You are on page 1of 5

TA4 9:15 -

MODELBASEDFAULTDETECTION
AND DIAGNOSIS
METHODS
Rolf Isermann
Institute of Automatic Control, Technical University of Darmstadt, Landgraf-Georg-Str. 4, D-64283 Darmstadt, Germany
Phone: +49 6151 1632114, Fax: +49 6151 293445, E-mail: isermanncairtl. rt .e - t e c h n i k . t h - d a r m s t a d t .de

ABSTRACT the operating point. In the case of closed loops changes in the
For the fault detection of technical processes different methods process are covered by control actions and cannot be detected
can be applied based on the information extracted from direct from the control variable, as long as the manipulated process
measured signals, from signal models and process models. inputs remain in the normal range and not touch restrictions.
Examples for signal model based fault detection methods are The big advantage of the classical limit-value based supervision
spectral analysis or parameter estimation of ARMA models, methods is their simplicity and reliability. However, they are
examples for process model based methods are parameter esti- only able to react after a relatively large change of a feature,
mation, state estimation or parity equation approaches. A com- i.e. after either a large sudden fault or a long lasting gradual
parison of these methods shows that they have different prop- increasing fault. In addition an in-depth fault diagnosis is
erties with regard to the detection of faults in the process, the usually not possible.
actuators and the sensors. By a proper integration of different Therefore (c) advanced methods of supervision and fault
fault detection methods mainly their advantages can be used to diagnosis are needed which satisfy the following requirements:
generate a number of different analytical symptoms. (i) Early detection of small faults with abrupt or incipient time
For fault diagnosis a knowledge based procedure is required, behaviour.
because also qualitative information in form of heuristic symp- (ii) Diagnosis of the faults in the actuator, process components
toms have to be taken into account. Based on heuristic process or sensors.
knowledge as fault-symptom causalities and a unified represen- (iii) Detection of faults in closed loops.
tation of all symptoms an integratedfault diagnosis can be per- (iv) Supervision of processes in transient states.
formed. This comprises the treatment of the symptoms as un- The goal for the early detection and diagnosis is to have
certain facts and approximate diagnostic reasoning via if-then enough time for counteractions like other operations, triggering
rules either in a probabilistic or a fuzzy-logic (possibilistic) of redundance, configuration, maintenance or repair. The earlier
frame. detection can be reached by gathering more information,
The described methodology was verified by experiments with especially by using the relationship between the measurable
several technical processes like electr. motors, actuators, quantities in the form of mathematical models. For fault diag-
pumps, machine tools, robots, heat exchangers, combustion nosis the knowledge of cause-effect relations has to be used. A
engines and vehicles general scheme for all supervisory functions and resulting
Keywords. Fault detection, supervision, monitoring, parameter actions is given in [l].
estimation, state estimation, parity equations, fault diagnosis,
diagnostic reasoning, fuzzy reasoning. 2. FAULTDETECTION AND FAULTSDIAGNOSIS
Fig. 1 shows an overall scheme of knowledge based fault
1. INTRODUCTION detection and diagnosis. The main tasks can be subdivided in
Within automatic control oftechnical systems supervisory func- fault detection by analytic and heuristic symptom generation
tions serve to indicate undesired or unpermitted process states and fault diagnosis [2].
and to take appropriate actions in order to maintain the
operation and to avoid damages or accidents. Following 2.1. Analvtic symptom generation
functions can be distinguished: The analytical knowledge on the process is used to produce
quantifiable,analytical information.To this based on measured
(a) monitoring:measurable variables are checked with regard to process variables a data processing has to be performed to
tolerances and alarms are generated for the operator. generate first characteristic values by
(b) automatic protection: in the case of a dangerous process - limit value checking of direct measurable signals. Charac-
state, the monitoring function initiates automatically an teristic values are exceeded signal tolerances.
appropriate counteraction. - signal analysis of directly measurable signals by use of signal
(c) supervision with fault diagnosis: based on measured models like correlation functions, frequency spectra,
variables features are calculated, symptoms are generated via autoregressive moving average (ARh4A). Characteristic values
change detection, a fault diagnosis is performed and decisions are e.g. variances, amplitudes, frequencies or model parameters.
are made for counteractions. - process analysis by using mathematical process models
together with parameter estimation, state estimation and parity
The classical methods (a) and (b) are suitable for the overall equation methods. Characteristic values are parameters, state-
supervision of the processes. Most frequently simple limit variables or residual.
value checking is applied which works especially well if the In some cases then specialfeatures can be extracted from these
process operates approximately in a steady-state. However, the characteristic values, e.g. physical defined process coefficients
situation becomes more involved if the process changes rapidly or special filtered or transformed residuals. These features are

1605
ANALYTICAL
KNOWLEDGE
\

HEURISTIC
KNOWLEDGE
\

Fig. 1 Scheme of a knowledge-based fault detection and diagnosis [9].

then compared with the normal features of the non-faulty process including actuators and sensors by measuring the
process. For this, methods of change detection and available input and output variables u(t) and Y(t). The process
classification are applied. As analytic symptoms the resulting is considered to operate in open loop. A distinction can be
changes (discrepancies) of the described direct measured made between static and dynamic, linear and nonlinear process
signals, signal models or process models are considered. models.

2.2 Heuristic symptom generation 3.1 Fault modeling


In addition to the symptom generation with quantifiable Processes with lumped parameters which can be linearized
information heuristic symptoms can be produced by using around one operating point are usually described by an
qualitative information from human operators. Through human ordinary differential equation equation
observation and inspection heuristic characteristic values in
y(t) + a,y(')(t) + ... + ay(")(t)
form of special noise, colour, smell, vibration, wear and tear,
etc. are obtained. The process history in form of performed = b,u(t) + b,u(l)(t) + ... + b,u('")(t) (1)
maintenance, repair, former faults, lifetime, load measures
constitute a further source of heuristic information. Statistical y(t) = Y(t) - Y, , u(t) = U(t) - U, (2)
data (e.g. MTBF, fault probabilities) achieved from experience where Uoo,Y,, are known steady-state (or d.c.) values and
with the same or similar processes can be added. By this way y("'(t)=d"y(t)/dt".
heuristic symptoms are generated which can be represented as The process model can be written in vector form
linguistic variables (e.g. small, medium, large) or as vague Y(t) = *=to 8 (3)
numbers (e.g. around a certain value).
with the parameter vector and data vector
2.3. Fault diagnosis -O T = [ U , ... U" bo ... b, ] (4)
The task of fault diagnosis consists in determining the type, $'(t) = [ -y(l)(t) ... -y(")(t)u(t) ... u("(t) ] (5)
size and location of the fault as well as its time of detection
based on the observed analytical and heuristic symptoms. With Offset faults f,(t) and f,(t) at the input and the output signal
the aid of heuristic knowledge in form of heuristic process then can be modelled by
models (qualitative models), fault-symptom causalities and ~ ( t =) S'Ct) 8 AgT(t)B f,(t)
+ +
(6)
weighting of effects different diagnostic reasoning strategies
can be applied. Finally by a fault decision the type, size and AST(t) = [ -y")(t) ... -y'")(t) f,(t) ... f ) ( t ) ] ('1
location of the most possible fault as well as its time of detec-
tion is indicated. For parameter faults A€),, it holds
y(0 = S'(0 C Q + AW) 1 (8)
3. MODEL-BASED FAULT DETECTION METHODS The considered linear process in state-space form is
Different approaches for fault detection by using mathematical
models were developed in the last 20 years, see. e.g. P (0 = c I (Q (10)
PI, i41, [si,[61>[71>
181. with p input signals g(t) and r output signals y(t). Offset
The task consists in the detection of faults in the technical changes &(t) of the states &() and &(t) of the output y(t) are

1606
then modelled by with reference to the normal values. Usually the time instant T,
a0 = A_ d t ) B d t ) + L fL(0
+ (11) of faults occurrence is unknown. In order to detect these
Y (0 = € x (0 "fM (0 + (12) changes methods of change detectzon are used, e.g. as a
likelihood-ratio-test or Bayes decision, a run-sum test or a two-
If faults appear as parameter changes AAA AB, AC, it holds probe t-test.
2(t) = [ 4 + AA 3 x(t) + [ B + AB ] u(t) (13) When several variables change, classiJicationmethods are used.
r(0 = [ C A€ 1 &(t) + (14) In a multidimensional space the symptom vector
AS'.= [AS AS, ... AS,,] (17)
3.2 Fault detection methods moves in certain &ections which depend on the fault. In this
Three important model based fault detection methods are case the process of symptoms generation consists of
shown in Table 1. For a detailed description of these methods determining the direction as well as the distance of & from
see e.g. [1],[6],[7],[8]. They generate residuals of the following the origin or another reference vector Soof the normal state. To
form: this geometrical distance methods or artificial neural networks
(9 parameter estimation: can be applied.
changes of parameter estimates Ag
or viaj=f(Q changes of process coefficients AE 3.4 Combination of different fault detection methods
(ii) state estimation: As shown in [6] parameter estimation on the one side and state
changes of state estimates A& (t) estimation and parity equations on the other side show
or output errors e(t) = y(t) - C A ( ) advantages and disadvantages with regard to the detection of
or filtered output errors r(t) = W e(t) the various types of faults. Therefore, if all faults should be
(iii) parity equations: detectable, different detection methods should be integrated
output error e'(s) = y(s) - G&) u(s) properly in order to mainly use their advantages. As in most
polynomial error e(s) = AM@)y(s) - BM(s) u(s) cases the model parameters are unknown anyhow it is quite
With regard to these residuals the faults can be classified into natural to apply first parameter estimation. Then following
two catagories: combinations of model-based detection methods result [6].
(a) Additive faults (signal changes) I) Sequential parameter and state estimation
- differential equation: f,(t): input fault; $(t): output fault - parameter estimation to obtain the model
- state-space vector equation: &(t): input fault or state fault; - state estimation for fast change detection
fJt): output fault. - parameter estimation (on request) for deep fault
(b) Multiplicative faults (parameter changes) diagnosis
- differential equations: Aq(t), Ab, (t) 11) Sequential parameter estimation and parity equations [ 101
- state-space vector equation: AA(t), AB(t), Ac(t) - parameter estimation to obtain the model
Additive faults change the normal deviations of the residuals by - parity equations for change detection with less
a summation and multiplicative faults by a product with time computations [181
dependent process variables. - parameter estimation (on request) for deep fault
The suitability of the different fault detection methods was diagnosis
discussed in [6]. For single-input single-output processes the 111) Parallel parameter and state estimation
results can be summarized as follows. As parameter estimation - for multiplicative and additive faults
is especially suitable for multiplicative faults this detection - depending on input excitation
method can be primarily be recommended for corresponding The way of integration depends very much on the process, the
faults in the processes and faults which change the dynamics of faults to be detected and the allowable computational effort.
actuators and sensors. But also additive faults at the input and In some cases also the integration of process model based and
output can be included in the parameter estimation, as for static signal model based detection methods gives a good overall
actuator and sensor faults. State estimation and parity equations information:
have their advantages for additive faults and are therefore IV) Parameter estimation and vibration analysis
feasible for corresponding faults in the sensors, actuators and in - parameter estimation for parameter mapping faults
some cases for processes. For multi-input multi-output - vibration analysis for other type of faults like
processes the analytical redundancy between the measured unbalance, knocking, chattering
inputs and output increases. This seems to be especially (This is especially attractive for rotating machines)
advantageous for the detection of sensor faults where the real
input signal is unknown and for actuator faults, if the actuator 4. FAULT DIAGNOSIS METHODS
output is not measurable. However, on the other hand it is The task of fault diagnosis consists in the determination of the
more difficult to obtain precise process models with all type of fault with as much as possible details such as the fault
crosscouplings for multivariable processes. size, location and time of detection. The diagnosis procedure is
based on the observed analytical and heuristic symptoms and
3.3 ChanPe detection and symptom generation the heuristic knowledge on the process, as shown in Fig. 1.
The measured or estimated quantities like signals, parameters, In this section the heuristic part of the knowledge and inference
state-variables or residuals are usually stochastic variables Si(t) mechanisms for diagnosis are described in order to build up
with mean value and variance on-line expert systems for fault dzagnoszs [2],[9].
-
si = E(si(t)) ; .: = 4]*)(15)
E{[S,(t)
4.1 Symptom representation
as normal values for the non-faulty process. Analytic symptoms The inputs of the knowledge based fault inferencing mechanism
are then obtained as changes are all available symptoms as facts and the fault relevant
ASi E(si(t) - ;A = EIa,(t) - ZJ t > TF(16) knowledge on the process, mostly in heuristic form.

1607
Table 1 Model based fault detection methods for linear processes
lockdiagram asic mathematic JEstimator pemarks
lode1 besiduals I

!!=[a,...b,...] cients:
1) -g
A&+ l)=b(k+ p=f -m
A&+ l)=fi(k+1) -E

fault detection

.:. observe

qs)=&= Output error:


4s) e'(s)=y(s)-G&s)u(s)
_..
B&)- bo+blS+... olynomial error:
1+ais+... -B&s)u(s)

I
e 4Js)
r(s)=GAs)e(s) Specific properties I.
filtering with G,

a) Analytical symptoms Fig. 1.


The analytical symptoms ASa are the results of the limit value
checking of measurable signals, signal or process model fault 4.2 Heuristic knowledge representation
detection methods and change detection, as described in sect. 3 . For establishing heuristic knowledge bases for diagnosis several
b) Heuristic symptoms approaches do exist, see e.g. [11],[12]. In general specific rules
Heuristic symptoms AShl are observations of the operating are applied in order to set up logical interactions between
personal in form of acoustic noise, oscillations or optical observed symptoms (effects) and unknown faults (causes). The
impressions like colours or smoke obtained by inspection. propagation from appearing faults to observable symptoms in
These empirical facts can usually only be represented in form general follows physical cause-effect relationships where
of qualitative measures, e.g. as linguistic expressions like physical properties and variables are connected to each other
"little", "medium" or" much". quantitatively and also as functions of time. However, the
c) Process history and fault statistics underlying physical laws are frequently not known in analytical
A third category of facts depends on the general status based form or too complicated for calculations. Therefore heuristic
on the history (past life) of the process. This process history knowledge in form of qualitative process models can be
includes the past information of running time, load measures, expressed in form of rules like
last maintenance or repair. Dependent on the quality of these
measures they can be used as analytical or heuristic symptoms. IFxcondition> THEN <conclusion> (18)
The symptoms based on process history will be called ASp,. The condition part (premise) contains facts in the form of
d) Unified symptom representation observed symptoms ASi as inputs and the conclusion part
For the processing of all symptoms in the inference mechanism includes events E, and faults F, as a logical cause of the facts.
it is advantageous, to use a unified representation. One This procedure results in fault-symptom trees ("directed
possibility is to present the analytic and heuristic symptoms graphs"), relating symptoms to events and faults.
with confidence numbers 0 I c(SJ I 1 and treatment in the Then symptoms or events are associated by AND and OR
sense of probabilistic approaches known from reliability theory connectives and the rules become of form
[2]. Another possibility is the representation as membership IF < [ t ,AND t,+lAND ... 4,] OR
function 0 I p(S,) I 1 of fuzzy sets [9], Fig. 2. By these fuzzy
[ € , I AND t , ~AND+ ~ t,/+ll OR (19)
sets and corresponding membership functions all analytic and
[ ... ] > THEN < [qk] >
heuristic symptoms can be represented in a unified way within
the range 0 2 ~ ( s , 5 ) 1. This integrated symptom In the classical fault-tree analysis the symptoms and events are
representations are then the inputs for the inference mechanism, considered as binary variables with cl=l for happened and E,,=O

1608
connectives and the rules become of form (24)
"P(q) = 1- P(4,)
IF < [ti AND AND OR ... e,]
[ € i / AND Sil+i AND €,/+11 OR (19) In case of the simplified fuzzy-logic evaluation of the
[ ...
] THEN 4 [qk] + symptoms the output q of a rule is only determined by the
fuzzy-logic connections of the inputs G,, i.e. the conclusion is
In the classical fault-tree analysis the symptoms and events are identical with the fuzzy logic operation on the condition.
considered as binary variables with &=I for happened and 5,=0 Therefore no specifications of fuzzy sets (membership
P.f 1)
functions) of the outputs (events, faults) are required.
b) Approximate reasoning with fuzzy logic
If the relation between one event E, and one symptom AS, is
considered as a cause-effect relationship, fuzzy subsets El and
3,are defined by stating membership functions pEl (e) and
.u f d. pcs,(s). The approximate reasoning then follows the
Increased
compositional rule of inference [14]. The whole procedure is
described in [15]. For universal sets pEk (e) = 1 the same
C)
results are obtained as for a)
0 A refinement of the diagnosis can be achieved by backward
E, chaining [ 161.
I
- A I I I +AI,

Fig3 Membership functions of symptoms Si = Asi 5. APPLICATIONS


represented as fuzzy sets The desribed methodology was applied to different technical
a) symptom increases, b) symptom decreases, c) symptom processes like d.c.motors, el. and pneumatic actuators, pumps,
increases or decreases with linguistic terms (labels) robots, machine tools, heat exchanger, vehicles, see [11,[21,[51,
161,[101,[161,[~71.
for not happened [131.
6. REFERENCES
4.3 Diagnostic reasoning [ 11 Isermann, R. (1993). Fault diagnosis of machines via parameter
Based on the available heuristic knowledge in form of heuristic estimation and knowledge processing. Automatica, Vol. 29, No. 4,
process models and weighting of effects different diagnostic pp. 815-835.
forward and backward reasoning strategies can be applied. [2] Isermann, R.; B. Freyermuth (1991). Process fault diagnosis
Finally the diagnostic goal is achieved by a fault decision based on process model knowledge. Trans. ASME (1991), Vol.113,
which specifies the type, size and location of the fault as well pp. 620-626, 627-633.
as its time of detection. A review of recent developments in [3] Willsky, A.S. (1976). A survey of design methods for failure
reasoning oriented approaches for diagnosis was given in [9]. detection systems. Automatica, Vol 12, pp. 601-611.
(41 Himmelblau, D.M. (1978). Fault detection and diagnosis in
By using the strategy of forward chaining a rule, the facts are
chemical and petrochemical processes. Amsterdam, Elsevier, pp.
matched with the premise and the conclusion is drawn based
343-393.
on the logical consequence (Modus ponens). Therefore with the
[5] Isermann R. (1984). Process fault detection based on modelling
symptoms s,=AS,/S,max as inputs the possible faults f,=AFJ/FJmax and estimation methods-a survey. Automatica, V01.20, No.4, pp.
are determined using the heuristic causalities. 387-404.
In general the symptoms have to be considered as uncertain [6] Isermann, R. (1994). Integration of fault detection and
facts. Therefore a representation of all observed symptoms as diagnosis methods. IFAC Symposium. SAFEPROCESS, Helsinki.
membership functions p(s,) of fuzzy sets in the interval [0,1] is [7] Gertler, J. (1988). Survey of model-based failure detection and
feasible, especially in unified form as described in section 4.1. isolation in complex plants. IEEE Control Systems Magazine.
a) Symptom-fault processing by simplified fuzzy-logic [8] Frank, P.M. (1990). Fault diagnosis in dynamic systems using
operations analytical and knowledge-based redundancy. Automatica 26, pp.
By representing the symptoms as fuzzy sets with membership 459-474.
functions 0 I ~(5,) I 1 and the events (faults) with 0 I p(qJ [9] Isermann, R.; M. Ulieru (1993). Integrated fault detection and
< 1 the conditional part of the IF-THEN-rules can then be diagnosis. IEEWSMC Conference "Systems Engineering in the
determined by fuzzy-logic operations. For the determination of Service of Humans", Le Touquet, France.
the membership function p(q) of the output in dependence on [lo] HOfling, T. ; R. Isermann (1995). Parameter Estimation
the inputs ~(5,)different t-norms and t-conorms can be Triggered by Continous Time Parity Equations. ACC 1995.
applied. Relatively simple operations are obtained by max-min- [ 111 Frost, R.A. (1986). Introduction to knowledge base systems.
operation London, Collins.
[12] Torasso, P.; L. Console (1989). Diagnostic Problem Solving.
FuzzY-AND: P(+&[P(FJ ,.-,P(EJ (20)
North Oxford Academic, U.K.
[13] Lee, W.S.; D.L. Grosh; F.A. Tillmann; C.H. Lie (1985). Fault
Fuzzy-OR:. P(?)=-[P(t;l) 9.*.Y P(€JI (21) tree analysis, methods and applications - a review. IEEE
Another possibility is to use the prod-sum-operation
Transactions on Reliability, Vol. R-34, No. 3, pp. 194-202.
FuzzY-AND: ~ ( t l ) = ~ ( €~ i( )€ 2 ) ~ ( f , ) (22) [141 Zadeh, LOA,(19781, Fuzzy sets as a basis for a theory of
possiblity. Int. Journal of Fuzzy Sets and Systems. No.1, pp. 3-28.
[15] Ulieru, M. (1994). Diagnosis by Approximate Reasoning on
Dynamic Fuzzy Fault Trees. IEEE'94, Orlando, Florida.
In this case all values are represented in the result. Note also [16] Freyermuth, B. (1991). Knowledge based incipient fault
the similarity to equations of probability theory. diagnosis of industrial robots. IFAC-Symposium SAFEPROCESS,
The NOT-Operation is in both cases Baden-Baden. Proc.Pergamon Press, Oxford.

1609

You might also like