Automatica 37 (2001) 687}699

Analytical redundancy relations for fault detection and isolation

in algebraic dynamic systems夽
M. Staroswiecki*, G. Comtet-Varga
LAIL-CNRS UPRESA 8021, EUDIL, Universite& des Sciences et Technologies de Lille, 59655 Villeneuve d'Ascq Cedex, France
Received 28 July 1998; revised 31 August 2000; received in "nal form 26 October 2000

Abstract

This paper extends the analytical redundancy techniques, well developed for linear systems, to FDI in non-linear dynamic systems
modeled by polynomial di!erential algebraic equations. The residual evaluation form is developed for sensor, actuator and process
component faults. Fault detectability and isolability conditions are given, and the design of robust structured residuals is ad-
dressed.  2001 Elsevier Science Ltd. All rights reserved.

Keywords: Fault detection and isolation; Non-linear dynamic systems

1. Introduction are considered, but the approach obviously extends to

Most work on fault detection and isolation in non-
linear systems is based on non-linear observers (Seliger
& Frank, 1991; Hammouri, Kinnaert, & Yaagoubi, 1999;
Yu & Shields, 1995). Parity space methods have only
been investigated for linear (Chow & Willsky, 1984),
bi-linear (Yu, Williams, Shields, & Gomm, 1995) or state
a$ne (Comtet-Varga, Cassar, & Staroswiecki, 1997) sys-
tems. The aim of this paper is to extend the parity space
approach to non-linear systems, focusing on the issues of
robustness, fault detectability and isolability.
Analytical redundancy relations (ARR) have to be sen-
sitive to faults and insensitive to perturbations. The ef-
fects of di!erent faults have to be di!erent. This can be
achieved by designing robust structured residuals. Con-
siderable attention has been paid to this problem for
linear systems, including optimization approaches when
only approximate properties can be obtained (Patton,
1994). In this paper, ARR design is discussed for the class
of algebraic dynamic systems. Continuous-time models
夽
This paper was not presented at any IFAC meeting. This paper was
recommended for publication in revised form by Associate Editor
M. Basseville under the direction of Editor Torsten SoK derstroK m.
* Corresponding author. Tel.: #33-3-20-33-71-90; fax: #33-3-20-
33-71-89.
E-mail address: marcel.staroswiecki@univ-lille1.fr
(M. Staroswiecki).
discrete-time models.
The paper is organized as follows. Section 2 recalls
some results on analytical redundancy relations for poly-
nomial dynamic systems. Section 3 concerns fault detec-
tion and isolation. The residual evaluation form is
developed for sensor, actuator or component faults, and
the issues of fault detectability and fault isolability are
discussed. Section 4 addresses the design of robust
structured residuals. An induction motor example is
given in Section 5.
2. Analytical redundancy relations in algebraic dynamic
systems
We consider continuous-time, time-invariant systems
described by
x
"f (x, z, ), (1)
y"h(x, z, ), (2)
where x3RL is the state vector, z3RK is the input vector,
y3RN is the output vector, 3RJ is a constant parameter
vector; z"(u, v, , ) where u is the control vector, v is
some unknown input vector,  is some fault vector and
 is some stochastic vector (v"0 means no perturbation,
"0 means no fault and "0 means no noise). The
components of f and h are polynomial functions over R of
their arguments. The following notation will be used:

0005-1098/01/$ - see front matter  2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 0 5 - 1 0 9 8 ( 0 1 ) 0 0 0 0 5 - X
688 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699

being some vector,
M I is the vector whose components Corollary 1. rH3N, rH4n such that
are
and its time derivatives up to order k;
M stands for


s#1, s(rH,

and its time derivatives up to some (unspeci"ed) order. rH "
Q rH, s5rH.
2.1. Existence of analytical redundancy relations
Therefore, ARR exist only if s5rH (obviously,
Analytical redundancy relations (ARR) involve only s#1!rH "0 when s(rH).
Q
y and z . For linear systems, the unknown state is elimi-
nated from (1), (2) using the parity space approach. Remark 1. In Fliess (1989) rH is the algebraic dimension
Projection techniques can also be used for state-a$ne of the single output system (1), (3) i.e. the dimension of the
systems which are polynomial in the input (Comtet- minimal state according to realization theory.
Varga, Cassar, & Staroswiecki, 1997). Such a class obvi-
ously includes bi-linear systems. In the case of algebraic Remark 2. Corollary 1 is the analogue of the Cayley}
dynamic systems, elimination theory provides the most Hamilton theorem for non-linear algebraic systems.
straightforward and understandable tool for state elim-
ination. Note that ARR are input}output system descrip- Remark 3. The ARR of a polynomial system form a dif-
tions (although in some cases output}output relations ferential ideal.
exist, e.g. when redundancy is present among the system
sensors). Many authors have addressed their existence 2.1.2. Inter-redundancy
and computation (Fliess, 1989; Glad, 1989; Diop, 1991), ARR which involve several outputs are called inter-
their use for the analysis of structural properties (Diop redundancy relations when they cannot be inferred from
& Wang, 1993; Ljung & Glad, 1994) and some recent auto-redundancy relations.
papers have considered fault detection and isolation
issues (Guernez, Cassar, & Staroswiecki, 1997; Zhang, Theorem 2. Let (s , s ,2, s )3NN with s (rH ∀j3J
  N H
Basseville, & Benveniste, 1998; Shumsky, 2000). This and let r   2 N be the generic rank of
Q  Q   Q 
section recalls some results about the existence of ARR [J (G ) 2 J (GNN )]. Then there are exactly
V Q V Q
for polynomial dynamic systems.
"N s #p!r   2 N independent ARR involving
H H Q  Q   Q 
y Q ,2, y QN ,  and z in the form
 N
2.1.1. Auto-redundancy
w (y Q ,2, y QN , z , )"0, k"1,2,
(5)
For the jth output, j3J"1,2, p, Eq. (2) is I  N
with the following rank property:
y "h (x, z, ).

 
(3)
H H


w
From (1) and (3) one has y
"(h /x)f#(h /z)z
and it  (y , z , )
H H H y
can be proven that for all ( j, s)3J;N there exists
gH a polynomial function over R such that
Q rank "

yQ"gH (x, z Q, ). Let gH "h and GH"((gH )(gH )

H Q  H Q  


2 (gHQH )). Let rHQ be the generic rank of the Jacobian w
E (y , z , )
matrix J (GH ). y
V Q
for almost any y "(y Q ,2,y QN ), z and  which satisfy (5).
Theorem 1. For any xxed s3N there are exactly  N
s#1!rH ARR: Corollary 2. ∀(s ,2, s )3NN with s (rH ∀j3J,
Q  N H
w (y Q, z Q, )"0, k"1,2, s#1!rH


(4) N
I H Q r  2 N 4Min  s #p, n .
which are independent in the following sense: the Jacobian Q   Q  H
H
matrix When equality holds, the choice N s #p"n#1
H H

 


w minimizes the highest order of derivatives for each out-
 (y Q, z Q, ) put. Only one ARR exists.
y Q H
H
J  H (=)" 2.2. Computational issues
W


w 2.2.1. (Ow-line) state elimination
Q>\PHQ (y Q, z Q, )
y Q H Computing the ARR amounts to eliminating the state
H
in the system:
is of rank s#1!rH for almost any vector (y Q,z Q,) which
Q H
satisxes (4). y QH "GHH (x, z QH , ), j"1,2, p. (6)
Q
 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699
There are, basically, three techniques. All three require
the components of the state to be eliminated according to
some selected order.
Elimination theory rests on Euclidean division and deri-
vation. Since (6) contains no derivative of the variables to
be eliminated, only Euclidean division is used for ARR
computation. Details on elimination theory may be
found in Diop (1991) and Seidenberg (1956).
Gro( bner bases rest on Buchberger's algorithm (Buch-
berger, 1985) which uses Euclidean division and the com-
putation of the so-called S-polynomials. Details and
de"nitions can be found in Cox, Little, and O'Shea
(1992). For an application of GroK bner bases to the calcu-
lation of ARR for discrete-time polynomial systems, see
Guernez, Cassar, & Staroswiecki (1997).
Characteristic sets (also called Ritt's algorithm) rest on
Euclidean division and derivation. The state is directly
eliminated from system (1), (2), and ARR with minimum
derivative order can be obtained. The reader is referred
to Glad (1989) for an application to the calculation of
input}output relations and to Zhang, Basseville, and
Benveniste (1998) for an application to FDI.
2.2.2. (On-line) estimation of derivatives
ARR link the system inputs and outputs and their
derivatives. In the next section, they will be used to build
non-linear parity-space residuals. Because of the presence
of noise, it is di$cult in practice to compute the signal
derivatives, and one could argue that observer-based
approaches would not present this drawback. However,
when discrete-time models are available, the previous
developments still hold and lead to similar results, with
derivatives replaced by time-shifted variables. Note that
a discrete-time model is indeed used in observer-based
approaches, since the practical software used for integra-
tion actually rests on time discretization.
Evaluating derivatives from noisy signals can be done
using either observers or speci"c algorithms. Non-linear
observers can obviously be used to feed parity space
residuals with derivative estimates. Speci"c algorithms
are extensively reported in the literature (Fox & Parker,
1968; David, 1975; Dierckx, 1993). In the induction mo-
tor example of Section 5, the non-linear parity space
approach has been implemented using a polynomial ap-
proximation-based scheme, which has proved to work
satisfactorily.
Even when signal derivatives are estimated by non-
linear observers, parity space residuals still deserve some
interest. On the one hand, although the structuring of
observer-based residuals is considered in Hammouri,
Kinnaert, and El Yaagoubi (1999) for additive faults,
using parity residuals provides degrees of freedom (espe-
cially for the structuring w.r.t. parameters) which are easy
to obtain and understand using elimination theory (see
Section 4). On the other hand, ARR are a good tool for
the analysis of fault detectability conditions, which are
689
the concern of the next part of this paper. In the linear
case, parity space and observer-based approaches have
been shown to be equivalent (Magni & Mouyon, 1994)
and this result has been extended to non-linear state-
a$ne systems (KaboreH , Staroswiecki, & Wang, 1999) and
to a class of non-linear systems (Cocquempot & Chris-
tophe, 2000). For general non-linear systems such a result
is not available, but it can be conjectured that faults
which are found to be non-detectable through the ARR
analysis will also be non-detectable using observer-based
instead of ARR-based residuals.
3. Fault detection and isolation
3.1. ARR decomposition
Consider a set of ARR w(y , z , )"0. Remember that
z"(u, v, , ) and w(y , z , ) is a polynomial in the com-
ponents of z . The following decomposition holds:
w(y , u , v , M ,  , )"w (y , u , v , M , )!w (y , u , v , M ,  , ), (7)
 
where w (y , u , v , M , )"w (y , u , )!w (y , u , M , )!
  
w (y , u , v , M , ) is the deterministic part of w(y , u , v , M ,  , ) (a

polynomial of degree zero in the components of  ), w is

the stochastic part (a polynomial of degree at least one in
some component of  ). w (y , u , )!w (y , u , M , ) is the
 
perturbation-free part of w(y , u , v , M ,  , ) and w (y , u , ) is

the perturbation and fault-free part. Note that
w (y , u , 0, )"0, w (y , u , 0, M , )"0,
  (8)
w (y , u , v , M , 0, )"0.

ARR insensitivity w.r.t. unknown inputs results from
w (y , u , v , M , ) being the null polynomial and

(w /v )(y ,u ,v ,M , ,)"0. ARR might enjoy this property

either because there is no unknown input or because they
have been decoupled w.r.t. the unknown inputs (this is
a highly desirable property, and Section 4 addresses this
issue). From now on, only ARR insensitive w.r.t. un-
known inputs are considered. They are of the form
w(y , u , M ,  , )"w (y , u , )

!w (y , u , M , )!w (y , u , M ,  , )"0
 
and the following is always true:
w (y , u , )"w (y , u , M , )#w (y , u , M ,  , ). (9)
  
3.1.1. Residual computation form
Let "w (y , u , ).  is called the residual, and

w (y , u , ) is the residual computation form. It involves

only known variables and thus it can be computed on-
line. From (8), in the absence of noise and faults,  is
identically zero for any triple (y , u , ) which satis"es the
state and measurement equations (1), (2).
690 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699

3.1.2. Residual evaluation form
In real life, the behavior of  is unknown (because of )
and stochastic (because of ). The right-hand side of (9)
explains why  might be di!erent from zero. It is called
pansions of polynomials are exact, valid and "nite
(Lelong-Ferrand & Armandiès, 1978), and since
w (y , u , )"0, the residual evaluation form is




the residual evaluation form, and w (y , u , M , ) is its deter- A 1 w(y , u , ) I
 "  fM "w (y , u , fM ,). (10)
ministic part. k! y W  W
I

3.2. D-detectability 3.3.2. Actuator faults

A simple model of a faulty actuator is that of the
Fault detection procedures have to decide between non-faulty one driven by some unknown signal. Let
the two hypotheses H : "0 (healthy system) and u "u#f , where u is the (known) theoretical actuator
S

H : O0 (faulty system). For linear systems and additive input and f is some unknown input deviation. More
S

faults, the residual vector is "w (M , )#w ( , ) and realistic approaches would introduce parameterized ac-
 
w ( , ) is linear w.r.t.  , so that, assuming E( )"0, hy- tuator models in the state equations and consider system
 component faults described by parameter deviations (see
potheses H and H are characterized by
  next section).
H : "w ( , ) and E()"0,
  Actuator faults result in both state and output devi-
H : "w (M , )#w ( , ) and E()"w (M , ).
    ations. Let y be the actual output. Using the notation
Non-detectable faults are such that the residual vector y "y #y , the evaluation form is
exhibits the same stochastic behavior under both hy-
"w (y , u , )"w (y #y , u , )"w (y , u , y , ). (11)
potheses, thus: w (M , )"0 in spite of O0. In the non-   

linear situation "w (y , u , 0,  , ) under H , since Note that in (10) f represents the sensor faults, which are
  W
w (y , u , 0, )"0 from (8). However, no simple character- independent exogenous variables, while in (11) the inde-

ization of the residual's stochastic behavior can be pendent exogenous variables are f , and y represents the
S
provided since w (y , u , 0,  , ) is non-linear w.r.t.  . e!ect of actuator faults on the system outputs.

Note that in general E()"0 is not true. The faulty Because of the chosen fault model, the following re-
situation H is characterized by "w (y , u ,M , )# dundancy relations hold:
 
w (y , u , M ,  , ) which shows that the fault  in#uences both
 w (y , u , )"0 (12)
the deterministic and the stochastic parts of the residual 
vector. De"ning non-detectable faults as above seems and using Taylor expansion, one obtains
thus to be most impractical, and a stronger de"nition,
namely D-detectability, is proposed by analogy with the w (y , u , )!w (y , u , )"P(y , fM )"0, (13)
  S
linear case.
where P is a polynomial vector in the components of y
and fM , whose coe$cients are polynomials in the variables
De5nition. A D-detectable fault is such that the deter- S
ministic part of its residual evaluation form di!ers from
zero.
D-detectability is indeed a weaker concept, since non-
D-detectable faults might, nevertheless, in#uence the
stochastic part of the residual in such a way that they
could be detected.
(y , u , ), with the fundamental property fM "0Ny "0.
S
3.3.3. System component faults
System component faults are modeled as parameter
deviations  "#f where  is the actual parameter
F
vector, and  is its nominal value. This model only
considers component faults which leave the system struc-
ture unchanged. Parametric faults produce deviations in
3.3. The residual evaluation form the state: x "x#x and in the outputs: y "y#y,
and the residual computation and evaluation forms are
"w (y , u , )"w (y , u , y , ). Again, y is the output
In this section, di!erent fault hypotheses are con-  
sidered. Faults  are modeled by additive signals on the deviation which results from the parameter deviations,
inputs ( f ), the outputs ( f ) or the system parameters ( f ). and since
S W F
Since D-detectability is investigated, only deterministic w (y , u ,  )"0, (14)
systems are considered. 
then
3.3.1. Sensor faults w (y , u ,  )!w (y , u , )"Q(y , f )"0, (15)
Let y be the output associated with healthy sensors   F
and let y be the actual output. f "y !y is the fault where Q is a polynomial in the components of y and
W
model. Based on the actual outputs, the residual compu- f whose coe$cients are polynomials in (y , u , ) with the
F
tation form produces "w (y , u , ). Since Taylor ex- fundamental property f "0Ny "0.
 F
 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699 691

3.4. Fault sensitivities Then Z (y , u , )(y )" f O0 s.t. f "0 and necessary
W W WY
conditions for f (resp. f ) to belong to Z (y , u , )(y ) (resp.
S F S
For each fault, the deterministic part of the residual Z (y , u , )(y )) are
F
evaluation form is a polynomial in the deviation vector.
P(0, y , fM )"0 (resp. Q(0, y , fM )"0) ∀y .
A uni"ed representation is S F
"w (y , u , ) computation form, 3.4.2. Residual sensitivity
 (16)
"w (y , u , , y ) D-evaluation form Residual sensitivity is concerned with the transfer be-
 tween y and the D-evaluation form. Divide the residual
with vector  into two sub-vectors  and . The set of output
deviations to which  is insensitive is de"ned by
y!f "0 for sensor faults, (17a)
W
(y , u , )( )"y O0 s.t. w (y , u , , y )"0,
P(y , fM )"0 for actuator faults, (17b) 
S where w (y , u , , y ) is the D-evaluation form of the sub-
Q(y , f )"0 for component faults 
F
(17c) residual  .
and properties:
3.4.3. Structured residuals
w (y , u , , 0)"0, (18a) Consider the residual  and let S (i) (resp. S (i), S (i))
 G W S F
be the subset of the sensors (resp. the actuators, the
f ,0 N y "0, (18b)
W system components) that appear in its computation form.
f ,0 N y "0, Let y (resp. u ,  ) be the corresponding input (resp.
S
(18c) G G G
output, parameter) sub-vectors. S(i)"S (i)S (i)S (i) is
W S F
f ,0 N y "0. (18d) residual  's structure (Gertler & Luo, 1989). One has
F G
The presence of a fault is expected to deviate the D- 
G (y , u , ),0 ∀y , S (i),
evaluation form from zero. The analysis of D-detectabil- y I W
I
ity is concerned with two cases in which no such devi-

ation exists, namely output insensitivity and residual G (y , u , ),0 ∀u , S (i),
insensitivity. Output insensitive faults are charac- u I S
I
terized by y "0, so that, by (18a) the D-evaluation

form remains equal to zero in spite of the fault. On the G (y , u , ),0 ∀ , S (i).
 I F
contrary, residual insensitivity is characterized by I
w (y , u , , y )"0 in spite of y being non-zero. It is postulated that structured residuals are insensitive to

faults which do not belong to their structure. This is quite
3.4.1. Output sensitivity obvious for sensor faults in the linear case, and has been
Consider (17), and de"ne the following three sets: widely used in the DOS and GOS observer schemes
(Frank, 1993). For general non-linear systems, this prop-
Z (y , u , )" f O0 s.t. y "0,
W W erty has been used in the structural analysis approach
Z (y , u , )" f O0 s.t. P(y , fM )"0Ny "0, (Cocquempot, Cassar, & Staroswiecki, 1991) but not
S S S demonstrated. Though the proof is very simple, the prop-
Z (y , u , )" f O0 s.t. Q(y , f )"0Ny "0, erty cannot be directly derived from (16) and (17) except
F F F
for sensor faults since for any subset of faulty sensors,
Z (y , u , ) is obviously empty (sensor faults are always
W none of which belongs to S (i), (17a) gives y "f G ,0.
output sensitive), while Z (y , u , ) (resp. Z (y , u , )) are the W G W
S W Consider now a subset of faulty actuators none of
sets of output insensitive actuator faults (resp. the set of
which belongs to S (i). In this case, f G ,0 but (17b) does
output insensitive component faults). A necessary condi- S S
not imply y ,0. In fact,
tion for a fault f (resp. for a fault f ) to belong to G
S F
Z (y , u , ) (resp. to Z (y , u , )) is that it satis"es the di!er- P (y ,0)"0,
S F G G
ential equations P(0, fM )"0 (resp. Q(0, f )"0). P (y , fM H )"0, jOi
S F H H S
More generally, divide the outputs into two subsets
y"((y ) (y)) and de"ne output sensitivity with respect shows that such faulty actuators might, nevertheless,
to y : lead to y O0 (indeed, one might have jOi s.t.
G
S ( j)S (i)O and y O0). However, Eq. (12), which
Z (y , u , )(y )" f O0 s.t. y "0, W W H
W W is of the form w (y #y , u ,  )"0, shows that the
G G G G G
Z (y , u , )(y )" f O0 s.t. P(y ,fM )"0Ny "0, D-evaluation form of  would still be zero. When
S S S G
component faults are considered, similar conclusions
Z (y , u , )(y )" f O0 s.t. Q(y , f )"0Ny "0. are derived from (14).
F F F
692 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699

3.5. D-detectability condition 4.1. Robustness with respect to unknown inputs

For simplicity, only homogeneous subsets of faults The deterministic state and measurement equations
(sensor, actuator or component) are considered. Let (1), (2) are
 stand for y, u or  and consider a subset of faults such
x
"f (x, u, v, , ), (19)
that f "( f  f  ), f O0, f  ,0. A necessary and su$-
J JY JY JY J
cient condition for the subset of faults f to be D-detect-
JY y"h(x, u, v, , ). (20)
able is
The goal is to design ARR that do not depend on v, i.e.
(1) y s.t. ( f  0),Z (y , u , )(y ), for some "xed multi-index (s , s ,2, s )3NN, both x and
JY J   N
(2) y , (y , u , )(), v Q, where s"max s , are to be eliminated from system
G G
where y is a subset of sensors and y is the output (6).
deviation which results from the fault vector f . The "rst
JY Theorem 3. Let r夹  2 N "rank[J  (G ) 2
condition expresses that there exists at least one output Q  Q   Q  V T Q
J  (GNN )], then there are exactly
"N s #p!
which is sensitive to the faults f , while the second condi-
JY 夹V T Q G G
r   2 N independent ARR involving y Q ,2, y QN , u , M
tion expresses that at least one residual is sensitive to the Q  Q   Q   N
resulting output deviation. and  in the form
D-detectability conditions de"ne the set of faults which w (y Q ,2, y QN , u , M , )"0, k"1,2,
(21)
cannot be detected since they have no e!ect on the outputs I  N
or on the residuals. The residual evaluation form is indeed with the following rank property:

 


a non-linear function of the system trajectories and faults, w
 (y Q ,2, y QN , u , M , )
therefore ARR-based residuals might fail to detect some y  N
faults. D-detectability conditions can be used in di!erent
ways, e.g. to characterize the set of faults which cannot be rank "

detected, for given nominal trajectories; to characterize the


set of faults which can (or cannot) be detected for any w
E (y Q , 2, y QN , u , M , )
system trajectory; to check the FDI sensitivity along the y  N
nominal system trajectories (or on-line along the current for almost any vector (y Q ,2, y QN , u , M , ) which satisxes
system trajectories), for given faults to be detected.  N
(21).
3.6. Fault isolability Remark 4. The existence of robust ARR is not ensured,
even for high orders of derivation, since there is no
In the structured residual approach, a given fault sig- equivalent of Corollary 2 to give an upper bound of
nature is de"ned by the subset of those residuals whose r夹  2 N
structure the faulty element belongs to (Gertler & Luo, Q Q   Q 
1989). Consider the set of all the possible fault signatures, 4.2. Structured residuals
and de"ne a dissimilarity index between signatures (for
example the Hamming distance). The set of fault signa- Let us consider a subset of sensors J LJ and, for
tures can be decomposed into equivalence classes, each of simplicity, renumber them so that J "1,2, p . The
them gathering signatures with dissimilarity index zero. output vector is split into y and y. Consider also a sub-
These classes are the subsets of faults which are not set of actuators (resp. of parameters) such that the input
isolable from each other using the given residual. In other vector u (resp. the parameter vector ) is split into two
words, a necessary and su$cient condition for two faults sub-vectors u and u (resp.  and ). A "rst (weak)
to be isolable by a given residual is that they do not setting of the FPRG is to design residuals whose struc-
belong to the same equivalence class. ture does not contain y, u and . A second (strong)
setting is to ensure that the residual structure contains all
components of y , u and  .
4. The design of robust structured residuals
4.2.1. Weak residual generation problem
The fundamental problem of residual generation
In order to address the WRGP, the extended state
(FPRG) consists of the design of residuals which do not
(x, ) is considered:
depend on any unknown input and whose structure is
de"ned a priori. Robust structured residuals avoid false x
"f (x, u , u, v, ,  , ),
alarms due to unknown inputs and uncertain para-
Q "0,
meters, and solve the fault isolation problem in a simple
way. y "h (x, u , u, v, ,  , ).
 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699 693


Gathering x and  in a single vector  (v and u in ¹here exists no subset y Ly
a single vector
), one has and multi-index ( , j3JI )
(c) H (23)
Q "
(, u ,
, ,  ), such that properties (a) and (b)

Q "0, hold for  4s ∀j3J JI

H H
y "h (, u ,
,, ) (22) Indeed, let l be the dimension of . There exist two
integers r4n#l and r4n#l!1 such that
and the WRGP simply consists in eliminating the state I


 and the unknown inputs
in (22).
 s #1 if  s (r,
G G
rFG " GZ'Y GZ'Y (24)
Q  GZ'Y
Theorem 4. A residual with weak parameter structure r if  s 5r,
(y , u , ) exists at the orders of derivation (s , j3J ) of the G
H GZ'Y
outputs y if and only if


 s #1 if  s (r ,
G G I

(J , u ,  )"  s #p !rIH rFG ( )" GZ'Y GZ'Y (25)

H Q  HZ(Y
'0, Q  GZ'Y I
r if  s 5r .
HZ(Y I G I
GZ'Y
where
From (24), it appears that (23a) can hold only if

 
J (G ) J  (G ) J    (G )  s 5r, which means that there exist exactly
V Q F Q T S Q HZ(Y H

 " s #1!r independent ARR possibly in-
rIH "rank . (Y HZ(Y H
Q  HZ(Y volving (y QH , j3J ), u , M ,  and no others. Property (23b)
H
J (GNYNY )
V Q
J  (GNYNY )
F Q
J    (GNYNY )
T S Q means that for each parameter  of  , there exist
 !1
I (Y
independent ARR possibly involving (y QH , j3J ),
H
Note that, from Theorem 2 and Corollary 2, the state u ,M ,     and no others.
I
x and the parameters  can be eliminated, since the rank For  s "r (23a) and (23b) give a necessary and
HZ(Y H
of the Jacobian sub-matrix J  is bounded, while noth- su$cient condition for the existence of one residual de-
V F
ing can be said about the actual number of ARR which signed from the subset (y , j3J ) at orders of derivation
H
are robust w.r.t. v and u (see Remark 4). (s , j3J ) whose structure contains all the parameters of
H
 and none of the parameters of . Condition (23c)
4.2.2. Strong residual generation problem means that this result cannot be obtained from any
The SRGP is concerned with the actual presence of subset of sensors y Ly , except possibly using higher
each component of (y , u ,  ) in the residuals. Since most orders of derivation. Thus, all the sensors of y really
of the development is similar, only the design of strongly belong to the structure of the ARR.
structured residuals w.r.t. (y ,  ) will be considered, start-
ing with the part of the system robust w.r.t. the unknown Remark 5. The existence of a residual with parameter
inputs. structure  is a generalization of the identi"ability
For each component  of  , de"ne the following problem, in which  contains only one component (for
I a reference study on identi"ability, see Ljung and Glad
ranks:
(1994))

 
J (G ) J  (G )
V Q F Q

rF "rank , Remark 6. From (24) and (25), if there exists a parameter
Q  HZ(Y
H
 such that r "r, then  cannot belong to any struc-
J (GNYNY ) J  (GNYNY ) I I I
V Q F Q ture which would contain no parameter of , at any
order of derivation of the considered outputs. In the

 
J (G ) J  (G ) J I (G ) special case  "  (i.e. "k), this corresponds
V Q F Q F Q I
to  being non-identi"able through the observation of
rF
( )"rank . I
Q  HZ(Y I
H the outputs y , at any order of derivation of these outputs.
J (GNYNY ) J  (GNYNY ) J I (GNYNY )
V Q F Q F Q
Remark 7. Suppose there exists a subset of parameters 
Theorem 5. A residual with strong parameter structure such that r"r. Then no parameter of  could belong to
(y ,  ) exists at the orders of derivation (s , j3J ) of the a structure which would not contain any component of x.
H
outputs y if and only if The conclusion is that it is not possible to design any
parity-space residual sensitive to a parameter fault in ,
(a)  s #1'rFH ,
HZ(Y H Q  HZ(Y using the subset of outputs y , at any order of derivation
 
(b) rF H ( )"rF H #1 ∀ 3 , of these outputs.
Q  HZ(Y I Q  HZ(Y I
694 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699

Fig. 1. Residual evolutions: non-faulty case; C from 15 to 25 N m.



5. Example MR R
x
"  x !  x #px x ,
 ¸  ¸   
5.1. System model  


M R 1
x
"x #  x #px x # v ,
Consider the (d, q) model of the induction motor (for   ¸¸ ¸    ¸  B
   
an extensive development of this example, see Chris- (26)


tophe, Cocquempot, and Staroswiecki (1999) and Com- M R 1
x
"x #  x !px x # v
tet-Varga, Christophe, Cocquempot, and Staroswiecki   ¸¸ ¸    ¸  O
   
(1999)). The state variables are the rotor angular speed
 and the (d, q) projections of the stator current and rotor with state vector
#ux: i , i ,  and  . The control inputs are the (d, q) x"(x , x , x , x , x )"(, , , i , i ),
B O B O      B O B O
projections of the stator voltage: v and v . The load
B O and measurement vector
torque C is unknown. The outputs are the only state

variables that can be measured, i.e. , i and i . Under
B O y"(y , y , y )"(x , x , x ).
     
usual hypotheses the model equations are
R , ¸ are the stator resistance and inductance, R , ¸ are
pM K C    
x
" (x x !x x )! x !  , the rotor resistance and inductance, M is the mutual
 J¸     J  J inductance between stator and rotor, p is the number of

pole pairs, K is the damping coe$cient, J is the moment
x
"
MR R
 x !  x !px x , of inertia, "1!M/¸ ¸ and "!(1/)(R /¸ #
  ¸       
¸ (1!)R /¸ ).
   
 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699 695

Fig. 2. Residual mean values: non-faulty case and speed sensor fault (1, 3 and 5%).

5.2. Residual design has rank 6: there exist 2 ARR robust w.r.t. the load
torque C :

5.2.1. State elimination  "AQ [#(py )]
  
Deriving speed and currents at order 1, the Jacobian
matrix 1!
! [y #y y ][#(py )]
    

 
J (G )
V  ! [A #A py ][!2py #py
]
    
J (G )
V  ![A !A py ][(py )!],
J (G )    
V   "AQ [#(py )]
  
has rank 5. Thus, according to Theorem 2, there exists 1!
exactly one residual involving (y , y , y ): ! [y #y y ][#(py )]
       
 "p¸ A (py y !y )#p¸ A (py y #y ) ! [A #A py ][!2py #py
]
               
# [#(py )][Jpy
#Kpy #pC ] ![A !A py ][(py )!].
   
   
with A "y
!y !(1/¸ )v , A "y
!y ! The units of  and  are A rad s\.
    B     
(1/¸ )v and "R /¸ . Note that  (unit: A rad s\)
 O   
is not robust w.r.t. the load torque. 5.2.3. Structured residuals
Let us design residuals robust w.r.t. the load torque,
whose structure contains either parameter R or para-
5.2.2. Robustness w.r.t. the unknown input 
meter R . In Christophe, Cocquempot, and Staroswiecki
Using higher-order derivatives (speed at order 1 and 
(1999), it is shown that the residual with structure
currents up to order 2), the Jacobian matrix
R (unit: A rad s\), is given by the determinant of the


 
J (G ) J  (G ) Sylvester matrix:
V  ! 

 
J (G ) J  (G ) S? S?
V  !   "   ,
J (G ) J  (G )  S? S?
V  !   
696 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699

Fig. 3. Residual mean values: non-faulty case and rotor resist. fault (5, 15 and 25%).

 
where S@ 0 0 S@ 0 0
 
S@ S@ 0 S@ S@ 0
S? "(#(py ))(!y
!y #py y )    
      S@ S@ S@ S@ S@ S@
 "         ,
 S@ S@ S@ S@ S@ S@
#py
(py y !y ),        
   
0 S@ S@ 0 S@ S@
    
S? "(#(py ))(!y !y
!py y )
      0 0 S@

0 0 S@

#py
(py y #y ),
    where
v
S? "(#(py ))(!y
!yK #py y
) S@ "!y !y
# O ,
         ¸

S@ "py B !BQ !py
y (1!),


v
v v      
#(#(py )) O !py B # O S@ "!py
B #(py )(S@ )#py
y (1!),
 ¸ ¸ ¸       
  
S@ "(py )B #py
B !(py )BQ ,
      


v v S@ "!y !y
#v /¸ ,
#py
(1!)(py y !y )#  B !py O ,    B 
    ¸  ¸
  S@ "!py B !BQ #py
y (1!),
     
S? "(#(py ))(!y
!py y
!yK ) S@ "py
B #(py )(S@ )#py
y (1!),
            
S@ "!(py )B #py
B !(py )BQ ,
      


v
v v
#(#(py )) B # B #py O and B "y #y
!v /¸ B "y #y
!v /¸ .
 ¸ ¸ ¸    B     O 
  


v v
#py
((1!)py y #y )!  O #py B 5.3. Simulation results
    ¸  ¸
 
5.3.1. Simulation parameters
and the residual with structure R (unit: A s\) is given The control strategy is a vector control law which

by the determinant cancels the rotor #ux  in order to control the angular
B
 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699 697

Fig. 4. Residual mean values: non-faulty case and stator resist. fault (15, 30 and 50%).

speed by acting on the stator current. The parameters are 5.3.3. FDI performance
Five di!erent situations have been considered:
R "10 , ¸ "0.38 H, R "3.5 , ¸ "0.3 H,
   
M"0.3 H, 1. normal load torque (C 420 N m)

15, 17.5 and 20 N m,
J"0.02 kg m, 2. external fault (C '20 N m)

K"0.04 N m s rad\ and p"2. 22.5 and 25 N m,
3. speed sensor fault (at C "20 N m)
The sampling interval is 0.1 ms and the simulation is 
1, 3 and 5% error,
performed for 10 s. Gaussian white noise with zero mean 4. rotor resistance fault (at C "20 N m)
is added to the sensor outputs. The noise standard devi- 
5, 15 and 25% variation,
ation equals 1 for the current and speed sensors and 10 5. stator resistance fault (at C "20 N m)
for the voltage sensors, leading to the signal-to-noise ratios: 
15, 30 and 50% variation.



s s
"18 dB, "25.5 dB, In all runs, the load torque increases from zero to its
n n
  steady-state value during the "rst 2 s. Rotor and stator


s resistance faults are abrupt changes at t"5 s. The speed
"38.4 dB. sensor fault increases from zero to the fault value from
n
 
The speed measurement is "ltered with a lowpass nineth-
order Butterworth "lter. The cut-o! frequency is 200 Hz.
The (unknown) load torque is normally less than 20 N m.
5.3.2. Estimation of the derivatives
The signal derivatives are calculated using polynomial
approximation, i.e. the variables are approximated in the
least-squares sense, on a given time window, by a poly-
nomial function of time. In this example, the time win-
dow is 1.51e-2 s long and the approximating polynomial
is of degree 5. The signal derivative is estimated by the
polynomial derivative in the middle of the time window.
t"4 to 5 s and remains constant after t"5 s.
Fig. 1 displays the time evolution of residuals  } in
 
the "rst two cases. Residual  can be used to estimate

C and to detect external faults, while residuals  } ,
  
which are robust w.r.t. C , are not a!ected by its vari-

ations (all trajectories overlap).
Figs. 2, 3 and 4 show the time evolution of the residual
mean values (computed over a successive 1000 sample
points) in fault cases 3, 4 and 5, for di!erent fault
sizes. Only the robust residuals  } are used.
 
It can be checked that  is insensitive to the

rotor resistance deviation and that  is insensitive

to the stator resistance deviation. It also appears that
698 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699
the fault on the speed sensor does not a!ect the mean
value of residual  . It can indeed be proved that
 Automatic Control, AC-29, 603}614.
the speed sensor fault is not D-detectable using this
residual.Two decision thresholds (resp. 3 and 5, with
 the standard deviations in the non-faulty case) are
displayed. The di!erences in the mean values between the
non-faulty and the faulty cases are signi"cant enough,
which gives in practice very good false-alarm and
missed-detection rates, in spite of the heuristic decision
procedure.
6. Conclusion
Robust structured analytical redundancy relations can
be designed for algebraic dynamic systems, thus extend-
ing the parity-space approach. Further extensions to
rational systems and algebraic discrete-time systems can
be done. It should also be noticed that non-linear systems
that are not algebraic can be transformed into algebraic
systems, especially in cases where the non-linearities ap-
pear through sines, cosines, logarithms, or exponentials.
Detectability issues have been discussed, and the def-
inition of D-detectability has been introduced. Condi-
tions for the design of structured residuals and their
decoupling w.r.t. unknown inputs have been developed.
The approach has been illustrated on the induction mo-
tor example.
FDI theory for non-linear systems still requires more
research. For practical applications, the calculation of
the residuals arises two kinds of problems. On the one
hand, the time requirements of formal computational
software heavily depend on the a priori chosen order of
elimination of the system variables (Faugère et al., 1993).
In some cases, the problem is unmanageable with the
available software. On the other hand, analytical redund-
ancy residuals involve successive derivations of the sys-
tem inputs and outputs. For practical applications, the
real-time estimation of these derivatives still requires
tools to be developed.
The design of structured residuals and the problem of
unknown inputs are solved through strong decoupling
approaches. However, in some cases, approximate de-
coupling has to be considered. The theory of approximate
decoupling has been well developed for linear systems,
because of the simple de"nition of sensitivity and robust-
ness indices, which are constant. The non-linear situation
is much more complex, since sensitivities are non-linear
functions of the system trajectories and faults.
References
Buchberger, B. (1985). GroK bner bases: An algorithmic method in poly-
nomial ideal theory. In: N. K. Bose (Ed.), Multidimensional systems
theory (pp. 184}232) Dordrecht: Reidel.
Chow, E. Y., & Willsky, A. S. (1984). Analytical redundancy and the
design of robust failure detection systems. IEEE Transactions on
Christophe, C., Cocquempot, V., & Staroswiecki, M. (1999).
Robust residual generation for the induction motor using
elimination theory. Proceedings of the IEEE SDEMPED+99, Gijon,
Spain.
Cocquempot, V., Cassar, J. Ph., & Staroswiecki, M. (1991). Generation
of robust analytic redundancy relations. Proceedings of the
European control conference (ECC+91), Grenoble, France.
Cocquempot, V., & Christophe, C. (2000). On the equivalence between
observer-based and parity space approaches for F.D.I. in non-
linear systems. Proceedings of the IFAC safeprocess, Budapest
(pp. 232}237).
Comtet-Varga, G., Cassar, J. Ph., & Staroswiecki, M. (1997).
Analytic redundancy relations for state a$ne systems. Proceedings
of the fourth European control conference (ECC+97), Brussels,
Belgium.
Comtet-Varga, G., Christophe, C., Cocquempot, V., & Staroswiecki, M.
(1999). F.D.I. for the induction motor using elimination theory.
Proceedings of the xfth European Control Conference (ECC+99),
Karlsruhe, Germany.
Cox, D., Little, J., & O'Shea, D. (1992). Varieties and algorithms. New
York: Springer.
David, P. J. (1975). Interpolation and approximation. New York: Dover
Publications.
Dierckx, P. (1993). Curve and surface xtting with spline. Oxford, New
York: Clarendon.
Diop, S. (1991). Elimination in control theory. Mathematical Control
Signals Systems, 4, 17}32.
Diop, S., & Wang, Y. (1993). Equivalence between algebraic observabil-
ity and local generic observability. Proceedings of the 32nd IEEE
conference on decision and control, San Antonio (pp. 2864}2865).
Faugère, J. C., Gianni, P., Lazard, D., & Mora, T. (1993). E$cient
computation of GroK bner bases by change of orderings. Journal of
Symbolic Computations, 16, 329}344.
Fliess, M. (1989). Automatique et corps di!eH rentiels. Forum
Mathematicum, 1, 227}238.
Fox, L., & Parker, I. B. (1968). Chebyshev polynomials in numerical
analysis. London, New York: Oxford University Press.
Frank, P. M. (1993). Advances in observer-based fault diagnosis.
Proceedings of the international conference on fault diagnosis
(TOOLDIAG+93), Toulouse, France.
Gertler, J. J., & Luo, Q. (1989). Robust isolable models for failure
diagnosis. A.I.Ch.E. Journal, 35(11), 1856}1861.
Glad, S. T. (1989). Non-linear state space and input-output descriptions
using di!erential polynomials. In J. Descusse, M. Fliess, A. Isidori,
& D. Leborgne (Eds.), New trends in nonlinear control theory,
Lecture notes in control and information science, vol. 122 (pp.
182}189). Berlin: Springer.
Guernez, C., Cassar, J. Ph., & Staroswiecki, M. (1997). Extension of the
parity space to non-linear polynomial dynamic systems. Proceed-
ings of the IFAC Safeprocess +97, Hull.
Hammouri, H., Kinnaert, M., & El Yaagoubi, E. H. (1999). Observer
based approach to fault detection and isolation for non linear
systems. IEEE Transactions on Automatic Control, 44, 10.
KaboreH , P., Staroswiecki, M., & Wang, H. (1999)c. On parity space and
observer-based approaches for FDI in state-a$ne systems. In
Proceedings of the 38th IEEE conference on decision and control,
Phoenix, AZ, USA.
Lelong-Ferrand, J., & Armandiès, J.M. (1978). Cours de Mathe& matiques,
Tome 1, Algèbre. Paris: Dunod UniversiteH .
Ljung, L., & Glad, S. T. (1994). On global identi"ability for arbitrary
model parameterizations. Automatica, 30, 265}276.
Magni, J. F., & Mouyon, Ph. (1994). On residual generation by observer
and parity space approaches. IEEE Transactions on Automatic
Control, 39(7), 441}447.
 M. Staroswiecki, G. Comtet-Varga / Automatica 37 (2001) 687}699 699

Patton, R. J. (1994). Robust model based fault diagnosis: The M. Staroswiecki graduated in 1968 from
state of the art. Proceedings of the IFAC Safeprocess+94, Helsinki, the Ecole Nationale SupeH rieure d'Arts et
Finland. MeH tiers in Mechanical and Electrical Engi-
Seidenberg, A. (1956). An elimination theory for di!erential algebra. neering. He received his Ph.D. degree in
University of California Publication Mathematics, 3, 31}65. 1970 and his Dr es Physical Sciences in
1979. He is a professor in Control Engineer-
Seliger, R., & Frank, P. M. (1991). Fault diagnosis by disturbance ing at the Ecole Universitaire d'IngeH nieurs
decoupled non-linear observers. Proceedings of the 30th IEEE de Lille. He is currently the director of
conference on decision and control, Brighton (pp. 2248}2253). the Laboratory of Automatic Control and
Shumsky, A. (2000). Continuous-time parity relation method for fault Computer Science at the Lille University,
diagnosis in non-linear uncertain systems. Proceedings of the IFAC and heads the group on Fault Detection
safeprocess, Budapest (pp. 998}1003). and Isolation and Fault Tolerant Control.
Yu, D. N., & Shields, D. N. (1995). Fault diagnosis in bi-linear systems
* A survey. Proceedings of the third European control conference, Gilles Comtet-Varga graduated in 1993
Rome (pp. 360}366). from Ecole Centrale Paris in applied
Yu, D. N., Williams, D., Shields, D. N., & Gomm, J. B. (1995). mathematics. He received his Ph.D. in
A parity space method of fault detection for bi-linear systems. Control Theory from University of
Sciences and Technologies of Lille, in 1997,
Proceedings of the American control conference, Seattle
at the Laboratory of Automatic Control
(pp. 1132}1133). and Computer Science. He then joined
Zhang, Q., Basseville, M., & Benveniste, A. (1998). Fault detection and Syntegra as an Information Technology
isolation in nonlinear dynamic systems: A combined input}output engineer in 1998 and he is currently a sys-
and local approach. Automatica, 34(11), 1359}1374. tems engineer at Steria.