Design of observers for Takagi –Sugeno descriptor

systems with unknown inputs and application to

fault diagnosis
B. Marx, D. Koenig and J. Ragot

Abstract: A method for state-estimation of Takagi – Sugeno descriptor systems (TSDS) affected by
unknown inputs (UI) has been presented here. For ease of implementation’s sake, the proposed
observers are not in descriptor form, but in usual form. Sufficient existence conditions of the
unknown input observers (UIOs) are given and strict linear matrix inequalities are solved to deter-
mine the gain of the observers. If the perfect UI decoupling is not possible, the UIO is designed in
order to minimise the L2-gain from the UI to the state estimation error. The two previous objectives
can be mixed in order to decouple the estimation to a subset of the UI, while attenuating the L2-gain
from the other UI to the estimation. The proposed UI observers are used for robust fault diagnosis.
Fault diagnosis for TSDS is performed by designing a bank of observers. A simple decision logic
and thresholds setting allow to determine the occurring fault. The results are established for both
the continuous and the discrete time cases. The proposed method is illustrated by a numerical
example.

1 Introduction Robust output feedback, and H1 -control are considered for

TSDS in [8] and [9], respectively. The study of TSDS is envi-
The Takagi – Sugeno (TS) model proposed by Takagi and saged with interval methods in [10], in order to account the
Sugeno [1] is a well-known structure to represent nonlinear different operating points. Unfortunately, the problem of
systems into several linear fuzzy models. In the last two observer design, and especially the design of unknown input
decades, the control and the observation of TS systems observers (UIOs), has resulted in very few works.
have become challenging problems that received a con- The design of UIO is a crucial problem since, in many
siderable amount of attention. In [2], stability analysis and practical cases, all input signals cannot be known.
controller design are addressed, solutions are derived in Moreover, this class of observers is widely used in the
the linear matrix inequality (LMI) formalism. Relaxed suffi- area of fault diagnosis, even if all the inputs are known
cient conditions for fuzzy controllers and fuzzy observers (see Chapter 3 in [11]). The design of UIO has received
are proposed in [3, 4] via a multiple Lyapunov function considerable attention in the case of usual (in opposition to
approach. descriptor) linear systems [12], descriptor systems [13 – 15]
The descriptor formalism is very attractive for system or TS systems [16]. Unfortunately, to the best of the
modelling, as pointed out in [5], since it describes a wider authors’ knowledge, the design of UIO has not been treated
class of systems including physical systems with non- in the generic case of TSDS. The aim of this paper is not
dynamic constraints (e.g. algebraic relations induced in only to generalise the existing works on UIO design to
interconnected systems such as power transfer networks TSDS, but also to apply this new observer in the field of
or water distribution networks) or jump behaviour. The fault diagnosis of TSDS, which has not been treated so far.
enhancement of the modelling ability is because of the This paper gives a simple extension to TSDS of the
structure of the dynamic equation, which encompasses not design of observers for the state estimation in the presence
only dynamic equations, but also algebraic relations. of UI. Under some sufficient conditions, the design of
Since both TS and descriptor formalisms are attractive in the the observer is reduced to the determination of a matrix.
field of modelling, thus the TS representation has been gener- The choice of this parameter is performed by solving
alised to descriptor systems. The stability and the design of strict LMIs. If the estimation error cannot be decoupled
state-feedback controllers for TS descriptor systems (TSDS) from the UI, an L2 observer is proposed to minimise the
are characterised via LMI in [6, 7], in particular, the influence of the UI on the state estimation. The two
problem of nonlinear model following is treated in [7]. design objectives can be mixed by decoupling the state esti-
mation from a subset of the UI, and minimising the L2-gain
between the other UI and the state estimation error. The
# The Institution of Engineering and Technology 2007
designed observers are used for fault diagnosis, since the
doi:10.1049/iet-cta:20060412
UI can encompass the faults and the disturbances affecting
Paper first received 28th September 2006 and in revised form 5th March 2007
the system. Designing several observers attenuating the dis-
B. Marx and J. Ragot are with the Centre de Recherche en Automatique de turbance effect, and decoupling the estimation from all
Nancy UMR 7039, CNRS—Nancy Université, 2 avenue de la forêt de Haye,
54516 Vandoeuvre-lès-Nancy, Cedex, France faults, but one lead to the well-known generalised observer
D. Koenig is with the Laboratoire d’Automatique de Grenoble UMR 5528 scheme (GOS) for fault diagnosis [11]. The design of obser-
CNRS-INPG-UJF, BP 46, 38402 Saint Martin d’Hères, Cedex, France vers is detailed both in the continuous time case and in the
E-mail: [email protected] discrete time case.
IET Control Theory Appl., 2007, 1, (5), pp. 1487 –1495 1487
 The paper is organised as follows: the class of studied as unknown in order to isolate the default corrupting this
systems is defined in Section 2 and the main results about particular signal (see Chapter 3 in [17]).
UIO design are detailed in Section 3. Firstly, the definition In this study, the proposed observers are not in descriptor
of the UIO and the sufficient existence condition are estab- form, in order to reduce the implementation complexity. In
lished. Secondly, the computation of the gains of the obser- the continuous-time case, the proposed multiple UIO is
ver is established. The design of L2 observers is treated in defined by
Section 4. Section 5 deals with the design of observers for
both UI decoupling and disturbance attenuation. The appli- X
r
z_ (t) ¼ hi (w(t))(Ni z(t) þ Mi u(t) þ Li y(t)) (5)
cation to fault diagnosis is studied in Section 6. Section 7 is
i¼1
devoted to a numerical example.
x^ (t) ¼ z(t) þ T2 y(t) (6)
2 Takagi –Sugeno descriptor systems
In the discrete-time case, the proposed multiple UIO is
To begin with, the class of systems considered in this defined by
paper is described. In the continuous time case, a TSDS is X
r
defined by zkþ1 ¼ hi (wk )(Ni zk þ Mi uk þ Li yk ) (7)
i¼1
X
r
E_x(t) ¼ hi (w(t))(Ai x(t) þ Bi u(t) þ Di d(t)) (1) x^ k ¼ zk þ T2 yk (8)
i¼1

y(t) ¼ Cx(t) þ Gd(t) (2) The problem of unknown input decoupling observer
(UIDO) design is to find the gains of the UIDO (5 and 6)
In the discrete time case, a TSDS is defined by [resp. (7 and 8)], namely Ni , Mi , Li and T2 , in order that
the estimated state x^ asymptotically tends to the state of
X
r (1 and 2) [resp. (3 and 4)]. In other words, the objective is
Exkþ1 ¼ hi (wk )(Ai xk þ Bi uk þ Di d k ) (3) that the estimation error defined by e(t) ¼ x(t)  x^ (t)
i¼1 [resp. ek ¼ xk  x^ k ) tends to zero when t ! 1 (resp.
yk ¼ Cxk þ Gd k (4) when k ! 1), regardless of the unknown input the
control input and the initial state.
where x [ Rn is the state variable, u [ Rnu is the control Firstly, a sufficient rank condition for UI decoupling is
input, d [ Rq is the unknown input (disturbance, actuator given in Lemma 1. Secondly, a sufficient LMI condition
noise or hidden message in the recovering framework) for the convergence of the continuous-time UIO is given
and y [ Rm is the measured output. The matrices in Lemma 2 (it is extended to the discrete-time case in
E, Ai , Bi , Di , C and G are supposed to be real, known, con- Corollary 1). Finally, the results are gathered in Theorem
stant and with appropriate dimensions according to the defi- 1 and a design algorithm is given.
nition of the signals. The matrix E may be singular. The
activating functions, denoted hi (w(t)), for i ¼ 1, . . . , r, are Lemma 1: There exists a continuous-time (resp. discrete-
normalised, and satisfy the following constraints time) UIDO (7 and 8) for (1 and 2) [resp. (7 and 8) for (3
and 4)] if the following condition holds
X
r


0
hi (w(t))
1, hi (w(t)) ¼ 1, 8t D    Dr
i¼1
rank X ¼ rank 1 þ n þ rank G (9)
Ir  G
X
r
0
hi (wk )
1, hi (wk ) ¼ 1, 8k Proof: The estimation error e(t) ¼ x(t)  x^ (t) is given by
i¼1
e(t) ¼ x(t)  x^ (t)
The decision variable w(t) (or wk ) is supposed to be real-
time accessible, depending on the control input or on the ¼ x(t)  z(t)  T2 Cx(t)  T2 Gd(t)
measured output. Assume that the same matrix E appears
in all the different sub-models is not restrictive if we con- Assume that there exist T1 and T2 such that, the following
sider that the structure of the differential or algebraic equations hold
relations is imposed by the physical structure of the T1 E þ T2 C ¼ In (10)
system, which generally does not change with time. This T2 G ¼ 0 (11)
formalism still encompasses the varying parameters or the With (10) and (11), the estimation error becomes
nonlinearities since the matrices Ai are different one from e(t) ¼ T1 Ex(t)  z(t). Its time derivative is given by
another, and since the activating functions introduce the
nonlinear dynamics. An analogous argument justifies the e_ (t) ¼ T1 E_x(t)  z_ (t)
single nature of the output matrix C. The available measure-
ments are determined by the location and the nature of the X
r

sensors, which generally do not change (the sensors are ¼ hi (w(t))[T1 (Ai x(t) þ Bi u(t) þ Di d(t))
i¼1
not removed during the operating time).
 Ni z(t)  Mi u(t)  Li y(t)]
3 Design of UI decoupling observers X
r
¼ hi (w(t))[Ni e(t)
In this section, our aim is to design a multiple UIO. The UIO i¼1
is used widely in the field of fault detection and isolation for þ (T1 Ai  Ni T1 E  Li C)x(t)
dynamic systems, because the fault signals are generally
unknown. Moreover, a measured signal can be considered þ (T1 Bi  Mi )u(t) þ (T1 Di  Li G)d(t)] (12)

1488 IET Control Theory Appl., Vol. 1, No. 5, September 2007

The time derivative of the estimation error is given by verifying the following LMI for i ¼ 1, . . . , r
X
r
(YXþ Yi )T P þ PYXþ Yi þ (X? Yi )T ZT þ ZX? Yi , 0 (31)
e_ (t) ¼ hi (w(t))Ni e(t) (13)
i¼1 where  is the Kronecker product. The matrices X and Y
if the following constraints hold for i ¼ 1, . . . , r are defined by (29) and (30), respectively, and
Yi [ R(nþm(rþ1))n are defined by
In ¼ T 1 E þ T 2 C (14) 2 3
Ai
0 ¼ T2 G (15) 6 7
Yi ¼ 4 0mn 5
0 ¼ T1 Ai  Ni T1 E  Li C (16)
ei  C
0 ¼ T1 Bi  Mi (17)
r1
0 ¼ T 1 Di  L i G (18) where ei [ R is the column vector with all its com-
ponents equal to 0, except the ith equal to 1.
In order to find the gains of the UIDO, according to the con- Proof: Suppose that (9) is satisfied, then (25) is solvable and
straints (14 – 18), new parameters Ki ¼ Ni T2  Li are intro- the solutions Q are given by
duced in (16). Then, the UIDO exists if, for i ¼ 1, . . . , r, the
following statements are true Q ¼ YXþ þ ZX? (32)
N i ¼ T 1 A i þ Ki C (19) where Z [ Rn(nþm(rþ1)) is an arbitrary matrix.
In ¼ T 1 E þ T 2 C (20) With (26) and (32), the matrices Ni are defined by
Ni ¼ YXþ Yi þ ZX? Yi . The state estimation error tends to
0 ¼ T2 G (21) zero if the polytopic system (13) is stable. A well-known stab-
0 ¼ T 1 D i þ Ki G (22) ility condition for polytopic system (see [18]) is the existence
Mi ¼ T1 Bi (23) of a symmetric positive definite matrix P verifying
NTi P þ PNi , 0 for i ¼ 1, . . . , r. Then the UIDO provides
Li ¼ Ni T2  Ki (24) an estimate of the system state if there exists a matrix Z such
Verifying the constraints (19 – 22) reduces to finding that P(YXþ Yi þ ZX? Yi ) þ (YXþ Yi þ ZX? Yi )T P , 0, for
Q [ Rn(nþm(rþ1)) such that all i ¼ 1, . . . , r. Setting Z ¼ PZ then (31) is obtained,
which completes the proof. A
QX ¼ Y (25) This result is extended to the discrete-time case.
Ni ¼ QYi (26)
Corollary 1: The estimation error of the UIO (7 and 8) for
where Q, is given by (3 and 4) tends to zero if there exists a symmetric positive
Q ¼ [T1 T2 jK1 K2  Kr ] (27) definite matrix P [ Rnn and a matrix Z [ Rn(nþm(rþ1))
verifying the following LMI for i ¼ 1, . . . , r
Once Q is known, Mi and Li are deduced from (23) and

(34), respectively. Equation (25) is solvable in the variable Fi (X? Yi )T ZT
,0 (33)
Q if the following condition holds Z(X? Yi ) P


X where Fi is defined by
rank ¼ rank X (28)
Y
Fi ¼ (X? Yi )T ZT (YXþ Yi ) þ (YXþ Yi )T Z(X? Yi )
where the matrices X [ R(nþm(rþ1))(nþq(rþ1)) and
Y [ Rn(nþq(rþ1)) are defined by þ (YXþ Yi )T P(YXþ Yi )  P
2  3
E 0nq  D1    Dr
 Proof: Proof is similar to the continuous-time case, apart
X¼6
6 C G  0mq . . . 0mq 7
7 (29) from the condition for the stability of the state estimation
4  5 error. In the discrete-time case, ek tends to zeros if there

0rmn 0rmq  Ir  G exists a symmetric positive definite matrix P such that
 NTi PNi  P . 0 for i ¼ 1, . . . , r. With PZ ¼ Z, the LMI
Y ¼ [In 0nq 0nrq ] (30) (33) follows. A
Obviously, with (30) and (29), the condition (28) becomes


 Theorem 1: There exists a continuous-time (resp. discrete-
X D    Dr time) UIO (5 and 6) for (1 and 2) [resp (7 and 8) for
rank ¼ n þ rank G þ rank 1 (3 and 4)] if the condition (9) is satisfied and if there
Y Ir  G
2  3 exists a symmetric positive definite matrix P [ Rnn and
E 0  D1    Dr a matrix Z [ Rn(nþm(rþ1)) verifying (31) [resp. (33)], for
6C G 0  0 7 i ¼ 1, . . . , r.
rank X ¼ rank6  7
4  5

0 0  I G Finally, the design of UIO for continuous-time (resp.
r
discrete-time) TSDS is reduced to the following procedure.
Then (28) is equivalent to (9). In the discrete-time case, the
proof is very similar, thus it is omitted. A Step 1. Verify the existence condition (9).
Step 2. Solve the LMI (31) [resp. (33)] in P and Z.
Lemma 2: The estimation error of the UIO (5 and 6) for Step 3. Compute Z with Z ¼ P1 Z. For a given Z, Q is
(1 and 2) tends to zero if there exists a symmetric positive deduced from (32), the matrices Ni , Mi and Li are derived
definite matrix P [ Rnn and a matrix Z [ Rn(nþm(rþ1)) from (19), (23) and (24), respectively.
IET Control Theory Appl., Vol. 1, No. 5, September 2007 1489
 This result unifies the results obtained, on the one hand, " #
in the field of the descriptor systems with UI [19 – 21] ATi PAi þ CTi Ci  P ATi PBi
and, on the other hand, in the field of the TS systems with ,0 (40)
BTi PAi BTi PBi  g2 I
UI [16]. It is useful because a TSDS cannot be reduced,
either to a single singular system (it would not handle the For a given real-positive g, an observer is said to be an
nonlinearities because of the weighting functions hi) nor unknown input attenuating observer (UIAO) of L2-gain g,
to a regular TS system (it would not handle the algebraic if the state estimation error, e, and the unknown input, d,
relation between the state variables). The existence con- satisfy kek2 , gkdk2 .
dition (9) can be linked to previous works concerning
single descriptor systems [20, 21]. Considering (9) for a Theorem 2: There exists a UIAO (51 and 61), with an
single descriptor system, would lead to the condition L2-gain lower than g, for the system (1 and 2), if the con-
2 3 dition (41) is satisfied, and if there exist a symmetric posi-
E D1 0

D tive definite matrix P [ Rnn and matrices Z [ Rn(nþm)
rank4 C 0 G 5 ¼ rank 1 þ n þ rank G (34) and Ki [ Rnm , verifying the LMI (42) for i ¼ 1, . . . , r.
G
0 G 0

E 0
One can note that (34) is equivalent to the condition (21) or rank ¼ n þ rank G (41)
C G
(31) of [20], and is also equivalent to the condition (A3a) " #
in [21]. Moreover, the present paper gives only sufficient Ci,1 Ci,2
conditions, whereas [20] gave necessary and sufficient con- ,0 (42)
CTi,2 g2 Iq
ditions. This difference appears because the present paper is
basically written for TS systems, thus the weighting
P func- where the matrices Ci,1 , and Ci,2 are given by
tions cause conservatism since the matrix ri¼1 hi (w(t))Ai
can take all the possible values in the polytope defined by Ci,1 ¼ PYXþ ?
1 Ai þ ZX1 Ai þ Ki C
its vertices Ai .
T
þ (PYXþ ?
1 Ai þ ZX1 Ai þ Ki C) þ In
4 Design of UIAOs
Ci,2 ¼ PYXþ ?
1 Di þ ZX1 Di þ Ki G
In this section, the aim is to design an observer for TSDS in (nþq)n (nþq)m (nþm)n
order to minimise the influence of the UI on the state esti- where Xþ
1 [R , Xþ
2 [R , X?
1 [R and
? (nþm)m
mation when the perfect decoupling is not possible. The X2 [ R are defined by
chosen criterion to minimise is the L2-gain between the

unknown input and the state estimation error. This approach E 0 þ  þ 
¼ X1 Xþ 2
is less restrictive than the design of a UIO since the structural C G
condition (9) is partially relaxed.

þ
As pointed out in the section of UIO design, the esti- E 0 E 0  
 Inþm ¼ X? X?
mation error e is governed by a non-singular TS system C G C G 1 2
(12), thus in order to bound the L2-gain from the UI to e,
and establish the sufficient conditions of the so-called Proof: If (41) is satisfied, then there exist T1 and T2 such
L2-observer, the following lemma concerning L2-gain of that
TS-systems is needed.
[T1 T2 ]X ¼ Y
Lemma 3 [18]: Consider the continuous-time TS-system where X and Y are given by
defined by

E 0
X
r X¼ , Y ¼ [ In 0nm ]
x_ (t) ¼ hi (w(t))(Ai x(t) þ Bi u(t)) (35) C G
i¼1
and, for any arbitrary matrix Z, T1 and T2 are given by
X
r
y(t) ¼ hi (w(t))Ci x(t) (36) T1 ¼ YXþ ?
(43)
1 þ ZX1
i¼1
T2 ¼ YXþ
2 þ ZX?
2 (44)
and the discrete-time TS-system defined by
X
r Following the proof of Theorem 1, if (19), (23) and (24)
xkþ1 ¼ hi (wk )(Ai xk þ Bi uk ) (37) hold, the state estimation error e(t) ¼ x(t)  x^ (t) is
i¼1 governed by
X
r X
r
yk ¼ hi (wk )Ci xk (38) e_ (t) ¼ hi (w(t))((T1 Ai þ Ki C)e(t) þ (T1 Di þ Ki G)d(t))
i¼1 i¼1

The system (35 and 36) [resp. (37 and 38) is stable and veri- (45)
fies kyk2 , gkuk2 if there exists a symmetric positive defi-
nite matrix P [ Rnn such that (39) [resp. (40)] is satisfied According to Lemma 3, ke(t)k2 , gkd(t)k2 if there exists a
for i ¼ 1, . . . , r. symmetric positive definite matrix P such that the following
LMI hold for i ¼ 1, . . . , r
" #

ATi P þ PAi þ CTi Ci PBi Ci þ In PT1 Di þ PKi G
,0 (39) ,0
BTi P g 2 I DTi TT1 P þ GT KTi P g2 Iq

1490 IET Control Theory Appl., Vol. 1, No. 5, September 2007

where Ci is given by The designs of UIDO and UIAO are combined to derive
T
the design of a UI decoupling/attenuating observer
Ci ¼ (T1 Ai þ Ki C) P þ P(T1 Ai þ Ki C) (UIDAO). The sufficient existence conditions are given in
With Ki ¼ PKi and Z ¼ PZ, the LMI (42) follows, which the following theorem.
completes the proof. A Theorem 3: There exists an observer (5 and 6) ensuring
perfect decoupling to d(t) and maximally robust to d(t) if
Remark 1: Obviously, condition (41) is less restrictive
condition (49) is satisfied and if there exist a symmetric
than (9).
positive definite matrix P [ Rnn and a matrix
Remark 2: g2 can be considered as a variable to be mini- Z [ Rn(nþ(rþ1)m) solution of the minimisation of g under
mised during the LMI optimisation, to obtain an optimal the LMI constraint (50) for i ¼ 1, . . . , r.
UI attenuation

D1    Dr
rank X ¼ n þ rank ¼ [G G] þ rank (49)
Corollary 2: There exists a UIAO (7 and 8) with an L2-gain Ir  G
lower than a given real positive g for the system (3 and 4), if " #
Ci,1 Ci,2
the condition (41) is satisfied, and if there exist a symmetric T ,0 (50)
positive definite matrix P [ Rnn , and matrices Ci,2 g2 I
Z [ Rn(nþm) and Ki [ Rnm , verifying the following
LMI for i ¼ 1, . . . , r where Ci,1 and Ci,2 are given by
2 3 þ ?
In  P 0 FTi,1 Ci,1 ¼ PYX Yi þ ZX Yi
6 7
4 0 g2 In FTi,2 5 , 0 (46) þ ?
þ (YX Yi )T P þ (X Yi )T Z þ In
T

Fi,1 Fi,2 P
þ ?
Ci,2 ¼ PYX Yi þ ZX Yi
where Fi,1 and Fi,2 are defined by
þ ?
where X is the pseudo-inverse of X, X ¼ I  XXþ , and
Fi,1 ¼ PYXþ ?
1 Ai þ ZX1 Ai þ Ki C where X, Y, Yi and Yi are given by
2  3
Fi,2 ¼ PYXþ ?
1 Di þ ZX1 Di þ Ki G E 0nq 0n
q  D1  Dr
6  7
X¼4 C G G  0mq    0mq 5
Proof: The proof follows the lines of the proof of Theorem 
0 0 0 
2 with a Schur complement and is therefore omitted. A rmn rmq I
rm
q r G

Finally the design of UIAO for continuous-time (resp. Y ¼ [In 0nqþq j0nrq ]
discrete-time) TSDS is reduced to the following procedure. 2 3 2 3
Ai Di
Step 1. Verify the existence condition (41). 6 7 6 7
Yi ¼ 4 0mn 5, Yi ¼ 4 0mq 5
Step 2. Solve the LMI (42) [resp. (46)] in P, Z and Ki .
Step 3. Compute Z and Ki with Z ¼ P1 Z and Ki ¼ P1 Ki , ei  C ei  G
respectively. The matrices T1 and T2 are obtained by (43)
and (44). The matrices Ni , Mi and Li are derived from Proof: The state estimation error e(t) is governed by the
(19), (23) and (24), respectively. following system
X
r
5 Design of unknown input decoupling and e_ (t) ¼ hi (w(t))(Ni e(t) þ (T1 Ai  Ni T1 E  Li C)x(t)
attenuating observer i¼1

þ (T1 Bi  Mi )u(t) þ (T1 Di  Li G)d(t)

If the UIs are too numerous or if their distribution structure
makes the perfect UI decoupling of the estimation imposs- _
_ þ T2 Gd(t))
þ (T1 Di  Li G)d(t) þ T2 Gd(t) (52)
ible [i.e. if the structural condition (9) is not satisfied] a
compromise can be made in order to design an observer Following a similar argument to that in the proof of
ensuring two complementary objectives with less restrictive Theorem 1, if the observer parameters satisfy the constraints
existence conditions. Firstly, the state estimation is per- (19 – 24) and T2 G ¼ 0, e(t) is governed by
fectly decoupled to a subset of the UI denoted d(t).
Secondly, the L2-gain between the other UIs, denoted X
r


d(t), to the state estimation error is minimised, thus the e_ (t) ¼ hi (w(t))(Ni e(t) þ (T1 Di þ Ki G)d(t)) (53)
i¼1
state estimation is made maximally robust to these UI.

the system (1 and 2)
Partitioning the UI into d(t) and d(t) These constraints can be written as QX ¼ Y, with Q
can be written as defined by (27). This equation can be solved if and only if
X r
E_x(t) ¼ hi (w(t))(Ai x(t) þ Bi u(t) rank[ XT YT ]T ¼ rank X, which is equivalent to the
i¼1 condition (49). If condition (49) is satisfied, then, for any
arbitrary matrix Z, u is given by
þ Di d(t) þ Di d(t)) (47)
þ ?
Q ¼ YX þ ZX (54)
y(t) ¼ Cx(t) þ Gd(t) þ Gd(t) (48)
Then, the only parameter to be found is the matrix Z. Since
where d(t) [ Rq and d(t) [ Rq . The partition of the UI into the matrices Ni and (T1 Di  Li G) can be written as

is such that the perfect decoupling condition is
d(t) and d(t)
satisfied for (E, C, G, D1 , . . . , Dr ), then the L2-gain from þ ?
d(t) to the state estimation error is minimised. Ni ¼ QYi ¼ YX Yi þ ZX Yi

IET Control Theory Appl., Vol. 1, No. 5, September 2007 1491

 þ ?
f k [ Rq and disturbances wk [ Rq is defined by
T1 Di þ Ki G ¼ QYi ¼ YX Yi þ ZX Yi
X
r
Exkþ1 ¼ hi (wk )(Ai xk þ Bi uk þ Dfi f k þ Dwi wk )
the stability condition of (53) follows the same lines as in i¼1
the proof of Theorem 2. Rewriting the stability condition
(39) for the triplet (Ni , (T1 Di  Ki G), In ), and setting yk ¼ Cxk þ Gf f k þ Gw wk
PZ ¼ Z, the LMI condition (50) follows. A
It is assumed that each component of the disturbance vector
Remark 3: Obviously, condition (49) is less restrictive than is bounded, and that the value of this bound is known :
(9). To obtain perfect decoupling to all the UI of (47 and jwi (t)j , ni (resp jwik j , ni in the discrete-time case) for
48), Di and G should be replaced by [ Di Di ] and i ¼ 1, . . . , q) for all t resp. for all k). The well-known
[ G G ], respectively, in (49) and (51), which would lead GOS [11] can be applied to propose a method for the
to a more restrictive existence condition. fault diagnosis of TSDS. In this approach, q UIDAO
A similar result can be given for discrete-time systems are designed. The ‘th UIDAO is designed by considering
defined by the ‘th fault as a UI. A subset, denoted S‘ , of the disturb-
ances can also be considered as UI provided the existence
X
r   condition (49) is satisfied. In other words, the ‘th observer
Exkþ1 ¼ hi (wk ) Ai xk þ Bi uk þ Di d k þ Di d k (55) is designed for the system (47 and 48) [resp. (55 and 56)],
i¼1
with
yk ¼ Cxk þ Gd k þ Gd k (56) h i  
Di ¼ D‘fi jDjwi , j [ S‘ , Di ¼ Djwi , j [ S‘
h i  
Corollary 3: There exists an observer (7 and 8) ensuring
G ¼ G‘f jGjw , j [ S‘ , G ¼ Gjw , j [ S‘
perfect decoupling to d k and maximally robust to d k if con-
dition (49) is satisfied and if there exist a symmetric positive
definite matrix P [ Rnn and a matrix Z [ Rn(nþ(rþ1)m) where M ‘ denotes the ‘th column of the matrix M, and S‘
solution to the minimisation of g under the following LMI denotes the complementary to S‘ in {1,2, . . . , q}.
constraint for i ¼ 1, . . . , r. As a consequence, the output estimation of the ‘th obser-
ver will be sensitive to all the faults but the ‘th, insensitive
2 3
In  P 0 FTi,1 to the ‘th fault and to a subset of the disturbances, denoted
6 7 S‘ , and maximally robust to the other disturbances belong-
4 0 g2 I FTi,2 5 , 0 (57) ing to the subset S
‘ . The subsets of the UI are determined so
Fi,1 Fi,2 P that the decoupling condition (49) is satisfied for all the dis-
turbances in S‘ and so that the L2-gain from the disturb-
where Fi,1 and Fi,2 are given by ances in S
‘ to the output estimation error is minimised. In
þ ?
other words, the output estimation error is a residual
Fi,1 ¼ PYX Yi þ ZX Yi signal. A classical method for observer-based fault diagno-
þ ?
sis is to suppose the occurrence of the ‘th fault if all residual
Fi,2 ¼ PYX Yi þ ZX Yi signals, except the ‘th, are significantly different from zero.
The problem is then to quantify the term significantly. In
Proof: The proof follows the lines of the proof of Corollary order to discriminate between the influence of the disturb-
3 and Theorem 3, and is therefore omitted. A ances and the ‘th fault, one can compute the L2-gain from
the disturbances to each output estimation error, as
Finally the design of the observer for continuous-time described in the following procedure.
(resp. discrete-time) TSDS is reduced to the following
procedure. † For each fault f‘(t) (resp. f‘k in the discrete-time case)
Design the UIDAO, sensitive to all faults except f ‘ (t) (resp.
Step 1. Verify the existence condition (49). f ‘k ), insensitive to wi (t) (resp. wik ) i [ S‘ and maximally
Step 2. Solve the LMI (50) [resp. (57)] in P and Z. robust to wi (t) (resp. wik ), i [ S‘ .
Step 3. Compute Z with Z ¼ P1 Z. For a given Z, Q is † Compute the norm-bound of the attenuated disturbances,
given by (54), then the matrices T1 , T2 and Ki are obtained. denoted r‘
The matrices Ni , Mi and Li are derived from (19), (23) and sffiffiffiffiffiffiffiffiffiffiffiffi
X ffi
(24), respectively. r‘ ¼ n2i
i[S‘
6 Application to fault diagnosis
† For each component of the output yj (t) (resp. yjk ), compute
In this section, the UI decoupling and attenuating observers the L2-gain, denoted g‘j , from the attenuated disturbances
are used to perform fault diagnosis. Consider a continuous- {wi (t)ji [ S‘ } (resp. {wik ji [ S‘ }) to the jth output esti-
time TSDS affected by faults f (t) [ Rq and disturbances mation error and compute the boolean vector b‘ (t) ¼ [ b‘1 (t)
w(t) [ Rq defined by b‘2 (t)    b‘m (t)] (resp. b‘k ¼ [b‘1k b‘2k  b‘mk ]),
where b‘j (t) (resp. b‘jk ) is defined by
X
r
E_x(t) ¼ hi (w(t))(Ai x(t) þ Bi u(t) 1, if j^y‘j (t)  yj (t)j . ag‘j r‘
i¼1 b‘j (t) ¼
0, else
þ Dfi f (t) þ Dwi w(t))
1, if j^y‘jk  yjk j . ag‘j r‘
y(t) ¼ Cx(t) þ Gf f (t) þ Gw w(t) b‘jk ¼
0, else
In the discrete-time case, the TSDS affected by faults where y^ ‘j (t) (resp. y^ ‘jk ) is the jth component of the estimated
1492 IET Control Theory Appl., Vol. 1, No. 5, September 2007
output given by the ‘th observer. The positive scalar a allows 0 1
5 0
the designer to handle the compromise between non-detection B0
and false alarm (e.g. considering the accuracy of the model). B 0CC
B1 ¼ Df 1 ¼B C
† Compute the alarm a‘ (t) (resp. a‘k ), affected to f ‘ (t) @0 4A
(resp. f ‘k ), defined in the continuous-time case by 0 0
8
< 1, if (bi (t)bTi (t)  1, 8i = ‘) 0 1
a‘ (t) ¼ & (b‘ (t)bT‘ (t) ¼ 0) 0:1 0:2 0:4 0:9
: B 0:2
0, else B 0:6 0:2 0:7 CC,
A2 ¼ @
0:5 0:7 0:7 0:6 A
or in the discrete-time case by 0:7 0:4 0:4 3:6
0 1 0 1
6 0 0 0
1, if(bik bTik  1, 8i = ‘) & (b‘k bT‘k ¼ 0) B0 0C B 0:8 0C
a‘k ¼ B2 ¼ Df 2 ¼ B C B C,
0, else @ 0 3 A Dw1 ¼ @ 0 0A
0 0 1 0
This approach can be conservative, since the only avail- 0 1
able information about the disturbances is their amplitude 0 0
B1 0 C 1 1 1 0
bound and the L2-gain of their influence onto the output Dw2 ¼ B C
@ 0 0:1 A C ¼ 0 1 0
estimation error, thus it may imply non-detection. This 1
effect can be limited by the use of the parameter a, which 1 0
can be adjusted according to measurements of the system
under healthy operation. Nevertheless, comparing each One can notice that the subsystem (E, A1 ) is impulsive. The
component of the estimation error with a threshold makes finite spectrum of the two subsystems is significantly differ-
it possible to avoid distributing a significant error affecting ent since we have sf (E, A1 ) ¼ {0:080, 0:564} and
a component on all the various components, and then sf (E, A2 ) ¼ {0:935, 0:126, 0:996}, thus the global
reduces the non-detection. system is not close to a linear system. The sampling time
One should note that the GOS is an efficient structure of is ts ¼ 0.03 s. The activating functions hi (wk ) are defined
diagnosis in order to detect and isolate single faults. In the by h1k ¼ (1 þ tanh(u1k =10))=2 and h2k ¼ 1  h1k . The dis-
case of simultaneous faults, two faults may cause non-zerol turbances w1k and w2k are bounded centred white noise,
residue responses in all observers. In this case, the dedicated with norm bound n1 ¼ n2 ¼ 1. The fault signals represent
observer scheme can be considered as an alternative, but the control input dysfunctions and they are defined by
decoupling conditions become much more restrictive since
all the fault inputs, but the ‘th have to be decoupled from the 0:8 u1k , if 35
tk
40
f1k ¼
‘th residue. This scheme is not detailed here, but can readily 0, else
be applied since it suffices to change the definition of the
matrices Di , Di , G and G: the ‘th observer should be 0:8 u2k , if 40
tk
45
f2k ¼
designed for the system (47 and 48) [resp. (55 and 56)], with 0, else
h i
The first UIADO is designed with the first control input as
Di ¼ D1fi    D‘1
fi D‘þ1
fi  Dqfi j Djwi , j [ S‘
UI, the L2-gain from w to the first and second output
estimation error are g11 ¼ 0:098 and g12 ¼ 0:050, respect-
ively. The second UIADO is designed with the second
h i
control input and the second disturbance as UI,
G ¼ G1fi    G‘1
fi G‘þ1
fi    Gqfi j Gjw , j [ S‘ e
the L2-gain from w1 to the first and second output
estimation error are g21 ¼ 0:052 and g22 ¼ 0:009,
    respectively.
Di ¼ Djwi , j [ S‘ , G ¼ Gjw , j [ S‘

7 Design example

In this section, the proposed approach for fault diagnosis is

illustrated. Consider a discrete-time TSDS defined by

X
2
Exkþ1 ¼ hi (wk )(Ai xk þ Bi uk þ Dfi f k þ Dwi wk )
i¼1

yk ¼ Cxk

with E ¼ diag 1 1 1 0 and

0 1
0:5 0:5 0:2 0:2
B 0:9 0:1 0:4 0:7 C
B C
A1 ¼ B C,
@ 0:2 0:7 0 0 A
0:2 0:4 0:4 0 Fig. 1 Inputs and activating functions of the simulated system
IET Control Theory Appl., Vol. 1, No. 5, September 2007 1493
Fig. 2 Original and estimated x1k and x2k obtained with an Fig. 5 Original and estimated x3k and x4k obtained with an
observer decoupling the first input observer decoupling the second input

Fig. 3 Original and estimated x3k and x4k obtained with an Fig. 6 Output estimation errors obtained with an observer
observer decoupling the first input decoupling the first input

Fig. 4 Original and estimated x1k and x2k obtained with an Fig. 7 Output estimation errors obtained with an observer
observer decoupling the second input decoupling the second input

In Fig. 1, the inputs and the activating functions are dis- estimation, whereas the estimation is sensitive to the
played. Figs. 2 and 3 display the comparison of the state fault f2k present between tk ¼ 40 s and tk ¼ 45 s. Figs. 4
variables and their estimates supplied by the UIDAO and 5 display the comparison of the state variables and
insensitive to the first fault. The fault f1k appearing their estimates supplied by the UIDAO insensitive to the
between tk ¼ 35 s and tk ¼ 40 s does not affect the fault f2k and affected by f1k . The residual signals and
1494 IET Control Theory Appl., Vol. 1, No. 5, September 2007
their corresponding threshold, for a ¼ 1, are displayed in 3 Tanaka, K., Ikeda, T., and Wang, H.O.: ‘Fuzzy regulators and fuzzy
Figs. 6 and 7. observers: relaxed stability conditions and lmi-based designs’, IEEE
Trans. Fuzzy Syst., 1998, 6, pp. 250–265
The residual signals computed with the output estimation 4 Tanaka, T., Hori, T., and Wang, H.O.: ‘A multiple Lyapunov function
error of the first observer are sensitive to f2k and insensitive approach to stabilization of fuzzy control systems’, IEEE Trans. Fuzzy
to f1k whereas the residual signals computed with second Syst., 2003, 11, pp. 582–589
UIADO are sensitive to f1 and insensitive to f2 . The 5 Dai, L.: ‘Singular control systems’ (Springer, Berlin, 1989)
6 Taniguchi, T., Tanaka, K., Yamafuji, K., and Wang, H.O.: ‘Fuzzy
L2-gains g11 , g12 , g21 and g22 are good thresholds for descriptor systems : stability analysis and design via LMIs’. Proc.
fault isolation, with a ¼ 1, since the fault f2k (occurring Amer. Control Conf., San Diego, USA, 1999, vol. 3, pp. 1827– 1831
for 40
tk
45) is isolated at t ¼ 41.5 s, and the fault f1k 7 Taniguchi, T., Tanaka, K., and Wang, H.O.: ‘Fuzzy descriptor systems
(occurring for 35
tk
40) is isolated at t ¼ 35.1 s. and nonlinear model following control’, IEEE Trans. Fuzzy Syst.,
The sudden appearance or disappearance of a fault may 2000, 8, pp. 442– 452
8 Lian, K.Y., Liu, P., Liou, J.J., and Wu, T.C.: ‘Robust output feedback
cause abrupt changes of the state variables that the estimate control for fuzzy descriptor systems’. Proc. IEEE Int. Fuzzy Syst.
cannot follow instantaneously. Thus, even if the estimation Conf., Melbourne, Australia, 2001, pp. 904–907
is decoupled from the occurring fault, a residual signal may 9 Yoneyama, J., and Ichikawa, I.: ‘H1-control for takagi-sugeno fuzzy
transiently be higher than the threshold. This phenomenon descriptor systems’. Proc. IEEE Int. Conf. Syst. Man Cybern,
Tokyo, Japan, 1999, vol. 3, pp. 28– 33
appears in Fig. 7 at t ¼ 45 when f2k disappears and causes 10 Wang, Y., Zhang, Q.L., and Liu, W.Q.: ‘Stability analysis and design
a brief overshot of the output estimation error. for T-S fuzzy descriptor systems’. Proc. IEEE Conf. Decision Control,
Orlando, USA, 2001, vol. 4, pp. 3962–3967
8 Conclusion 11 Patton, R.J., Frank, P., and Clark, R.: ‘Fault diagnosis in dynamic
systems - theory and application’ (Prentice-Hall, London, 1989)
12 Darouach, M., Zasadzinski, M., and Xu, S.J.: ‘Full-order observers for
In this paper, a simple method is proposed to design UIO for linear systems with unknown inputs’, IEEE Trans. Autom. Control,
TSDS. Sufficient existence conditions were given, and the 1994, 39, pp. 606 –609
determination of the observer parameters is based on 13 Darouach, M., and Boutayeb, M.: ‘Design of observers for descriptor
solving a system of strict LMI. If the unknown input decou- systems’, IEEE Trans. Autom. Control, 1995, 40, pp. 1323– 1327
pling condition is not satisfied, it has been proposed to 14 Kawaji, S., and Kim, H.S.: ‘Full order observer for linear descriptor
systems with unknown-inputs’. Proc. Conf. on Decision and
design an L2-observer in order to minimise the L2-gain Control, New Orleans, USA, 1995, pp. 2366–2368
from the UI to the estimated state. A compromise 15 Koenig, D., and Mammar, S.: ‘Design of proportional integral
between perfect unknown input decoupling and unknown observer for unknown input descriptor systems’, IEEE Trans.
input attenuation can be made to design observer ensuring Autom. Control, 2002, 47, pp. 2057–2062
16 Akhenak, A., Chadli, M., Ragot, J., and Maquin, D.: ‘State estimation
perfect decoupling face to a subset of the unknown inputs, via multiple observer with unknown input. Application to the
and robustness face to the other unknown inputs. The three-tank system’. Proc. 5th IFAC SAFEPROCESS, Washington,
three observer designs are treated in both continuous and USA, 2003, pp. 245–251
discrete-time cases. The proposed observers are used to 17 Chen, J., and Patton, R.J.: ‘Robust model-based fault diagnosis for
perform fault diagnosis. Designing a bank of observers, dynamical systems’ (Kluwer Academic Publishers, London, 2003)
18 Boyd, S., El Ghaoui, L., Feron, E., and Balakrishnan, V.: ‘Linear
where each observer considers a fault as a UI, the GOS matrix inequalities in system and control theory’ (SIAM Studies in
can be extended to TSDS. Applied Mathematics, 1994)
19 Darouach, M., Zasadzinski, M., and Hayar, M.: ‘Reduced-order
9 Reference observer design for descriptor systems with unknown inputs’, IEEE
Trans. Autom. Control, 1996, 41, pp. 1068– 1072
1 Takagi, T., and Sugeno, M.: ‘Fuzzy identification of systems and its 20 Koenig, D.: ‘Unknown input proportional multiple-integral observer
application to modelling and control’, IEEE Trans. Syst., Man design for linear descriptor systems: application to state and fault
Cybern., 1985, 15, pp. 116 –132 estimation’, IEEE Trans. Autom. Control, 2005, 50, pp. 212–217
2 Wang, H.O., Tanaka, K., and Griffin, M.F.: ‘An approach to fuzzy 21 Koenig, D.: ‘Observer design for unknown input nonlinear descriptor
control of nonlinear systems : stability and design issues’, IEEE systems via convex optimization’, IEEE Trans. Autom. Control, 2006,
Trans. Fuzzy Syst., 1996, 4, pp. 14–23 51, pp. 1047–1052

IET Control Theory Appl., Vol. 1, No. 5, September 2007 1495