruction. The methods are demonstrated using a chemical process example. ? 2002 Elsevier Science Ltd. All rights reserved.

Keywords: Sliding mode; Observer; Fault detection and isolation; Fault reconstruction; Linear matrix inequalities

1. Introduction FDI. Sreedhar, Fernandez, and Masada (1993) used sliding

mode observers for fault detection, assuming that all states
Fault detection and isolation (FDI) is becoming an ever are available. Hermans and Zarrop (1996) designed a slid-
increasingly important area. A fault is deemed to occur when ing mode observer such that, in the presence of faults, the
the system experiences an abnormal condition, such as a sliding motion is destroyed. Edwards, Spurgeon, and Patton
malfunction in the actuators=sensors. The fundamental pur- (2000) used the same observer to reconstruct the faults using
pose of an FDI scheme is to generate an alarm when a fault the so-called ‘equivalent output estimation error injection’
occurs and also to identify its location. Overviews of work concept.
in this area appear in Patton, Frank, and Clark (1989), Frank This paper builds on the work of Edwards et al. (2000)
(1996) and Chen and Patton (1999). Observer-based meth- in two ways. Firstly, by considering a linear matrix inequal-
ods are commonly used, where the observer signals are used ity (LMI) observer design method proposed by Tan and
to infer information about the fault. Residual generation ap- Edwards (2000), more complicated examples can be con-
proaches, using linear observers, have been widely used, sidered. Previous examples in Edwards et al. (2000) and
where the di;erence between the system output and observer Edwards and Spurgeon (2000), although challenging from
output is processed to form the so-called residuals. Ideally, certain view points, were of low order and allowed for
these will be zero during fault-free operation but will give straightforward tuning of the gains in the observer. Secondly,
a speci2c response when a certain fault occurs. Residual in the work of Edwards et al. (2000), only the steady-state
generation techniques have been demonstrated by Zhang, components of sensor faults were replicated. This paper will
Basseville, and Benveniste (1998) and Magni and Mouyon present methods to reconstruct sensor faults. The concept
(1994). Recently, sliding mode observers have been used for of reconstructing faults will enable the detection of subtle
drifts in the sensor measurements, which are diFcult to de-

This paper was not presented at any IFAC meeting. This paper was tect using other methods. In this paper, two approaches will
recommended for publication in revised form by Associate Editor Rene be considered.
Boel under the direction of Editor Tamer Basar. The structure of the paper is as follows: the preliminaries
∗ Corresponding author. Tel.: +44-116-223-1303; fax: +44-116-
and background work for this paper will be presented 2rst,
252-2619. introducing the sliding mode observer and its design method
E-mail addresses: tan.chee.pin@engsci.monash.edu.my (C.P. Tan),
ce@sun.engg.le.ac.uk (C. Edwards).
(Tan & Edwards, 2000) followed by the method to recon-
1 Present address: School of Engineering and Science, Monash Univer- struct actuator faults by Edwards et al. (2000). Next, the
sity Malaysia, 2 Jalan Kolaj, Bandar Sunway, 46150 Petaling, Malaysia. 2rst method to reconstruct sensor faults will be presented,

0005-1098/02/$ - see front matter ? 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 5 - 1 0 9 8 ( 0 2 ) 0 0 0 9 8 - 5
1816 C.P. Tan, C. Edwards / Automatica 38 (2002) 1815–1821

utilizing LMIs to ensure that the sliding motion of a Minimize trace(X ) w.r.t. P, X and Y subject to
‘secondary observer’ is stable. The second method to re-  
PA + AT P − YC − (YC)T P Y
construct sensor faults will then be presented. Finally, both  
 −W −1 0 
methods will be demonstrated with an example, which is a  P  ¡ 0; (6)
17th-order model of a chemical plant.
YT 0 −V −1
 
−P I
2. Preliminaries ¡ 0; (7)
I −X
This section introduces the preliminaries and background where X; P ∈ Rn×n are s.p.d. matrix variables and the matri-
work that is vital to the work in this paper. ces W ∈ Rn×n and V ∈ Rp×p are user de2ned s.p.d. weights.
The matrix variable P has the structure
2.1. The sliding mode observer  
P11 P12
P= T
¿ 0; P12 := [ P121 0 ]; (8)
Consider the dynamical system P12 P22

ẋ(t) = Ax(t) + Bu(t) + Ffi (t); (1) where P11 ∈ R(n−p)×(n−p) and P121 ∈ R(n−p)×(p−q) . Stan-
dard LMI software can be used to synthesize P, X and Y .
y(t) = Cx(t); (2)
Given the LMI variables P11 ; P12 ; P22 ; Y , de2ne
where x ∈ Rn , u ∈ Rm and y ∈ Rp . Assume that C and F −1
L = [P11 P121 0(n−p)×q ] (9)
are full rank. The signal fi ∈ Rq represents an actuator fault
and is unknown but bounded so that fi (t) 6 (t; y; u)−0 then the observer matrix P0 in (4) is chosen as
where  : R+ ×Rp ×Rm → R+ is known and 0 is a positive T −1
scalar. Assuming that p ¿ q, Edwards and Spurgeon (1994) P0 = T (P22 − P12 P11 P12 )T T (10)
proposed an observer and the gains
z˙ = Az + Bu − Gl ey + Gn ; (3)
Gl = P −1 Y; GnT = [ − TLT T ]: (11)
where the discontinuous term  is de2ned by
 The motivation for this design is given in Tan and Edwards
 −(t; y; u)CF P0 ey
 if ey = 0; (2000). Satisfying conditions A1 and A2 guarantees the ex-
= P0 ey  (4) istence of an observer despite the restriction on the structure

0 of P12 . This does, however, introduce a degree of conser-
otherwise;
vatism into the design. The restriction on P12 arises from
where ey := Cz − y and P0 ∈ Rp×p is a symmetric positive the requirement of satisfying a structural constraint between
de2nite (s.p.d.) matrix. The matrices P0 , Gl and Gn are to P, F and C. For details, see Edwards and Spurgeon (1994).
be determined. Edwards and Spurgeon (1994) argue that an Additional LMIs to (6) and (7) may be incorporated to force
observer of the form given in (3) which rejects the e;ect of '(A−Gl C) to lie in certain convex regions (Tan & Edwards,
fi on the estimation error, exists i; 2000).
Applying the change of coordinates
A1. rank(CF) = q,  
A2. invariant zeros of (A; F; C) are stable. In−p L
TL = (12)
Edwards and Spurgeon (1994) have proven that if A1 is 0 T
satis2ed, then there exists a change of coordinates in which
to (A; F; C) and Gn yields, respectively,
the triple (A; F; C) has the following structure:      
    A11 A12 0 0
A11 A12 0 A= ; F= T
; C = ;
A= ; F= ; C = [0 T ]; (5) A21 A22 F2 Ip
A21 A22 F20
 
where A11 ∈ R(n−p)×(n−p) ; F20 ∈ Rq×q is non-singular and 0
Gn = ; (13)
T ∈ Rp×p is orthogonal. De2ne A211 as the top p − q rows Ip
of A21 . It can be shown that the pair (A11 ; A211 ) is de-
tectable and its unobservable modes are the invariant ze- where F2 = TF2 and A11 = A11 + LA21 is stable.
ros of (A; F; C). De2ne F2 as the bottom p rows of F.
Tan and Edwards (2000) presented a method for design- 2.2. Reconstruction of faults
ing this observer using LMIs. A slight modi2cation to this
method will be used in this paper. The design method can be Edwards et al. (2000) proposed a method to use sliding
summarized as: mode observers to reconstruct the fault fi . Assume that
 C.P. Tan, C. Edwards / Automatica 38 (2002) 1815–1821 1817

observer (3) has been designed, and a sliding motion has

occurred (ey = ė y = 0). De2ne eq as the equivalent output
error injection required to maintain a sliding motion (Utkin,
1992). As argued by Edwards et al. (2000) this can be ap-
proximated online to any required accuracy. During the slid-
ing motion (governed by A11 ), the estimation error system
in the coordinates of (13) is given by
ė 1 = A11 e1 ; (14)

0 = A21 e1 + eq − F2 fi : (15)

Since A11 is stable, the reconstruction signal
f̂i := (FT2 F2 )−1 FT2 eq → fi (16) Fig. 1. Schematic of the fault detection structure using the secondary
observer.
and can be computed online, from eq .
Now consider the system in (1) subject to sensor faults.
In this case, fi = 0 and Eq. (2) is replaced by The signal zf is available since eq is computable online.
Eqs. (22) and (23) are now in the form of (1) and (2), where
y = Cx + f0 ; (17)
the ‘actuator fault’ is f0 and the ‘output’ is zf . A sliding
where f0 ∈ Rp is the sensor fault vector. Assume a sliding mode observer as described in Section 2.1 can therefore be
motion has been attained, the error system in the partition used to reconstruct the sensor fault f0 . This will be termed
of (13) satis2es the ‘secondary observer’. The triple (A0 ; B0 ; C0 ) is a square
system, hence no design freedom exists associated with the
ė 1 = A11 e1 + A12 f0 ; (18) sliding motion (Edwards & Spurgeon, 1994). By applying
the following change of coordinates:
ḟ0 = A21 e1 + A22 f0 + eq : (19)  
In−p A12 A−1 −1
22 Af
As argued by Edwards et al. (2000) if A is full rank, then T0 = (24)
det(A22 − A21 A−1 0 Ip
11 A12 ) = 0 and at (pseudo) steady state
to the triple (A0 ; F0 ; C0 ) to obtain the canonical form (13),
f0 ≈ −(A22 − A21 A−1 −1
11 A12 ) eq : (20)
it can be shown that the sliding motion of the secondary
observer is governed by '(A11 −A12 A−1 22 A21 ). Notice that
C0 F0 =−Af A22 and so is full rank i; A22 is full rank. Thus,
3. Reconstruction of sensor faults I
necessary requirements for the existence of an appropriate
secondary observer are that (A11 − A12 A−1 22 A21 ) is stable
This paper seeks to improve on the estimate for f0 given
and A22 is invertible. The equivalent of Eq. (16), f̂0 :=
above. Consider a new state zf ∈ Rp which is de2ned by
− A−122 Af eq; o where eq; o is the output error injection of
z˙f = −Af zf + Af eq ; (21) the secondary observer, represents a reconstruction of f0 .
The complete FDI scheme is shown in Fig. 1.
where −Af ∈ Rp×p is a stable 2lter matrix. (Typically, Af
represents a diagonal positive de2nite matrix where the di- Remark 1. From the Schur complement; if det(A) = 0 and
agonal elements represent inverse time constants.) Assume A22 is non-singular; then det(A11 − A12 A−1
22 A21 ) = 0 and
the sensor fault represents a slow incipient drift ḟ0 ≈ 0 (step so no stable sliding motion exists.
sensor failures are relatively easy to detect because they usu-
ally break the sliding motion (Edwards & Spurgeon , 2000)). Remark 2. A key observation is that the sliding motion is
Combining (18), (19) and (21), the following state-space independent of Af ; so whilst Af will a;ect the gains of the
representation can be obtained observer; theoretically it has no e;ect on the reconstruction
      
ė 1 A11 0 e1 A12 signal. In practise; the values of the diagonal elements of
= + f0 ; Af have been chosen in an e;ort to maintain=establish good
z˙f −Af A21 −Af zf −Af A22 numerical conditioning in the system matrices in (22) about
   
A0 F0 which the secondary observer is designed.
(22)
  3.1. Designing the secondary observer using LMIs
e1
zf = [0 Ip ] : (23) From the previous section, the problem now is to make
  zf
C0 A22 invertible and A11 − A12 A−1
22 A21 stable by choice of
1818 C.P. Tan, C. Edwards / Automatica 38 (2002) 1815–1821

L, whilst retaining the property that A11 is stable. This will Lemma 4. The matrix K is invertible if A is invertible.
be achieved in two stages. Write L = L1 + L2 where L1 and
L2 do not necessarily have the same structure as L in (9). Proof. If A is invertible; Ã in (26) is invertible for any L1 .
Decompose TL from (12) as It can be seen from (31) that K is a Schur complement for
   Ã. By design; L1 is chosen to make Ã22 invertible; and hence
I L2 I L1
TL = TL; 2 TL; 1 = : (25) K is invertible.
0 T 0 I
Applying TL; 1 to the matrix A from (5) will yield From Lemma 1, making A11 − A12 A−1 22 A21 stable is
  equivalent to making KJ −1 stable, which is achievable by
−1
Ã11 Ã12 making JK −1 stable. (This implies that JK −1 will be invert-
à = TL; 1 ATL; 1 = ; (26)
Ã21 Ã22 ible, implying that J will be invertible, and from Lemma 3,
A22 will be invertible.) Selecting L2 so that JK −1 is sta-
where Ã22 = A22 − A21 L1 . Applying TL; 2 to Ã, the elements ble is equivalent to 2nding a s.p.d. P̃ ∈ R(n−p)×(n−p) and L2
of A in (13) can then be written as satisfying
A11 = Ã11 + L2 Ã21 ; (27) P̃JK −1 + (P̃JK −1 )T ¡ 0: (33)
−1 −1
A12 = ((Ã12 + L2 Ã22 ) − (Ã11 + L2 Ã21 )L2 )T T ; (28) De2ne M =(I +L1 Ã22 Ã21 )K −1 and N =Ã22 A21 K −1 . Choose
P̃ =P11 from the Lyapunov matrix in (8). Then recalling that
A21 = T Ã21 ; (29) L = L1 + L2 , and using the switch of variables P11 L = P12
from Section 2, inequality (33) becomes
A22 = T (Ã22 − Ã21 L2 )T T : (30)
P11 M + (P11 M )T − P12 N − (P12 N )T ¡ 0: (34)
Based on these de2nitions the following lemmas hold:
Therefore, if inequality (34) and P11 ¿ 0 have a feasible
Lemma 1. The matrix (A11 − A12 A−1 solution, then A11 − A12 A−1 22 A21 is stable by choice of L.
22 A21 ) can be ex-
pressed as KJ −1 where Inequality (34) can be added to inequalities (6) and (7) when
−1 −1 designing the primary observer. The LMI variables involved
K := Ã11 − Ã12 Ã22 Ã21 ; J := I − L2 Ã22 Ã21 : (31) here are P11 and P12 , which are a subset of the variables in
the optimization problem in Section 2. The design problem
Proof. This follows from straightforward algebra. may be summarized as:
Minimize trace(X ) with respect to the LMI variables
Lemma 2. If det(A) = 0 then there exists an L1 such that P11 ; P12 ; P22 ; X and Y subject to (6), (7) and (34).
det(A22 − A21 L1 ) = 0.
Remark. Since P11 JK −1 is aFne in P11 and P12 ; pole clus-
Proof. Under the assumption that det(A) = 0 it follows tering methods as described in (Chilali & Gahinet; 1996)
rank[A21 A22 ] = p. Thus; the matrix pencil [sI − A22 A21 ] can be incorporated to force '(JK −1 ) to lie in certain convex
associated with the PHB controllability test for (A22 ; A21 ) regions of the complex plane. For example; if the eigenval-
has full rank at s = 0. This implies that s = 0 is a control- ues of JK −1 lie inside a disc centred at (q̃c ; 0) with a radius
lable mode of (A22 ; A21 ). Hence; L1 can always be chosen r̃ c where q̃c = qc =(qc2 − rc2 ); r̃ c = rc =(qc2 − rc2 ) with qc ¡ 0 and
so that Ã22 is full rank. rc ¡ |qc |; then '(A11 − A12 A−1 22 A21 ) will lie in the disc
centred at (qc; ; 0) with a radius rc .
Assuming Ã22 is invertible the following can be proven:
3.2. An example
Lemma 3. J is invertible i4 A22 is invertible.
Now the method described above for sensor fault re-
Proof. The matrix J in (31) is a Schur complement of
  construction will be demonstrated with a chemical process
I L2 plant, which is a 270 state HDA-plant that produces ben-
Js := (32) zene. Hermann, Spurgeon, and Edwards (2000) linearized
Ã21 Ã22
and model-reduced the plant to a 17th-order system. The
since Ã22 is non-singular. Hence J is invertible i; Js is in- system has 2ve outputs, 2ve control inputs and is subject
vertible. However Js is invertible i; det(Ã22 − Ã21 L2 ) = 0 to two disturbance signals. Due to space constraints the pa-
(a Schur complement of Js ). From the de2nitions of L; Ã21 rameters of the plant and the observers will not be shown. It
and Ã22 ; can be veri2ed from Hermann et al. (2000) that CF is full
rank and (A; F; C) is minimum phase, where F represents
det(Ã22 − Ã21 L2 ) = det(T T A22 T ) = det(A22 )
the disturbance distribution matrix. For the main observer,
and hence det(J ) = 0 ⇔ det(A22 ) = 0. the weighting matrices W = I17 and V = I5 and the 2lter
 C.P. Tan, C. Edwards / Automatica 38 (2002) 1815–1821 1819

0.16 0.2
0.14 0
0.12
0.1 -0.2
0.08 -0.4
0.06
0.04 -0.6
0.02 -0.8
0
-0.02 -1
-0.04
0 10 20 30 40 50 0 10 20 30 40 50
Hours Hours

Fig. 2. The left sub2gure is a fault on sensor 1 Sash inlet temperature ◦F). The right sub2gure is a fault on sensor 2 (production rate (lb mol=h)).

0.16 0.2
0.14 0
0.12
0.1 -0.2
0.08 -0.4
0.06
0.04 -0.6
0.02 -0.8
0
-0.02 -1
-0.04
0 10 20 30 40 50 0 10 20 30 40 50
Hours Hours

Fig. 3. The left sub2gure is the reconstruction of the fault on sensor 1. The right sub2gure is the reconstruction of the fault on sensor 2.

0.16 0.2
0.14 0
0.12
0.1 -0.2
0.08 -0.4
0.06
0.04 -0.6
0.02 -0.8
0
-1
-0.02
-0.04
0 10 20 30 40 50 0 10 20 30 40 50
Hours Hours

Fig. 4. The left sub2gure is the ‘reconstruction’ of the fault on sensor 1. The right sub2gure is the ‘reconstruction’ of the fault on sensor 2.

Af = 10I5 . For the secondary observer, the weighting ma- 4. Reconstruction of sensor faults II
trices W0 = 0:01I17 and V0 = 100I5 . An additional LMI con-
straint was added to force the eigenvalues of the secondary In Section 3, the analysis ignored the e;ect of ḟ0 . De-
observer to lie to the right of a vertical line through −100 pending on the system and the size of ḟ0 , this may be
in the complex plane. Note that for this high-order example signi2cant. This section presents a method to fully recon-
it would have been impossible to do the design by hand us- struct the sensor fault, hence improving on the method in
ing pole placement design. In the following simulations the Section 3. Consider new states zg ∈ Rp that represent a
sensors were subject to white noise. 2ltered version of eq from the primary observer. Hence
Fig. 2 shows the faults acting on the sensors, and Fig. 3 from (19),
shows that the secondary observer reconstructs the fault
faithfully. In comparison, Fig. 4 shows the fault recon- z˙g = −Ag zg − Ag A21 e1 − Ag A22 f0 + Ag ḟ0 ; (35)
struction using the ‘steady-state’ method by Edwards et al.
(2000), approximating the fault by (20), and using only the where −Ag ∈ Rp×p is a stable 2lter matrix as in Section 3.
primary observer. It can be seen that the method in this pa- Now 2lter zg to obtain a new state zh
per is an improvement: the shape of the fault is preserved
with higher 2delity. z˙h = −Ah zh + Ah zg ; (36)
1820 C.P. Tan, C. Edwards / Automatica 38 (2002) 1815–1821

where −Ah ∈ Rp×p is a stable 2lter matrix, and de2ne a

new state w ∈ Rp as
w = zg − Ag f0 : (37)
Di;erentiating (37) and combining with (35), (18) and
(36) an augmented system of order n + p can be formed
    
ė 1 A11 0 0 e1
    
 ẇ  =  −Ag A21 −Ag 0  w 
 
   
z˙h 0 Ah −Ah zh
 
Aa
  Fig. 5. Schematic of the fault detection structure using the secondary
−A12 observer.
 
− 
 Ag (I + A22 ) f0 ; (38)
The particular observer gain will be a;ected but not the
A h Ag reconstruction signal calculations. This FDI scheme is
 
Fa schematically described in Fig. 5.
 
e1
  4.1. An example
zh = [0 0 Ip ]  w : (39)
   
Ca zh The method described in this section will be demonstrated
with the chemical plant in Section 3.2. For the main ob-
Eqs. (38) and (39) are now in the form of (1) and (2) and server, the matrices were the same as those in Section 3.2.
represent a system with an actuator fault f0 . Hence, the The 2lter matrices from (35) and (36) Ag = Ah = 10I5 . The
observer in Section 2.1 can be used, driven by the signal zh , design matrices for the secondary observer Wa = 0:01I22 and
to reconstruct f0 . Again, since (Aa ; Fa ; Ca ) is square, there is Va = 100I5 . The system was subjected to the same con-
no freedom to select the sliding motion. The canonical form ditions (noise=faults) as in Section 3.2.
(13) for the secondary observer can be directly obtained by Fig. 6 shows that the secondary observer in this case does
applying the transformation Ta to (Aa ; Fa ; Ca ) where reconstruct the sensor fault faithfully. It can be seen that this
 
In−p 0 −A12 A−1 g Ah
−1 method is an improvement on the one in Section 3.
 
Ta =  0 −A−1 g −(Ag + A22 )A−1 −1 
g Ah  (40)
0 0 Ip 5. Conclusion

and it can be shown that the sliding motion poles are This paper has proposed two methods to reconstruct
'(A), hence this method is only applicable to open-loop sensor faults using sliding mode observers. The paper
stable systems. It can be seen that Ca Fa is full rank. Af- builds on the work of Edwards et al. (2000) and Tan and
ter transformation (40), the fault f0 can be reconstructed Edwards (2000). Here the equivalent output estimation er-
−1
as f̂0 := A−1
g Ah eq; a where eq; a is the equivalent output ror injection is 2ltered to form a 2ctitious system in which
error injection of the secondary observer. Again the 2lter the sensor faults appear as actuator faults. The method to
matrices have no theoretical e;ect on the sliding motion. obtain a good approximation of actuator faults described by

0.16 0.2
0.14 0
0.12
0.1 -0.2
0.08 -0.4
0.06
0.04 -0.6
0.02 -0.8
0
-0.02 -1
-0.04
0 10 20 30 40 50 0 10 20 30 40 50
Hours Hours

Fig. 6. The left sub2gure is the reconstruction of the fault on the sensor 1. The right sub2gure is the reconstruction of the fault sensor 2.
 C.P. Tan, C. Edwards / Automatica 38 (2002) 1815–1821 1821

Edwards et al. (2000) is then used to estimate the original Frank, P. M. (1996). Analytical and qualitative model-based fault
sensor faults. This approach therefore requires the design of diagnosis—a survey and some new results. European Journal of
a secondary sliding mode observer. This places additional Control, 2, 6–28.
Hermann, G., Spurgeon, S. K., & Edwards, C. (2000). Model-based
design requirements on the primary sliding mode observer. control of the HDA-plant, a non-linear, large scale chemical
The paper shows how LMI methods can be used to solve process, using sliding mode and H∞ approaches. UKACC Control
this problem and to synthesize the required observer gains Conference, Cambridge, 2000.
eFciently. Hermans, F. J. J., & Zarrop, M. B. (1996). Sliding mode observers for
robust sensor monitoring. 13th IFAC world congress, San Francisco
(pp. 211–216).
References Magni, J. F., & Mouyon, P. (1994). On residual generation by observer
and parity space approaches. IEEE Transactions on Automatic
Chen, J., & Patton, R. J. (1999). Robust model-based fault diagnosis for Control, 39, 441–447.
dynamic systems. Dordrecht: Kluwer Academic Publishers. Patton, R. J., Frank, P. M., & Clark, R. N. (1989). Fault diagnosis in
Chilali, M., & Gahinet, P. (1996). H∞ design with pole placement dynamic systems: Theory and application. New York: Prentice Hall.
constraints: An LMI approach IEEE Transactions on Automatic Sreedhar, R., Fernandez, B., & Masada, G. Y. (1993). Robust fault
Control, 41, 358–367. detection in nonlinear systems using sliding mode observers. IEEE
Edwards, C., & Spurgeon, S. K. (1994). On the development of conference on control applications, Vancouver (pp. 715 –721).
discontinuous observers. International Journal of Control, 59, Tan, C. P., & Edwards, C. (2000). An LMI approach for designing sliding
1211–1229. mode observers. IEEE conference on decision and control, Sydney
Edwards, C., & Spurgeon, S. K. (2000). A sliding mode observer based (pp. 2587–2592).
FDI scheme for the ship benchmark. European Journal of Control, Utkin, V. I. (1992). Sliding modes in control optimization. Berlin:
6, 341–356. Springer.
Edwards, C., Spurgeon, S. K., & Patton, R. J. (2000). Sliding Zhang, Q., Basseville, M., & Benveniste, A. (1998). Fault detection and
mode observers for fault detection and isolation. Automatica, 36, isolation in nonlinear dynamic systems: A combined input-output and
541–553. local approach Automatica, 34, 1359–1373.