403-424, 1997
1997 Elsevier Science Ltd. All rights reserved
Printed in Great Britain
0959-1524/97 $17.00 ~- 0.00
ELSEVIER
PII: S0959-1524(97)00016- 4
Papers
Survey of robust residual generation and
evaluation methods in observer-based fault
detection systems
P. M. Frank* and X. Ding*
*Gerhard-Mercator-Universitat -GH-Duisburg, Fachgebiet Mess- und Regelungstechnik.
Bismarckstrasse 81, 47048 Duisburg, Germany
*FH Lausitz, FB Elektrotechnik, GrossenhainerStrasse 57, 01968 Senftenberg, Germany
The paper outlines recent advances of the theory of observer-based fault diagnosis in dynamic systems
towards the design of robust techniques of residual generation and residual evaluation. Emphasis will be
placed upon the latest contributions using frequency domain techniques including Ho~ theory, nonlinear
unknown input observer theory, adaptive observer theory, artificial intelligence including fuzzy logic,
knowledge-based techniques and the natural intelligence of the human operator. Two representative
examples illustrate the efficiencyof the observer-based approach. © 1997 Elsevier Science Ltd
Keywords: fault detection; robustness; unknown input observer; nonlinear observer; adaptive observer; obser-
ver-based fault detection; fuzzy logic
Associated with an increasing demand for higher per- On the other hand, the involvement of knowledge-
formance as well as for more safety and reliability of processing techniques leads to the concept of a know-
dynamic systems, fault diagnosis has received more and ledge-based observer that makes use of a qualitative
more attention. One area of active research is the process model (knowledge model). Its task in fault
development of model-based fault detection systems. A diagnosis is to reconstruct the symptoms corresponding
fault detection system processes on-line information of to the observations of the process that can be used for a
the process under observation, namely process input fault decision on the basis of the knowledge redundancy
and output signals. The output of the fault detection thus being created. Both analytical and knowledge
system may be simply an alarm signal that takes two redundancy may be subsumed under the term functional
values, high for fault and low for fault-free or, more redundancy.
sophisticatedly, knowledge of faults such as location, In the field of the quantitative model-based tech-
spectrum or amplitude. niques, the observer-based approach is in contrast to the
There exist a number of strategies to construct model- parameter estimation approach where the fault decision
based fault detection systems. One of them, using is performed by on-line parameter estimation. Both
observer techniques, has received much attention during approaches have advantages and disadvantages in dif-
the last years. Especially, substantial progress in control ferent respects, and there was much argument for and
theory and computer capability has made it possible to against each approach. Realistically one should admit
apply observer-based fault detection techniques to that both methods are to a certain degree complemen-
complex processes including nonlinear and time-varying tary and are therefore best applied in combination 1.
systems with considerable modelling uncertainty. Typi- The paper focuses on the observer-based approach.
cal for the classical observer-based approach is that one We also include in this framework the parity space
reconstructs measurements of the process with the aid methodology, because it was recently evidenced by sev-
of an observer using a quantitative mathematical model eral authors 2-5 that the parity space approach leads to
of the process and makes the decision on possible faults certain types of observer structures and is therefore
in the process on the basis of the analytical redun- structurally equivalent even though the design proce-
dancy thus being created. Even though there are a dures differ.
number of different design procedures, the core of the In theory, more attention has been paid to the obser-
resulting diagnostic systems is always observers or ver-based approach than to the parameter estimation
Kalman filters. approach. This is probably due to the fact that the
403
404 Robust residual generation and evaluation methods: P. M. Frank and X. Ding
existing parameter estimation theory can readily be where u(s) is the input vector and yo(S) is the nominal
applied to fault diagnosis without major modifications observation vector, G,(s) describes the transfer behavior
whereas diagnostic observers are different from the well- between the input vector u(s) and the nominal output
known control observers and therefore deserve particular vector yo(S). The effect of faults on the system dynamics
theoretical treatment. One of the essential differences is is modelled by Gf(s)f(s) with f representing a fault vec-
that diagnostic observers are primarily output observers tor and Gf(s) denoting a distribution transfer matrix,
rather than state observers as needed for control pur- respectively. Under the assumption that no false alarm
poses. This has often been overlooked in the literature is allowed, the threshold should be the maximal value of
and has misled many practitioners to the erroneous the evaluated output in the nominal process operating
opinion that for the observer-based approach the state (fault-free) for example:
knowledge of state-space theory would be indis-
pensable. ]] au u []e~-~ Jth for all possible u(s) and G,(s) (2)
Another most important difference is that whilst
control observers are used within a closed loop, diag- with II'lle denoting some evaluation function. This yields
nostic observers operate in an open-loop configuration.
Therefore, modelling errors of the process, which can by Jth = sup [1y lie--- sup I1yo lie (3)
no means be avoided in practice, are much more trou- f=0
blesome. This requires robustness with respect to model
uncertainties. Actually, satisfactory robustness is the It is clear that a fault can be detected only if it causes
indispensable precondition for the practical application the evaluated output y to be larger than the threshold:
of a diagnostic observer scheme. A great deal of theo-
retical work is therefore devoted to the robustness pro- [I Y lie =11 a.(s)u(s) + Gf(s)f(s) lie> Jth
(4)
blem, and the task of enhancing the robustness in the for all possible u(s) and G,(s)
face of considerable modelling uncertainty is the subject
of many publications in recent years2'4,6-26. which equivalently means (see below or Emami-Naeini
Despite the deficiency of unifying mono- et al. 8)
graphs 1~,22,27,2s, there is a solid theoretical foundation
to the analytical observer-based approach as far as the 1t a f f [lee 2Jth = 2 sup 11yo lie (s)
time domain and linear systems are concerned. How-
ever, extensions to the frequency domain design of lin- In practice, sup [[ Yo lie can be obtained by simulation,
ear observers 6--8'29-39 adaptive observers 32,36,48,58,59 and its value depending on the input signal u, and therefore
nonlinear robust observer schemes 11,36,39,43-45 have been could be very large so that, according to the above
elaborated upon only lately and research in these areas relationship, faults with smaller size become undetec-
is still going on. Some relevant contributions in these table. A well known way to solve this problem is the
areas for the design of robust residual generation and utilization of knowledge of the nominal process transfer
residual evaluation will be outlined in the paper. behavior. If one could model the dynamic processes
In recent years, there is also a clear trend towards an accurately so that the desired process variables could be
enlarged involvement of knowledge-based and artificial precisely estimated, the difference between the measure-
intelligence methods, including qualitative modelling for ment y(s) and its estimation, called residual, can be used
residual generation and fuzzy logic for residual evalua- instead of the output y for the purpose of fault detec-
tion 9'12"40-44. This issue will briefly be addressed towards tion. In this case, the influence of the process input sig-
the end of the paper. Finally, we will outline the obser- nals can be exactly eliminated so that Jth is nearly zero.
ver-based supervision of a three-tank system and an This means that every fault can theoretically be detec-
industrial robot as examples of a successful practical ted. This ideal case is, unfortunately, rare in a real
application of the observer-based methodology. technical process. Perfect models do not exist nor are
characteristics of possible model uncertainties, which
are unavoidable in real technical systems, available.
Denote the residual with
Background and problem formulation
r(s) = ar(s) + 6rAs)f(s) (6)
Background
where Ar(s) represents the effect of model uncertainty
It is very interesting to notice that in practice, instead of on the residual r(s). Consequently, a fault is detectable
residuals, output signals of the process under consid- only if
eration are often directly evaluated and compared with
a given threshold. Suppose that the process can be [I Grff ]]e~ 2Jth = 2 sup II Ar lie (7)
described by
Thus, the core of constructing a fault detection system is to
minimize the influence of disturbances and model uncer-
y(s) = yo(s) + CAs)f( ) = 6.(s)u(s) + aAs)f(s) (1) tainties on the residual, and this can be achieved either
Robust residual generation and evaluation methods. P. M. Frank and X. Ding 405
• by utilizing additional information about the pro- unknown inputs (disturbances, noise, modelling errors),
cess such as qualitative knowledge, or to which the detection system should be immune. The
• by applying a robust technique to the fault detec- ideal goal o f a residual generator is to generate a vector
tion system design. r(t)such that r(t) = 0 asf(t) = 0 and
1. r(t) ~ 0 asf(t) ~ 0 for fault detection
Problem formulation
2. ri(t) ~ 0 asf.(t) -¢ 0 for fault isolation
3. limt_~ ~(t) - r(t)] = 0 for fault identification
On the basis of the observation and discussion in the
last sub-section, we formulate the tasks of constructing where the f- represent the different faults to be isolated
a fault detection system as follows: and ri the corresponding subsets of residuals.
A number of methods for observer-based residual
• design a residual generator that eliminates the
generation have been proposed over the past two dec-
effects of process input signals and,
ades. The most significant approaches are the fault
if possible, also the effects of disturbances and
detection filter, the innovation test, the dedicated and
model uncertainties on the residual generated;
generalized observer scheme, and the unknown input
• design a residual evaluator by selecting a suitable
observer scheme. In the face of an overwhelming litera-
evaluation function [l[[e and
ture on the subject, we simply refer to Patton et al. 22
determining the threshold Jth;
and the papers cited therein.
• if a full elimination of the effects of disturbances
In parallel, there have been similar efforts to solve the
and model uncertainties on the residual is not
fault detection and isolation problem starting from the
possible, optimize the residual generator and eval-
parity equations, The most relevant contributions to
uator to achieve the maximum set o f detectable
redefine the parity space approach, discover the con-
faults.
nections to the observer-based approach and generate
structured residuals were recently made by Gertler 4"13"~4.
In recent years, the studies concentrated more and
Observer-based residual generation more on the design of robust residual generators that are
invariant or at least insensitive with respect to unknown
State of the art inputs. These studies have converged to a well founded
theoretical framework comprising a number of different
In the analytical observer-based approach, the genera- approaches 3,5,9,1°J6,23.25,26.39,45. It is noticeable, how-
tion of residuals reflecting the faults is done by estimat- ever, that the established theory is almost entirely
ing outputs of the process and using the estimation devoted to linear systems and to the design in the time
errors as the residuals. For the fault detection task, a domain, even though many processes in practice are
single observer or Kalman filter is sufficient whereas, for nonlinear and, on the other hand, frequency domain
the localization of the faults, properly structured sets of techniques have made big progress.
residuals are required. The latter can be generated by In this section, we will therefore focus upon the latest
using banks of the observers, so-called dedicated and attempts to approach the problem of residual genera-
generalized observer schemes (DOS and GOS) 9. tion in the frequency domain, and will then extend the
Depending on the circumstances, one may use linear or theory of linear unknown input observers to the non-
nonlinear, full or reduced-order, or fixed or adaptive linear and adaptive case. Finally, we will briefly outline
observers (or Kalman filters). the basic idea behind the concept of knowledge observer.
The basic concept of an analytical observer-based
residual generator is illustrated by the block diagram of Basic principle of residual generator construction
a linear full order observer in Figure 1. By f we denote
the vector o f faults to be detected, represented by Whilst the task of an observer for control purposes is to
(unknown) time functions, and by d the vector of reconstruct the states of the process, there is normally
no such need for diagnostic observers. Their task is to
reconstruct the outputs (i.e. the subset of the state vector
that is measurable) in order to create redundancy.
d
Therefore, linear diagnostic observers can readily sim-
MEASUREMENTS
PROCESS
ply be designed as output observers.
A direct way to construct an output observer is using
the input-output relation which is usually described by a
transfer function in the frequency domain. This schema,
MODEL
also called the frequency domain approach, was intro-
duced by Viswanadham et al. 45 and lately extended by
Ding and Frank 6,29"34'35,37. The major features of this
approach are, on the one hand, its potential to pro-
OBSERVER; KALMANFILTER
vide a complete solution and, on the other, its obvious
Figure 1 Full order observer for residual generation physical purpose of checking the input-output transfer
406 Robust residual generation and evaluation methods."P. M. Frank and X. Ding
relation as well as the use of transfer functions instead Remember that Q(s) is a parametrization matrix yet
of state-space descriptions. free to select. Thus, it follows from the results given in
To briefly outline the basic idea, consider a linear Ding et al. 7 that all residual generators can be con-
process described by structed by
yz(s) = G.(s)uL(s) + Ayt.(s) + Gf(s)ft(s) (8) rL (S) = Q (s) (yl. (s) - YL (s) ) (16)
where Gu, Gf are known transfer matrices from the with ~L(S) as an output estimation which is the output
input vector u E Rp, fault vector f E R q to the output of any type of output observers in the classical sense of
vector y E R m, and the subscript L denotes the Laplace observer theory, e.g. Luenberger-type output observer,
transformed time functions. Ay/.(s) is an unknown vec- or identity observer or parity space estimator. This
tor representing unknown disturbances and model means that every residual generator is indeed an output
uncertainty. estimation error system that is, if necessary, filtered by
Viswanadham et al. 46 have proposed to construct the Q(s). This fact explains why in practice a lot of simple
residual generator by factorization technique, which is fault detection systems, consisting of an output estima-
applicable for all processes (both stable and unstable) tor simply and experientially constructed, are success-
and the residual generator can then be brought into the fully used for fault detection.
form:
Frequency domain approach to robust residual generator
rL (S) = (S)(y/. (S) -- G. (s)UL (S))
(9) We now consider the effects of faults and model
= t.(s)yL(s) -- g.(S)UL(S)
uncertainty AyL(s) on the residual rL(S). For this pur-
pose, residual generators of form (13) are taken into
where A?/.(s) and Ar~(s) are left coprime factors of G.(s) account. Substituting the system Equation (8) into it
satisfying ~l~1(s)N~(s) = G.(s). Denoting the state gives
space realization of the nominal transfer matrix G.(s)
with rL(S) = Q(s)l(4.(s)[Gf(s)fL(s) + AyL(s)] (17)
k(t) = A x ( t ) + Bu(t), y(t) = Cx(t) + Du(t) (10) The model uncertainty Ay/.(s) can be divided into
structured and unstructured uncertainty. For the pur-
-~/'.(s) and N.(s) can be calculated as follows: pose of fault detection, we only need to consider the
following general form:
t~I(s) = I - C ( s I - A + L C ) - 1 L (11)
AyL(s) = Gd(s)dL(s) (18)
]Vu(S) = D + C(sI - A + LC) -1 (B - LD) (12)
with an unknown but bounded vector d
with L ensuring the stability of the matrix A - LC. II d I1=< aa (19)
Note that this form for rz is identical with the
(transformed) generalized parity vector introduced by
Lou et al. ~6 The relationship (9) was recently generalized If rank Ga(s) = m and no information in the frequency
by Ding and Frank 29 to domain is available, for instance Ga(s) = L the expres-
sion Ga(s)di.(s) represents unstructured uncertainty and
otherwise structured uncertainty. Thus, it follows from
rL(S) = Q(s)[~l.(s)yL(s) - 17.(S)UL(S)] (13)
Equation (17):
This is the effectual form of the residual generator that of the effects of the faults and unknown inputs as an
is used to determine the parametrization matrix Q(s) for additional criterion for their discrimination. Therefore,
satisfying desired specifications. from a more practical viewpoint, the generalized residual
generator of form (13) can be seen as a useful extension
Full decoupling. Perfect fault isolation and total invar- of both the generalized parity space approach and the
iance from the unknown inputs dL(s) require perfect robust observer-based approaches in the time domain.
decoupling not only among the faults but also between It was one of the most remarkable achievements of
the faults and the unknown inputs. advanced research in observer-based fault detection to
The latter can only be achieved if the uncertainty is show that there are indeed practical situations where
structured. In this case it follows from Equation (17) perfect unknown input decoupling can be reached. In
that the matrix Q must be chosen so that the following most situations, however, even if the uncertainty is
conditions are met: structured this will not be the case due to the restrictive
conditions (23), (24). Then only optimal approximations
Q(s)~l,(s)Gf(s) = diag((tl(s) ..... tq(S)) E Rnoc (21) of perfect decoupling can be achieved.
this problem. It is worth mentioning that in Murad et TEl = 0 . GE2 = 0 , L2E2 = 0, L ~ T + L 2 C = O (43)
al. is, Nett et al. ~9 and Tyler 51 the diagnostic problem
has been solved in connection with the solution of the
control problem. This leads to an integrated design of
rank(TK~) = rank(K~),rank( ( G ) K 2 ) =rank(KQ
L2
FDI and control systems.
(44)
An alternative strategy for robust residual generation
in the case of unstructured uncertainties was recently
suggested by Patton and Chen 24. The basic idea of this If these requirements can be fulfilled, the dynamics of
approach is to compute the distribution matrices such the residual is governed by
that an optimal approximation of disturbance decoup-
ling is achieved. This optimization problem is solved via b = Fe + GK2f2 - TKIflr = LIe + L2K2f2 (45)
a singular value decomposition method for the rank-
reduced approximation of a rectangular matrix, first The drawback of this elegant extension of the linear
introduced to fault diagnosis by Lou et al. 16 unknown input observer theory to a class of nonlinear
systems is that the class of systems described by models
Nonlinear unknown input observer approach matching (39), (40) is rather limited. Many technical or
physical systems cannot be modelled this way. If this is
Many nonlinear processes in practice cannot be repre- the case, the given physical model must be transformed
sented by linear models, in particular when they are not into the required form by a suitable nonlinear state-
operating at a fixed operation point. This is the normal space transformation. The existence conditions for these
case in fault detection, because at the occurrence of a transformations are very restrictive. Consequently, the
fault the process runs out of its operating point. Hence, class of models that are actually transformable is rather
if a linear residual generator was used for a nonlinear small. But even if the existence conditions can be satis-
process then, after the occurrence of a fault, the detec- fied, finding the transformation will be hampered by the
tion system would be submitted to increasing modelling necessity to solve nonlinear partial differential equations
errors and run out of its range of validity. As a result, it or the requirement of up to nth-order time derivatives of
would release false alarms rather than detect and isolate the input signal u, where n is the dimension of the
the actual faults. model.
Even though this is a key point concerning the prac- Therefore, a different approach that extends the class
tical applicability of model-based fault detection, little of transformable systems, because it requires weaker
work has been done so far to develop residual genera- existence conditions, has been proposed by Seliger and
tors using nonlinear observers. Recently, Frank, Frank 53"55. It is based on the following more general
Wfinnenberg, Seliger and Ding have intensively studied model:
this problem 26"39'53-5s. They have extended the theory of
linear unknown input observers for residual generation x = A(x) + B(x)u + E(x)d + K(x)~ y = C(x) (46)
to certain classes of nonlinear systems.
For a brief discussion of this approach, consider the where the unknown inputs are modelled as to represent
class of systems that can be described by parameter uncertainties. It is desirable to compute a
nonlinear transformation z = T(x), separating the dis-
2 = Ax + B(y, u) + E1dl + KI~ (39) turbed from the undisturbed portion of the model. This
separation can be achieved if, and only if
y = c x + E2d,. + K2f2 (40)
OT(x) E(x) = 0 (47)
The signals dl, d2 represent unknown inputs and the Ox
terms fi,f2 denote the faults• Note that the nonlinear
term B(y, u) depends only upon y and u, i.e. upon sig- This relation constitutes a system of l st-order linear
nals which are directly available by measurements, partial differential equations which are to be solved
'observable nonlinearity'. It is therefore possible to simultaneously by z = T(x). The theorem of Frobenius
compensate completely the nonlinearity by reproducing can be applied to derive necessary and sufficient exis-
it using an observer of the form: tence conditions for solutions of (47) 53`55.
Suppose solutions z = T(x) of (47) exist. On the
= F~ + JO', u) + Gy, r = LI2 + L2y (41) assumption that a relation x = ~ o ( z , y * ) exists, the
model can be rewritten as
The conditions which are to be met by the observer _ or(x)
matrices in order to provide robustness to the unknown " Ox (A(x) + B(x)u + K(x)J) IX=,o(:,y.) (48)
inputs and sensitivity to the faults can be stated as fol-
lows: where the output transformation y* = C*(y) denotes a
subset of the set of available measurements y = C(x)
TA-FT=GC, Fstable, J(y,u)= TB(y,u) (42) which is subject ~o the condition
410 Robust residual generation and evaluation methods: P. M. Frank and X. Ding
P
Observer Error Linearization (OEL) problem that can
Jc = a(x) + qo(X, u) + Z qi(x, u)Oi
i= I briefly be formulated as finding a co-ordinate transfor-
q mation that transforms a general nonlinear system into
(58)
+Z hi(x, u)f" + g(t) the AOCF. It is known that a one-to-one transformation
i=l exists only if some rigorous conditions are satisfied 6°
= a(x) + qo(X, u) + O(x, u)O + H(x, u)f+ g(t), which, unfortunately, do not hold in many practical
cases.
y = c(x) (59) In contrast, Ding et al. 3~ and Seliger and Frank 53
have shown that a reduced-order observer for which
such a one-to-one transformation is not necessary can
where x E R n is the state vector, u E R t the known input
release these hard existence conditions.
vector, y E R m the measurable output vector, and a : R"
It is well known 9,29 that a residual generator is the
R n, qi, hi : R n × RR l ---' R ~, g : R ---, R n and c : R ~ --*
error system of an output estimator which can usually
R m are assumed to be known and smooth enough.
be constructed by a reduced-order system. Thus, it is
O(t) E RP is an unknown vector which represents reasonable to use a bank of output observers instead of
unknown time-varying parameters, slowly varying
a state observer• Following this. the corresponding so-
faults or part of nonlinearities of the system. The deri-
called Output Observer Error Linearization (OOEL)
vative is bounded by
problem can be formulated. For the sake of simplicity,
we restrict ourselves to the description of the following
]O]<M,O<M<<oc (60)
problem.
Given a nonlinear system, Equations (58) and (59),
The vector 0 will be on-line estimated. Vector f E R q and an initial state Xo: find (if possible) a neighborhood
denotes a b r u p t changes (faults) in the system that are to U of Xo, and transformations z = T(x) E R", (hi < n),
be detected as early as possible. F(y) E R defined on U such that
The laws for the corresponding adaptive generator o f
the residual r may be written in the following general z,, = ? ( y ) = ?(c(x)), (63)
form:
P q
- = Fz + O0(y..) + ~ ~,(y, .i0i + ~ ~,(x, u).~
Residual generator" z = b ( L O , u,y), r = h(2,0, u,y) i=1 i=1
(61) := rz + ¢'o(Y, u) + CO(y, u)O + E(x, u)f
(64)
A d a p t a t i o n l a w ' 0 = d(L O, u, y) (62)
and
The structural diagram of the adaptive residual genera-
tor is depicted in Figure 3. rank E(x, u) = q, rank qJ()', u) = p (65)
In order to apply the adaptive observer scheme to
fault detection, the nonlinear system, Equations (58)
for all z E T(U) and u, and
and (59), has to be brought into the so-called adaptive
observer canonical form (AOCF). As shown by
Marino 6°, this requires the solution of the so-called
F= [i0 011
0
• --
-.-
0
0
Here, F E R p×p is an arbitrary positive definite matrix, F* = Q ( F - LC)Q -1, ~bT(t) = [~bT(t)...~bmv(t)],
and ~o7(t) = [O~oVi (t)]
Residual evaluation
Problem formulation
R = diag(R1 . . . . . Rm) E R ('-m)×(n-m)
The second step of a fault detection procedure is to
o evaluate the residuals. This is a decision-making process
which always comes down to a threshold logic of a
0 ... 0 -ra
EE R (ni-1)x(ni-l), decision function. If there are no uncompensated
Ri
unknown input effects on the residuals due to a perfect
decoupling, then the thresholds diminish to zero.
• .. 0 I -ri(m-l) d
Otherwise, thresholds different from zero have to be
ril assigned.
rz2 In practice, there is usually such a great number of
r i -~- unknown inputs that, in the face of the limitations of
available measurement information, a complete decoup-
ri(ni- 1) ling from all unknown inputs is hardly achievable even
if the model uncertainties are structured. The situation
becomes even worse if the uncertainties are unstruc-
if(y, u) = [ ~ ( y , u ) . . . ~mV(y, u)] T E R ("-m)xp tured, in which case a perfect decoupling in the residual
,i,(y, u) = ~Pn (Y, u) - riqJz2(y, u) E R (m-l)xp generation stage is basically impossible. Hence the resi-
u) = u) ... %.(y,
^s u)] v c R Xp, duals or any decision functions built from them always
deviate from zero even if no fault is present.
u) = qJa(Y, u) E R lxp In this case, robust residual evaluation is the only way
K = diag( K~ . . . . Kin) ~ R m~("-") to keep the false alarm rate small with an acceptable
sensitivity to faults. Robust residual evaluation can be
KT= [ 0 . . . 0 1 0 . . . 0 ] E R l×("'-r) accomplished in many ways, for example by statistical
data processing, data reconciliation, correlation, pattern
In the above equations, L and ri, (i = 1. . . . . m) have to recognition, fuzzy logic or adaptive thresholds.
be chosen such that the matrices F - L C and R are sta- Here, we restrict ourselves to the adaptive threshold
bile matrices. Notice that since L and r; are arbitrary, and fuzzy decision-making approaches.
the eigenvalues o f matrices F - L C and R can be
assigned arbitrarily. Robust residual evaluation using adaptive thresholds
The dynamics o f the residual r with respect to the
unknown parameter vector 0 and faults f are given by Basic idea. The crux with fixed thresholds is that
the following error system equation: choosing the threshold too low increases the rate of
false alarms, choosing it too large reduces the efficiency
of fault detection• Evidently, the optimal choice of the
magnitude of the threshold depends upon the nature of
(69) the system uncertainties and varies with the system
input. One may therefore use thresholds that adapt to
+ z(x,u)f the input.
Figure 4 shows a typical situation. Consider the shape
of the residual (or a decision function) with effects of
r = ~(y) - C2 (70)
model uncertainties from t = 0 on and a fault at t ~ tF.
With a fixed threshold, the increase of the residual due
= z- 2,g= o-o, .2, = Q ~ - v(t)o to an input maneuver leads to a false alarm at trA and
no detection of the fault. Using an adaptive threshold
Q = diag(Ql . . . . . Qm), depending upon the system input allows the avoidance
1 0 ... 0 --ril of the false alarm and the detection of the fault at tE.
0 1 ... 0 -ra The idea of adaptive thresholds was first introduced
by Clark 22. He has chosen the shape of the threshold as
E R nixni
Qi a function of the input of the process in a more or less
0 ... 0 1 -ri(ni_l)
intuitive manner. A broader theoretical foundation of
this strategy was given by Emami-Naeini et al. 8 in terms
0 ... 0 0 1
of the threshold selector.
Robust residual generation and evaluation methods." P. 114.Frank and X. Ding 413
I DECISION FCT.
]
rL(s) = O(s)iif/l.(s)Gr(s)fL(s ) + ~Ids)AG.(s)uL(s)I
t FA tF TIME (75)
Figure 4 Adaptive threshold test of the residual or decision function
As shown by Frank and Ding 52, one can find from (75)
Evaluation functions and thresholds. The first step of the following relation for the threshold
residual evaluation is to choose an evaluation function
and, based on it, to determine the corresponding J,h -II GQMuu lie (76)
threshold. Among a number of residual evaluation
functions, the so-called root mean square (rms) is often Here, II AG. tl< ~. denotes a known bound on AG. and
used in practice. The rms can be represented either in Q again the parametrization matrix. It is seen that the
the time domain: threshold is no longer fixed but depends upon the input
u, thus being adaptive to the system operation. A fault
may be declared if II r lie> Jth.
II r(t) t1~= J(r) = (r -t jrr(t)r(t)dt) 1/2 (71) In general, a threshold selector can be, according to
0 (73), established as follows:
II rLO'O ) = J(E) = J r*L(jo))rL(jog)d(.o) 1/2, Emami-Naeini et al. s have solved this problem under
(72) the assumption that the rms in the time domain is used
o) I
as residual evaluation function H r lie. The similar solu-
ff = dO2 - - O} 1
tion based on the frequency domain rms (72) was per-
formed by Ding and Frank 51. By a suitable choice of the
where z and c denote the detection window in the time frequency window ~ = a~2 - COl, one can find a threshold
and frequency domain, respectively, and H'][e stands for so that the robustness with respect to uncertainties can
an evaluation function. be increased. The expression for the threshold can be
The selection of residual evaluation functions often found by setting fL (s) = 0 in (20) which yields
plays an important role for fault detection, especially in
practice, and may strongly influence the performance of Jth = sup I[ Q(s)iQ.(S)Gd(s)d(s) lie (78)
a fault detection system. Unfortunately, this problem is d
To outline the basic idea o f the threshold selector, we This means, somewhat surprisingly, that the threshold is
consider unstructured modelling errors. A linear time a constant equal to e -~n. As a matter of fact, one may
414 Robust residual generation and evaluation methods: P. ltd. Frank and)(. Ding
expect a Jth that changes with the input signal. How- I] r [Ioc~ Jth =:> a f a u l t h a s o c c u r r e d (83)
ever, if we observe 8d(W) in detail, the reason becomes
evident. For instance, for a model with additive [t r ]1~ < Jth ~ no fault has occurred (84)
unstructured modelling uncertainties:
Here l] r II~ denotes the Lo~ norm of the residual r
d(s) = [ ZXGw( ) (s) ] which is defined by
I[ d(fio) [12< ~(co) + ~(09) t1 u(j09) ]l~:= 82(c0) (82) Suppose the dynamic model of the process under con-
sideration is given by
This shows that the information on the input signal is
2=A(x)+B(x)u+El(x, u)d+Kl(x,u)f (86)
included in ~d(o~) which is further processed during the
residual generator design. Figure 5 shows the resulting
block diagram for the overall procedure for robust fault y = C(x) + E2(x)d+ K2(x)f (87)
detection in the frequency domain.
where d denotes the vector of unknown input signals
Nonlinear threshold selector representing, for example, modelling errors, and f
denotes the vector of faults to be detected. Suppose,
Nonlinear threshold selection. In connection with the moreover, there exist state transformations z = T(x, u)
observer-based residual generation scheme for nonlinear as well as output transformations ~ = p(y) allowing for
uncertain systems, Seliger and Frank suggest a thresh- a system description according to:
old selection technique allowing for a robust residual
evaluation in the case that a complete disturbance = Pz + +(z, u, u) + Or(x, u____el
___A)(x, u)d
decoupling of the estimation error is not possible 54. The Ox
(88)
residual will then differ from zero even if no faults -+ OT(x, u) K1 (x, u)f
occur.
In order to avoid false alarms, the residual must be
evaluated by some algorithm before a decision on a f: = p(y) = S(x, d , f ) (89)
fault is made. Since we do not make assumptions about
the dynamic properties or the statistics of the unknown We assume that there are several possible state and
inputs, noise and faults, the only practical solution is the output transformations for the above system description
selection of a threshold Jth different from zero. The fol- with, in general, different matrices F and nonlinearities
lowing simple decision logic can be employed: ,i,(-).
For the system description (88), (89), the residual
generator takes the form:
y(0
, Process k
z = F~ + * ( L )5, u, u), r = *(~, )5, u) (90)
:: AdaptiveGeneration where
q Adaptive
Detection *( T(x, u), p(x, O, 0), u) = 0 V x, u (91)
Filter :,
I t
r(t,~)
Transformation .. of J(v.)
~ J(s) = Fe + l-'(e, t) - O T E * ( t ) d - O T K * ( t ) f ,
OX OX (92)
r = qJ*(e, t, d,f)
Generator Logic
where
::...................................................... If ................. **(0, t, 0 , 0 , ) = 0 Vt, r'(o,t)=0 vt (93)
?
Alarm
Figure 5 Block diagram of the robust residual generation and eva- The following abbreviations have been used in the
luation algorithm in the frequency domain above equations:
Robust residual generation and evaluation methods. P. M Frank and X. Ding 415
I'(e, t) = *( T(x, u) + e, S(x, d,f), u, it) provided that, in accordance with the small-gain theo-
- Yo(T(x, u), S(x, d,f), u, it) + ( f - F)z (94) rem, the inequality (103) holds
The goal that is being pursued by introducing the II r U~c< O/eac II e I1~ +~a~ IId I1~ +~c~ Ilfli~ (1o5)
transformations T(x, u), P(Y) and, if necessary, also a
suitable feedback, is to obtain a partially linear estima- Substituting Equation (103) into Equation (105) yield
tion error system as stated in (92) which is asymptoti-
cally stable for f = d = 0.
The estimation error e can be written as Hrlioc <_ l-flocy~cdx+°~a~ Ildll~
where
This inequality can conveniently be used to define a
gF(t) = L-I{GF(S)} = L-I{(sI - F) -1} (98) detection threshold which excludes completely the pos-
sibility of false alarms due to disturbances. Notice that
usually there are no disturbances or unknown inputs of
denotes the impulse response of the linear system por-
arbitrarily high magnitude. It is therefore reasonable to
tion and * is the convolution operator. The term e0
exploit some a priori knowledge of the process under
denotes the response to initial conditions. Using L ~
norms one can write consideration in order to define the set of possible
unknown inputs:
of a norm of r. To this end, one may take the root mean Notice that to ensure that the above relation holds, one
square (rms) defined by must choose the matrix K such that
lo +'r 1 - ~ II Gt I1~> 0
J(v) - ((l/r) J rY(t)r(t)dt) 1/2 :=ll r I1~ (109)
Io Following the inequalities
The condition for J(z) is that in the absence of faults it II IT"(t)~ [[r= max 6(Vr(t)fs(t)) [] 0 lit (120)
t~[to,to+r i
should stay less than a threshold Jrn:
where K is some matrix whose function will be men- j [I r(s) I1~ds)dt) 1/2+ [I K I[~fl r 113
tioned below. This leads to
to
f
(123)
e( t) = (o(t)(- I r~°-r(s)r(s)ds + aO( t) ) + Kr( t) (116) t0+r t
to
<11 r lit ( 1[ Kl]o~ +( I J 62(~0(/)F~pT(s)~0(s)
t0 l0 (124)
Gj (s) = (I + 6(s)IC)-~6(s) (117)
F~o-r(t))dsdt) z/2) + max 6(~ov(t)co(t))3o
tE[to,to+r]
F r o m the small gain theorem we know that if the fol-
lowing inequality Write
II e II~<~ ~ II r Ib +~ (118)
& =ll x" II~,
tO+T l
(125)
/33 = ( I 1 6"2(~0(t)FcpT(s)~p(s)F~PT(t))dsdt)l/2
holds for some or, 3(> 0), then for the above system
l0 t0
II ~ I1~_<11G~ tim (~+ II IS'(t)~ I1~)/(1 - ~ II G1 I1~) ~2 = max t~(tpT(t)~0(t)), ot = fl0ot2 (126)
(119) tE[toto+r]
Robust residual generation and evaluation methods." P. 114.Frank and X. Ding 41 7
We then have an estimation for flr[l~ in fault-free case: Figure 7 shows a characteristic shape of the residual
associated with a fixed threshold. This bears the danger
)~,r I1~< ~(~oa,. +c~ao)l(1 - ~ , . -/~1/33) (127) of false alarms due to the fact that the thresholds have
to be chosen as small as possible because any increase of
based on which the threshold is defined by the threshold is associated with a loss of sensitivity to
faults. As a remedy, one can replace the crisp threshold
Jth = fll (~0012 -~" ~1120)/( ] -- ~1/~2 -- /~1~3) (]2g) by a fuzzy threshold. This means that the line which
constitutes the discrimination between zero and one in
Notice that the thresholds are expressed in terms of a the decision logic is replaced by an interval with prop-
number of constants; some of them are known or can be erly chosen upper and lower bounds and membership
off-line calculated, such as ill, f12, Oto,fin, but the others, functions defining the variables {zero} and {one} in a
fl3,~lcr2, are only on-line achievable, since they are fuzzy sense.
dependent on y(t) and u(t) (see the definition of The crisp set {zero} is replaced by a fuzzy set {zero}
P(t),~o(t)). This may cause some troubles with the characterized by a membership function lz(x) as, for
implementation, since the computation of singular example, shown in Figure 8. The parameter an is chosen
values may be too expensive to be carried out on-line as the level of noise to ensure that residuals within this
and the dependence of thresholds on the output vector range are definitely interpreted as zero. The parameter
y(t) may reduce the sensitivity to the faults and even takes into account the effects of disturbances and mod-
prevent a successful fault detection. To overcome these elling uncertainties beyond the noise level. In our case
difficulties, estimations for ~3~1, or2 are in some cases we have chosen for the S interval the standard triangle
needed. type of membership function.
The block diagram of the resulting adaptive residual The membership function of the fuzzy set {one} may
generation and evaluation system is depicted in Figure 6. be assigned as illustrated in Figures 9 and 10. In Figure 9,
the classical way of threshold logic decision is shown
Fuzzy threshold logic
Output',
Observer j
~(t)~+ _r(t)
ao -~ -a£ a'o ao+~ x
Figure 8 Membership function of the fuzzy set {zero) under consid-
! : | Es~mator I- eration of disturbances and noise
z ................................................
Adaptive Residual Generator
X~ fault
disturbance /l Threshold
Generator]
_.._,_j Threshold i I II ~ ~ Alarm
Selector .1
Adaptive ResidualEvaluator t
Figure 6 Concept of adaptive fault detection Figure 9 Classical threshold evaluation of a residual
418 Robust residual generation and evaluation methods: P. M. Frank and X. Ding
p.(x) 1 0
Comparison of fault detection systems
. . . . . . .
t
Figure l0 Definition of a membership function for the fuzzy set {one} In the last two sections, we have outlined recent advan-
ces of the theory of observer-based fault diagnosis with
again. Let the first maximum indicate a disturbance, emphasis on the latest contributions using frequency
and the second maximum a fault. It can be seen that the domain techniques, nonlinear unknown input observer
first maximum stays below the threshold but, if the and adaptive observer theory. In practice, we often face
residual slightly increases, it will surpass the threshold the problem of choosing a suitable FDI scheme from a
and cause a false alarm. great number of well developed approaches. This pro-
This type of false alarm can be avoided with a fuzzy blem has been recognized and studied by Ding et al. 34'6
threshold as illustrated in Figure 10. Beyond the noise and Seliger and Frank 54.
level, the truth value one is characterized by the mem-
bership function/z(x), as shown on the right-hand side Basic idea
of Figure 10. Here again, a standard triangle type is
chosen for/z(x) for ease of explanation. To outline the basic idea, recall that there will always be
It can be seen that if the residual due to the distur- a trade-off between the avoidance of false alarms on one
bance (peak 1) in Figure 10 (left side) increases slightly, hand and the detection of small faults which do not
this will cause just a small effect in the truth value of the cause the residual to surpass the threshold on the other
alarm rate. In contrast to the conventional case where hand. That faults which exceed the given tolerance
the small increase of the residual causes a false alarm, should be detected is the major practical requirement on
we now get a weak indication of inconsistency. Hence, a fault detection system. It is clear that a fault can be
we have replaced the yes-no decision by a continuous detected only if it causes the residual evaluation func-
indication of a faulty situation. tion to surpass the threshold, i.e:
The resulting membership function diagram of the
fuzzy variables {zero} and {one} is shown in Figure 11. II r lie >- J,h (129)
As one can see, there is an overlapping of the member-
ship functions of the fuzzy sets zero and one, which is Notice, however, that a fault may, due to the model
typical for the application of fuzzy logic. Therefore, one uncertainties, have different influences on the residual
has to define and evaluate rules as a basis to make the and, furthermore, on evaluation function. Taking this
final decision on the occurrence of a fault in a concrete into account we say a fault is detectable if
situation. This task can be carried out either by the
computer (using artificial intelligence) or by the human inf ]l r Ile~ Jth (130)
Ay
operator (using natural intelligence). Notice the simila-
rities to the statistical decision-making concept where,
instead of the membership functions, the probabilities INPUT RESIDUAL
are used and the decisions are made, for example by
maximum-likelihood ratio tests. In our case it is not
necessary to define the membership functions in terms VISUAL
of probabilities.
The fuzzy logic approach is illustrated in Figure 12.
By this procedure the human operator can make the
final decision involving all kinds of additional know- DIRECTPROCESS I FUZZYLOGIC I
ledge and experiences he has concerning the process. INFORMATION
4t
~(x)
{zero} {one}
1
i NATURALINTELLIGENCE I
EXPERTKNCWLEGE
7 FAULTDECISION
o a0 ao'l- ~ X~x X
Sf := { f : inf II r ]]<_>Jth} (131) In case that frequency domain rms is used as an eval-
Ay
uation function, Ding et al. 6"34 have shown that
It is evident that the size of the set Sf depends on the 2 maxo~, 8(,~a(a))Q(joo):~Iu(joo)Gd(jo)))
constriction of the residual generator, i.e. Q(s), the tl fmi. I1<=
minoo~ g(Q (rio) M,, (jo))Gf(j~) )
selection of the residual evaluation function I ] lie as
(138)
well as the threshold. A reasonable evaluation basis for
a certain fault detection system is certainly given by the
size of set Sf because the larger it becomes, the more where g(Gjeo) denotes the minimum singular value of a
faults are going to be detected. In this sense, we can say transfer matrix G(jo)).
that a fault detection system is optimal if the size of set On the basis of this scheme, Ding et al. 6,34 have fur-
Sf reaches its maximum. thermore proposed an approach to designing an opti-
mal fault detection system. The core of it is to minimize
Comparison of linear fault detection systems the minimum detectable fault, which is equivalent to
solving the following optimization problem:
To avoid the difficulty of determining the size of Sf,
Ding et al. 6'34 suggested the use of the concept of mini- inf IIf~,;. lie = inf 2 maxo,~ 8(aa(w)R(joo)f4u(joo)Ga(jo)))
mum detectable faults which was introduced by Emami- Q(s) Q(s) min,oe~(Q(jo))~I,(joo)Gr(fio) )
Naeini et al. 8
Minimum detectable fault, f,,i,, is defined by
(139)
inf
Ay
II r li< = inf II Q~Io(GyfL + GadL) I1~
Suppose now that the right-hand side of Equation (106)
=1) Olfl~@fL I1~ - s u p 11QM~GddL lie which is the upper bound of the L:c-norm of the resi-
d
dual equals the detection threshold Jth; i.e.
(134)
( .,Socae~ )
and the threshold is given by l _ fl~g ca~c+ °taoc II d {{~
(142)
],h = sup II QM,,Ga& lie (135)
d + 1 7/7~--y~ cs~+°q~ IIf(l~=J*h
(145) where Q1 (t), Q2(t) are incoming mass flows, hi (t), h2(t),
{ f E R f : l l f l l ~ < fo} C SZ
h3 (t) are the water levels of each tank and measured and
QfI(t), Qf2(t) and Qf3(t) denote faults representing
These faults are therefore undetectable. In other words,
undesirable mass flows into the tanks caused by leaks or
f0 defines the minimum detectable fault. As mentioned
plugging in the various tanks or pipes. The three circular
in the last subsection, in order to reduce the size of the
tanks have the same cross-section A and are intercon-
set Sf, which is equivalent to increasing the set Sf of
nected via circular pipes with cross-sections s~3, s23. The
detectable faults, one must therefore reduce the mini-
outlet pipe is also circular with cross-section so. al, a2,
mum detectable fault f0 which, in this sense, can be
a3 are scaling constants and g is the gravity constant.
understood as a performance index describing the qual-
Based on the system description given above,
ity of the residual generator.
Wtinnenberg had developed a nonlinear observer-based
The performance index is a function of the constants
scheme to achieve a successful fault detection and iso-
which previously have been defined in order to evaluate
lation 26. The key to this scheme is the construction of
upper bounds for the nonlinearities and the residual.
three nonlinear observers described by
These constants are in turn directly related to the
transformations the residual generator is based on and 1
are therefore strongly affected by the choice of the Zl = - l l z l + A (Q1 - alsl3sgn(hl - h3) x/2g I hl -h3 ])
transformation. According to Equation (143), the per-
- llhl,rl = ll(hl - zi)
formance index will be mostly affected by the constants
Cd~c and c~d~ versus C f o c and ¢xfoo.
By minimizing this performance index (143), one
could find an optimal residual generator. Due to the
complexity of this minimization problem which is
mainly caused by the nonlinearity of the models, it is Pumpl Pump2
virtually impossible to derive a generally applicable,
Systematic optimization strategy. This restricts the per- Ql(t) A Q2(t)~ ~::
formance index to compare different residual generators
rather than to find systematically an optimal solution.
E x a m p l e s o f application
1
zz9 = --12Z2 + -'~ (Q2 - a3s23sgn(h3 - h2)v/2g { h3 - h2 I an industrial robot which was o p e r a t i n g under real-life
conditions•
- a2so 2 v / ~ 2 ) - 12h2, r2 = 12(h2 - z2) The mathematical model o f a r o b o t is normally given
in the form:
1
33 = -13z3 + -~ (alsl3sgn(hi - h3)x/2g [ h~ - h3 r J(q(t))ij(t) = xd(q(t), q(t)) + xg(q(t)) + f(u(t)) (149)
The resulting friction characteristics for two axes of sketched in Figure 18a could be detected reliably
the M A N U T E C r3 robot are shown in Figure 17. Dif- (Figure 18b), giving hints on critical situations and pos-
ferent measurement cycles are performed and repeated sibly severe crashes.
and the friction force is plotted versus generalized posi- The detection of external interaction proves to be
tions (positive friction branch corresponds to a positive robust enough so that it can even be used for fault
velocity and vice versa). In Figure 17a, partial wear-out identification, i.e. to measure external forces and tor-
at angles of q = 40 ° and q = 100° can be observed. This ques. From a comparison of reconstructed (observer
may result from former intensive and heavily loaded technique) with measured (force/torque sensor) torques
robot movements at specific configurations. Figure 17b that were externally applied, one can conclude that the
reveals a periodic oscillation which results from defects observer-based technique is a reliable method to gain
in train gears (here a ratio of n = 5) and which might be additional process information even in terms of quanti-
due to misplaced bearings, etc. Notice the perfect tative figures. This is demonstrated in Figure 19.
reproducibility of the friction characteristic. The Figure 19a shows the velocities of the test trajectories,
threshold can therefore be selected so as to adapt to this and Figure 19b gives the measurement result, where the
characteristic or, in other words, the characteristic can identified torques (dotted line) and measured torques
be compensated 61"62. (solid line) are plotted. The accuracy of the results
The observer-based fault detection system with com- obtained by identification lies within a :t: 2 Nm range.
pensated friction characteristic is next used to gain In summary, this example shows that the observer-
additional process information by detecting and moni- based technique can be applied successfully to robots to
toring external torques. If the friction characteristic is monitor and analyze friction and, using a compensation
compensated in the residual, then changes are solely scheme, detect small external torques immediately.
related to interactions with the environment. Such Combining both features, the observer-based technique
interactions may be contact forces during mounting or can be used as a powerful tool for fault detection and
fitting operations, soft collision (no emergency shut process supervision.
down), loss of pay load, etc.
In the laboratory of our industrial partner, a test
environment has been installed where a robot performs Strategies of implementation
some standard transport and mounting operations in
order to study the power of the detection system. A fault detection system can be implemented using
As a result, intentionally introduced faults could be either analytical or knowledge-based techniques or
detected. For example, a soft collision with a flexible combinations of both, (Figure 20). Analytical residual
spring brought into the regular path of the gripper as generators and evaluators are the core of analytical
(a) ~° (b) ~
6~
40
40
zg2~
._g o
t~ -20
~- -2Q
-°°I -40
]
-80 -60
-200 -1~0 -100 -50 0 50 200 -150 - ~oo - 50 0 50 1oo i bo
(a) (b) 50
E
b z lo
"o
5
g:
2 2.5 3 3.5 4 45
Time Is]
'i
6[-
/,'.,,^ ,: ,'i ,
E
Z
"i !! ipI
_aJ , . , , . . . . .
q -2i
0 1 2 3 4 5 6 7 8 9 ~0 C 2 5 4 5 6 7 8 9 '0
Figure 19 (a)Velocities of the test trajectory (axes 1 to 3) (b) Measurement (solid line) and fault identification (dotted line) of externally applied
torques to axis 1
tool-boxes. The inclusion of artificial intelligence leads to and practical applications. Though the practicability of
the concept of diagnosis expert systems in which com- the outlined methods is basically out of question, their
monly analytical and heuristic information and knowl- practical significance is still an open question. However,
edge processing are combined. While expert systems the examples discussed in the paper give rise to great
have not been very successful for control, they are encouragement and may help to motivate intensive future
widely used for fault diagnosis systems4°-42,44. An alter- efforts towards the practical application of these ideas.
native approach is the implementation of combining the
analytical techniques with the natural intelligence of the
human operator. This leads to a computer assistant Acknowledgements
human supervisory concept as suggested by Frank and
Kiupe112. This technique is the most powerful available The authors owe gratitude to Mr Kiupel, Mrs Seliger-
as long as natural intelligence is superior to artificial K6ppen and Dr Seliger for their scientific contributions
intelligence, which is still the case. and to Mrs Appelt and Mr G6bel for their support in
the production of the paper.
The paper is dedicated to the memory of Jtirgen
Conclusion Wfinnenberg, a wonderful person and most talented
scientist, who was much too early taken away from us
The paper presents a survey of advances in the theory of forever.
observer-based fault diagnosis. Because of the current
tremendous research activity in this field, it was
not possible to provide a comprehensive representation References
of the scene. We have therefore focused on those
1. Prock, J., Signalvalidierung mittels analytischer Redundanz.
contributions which we think close a gap in the existing Gesellschaft ffir Reaktorsicherheit (GRS). Rep. No. GRS-A-1482,
theory and may gain some relevance for future research 1988.
2. Frank, P. M., Robust model-based fault detection in dynamic
systems. Preprints of IFAC Symposium: On-line Fault Detection in
the Chemical Process Industries, Newark, Delaware, 1992, pp. l-13.
,' MODEL-BASED • ( MODEL-BASED • 3. Frank, P. M. and Wfinnenberg, J., Robust fault diagnosis using
~..,..~ROA OH unknown input observer schemes. In Fault Diagnosis in Dynamic
Systems, ed. R. J. Patton, P. M. Frank and R. N. Clark. Prentice
Hall, 1989, pp. 46-98.
4. Gertler, J., Analytical redundancy methods in fault detection and
HUMAN OPERATOR
• EXPERTSYSTEM DIGITAL COMPUTER isolation. Proceedings of the IFAC/IMACS Symposium SAFE-
PROCESS 91, Baden-Baden, 1991, pp. 9-21.
i COMPUTER+ COMPUTER + 5. Patton, R. J. and Chen, J., A review of parity space approaches to
NATURAL HUMAN HEURISTIC ANALYTICAL fault diagnosis. Proceedings of the IFAC/IMACS Symposium
RESOURCES KNOWLEDGE INFORMATION SAFEPROCESS, 91, Baden-Baden, 1991.
PROCESSING PROCESSING 6. Ding, X,, Frank, P. M. and Guo, L., An approach to residual
I
generator and evaluator design and synthesis. Proceedings 12th
i 1
NATURAL "i ( ARTIFICIAL "i (SYSTEM "1
IFAC World Congress, Sydney, Australia, 1993.
7. Ding, X., Guo, L. and Frank, P. M., Parameterization of linear
L'NTELLIGENCEJ [ INTELLIGENCEl , ~ THEORY J observers and its application to observer design. IEEE Trans.
Autom. Contr,, 1994, AC-39, 1648-1652.
8. Emami-Naeini, A., Akhter, M. M. and Rock, S. M., Effect of
;D'EXPERTi FANA 'C I model uncertainty of failure detection: the threshold selector.
( sY~M] j :~Di ! IEEE Trans. Autom. Contr., 1988, AC-33, 1106--1115.
L joQ Box iJ 9. Frank, P, M., Fault diagnosis in dynamic systems using analytical
and knowledge-based redundancy- a survey. Automatica, 1990,
t COMP~ER~AGSIS~ED i 26, 459-474.
HUMAN;:SU PE~VBoR ; Jl 10. Frank, P. M., Enhancement of robustness in observer-based fault
detection. Proceedings IFAC/IMACS Symposium SAFEPRO-
Figure 20 Strategies for implementation of fault diagnosis systems CESS "91, Baden-Baden, 1991, 1, pp. 275-287.
424 Robust residualgeneration and evaluation methods."P. M. Frank and)(. Ding
11. Frank, P. M. and Seliger, R., Fault detection and isolation in 37. Ding, X., Frequenzbereichsverfahren zur beobachtergestuetzten
automatic processes. In Control and Dynamic Systems, Advances Fehlerentdeckung. Diss. Universitaet Duisburg, Fachgebiet Mess-
in Theory and Applications, Vol. 49, Part 5 of 5, ed. W. Leondes, und Regelungstechnik 1991. VDI Fortschrittsbericht Reihe 8,
Academic Press, 1991, pp. 241-287. Mess-, Steurungs- und Regelungstechnik, Nr. 295, 1992.
12. Frank, P. M. and Kiupel, N., Fuzzy supervision for lean produc- 38. Marquez, H. J. and Diduch, C. P., Sensitivity robustness in fail-
tion. Proceedings 6th IAR Colloqu. Duisburg, 1992, pp. 163-177. ure detection: A frequency domain approach. Proceedings 29th
13. Gertler, 3. and Singer, D., A new structural framework for parity IEEE CDC, Honolulu, USA, 1990.
equation based failure detection and isolation. Automatica, 1990, 39. Frank, P. M., Advanced fault detection and isolation schemes
26, 381-388. using nonlinear and robust observers. Presented at 10th IFAC
14. Gertler, J., Structured residuals for fault isolation, disturbance World Congress, Munich, 1987.
decoupling and modelling error robustness. IFAC Symposium On- 40. Chang, S. J., Dicesare, F. and Goldbogen, G., Failure propa-
line Fault Detection and Supervision in the Chemical Process gation trees for diagnosis in manufacturing systems. IEEE
Industries, Newark, Delaware, April 1992. Transaction on Systems, Man, and Cybernetics, July 1991, 21(4),
15. Hou, M. and Muellers, P. C., Design of observers for linear systems 767-776.
with unknown inputs. 1EEE Trans. Aut. Contr., 1992, 37(6), 871-875. 41. Isermann, R., Fault diagnosis of machines via parameter estima-
16. Lou, X. C., Willsky, A. S. and Verghese, G. C., Optimally robust tion and knowledge processing. SAFEPROCESS '91, Baden-
redundancy relations for failure detection in uncertain systems. Baden, 1991, 1, 121-133.
Automatica, 1986, 22(3), 333-344. 42. Lunderstaedt, R. A., Knowledge-based sensor fault detection for
17. Massoumnia, M. A., A geometric approach to the synthesis of fail- gas turbines under consideration of model-based methods. IFAC/
ure detection filters. IEEE Tram. Aut. Contr., 1986, 31,839-846. IFIPI/MACS Symposium, Delft, June 1992.
18. Murad, G. A., Postlethwaite, I. and Gu, D.-W., A robust design 43. Sauter, D., Aubrun, C., Noura, H, and Robert, M., Fault diag-
approach to integrated controls and diagnostics. Proceedings 13th nosis and reconfiguration of systems using fuzzy logic; application
IFAC World Congress, 1996. to a thermal plant. Proceedings 6th 1AR Colloquium, University of
19. Nett, C. N., Jacobson, C. A. and Miller, A. T., An integrated Duisburg, 1992, pp. 177-190.
approach to controls and diagnostics. Proceedings ACC, 1988. 44. Tzafestas, S. G., System fault diagnosis using the knowledge-
20. Niemann, H. H. and Stoustrup, J. Filter design for failure detec- based methodology. In Fault Diagnosis in Dynamic Systems,
tion and isolation in the presence of modeling errors and distur- Theory and Application, 1989, Prentice Hall, 1989, pp. 509-527.
bances. Proceedings 35th CDC, 1996. 45. Ge, W. and Fang, C. Z., Detection of faulty components via
21. Olin, P. M. and Rizzoni, G., Residual generation for fault detec- robust observation. Int. J. Control 1988, 47(2), 581-599.
tion and isolation in the joint time-frequency domain. Proceed- 46. Viswanadham, N., Taylor, J. H. and Luce, E. C., A frequency
ings Int. Conference on Fault Diag., Toulouse, France, 1993. domain approach to failure detection and isolation with applica-
22. Patton, R. 3., P. M. Frank, and R. N. Clark (eds), Fault diagnosis tion. Control-theory and Advanced Technology, 1987, 3, 45-72.
in dynamic systems, theory and application. Prentice Hall, 1989. 47. Hou, M. and Patton, R. J., An LMI approach to tt/Ho~ obser-
23. Patton, R. J. and Kangethe, S. M., Robust fault diagnosis using vers. Proceedings IEE Conference Contr. '96, England, 1996.
eigenstructure assignment of observers. In Fault Diagnosis in 48. Hamelin, F. and Sauter, D., Robust residual generation for FDI
Dynamic Systems, ed. R. J. Patton, P. M. Frank and R. N. Clark, in uncertain dynamic systems. Proceedings the 34th CDC, New
Prentice Hall, 1989, pp. 99-154. Orleans, USA, 1995.
24. Patton, R. J. and Chen, J,, Optimal selection of unknown input 49. Khargoneckar, P. P. and Ting, T. L., Fault detection in the pre-
distribution matrix in the design of robust observers for fault sence of modeling uncertainty. Proceedings the 32rid CDC, San
diagnosis. Proceedings IFAC/IMACS Symposium, SAFEPRO- Antonio, USA, 1993.
CESS, "91, Baden-Baden, 1991, pp. 221-226. 50. Qiu, Z. and Gertler, J., Robust FDI systems and H~-Optimiza-
25. Viswanadham, N. and Srichander, R., Fault detection using tion. Proceedings the 32nd CDC, San Antonio, USA, 1993.
unknown input observers. Control Theory and Advanced Techno- 51. Tyler, M. L., Optimal and robust design of integrated control and
logy, 1987, 3(2), 91-101. diagnostic modules. Proceedings ACC, USA, 1994.
26. Wiinnenberg, 3, Observer-based fault detection in dynamic sys- 52. Frank, P. M. and Ding, X., Frequency domain approach to
tems. Diss. Universitaet Duisburg, Fachgebiet Mess- und Rege- optimally robust residual generation and evaluation for model-
lungstechnik 1990. Fortschrittsberichte VDI, Reihe 8, Nr. 222, based fault diagnosis. Automatica, 1994, 30, 789-904.
VDI Verlag Duesseldorf, 1990. 53. Seliger, R. and Frank, P. M., Fault diagnosis by disturbance
27. Brunet, J., 3aume, D., Labarrere, M., Rault, A. and Verge, M., decoupled nonlinear observers. In Proceedings 30th IEEE CDC,
Detection et diagnostic de pannes. Hermes, Paris, 1990. Brighton, England, 1991, Vol. 3, pp. 2248-2253.
28. Viswanadham, N., Sarma, V. V. S. and Singh, M. G., Reliability 54. Seliger, R. and Frank, P. M., Robust residual evaluation by
of computer control systems. North Holland Systems and Control threshold selection and a performance index for nonlinear obser-
Series, North-Holland, 1987, 8. ver-based fault diagnosis. International Conference on Fault
29. Ding, X. and Frank, P. M., Fault detection via factorization Diagnosis (Tooldiag '93), 1993, Toulouse, France.
approach. Syst. Contr. Lett, 1990, 14, 431-436. 55. Seliger, R. and Frank, P. M., Robust component fault detection
30. Ding, X and Frank, P. M., Fault identification filter design via Hoo- and isolation in nonlinear dynamic systems using nonlinear
optimization techniques. Proceedings IFAC Symposium on Identifi- unknown input observers. Preprints of the IFAC/IMACS Sympo-
cation and System Parameter Estimation, Budapest, July 1991. sium SAFEPROCESS "91, Baden-Baden, 1991, 1, pp. 313-318.
31. Ding, X., Frank, P. M. and Guo, L., Nonlinear observer design 56. Frank, P. M., Ding, X. and Wochnik, J., Model-based fault
via an extended observer canonical form. Syst. Contr. Lett., 1990, detection in diesel-hydraulically driven industrial trucks. Pro-
15, 313-322. ceedings ACC., AACC, Boston, 1991, 2, pp. 1528-1533.
32. Ding, X., Frank, P. M. and Guo, L., Fault detection via adaptive 57. Basseville, M., Detecting changes in signals and systems-A sur-
observers based on orthogonal functions. Proceedings IFAC vey. Automatica, 1988, 24, 309-326.
Symposium on Advanced Information Processing in Automatic 58. Sauter, D,, Contribution fi l'etude des methodes de detection de
Control Nancy, 1989. Pergamon, 1990, pp. 95-100. rupture de model. These de Docteur es Sciences Physiques, Uni-
33. Ding, X. and Frank, P. M., An adaptive observer-based fault versite de Nancy I, CRAN, 1991.
detection scheme for nonlinear dynamic systems. Proceedings 59. Bastin, G. and Gevers, M. R,, Stable adaptive observers for non-
12th IFAC World Congress, 1993. linear time-varying systems. IEEE Trans. Contr., 1988, AC-33,
34. Ding, X., Guo, L. and Frank, P. M., A frequency domain 650-658.
approach to fault detection of uncertain dynamic systems. Pro- 60. Marino, R., Adaptive observers for single output nonlinear syst-
ceedings 32nd CDC, 1993, pp. 1722-1727. ems. IEEE Trans. Autom. Contr, 1990, AC-35, 1054-1058.
35. Ding, X. and Frank, P. M., Fault detection and Identification via 61. Schneider, H. and Frank, P. M., Observer-based fault detection
frequency domain observation approaches. In Mathematical and of robots enhanced by characteristic curves. IMACS/SICE Inter-
Intelligent Models in System Simulation, ed. R. Hanus, P. Kool national Symposium on Robotics, Mechatronics and Manufac-
and S, Tzafestas. J. C. Baltzer AG Sc. Publ. Co., IMACS, 1991, turing Systems, 1992, Kobe, Japan.
pp. 471-476. 62. Schneider, H., Fritsch, M. C. and Niedermayr, E., Application of
36. Ding, X. and Frank, P. M., On-line fault detection in uncertain model-based force/torque observer to manufacturing processes.
systems using adaptive observers. European J. of Diagnosis and North American Manufacturing Research Conference, NAMRC
Safety in Automation, 1993, 3, 9-21. XXI, 1993, Stillwater, Oklahoma, USA.