Non Linear Parameter Varying Observer

based on Descriptor Modeling for Damper

Fault Estimation ⋆
Thanh-Phong Pham ∗ Olivier Sename ∗∗
Gia Quoc Bao Tran ∗∗,∗∗∗
∗
Faculty of Electrical and Electronic Engineering, The University of
Danang - University of Technology and Education, 550000 Danang,
Vietnam (e-mail: ptphong@ute.udn.vn).
∗∗
Univ. Grenoble Alpes, CNRS, Grenoble INP⊤ , GIPSA-lab, 38000
Grenoble, France ⊤ Institute of Engineering Univ. Grenoble Alpes
(e-mail: olivier.sename@grenoble-inp.fr).
∗∗∗
Centre Automatique et Systèmes (CAS),
Mines Paris, Université PSL, 75006 Paris, France
(e-mail: gia-quoc-bao.tran@minesparis.psl).

Abstract: This paper proposes an H∞ Non Linear Parameter Varying (NLPV) observer for
fault estimation in semi-active Electro-Rheological (ER) suspensions. The damper fault (a loss-
of-efficiency factor) is modeled as a lost force of unknown/free dynamics to be estimated.
Thanks to the parameter-dependent descriptor-form system modeling, there is no assumption
made on the fault dynamics, thus making this method applicable to all considered types of
damper faults. The nonlinearity in the damper model is bounded by its Lipschitz property,
while the road disturbance and the measurement noise are handled using the H∞ condition.
The observer is parameterized and then designed by solving Linear Matrix Inequalities (LMIs)
and is implemented in a polytopic gain scheduling approach. Synthesis results including Bode
plots and simulations illustrate the method in both the frequency and the time domains.

Keywords: Descriptor form, NLPV observer, fault estimation, semi-active suspension, Lipchitz
condition.

1. INTRODUCTION et al., 2022). In these works, the loss of effectiveness of

In the automotive field, semi-active (SA) suspension sys-
tems are a potential candidate since they offer advantages
in improving driving comfort with better performance
than passive suspension and reduced energy consumption
compared to active ones. There have been a lot of research,
development, and implementation of such SA suspensions
proposed in the literature (see (Savaresi et al., 2010) and
references therein). Thorough reviews about these systems
can be found in (Sename, 2021; Poussot-Vassal et al.,
2012). It is worth noting that the loss of effectiveness
of the SA suspension due to an electrical fault, physical
deformation, or oil leakage may significantly reduce vehicle
performances (Hernández-Alcántara et al., 2016; Morato
et al., 2020). Therefore, fault detection of this system is of
paramount importance for reliability.
For those purposes, a few studies have been concerned with
the modeling of loss of effectiveness of the SA damper,
as seen in (Hernández-Alcántara et al., 2016; Morato
et al., 2020). Furthermore, some recent works have been
dedicated to the study of damper fault estimation (Morato
et al., 2019; Nguyen et al., 2016; Do et al., 2018; Tran
⋆ This research is funded by Funds for Science and Technology
Development of the University of Danang under project number
B2020-DN06-21.
the damper is represented in the multiplicative or additive
fault representations to develope a fault observer; however,
the time derivative of the fault is assumed to be zero, which
simplifies the theoretical problem but is not realistic for
the considered system. Among fault estimation methods,
the Proportional Integral (PI) observer (Do et al., 2018;
Guzman et al., 2021) is interesting given that the fault
varies very slowly and can then be modeled as a constant
extra state. Besides, it is important to notice that the two
main requirements concerning the development of the fault
detection methods for the SA suspension are as follows: i)
The schemes have to be able to deal with the bi-viscous
nonlinearity in the SA suspension model; ii) The method
must be able to handle the effect of unknown road profile
disturbances and sensor noise on the estimation error.
To avoid using the slow-variation assumption of the fault
dynamics and handle the requirements mentioned above,
robust observers for descriptor Lipschitz systems are an
interesting approach since the nonlinear function in the
SA suspension model satisfies the global Lipschitz condi-
tion (see (6)). For the last two decades, many theoretical
contributions to designing the observers for descriptor
Lipschitz nonlinear systems have been proposed in the
literature (Koenig, 2006; Darouach et al., 2011; Darouach
and Boutat-Baddas, 2008; Darouach et al., 2017; Osorio-
Gordillo et al., 2019; Ha and Trinh, 2004; Delshad et al., the tire which is modeled as a spring of stiffness kt . From
2016). Those methods also account for disturbance min- Newton’s second law of motion, the system dynamics
imization by mean of H∞ , H2 or mixed H∞ /H2 crite- around the equilibrium are
rion. Thus an H∞ robust observer for singular nonlinear
(
ms z̈s = −Fs − Fdf
parameter-varying (NLPV) systems has been presented in (1)
(Do et al., 2020), but that method may be conservative due mus z̈us = Fs + Fdf − Ft ,
to assumptions such as bounded energy noise derivative. where Fs = ks zdef is the spring force (zdef = zs −
In this paper, we consider the problem of estimating addi- zus is the deflection); Ft = kt (zus − zr ) is the tire
tive damper faults, represented by a loss of effectiveness, force; the faulty damper force Fdf is given in (3); zs and
using accelerometers as inputs of the proposed observer zus are the displacements of the sprung and unsprung
only (to limit the cost over deflection sensors). To reduce masses, respectively; zr is the road displacement input.
the conservatism in (Do et al., 2020), an extension of The damper force Fd in the healthy case is as follows
(Delshad et al., 2016) is proposed using the S-procedure to

F = k0 zdef + c0 żdef +Fer
 d

integrate the Lipschitz condition into the H∞ condition.
 | {z }
Our main contributions are summarized as follows: Fpassive (2)
Ḟer = − 1 Fer + fc · u · tanh(k1 zdef + c1 żdef ).


• The semi-active suspension system is modeled as a de- τ τ
scriptor NLPV formulation without any assumption Taking the loss of efficiency of the damper in to account,
on the fault dynamics; the faulty damper force is
• The results presented in (Delshad et al., 2016) to de-
sign a reduced-order observer are extended to a class Fdf = Fd − αFd = Fd − f, (3)
of descriptor NLPV systems, for fault estimation; where α ∈ [0, 1] is the loss of effectiveness factor; f is the
• The proposed approach has been simulated on a lost damper force to be estimated. Substituting (3) into
quarter-car model built from our suspension testbed (1), we obtain the system dynamics considering the loss of
presented in (Pham et al., 2019). The observer per- effectiveness of the damper
formances are then assessed with simulation results

ms z̈s = −Fs − Fd + f
in the time and frequency domain. (4)
mus z̈us = Fs + Fd − f − Ft ,
2. SEMI-ACTIVE SUSPENSION MODELING Choosing the state x = (x1 , x2 , x3 , x4 , x5 , x6 )⊤ = (zs −
zus , żs , zus − zr , żus , Fer , f )⊤ ∈ R6 , the measured output
A nonlinear dynamical model of the Electro-Rheological y = (z̈s , z̈us )⊤ ∈ R2 , the scheduling variable ρ = u ∈ [0, 1],
(ER) suspension system is illustrated in Fig. 1. More we rewrite the system dynamics in the descriptor NLPV
details about this system, as the model parameters given form as
below and used for simulations are presented in (Pham 
E ẋ = Ax + B(ρ)Φ(Ex) + D1 ω
et al., 2019). (5)
y = Cx + D2 ω,
⊤
where ω = (żr ωn ) , in which żr is the road profile
derivative and ωn is the sensor noise.
The system nonlinearity
Φ(Ex) = tanh(k1 x1 + c1 (x2 − x4 )) = tanh(Γe x),
with Γe = (k1 c1 0 −c1 0 0), is globally Lipschitz, i.e., for
all (x, x̂) ∈ R6 × R6 ,
∥Φ(Ex) − Φ(E x̂)∥ ≤ ∥Γe (x − x̂)∥. (6)
The matrices in (5) are (with k = ks + k0 ):
 

1 0 0 0 0 0
 0 
0 0

0 1 0 0 0 0  0  0 0
E = 0 0 1 0 0 0 , B(ρ) =  0  , D1 = −1 0 ,
 
 
0 0 0 1 0 0  0  
0 0

ρ
0 0 0 0 1 0 fc 0 0
 τ 
0 1 0 −1 0 0
 −k −c0
0
c0 −1 1 
 ms ms ms ms ms   
−2
0  , D = 0 10
 0 0 0 1 0

A= 2 ,
 k c0 −kt −c0 1 −1  0 10−3
 
 mus mus mus mus mus mus 
−1
0 0 0 0 0
τ
−k −c0 c0 −1 1
 
Fig. 1. Quarter-car model with semi-active suspension. 0
C= mk
s ms
c0 −kt
ms ms ms  .
−c0 1 −1
The well-known quarter-car model in Fig. 1 consists of mus mus mus mus mus mus
 
the sprung mass ms , the unsprung mass mus , and the E
Note that here rank = 6.
suspension components located between these masses and C
 3. NLPV OBSERVER DESIGN Next, let M ∈ Rnx ×nx be an arbitrary matrix of full row
rank such that  
In this section, the reduced-order observer definition for
 
M TE
descriptor systems (Delshad et al., 2016) is extended to a rank = rank = nx . (16)
C C
class of descriptor NLPV systems (5) as follows
 Then there always exists a parameter matrix K such that
ż = N (ρ)z + J(ρ)y + H(ρ)Φ(E x̂)     
(7) TE I −K M
x̂ = Rz + Sy, = ⇐⇒ T E = M − KC
C 0 I C
where z ∈ Rnx −ny is the state variable of the reduced-  
E
order observer, x̂ is the estimate of x ∈ Rnx , y ∈ Rny is ⇐⇒ (T K) = M. (17)
C
the measurement, and ρ ∈ Rnρ is the scheduling parameter
(here nx = 6, ny = 2, and nρ = 1). The observer matrices A solution for (16) is given by
N (ρ), J(ρ), H(ρ), R, and S of appropriate dimensions have (T K) = M Σ+ , (18)
to be designed. Let us introduce the dynamic error +
where Σ = (E C) where Σ is any general inverse of
ϵ = z − T Ex, ∈ Rnx −ny (8) matrix Σ satisfying ΣΣ+ Σ = Σ. This is equivalent to
where the matrix T is an arbitrary matrix.    
Differentiating (8) with respect to time and using (5) and + I + 0
T = MΣ , K = MΣ . (19)
(7), one obtains 0 I
Besides, the solution set of (14) is given by


 ϵ̇ = N (ρ)ϵ + (N (ρ)T − T A + J(ρ)C)x  

 +(J(ρ)D2 − T D1 )ω + (H(ρ) − T B(ρ))Φ(E x̂) N (ρ) J(ρ)
 −T B(ρ)(Φ(Ex) − Φ(Êx)) R S


x̂ = Rϵ + (RT E + SC)x + SD ω.   +     + !
2 TA TE Z1 (ρ) TE TE
(9) = + I− ,
I C Z2 C C
It is obvious that if the decoupling conditions (20)
N (ρ)T E − T A + J(ρ)C = 0, (10)
 
Z1 (ρ)
H(ρ) − T B(ρ) = 0, (11) where is a free matrix of appropriate dimension.
Z2
RT E + SC = I, (12) This is equivalent to
are satisfied, the system (9) becomes N (ρ) = T Aα1 + Z1 (ρ)β1 , (21)

ϵ̇ = N (ρ)ϵ − T B(ρ)∆Φ + (J(ρ)D2 − T D1 )ω J(ρ) = T Aα2 + Z1 (ρ)β2 , (22)
(13)
e = Rϵ + SD2 ω, R = α 1 + Z 2 β1 , (23)
where e = x̂ − x is the state estimation error and ∆Φ = S = α 2 + Z 2 β2 , (24)
Φ(Ex) − Φ(E x̂). 
TE
+  
I

TE
+  
0
where α1 = , α2 = ,
The problem of the H∞ observer design is thus reduced to C 0 C I
determining the observer matrices N (ρ), J(ρ), H(ρ), R,   + !  
and S such that TE TE I
β1 = I − , and
C C 0
• All the conditions (10)-(12) are satisfied;
• The effect of the combined disturbance-noise ω on   + !  
the state estimation error e is minimized while ∆Φ is TE TE 0
β2 = I − .
bounded by the Lipschitz condition (6). C C I

3.1 Parameterization of the Observer Matrices Remark: If the matrices N (ρ), J(ρ), H(ρ), R, and S can
be chosen according to (21), (22), (11), (23), and (24),
First note that from (11), we get respectively, then all conditions (10)-(12) are fulfilled.
H(ρ) = T B(ρ). From the results of above parameterization, for brevity,
Therefore, H(ρ) will be given when T is chosen. Now, the matrices of system (13) can be rewritten as
in order to determine T and the observer matrices N (ρ), A(ρ) = N (ρ) = A11 + Z1 (ρ)A12 , (25)
J(ρ), R, and S of the proposed observer satisfying all the
conditions equalities (10)-(12), parameterization is made B(ρ) = J(ρ)D2 − T D1 = B11 + Z1 (ρ)B12 , (26)
by using the general solution of (10) and (12). W(ρ) = −T B(ρ), (27)
C = R = C11 + Z2 C12 , (28)
First, from (10) and (12), one obtains
     D = SD2 = D11 + Z2 D12 , (29)
N (ρ) J(ρ) TE TA
= . (14) where A11 = T Aα1 , A12 = β1 , B11 = T Aα2 D2 − T D1 ,
R S C I
B12 = β2 D2 , C11 = α1 , C12 = β1 , D11 = α2 D2 , and
The equation (14) is solvable if and only if D12 = β2 D2 . Notice that all the matrices A11 , A12 , B11 ,
TE B12 , C11 , C12 , D11 , and D12 are known and the matrix
 
 
C  TE W(ρ) is known at each vertex of ρ. Therefore, the observer
rank  = rank = nx . (15) design problem is reduced to determining Z1 (ρ) and Z2 ,
T A C
I which is discussed in the following part.
3.2 Polytopic H∞ Observer Design From (6), the following condition is obtained
(Φ(Ex) − Φ(E x̂))⊤ (Φ(Ex) − Φ(E x̂)) ⩽ e⊤ Γ⊤
e Γe e
Using (25)-(29), we rewrite the estimation error dynamics
(13) as ⇐⇒ (∆Φ)⊤ ∆Φ ⩽ (Cϵ + Dω)⊤ Γ⊤
e Γe (Cϵ + Dω)
⇐⇒ η ⊤ Q2 η ⩽ 0,

ϵ̇ = A(ρ)ϵ + W(ρ)∆Φ + B(ρ)ω (36)
(30)
e = Cϵ + Dω. −C⊤ Γ⊤ ⊤ ⊤
 
e Γe C 0 −C Γe Γe D
where Q2 =  0 I 0 .
After the parameterization step, while ∆Φ is bounded by ⊤ ⊤ ⊤ ⊤
−D Γe Γe C 0 −D Γe Γe D
the Lipschitz condition (6), the observer design problem is
now to determine the matrices Z1 (ρ) and Z2 such that By applying the S-procedure (Boyd et al., 1994) to the
• The system (30) is asymptotically stable for ω(t) = 0; inequalities (35) and (36), we have that V̇ +e⊤ e−γ 2 ω ⊤ ω <
• ∥e(t)∥L2 < γ∥ω(t)∥L2 for ω(t) ̸= 0; γ is minimized. 0 if there exists a scalar ϵl > 0 such that
V̇ + e⊤ e − γ 2 ω ⊤ ω − ϵl (η ⊤ Qη) < 0
In this paper, the design of the observer will be carried
out using the polytopic method. It means that we assume ⇐⇒ η ⊤ (Q1 (ρ) − ϵl Q2 )η < 0, (37)
that Z1 (ρ) depends in an affine way on the parameter ρ. The condition (37) is equivalent to
In such a case, the design method is restricted to solving Q1 (ρ) − ϵl Q2 < 0
the given problem only to get the vertices Z1,i of Z1 (ρ)  
formed when ρ varies within its bounds (Apkarian et al., Ωa (ρ) XW(ρ) Ωb (ρ)
1995). Theorem 1 then solves the observer design problem ⇐⇒ W⊤ (ρ)X −ϵl I 0  < 0, (38)
⊤
in an LMI framework. Ωb (ρ) 0 Ωc
Theorem 1. Consider the system model (5) and the ob- where Ωa (ρ) = A(ρ)⊤ X + XA(ρ) + C⊤ C + ϵl C⊤ Γ⊤
e Γe C,
server (7). The observer design problem is solved if there Ωb = XB(ρ) + C⊤ D + ϵl C⊤ Γ⊤ ⊤
e Γe D, and Ωc = D D +
exist matrices X = X ⊤ > 0, Yi , Z2 , and a scalar ϵl > 0 ϵD⊤ Γ⊤ Γ
e e D − γ 2
I.
minimizing γ such that
  Applying Schur’s complement to (38), one obtains
Ω11,i XW(ρi ) Ω13,i Ω14 Ω15
Ω1 (ρ) XW(ρ) XB(ρ) C⊤ C⊤ Γ⊤
 
⊤
W (ρi )X −ϵl I 0 0 0  e
 Ω⊤

0 −γ 2
I Ω Ω
 W⊤ (ρ)X −ϵl I 0 0 0 
13,i 34  < 0,
35  (31)  ⊤
 B X 0 2 ⊤ ⊤ ⊤
−γ I D D Γe  < 0, (39)
 Ω⊤ Ω⊤

14 0 34 −I 0 
0 −I 0
 
C D
Ω⊤15 0 Ω⊤
35 0 −ϵl I
Γe C 0 Γe D 0 −ϵl I
for ρi , i = 1, 2, . . . , 2 at the 2nρ vertices of ρ, where
nρ

Ω11,i = A⊤ ⊤ ⊤ where Ω1 (ρ) = A(ρ)⊤ X + XA(ρ).

11 X + XA11 + A12 Yi + Yi A12 ; Ω13,i = XB11 +
⊤ ⊤ ⊤ ⊤ ⊤ ⊤ ⊤
Yi B12 ; Ω14 = C11 + C12 Z2 ; Ω15 = C11 Γe + C12 Z 2 Γe ; Substituting (25)-(29) into (39) and letting Yi = −XZ1,i
⊤ ⊤ ⊤ ⊤ ⊤ ⊤ ⊤ ⊤
Ω34 = D11 + D12 Z2 ; Ω35 = D11 Γe + D12 Z2 Γe . Then, with Z1,i = Z1 (ρi ) at each vertex ρi of ρ, we obtain the
Z1,i for each vertex of ρ is found as Z1,i = −X −1 Yi . LMI (31) at the vertex ρi .
Proof. Consider the Lyapunov function candidate If (31) is satisfied, from (36), (37) implies that
V = ϵ⊤ Xϵ. (32) V̇ + e⊤ e − γ 2 ω ⊤ ω < 0. (40)
Differentiating V along the solution of (30) yields Following the steps in (Darouach et al., 2011), we get
V̇ = ϵ̇⊤ Xϵ + ϵ⊤ X ϵ̇ ∥e(t)∥2L2 < γ 2 ∥ω(t)∥2L2 . (41)
= (A(ρ)ϵ + W(ρ)∆Φ + B(ρ)ω) Xϵ ⊤ The proof is completed. □
+ ϵ⊤ X(A(ρ)ϵ + W(ρ)∆Φ + B(ρ)ω). (33) After solving Theorem 1 for the 2 vertices of Z1 (ρ) and for
Z2 , we solve (21), (22), (11), (23), and (24) for the observer
To satisfy the performance objective w.r.t the L2 gain matrices (at the vertices), then the polytopic convex
disturbance attenuation, we must satisfy the inequality computation of the matrix gain is used in implementation
V̇ + e⊤ e − γ 2 ω ⊤ ω < 0 ⇐⇒ (Apkarian et al., 1995).
ϵ⊤ (A(ρ)⊤ X + XA(ρ) + C⊤ C)ϵ + ∆Φ⊤ W⊤ (ρ)Xϵ
4. OBSERVER SYNTHESIS RESULTS
+ ϵ⊤ XW(ρ)∆Φ + ω ⊤ (B(ρ)⊤ X + D⊤ C)ϵ
+ ϵ⊤ (XB(ρ) + C⊤ D)ω + ω ⊤ (D⊤ D − γ 2 I)ω < 0. In this Section, the synthesis results of the NLPV observer
(34)
are shown. First, the observer proposed in Section 3 is
ϵ
!
designed for the system presented in Section 2. Solving
Defining η = ∆Φ , one obtains Theorem 1, so the LMIs (31) with vertices ρ1 = 0 and
ω ρ2 = 1, we obtain the minimum L2 -induced gain γ =
V̇ = η ⊤ Q1 (ρ)η < 0, (35) 2.0078, ϵl = 160, and the matrices Z1,i , i = 1, 2 and Z2 .
 According to (21)-(24) the observer matrices at each vertex
XW(ρ) XB(ρ) + C⊤ D

Ω(ρ) are obtained as follows: Ni = T Aα1 +Z1,i β1 , Ji = T Aα2 +
where Q1 (ρ) =  W⊤ (ρ)X 0 0 ,
Z1,i β2 , for i = 1, 2, R = α1 + Z2 β1 , and S = α2 + Z2 β2 .
B(ρ)⊤ X + D⊤ C 0 D⊤ D − γ 2 I Then the observer matrices N (ρ) and J(ρ) are deduced by
⊤ ⊤
where Ω(ρ) = A(ρ) X + XA(ρ) + C C. using convex interpolation.
In Fig. 2 and Fig. 3, the Bode diagrams of the estimation
error systems w.r.t the road profile derivative and sensor
noise are shown for the two vertex observers for parameter
values ρ = {0, 1}. These results emphasize the satisfactory
attenuation level (typically of the range -50 to -100 dB) of
the unknown road profile derivative, and of measurement
noise appearing at high frequencies, on the six estimation
errors e with scheduling parameter ρ1 = 0 (red dashed
line) and ρ2 = 1 (blue line).
Bode Diagram
-100
e1
⊤ ⊤
(0 0 0 0 0 0) for the system and z(0) = (0 0.01 0.2 1)
for the reduced-order observer.
Two simulation scenarios are used to evaluate the perfor-
mance of the observer as follows.
Simulation 1:
• The road profile is sinusoidal;
• The control u (recall ρ = u) is constant at u = 0.3;
• α (as in (3)) increases from 0 to 0.2 at 7s.
Simulation 2 :
• An ISO 8608 road profile of Type C is used;

-150
• The control u is obtained from a Skyhook controller.
It is important to note that in such a case, u varies
-150
e2

infinitely fast, which means that the use of the poly-

-200 topic approach is justified;
• α increases from 0 to 0.2 at 5s and then to 0.4 at 10s.
Magnitude (dB)

-150
e3

-200
0.4
-150
e4

0.01 0.3
-200 0.005
0.2
-150 0
e5

0.1
-0.005
-200
-50 -0.01 0
e6

-100 -0.015 -0.1

0 5 10 15 0 5 10 15
-1 0 1 2
10 10 10 10
Frequency (Hz) (a) (b)

1.5 1.5
Fig. 2. Transfer ∥e/żr ∥—Bode diagrams of NLPV observer 1 1
w.r.t the road profile derivative with ρ1 = 0 (red
Lost Force (N)

0.5 0.5

Error (N)
dashed line) and ρ2 = 1 (blue line). 0 0

-0.5 -0.5

-1 -1

Bode Diagram -1.5

0 5 10 15
-1.5
0 5 10 15
Time(s) Time(s)
-80
e1

(c) (d)
-100
-60
e2

-80 Fig. 4. Simulation 1: (a) Road profile, (b) Loss-of-efficiency

factor, (c) Fault estimation, and (d) Estimation error.
Magnitude (dB)

-80
e3

-90
-100
0.8
-80
e4

-90 0.01
0.6
-100
0.4
0
-100
e5

0.2
-120 -0.01
0

-20 0 5 10 15 0 5 10 15
e6

-30
-40
(a) (b)
100 101 102 103 104
Frequency (Hz) 10
10 1
0.5
Lost Force (N)

0 5
Fig. 3. Transfer ∥e/ωn ∥—Bode diagrams of NLPV ob- 5
Error (N)

-0.5
0.5 1 1.5
0
server w.r.t the measurement noise with ρ1 = 0 (red 0

dashed line) and ρ2 = 1 (blue line). -5 -5

-10 -10
0 5 10 15 0 5 10 15
Time(s) Time(s)
5. SIMULATION RESULTS
(c) (d)
To emphasize the effectiveness of the proposed approach,
simulations are now performed considering the nonlinear Fig. 5. Simulation 2: (a) Road profile, (b) Loss-of-efficiency
quarter-car model (5). The initial conditions are x(0) = factor, (c) Fault estimation, and (d) Estimation error.
The simulation results are shown in Fig. 4 and Fig. 5. It
can be seen that the asymptotic estimation is achieved
with small errors (see Table 1) and for various kinds of
variations in the fault, which highlights the advantage that
our method is appropriate for all fault dynamics.
Table 1. Normalized Root-Mean-Square Errors
(NRMSE).
Simulation NRMSE (-)
Scenario 1 0.0315
Scenario 2 0.0091
6. CONCLUSION
This paper presents an NLPV observer to estimate the
damper fault (modeled as the lost damper force) in SA
ER automotive suspensions. The descriptor NLPV mod-
eling avoids any assumption made on the fault dynamics,
making the approach suitable for any faults. While the
Lipschitz condition is used to bound the system nonlinear-
ity, the combined effects of unknown inputs (road profile
derivative and measurement noise) on the estimation error
are minimized using the H∞ condition. Both frequency-
domain analysis and time-domain simulations assess the
performance of the method.
REFERENCES
Apkarian, P., Gahinet, P., and Becker, G. (1995). Self-
scheduled H∞ Control of Linear Parameter-varying Sys-
tems: A Design Example. Automatica, 31(9), 1251–1261.
Boyd, S., El Ghaoui, L., Feron, E., and Balakrishnan,
V. (1994). Linear Matrix Inequalities in System and
Control Theory, volume 15. Siam.
Darouach, M. and Boutat-Baddas, L. (2008). Observers
for a Class of Nonlinear Singular Systems. IEEE
Transactions on Automatic Control, 53(11), 2627–2633.
doi:10.1109/TAC.2008.2007868.
Darouach, M., Amato, F., and Alma, M. (2017).
Functional Observers Design for Descriptor
Systems via LMI: Continuous and Discrete-
time Cases. Automatica, 86, 216–219. doi:
https://doi.org/10.1016/j.automatica.2017.08.016.
Darouach, M., Boutat-Baddas, L., and Zerrougui, M.
(2011). H∞ Observers Design for a Class of Nonlinear
Singular Systems. Automatica, 47(11), 2517–2525.
Delshad, S.S., Johansson, A., Darouach, M., and Gustafs-
son, T. (2016). Robust State Estimation and Unknown
Inputs Reconstruction for a Class of Nonlinear Systems:
Multiobjective Approach. Automatica, 64, 1–7.
Do, M.H., Koenig, D., and Theilliol, D. (2018).
Robust H∞ Proportional-Integral Observer for
Fault Diagnosis: Application to Vehicle Suspension.
IFAC-PapersOnLine, 51(24), 536–543. doi:
https://doi.org/10.1016/j.ifacol.2018.09.628. 10th
IFAC Symposium on Fault Detection, Supervision and
Safety for Technical Processes SAFEPROCESS 2018.
Do, M.H., Koenig, D., and Theilliol, D. (2020). H∞
Observer Design for Singular Nonlinear Parameter-
varying System. In 2020 59th IEEE Conference
on Decision and Control (CDC), 3927–3932. doi:
10.1109/CDC42340.2020.9303844.
Guzman, J., López-Estrada, F.R., Estrada-Manzo, V., and
Valencia-Palomo, G. (2021). Actuator Fault Estimation
based on a Proportional-integral Observer with Non-
quadratic Lyapunov Functions. International Journal
of Systems Science, 1–14.
Ha, Q.P. and Trinh, H. (2004). State and Input Simul-
taneous Estimation for a Class of Nonlinear Systems.
Automatica, 40(10), 1779–1785.
Hernández-Alcántara, D., Tudón-Martı́nez, J.C.,
Amézquita-Brooks, L., Vivas-López, C.A., and
Morales-Menéndez, R. (2016). Modeling, Diagnosis and
Estimation of Actuator Faults in Vehicle Suspensions.
Control Engineering Practice, 49, 173–186. doi:
https://doi.org/10.1016/j.conengprac.2015.12.002.
Koenig, D. (2006). Observers Design for Unknown Input
Nonlinear Descriptor Systems via Convex Optimization.
IEEE Transactions on Automatic control, (06), 1047–
1052.
Morato, M.M., Sename, O., Dugard, L., and Nguyen,
M.Q. (2019). Fault Estimation for Automotive
Electro-Rheological Dampers: LPV-based Observer Ap-
proach. Control Engineering Practice, 85, 11–22. doi:
https://doi.org/10.1016/j.conengprac.2019.01.005.
Morato, M., Pham, T.P., Sename, O., and Dugard, L.
(2020). Development of a Simple ER Damper Model for
Fault-tolerant Control Design. Journal of the Brazilian
Society of Mechanical Sciences and Engineering, 42(10),
502 (2020). doi:10.1007/s40430-020-02585-y.
Nguyen, M.Q., Sename, O., and Dugard, L. (2016). Com-
parison of Observer Approaches for Actuator Fault Es-
timation in Semi-active Suspension Systems. In 2016
3rd Conference on Control and Fault-Tolerant Systems
(SysTol), 227–232. IEEE.
Osorio-Gordillo, G.L., Darouach, M., Astorga-Zaragoza,
C.M., and Boutat-Baddas, L. (2019). Generalised Dy-
namic Observer Design for Lipschitz Non-linear De-
scriptor Systems. IET Control Theory & Applications,
13(14), 2270–2280.
Pham, T.P., Sename, O., and Dugard, L. (2019). Unified
H∞ Observer for a Class of Nonlinear Lipschitz Sys-
tems: Application to a Real ER Automotive Suspen-
sion. IEEE Control Systems Letters, 3(4), 817–822. doi:
10.1109/LCSYS.2019.2919813.
Poussot-Vassal, C., Spelta, C., Sename, O., Savaresi, S.M.,
and Dugard, L. (2012). Survey and Performance Evalu-
ation on some Automotive Semi-active Suspension Con-
trol Methods: A Comparative Study on a Single-corner
Model. Annual Reviews in Control, 36(1), 148–160.
Savaresi, S.M., Poussot-Vassal, C., Spelta, C., Sename, O.,
and Dugard, L. (2010). Semi-active Suspension Control
Design for Vehicles. Elsevier.
Sename, O. (2021). Review on LPV Approaches for
Suspension Systems. Electronics, 10(17), 2120.
Tran, G.Q.B., Pham, T.P., and Sename, O. (2022).
Multi-objective Grid-based Lipschitz NLPV PI
Observer for Damper Fault Estimation. In
11th IFAC Symposium on Fault Detection,
Supervision and Safety for Technical Processes
- SAFEPROCESS 2022. Paphos, Cyprus. URL
https://hal.archives-ouvertes.fr/hal-03611561.