Download as pdf or txt
Download as pdf or txt
You are on page 1of 14

Control Eng. Practice, Vol. 5, No. 5, pp.

639-652, 1997
Copyright © 1997 Elsevier Science Ltd
Pergamon Printed in Great Britain. All fights reserved
0967-0661/97 $17.00 + 0.00
PII:S0967-0661(97)00046-4

SUPERVISION, FAULT-DETECTION AND FAULT-DIAGNOSIS


METHODS - AN INTRODUCTION
R. lsermann
Institute of Automatic Control, Laboratoryof Control Engineering and ProcessAutomation, Darmstadt Universityof
Technology, Landgraf-Georg-Str.4, D-64283 Darmstadt, Germany(isermann@irtl.rt.e-technik.th-darmstadt.de)

(Received December 1996; in final form March 1997)

Abstract: The operation of technical processes requires increasingly advanced supervision and
fault diagnosis to improve reliability, safety and economy. This paper gives an introduction to
the field of fault detection and diagnosis. It begins with a consideration of a knowledge-based
procedure that is based on analytical and heuristic information. Then different methods of fault
detection are considered, which extract features from measured signals and use process and
signal models. These methods are based on parameter estimation, state estimation and parity
equations. By comparison with the normal behaviour, analytic symptoms are generated. Human
operators are another source of information, and support the generation of heuristic symptoms.
For fault diagnosis, all symptoms have to be processed in order to determine possible faults.
This can be performed by classification methods or approximate reasoning, using probabilistic
or possibilistic (fuzzy) approaches based on if-then-rules.
Copyright © 1997 Elsevier Science Ltd

Keywords: Fault detection, supervision, monitoring, parameter estimation, state estimation,


parity equations, technical processes, actuators, sensors, fault diagnosis, diagnostic reasoning,
fuzzy reasoning.

1. INTRODUCTION The classical methods (a) and (b) are suitable for the
overall supervision of the processes. To set the toler-
Within the automatic control of technical systems, ances, compromises have to be made between the
supervisory functions serve to indicate undesired or detection size of abnormal deviations and unnecessary
unpermitted process states, and to take appropriate alarms because of normal fluctuations of the vari-
actions in order to maintain the operation and to ables. Most frequently, simple limit value checking is
avoid damage or accidents. The following functions applied, which works especially well if the process
can be distinguished: operates approximately in a steady state. However,
the situation becomes more involved if the process
(a) monitoring: measurable variables are operating point changes rapidly. In the case of closed
checked with regard to tolerances, and loops, changes in the process are covered by control
alarms are generated for the operator. actions and cannot be detected from the output sig-
(b) automatic protection: in the case of a dan- nals, as long as the manipulated process inputs remain
gerous process state, the monitoring function in the normal range. Therefore, feedback systems
automatically initiates an appropriate coun- hinder the early detection of process faults.
teraction.
(c) supervision with fault diagnosis: based on The big advantage of the classical limit-value-based
measured variables, features are calculated, supervision methods is their simplicity and reliability.
symptoms are generated via change detec- However, they are only able to react after a relatively
tion, a fault diagnosis is performed and large change of a feature, i.e. after either a large sud-
decisions for counteractions are made. den fault or a long-lasting gradually increasing fault.

639
640 R. lsermann

In addition, an in-depth fault diagnosis is usually not are described in more detail in the subsequent papers.
possible.

Therefore (c) advanced methods of supervision and 2. FAULT DETECTION AND FAULT DIAGNOSIS
fault diagnosis are needed,which satisfy the following
requirements: Fig. 2 shows an overall scheme of knowledge-based
fault detection and diagnosis. The main tasks can be
(i) Early detection of small faults with abrupt or subdivided into fault detection by analytic and heuris-
incipient time behaviour. tic symptom generation, and fault diagnosis.
(ii) Diagnosis of faults in the actuator, process
components or sensors. 2.1 Analytic symptom generation
(iii) Detection of faults in closed loops.
(iv) Supervision of processes in transient states. The analytical knowledge about the process is used to
produce quantifiable, analytical information. To do
The goal for the early detection and diagnosis is to this, data processing based on measured process vari-
have enough time for counteractions such as other ables has to be performed, to generate first the char-
operations, reconfiguration, maintenance or repair. acteristic values by
The earlier detection can be achieved by gathering
more information, especially by using the relationship limit value checking of direct, measurable
between the measurable quantities in the form of signals. The characteristic values are the
mathematical models. For fault diagnosis, the know- exceeded signal tolerances.
ledge of cause-effect relations has to be used. A gen- signal analysis of directly measurable signals
eral scheme for all the supervisory functions and the by the use of signal models like correlation
resulting actions is given in Fig. 1. functions, frequency spectra, autoregressive
moving average (ARMA) or the characteris-
This paper introduces some basic problems and tic values (e.g., variances, amplitudes,
methods in supervision, fault detection and fault diag- frequencies or model parameters).
nosis. In Section 2 the main tasks of fault detection process analysis by using mathematical pro-
and fault diagnosis are considered. Then an overview cess models together with parameter estima-
of model-based fault-detection methods is given in tion, state estimation and parity equation
Section 3. This is followed by an introduction to fault methods. The characteristic values are para-
diagnosis procedures, Section 4. The various methods meters, state variables or residuals.

Hazard- Faults Symptoms Features i


classes \ ~ Sup.er-
I VISIOn
- ~ - - D ~ i s i o ~ - - ~ e F a ~ - ~ i - o ~ diaFgnaUlotsis £~t~gen p~eeg~assling~I with fault
I~diagnosis
I Super-
, visory
I
Protection level

Stop ~ Signal
A
iF___qop tion[ I Alarm ~ e v a l u a t i o n ~
Moni-
toring
~ ..... 1. . . .

II "10perati°n
/ .~ F Faults Measure-
Me;
merits
Control
+
Process
II J Re,con- ~ , I W r--~--Ll ~ ~C- ] r[I/,~
Levels
, I figurati°nl
--" q : ~ C 1=~ P ~-"~
J

. . . . . I
\Piooos
Ii - Con+trol

"" Actions

Fig. 1 General scheme of different supervision methods


Supervision, Fault-Detection and Fault-Diagnosis Methods 641

In some cases, special features can then be extracted type, size and location of the fault as well as its time
from these characteristic values, e,g. physically of detection based on the observed analytical and heu-
defined process coefficients, or special filtered or ristic symptoms. If no further knowledge of fault-
transformed residuals. These features are then com- symptom causalities is available, classification
pared with the normal features of the non-faulty pro- methods can be applied which allow a mapping of
cess. For this, methods of change detection and clas- symptom vectors into fault vectors. To this end,
sification are applied. The resulting changes (discrep- methods like statistical and geometrical classification
ancies) in the described directly measured signals, or neural nets and fuzzy clustering can be used. If,
signal models or process models are considered as however, a-priori knowledge of fault-symptom causa-
analytic symptoms. lities is available, e.g. in the form of causal networks,
diagnostic reasoning strategies can be applied. For-
2.2 Heuristic symptom generation ward and backward chaining, with Boolean algebra
for binary facts and with approximate reasoning for
In addition to the symptom generation using probabilistic or possibilistic facts, are examples. Final-
quantifiable information, heuristic symptoms can be ly a fault decision indicates the type, size and location
produced by using qualitative information from of the most possible fault, as well as its time of de-
human operators. Through human observation and tection.
inspection, heuristic characteristic values in the form
of special noises, colours, smells, vibration, wear and The terminology used in this field is described in
tear, etc., are obtained. The process history, in the (Isermann and Ballr, 1997), based on definitions of
form of maintenance performed, repairs, former the IFAC Technical Committee SAFEPROCESS.
faults, lifetime and load measures, constitutes a fur-
ther source of heuristic information. Statistical data
(e.g. MTBF, fault probabilities) achieved from experi- 3. MODEL-BASED FAULT DETECTION
ence with the same or similar processes can be added. METHODS
In this way heuristic symptoms are generated, which
can be represented as linguistic variables (e.g., small, Different approaches for fault detection using math-
medium, large) or as vague numbers (e.g., around a ematical models have been developed in the last 20
certain value). years, see. e.g. (Willsky, 1976; Himmelblau, 1978;
Isermann, 1984, 1993, 1994a; Gertler, 1988; Frank,
2.3 Fault diagnosis 1990). In this section, basic methods are briefly
described.
The task of fault diagnosis consists of determining the

~ FAULTS
olmerved
HEURISTIC HEURISTIC
ANALYTICAL ANALY'IICALsyMPTOM ~ -m'~kJ
KNOWLEDOE

' tL
OBR~VATIOI~
+ MODm .

i FILTI~.ING, NORMAL
BBi~'VIINJR I
~TIMAII~q IIX'rlIACHI~I

ItlSIUItY
WlIII31t'I'II~
Vl Dm'Bml~ 8T/~.STICS
~V lamukl~

r
g
Clt[~,Jl,l~o I~CIIilON
FAULT-
DIAOIqC~II8
0---~

Fig. 2 Scheme of knowledge-based fault detection and diagnosis


642 R. Isermann

The task consists of the detection of faults in the pro- least one characteristic property of a variable from an
cesses, actuators and sensors by using the depen- acceptable behaviour. Therefore, the fault is a state
dencies between different measurable signals. These that may lead to a malfunction or failure of the sys-
dependencies are expressed by mathematical process tem.
models. Fig. 3 shows the basic structure of model-
based fault detection. Based on measured input sig- The time dependency of faults can be distinguished as
nals U and output signals Y__,the detection methods shown in Fig. 5:
generate residuals b parameter estimates 19 or state
estimates ~ which are called features. By comparison abrupt fault (stepwise)
with the normal features, changes of features are incipient fault (drift-like)
detected, leading to analytical symptoms s. intermittent fault.
FALIL'I~
With regard to the process models, the faults can be
further classified. According to Fig. 6 additive faults
influence a variable Y by an addition of the fault f,
and multiplicative faults by the product of another
variable U with f. Additive faults appear, e.g., as off-
sets of sensors, whereas multiplicative faults are para-
meter changes within a process.

t
.......
f
l¢0mhaAL

i ,amd~ S~pto~ a b

Fig. 3 General scheme of process-model-based fault


detection u

For the application of model-based fault detection


methods, the process configurations according to Fig.
4 have to be distinguished. With regard to the
inherent dependencies used for fault detection, and r
the possibilities for distinguishing between different
faults, the situation improves greatly from case A) to
B) or C), by the availability of some more measure- Fig. 5 Time-dependency of faults
ments. a) abrupt b) incipient c) intermittant

Now, lumped-parameter processes are considered,


,_. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . which operate in an open loop. The static behaviour
A B (steady states) can frequently be expressed by a non-
linear characteristic.

~ Iy2 Y=flo +ill U + . ~ 2 U2 +"" +J~qUq (1)

L. . . . . . . . . . . . . . . . . . . . . . ~ y =~ r ~ (2)

~. . . . . . . . . . . . . . . . . . . J oT= LS~...~q] (3)


$
C D
Fig. 4 Process configurations for model-based fault 3gr= [ 1 UU:... U q] (4)
detection
A) SISO (single-input single-output) Changes of the parameters Bi are multiplicative faults.
B) SISO with intermediate measurement Input signal faults fu and output signal faults fv are
C) SIMO (single-input multi-output) additive faults, compare Fig. 7.
D) MIMO (multi-input multi-output)
By considering small signal deviations around an
3.1 Process models and fault modeling operating point (Y~x/UcM~)the input/output behaviour of
processes can frequently be described by ordinary
A fault is defined as an unpermitted deviation of at linear differential equations
Supervision, Fault-Detection and Fault-Diagnosis Methods 643

Y(0 +atY°)(0 +... +a.Y(n)(O (5) Y(O =crx(t). (12)


=bou(O + b~uO)(O +... + b.u{')(O As shown in Fig. 9, additive faults are usually
modelled as input or state faults ft or output faults fro.
y(O=Y(t)-yoo;u(t)=U(O-Uoo (6) Parameter faults Abj, Aai or Acj are multiplicative
faults.

where y(")(t)--d"y(t) /dt" are derivatives.

I f=-Aa
U . ~ - q y = (a + aa)O
Y. Y=Y.+ f "1__~ " = a l l + f U

a) b)
Fig. 6 Basic models of faults
a) additive faults b) multiplicative faults Fig. 9 State-space model for a SISO process and
faults

Similar representations hold for nonlinear differential


equations and state space models,and for multi-input
and multi-output processes, also in discrete time.
U Based on these process and fault models, some basic
model-based fault-detection methods can be briefly
described.

Fig. 7 Nonlinear static process model


3.2 Fault detection with parameter estimation
The process model can be written in vector form
In most practical cases the process parameters are not
Y(0 =~r(t) O (7) known at all, or are not known exactly enough. Then
they can be determined with parameter estimation
o r = [a 1...a n bo...b=] (8) methods by measuring input and output signals if the
basic structure of the model is known.
~Zr(t)=[_yO)(t)..._y(n)(t) u (O ... u fm)(O ] (9)
a) Equation error methods
The process model is written in the vector form, eq.
The corresponding transfer function becomes, through (7), with the parameter and continuous-time data
Laplace transformation, vectors, eqs (8), (9). For parameter estimation the
equation error e(t) is introduced, compare Fig. 10a):
_ y(s) _ n(s) _ bo+b:+...+b_____~m_ (10)
e(0 =y(0-~r(00 (13)
Gp(s) - ~ ) A(s) l +ats+...+an sn or

e(s) =~(s)u(s) -~i(s)y(s).


As shown in Fig. 8, the input signal faults f. and out-
put signal faults fy are additive faults, and parameter
faults Aa. Abi are multiplicative faults, After sampling with the discrete time k=t/T0=0,1,2 ....
(for parameter estimation only) with T O the (small)
sampling time minimization of the sum of least
squares,
N
u B(s) V = ~ e2(k) = ere (14)
Gp(s) -X (s) k=l

dr/dO = 0 (15)
Fig. 8 Linear input/output model and faults leads to the least-squares (LS) estimate

The process model in state-space form becomes ~(N) = ['t'rv]-l~l'rX (16)

~(t) =A_x(0 +bu(0 (11) or in recursive form (RLS)


644 R. Isermann

_~(k+1) =I~(k)+~L(k)[y(k+ 1) -gff(k + 1){~(k)]


:Z(k) =
1
_q1.r(k+1) P(k)_~(k+ 1) + 1
e(k)~(k + 1)
u
A
d B(s) ] . y.~
_P(k+l) =[I-:z(k)_,r(k+ 1)]/~(k) (17)
"1 A(s) [ I
see e.g. (Isermann, 1992a). For the improvement of
the numerical properties, square-root filter algorithms
are recommended, and for the determination of the
signal derivatives in the data vector ~+(k) state vari-
able filters. An appropriate realization of the state a)
variable filter is a keypoint for obtaining correct para-
meter estimates (Isermann et al, 1992).

b) Output error methods


Instead of the equation error, eq. (13), the output
error
I
e/(t) = y(t) - yM(0, t) (18)

where
y~(O,s) = ~(s) u(s) (19)
A(s)
b)
is the model output, can also be used, see Fig. t0b).
Fig. 10 Model structures for parameter estimation
But then no direct calculation of the parameter esti-
a) equation error b) output error
mate O is possible, because e'(t) is nonlinear in the
parameters. Therefore the loss function as eq. (14) is
minimized by numerical optimization methods. The 3.3 Fault detection with state estimation and ob-
computational effort is then much larger, and on-line servers
real-time application is in general not possible. How-
ever, relatively precise parameter estimates may be The linear process under consideration is described in
obtained. If a fault within the process changes one or state-space form
several parameters by A0j, the output signal changes x_'(t) = A_x(t)+Bu(t) (23)
for small deviations according to
)~(t) = Cx_(t). (24)
Ay(0 =*r(OAO(t) + A~r(t) O + A gtr(t) A0(t), (20)
Here, p input signals u(t) and r output signals y(t) are
and the parameter estimator indicates a change A_0. assumed, as the methods described are especially
suitable for multivariable processes. Assuming that as
Generally, the process parameters .9_0depend on physi- well as the structure, all process parameters A, B, C
cal process coefficients P (like stiffness, damping are known (which is very restrictive) a state observer
factor, resistance) is used to reconstruct the unmeasurable state variables
0 = f(~) (21) based on the measured inputs and outputs

via nonlinear algebraic equations. If the inversion of g(0 = A_~(0 +~R(0 +H£(0 (25)
this relationship
¢(0 = 3(0 -C~(0. (26)
/2 = f-l(0) (22)
Compare Fig. 11, where e(t) is the output error. For
exists (Isermann, 1992b), changes Api of the process the state estimation error it follows from eq. (25) and
coefficients can be calculated. These changes in the (26), that
process coefficients are in many cases directly related
to faults. Therefore, the knowledge of Api facilitates (27)
fault diagnosis, but is not necessary for fault detection g(0 = ~(0 -g(0
only. Parameter estimation can also be applied for 4(0 = ~_-/-/~g(0.
nonlinear static process models (Isermann, 1993). The state error vanishes asymptotically

limx_-(t) = _0 (28)
Supervision, Fault-Detection and Fault-Diagnosis Methods 645

if the observer i s stable, which can be achieved by deviate from zero. x(t) as well as e(t) show a dynamic
proper design of the observer feedback H. The pro- behaviour which is different for f_L(t) and f.~(t). Both
cess is now influenced by disturbances and faults as x(t) or e(t) can be taken as residuals. In particular, the
follows, compare Fig. 12, residual e_(t) is the basis for different fault-detection
methods based on state estimation. For the generation
x_'(t) = A__~(t)+_Bu(t) +_Fv(t) +L_fL(t) (29) o f special properties, the design of the observer feed-
back H is of interest. Limiting conditions are the
stability and the sensitivity against output distur-
)~(t) = C~(t) +N_n(t) +M_M_fAf(t). (30) bances, n(t). If the signals are stochastic, Kalman-
Bucy filters are applied instead of observers.
Hereby are
v(t) non-measurable disturbances at the input If faults appear as changes AA~ AB or AC of the para-
n(t) non-measurable disturbances at the output meters, the process behavior becomes
f_L(t) fault signals at the input, acting through L on
x(t) (e.g., additive actuator or process faults) ~(t) = [A_+AA_].X(t)+[B_+A__Blu(t) (33)
f_.~(t) fault signals at the output, acting through M as
output changes Ay(t) (e.g. additive sensor )~(t) = [_C+ A ~Z]x(t) (34)
faults)
and the state estimation error
(35)
j--[,~ - ~ Z ( t ) + [aA_ - L/A £1~(t) + a ~ u ( t )
u Y

e(t) = Cg(t) + A Cx(0. (36)

The faults AA, AB and AC are then multiplicative


faults. In this case the changes in the residuals depend
on the parameter changes, and input and state variable
Y_M changes. Hence, the influence of parameter changes
on the residuals is not as straightforward as in the
case of the additive faults f_L(t) and f_.~(t).

The following fault-detection methods are known:


Fig. 11 Process and state observer

a) Dedicated observers for multi-output processes


Observer~ excited by one output: One observer
is driven by one sensor output. The other out-
puts 2£ are reconstructed and compared with
measured outputs y. This allows the detection
of single sensor faults (Clark, 1978a).
Kalman filter~ excited by all outputs: The
residuum (innovation) e(t) changes the charac-
ter of zero mean white noise with known
Fig. 12 Multivariable process with disturbances v, n
covariance if a fault appears. This is detected
and fault signals f_L,
by a hypothesis test (Mehra and Peschon,
1971; Willsky, 1976).
For the state estimation error the following equations Bank of observers~ excited by all outputs: Sev-
hold if the disturbances v(t)=0 and n_(t)--0 eral state observers are designed for a definite
fault signal and detected by a hypothesis test
(Willsky, 1976).
~(0 = t a - ~ g ( 0 +MLC0-~Mf~0 (31) - Bank of observers~ excited by single outputs:
Several observers for single sensor outputs are
and the oiatput error becomes used. The estimated outputs y are compared
with the measured outputs 2£- This allows the
e(t) = (7~(t) + Mfu(t). (32) detection of multiple sensor faults (Clark,
1978b) (dedicated observer scheme).
f_L(t) and f_u(t) are additive faults, because they influ- Bank of observers~ excited by all outputs
ence x(t) and e(t) by a summation. except one: As before, but each observer is
excited by all outputs except one sensor output
In the case of suddenly appearing and permanent fault which is supervised (Frank, 1987).
signals f_L(t) and f_.M(t)the state estimation errors will
646 R. Isermann

b) Fault-detection filters (fault-sensitive filters) for are independent of the process states _~(t), the known
multi-output processes input u and unknown inputs v (Fig. 12). In this way
- The feedback H of the state observer is chosen the residuals are dependent only on additive input
so that particular fault signals f_L(t)change in faults f~ and output faults f_u, see Fig. 12 and
a definite direction, and fault signals ~(t) in a (Hfifiing, 1996).
definite plane (Beard, 1971; Jones, 1973).

These fault-detection methods mostly require several


measurable output signals and make use of the inter-
nal analytical redundancy of multivariable systems. In
recent years it was proposed to improve their robust-
ness with regard to process parameter changes and a) b)
unknown input signals v(t), see e.g. (Frank, 1990; Fig. 14 Parity equation methods
Patton and Chen, 1991). This can be reached, for a) output error b) equation error
example, through filtering the output error of the
observer by
L'(t) = .~e(0 (37) 3.4 Fault detection with parity equations

together with a special design of the observer feed- A straightforward model-based method, of fault detec-
back H. tion is to take a fixed model G Mand to run it parallel
to the process Gp, thereby forming an output error
c) Output observers
Another possibility is the use of output observers (or r ( s ) = [Ge(s ) -Gu(s)]u(s ) (42)
unknown input observers) if the reconstruction of the
state variable ~(t) is not of primary interest (Frank, as shown in Fig. 14a). However, as for observers, the
W/innenberg, 1989; Tsui, 1993). Through a linear process model parameters have to be known a priori.
transformation If Gp(s)--GM(s), the output error then becomes for
additive input and output faults, Fig. 8,
z(t) =/?~(t) (38)
r' (s) =Gt,(s)fu(s ) +fy(s). (43)
the state representation of the observer becomes (Fig.
13) Another possibility is to generate a polynomial error,
Fig. 14b),
~(0 =Eg(0 +du(O+Gx(O (39) r(s) =au(s)y(s) -B~(s)u(s) (44)
=Bp(s)fu (s) +Ap(s)fy(s).
and the residual is determined by
In both cases different time responses are obtained for
(40) an additive input or output fault, r'(s) corresponds to
the output error of parameter estimation, eq. (19),and
r(s) to the equation error, eq. (13). Eqs (43) and (44)
u generate residuals and are called parity equations
(Gertler, 1991). To generate specific properties the
residuals can be filtered

r:(s) = G r (s) r (s) (45)

(Gertler, 1991; Patton and Chen, 1991).

However, for SISO processes only one residual can


be generated, and it is therefore not easy to distin-
guish between different faults.

More freedom in the design of parity equations can


Fig. 13 Process and output observer be obtained if for SISO processes intermediate signals
can be measured (Fig. 4), or for MIMO processes.
The state estimation error Then a state-space model is appropriate, as shown by
(Gcrtler, 1995) for discrete-time models. For a con-
e(t) --g(t~ -z(t) =g(t) - Z'x(t) (41) tinuous-time model
~(t) =Ax(t) +__Bu(t) (46)
and the residuals r_(t) are then designed, such that they
Supervision, Fault-Detection and Fault-Diagnosis Methods 647

Y(0 =Cx(t). (47) faults (turbo machines), knocking (Diesel engines) or


chattering (metal-grinding machines). But also signals
Eq. (47) is inserted in eq. (46) and differentiated p from many other sensors, like electrical current, posi-
times, resulting in the equation system tion, speed, force, flow and pressure, may show oscil-
lations with a variety of higher frequencies than the
usual process dynamic responses. The extraction of
fault-relevant signal characteristics can in many cases
be restricted to the amplitudes y0(to) or amplitude den-

I i11 ° °°1
Yv(0 =~x(0 +t2 Uv(0.
(48) sities [y(ito) [ within a certain bandwidth ~t)min~f.l)~O.)ma x
of the signal by using of bandpass filters (Stearns,
1975).

Other possibilities are the determination of


autocorrelation functions or the Fourier Transform
In order to remove the nonmeasurable states x(t), eq. (especially in the form of the Fast Fourier Transform,
(48) is multiplied by W_~ such that FFr) or the spectral density (Barschdorff, 1987). Cor-
relation functions and spectral densities are especially
~_T=Q. (49) suited to separating stochastic and periodic signal
components. If the frequencies are known, cross-
This leads to residuals correlation with e.g. the known basic and higher-order
harmonics can be performed.
t(O = VCZv(O - ~ O Ur(O (50)
However, a correlation function or a frequency spec-
compare Fig. 15. trum forms a nonparametric representation of the
signal and is (both for harmonic and stochastic sig-
nals) not well suited to direct automatic fault detec-
tion if the frequencies are unknown. Therefore, para-
u I metric signal models should be preferred, which allow
the main frequencies and their amplitudes to be di-
rectly estimated, and which are especially sensitive to
D" D I small frequency changes. This is possible by
modeling the signals as a superposition of damped si-
nusoids in the form of discrete-time ARMA (au-
toregressive moving average) models (Burg, 1968).
By means of a two-step maximum-entropy estimator
via correlation functions, a selected number of dis-
Fig. 15 Parity equation methods for a MIMO tinct, unknown frequencies taj can be estimated, as
state-space model; D ' : differentiation well as their amplitudes Yoj (Neumann, 1991 and
filter Janik, 1993). Symptoms are then generated by
changes of estimated frequencies and amplitudes.

The derivatives in Y.v(t) and U__v(t)can be obtained by


digital state variable filters for orders n < 3. The 3.6 Change detection and symptom generation
design of the matrix W gives some freedom to gener-
ate a structured set of residuals. One possibility is to The measured or estimated quantities, such as signals,
select the elements of W such that one measured parameters, state variables or residuals are usually
variable has no impact on a specific residual. Then stochastic variables Si(t), with mean value and vari-
this residual remains small in the case of an additive ance
fault on this variable, and the other residuals increase.
s~ -- gks,(0} ; 5~ = g/[s~(0 _~]2} (51)
3.5 Fault detection with signal models
as normal values for the non-faulty process. Analytic
Many measured signals y(t) show oscillations that are symptoms are then obtained as changes
of either harmonic or stochastic nature, or both. If
changes in these signals are related to faults in the aSi=EtSi(0-S.); a oi=E{oi(0-~) t ~ Tr(52)
process, actuator or sensor, a signal analysis is a fur-
ther source of information. Especially for machine with reference to the normal values. Usually, the time
vibration, sensors for position, speed or acceleration instants T F of fault occurrences are unknown.
are used to detect, for example, imbalance and beating
648 R. Isermann

To separate normal from faulty behaviour usually a 4.1 Symptom representation


fixed threshold
The inputs to the knowledge-based fault-inferencing
A S tot =C-Os c ~2
mechanism are all available symptoms as facts, and
has to be selected. By this means, a compromise has the fault-relevant knowledge about the process, most-
to be made between the detection of small faults, and ly in heuristic form.
false alarms because of short time exceeding (Fig.
16a). Methods of change detection, e.g., a likelihood- a) Analytical symptoms
ratio test or Bayes decision, a run-sum or a two-probe The analytical Symptoms Sa~ are the results of the
t-test (Basseville and Benveniste, 1985) may improve limit value checking of measurable signals, signal or
the (binary) decision. process-model fault-detection methods and change
detection, as described in Section 3.
Another possibility is to represent the feature as fuzzy
set Its(S) (e.g. a triangle, with the center as the mean b) Heuristic symptoms
and lower width A=~o s, ~:=2,3..... Fig. 16b). By Heuristic symptoms Sh~ are the observations of the
matching the current value las.(S) with the feature's operating personnel in the form of acoustic noise,
membership function, the "increased" Its÷(S), oscillations or optical impressions like colours or
smoke, obtained by inspection. These empirical facts
~s=maxs[min(~ts,(S),las+(S))], (53) can usually only be represented in form of qualitative
measures, e.g. as linguistic expressions like "little",
one obtains a gradual measure for exceeding a thresh- "medium" or "much".
old, a fuzzy threshold (Isermann, 1994b).
c) Process history and fault statistics
The generation of analytic symptoms concludes the A third category of facts depends on the general
task of model-based fault detection, Fig. 2. status, based on the history (past life) of the process.
This process history includes the past information of
running time, load measures, last maintenance or
p(S)~ AS
repair. If fault statistics exist (e.g., from "statistical
I.tb I : .tq
process control") they describe the frequency of cer-
tain faults for the same or similar processes. Depend-
s
ing on the quality of these measures, they can be used
as analytical or heuristic symptoms. However, the
information on the process history in general is
1 vague, and their facts have to be taken as heuristic
0,6 ....
symptoms. The symptoms based on process history
0 will be called Spi.
S
a) AS~
b)
The knowledge about the symptoms can be represen-
Fig. 16 Symptom detection of simple features ted, e.g., in the form of data strings, and can include,
a) Fixed threshold b) Fuzzy threshold for example: number, name, numerical value, refer-
ence value, calculated confidence or membership
value, time of detection, explanatory text
(Freyermuth, 1991).
4. FAULT-DIAGNOSIS METHODS
d) Unified symptom representation
The task of fault diagnosis consists of the determina- For the processing of all symptoms in the inference
tion of the type of fault with as many details as poss- mechanism, it is advantageous to use a unified repre-
ible such as the fault size, location and time of detec- sentation. One possibility is to present the analytic
tion. The diagnostic procedure is based on the and heuristic symptoms with confidence numbers
observed analytical and heuristic symptoms and the 0<c(s~)<l and treatment in the sense of probabilistic
heuristic knowledge of the process, as shown in Fig. approaches known from reliability theory (Freyer-
1. muth, 1991). Another possibility is the representation
as membership functions 0<it(si)<l of fuzzy sets
In this section the heuristic part of the knowledge and (Isermann and Ulieru, 1993). Fig. 17 shows some
inference mechanisms for diagnosis are described, in examples for cases where the symptom ASi, eq. (52),
order to build up on-line expert systems for fault either increases or decreases. Fig. 17a),b) has the ad-
diagnosis. vantage that only one membership function has to be
processed, in contrast to Fig. 17 c) where 5 member-
ship functions for linguistically expressed changes
have to be processed.
Supervision, Fault-Detection and Fault-Diagnosis Methods 649

By these kinds of fuzzy sets and corresponding mem- If no further knowledge is available for the relations
bership functions, all the analytic and heuristic symp- between features and faults, classification or pattern
toms can be represented in a unified way within the recognition methods can be used, Fig. 18. Here, refer-
range 0 < ~t(si) < 1. These integrated symptom repre- ence vectors S~ are determined for the normal behav-
sentations are then the inputs for the inference mech- iour. Then the corresponding input vectors S of the
anism, Fig. 1. features are determined experimentally for certain
faults Fj. The relationship between F u n d S is there-
fore learned (or trained) experimentally and stored,
forming an explicit knowledge base. By comparison
a) 0/ b) I of the observed S with the normal reference S_,, faults
' ~s, X ~ ''.as, F can be concluded.
~(S)I decreased increased
I I~FERENCE,-
PATTERN ]
I~, S, S
'r LASSIFICATIO ~
-AS +as
Fig. 17 Membership functions of symptoms Si=ASi
represented as fuzzy sets $1

e) Fault-symptom relationships
The propagation of faults to observable features or
symptoms in general follows physical cause-effect
relationships, where physical properties and variables Fig. 18 Fault diagnosis using classification methods
are connected to each other quantitatively and also as
functions of time. However, the underlying physical One distinguishes between statistical or geometrical
laws are frequently not known in analytical form, or classification methods, with or without certain prob-
are too complicated for calculations. If no information ability functions (Tou and Gonzalez, 1974). A further
is available on the fault-symptom causalities, experi- possibility is the use of neural networks because of
mentally trained classification methods can be applied their ability to approximate nonlinear relations and to
for fault diagnosis. This leads to an unstructured determine flexible decision regions for F in continu-
knowledge-base. In the case that the fault-symptom ous or discrete form (Barschdorff and Becker, 1990;
causalities can be expressed in the form of if-then Leonhardt, 1996). By fuzzy clustering the use of fuzzy
rules reasoning methods can be used. In the following separation areas is possible (Halgamuge, 1996). (If
subsections, both diagnostic procedures will be briefly features S are used for classification according to Fig.
described. As the timely behaviour (dynamics) 18 the generation of symptoms AS_ais not necessary
between causes and effects is not known in most because this is included in the classification pro-
cases, the relations are considered to be static. cedure, together with the reference S~)

4.2 Diagnosis using classification methods 4.3 Diagnosis using reasoning methods

In Section 3 it was shown how analytic and heuristic For some technical processes, the basic relationships
features S (or symptoms ASi) can be generated. These between faults and symptoms are at least partially
features are now represented in a feature vector known. Then this a-priori knowledge can be re-
presented in causal relations
$r=[$1,$2 .... ,S,]" (54)
fault --~ events ~ symptoms.
The corresponding faults are assumed to be known,
also in vector form Fig. 19a) shows a corresponding causal network, with
the nodes as states and edges as relations. The esta-
F T = IF1, F2 ..... Fm] (55) blishment of these causalitites follows the fault-tree
analysis (FTA), proceeding from faults through inter-
mediate events to symptoms (the physical causalitites)
The elements of F may be binary Fja [0,1] expressing or the event-tree analysis (ETA), proceeding from the
the faults as either "happened" or "not happened". symptoms to the faults (the diagnostic forward-chain-
They may also represent gradual measures for the ing causalities), see e.g. (Lee et al, 1985). To perform
size of the faults Fie [0...1]. a diagnosis, this qualitative knowledge can now be
expressed in form of rules
650 R. Isermann

IF<condition>THEN<conclusion>. c(Si), membership functions of fuzzy sets ~l(Si) or


probability density functions p(Si).
The condition part (premise) contains facts in the
form of symptoms Si as inputs, and the conclusion The probabilistic reasoning is based on Bayesian
part includes events F_~and faults Fj as a logical cause networks with conditional probabilities for the causal-
of the facts. If several symptoms indicate an event or ities (Pearl, 1988). However, in order to reduce the
fault, the facts are associated by AND and OR con- computational effort, statistically independent symp-
nectives, leading to rules in the form toms have to be assumed. Then the formulae are very
similar to the prod-sum operation in fuzzy logic
IF<SjANDS 2>THEN<E~> (Ulieru, 1996).
IF < E 1 OR E 2 > THEN < FI >
Possibilistic reasoning with fuzzy logic is more
compare Fig. 19b). straightforward. The features S~ or symptoms ASi are
represented as fuzzy sets with linguistic meanings as
"small", "medium", "large"; see Fig. 17. The fuzzy
logic reasoning then follows the principles of fuzzy-
IF-THEN-rule systems with the steps: fuzzification,
inference, accumulation and defuzzification. By
chaining these rules, several levels of rules can be
considered. The result is then the possibility of
a) measuring a fault and its size. For more details see
(Isermann, 1994b; Ulieru, 1996).

Because of the tedious acquisition of the rule-base


further developments consider adaptive fuzzy systems
for approximate reasoning. By using similarities
between fuzzy systems and neural networks,
especially with radial basis functions, Ayoubi (1996)
b) has proposed an adaptive neuro-fuzzy system, where
the membership functions of the symptoms and the
Fig. 19 Fault diagnosis using inference methods degree of AND OR connectives are tuned by experi-
a) causal networks ment.
b) fault symptom trees
The strategy of backward chaining assumes the con-
clusion as known, and searches for all relevant prem-
For the establishment of this heuristic knowledge ises (modus tollens). This is especially of interest if
several approaches exist, see (Frost, 1986; Torasso the symptoms are not complete. Therefore, the con-
and Console, 1989). cluded events and faults are displayed to the operator
after forward chaining, with all known symptoms. A
In the classical fault-tree analysis the symptoms and refinement of the diagnosis can then be achieved by
events are considered as binary variables, and the selecting the most plausible events and faults as hypo-
condition part of the rules can be calculated by theses, and applying backward chaining by asking for
Boolean equations for parallel-serial-connection, see missing symptoms. Then forward chaining is
e.g. (Barlow and Proschan, 1975; Freyermuth, 1993). restarted. This procedure is best implemented within
However, this procedure has not proved to be suc- an interactive dialogue, and repeated until terminated
cessful because of the continuous nature of faults and by the operator (Freyermuth, 1991, 1993). Other
symptoms. chaining strategies are Establish-Refine, Hypothesize
and Test, Depth and Width searching, etc, see
For the fault diagnosis of technical processes, methods (Harmon and King, 1986).
of approximate reasoning are more appropriate. A
review of the development in reasoning approaches
for diagnosis is given in (Isermann and Ulieru, 1993) REFERENCES
and (Ulieru, 1996).
Ayoubi, M. (1996). Fuzzy systems design based on
By using the strategy of forward chaining a rule, the hybrid neural structure and application to the
facts are matched with the premise and the conclusion fault diagnosis of technical processes. Control
is based on the logical consequence (modus ponens). Engineering Practice, Vol. 4, No. 1, pp. 35-
For approximate reasoning, the symptoms are con- 42.
sidered as uncertain facts. This can be performed by Barlow, R.E. and F. Proschan (1975). Statistical the-
representing the symptoms as confidence functions ory of reliability and life testing. Holt,
Supervision, Fault-Detectionand Fault-DiagnosisMethods 651

Rinehart & Winston, Inc. of Fuzzy System and Neural Networks in Intel-
Barschdorff, D. (1987). Gearbox failure diagnostics. ligent Data Processing. Fortschr. Ber. VDI
VDI-Berichte Nr. 644, 241-248. VDI-Verlag: Reihe 10 Nr. 401. VDI-Verlag: Diisseldorf.
Dtissddorf. Harmon, P. and D. King (1986). Expertensysteme in
Barschdofff, D. and D. Becker (1990). Neuronale der Praxis - Perspektiven, Werkzeuge, Erfahr-
Netze als Signal- und Musterklassifikatioren. ungen. Oldenbourg-Verlag: Mtinchen.
Technisches Messen, Vol. 57, pp. 437-444. Himmelblau, D.M. (1978). Fault detection and diag-
Basseville, M. and A. Benveniste, Eds. (1985). De- nosis in chemical and petrochemical processes,
tection of abrupt changes in signals and pp. 343-393. Elsevier: Amsterdam.
dynamical systems. Lecture notes in Control H/Sfling, T. (1996). Methoden zur Fehlererkennung
and Information Sciences. Vol. 77. Springer- mit Parameterschi~tzung und Parittitsglei-
Verlag: Berlin. chungen. Fortschritt-Ber. VDI Reihe 8 Nr.
Beard, R.V. (1971). Failure accommodation in linear 546, VDI-Verlag: Diisseldorf.
systems through self-reorganization. Rept. Isermann, R. (1984). Process fault detection based on
MVT-71-1. Man Vehicle Laboratory, Cam- modelling and estimation methods - a survey.
bridge, Mass. Automatica, Vol. 20, No. 4, pp. 387-404.
Burg, J.P. (1968). A new analysis technique for time Isermann, R. (1992a). Identifikation dynamischer
series data. NATO Advanced Study Institute on Prozesse. Vol. 1 and 2. 2nd edition. Springer-
Signal Processing with Emphasis on Under- Verlag: Berlin.
water Acoustics, Aug. 12-23. Isermann, R. (1992b). Estimation of physical parame-
Clark, R.N. (1978a). A simplified instrument detec- ters for dynamic processes with application to
tion scheme. IEEE Trans. Aerospace Electron. an industrial robot. Int. J. Control, Vol. 55, pp.
@st., Vol. 14, pp. 558-563. 1287-1298.
Clark, R.N. (1978b). Instrument fault detection. IEEE Isermann, R.; K.-H. Lachmann and D. Matko (1992).
Trans. Aerospace Electron. Syst., Vol. 14, No. Adaptive Control Systems. Prentice Hall: New
3, pp. 456-465. York.
Frank, P.M. (1987). Advanced fault detection and Isermann, R. (1993). Fault diagnosis of machines via
isolation schemes using nonlinear and robust parameter estimation and knowledge process-
observers, lOth IFAC Congress, Miinchen, ing. Automatica, Vol. 29, No. 4, pp. 815-835.
Vol. 3, pp. 63-68. Isermann, R. and M. Ulieru (1993). Integrated fault
Frank, P.M. (1990). Fault diagnosis in dynamic sys- detection and diagnosis. IEEE/SMC Confer-
tems using analytical and knowledge-based ence "Systems Engineering in the Service of
redundancy. Automatica, Vol. 26, pp. 459-474. Humans", Le Touquet, France. Vol. 1, pp.
Frank, P.M. and J. Wtinnenberg (1989). Robust fault 743-748.
diagnosis using unknown input observer Isermann, R., Ed. (1994a). (lberwachung und Fehler-
schemes, Chapter 3 in Patton, R.J.; P.M. Frank diagnose - Moderne Methoden und ihre
and R.N. Clark; Eds. (1989). Fault diagnosis Anwendungen bei technischen Systemen. VDI-
in dynamic systems - theory and applications. Verlag, Dtisseldorf.
Prentice Hall: London. Isermann, R. (1994b). On Fuzzy Logic Application
Freyermuth, B. (1991). Knowledge based incipient for Automatic Control, Supervision and Fault
fault diagnosis of industrial robots. IFAC Diagnosis. Third European Congress on Fuzzy
SAFEPROCESS Symposium, Baden-Baden. and Intelligent Technologies EUFIT'95,
Vol. 2, pp. 31-37. Pergamon Press: Oxford. Aachen, Vol. 2, pp. 738-753.
Freyermuth, B. (1993). Wissensbasierte Fehlerdia- Isermann, R. and P. Ball6 (1997). Trends in the
gnose am Beispiel eines Industrieroboters. Application of Model Based Fault Detection
Dissertation TH Darmstadt. Fortschr. Ber. VDI and Diagnosis of Technical Processes. Control
Reihe 8 Nr. 315. VDI-Verlag: Dtisseldorf. Engineering Practice, Vol. 5, No. 5.
Frost, R.A. (1986). Introduction to knowledge base Janik, W. (1993). Fehlerdiagnose des Aussenrund-
systems. Collins: London. Einstechschleifens mit Prozess- und Signal-
Gertler, J. (1988). Survey of model:based failure modellen. Dissertation TH Darmstadt. Fortschr.
detection and isolation in complex plants. Ber. VDI Reihe 2 Nr. 288. VDI-Verlag:
IEEE Control Systems Magazine. Dtisseldorf.
Gertler, J. (1991). Analytical redimdancy methods in Jones, H.L. (1973). Failure detection in linear sys-
fault detection and isolation. IFAC SAFEPRO- tems. Ph.D. Thesis, Dept. of Aeronautics,
CESS Symposium, Baden-Baden. Vol. 1, pp. 9- M.I.T., Cambridge, Mass.
21. Pergamon Press: Oxford. Lee, W.S., D.L. Grosh, F.A. Tillmann and C.H. Lie
Gertler, J. (1995). Generating Directional Residuals (1985). Fault tree analysis, methods and appli-
with Dynamic Parity Equations. Automatica, cations - a review. IEEE Transactions on
Vol. 31, No. 4, PI~. 627-635. Reliability, Voi. R-34, No. 3, pp. 194-202.
Halgamuge, S. (1996). Advanced Methods for Fusion
652 R. Isermann

Leonhardt, S. (1996). Modellgestiitzte Fehlererken- Morgan Kaufmann: San Mateo, Calif.


nung mit neuronalen Netzen - Oberwachung Steams, S.D. (1975). Digital Signal Analysis. New
von Radaufhgingungen und Diesel-Einspritzan- Jersey: Hayden Book Company.
lagen. Dissertation TH Darmstadt. Fortschr. Torasso, P. and L. Console (1989). Diagnostic Prob-
Ber. VDI Reihe 12. VDI-Verlag: DUsseldorf. lem Solving. North Oxford Academic, U.K.
Mehra, R.K. and J. Peschon (1971). An innovations Tou, J.T. and R.C. Gonzalez (1974). Pattern recogni-
approach to fault detection and diagnosis in tion principles. Addison-Wesley Publishing:
dynamic systems. Automatica, Vol. 7, pp. 637- Reading, Mass.
643. Tsui, C.-C. (1993). A general failure detection, isola-
Neumann, D. (1991). Fault diagnosis of machine- tion and accomodation system with model
tools by estimation of signal spectra. IFAC uncertainty and measurement noise. 12th IFAC
SAFEPROCESS Symposium , Baden-Baden. Worm Congress, Sydney, Australia. Vol. 6, pp.
Vol. 1, pp. 73-78. Oxford: Pergamon Press. 231-238
Patton, R.J. and J. Chen (1991). A review of parity Ulieru, M. (1996). Fuzzy logic in diagnostic decision:
space approaches to fault diagnosis. IFAC Possibilistic Networks. Dissertation, Technical
SAFEPROCESS Symposium, Baden-Baden. University Darmstadt.
Vol. 1, pp. 239-255. Oxford: Pergamon Press. Willsky, A.S. (1976). A survey of design methods for
Pearl, J. (1988). Probabitistic Reasoning in Intelligent failure detection systems. Automatica, Voi. 12,
Systems: Networks of Plausible Inference. pp. 601-611.

You might also like