International Journal of Control, Automation, and Systems 21(3) (2023) 697-710 ISSN:1598-6446 eISSN:2005-4092

http://dx.doi.org/10.1007/s12555-022-0180-y http://www.springer.com/12555

Descriptor Observers for Robust Fault Reconstruction in a Class of Non-

linear Descriptor Systems
Joseph Chang Lun Chan , Wen-Shyan Chua, Tae H. Lee* , and Chee Pin Tan 

Abstract: This paper presents a robust fault reconstruction scheme for a class of nonlinear descriptor systems that
is affected by disturbances. The fault enters the system both linearly and through nonlinear functions in both the
state and output equations. The system is first re-expressed into a form which is easier to manipulate. Two cases
are considered: in the first case, the fault is present in the nonlinear function of the output equation, whereas in
the second case, the nonlinear function of the output equation does not contain the fault. In each case, a nonlinear
descriptor observer is used to reconstruct the fault, where the gains of the observer are designed using linear matrix
inequalities such that the effect of disturbances on the fault reconstruction is bounded. As a result of their structure,
the proposed descriptor observers do not require restrictions on the rates of the input and disturbances (unlike in
observers using a state-space structure) which makes the proposed observers more widely applicable. Finally, two
simulation examples are performed to verify the performance of the proposed observers.

Keywords: Descriptor systems, fault estimation, nonlinear systems, observers, uncertain systems.

1. INTRODUCTION be designed such that the effects of these disturbances on

the fault reconstruction are bounded, thus achieving robust
The descriptor system representation (which is synony- fault reconstruction [15].
mous with singular systems [1]) has been used to rep- Additionally, practical systems are often nonlinear [16].
resent systems containing both differential and algebraic Various observers in [4,17-20] however assume the de-
equations, and as such is more general than state-space scriptor system to be linear, which limits their applica-
systems (which only contain differential equations) [2]. bility. Motivated by this limitation, observers have been
Therefore, descriptor systems have been used to model proposed for nonlinear descriptor systems [21-23], but
various practical systems across different engineering do- many of these observers do not consider fault reconstruc-
mains, including electronic circuits [3], chemical mixing tion. Furthermore, most observers for nonlinear systems
systems [4], and Takagi-Sugeno systems [5]. Practical sys- [15,24-26] only consider faults entering the system lin-
tems are however prone to encounter faults during their early. To overcome this, some works considered systems
operation, which could lead to deterioration in perfor- where the fault is linearly multiplied by a nonlinear func-
mance and costly downtime [6]. Hence it is important to tion [11,27-30]; this approach however requires the out-
reconstruct these faults accurately as they occur for timely puts to be a linear combination of only the states, which
corrective action to be taken [7,8]. reduces the applicability of the observers.
Observers have been widely utilised to reconstruct To reconstruct faults, Ha and Trinh [31] proposed an-
faults in practical systems [9-11]. The design process other method for state-space systems, which is to form an
of the observer however requires a mathematical model augmented descriptor system which includes the fault as
of the system [12], which may not account for unmod- part of its state. Chua et al. [10] then considered state-
elled dynamics and uncertainties (both of which will here- space systems with disturbances and nonlinearities in the
after be referred to as disturbances) in the actual sys- outputs, relaxing the restriction in the other approach by
tem [13,14]. These disturbances may raise false alarms, [11,27-30]. Pan et al. [32] then utilised the augmented
or even mask ongoing faults. Hence the observer needs to state approach in nonlinear descriptor systems affected

Manuscript received February 24, 2022; revised June 6, 2022; accepted July 29, 2022. Recommended by Associate Editor Mien Van under
the direction of Senior Editor Bin Jiang. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the
Korea government (MSIT) (No.2020R1C1C1012707) and also was supported by National University Development Project in 2020.

Joseph Chang Lun Chan and Tae H. Lee are with the Division of Electronic Engineering, Jeonbuk National University, Jeonju 54896, Korea
(e-mails: joseph.chan1@monash.edu, fesuselee@gmail.com). Wen-Shyan Chua is with the Selangor Human Resource Development Centre,
Shah Alam 40100, Malaysia (e-mail: chua.wen.shyan@shrdc.org.my). Chee Pin Tan is with the Advanced Engineering Platform, Monash
University Malaysia, Bandar Sunway 47500, Malaysia (e-mail: tan.chee.pin@monash.edu). Joseph Chang Lun Chan, Wen-Shyan Chua, and
Chee Pin Tan are also with the School of Engineering, Monash University Malaysia, Bandar Sunway 47500, Malaysia.
* Corresponding author.

©ICROS, KIEE and Springer 2023

698 Joseph Chang Lun Chan, Wen-Shyan Chua, Tae H. Lee, and Chee Pin Tan

q
lim 1 τ kw(t)k2 dt, the empiri-
R
by faults entering through nonlinear functions in the state is defined as R(w) =
τ→∞ τ 0
equation. Finally, Chan et al. [33] extended these results cal RMS q (at time instant T ) of a signal g is defined as
into nonlinear descriptor systems with disturbances, as
RT (g) = T1 0T kg(t)k2 dt [15], ∗ denotes the transposed
R
well as faults entering through nonlinear functions, in both
the state and output equations. The utilised state-space ob- term in the symmetric positions of a matrix, r(t) is the unit
server however requires the rates of the input and distur- ramp function, and a square wave with an amplitude of a
bance to be bounded. Since disturbances are generally un- units and frequency of b Hz is denoted as S (a, b).
known signals [15] and restrictions on the input also limit
the feasibility of potential controllers, these requirements 2. PRELIMINARIES
reduce the applicability of the state-space observer.
To address the aforementioned shortcomings, this paper Consider the following nonlinear descriptor system
proposes a robust fault reconstruction scheme for nonlin-
ear descriptor systems, with disturbances in both the state Eo ẋo = Ao xo + Bo u + Mo f + Hxo h (To x̆o ) + Qo ξ , (1)
and output equations. Moreover, the fault affects the state yo = Co xo + Do u + No f + Hyo h (To x̆o ) + Ro ξ , (2)
and output equations both linearly and through nonlin-
ear functions. The original descriptor system is first re- where Eo , Ao ∈ Rn×n , Bo ∈ Rn×m , Mo ∈ Rn×q , Hxo ∈ Rn×b ,
formulated such that design freedom is separated from Qo ∈ Rn×h , Co ∈ R p×n , Do ∈ R p×m , No ∈ R p×q , Hyo ∈
other parts of its structure. Two approaches for design- R p×b , Ro ∈ R p×h , and To ∈ Rb×(n+q) are known and con-
ing descriptor observers are then developed depending on stant matrices, b is the number of arguments into h (·),
the system requirements: in the first case, faults are also while h (·) : Rb 7→ Rb is a known nonlinear function, whilst
present in the nonlinear functions in the output equation, xo ∈ Rn , u ∈ Rm , yo ∈ R p , f ∈ Rq , and ξ ∈ Rh represent
while in the second case the faults are absent from these the states, inputs, outputs, faults, and disturbances, respec-
 T
functions, but meet a certain rank condition instead. In tively, and x̆o = xoT f T . Assume generally that Eo is
each case, a descriptor observer is utilised to reconstruct rank-deficient and Co is full row-rank, i.e., rank (Eo ) = r <
the fault. Linear matrix inequalities (LMIs) are used to n, rank (Co ) = p. The objective is to reconstruct f using
design the observer gains to bound the effect of the dis- only u and yo .
turbances on the fault reconstruction. The proposed de- The following assumptions are first made as a founda-
scriptor observers do not require the input and disturbance tion for later portions.
rates to be bounded, making it more widely applicable Assumption
 1: System (1)-(2) is infinitely observable,
than state-space observers (which need these restrictions). i.e., rank EoT CoT = n.
The contribution of our work over existing schemes in the Assumption 2: The RMS of ξ is finite.
literature is summarised as follows:
Remark 1: Assumption 1 is commonly assumed in ob-
1) Two methods to design descriptor observers are pro- server design for descriptor systems [4,20,22,26,29,34],
posed for robustly reconstructing faults in nonlinear and will be used to re-express system (1)-(2). Assump-
descriptor systems with faults entering through lin- tion 2 is necessary to quantify the performance of the ob-
ear and nonlinear functions, where the applicability servers, and merely implies that the disturbances have a
of each method depends on the system structure. finite ‘average magnitude’. This assumption is applica-
2) The descriptor observers only require the disturbance ble to various practical signals (e.g., combinations of step
to be bounded (whereas the observers in [31-33] need functions, triangular waves, and sinusoids), and is more
additional restrictions on the dynamics of the distur- relaxed than those made in other schemes (e.g., ξ is also
bance and control input). required to be differentiable with respect to time, ξ˙ also
needs to have a finite RMS [12,25,33]).
The paper is organised as follows: Section 2 introduces
the problem and details of the preliminary transforma- We now re-express system (1)-(2) into a form facilitat-
tions. Sections 3 and 4 then present the observer for robust ing further analysis using the following proposition.
fault reconstruction in the case where the fault is present Proposition 1: Suppose Assumption 1 is satisfied.
in the nonlinear functions of the output equation, and in Then system (1)-(2) can be re-expressed as
the case where the nonlinear functions in the output equa-    
tions do not contain the fault, respectively. Section 4 also Ir 0 H
ẋ = Ax + Bu + M f + x1 h (T x̆) + Qξ , (3)
makes some general remarks discussing the observers, and 0 0 Hx2
| {z } | {z }
outlines a set of design procedures for the proposed fault E Hx
reconstruction scheme. Section 5 then verifies the perfor-  
H
y = 0 Ip x + Du + N f + y1 h (T x̆) + Rξ ,
 
mance of the scheme using two simulation examples, and (4)
Hy2
Section 6 draws some conclusions.
| {z }
C | {z }
Notation: The root-mean-square (RMS) of a signal w Hy
 Descriptor Observers for Robust Fault Reconstruction in a Class of Nonlinear Descriptor Systems 699

where Hx2 ∈ R(n−r)×b , Hy1 ∈ R(r−n+p)×b , Hy2 ∈ in the output equation (i.e., Hy h (T x̆)) contains f . The fol-
T T
lowing assumption is made in this section.
 
R(n−r)×b , x̆ = xT f T , and x = xaT xbT , xa ∈ Rr ,
while T and y will be defined in the following proof. Assumption 5: The rate of the fault f has a finite RMS,
i.e., R f˙ < ∞.


 T
Proof: Define a non-singular matrix Ta = NC CoT , Remark 3: Assumption 5 implies that f is energy-
where NC ∈ Rn×(n−p) spans the right-nullspace of Co , bounded [19], and is necessary to guarantee the perfor-
i.e.,
 Co NC = 0 and rank (NC ) = n − p. Let ETa−1 , mance of the observer in this section. This assumption has
Ea1 Ea2 , where Ea2 ∈ Rn×p . Assumption 1 implies also been applied in various other works [7,13,26,28,34],
rank (Ea1 ) = n − p, and therefore by applying QR de- and is not overly restrictive as a wide range of functions
 T
composition on Ea1 we obtain Tb Ea1 = In−p 0 , encountered in practical systems satisfy this assumption
where Tb ∈ Rn×n is non-singular. Then let Tb Eo Ta−1 ,
 (e.g., periodic and sinusoidal signals, and piecewise con-
In−p Eb1 tinuous polynomial functions) [10].
, where Eb2 ∈ R p×p . Since rank (Eo ) = r,
0 Eb2 Consider the following observer for system (3)-(4):
it can be seen that rank (Eb2 ) = r − n + p, and hence
E x̄˙ˆo = Ax̄ˆo + Bu + M fˆ¯ + Hx h T x̄ˆ + L̄1 ēy ,


apply singular value  decomposition (SVD) on Eb2 to (5)
−1 Ir−n+p 0 ˙ˆ¯
get Tc11 Eb2 Tc21 = , where Tc11 , Tc21 ∈ R p×p f = L̄2 ēy , (6)
0 0
are non-singular.
 Next, define two non-singular  matri- ˆ
ēy = Cx̄ˆo + Du + N f¯ + Hy h T x̄ˆ − y,


(7)
In−p 0 In−p 0
ces Tc1 = and Tc2 = , and let
0 Tc11 0 T x̄ˆa l r
 
  c21 ˆx̄o
 
In−p Ec1 Ec2 where x̄ˆ = ˆ¯ = x̄ˆb  l n − r , and L̄1 ∈ Rn×p and
Tc1 Tb2 Eo (Tc2 Ta )−1 ,  0 Ir−n+p 0 . Then define f
fˆ¯ l q
0 0 0 L̄2 ∈ R q×p
are design matrices, with L̄1 partitioned as
 
In−p −Ec1 −Ec2  
Td−1 =  0 Ir−n+p 0 . L̄
L̄1 = 111
L̄112
, (8)
0 0 In−r L̄121 L̄122
Therefore, pre-multiply both sides of (1) with Tc1 Tb , trans-
form xo such that xo 7→ Td Tc2 Ta xo = x (i.e., post-multiply where L̄111 ∈ Rr×(r−n+p) and L̄122 ∈ R(n−r)×(n−r) . To en-
Eo , Ao ,Co with (Td Tc2 Ta )−1 ), transform yo such that yo 7→ sure no implicit

solutions of x̄ˆb are present, set L̄121 = 0 if

Tc21 yo = y (i.e.,pre-multiply both Hy1 h T x̄ contains x̄ˆb ; similarly, set L̄122 = 0 if Hy2 h T x̄ˆ
ˆ
 sides of (2) with Tc21 ), ˆ¯
(Td Tc2 Ta )−1 0 contains x̄ˆb . Next, ˆ

define the errors ēx = x̄o −x, ē f = f − f ,
and let T = To to obtain the structures
0 Iq and h̄ = h T x̄ˆ − h (T x̆). Using (3)-(4) and (5)-(7) yields
in (3)-(4), completing the proof. 
E ē˙x = Aēx + M ē f + Hx h̄ − Qξ + L̄1 ēy , (9)
Assumption 3: The nonlinear function h (·) is Lips- ē˙ f = L̄2 ēy − f˙. (10)
chitz with respect to T x̆ in Rb , i.e., for all T x̆1 , T x̆2 ∈
ēy = Cēx + N ē f + Hy h̄ − Rξ . (11)
Rb , there exists a known positive constant l f satisfying
kh (T x̆1 ) − h (T x̆2 )k ≤ l f kT x̆1 − T x̆2 k. Then define the error ē = x̄ˆ − x̆, and by substituting (11)
Assumption 4: The term Hx2 h (T x̆) does not contain into (9)-(10), we obtain the following error system:
elements of xb .      
E 0 ˙ A + L̄1C M + L̄1 N H + L̄1 Hy
Remark 2: Assumption 3 is typically made for nonlin- ē = ē + x h̄
0 Iq L̄2C L̄2 N L̄2 Hy
ear observers [12,16,23,24,29], and will be used to design | {z } | {z } | {z }
the observers in the succeeding sections. Assumption 4 on Ē Ā H̄
  
the other hand ensures (4) does not have implicit solutions − (Q + L̄1 R) 0 ξ
+ . (12)
for xb , and has been utilised in various works for descrip- −L̄2 R −Iq f˙
tor systems [3,5,12,13,15,17,20-23,26,29,33].
| {z } |{z}
Q̄ ξ¯
We now present two observers to reconstruct f . The ap-
plicability of each observer would depend on whether f is It can be seen that observer (5)-(7) reconstructs f
present in Hy h (T x̆). through fˆ¯, but from (12), ē f (and therefore, fˆ¯) is corrupted
by ξ¯ . Additionally, error system (12) is a nonlinear de-
3. THE FIRST OBSERVER FOR ROBUST FAULT scriptor system, and therefore observer (5)-(7) also needs
RECONSTRUCTION to be designed to be solvable. Hence the objective now is
to design observer (5)-(7) such that it is solvable, and the
We first consider the case where the nonlinear function effects of ξ¯ on ē f (and thus, fˆ¯) are bounded.
700 Joseph Chang Lun Chan, Wen-Shyan Chua, Tae H. Lee, and Chee Pin Tan

 
We first quantify the performance of observer (5)-(7) x̌ l q+r
Re-express (17) using x̌ 7→ K x̄ = 1 −1 ˆ
:
using the following definitions and lemma: x̌2 l n − r


Definition 1 [2]: The pair Ě, Ǎ is said to be
x̌˙1 = Ā1 x̌1 + J1 Ȟ (x̌, u) , (18)
i) regular, if there exists some s ∈ C such that
0 = x̌2 + J2 Ȟ (x̌, u) . (19)
sĚ − Ǎ 6= 0,



ii) impulse-free, if deg sĚ − Ǎ = rank Ě ,
   
x̌ x̌
Define x̌a = 1 and x̌b = 1 , and let x̄ˆa = K x̌a and
iii) stable, if all roots of sĚ − Ǎ 6= 0 belong to C− , and x̌2a x̌2b
iv) admissible, if it is regular, impulse-free, and stable. x̄ˆb = K x̌b . From (17) and Assumption 3, we obtain


Lemma 1 [33]: The pair Ě, Ǎ is said to be admissible
J2 Ȟ (x̌a , u) − J2 Ȟ (x̌b , u)
if and only if there exists a matrix X̌ (with the appropriate



dimensions) satisfying Ě T X̌ = X̌ T Ě ≥ 0, ǍT X̌ + X̌ T Ǎ < 0. ≤ kJ2 k H̄ h T x̄ˆa − h T x̄ˆb
Definition 2 [10]: Consider the two signals zi : ≤ l f kJ2 k kH̄k kT K (x̌a − x̌b )k
[0, ∞) 7→ R and zo : [0, ∞) 7→ R, and let Z be the set of ≤ l f τ kJ2 k kH̄k kK2 k kx̌2a − x̌2b k . (20)
all admissible signals zi . 
The RMS gain from zi tozo is
LMI (13) implies kJ2 k ≤ 1 [20] and also l 2f K2T K2 < In−r ,
R(zo )
defined as ε (zi , zo ) = sup R(zi ) : R (zi ) > 0, zi ∈ Z . which further implies kK2 k ≤ l √11+l , where lo is a suffi-
f o

Theorem 1 presents the main results for this section. ciently small positive scalar [33]. The Schur complement
of LMI (14) then implies H̄ T H̄ ≤ τ12 Ib , which in turn im-
Theorem 1: Suppose Assumptions 2, 3, and 5, and 2
plies λmax H̄ T H̄ = kH̄k ≤ τ12 , i.e., kH̄k ≤ τ1 . Therefore


Proposition 1 are satisfied, and that for the given positive
scalar l f there exist any matrices L̄1 and L̄2 , and positive (20) becomes
scalars η̄ and ν̄ satisfying 1
J2 Ȟ (x̌a , u) − J2 Ȟ (x̌b , u) ≤ √ kx̌2a − x̌2b k .
Ā + ĀT + l 2f + 1 In+q < 0, 1 + lo


(13)
  (21)
−Ib ∗
≤ 0, (14)
H̄ − τ12 In+q Using the fixed-point principle [3], (21) implies that there

Ā + ĀT + η̄l 2f T T T ∗ ∗
 exists a unique solution to (19), thereby guaranteeing the
 H̄ T −η̄Ib ∗  ≤ 0, (15) solvability of x̌2 . We now show the solvability of x̌1 . The
T
Q̄C 0 −ν̄I2q+h Picard-Lindelöf theorem states that for (18) (and there-
fore x̌1 ) to be solvable, Ȟ needs to be Lipschitz with
T T
   
where τ = kT k , Q̄C = Q̄ C̄T , and C̄ = 0q×n Iq .
 T 
respect to x̌1 [3]. Hence define x̌c = x̌c1 x̌c2 , x̌d =
Then observer (5)-(7) is solvable, and the RMS gain from  T T T

x̌d1 x̌d2 , x̄ˆc = K x̌c , x̄ˆd = K x̌d , and ∆Ȟ = Ȟ (x̌c , u) −
ξ¯ to ē f is bounded by ν̄.
Ȟ (x̌d , u), and it can be seen that
Proof: The proof is divided into two parts. Firstly, we



∆Ȟ = H̄ h T x̄ˆc − h T x̄ˆd
prove that satisfying LMIs (13)-(14) results in the solv-
ability of observer (5)-(7). Next, the satisfaction of LMI ≤ l f τ kH̄k kK (x̌c − x̌d )k
(15) is shown to guarantee that the RMS gain from ξ¯ to ē f ≤ l f kK1 (x̌c1 − x̌d1 ) + K2 (x̌c2 − x̌d2 )k
is bounded by ν̄. ≤ l f (kK1 k kx̌c1 − x̌d1 k + kK2 k kx̌c2 − x̌d2 k) .
Proposition 1 re-expresses system (1)-(2) to have the (22)
structures in (3)-(4). From (3) and (12), it can be seen
that Ē = Ē T ≥ 0. LMI
(13) implies Ā + ĀT < 0, and thus From (19), we get x̌c2 = −J2 Ȟ (x̌c , u) and x̌d2 =
from Lemma

1, Ē, Ā is admissible. The admissibility −J2 Ȟ (x̌d , u), and using kJ2 k ≤ 1 yields kx̌c2 − x̌d2 k ≤
of Ē, Ā then implies there exist two non-singular matri- kJ2 k ∆Ȟ ≤ ∆Ȟ . Since LMI (13) implies kK2 k ≤
T
√1
(n+q)×(n+q)

 J ∈ R
ces = J1T J2T and K ∈ R(n+q)×(n+q) = l 1+l
f o
, then (22) can be rearranged to give
K1 K2 such that √
l f kK1 k 1 + lo
∆Ȟ ≤ √ kx̌c1 − x̌d1 k , (23)
   
I 0 Ā 0
J ĒK = q+r , J ĀK = 1 , (16) 1 + lo − 1
0 0 0 In−r
that is, Ȟ is Lipschitz with respect to x̌1 , and therefore (17)
where J1 ∈ R(q+r)×(n+q) and K1 ∈ R(n+q)×(q+r) [20]. Next, is also solvable. Thus if LMIs (13)-(14) are satisfied, then
it can be seen that observer (5)-(7) can be re-expressed x̌ (and therefore x̄ˆ in observer (5)-(7)) is solvable, com-
using (12) as pleting the first part of the proof.
 
B + L̄1 D In the following final part of the proof, we show how LMI
Ē x̄˙ˆ = Āx̄ˆ + 1


u + H̄h T x̄ˆ . (17) (15) guarantees the RMS gain from ξ¯ to ē f is bounded
L̄2 D
 Descriptor Observers for Robust Fault Reconstruction in a Class of Nonlinear Descriptor Systems 701

by ν̄. Define the Lyapunov candidate function W̄ = ēT Ē ē, Remark 5: Assumption 6 is needed for further re-
and differentiate it with respect to time to get expressing system (3)-(4). It is also not overly restrictive,
and has commonly been used in [6,7,13,35].
W̄˙ = ēT Ā + ĀT ē + 2ēT H̄ h̄ + 2ēT Q̄ξ¯


We now re-express system (3)-(4) to facilitate further
Ā + ĀT ∗ ∗
 
analysis in this section with the following proposition.
= φ̄ T  H̄ T 0 ∗ φ̄ , (24) Proposition 2: Suppose Assumption 6 is satisfied.
Q̄T 0 0 Then system (3)-(4) can be re-expressed as
T
ξ¯ T
        
where φ̄ = ēT h̄T . Assumption 3 implies there E 0 A M B Hx
   0 0 x̆˙ = C1 −Iq  x̆ + D1  u + H̃y1  h (T x̆)
exists η̄ ∈ R+ such that η̄ (T ē)T (T ē) l 2f − h̄T h̄ ≥ 0,
 2 T  0 0 C2 0 D2 H̃y2
η̄l f T T ∗ ∗    
Q 0 0
i.e., W̄1 = φ̄ T  0 −η̄Ib ∗ φ̄ ≥ 0. Then apply
+ R1  ξ + −Iq 0  ỹ, (28)
0 0 0
Schur’s complement on LMI (15) to get R2 0 −Ip−q
where ỹ will be defined later in the following proof.
Ā + ĀT + η̄l 2f T T T + ν̄1 C̄T C̄ ∗
 
∗
Ψ̄ =  H̄ T −η̄Ib ∗ , Proof: Assumption 6 implies there exists a non-
T
Q̄T 0 −ν̄Iq+h

singular matrix Te ∈ R p×p such that Te N = −Iq 0 .
(25) Therefore, transform y such that y 7→ Te y = ỹ (i.e., pre-
multiply both sides of (4) with Te ) to obtain the structures
where Ψ̄ ≤ 0. Hence satisfying LMI (15) implies in (28). Thus, the proof is complete. 
1 Next, consider the following observer:
φ̄ T Ψ̄φ̄ = W̄˙ + W̄1 + ēTf ē f − ν̄ ξ¯ T ξ¯
ν̄
E x̃˙ˆo = A + L̃1C1 + L̃2C2 x̃ˆo + −L̃1 −L̃2 ỹ

 
˙ 1 T
≤ W̄ + ē f ē f − ν̄ ξ¯ T ξ¯ ≤ 0. (26)
ν̄ + B + L̃1 D1 + L̃2 D2 u + M − L̃1 fˆ˜

Then integrate (26) from time t = 0 to t = T , divide by T , + Hx + L̃1 H̃y1 + L̃2 H̃y2 h T x̃ˆ ,



(29)
and use W (T ) ≥ 0 to obtain ˆf˜ = C + L̃ C
x̃ˆ + H̃ + L̃ H̃
h T x̃ˆ

1 3 2 o y1 3 y2
W (0) W (T ) −W (0)

+ D1 + L̃3 D2 u + −Iq −L̃3 ỹ,

(30)
− ≤
T T
ν̄ T ¯
Z
2 1
Z T " # x̃ˆ  l r
a
≤ ξ (t) dt − kē f (t)k2 dt. x̃ˆo
T 0 ν̄T 0 where x̃ˆ = ˆ˜ = x̃ˆb  l n − r , and L̃1 ∈ Rn×q , L̃2 ∈
f
(27) fˆ˜ l q
Rn×(p−q) , and L̃3 ∈ Rq×(p−q) are design parameters. The
Assumptions 2 and 5 then imply R ξ¯ is bounded. Hence


matrices L̃1 ∈ Rn×q and L̃2 ∈ Rn×(p−q) are partitioned as
R(ē )
let T → ∞ and take square roots to obtain R ξ¯f ≤ ν̄. Using
( )  

L̃11 L̃21

l r
Definition 2, the RMS gain from ξ¯ to ē f is bounded by ν̄, L̃1 L̃2 =
L̃12 L̃22 l n−r
. (31)
thus completing the proof. 
Similar to Section 3, to ensure no implicit solutions of
Remark 4: In the absence of disturbances and if the x̃ˆb are present, set L̃12 = 0 if
H̃y1 h T x̃ˆ contains x̃ˆb ; simi-


fault is slowly varying (i.e., ξ = 0, f˙ ≈ 0), then (26) larly, set L̃22 = 0 if H̃y2 h T x̃ˆ contains x̃ˆb . Observer (29)-
becomes W̄˙ ≤ 0. From W̄ = ēT Ē ē, this implies that as (30) can be re-expressed as
t → 0, ē f → 0, and therefore fˆ¯ → f . Hence if disturbances    
were absent and the fault varies slowly, observer (5)-(7) E 0 ˙ˆ A + L̃1C1 + L̃2C2 M − L̃1 ˆ
x̃ = x̃
can asymptotically reconstruct f . 0 0 C1 + L̃3C2 −Iq
| {z } | {z }
Ẽ Ã
4. THE SECOND OBSERVER FOR ROBUST
   
B + L̃1 D1 + L̃2 D2 −L̃1 −L̃2
FAULT RECONSTRUCTION + u+ ỹ
D1 + L̃3 D2 −Iq −L̃3
| {z } | {z }
In this section, we consider the case where the nonlinear B̃ L̃
 
function in the output equation (i.e., Hy h (T x̆)) does not H + L̃1 H̃y1 + L̃2 H̃y2
+ x h T x̃ˆ .


(32)
contain f . The following is assumed in this section: H̃y1 + L̃3 H̃y2
Assumption 6: rank(N) = q.
| {z }
H̃
702 Joseph Chang Lun Chan, Wen-Shyan Chua, Tae H. Lee, and Chee Pin Tan

  T
In L̃1 L̃2
 T
Let L̃o = , and pre-multiply both sides of e h̃T ξT . Assumption 3 then implies there exists
0 Iq L̃3  2 T
η̃l f T T ∗ ∗

(28) with L̃o to obtain η̃ ∈ R+ such that W̃1 = φ̃ T  0 −η̃Ib ∗ φ̃ ≥ 0.
  0 0 0
˙ Q+L̃1 R1 +L̃2 R2
Ẽ x̆ = Ãx̆+B̃u+H̃h (T x̆)+L̃ỹ+ ξ. By applying similar reasoning as in (25)-(27) and using
R1 +L̃3 R2 R(ẽ )
Assumption 2, LMI (37) implies R(ξf) ≤ ν̃. Then from
(33)
Definition 2, the RMS gain from ξ to ẽ f is bounded by ν̃,
Define the errors ẽ = x̃ˆ − x̆, ẽ f = fˆ˜ − f , and h̃ = h T x̃ˆ −

thereby completing the proof. 
h (T x̆). Deducting (33) from (32) results in Remark 6: In the absence of disturbances (i.e., ξ = 0),

− Q + L̃1 R1 + L̃2
R2

 then by using a similar reasoning as in (25)-(27), we ob-
˙
Ẽ ẽ = Ãẽ + H̃ h̃ + ξ. (34) tain W̃˙ ≤ 0, and since W̃ = ẽT Ẽ ẽ, then t → 0, ẽ f → 0,
− R1 + L̃3 R2
| {z } resulting in fˆ˜ → f . Therefore, if disturbances were ab-
Q̃
sent, observer (29)-(30) can asymptotically reconstruct the
Similar to the previous section, observer (29)-(30) re- fault.
constructs f through fˆ˜, but from (34), ξ corrupts ẽ f (and We now make some remarks discussing the two ob-
therefore, fˆ˜). Additionally, error system (34) is a nonlinear servers in Sections 3 and 4.
descriptor system, and thus observer (29)-(30) also needs Remark 7: As pointed out in [33], the feasibility of de-
to be designed to be solvable. We now design observer signing state-space observers for system (1)-(2) requires
(29)-(30) such that it is solvable, and the effects of ξ on the rates of the input, disturbance, and nonlinearities (i.e.,
ẽ f (and hence, fˆ˜) are bounded.

u̇, ξ˙ , dh(Tdto x̆o ) ) to be bounded. Assuming that R ξ˙ <
The design goals for observer (29)-(30) are similar to ∞ would reduce the practicality of the state-space ob-
that of observer (5)-(7). In fact, observer (29)-(30) in the servers since ξ is generally an unknown but bounded sig-
form of (32) and its error system (34) have similar struc- nal (which could include step signals) [15]. Additionally,
tures as those in (17) and (12), respectively. Theorem 1 requiring R (u̇) < ∞ would restrict the controller design
can therefore be adapted into the following. by excluding step control inputs, limiting the applicabil-
Theorem 2: Suppose Assumptions 2-3, and Proposi- ity of the state-space observer. The observers proposed in
tions 1-2 are satisfied, and that for a given positive scalar this paper however only require ξ to be bounded (in As-
l f there exist any matrices L̃1 , L̃2 , and L̃3 , and positive sumption 2), while the observer in Section 3 also requires
scalars η̃ and ν̃ satisfying the rate of the fault (i.e., f˙) to also be bounded (in As-
sumption 5), with no additional restrictions required for
à + ÃT + l 2f + 1 In+q < 0,


(35) ξ , ξ˙ , u, and u̇. These requirements are more relaxed than
those in state-space observers, are not restrictive, and are
 
−Ib ∗ also widely used (as mentioned in Remarks 2 and 3), thus
≤ 0, (36)
H̃ − τ12 In+q increasing the applicability of the proposed observers.
Remark 8: Theorems 1 and 2 are valid even if
à + ÃT + η̃l 2f T T T W̄ (0), W̃ (0) > 0; this implies that system (1)-(2) and its
 
∗ ∗
 H̃ T −η̃Ib ∗  ≤ 0, (37) respective observer do not need to have identical initial
T
Q̃C 0 −ν̃Iq+h conditions. If W̄ (0) = 0 (i.e., both the system and observer
(5)-(7) have identical initial conditions), then the L2 gain
from ξ¯ to ē f is bounded by ν̄ instead. This is also true
   
where Q̃C = Q̃ C̃T , C̃ = 0q×n Iq , and τ has been
defined in Theorem 1. Then observer (29)-(30) is solvable, when W̃ (0) = 0, (i.e., both the system and observer (29)-
and the RMS gain from ξ to ẽ f is bounded by ν̃. (30) have identical initial conditions); the L2 gain from ξ
to ẽ f is bounded by ν̃ in this case.
Proof: Propositions 1-2 allow system (1)-(2) to be re-
expressed with the structures in (28), which in turn can be Remark 9: Both observers (5)-(7) and (29)-(30) are
re-expressed as (33). Then observer (29)-(30) can be re- also applicable when rank (Eo ) = n (e.g., Eo = In ), i.e.,
expressed as (32), and error system (34) can be obtained. they can also be applied on state-space systems. This
Using similar arguments as in (16)-(23), it can be seen that makes the presented work more general than those appli-
LMIs (35)-(36) imply that observer (32) (and therefore, cable to only state-space systems [10,11,27-29,31].
(29)-(30)) is solvable. A summarised set of design procedures for the proposed
Next, define the Lyapunov candidate function W̃ = ẽT Ẽ ẽ, fault reconstruction scheme is given in Table 1.
and differentiating

it with respect to time yields W̃˙ =
ẽ Ã + Ã ẽ + 2ẽ H̃ h̃ + 2ẽT Q̃ξ = φ̃ T Ψ̃φ̃ , where φ̃ =
T T T
 Descriptor Observers for Robust Fault Reconstruction in a Class of Nonlinear Descriptor Systems 703

Table 1. Design procedure. seen that Assumptions 1-2 are satisfied.

1: Check that Assumptions 1-2 are valid for system (1)-(2). If 2-3: The matrices Ta , Tb, Tc1 , Tc2 , and Td are found to
not, stop as the scheme is not applicable. 0 1 0
2: Calculate Ta , Tb , Tc1 , Tc2 , and Td to obtain the system in the be Ta = Tb = 1 0 0 and Tc1 = Tc2 = Td = I3 . Let
form of (3)-(4). Verify if Assumptions 3-4 hold. If they do
not, the scheme cannot be applied. 0 0 1
 T  T
3: Obtain the smallest value of l f satisfying Assumption 3. x = Td Tc2 Ta xo = x1 x2 x3 and x̆ = xT f . We
 
4: Determine from (4) if Hy h (T x̆) contains f .   0.3 sin (x3 )
then obtain T = 02 I2 , h (T x̆) = , and
5: if Hy h (T x̆) contains f , then observer (5)-(7) in Section 3 is 0.3 cos ( f )
utilised: Hx2 h (T x̆) = 0.3 cos ( f ). Hence, Assumptions 3-4 are also
6: Ensure that Assumption 5 is satisfied. If not, the scheme
is inapplicable. satisfied, with l f = 0.4243.


7: If Hy1 h T x̄ˆ contains x̄ˆb , set L̄121 = 0 in (8). 4: The term Hy h (T x̆) is given by Hy h (T x̆) =


8: Similarly, if Hy2 h T x̄ˆ contains x̄ˆb , set L̄122 = 0 in (8).  T
0 0.3 cos ( f ) , i.e., Hy h (T x̆) contains f . Therefore,
9: Use a LMI solver on LMIs (13)-(15) to find L̄1 , L̄2 , η̄, the design of the observer would follow the procedure in
and ν̄.
10: Reconstruct f using (6). lines 6-10.
11: else (i.e., if Hy h (T x̆) does not contain f ) observer (29)-(30) 6: The fault f is assumed to be a combination of ramps
in Section 4 is used: and sinusoids in this case, satisfying Assumption 5.
12: Verify if Assumption 6 is satisfied. If not, the scheme is
not applicable. h iT
13: Find Te to get the system with the structures in (28).
7-8: Let x̄ˆ = x̄ˆ1 x̄ˆ2 x̄ˆ3 fˆ¯ , and from (7), x̄ˆb = x̄ˆ3 .

h  iT
If H̃y1 h T x̃ˆ contains x̃ˆb , set L̃12 = 0 in (31).


14: Since Hy h T x̄ˆ = 0 0.3 cos f¯ˆ does not contain x̄ˆ3 ,
Likewise, if H̃y2 h T x̃ˆ contains x̃ˆb , set L̃22 = 0 in (31).


15:
L̄1 does not need to be assigned any special structure.
16: Solve LMIs (35)-(37) to obtain L̃1 , L̃2 , L̃3 , η̃, and ν̃.
17: Reconstruct f using (30). 9: Applying the SeDuMi solver for YALMIP in MATLAB
18: end if on LMIs (13)-(15) yields
 
−0.9919 0.0275
5. SIMULATION EXAMPLES L̄1 =  1.522 0  , η̄ = 4.024,
In this section, we perform two simulation examples to 0.1000 −0.6315
 
demonstrate the effectiveness of our proposed observers. L̄2 = −1.784 0.5833 , ν̄ = 2.457. (39)

5.1. Example 1 10: Finally, f is reconstructed using f¯ˆ, which is given by

˙ 
To first demonstrate the efficacy of the observer pro- fˆ¯ = −1.784 0.5833 ēy .

posed in Section 3, consider a system in the form of (1)
The design of the observer is complete. To demonstrate
and (2) with the matrices
    the performance of the scheme, we simulate two scenar-
  −6 0 0 1 ios. The first scenario is disturbance-free, and the initial
I 0
Eo = 2 , Ao =  1 −5 0  , Bo = 0 , condition of the system and the fault are set as
0 0
0 −1 −4 1
      xo (0) = {1, −1, 1.046},
0 1 0 1
Mo = 1 , Hxo = 0 0 , Qo = 1 , f¯n = 0.9 − 17 (r (t − 10) − r (t − 10.1))
0 0 1 0 + 10 (r (t − 20) − r (t − 20.1)) . (40)
       
1 0 0 0 1 1
Co = , Do = , No = , Ro = , In the second scenario, the system initial condition,
0 0 1 1 0 1
    fault, and disturbance are set as
0 0 0.3 sin (xo3 )
Hyo = , h (To x̆o ) = , (38)
0 1 0.3 cos ( f ) xo (0) = {1, −1, 1.063},
f¯t = 0.7 sin(0.8t + 7π/6) + 0.03(r(t − 5) − r(t − 15))
where the states, input, fault, and disturbance are labelled
 T
as xo = xo1 xo2 xo3 , u, f , and ξ , respectively, while − 0.08(r(t − 23) − r(t − 28)) − 0.2,
 T  
x̆o = xoT f and To = 02 I2 . We now design the fault ξ = S (0.1, 0.4) + 0.1 sin (0.8t + π/4) . (41)
reconstruction scheme according to the procedure in Ta-
In both scenarios, the initial conditions of the observer
ble 1.
T are set to be zero.
Fig. 1 shows the fault f¯n and its reconstruction using

1: From (38), we obtain rank EoT CoT = n = 3. The
disturbance ξ is also assumed to be bounded. It can be the observer in Section 3 for the disturbance-free scenario.
704 Joseph Chang Lun Chan, Wen-Shyan Chua, Tae H. Lee, and Chee Pin Tan

1 5
20
10
0.5 4
0
0 0.0001
3
0
2
-0.5
1

-1
0 5 10 15 20 25 30 0
0 5 10 15 20 25 30
Fig. 1. The fault (solid) and its reconstruction using the
observer in Section 3 (dashed) for the disturbance- Fig. 3. The empirical RMS gain ν̄E (solid) and its bound
free scenario in Example 1. ν̄ (dashed) in Example 1.

1
Therefore, the design for this example would follow lines
12-17.
0.5  T
12: We obtain N = 1 0 , which is full-column rank,
satisfying Assumption 6.
0  
−1 0
13: The matrix Te is calculated to be Te = .
0 1
-0.5
h iT
14-15: Let x̃ˆ = x̃ˆ1 x̃ˆ2 x̃ˆ3 fˆ˜ , and from (30), x̃ˆb = x̃ˆ3 .
-1
T
Since H̃y1 h T x̃ˆ = 0.3 sin x̃ˆ3 0 contains x̃ˆ3 , set L̃12


0 5 10 15 20 25 30
as L̃12 = 0.
Fig. 2. The fault (solid) and its reconstruction using the
observer in Section 3 (dashed) for the scenario with 16: Using the SeDuMi solver for YALMIP in MATLAB
non-zero disturbances in Example 1. on LMIs (35)-(37) yields
 T  T
L̃1 = 0 1 0 , L̃2 = −0.1394 0.1044 0.6668 ,
The fault reconstruction faithfully follows f¯n after some L̃3 = 0.5881, η̃ = 3.345, ν̃ = 2.297. (42)
initial transients, as stipulated in Remark 4.
Fig. 2 then shows the performance of the observer in 17: Lastly, f would be reconstructed using fˆ˜, which is
Section 3 for the scenario with non-zero disturbances. The calculated using
fault reconstruction is visibly affected by disturbances in
fˆ˜ = 0 −1 0.5881 x̃ˆ + 0.5881u − 0.3 sin x̃ˆ3
 

this case. To quantify the effect of the disturbances on the
fault reconstruction, we use the empirical RMS gain (at  
− 1 0.5881 ỹ. (43)
time instant T ) ν̄E , which for this scenario is defined as
h iT
ν̄E = RT ξ¯f , where ē f = fˆ¯t − f¯t and ξ¯ = ξ f˙¯t .
R (ē )
The design of the observer is now complete. Similarly,
T( )

Fig. 3 shows ν̄E against time. Initially, ν̄E is large due to demonstrate the robustness of the scheme, we simulate
to the mismatched initial conditions of the system and the two scenarios: the first scenario is disturbance-free, with
observer. The value of ν̄E however quickly falls below ν̄, system initial conditions and fault set as
i.e., ν̄E is ultimately bounded by ν̄, as stated in Theorem
xo (0) = {−1, 2, 0.2884},
1. Thus, the performance of the observer in Section 3 has
been conclusively demonstrated. f˜n = −0.2 + S (0.4, 0.4) + 0.5 sin (1.2t + 5π/3) .
Next, to showcase the effectiveness of the observer in (44)
Section 4, we again consider system (1)-(2) withthe ma-
The second scenario has system initial conditions, fault
1 0
trices in (38), but we now set Hyo as Hyo = . and disturbances set as
0 0
1-3: These steps are the same as in the previous case. xo (0) = {−1, 2, 0.2884},
4: The term Hy h (T x̆) is now given by Hy h (T x̆) = f˜t = −0.2 + S (0.4, 0.4) + 0.5 sin (1.2t + 5π/3) ,
 T
0.3 sin (x3 ) 0 , i.e., Hy h (T x̆) does not contain f . ξ = 0.1 + S (0.1, 0.3) + 0.2 sin (0.8t + π/4) . (45)
 Descriptor Observers for Robust Fault Reconstruction in a Class of Nonlinear Descriptor Systems 705

1 6

0 4

-1 2

-2 0
0 2 4 6 8 10 0 2 4 6 8 10

Fig. 4. The fault (solid) and its reconstruction using the Fig. 6. The empirical RMS gain ν̃E (solid) and its bound
observer in Section 4 (dashed) for the disturbance- ν̃ (dashed) in Example 1.
free scenario in Example 1.

for state-space systems, and therefore cannot be ap-

1
plied to descriptor systems.
2) The schemes in [4,6-9,14,18-20,24,34,35] require the
0 system to be linear, and are hence inapplicable.
3) The output equation is assumed to be free of faults,
disturbances, and/or nonlinearities in [4-7,9,11,
-1 13-15,18-23,26-32,34,35,38], which means these
schemes are not applicable.
4) The work in [1,11,12,17,19,25,29,33,37,38] require
-2
0 2 4 6 8 10 additional assumptions on the dynamics of the dis-
turbances.
Fig. 5. The fault (solid) and its reconstruction using the 5) Faults are assumed not to enter through nonlin-
observer in Section 4 (dashed) for the scenario with ear functions in [4,6,7,9,13,15,25,26,34,35], render-
non-zero disturbances in Example 1. ing them inapplicable.
This conclusively demonstrates the greater applicability of
In both scenarios, the observer is again set to have zero our proposed scheme over existing observers in the litera-
initial conditions. ture.
Fig. 4 shows the performance of the observer in Sec-
tion 4 for the disturbance-free scenario. After some initial 5.2. Example 2
transients due to the mismatch between the initial condi- In this subsection, we consider the petroleum cracking
tions of the system and observer, the fault f˜n is accurately process [36], which is a complex, widely-studied process
reconstructed, as per Remark 6. relating regulated variables (such as regenerator temper-
Fig. 5 shows the fault and its reconstruction using the ature, valve position, and blower capacity) with adminis-
observer in Section 4 for the scenario with non-zero dis- tration states (e.g., business benefits, administration, and
turbances. It can be seen that the reconstruction is now policy actions) [37], where the nonlinear functions could
visibly affected by disturbances. We similarly utilise the also include faults [38]. A simplified model of the process
empirical RMS gain to quantify the effect of the distur- is given in [38], which has a similar structure as that of
bances on the fault reconstruction, which is defined in this system (1)-(2). Suppose also that only the rate of the first
scenario as ν̃E = RTT (ξf) , ẽ f = fˆ˜t − f˜t .
R (ẽ )
state (i.e., ẋo1 ), and the second state xo2 are available for
Fig. 6 shows ν̃E against time. The initial value of ν̃E measurement.
is again large due to the mismatched initial conditions of To show the efficacy of the observer in Section 3, sup-
the observer and system, but ν̃E then goes below ν̃, which pose the nonlinear disturbance affecting the system is
shows that ν̃E is ultimately bounded by ν̃, as per Theorem given by
2. Thus, the performance of the proposed scheme has been  
demonstrated. f + 0.5 cos( f ) + 0.5ξ
ω(t) = , (46)
Remark 10: Existing results in the literature cannot be f +ξ
applied to system (38) for the following reasons:
where f are the fault signals of interest, and ξ represents
1) The observers in [6-14,24,27-31,35] are exclusively the uncertainties inherent in the system. By selecting the
706 Joseph Chang Lun Chan, Wen-Shyan Chua, Tae H. Lee, and Chee Pin Tan

drift matrix Ao accordingly, we obtain the system in the 0.6

form of (1)-(2) with the matrices
      0.4
1 0 −1 1 0
Eo = , Ao = , Bo = , 0.2
0 0 0.5 −1 1
     
1 1 0.5 0
Mo = , Hxo = , Qo = ,
1 0 1 -0.2
       
−1 1 0 1 0.5
Co = , Do = , No = , Ro = , -0.4
0.5 0 1 1 1
 
1 0 5 10 15 20 25 30
Hyo = , h (To x̆o ) = 0.5 cos ( f ) , (47)
0
 T Fig. 7. The fault (solid) and its reconstruction using the
where the states and input are labelled as xo = xo1 xo2 observer in Section 3 (dashed) for the disturbance-
 T free scenario in Example 2.
and
 u, respectively,
 while x̆o = xoT f and To =
0 0 1 . We now design the fault reconstruction
scheme according to the procedure in Table 1.
0.8
 T
1: From (47), we obtain rank EoT CoT = n = 2. The
disturbance ξ is also assumed to be bounded. It can be 0.6
seen that Assumptions 1-2 are satisfied.
0.4
2-3: The matrices
 Ta , Tb , Tc1 , Tc2 , and Td are found

−1 1 −0.5 0 0.2
to be Ta = , Tb = Td = I2 , Tc1 =
0.5 0 0 1
  0
0 −1  T
and Tc2 = . Let x = Td Tc2 Ta xo = x1 x2 and
1 0 0 5 10 15 20 25 30
 T T  
x̆ = x f . We then obtain T = 0 0 1 , h (T x̆) =
0.5 cos ( f ), and Hx2 h (T x̆) = 0. Hence, Assumptions 3-4 Fig. 8. The fault (solid) and its reconstruction using the
are also satisfied, with l f = 0.5. observer in Section 3 (dashed) for the scenario with
non-zero disturbances in Example 2.
4: The term Hy h (T x̆) is given by Hy h (T x̆) =
 T
0 0.5 cos ( f ) , i.e., Hy h (T x̆) contains f . Therefore,
the design of the observer would follow the procedure in fs1 = − 0.3 + 9 (r (t − 6) − r (t − 6.1))
lines 6-10.
− 10.5 (r (t − 15) − r (t − 15.1))
6: The fault f is assumed to be a combination of ramps + 6.5 (r (t − 25) − r (t − 25.1)) . (49)
and sinusoids in this case, satisfying Assumption 5.
h iT In the second scenario, the system initial condition,
7-8: Let x̄ˆ = x̄ˆ1 x̄ˆ2 fˆ¯ , and from (7), x̄ˆb = x̄ˆ2 . Since fault, and disturbance are set as

h  iT
Hy h T x̄ˆ = 0 0.5 cos fˆ¯ does not contain x̄ˆ2 , L̄1 xo (0) = {−0.8, −0.1933},
does not need to be assigned any special structure.
fs2 = 0.3 sin (1.5t + 11π/6) + 0.03r (t − 6)
9: Applying the SeDuMi solver for YALMIP in MATLAB − 0.05r (t − 15) + 0.03r (t − 20)
on LMIs (13)-(15) yields
  − 0.01r (t − 25) + 0.2,
−1.417 −0.1442 ξ = S (0.1, 0.3) + 0.05 sin (0.7t − 4π/3) . (50)
L̄1 = , η̄ = 2.273,
−0.4975 −0.1666
In both scenarios, the initial conditions of the observer
 
L̄2 = 1.047 −0.7395 , ν̄ = 2.027. (48)
are set to be zero.
10: Finally, f is reconstructed using f¯ˆ, which is given by Fig. 7 shows the fault fs1 and its reconstruction using
˙ 
fˆ¯ = 1.047 −0.7395 ēy .

the observer in Section 3 for the disturbance-free scenario.
The fault reconstruction faithfully follows fs1 after some
The design of the observer is complete. To demonstrate
initial transients following each change in the fault, as
the performance of the scheme, we simulate two scenar-
stipulated in Remark 4.
ios. The first scenario is disturbance-free, and the initial
Fig. 8 then shows the performance of the observer in
condition of the system and the fault are set as
Section 3 for the non-zero disturbance scenario. The fault
xo (0) = {−0.8, −0.4}, reconstruction is noticeably affected by disturbances, and
 Descriptor Observers for Robust Fault Reconstruction in a Class of Nonlinear Descriptor Systems 707

2.5 4 are also satisfied. We set l f = 0.3 to ensure the nonlin-

earity is captured sufficiently.
2
4: The term Hy h (T x̆) is given by Hy h (T x̆) =
 T
1.5 0.15 sin (0.4x1 ) 0 , i.e., Hy h (T x̆) does not contain
f . Therefore, the design of the observer would follow the
1
procedure in lines 12-17.
 T
0.5 12: We obtain N = 0.5 −1 , which is full-column
rank, satisfying Assumption 6.
0  
0 5 10 15 20 25 30 −2 0
13: The matrix Te is calculated to be Te = .
2 1
Fig. 9. The empirical RMS gain νe1 (solid) and its bound h iT
νE1 (dashed) in Example 2. 14-15: Let x̃ˆ = x̃ˆ1 x̃ˆ2 fˆ˜ , and from (30), x̃ˆb = x̃ˆ2 .
Since H̃y1 h T x̃ˆ = 0.15 sin (0.4x1 ) does not contain x̃ˆ2 , L̃12

and L̃22 do not need to have any special structure.

so we quantify the effect of the disturbances on the fault
reconstruction again using the empirical RMS gain (at 16: Using the SeDuMi solver for YALMIP in MATLAB
time instant T ) νe1 . In this scenario, νe1 is defined as on LMIs (35)-(37) yields
R (ē ) T
νe1 = RT ξ¯f , where ē f = fˆs2 − fs2 and ξ¯ = ξ f˙s2 .

T( )
   
L̃1 = 2.571 2.801 , L̃2 = 1.385 1.848 ,
Fig. 9 shows νe1 against time. The value of ν̄E remains
L̃3 = 1.158, η̃ = 4.604, ν̃ = 0.7029. (52)
below νE1 = ν̄, as stated in Theorem 1. Thus, the perfor-
mance of the observer in Section 3 has once again been
17: Lastly, f would be reconstructed using fˆ˜, which is
conclusively demonstrated.
calculated using
Next, to prove the efficacy of the observer in Section 4,
we consider a different set of nonlinear disturbances and
fˆ˜ = 0.3160 1.158 x̃ˆ + 0.1580u − 0.0474 sin x̃ˆ1
 

system parameters, which in the form of system (1)-(2) is  
given by − 1 1.158 ỹ. (53)

1 0
 
−1 −0.3
  
0 The design of the observer is now complete. Similarly,
Eo = , Ao = , Bo = , to demonstrate the robustness of the scheme, we simulate
0 0 5 −2 1
        two scenarios: the first scenario is disturbance-free, with
1 0 0.5 0.5 system initial conditions and fault set as
Mo = , Hxo = , Qo = , Ro = ,
1 1 1 0.5
xo (0) = {0.2, 0.6548},
     
−1 −0.3 0 1
Co = , Do = , No = ,
2.5 0 0.5 0.5 fs2 = − 0.2 + S (0.15, 0.2) + 0.3 sin (1.1t + π/2) .
 
0 (54)
Hyo = , h (To x̆o ) = 0.3 sin (xo1 ) , (51)
0.5
The second scenario has system initial conditions, fault
and disturbances set as
where the states, input, fault, and disturbance are again la-
 T
belled as xo = xo1 xo2 , u, f and ξ , respectively, while xo (0) = {0.2, 0.6445},
 T  
x̆o = xoT f and To = 1 0 0] . We design the fault f f 2 = −0.2 + S (0.15, 0.2) + 0.3 sin (1.1t + π/2) ,
reconstruction scheme according to the procedure in Table
ξ = 0.03 + S (0.1, 0.1) + 0.07 sin (0.5t + 3π/4) .
1 once more.
 T (55)
1: From (51), we obtain rank EoT CoT = n = 2. The
disturbance ξ is also assumed to be bounded. It can be In both scenarios, the observer is again set to have zero
seen that Assumptions 1-2 are satisfied. initial conditions.
Fig. 10 shows the performance of the observer in Sec-
2-3: The matrices Ta , Tb , Tc1 , Tc2 , and Td are found to
 tion 4 for the disturbance-free scenario. After some initial
−1 −0.3 2.5 0 transients due to the mismatch between the initial condi-
be Ta = , Tb = Td = I2 , Tc1 =
2.5 0 0 1 tions of the system and observer, the fault fs2 is accurately
 
0 1  T reconstructed, as per Remark 6.
and Tc2 = . Let x = Td Tc2 Ta xo = x1 x2 and
−1 0 Fig. 11 shows the fault and its reconstruction using the
 T  
x̆ = xT f . We then obtain T = 0.4 0 0 , h (T x̆) = observer in Section 4 for the scenario with non-zero dis-
0.3 sin (0.4x1 ), and Hx2 h (T x̆) = 0. Hence, Assumptions 3- turbances. It can now be seen that the reconstruction is
708 Joseph Chang Lun Chan, Wen-Shyan Chua, Tae H. Lee, and Chee Pin Tan

6. CONCLUSION
0.5

A robust fault reconstruction scheme for nonlinear de-

scriptor systems that are affected by disturbances has been
0 presented. The fault affects the system both linearly and
through nonlinear functions in the state and output equa-
tions. The system was first re-expressed to have structures
that facilitate analysis. Two cases were then considered:
-0.5
in the first case, the nonlinear function in the output equa-
tion contains the fault, while the second case, the fault
0 5 10 15 20 25 30
is absent from the nonlinear function in the output equa-
Fig. 10. The fault (solid) and its reconstruction using tion. In each case, a nonlinear descriptor observer is ap-
the observer in Section 4 (dashed) for the plied to reconstruct the fault, and LMIs were utilised to
disturbance-free scenario in Example 2. design the observer such that the RMS gain from the dis-
turbances to the fault reconstruction error is bounded. The
proposed descriptor observers do not require the rates of
the input and disturbance signals to be bounded, increas-
0.5
ing the applicability of the scheme over state-space ob-
servers (which need these restrictions). Lastly, two simu-
lation examples were carried out and their results demon-
0
strate the effectiveness of the proposed scheme. Future in-
vestigations could explore fault reconstruction in systems
where the fault is not energy-bounded, or where the fault
-0.5
distribution matrix in the output equation is not full-rank.

0 5 10 15 20 25 30
CONFLICTS OF INTEREST
Fig. 11. The fault (solid) and its reconstruction using the The authors declare that they have no conflicts of inter-
observer in Section 4 (dashed) for the scenario est.
with non-zero disturbances in Example 2.

1.5
REFERENCES
15
[1] W. Zheng, Z. Zhang, H. B. Wang, and M. Q. Zhang, “Quan-
0
tized sliding mode H∞ control for singular Markovian jump
1 0 2 4 systems with sensor failure and randomly changing struc-
ture,” International Journal of Control, Automation, and
Systems, vol. 19, no. 12, pp. 3890-3902, 2021.
0.5 [2] S. Xu and J. Lam, Robust Control and Filtering of Singular
Systems, Springer-Verlag, Berlin/Heidelberg, 2006.
[3] J. Zhou, Q. Zhang, B. Men, and S. Huang, “Input-to-state
0 stability of a class of descriptor systems,” International
0 5 10 15 20 25 30 Journal of Robust and Nonlinear Control, vol. 24, no. 1,
pp. 97-109, 2014.
Fig. 12. The empirical RMS gain νe2 (solid) and its bound [4] Y. Wang, V. Puig, F. Xu, and G. Cembrano, “Robust fault
νE2 (dashed) in Example 2. detection and isolation based on zonotopic unknown in-
put observers for discrete-time descriptor systems,” Jour-
also affected by disturbances, so we utilise the empirical nal of the Franklin Institute, vol. 356, no. 10, pp. 5293-
5314, 2019.
RMS gain νe2 to quantify the effect of the disturbances on
the fault reconstruction. In this scenario, νe2 is defined as [5] J. Li, Q. Zhang, X. G. Yan, and S. K. Spurgeon, “Observer-
R (ẽ ) based fuzzy integral sliding mode control for nonlinear de-
νe2 = RTT (ξf) , ẽ f = fˆf 2 − f f 2 .
scriptor systems,” IEEE Transactions on Fuzzy Systems,
Fig. 12 shows νe2 against time. The initial value of νe2 vol. 26, no. 5, pp. 2818-2832, 2018.
is large due to the mismatched initial conditions of the [6] X. Li, C. K. Ahn, D. Lu, and S. Guo, “Robust simultane-
observer and system, but νe2 then goes below νE2 , which ous fault estimation and nonfragile output feedback fault-
shows that νe2 is ultimately bounded by νE2 , as per Theo- tolerant control for Markovian jump systems,” IEEE Trans-
rem 2. Thus, the performance of the proposed scheme has actions on Systems, Man, and Cybernetics: Systems, vol.
been demonstrated. 49, no. 9, pp. 1769-1776, 2019.
 Descriptor Observers for Robust Fault Reconstruction in a Class of Nonlinear Descriptor Systems 709

[7] T. H. Lee, C. P. Lim, S. Nahavandi, and R. G. Roberts, [20] Z. Gao and S. X. Ding, “Fault estimation and fault-tolerant
“Observer-based H∞ fault-tolerant control for linear sys- control for descriptor systems via proportional, multiple-
tems with sensor and actuator faults,” IEEE Systems Jour- integral and derivative observer design,” IET Control The-
nal, vol. 13, no. 2, pp. 1981-1990, 2019. ory & Applications, vol. 1, no. 5, pp. 1208-1218, 2007.
[8] C. Deng, X. Z. Jin, W. W. Che, and H. Wang, “Learning- [21] T. Binazadeh and M. S. Asadinia, “A delay-dependent ap-
based distributed resilient fault-tolerant control method for proach to finite-time H∞ control of nonlinear descriptor
heterogeneous MASs under unknown leader dynamic,” systems with state delay via observer-based control,” Cir-
IEEE Transactions on Neural Networks and Learning Sys- cuits, Systems, and Signal Processing, vol. 39, no. 11, pp.
tems, vol. 33, no. 10, pp. 5504-5513, 2022. 5454-5474, 2020.
[9] F. Zhu, Y. Shan, and Y. Tang, “Actuator and sensor fault de- [22] M. C. Nguyen, H. Trinh, and P. T. Nam, “Non-linear ob-
tection and isolation for uncertain switched nonlinear sys- server design for a class of singular time-delay systems
tem based on sliding mode observers,” International Jour- with Lipschitz non-linearities,” IMA Journal of Mathemat-
nal of Control, Automation, and Systems, vol. 19, no. 9, pp. ical Control and Information, vol. 34, no. 3, pp. 919-935,
3075-3086, 2021. 2017.
[10] W. S. Chua, J. C. L. Chan, C. P. Tan, E. K. P. Chong, and S. [23] T. Berger, “On observers for nonlinear differential-
Saha, “Robust fault reconstruction for a class of nonlinear algebraic systems,” IEEE Transactions on Automatic Con-
systems,” Automatica, vol. 113, 108718, 2020. trol, vol. 64, no. 5, pp. 2150-2157, 2019.
[11] S. Laghrouche, J. Liu, F. S. Ahmed, M. Harmouche, and [24] S. Liu, X. Li, H. Wang, and J. Yan, “Adaptive fault esti-
M. Wack, “Adaptive second-order sliding mode observer- mation for T-S fuzzy systems with unmeasurable premise
based fault reconstruction for PEM fuel cell air-feed sys- variables,” Advances in Difference Equations, vol. 2018,
tem,” IEEE Transactions on Control Systems Technology, no. 1, pp. 105-117, 2018.
vol. 23, no. 3, pp. 1098-1109, 2015. [25] F. Nemati, S. H. S. Mostafa, and A. Zemouche, “A non-
[12] Y. Gao, J. Liu, G. Sun, M. Liu, and L. Wu, “Fault devi- linear observer-based approach to fault detection, isolation
ation estimation and integral sliding mode control design and estimation for satellite formation flight application,”
for Lipschitz nonlinear systems,” Systems & Control Let- Automatica, vol. 107, pp. 474-482, 2019.
ters, vol. 123, pp. 8-15, 2019. [26] A. H. Tahoun and M. Arafa, “A new unmatched-
[13] R. Shahnazi and Q. Zhao, “Adaptive fuzzy descriptor slid- disturbances compensation and fault-tolerant control for
ing mode observer-based sensor fault estimation for uncer- partially known nonlinear singular systems,” ISA Transac-
tain nonlinear systems,” Asian Journal of Control, vol. 18, tions, vol. 104, pp. 310-320, 2020.
no. 4, pp. 1478-1488, 2016. [27] B. Jiang and F. N. Chowdhury, “Parameter fault detection
[14] J. Cheng, J. H. Park, and Z. G. Wu, “Observer- and estimation of a class of nonlinear systems using ob-
based asynchronous control of nonlinear systems with servers,” Journal of the Franklin Institute, vol. 342, no. 7,
dynamic event-based try-one-discard protocol,” IEEE pp. 725-736, 2005.
Transactions on Cybernetics, pp. 1-11, 2021. DOI: [28] C. Martínez-García, C. Astorga-Zaragoza, V. Puig, J.
10.1109/TCYB.2021.3104806 Reyes-Reyes, and F. López-Estrada, “A simple nonlinear
[15] J. C. L. Chan, T. H. Lee, and C. P. Tan, “A sliding mode observer for state and unknown input estimation: DC motor
observer for robust fault reconstruction in a class of non- applications,” IEEE Transactions on Circuits and Systems
linear non-infinitely observable descriptor systems,” Non- II: Express Briefs, vol. 67, no. 4, pp. 710-714, 2020.
linear Dynamics, vol. 101, no. 2, pp. 1023-1036, 2020. [29] A. Sassi, H. S. Ali, M. Zasadzinski, and K. Abderrahim,
[16] S. A. Nugroho, V. Hoang, M. Radosz, S. Wang, and “Adaptive observer design for a class of descriptor non-
A. F. Taha, “New insights on one-sided Lipschitz and linear systems,” European Journal of Control, vol. 44, pp.
quadratically-inner bounded nonlinear dynamic systems,” 90-102, 2018.
Proc. of the American Control Conference, pp. 4558-4563, [30] H. Kazemi and A. Yazdizadeh, “Optimal state estima-
2020. tion and fault diagnosis for a class of nonlinear systems,”
[17] J. Li, Y. Liu, J. Yu, Y. Sun, and M. Liu, “A new result of IEEE/CAA Journal of Automatica Sinica, vol. 7, no. 2, pp.
terminal sliding mode finite-time state and fault estimation 517-526, 2020.
for a class of descriptor switched systems,” Applied Math- [31] Q. Ha and H. Trinh, “State and input simultaneous estima-
ematics and Computation, vol. 402, 126085, 2021. tion for a class of nonlinear systems,” Automatica, vol. 40,
[18] M. Li, J. Xia, J. Zhao, H. Zhang, and H. Shen, “Event- no. 10, pp. 1779-1785, 2004.
triggered extended dissipative control for networked singu- [32] J. Pan, M. Meng, and J. Feng, “A note on observers design
lar systems,” International Journal of Control, Automation, for one-sided Lipschitz nonlinear systems,” Proc. of the
and Systems, vol. 19, no. 1, pp. 382-391, 2021. 34th Chinese Control Conference, pp. 1003-1007, 2015.
[19] J. C. L. Chan, T. H. Lee, and C. P. Tan, “Observer-based [33] J. C. L. Chan, T. H. Lee, C. P. Tan, H. Trinh, and J. H. Park,
fault-tolerant control for non-infinitely observable descrip- “A nonlinear observer for robust fault reconstruction in
tor systems,” in: J. H. Park (Ed.), Recent Advances in one-sided Lipschitz and quadratically inner-bounded non-
Control Problems of Dynamical Systems and Networks, linear descriptor systems,” IEEE Access, vol. 9, pp. 22455-
Springer, Cham, pp. 123-145, 2021. 22469, 2021.
710 Joseph Chang Lun Chan, Wen-Shyan Chua, Tae H. Lee, and Chee Pin Tan

[34] D. Kharrat, H. Gassara, A. El Hajjaji, and M. Chaabane, Tae H. Lee received his B.S., M.S., and
“Adaptive observer and fault tolerant control for Takagi- Ph.D. degrees in electrical engineering
Sugeno descriptor nonlinear systems with sensor and actu- from Yeungnam University, Korea, in
ator faults,” International Journal of Control, Automation, 2009, 2011, and 2015, respectively. He
and Systems, vol. 16, no. 3, pp. 972-982, 2018. was a postdoctoral researcher in the same
[35] D. Du, Y. Yang, H. Zhao, and Y. Tan, “Robust fault diagno- university from 2015 to 2017 and a Alfred
sis observer design for uncertain switched systems,” Inter- Deakin Postdoctoral Research Fellow at
national Journal of Control, Automation, and Systems, vol. Institute for Intelligent Systems Research
18, no. 12, pp. 3159-3166, 2020. and Innovation, Deakin University, Aus-
tralia in 2017. He began with Jeonbuk National University,
[36] L. Dai, Singular Control Systems, Springer-Verlag, Berlin, Korea in 2017, where he is currently an Associate Professor. His
1989. research interests are in the field of stability analysis, sampled-
[37] W. Zhang, Y. Zhao, and L. Sheng, “Some remarks on sta- data control, dynamics of systems, and control issues in network
bility of stochastic singular systems with state-dependent systems. He is an author/coauthor of several journal papers and
noise,” Automatica, vol. 51, pp. 279-277, 2015. books. Dr. Lee currently serves as an Associate Editor for the
Applied Mathematics and Computation.
[38] J. Li and Q. Zhang, “An integral sliding mode control ap-
proach to observer-based stabilization of stochastic Itô de-
scriptor systems,” Neurocomputing, vol. 173, no. 3, pp. Chee Pin Tan received his B.Eng. (Hons.)
1330-1340, 2016. and Ph.D. degrees from University of Le-
icester, Leicester, UK, in 1998 and 2002,
respectively. He is currently a Professor
Joseph Chang Lun Chan received his at the School of Engineering, Monash
B.Eng. (Hons.) and Ph.D. degrees from University Malaysia, Bandar Sunway,
Monash University Malaysia, in 2013 and Malaysia. He has authored more than 60
2019, respectively. He was a postdoctoral internationally peer-reviewed research ar-
researcher from 2019 to 2021 in Jeonbuk ticles, including a book on fault recon-
National University, Korea, where he was struction. His research interests include robust fault estimation
also a Brain Korea Plus (BK21+) Postdoc- and observers. He is currently an Associate Editor of the Journal
toral Fellowship Scholar. Since 2021, he of the Franklin Institute and the International Journal of Systems
has been a postdoctoral research fellow in Science.
Monash University Malaysia. His research focusses on applica-
tion of sliding mode observers in non-infinitely observable de- Publisher’s Note Springer Nature remains neutral with regard
scriptor systems and other practical systems, as well as fault re- to jurisdictional claims in published maps and institutional affil-
construction in non-linear systems. iations.

Wen-Shyan Chua obtained his Ph.D. de-

gree in control systems engineering and a
B.Eng. degree in mechatronics engineer-
ing from Monash University Malaysia and
is a registered Professional Technologist
under the Malaysian Board of Technolo-
gists (MBOT). He is currently the Head
of Malaysian Smart Factory 4.0 @ Selan-
gor Human Resource Development Cen-
tre (SHRDC) which is a recognized technology centre by the
Malaysian Productivity Corporation (MPC) to support and accel-
erate Industry 4.0 technology adoption within the Malaysian In-
dustries. He currently focuses on developing community-driven
and open-source technology integration and use cases to boost
adoption of Industry 4.0 within the industries and community in
Malaysia.