Robust PI observer design for hybrid

heterogeneous multiple models with

delayed measurements
Rodolfo Orjuela, Benoı̂t Marx, José Ragot
and Didier Maquin

Centre de Recherche en Automatique de Nancy

UMR 7039, Nancy-Université, CNRS
2, Avenue de la Forêt de Haye
54516 Vandœuvre-lès-Nancy, France

Abstract: This paper deals with the problem of state estimation for hybrid systems modelled
via multiple models and subject to delayed measurements. In contrast to the most popular
results found in the multiple model literature, we consider heterogeneous multiple models, also
known as decoupled multiple model, which allow to use submodels whose state space dimension
can be different. On the basis of this multiple model a proportional-integral observer (PIO) is
designed in order to cope with the robust state estimation with respect to disturbances. Sufficient
delay-independent conditions for ensuring robust performances (attenuation level) and dynamic
performances (exponential convergence) of the estimation error are provided in terms of LMIs
using the Lyapunov-Krasovskii method. The validity of the proposed methodology is illustrated
by an academic example.

Keywords: heterogeneous multiple models; state estimation; proportional-integral observer;

delayed measurements

1. INTRODUCTION nonlinear model representations must be used to cope

effectively with the global modelling of these systems. A
Context. Time-delay or dead time is frequently encoun- number of quite diverse similar modelling frameworks such
tered in several real-world processes, such as chemical as hybrid systems, switched systems, multiple models, etc.
and thermal processes, electrical and communications net- are proposed to dealing with the modelling of systems that
works, etc. Indeed, any system has an amount of time, evolves according to different dynamics in the operating
more or less important, between its expected response space.
and its actual response. In many industrial applications Tools and proposed method. The nonlinear behaviour of
time-delays must unavoidably be considered as an inher- a system can accurately be represented using the multi-
ent part of the system, and consequently, they must be ple model approach [Murray-Smith and Johansen, 1997,
accounted for in the mathematical model of the system Boukhris et al., 1999]. Multiple models have been rec-
under investigation [Smith and Corripio, 1985]. Therefore, ognized as a popular and a powerful modelling tool of
considerable attention has been paid in the past decades complex nonlinear systems. In this approach the operating
to systems with delays and appropriate theoretical tools, space of the system under investigation is decomposed
in the frequency and in the time domain, for the analysis, into a number of operating zones, each of them being
control and state estimation are well established [Bliman, characterised by a submodel of reduced dimension. The
2001, Sename, 2001, Richard, 2003]. Among them, the well interpolation of submodels with the help of weighting
known Lyapunov-Krasovskii functional method is largely functions, associated to each operating zone, allows the
adopted for stability analysis of linear or nonlinear delayed global representation of the nonlinear system. Let us notice
systems. This method provides generally sufficient condi- finally that the basis of the multiple model approach can
tions in a LMI form which can be solved in polynomial- be easily related to the hybrid system philosophy in which
time using appropriated standard convex optimization al- the interaction of a set of submodels changes according to
gorithms [Boyd et al., 1994]. a dynamic law. Indeed, smooth as well as crisp (switching)
On the other hand, complex industrial systems generally transitions between the submodels can be considered for
present a nonlinear dynamic behaviour in particular when more flexibility in modelling stage.
the whole operating range (global behaviour) must be At this point, it is important to point out that the interpo-
considered instead of a reduced operating range in the lation of submodels can be operated using many schemas
neighbourhood of an operating point (local behaviour). [Filev, 1991, Gregorcic and Lightbody, 2008], among these
This situation is commonly encountered in the state esti- realisations two schemas can be distinguished: via an in-
mation problems because all admissible trajectories of the terpolation of parameters (e.g. Piecewise Affine Systems
system must be taken into consideration. Hence, adequate
or Takagi-Sugeno multiple model ) or via an interpolation We shall simply write µi (ξ(t)) = µi (t) and x(t − τ (t)) =
of outputs (e.g. decoupled multiple model ). Many contri- x(∇) where τ (t) > 0 is a variable time-delay.
butions concern the analysis, the control and the state
estimation of the first multiple model realisation (with 2. DECOUPLED MULTIPLE MODEL
or without delays) [Murray-Smith and Johansen, 1997,
Babuska, 1998, Tanaka and Wang, 2001, Lin et al., 2007], The structure of the decoupled multiple model can be
much less are devoted to observer design using decoupled viewed as a parallel interconnection of several submodels
multiple models. In this work, the latter realisation will be via a weighted sum of their outputs [Gatzke and Doyle III,
employed for dealing with the problem of state estimation 1999]. In other words, as a particular class of hybrid
for nonlinear hybrid systems with delayed measurements heterogeneous models. By considering a state space repre-
as an extension of the ideas proposed previously [Orjuela sentation, this structure takes the following form [Orjuela
et al., 2007, 2008]. et al., 2008]:
Related work. Uppal et al. [2006] have shown a successful
implementation of this structure in order to cope with the ẋi (t) = Ai xi (t) + Bi u(t) + Di w(t) , (1a)
state estimation for fault detection and isolation. How- yi (t) = Ci xi (t) , (1b)
ever, only a note of the convergence estimation error is L
proposed in this work and neither any analytic proof of X
z(t) = µi (ξ(t))yi (t) , (1c)
the convergence of the estimation error is given, nor the
i=1
time delay is taken into account. On the other hand, the
state estimation in presence of delayed measurements is y(t) = z(t − τ (t)) + W w(t) . (1d)
investigated in Orjuela et al. [2007]. Sufficient conditions, ni p
where xi ∈ R and yi ∈ R are respectively the state
under a LMI form, are given for ensuring the state estima- vector and the output of the ith submodel; u ∈ Rm is the
tion convergence towards zero. However, the robust state multiple model input, z ∈ Rp is the multiple model output
estimation with respect to disturbances (e.g. noise) acting and y ∈ Rp the measured output subject to a variable
on the outputs has not been addressed in this work. time-delay τ (t) and a perturbation w ∈ Rr . The matrices
Ai ∈ Rni ×ni , Bi ∈ Rni ×m , Di ∈ Rni ×r , Ci ∈ Rp×ni and
We introduce a so-called proportional-integral observer
W ∈ Rp×r are known and constant.
(PIO) which provides very interesting robustness proper-
ties with respect to perturbations [Beale and Shafai, 1989, Assumption 1. The perturbation is bounded energy sig-
Weinmann, 1991, Busawon and Kabore, 2001]. This ob- nal, i.e. kw(t)k22 < ∞.
server is characterised by the use of two corrective injection
terms, proportional and integral, instead of the conven- The variable time-delay τ (t) acting on the output is
tional proportional correction frequently employed in the perfectly modelled by a continuous function that satisfies
Luenberger observer and/or in its classical extensions. The the following conditions [Wu et al., 2004]:

PIO has also been successfully employed in the synchro- 0 ≤ τ (t) ≤ τmax
(2)
nization of a chaotic system by Hua and Guan [2005]. τ̇ (t) ≤ β < 1
The extension of the PIO design, based on dissipativity where τmax and β are constant. This time-delay can be
framework, to a particular nonlinear system whose non due for example to the inherent nature of the system, the
linearity is assumed to satisfy a sector bounded constraint, sensor reactivity or the transmission data time-delay.
has been recently proposed in Jung et al. [2007]. However,
the PIO design for standard delayed linear system and for The relative contribution of each submodel to the global
delayed multiple models (Takagi-Sugeno and/or decoupled model according to the current operating point of the sys-
multiple models) seems poorly investigated in the litera- tem is quantified by the weighting functions µi (ξ(t)) (i.e.
ture. the transition law). They are associated to each operating
zone and satisfy the following convex sum properties:
Contribution and paper organisation. In this communica-
tion, the PIO design for decoupled multiple models with L
delayed measurements is investigated. The outline of this
X
µi (ξ(t)) = 1 and 0 ≤ µi (ξ(t)) ≤ 1 , ∀i = 1...L, ∀t . (3)
paper is as follows. A brief description of the decoupled i=1
multiple model is presented in section 2. The state estima-
The current operating point is taken into account by
tion problem is presented in section 3 and sufficient con-
means of the so-called decision variable ξ(t) which is a
ditions, on the basis of the Lyapunov-Krasovskii method,
characteristic variable of the system (e.g. inputs and/or
are established in order to ensure the stability of the PIO.
measured variables).
Finally, in section 4, an academic example illustrates the
state estimation of a decoupled multiple model. In this multiple model the contribution of each submodel
is taken into account via a weighted sum of the submodel
Notations: The following standard notations will be used
outputs. Consequently, heterogeneous submodels (i.e. sub-
throughout the paper: P > 0 (P < 0) denotes a positive
models of different dimensions) can be considered in the
(negative) definite matrix P ; X T denotes the transpose of
modelling stage as a supplementary degree of flexibility
matrix X; I is the identity matrix of appropriate dimension
and generality. Therefore, this structure is well suited
and diag{} is a block diagonal matrix of appropriate
for modelling nonlinear hybrid systems whose structure
dimension. The L2 −norm of a signal, quantifying its
R∞ varies with the operating regime, for example when the
energy is denoted and defined by ke(t)k22 = eT (t)e(t)dt. complexity of the dynamic behaviour is not uniform across
0 the operating space [Gregorcic and Lightbody, 2008].
Notice that the multiple model approach is an efficient way 3.2 PIO structure
to deal with hybrid systems: no assumption is made on the
continuity of the weighting functions blending the outputs In a PIO, the classic used proportional action given by the
of the submodels. Then switches between submodels can gains Ki in (4) is replaced by the use of two correction
be modelled by discontinuous weighting functions taking actions: proportional and integral. For this purpose, a
boolean values in the set [0, 1]. Rt
supplementary integral variable yI (t) = y(η)dη must
3. STATE ESTIMATION PROBLEM 0
be taken into consideration in the PIO architecture. Note
In a previous work [Orjuela et al., 2007], the state esti- however that this integration can be replaced by a more
mation of a decoupled multiple model subject to delayed general action, for example, by considering the filtered
measurements has been investigated on the basis of a output signal as follows:
proportional observer given by :
ẏI (t) = N yI (t) + y(t) (7)
x̂˙ i (t) = Ai x̂i (t) + Bi u(t) + Ki (y(t) − ŷ(t)) , (4a) where the matrix N is a fading effect coefficient matrix
L
X that regulates the transient response of yI (t) [Jung et al.,
ẑ(t) = µi (ξ(t))Ci x̂i (t) , (4b) 2007].
i=1
Let us notice that the equations (5) and (7) can be
ŷ(t) = ẑ(t − τ (t)) , (4c) gathered as follows:
where Ki ∈ Rni ×p is the gain associated to ith observer.
Sufficient conditions for ensuring the exponential con- T T T
ẋa (t) = C 1 Ã C 1 + C 2 N C 2 xa (t) + C 2 C̃(∇) C 1 xa (∇)
vergence towards zero of the estimation errors ei (t) =
xi (t) − x̂i (t) have been established. Note however that + C 1 B̃u(t) + C 2 W + C 1 D̃ w(t) , (8a)
disturbances acting on the systems are not taken into
consideration. Consequently, the sensitivity problem of the T
y(t) = C̃(∇)C 1 xa (∇) + W w(t) , (8b)
state estimation with respect to disturbances has not been T
previously addressed in this work. On the other hand, the yI (t) = C 2 xa (t) (8c)
proportional observer offers only one degree of freedom Ki
where the augmented vector xa ∈ R(n+p) and the constant
for reducing the influence of the disturbance and providing
bloc matrices C 1 and C 2 of appropriate dimensions are
at the same time good dynamic performances (two antag-
respectively given by:
onist design goals).

In this paper, the proportional observer (4) is replaced by x(t) I 0
a more general observer, known as proportional-integral xa (t) = , C 1 = (n×n) , C 2 = (n×p) . (9)
yI (t) 0(p×n) I(p×p)
observer (PIO), which provides very interesting robustness
properties with respect to perturbations. In particular, The structure of the proposed PIO can be obtained on the
good trade-off between dynamic and robust performances basis of the augmented form (8):
of the observer can be obtained with the PIO because two
degrees of freedom are available for the observer design. T T
T
x̂˙ a (t) = C 1 Ã C 1 + C 2 N C 2 x̂a (t) + C 2 C̃(∇) C 1 x̂a (∇)
3.1 Preliminaries
+K̃P (y(t) − ŷ(t)) + K̃I (yI (t) − ŷI (t)) + C 1 B̃u(t) , (10a)
T
For the simplicity of manipulations, the decoupled multi- ŷ(t) = C̃(∇)C 1 x̂a (∇) , (10b)
ple model (1) is rewritten in the following compact form T
[Orjuela et al., 2007, 2008]: ŷI (t) = C 2 x̂a (t) . (10c)
where the gains K̃P and K̃I must be designed and the
ẋ(t) = Ãx(t) + B̃u(t) + D̃w(t) , (5a) matrices Ã, B̃ and C̃(t) have been previously defined
z(t) = C̃(t)x(t) , (5b) in (6). Let us notice that an additional injection term,
given by the filtered (or integral) output estimation error
y(t) = z(∇) + W w(t) , (5c) yI (t) − ŷI (t), is included in the dynamic equation of the
where: observer. Hence some degrees of freedom are in this way
L
introduced for observer design. The use of the two gains
T K̃P (proportional action) and K̃I (integral action) is at
X
n
T T T
x(t) = x1 (t) · · · xi (t) · · · xL (t) ∈ R , n = ni , (6a)
the origin of the name Proportional-Integral Observer.
i=1

à = diag {A1 · · · Ai · · · AL } ∈ R(n×n) , (6b)

T 3.3 Problem formulation
B̃ = B1 T · · · Bi T · · · BL T ∈ R(n×m) ,

(6c)
T The dynamics of the estimation error
D̃ = D1 T · · · Di T · · · DL T ∈ R(n×r) ,

(6d)
C̃(t) = [µ1 (t)C1 · · · µi (t)Ci · · · µL (t)CL ] ∈ R(p×n) . (6e) ea (t) = xa (t) − x̂a (t) (11)
The matrix C̃(t) is time-varying because the weighting can be easily established by considering the gathered forms
functions are taken into consideration in this matrix. of the decoupled multiple model (8) and the PIO (10):
ėa (t) = Λ1 ea (t) + Λ2 (∇)ea (∇) + Λ3 w(t) , (12) functional method is employed in order to provide delay-
where independent sufficient conditions, in LMIs terms [Boyd
et al., 1994], for ensuring the robust performances (16).
T T T
Λ1 = C 1 Ã C 1 + C 2 N C 2 − KI C 2 , (13a)
T
Theorem 1. Consider the decoupled multiple model (1)
Λ2 (∇) = C 2 − KP C̃(∇) C 1 , (13b) with time-delay (2) and assumption 1. There exists a
PIO (10) ensuring the objectives (16) if there exists
Λ3 = C 2 − KP W + C 1 D̃ . (13c)
two symmetric, positive definite matrices P and Q, two
Let us remark, from equations (13a) and (13c), that matrices GP and GI and a positive scalar γ solution of
the proportional gain KP can be used to modulate the the constrained optimisation problem:
influence of the disturbances on the estimation error ea (t)
and the observer dynamics can be improved with the help min γ subject to
 T T

of the integral gain KI . H H + Λ1 P + P Λ 1 + Q Λ̃2,i Λ3
Remark 1. Note, from equation (13c), that the distur-  Λ̃T2,i −(1 − β)e−2ατmax Q 0  < 0
bance w(t) can be totally decoupled (i.e. its influence is T
Λ3 0 −γI
null) if and only if KP W = C 2 W + C 1 D̃. Notice that for i = 1 · · · L where
T
T T
D̃ KP 1 Λ1 = C 1 Ã C 1 + C 2 N C 2 + αI − GI C2 ,
= KP W = W (14)
W I T
Λ̃2,i = P C 2 − GP C̃i C 1 ,
and consequently, KP 1 = D̃W T (W W T )−1 . So, this condi-
tion cannot be easily satisfied in practice because W is not Λ3 = P C 2 − GP W + P C 1 D̃ ,
necessary of complete rank. Besides, the total disturbance C̃i = [0 · · · Ci · · · 0] .
decoupling case reduce inevitably the freedom degrees to
for a given decay rate α and for prescribed matrices H
adjust the dynamic performances.
and N . The observer gains are given by KP = P −1 GP
and KI √= P −1 GI ; the L2 gain from w(t) to ν(t) is given
The robust PIO design problem can thus be formulated as by γ = γ.
finding the matrices K̃P and K̃I such that the influence Proof. The exponential convergence of the estimation
of w(t) on the estimation error (11) is attenuated and the error (11) is investigated via a Lyapunov-Krasovskii func-
state estimation error remains globally bounded for any tional proposed by Mondié and Kharitonov [2005]:
blend between the submodels. To this end, the following
objective signal ν(t) which only depends on the estimation Z0
error ea (t) is introduced:
V (t) = eTa (t)P ea (t) + eTa (t + θ)e2αθ Qea (t + θ)dθ
ν(t) = Hea (t) , (15) −τ (t)

where H is a matrix of appropriate dimension chosen (17)

by the designer. The weighting matrix H in (15) can be where P and Q are symmetric, positive definite matrices.
selected for totally or partially taking into consideration Mondié and Kharitonov [2005] propose this functional
the transfer from w(t) to ea (t). in order to provide sufficient delay-independent condi-
tions for the exponential stability of linear systems with
Finally, the expected performances of the PIO can be constant time-delay. We use the same functional but by
formulated by the following robust performances: considering a time-varying delay. On the other hand, the
Lyapunov-Krasovskii functional (17) can be rewritten as
lim ea (t) = 0 for w(t) = 0 , (16a) follows:
t→∞
kν(t)k22 ≤ γ 2 kw(t)k22 for w(t) 6= 0 and ν(0) = 0 , (16b) T
ea (t) P 0 ea (t)
where γ is the L2 gain from w(t) to ν(t) to be minimized. V (t) = (18)
ea (∇) 0 0 ea (∇)
Notice that condition (16a) ensures convergence towards
Z0
zero of the estimation error in the perturbation free case.
On the other hand, condition (16b) guarantees attenuation + eTa (t + θ)e2αθ Qea (t + θ)dθ .
level on the estimation error with respect to perturbations. −τ (t)
In the following, we will investigate the exponential con-
vergence of the estimation error. Indeed, the exponential On the other hand, the robust state estimation objectives
convergence is a strong form of convergence that guar- (16) are guaranteed if there exists a Lyapunov-Krasovskii
antees dynamic performances of the estimation error, in functional (17) such that [Boyd et al., 1994]:
particular a convergence velocity via a decay rate.
V̇ (t) + 2αV (t) < γ 2 wT (t)w(t) − ν T (t)ν(t) (19)
3.4 Computation of the gains K̃P and K̃I where α is the so-called decay rate for convergence velocity
and γ is the attenuation level from w(t) to ν(t) for robust
This section focuses on the computation of the PIO gains estimation. So, inequality (19) must be ensured in order to
K̃P and K̃I . For this purpose, the Lyapunov-Krasovskii provide a solution of the robust state estimation problem.
Let us consider initially the time-derivative of the func- However, the time-delayed varying matrix C̃(∇) in Λ2 (∇)
tional (17) evaluated via Leibniz-Newton formula: can be rewritten as a weighted sum of constant matrices
as follows:
V̇ (t) = ėTa (t)P ea (t) + eTa (t)P ėa (t) + eTa (t)Qea (t)
L
− (1 − τ̇ (t))e−2ατ (t) eTa (∇)Qea (∇) (20)
X
C̃(∇) = µi (∇)C̃i , (26)
Z0 i=1
− 2α eTa (t + θ)e2αθ Qea (t + θ)dθ where C̃i is a constant block matrix given by:
−τ (t)

that can be upper bounded, using the a priori knowledge C̃i = [0 · · · Ci · · · 0] (27)
(2) of the time-delay as follows: such that the term Ci is found on the i block column of th

C̃i . By considering (26) then the inequality (25) becomes

V̇ (t) ≤ ėTa (t)P ea (t) + eTa (t)P ėa (t) + eTa (t)Qea (t)
L
" #
2αP +H T H+Λ̃T
− (1 − β)e−2ατmax eTa (∇)Qea (∇) (21) X 1 P +P Λ̃1 +Q P Λ̃2,i P Λ3
µi (∇) Λ̃T
2,i P −(1−β)e−2ατmax Q 0 < 0.
Z0 ΛT 0 −γ 2 I
i=1 3 P
− 2α eTa (t + θ)e2αθ Qea (t + θ)dθ , (28)
−τ (t)
where the constant matrices Λ̃2,i are given by:
which becomes finally by considering the dynamics of the
estimation error (12): T
Λ̃2,i = C 2 − KP C̃i C 1 . (29)
Z0 According to the properties (3) of the weighting functions,
V̇ (t) ≤ −2α eT (t + θ)e2αθ Qe(t + θ)dθ the above matrix inequality (28) can be satisfied by con-
sidering the simultaneous solution of the set of inequalities
−τ (t) expressed for each upper bound of µi (∇). On the other
+w (t)ΛT3 P ea (t) + eTa (t)P Λ3 w(t)
T
(22) hand, the linearisation of the matrix inequality (28) can be
h
ea (t)
iT h T
Λ1 P +P Λ1 +Q P Λ2 (∇)
ih
ea (t)
i achieved by the standard change of variables GP = P K̃P ,
+ ea (∇) ΛT −(1−β)e−2ατmax Q ea (∇) . GI = P K̃I and γ 2 = γ. Hence, sufficient conditions for
2 (∇)P
ensuring the robust performances (16) are obtained under
Hence, the left-hand side of the inequality (19) is ensured a LMI form which can be solved via classical LMI tools.
by the following inequality: The proof of theorem 1 is completed. 2
#T 4. SIMULATION EXAMPLE
ea (t)
"
2αP 0 P Λ3
V̇ + 2αV ≤ ea (∇) 0 0 0 (23)
ΛT3 P 0 0 Let us consider the state estimation of a hybrid hetero-
w(t)
" ea (t) # geneous multiple models with L = 2 submodels. The
parameters of the submodels are:
T
Λ1 P +P Λ1 +Q P Λ2 (∇) 0
+ ΛT
2 (∇)P −(1−β)e−2ατmax Q 0 ea (∇) . " #
−2 0.5 0.6
0 0 0 w(t)
h i
−0.2 −0.6
A1 = −0.3 −0.9 0 , A2 = ,
0.3 −1
Now, by considering (23) then the inequality (19) can be −1.3 0.6 −0.8
guaranteed if the following inequality holds:
B1 = 1 0.2 0.5
T
,

B2 = −0.5 0.8
T
,
T T
#T D1 = 0.1 0.2 −0.3 , D2 = −0.1 0.1 ,
ea (t)
"
ΛT

1 P +P Λ1 +Q P Λ2 (∇) 0
ea (∇)
h i h i
T
Λ2 (∇)P −(1−β)e −2ατ max Q 0 1 0.8 0.5 0.7 0.3
C1 = , C2 = ,
w(t) 0 0 0 0.2 0.3 −0.5 0.2 −0.5
T
ea (t) W = 0.2 −0.3 .
" #
2αP +H T H 0 P Λ3
+ 0 0 0 ea (∇) < 0 . (24) The time-delay function appearing in the output of the
ΛT
3 P 0 −γ 2 I
w(t) systems is given by τ (t) = 0.5 + 0.45 sin(0.5t) with τmax =
Let us notice that the above inequality is a quadratic form 0.95. The upper bound of its derivative is β = 0.225. Here,
T
in [ ea (t) ea (∇) w(t) ] . Consequently, the inequality (19) for the decision variable ξ(t) is the input signal u(t) ∈ [0, 1].
ensuring robust objectives (16) is finally guaranteed if the The weighting functions are obtained from normalised
following inequality holds: Gaussian functions:
" T T
# L
2αP +H H+Λ1 P +P Λ1 +Q P Λ2 (∇) P Λ3 X
ΛT −(1−β)e−2ατmax Q 0 <0 . µi (ξ(t)) = ωi (ξ(t))/ ωj (ξ(t)) , (30)
2 (∇)P
ΛT 0 −γ 2 I j=1
3 P

2
(25) ωi (ξ(t)) = exp −(ξ(t) − ci ) /σ 2 , (31)
Note that a time-varying matrix Λ2 (∇), given by (13b), with σ = 0.4, c1 = 0.3 and c2 = 0.7.
appears in this inequality and its resolution cannot be
computed using standard LMI algorithms. A solution that satisfies the theorem 1, is given by:
K̃P = 0.075 0.144 −0.220 −0.074 0.075 0.999 0.002 T

, S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Lin-
−0.237 −0.444 0.678 0.225 −0.238 0.002 0.992
ear Matrix Inequalities in System and Control Theory.
0.139 −0.068 −0.054 0.046 2.269 −0.045 T
K̃I = −0.431

−0.141 −0.434 −0.114 −0.022 0.011 0.042 2.252 , SIAM Studies in Applied Mathematics 15, Philadelphia,
2 1994.
with a decay rate α = 0.2 and an attenuation level γ = 0.5 K.K. Busawon and P. Kabore. Disturbance attenuation
and for H = I and N = −0.3. using proportional integral observers. International
In the simulation the disturbance acting on the output is a Journal of Control, 74(74):618–627, 2001.
normally distributed random signal. The state estimation D. Filev. Fuzzy modeling of complex systems. Interna-
error is plotted in figure 1 (left) where ei is the ith com- tional Journal of Approximate Reasoning, 5(3):281–290,
ponent of the estimation error vector (11). The measured 1991.
and the estimated outputs are shown in figure 1 (right). E. P. Gatzke and F. J. Doyle III. Multiple model approach
In both cases, the error around the origin time is due to for CSTR control. In 14th IFAC World Congress,
the differences between initial conditions of the multiple Beijing, China, 1999.
model and the observer. G. Gregorcic and G. Lightbody. Nonlinear system iden-
tification: From multiple-model networks to gaussian
1
e1 (t)
1.5
y(t)
processes. Engineering Applications of Artificial Intelli-
0.8
e2 (t)
1
ŷ(t) gence, 21(7):1035–1055, 2008.
0.6 e3 (t)
0.4 e4 (t) C. Hua and X. Guan. Synchronization of chaotic systems
0.2 e5 (t) 0.5
based on PI observer design. Physics Letters A, 334
0

−0.2
0 (5-6):382–389, 2005.
−0.4 −0.5
J. Jung, K. Huh, and T. Shim. Dissipative proportional
−0.6 integral observer for a class of uncertain nonlinear
−0.8
0 5 10 15 20
−1
0 5 10 15 20 systems. In American Control Conference, ACC, New
time (s) time (s)
York City, USA, 2007.
Figure 1. State estimation errors (left). Actual and esti- C. Lin, Q.G. Wang, T.H. Lee, and Y. He. LMI Approach to
mated outputs (right). Analysis and Control of Takagi-Sugeno Fuzzy Systems
with Time Delay. Springer-Verlag Berlin Heidelberg,
2007.
5. CONCLUSION S. Mondié and V.L. Kharitonov. Exponential estimates
for retarded time-delay systems: An LMI approach.
A proportional-integral observer (PIO), on the basis of a IEEE Transactions on Automatic Control, 50(2):268–
decoupled multiple model, is proposed in this contribution 273, 2005.
in order to cope with the state estimation problem of a R. Murray-Smith and T.A. Johansen. Multiple model
nonlinear hybrid system in presence of delayed and per- Approaches to Modelling and Control. Taylor & Francis,
turbed measurements. The decoupled multiple model is an London, 1997.
interesting structure for modelling nonlinear systems with R. Orjuela, B. Marx, D. Maquin, and J. Ragot. A de-
variable structure because the dimensions of the employed coupled multiple model approach for state estimation of
submodels can be different. On the other hand, the PIO nonlinear systems subject to delayed measurements. In
offers more degree of freedom for robust state estimation 3rd IFAC Advanced Fuzzy and Neural Network Work-
with respect to the classic proportional observer previously shop, AFNC, Valenciennes, France, 2007.
proposed. The robust stability problem of the estimation R. Orjuela, B. Marx, D. Maquin, and J. Ragot. State
error is investigated using the Lyapunov-Krasovskii func- estimation for nonlinear systems using a decoupled
tional method and delay-independent sufficient conditions, multiple mode. International Journal of Modelling
under a LMI form, are established. Identification and Control, 4(1):59–67, 2008.
J.P. Richard. Time-delay systems: an overview of some
Further research will be to investigate the state estimation recent advances and open problems. Automatica, 39(10):
by considering distributed delays in the submodel dynamic 1667–1694, 2003.
equations. O. Sename. New trends in design of observers for time-
delay systems. Kybernetika, 37(4):427–458, 2001.
REFERENCES C.A. Smith and A.B. Corripio. Principles and practice of
automatic process control. John Wiley & Sons, 1985.
R. Babuska. Fuzzy Modeling for Control. Kluwer Aca- K. Tanaka and H. Wang. Fuzzy Control Systems design
demic Publishers, London, 1998. and analysis. John Wiley & Sons, New York, 2001.
S. Beale and B. Shafai. Robust control system design with F.J. Uppal, R.J. Patton, and M. Witczak. A neuro-fuzzy
a proportional integral observer. International Journal multiple-model observer approach to robust fault diag-
of Control, 50(1):97–111, 1989. nosis based on the DAMADICS benchmark problem.
P. Bliman. LMI characterization of the strong delay- Control Engineering Practice, 14(6):699–717, 2006.
independent stability of linear delay systems via A. Weinmann. Uncertain Models and Robust Control.
quadratic Lyapunov-Krasovskii functionals. Systems Springer-Verlag Wien New York, Vienne, 1991.
and Control Letters, 43(4):263–274, 2001. M. Wu, Y. He, J.H. She, and G.P. Liu. Delay-dependent
A. Boukhris, G. Mourot, and J. Ragot. Non-linear criteria for robust stability of time-varying delay sys-
dynamic system identification: a multiple-model ap- tems. Automatica, 40(8):1435–1439, 2004.
proach. International Journal of Control, 72(7/8):591–
604, 1999.