Aut~matica. Vol. 29. No. 4. pp. 815-835. 1993 0005-1098/93 $6.00+0.

00
Printed in Great Britain. (~) 1993 Pergamon Pre-~ lad

Tutorial Paper

Fault Diagnosis of Machines via Parameter

Estimation and Knowledge Processing
Tutorial Paper*
ROLF ISERMANNt

A fault diagnosis methodology is described which also processes analytical

symptoms obtained by parameter estimation as heuristic symptoms via
causal fault-symptom trees and confidence measures with applications to a
machine tool feed drive and other machines.
Key Words--Fault detection; supervision; monitoring; safety; parameter estimation; knowledge base;
diagnostics; fault trees; inference mechanism, forward and backward chaining, feed drive; machines.

A l ~ t r l e t - - F o r further improvement of the reliability and 1. INTRODUCTION

safety of machines the automatic early detection and
localization of faults is of high interest. The conventional AN ON-GOING TASK in e n g i n e e r i n g is t o i n c r e a s e
approach is to monitor some important variables like the reliability, availability and safety of technical
temperature, pressure, vibration and to generate alarms if processes. Therefore much effort goes into the
certain limits are exceeded. However, developing internal
faults are in this way only detected at a rather late stage. design and the manufacturing of products
By applying static and dynamic process models and common including the further development of materials,
process input and output measurements the inherent the testing of all elements, quality control, etc. If
relationships and redundancies can be used to detect faults
earlier and to localize them better. the process is then in operation the range of
Changes in process and signal parameters are very well process variables, and the health or condition of
suited for fault detection. The paper describes a general the process elements has to be observed, if
methodology for machines and other processes by using few
measurements, dynamic process and signal models and possible by automatic means. This holds
parameter estimation to generate analytical symptoms. especially for automatically controlled processes
Heuristic symptoms observed by the operator or based on, because the direct contact with the human
e.g. the process history are another source for fault
diagnosis. Both kinds of symptoms are weighted by operator is then reduced, or even missing, and
confidence measures and are fed into a knowledge based automatic systems are usually built to suppress
fault diagnosis procedure. The diagnosis is performed by disturbances and some faults to a certain degree.
forward and backward chaining according to causal
fault-symptom trees. If the basic tasks of process automation, the
The considered machines consist of a motor, a drive chain feedforward and feedback control, are dedicated
and a working process or load. They may be electrical motor to a first automation level, the various tasks with
or combustion engine driven pumps, fans or machine tools
with gear or belt drive chains. As one example, experimental supervisory functions can be considered as
results are shown for a DC motor powered feed drive of a forming a second level. These tasks are
machine tool. A summary of practical results with other described in a coherent way.
machines is also given.
Supervisory functions serve to indicate unde-
sired or unpermitted process states and to take
*Received 16 July 1991; revised 28 September 1992; appropriate actions in order to avoid a damage
received in final form February 1993. The original version of
this paper was presented at the IFAC Symposium on Fault of the process or an accident with human beings.
Detection, Supervision, and Safety for Technical Processes Figure 1 shows a scheme of these functions in
which was held in Baden-Baden, Germany during September more detail.
1991. The Published Proceedings of this IFAC Meeting may
be ordered from: Pergamon Press Ltd, Headington Hill Hail, It is assumed that faults affect the technical
Oxford OX3 0BW, U.K. This paper was recommended for process and its control. Generally a fault is to be
publication in revised form by Associate Editor C. C. Hang understood as a nonpermitted deviation of a
under the direction of Editor P. C. Parks. Corresponding
author Isermann's Telex 419 579; Telephone (6151) 16-21 14; characteristic property of the process itself, the
Fax (6151) 29 34 45. actuators, the sensors and controllers. If these
t Technical University of Darmstadt, Institute of Automa- deviations influence the measurable variables
tic Control, Laboratory of Control Engineering and Process
Automation, Landgraf Georg Str. 4, D-6100 Darmstadt, U(t) and y(t) of the process shown in Fig. 1 they
Germany. may be detected by an appropriate signal
815
816 R. ISERMANN

HUIIAN HAZARD- FAULTS SYMPTOMS FEATURES

rsuP R- - - -
VISION I
WITH I

' PROTEC- SUPER-

VI SORY
F- . . . . . . l ..... LEVEL

I 'MONI -
TORINO

,U,
Id,
: : ~ p ~ R _l:]~----Jl

,U]I W
. . . . .

~
O F FAULTS

U.
. .

.
. .

~ I
.....

/ MEASURE- CONTROL
l' C~FIOUR-~ ~ b ~ I ~>MENTS PROCESS
IAT!ON I II r I 1[ " " --~W [U,Y.] LEVELS

II " I ~i CONTROL

L I I I
x(ACTIONS
FIG. 1. Scheme for supervisory functions and resulting actions.

evaluation. The corresponding functions are necessarily improved. Also many faults cannot
called monitoring, if the measurable variables be detected directly by available sensor
are checked with regard to a certain tolerance of technology.
the normal values (limit or trend checking) and In recent years research efforts have shown
alarms are triggered if the tolerances are that process changes due to faults can be
exceeded. Based on these alarms the operator detected in an early state by using process
takes appropriate actions. In cases where the models and common sensors. Then non-
limit value violation signifies a dangerous process measurable quantities like process state variables
state an appropriate action can be initiated and parameters can be estimated. With this
automatically. This is called automatic protec- improved knowledge a process supervision with
tion. Both supervisory functions may be applied fault diagnosis (also called condition monitoring)
directly to the measured signals or to the results becomes possible. Figure 1 shows the cor-
of a following signal analysis, as e.g. in the case responding steps (Isermann (1984)). Signal
of frequency spectra of vibrations for rotating processing now provides features like direct
machines (cf. Collacott (1977), Schneider- measurable quantities or nonmeasurable quan-
Fresenius (1985), Sturm and F6rster (1988)). tities in the form of state variables or parameter
These classical ways of limit value checking of estimates. By comparison with normal values,
some important measurable variables are ap- changes are detected resulting in symptoms. A
propriate for the overall supervision of the knowledge based fault diagnosis indicates its
processes. However, developing internal process cause and location, see also Himmelblau (1978)
faults are only detected at a rather late stage and and Pau (1981). The next step is a fault
the available information does not allow an evaluation, that means an assessment is made of
in-depth fault diagnosis. This is one of the how the fault will affect the process. The faults
reasons that process operators are still required are then divided into different hazard classes
for the supervison of important processes. These according to an incident/sequence or a fault tree
human operators use their own sensors, analysis. Then the following action can be
datarecords, reasoning and long term ex- decided. If the fault is evaluated to be tolerable,
perience to obtain the required information on the operation may continue and if it is
process changes and its diagnosis. conditionally tolerable, a change of operation, a
If the supervision is going to be improved and reconfiguration of process parts or just main-
automated, a natural first step consists of adding tenance has to be performed. However, if the
more sensors and a second step to transfer the faults are intolerable an immediate stop is
operators knowledge into computers. Here it is required and the fault must be eliminated, e.g.
usually desirable to add such sensors which by repair. Figure 1 shows that a looped signal
directly indicate faults. Because the number of flow exists from the measured signals through
sensors, transmitters and cables increases, the the different actions back to the process. It is
cost goes up and the overall reliability is not therefore possible to refer to supervisory loops.
 Machine fault diagnosis 817

With such advanced supervisory, functions it is with

possible to improve the further automation of y(t)= Y(t)- YM~ and u ( t ) = U(t) - Ucx~, (3)
technical processes with the following features:
• early detection and diagnosis of developing where Uc~, Yx~ are steady-state (or DC) values
faults (also called incipient fault diagnosis); and y<")(t)=d"y(t)/dt". If nonlinear dynamic
• early detection and diagnosis of sudden process models are taken into account, special
(abrupt) faults; classes have to be considered. With regard to
• prevention of further fault expansion; parameter estimation with direct methods of
• preventive maintenance; least squares the class of nonlinear differential
• maintenance on request; equations which is linear in the parameters is of
° telediagnosis by using modern communication significance. One example of this type is
nets (without local inspection). Y(t) + a I,, Y(t)(t) + a lt [ Y(I)(t)]2
In this paper methods for the model based fault
diagnosis via parameter estimation are briefly -t- a12 sign Y~°(t) + a13 Yt2)(t)
described. Theoretical modeling and parameter =b~U(t)+bo~U(t)Y(t) +... (4)
estimation is the background for an automatic
However, also many other classes of nonlinear
supervision using analytic and heuristic knowl-
models are appropriate, if the output error is
edge of the process. Section 2 summarizes the minimized, using iterative operating parameter
methods for parameter based fault detection. In
estimation methods.
Section 3 static and dynamic models for
The process models (equations (1), (2) and (4))
machines as dependent on different operation
can be written in a unified form
modes are given. Then the symptom generation
by parameter estimation is described for Y(t) = ~p~rOs, (5)
dynamic and static processes and for vibration y(t) = ~'dr[,i, (t)0d, (6)
signals in Section 4. The processing of symptoms
T
is based on heuristic knowledge in order to Y(t) = ~d[,o,,i, (t)_0d, (7)
perform a fault diagnosis. Section 5 describes the with the parameter vectors
symptom and knowledge representation and an
inference mechanism. One example is shown, 0S = [0,, #q], (8)
where various methods described in this paper _0rl,,. = [a, . . . a,, bo " " b , l , (9)
were applied to a machine tool feed drive. Finally
a summary of experimental results with other _0drl,o,,i, = [amo all a,2 a,3 b~, b 0 , " "], (10)
machines is given. Hence, with reference to Fig. and the data vectors
1 this paper concentrates within the supervisory
level on supervision with fault diagnosis. [u v 2... vq], (11)
~dr[,i. = [--y(I)(t) . . . . yt")(t),
u(t).', u<m)(t)], (12)
2. MODEL BASED FAULT D E T E C T I O N - -
ANALYTICAL R E D U N D A N C Y T
~ d [nonlin = [ - - y t l ) ( t ) _ [y(1)(/)]2 _ sign Y t ° ( t )
The task consists in the detection of faults in -Yt2)(t) U(t) U ( t ) Y ( t ) . . . ] . (13)
the process including actuators and sensors by
measuring the input and output variables U(t) In the following the unified dynamic model
and Y(t). T h e process may hereby operate in an y(t) = ~ r ( t ) O (14)
open or closed loop. Using mathematical process is used for convenience instead of the two types
models a distinction can be made between static in equations (6) and (7).
and dynamic process models. Static process The model parameters _0s and 0d are now
models can, e.g. be expressed as nonlinear more or less intricate relationships of several
polynomials physical process coefficients p
(1) 0 =/(e). (15)
The process coefficients are, e.g. stiffness,
For processes with lumped parameters which damping factors, friction factors, resistance,
can be linearized around one operating point capacitance, inductance and contain valuable
dynamical process models usually consist of information on process faults. They may be
ordinary linear differential equations calculated from the inverse relationship
y(t) + a l y ° ) ( t ) + " " + a.Y<")(t) e =y-'(_O), (16)
= b . u ( t ) + b,ut')(t) + ' " " + b,,ut")(t), (2) if 0 is known.
818 R. ISERMANN

Another description of linearized dynamic are therefore called multiplicative faults. Statc
processes is the state variable representation changes are modelled by
£_(z) = A_x_(t) + _bu(t), (17) _£(t) = _A[,s(t) + A4:_(t)] + b_u(t), (23)
y ( t ) = _cT~(t), (18) y(t) = _cr[_x(/) + As(t)], (24)
where _A, _b and _c contain the process model resulting in additions to the variables and are
parameters _0. This model is of interest, if the therefore called additive faults.
state variables ~(t) are used for fault detection, If the structure of the process model is known
e.g. by applying state observers or state by theoretical modeling, the mostly non-
estimators. measurable process parameters _0(t) and state
For nonlinear dynamic processes the state variables _x(t) can be calculated by applying
variable presentation becomes parameter estimation and state estimation
methods based on the measurable input and
_X(t) =f(_X(t), U(t), _0), (19)
output variables U(t) and Y(t). Through these
_Y(t) =g(_X(t), U(t), _0). (20) model orientated detection procedures several
process, actuator and sensor faults may be
Many special forms of this general description do detected using only a few (robust) sensors and
exist.
the physically based dynamic causalities between
The time dependent state variables _x(t) are the signals expressed by the process model. Thus
either well-defined physical variables like the these methods use the analytic redundancy
position x~(t), the speed x~:)(t) and the which is inherent in the measurements.
acceleration x~2)(t) of a mechanical element For the detection of changes, the estimated
described in the controllable canonical form, or quantities g(t) and _0(t) are compared with
more complicated functions of the input and corresponding quantities of the normal model
output signals as, e.g. in the observable (i.e. previous model) resulting in changes
canonical form.
If a fault now appears in the process this Ag_ (t) = g(t) - g0(t), (25)
causes a remaining change (offset) Aps(t ) in a A_0(t) = _0(t) - _0,,(/), (26)
process coefficient if it is directly affected by the
fault. Examples are an increase due to the or by observing error signals or residuals, e.g.
resistance due to plugging, or in the damping by e(t) = y(t) - ~or(t)O(t), (27)
friction, or in the sensor gain. This leads to
changes of model parameters A_0(t) and due to ~(t) = y ( t ) -- _C'g(t). (28)
the process dynamics more or less delayed to
These changes are then symptoms and appear
changes A y ( t ) ~ ~T(I)A_O(t) of the measurable
e.g. as
output signal y ( t ) (Fig. 2(a)). A fault results in
(i) changes (offsets) in definite directions;
remaining changes (offsets) Axi(t) of a state
(ii) changes due to certain patterns;
variable, if the fault directly affects the state
(iii) changes with certain time dependencies;
variable like a leakage of a mass flow or of an
(iv) changes of variances and covariances.
electrical current or a sensor offset. Then the
Fundamental methods of model based fault
measurable output signal changes by A y ( t ) =
c_r A s ( t ) (Fig. 2(b)). detection are described for the state estimation
In the state representation parameter changes approach in Wiilsky (1976), lsermann (1984),
A_A, A b and A_c are modelled by Viswanadham and Strichander (1987), Gertler
(1988), Frank (1990), and the parameter
~(t) = [_A + A_Al_x(t) + [_b + A_b]u(t), (21) estimation approach lsermann (1984, 1989a),
y ( t ) = [_c + A_c]Z~S(t), (22) Filbert and Metzger (1982) and Rault et al.
(1984). See also Tsefestas et al. (1986) and
leading to multiplications with the variables and Patton et al. (1989).

~ FAULTS ~7 FAULTS
dA~J.
u(l} ylt) =",~-'T(I)e ] v uitl t
iltl Axltl* bultl
;,t) : cTx(t) [ ytt)"~ y"~ml| t

a} ~, o.Ae b) ~ x_(t)oAx_(tl
FIG. 2. Scheme of a (linear) process influenced by faults: (a) faults effect parameters; (b) faults effect state
variables.
 Machine fault diagnosis 819

Another approach is the use of parity based fault diagnosis. The analytical problem
equations. With the assumption that the solution comprises data-processing, feature ex-
input-output model GM(s) of a process Gr(s) is traction and the detection of changes as
known, an output error symptoms, Isermann (1988, 1989a). The cor-
e(s) = y(s) - yM(S) = y(s) -- GM(S)U(S) (29) responding process knowledge base consists of
two parts:
can be calculated, or with GM(S)=BM(S)/ (a) analytical knowledge
AM(S) the polynomial equation error • analytical process models from theoretical
modeling (physical laws);
e(s) = AM(S)y(s) -- BM(S)U(S). (30)
• estimation methods for parameters and state
Equations (29) and (30) are called parity variables;
equations. The residuals e(s) can be filtered by • normal process behaviour;
r(s) = Gv(s)e(s), • process history and fault statistics, if
quantifiable.
to generate special properties (Gertler (1991); (b) heuristic knowledge
Patton and Chen (1991)). • fault trees (connection of symptoms and
Presently investigations are still going on to faults);
show the advantages and disadvantages of the • process history and fault statistics, if only
different model based fault detection methods. It qualitatively known.
seems that parameter estimation has advantages The heuristic knowledge is partially founded on
mainly for muitiplicative faults and state physical laws, and partially on not exactly
estimation, and parity equations mainly for quantifiable rules gained from experience.
additive faults. Hence, a proper combination of Within the inference mechanism first a fault
these methods will be especially attractive, detection is performed based on the observed
depending on the type of faults. symptoms and available fault trees. Resulting
The following fault diagnosis now requires a possible faults then undergo a fault decision
knowledge based treatment. The intention is to where process history and fault statistics are
transfer the existing knowledge of engineers, considered. Then suggestions for fault type, size
operators and maintenance staff in the supervi- and cause are given.
sion methodology and to develop on-line expert Thus this knowledge based fault diagnosis
systems for fault diagnosis. forms an on-line engineering expert system
Figure 3 shows a scheme for the knowledge (Nevins et al. (1987)), which is a combination of

~ FAULT
KNOWLEDGE KNOWLEDGE
BASE
ACQUISITIO N
:^N,,,,.~ KNO*LEO.:°~ i................. I;I".L"
t ANALYTIC
ENGINEER ANALYTI
PROCESS
MODEL
CAL
/ OAiA-
PROCESSING
P.o.,E.
SOLUTION
PARAM./
STATES
: { PARAMETER-/
STATE- FEATURE
PHYSICAL ESTIMATION EXTRACTION
LAWS
I PROCESS
NORMAL ~ COEFFIC,
PROCESS DETECTION
BE HAVIOUR OF CHANGES
OTHER ]
PROCESSES / :HEURIST. KNOWLEDGE: 8
i FAULT- "...................
- - ~ "!
TREES FAULT :INFERENCF
i DETECTION
i i " MECHANISM
HISTORY ] POSSIBLE
(FAULT-
FAULT8 ! DIAQNOSlSJ
[FAULT- FAULT I i
STATISTICS DECISION i
' l
l
;AuLi-SI
~T~;
J
ZEP
-LOCATION
FIG. 3. Scheme of the knowledge based fault diagnosis for technical processes (Isermann (1989a)).
820 R. ISERMANN

formal engineering analysis and common expert 3. STATIC AND DYNAMIC MODELS OF
systems. For the last part of the problem MACHINES FOR DIFFERENT OPERATION MODES
solution a certain amount of human expertise 3.1. Mathematical process models
expressed in rules and facts has to be used. This Many machines and also vehicles consist of a
can be formulated for example by different levels motor, a drive chain and a working process or
of diagnostic reasoning (Milne (1987, 1991)) or other load. The drive chain may have a belt
different kind of diagnostic models (Fink and drive, elastic clutch with damper, shaft or
Lusth (1987)). spindle and a gear. Then a scheme as in Fig.
From this it follows that the development and 4(a) results. The motor and the load are coupled
application of process fault diagnosis depends by spring-mass-damper systems. These ele-
greatly on process specific properties and ments can be represented as connected four
corresponding knowledge and that the further terminal networks (Fig. 4(b)).
research work requires the consideration of Tables 1-3 show mathematical models of
concrete case studies. This paper now sum- some motors, drive chain elements and working
marizes developments and experiences within processes, obtained by theoretical modeling. The
several research projects. Theoretical and shafts can be modeled as two mass rotatory
experimental results for machines and drives are elements and can be represented for small
presented by using parameter estimation meth- deviations qg(t) as linear four terminal networks
ods as a basis. according to Fig. 5. The friction of many

W
o}

co.m~nd l~d
signal signal

b)
control electric belt shaft gear staff/ cutting
i:
drive drive I spindle protest I
elastic mnchine I
clutch vehicle

W~j~ ~1 o~

control electric belt, shaft, cutting

drive gear, spindle process I
machine I
vehit le

FIG. 4. Schemes of machines with motor, drive chain and working process/load: (a) rotating
spring-mass-damper systems; (b) representation in form of four terminal networks; (c) simplified model for
parameter estimation.

T A B L E I. M A T H E M A T I C A L MODELS OF MOTORS

~OTOR D~E I~,~10R

D.C.~t0¢
o
JM~'M : M ~ " M, JM momentof inertia
MM : ~'I~'MF L. anemmt t ~
!u'i- Mr : MFo ' Mr, ~.~ R,,
.
trmtm reemce
flex I = ~ e

Die=t-En=e
¢~,a r.010tglua
I , ! MFo ceelmbfn~0n
MM : c~(~,. ~,.., ~ ) U s , MF
Mr, tiscemfrictm
Mr : M.o.MF. ~M "MF2 ~ *
M,.q M~ sOm fnctm
 Machine fault diagnosis 821

TABLE 2. MATHEMATICALMODELSOF DRIVECHAINCOMPONENTS

DRIVECHAIN DYNAMICBEHAVIOR i COEFRCIEN~

J, ,~,: Cs(q,2" q,, ) ods(~2" ~o,)-M, Jl,J2 rnomen~of inertia

Cs springrate
J, 'i~: c, (q,,- ,e~ ) ' d, (,i,,- ,i',) - M, ds damperrate

Belt-dine s slip

J~~J'4 ,l-, ~.,. CB( v,.4- e3)' d. (.,,, ,e.- ~)" M-,

J,,,~. = ICe(*3 -v,,e.), de( *3 - v.e4)] Vo-M,

r3/4
:re/r.
radius
[ transfer
,,,:,,,~l-s) J racto~
o,,,: withom/with load
Gear-box ~, =r6/rs tra~ferfactor
J, ,~,: c. ( .ed,,. - ,e,). d.( .i,d'¢. - ,i,,). M,
J6 ,~6 = [c,(,ps - qp61v,) * d , ( ~ , - ~ 6 1 ~ , ) ] / v . - M6

Oil,l~ariel
mm, mn friction~ts
M,. (m~o' m,,, 'i",) FL FL load force

Rolli~-l~g fo, f, depends,onmaterial

geomemcs
M.: fo ,~2:3 fl ~.FR t.t frictionconstant
FR radialforce

T A B L E ,3. MATHI~MATICAL M O D E L S OF W O R K I N G MACHINES (AND A VEHICLE)

MkCIIR DYI~ WlkVlO[ L'~'F~NTS

C'.mk6m ! mt of i~mia
p Ikkl ~ 1 ;
J~e" Me- p(l~ ~eQ-[t0 0'* it. ~JI
H • h.~-J. +,O-h.O'
ko. k,. k# ~ " ~ c d Imqw

Cuoq rc ndm d' ~ warlr.wkt,

lea. [,-q) to, ~m~'c,:lr~q fn
Mc • r¢ ko, b kh'" mc moumlcram!
b kl~dm

Vekide
• vdmk an...
n , l , v . 1~4" M,
~ , • co. c. ,i~v ' Ce ('V ¢, I nmmce
822 R. ISERMANN

models, a DC motor is used as an example, cf,

°@2 Table 1.
(A) Wide range dynamic operation.
The most general mode of operation is
described by the nonlinear differential equation
a)
JM~M(t) = qJ(IE)la(t) -- MF,, sign ~M(t)
FK;. 5. Two-mass rotatory element (a) scheme; (b) four -- MF, CM(t) -- M, (t), (36)
terminal network (MF= 0)
LA,/A(t) = --RAIA(t) -- qJ(IA)~0M(t) + UA(t), (37)
GII($ ) = ~2(s) = Ml(s)
with M~ as Coulomb and Mm as viscous friction
_ q°2(s) ~ ' , - " Ml(s) coefficients. Without load, i.e. for idle running
c,~(s)-M-~- j ; o _ , , ( O = ~ M~,, Ml(t) = 0 and with load the torque equation of
the load has to be added.
(B) Small range dynamic operation.
bearings is, for example, described by For small deviations around an operating
Mm(t) = MFj0 sign ~(t) point ~MU> 0 the equations can be linearized
+ MFll(p(t)(l~p(t)l 4:0), (31) JM~M(/) = ~,, A/A(t) -- M,(pm(t) - AMI(t), (38)
where MFm is a Coulomb friction and ME11 a LAIA(t) = --RA A/A(t) -- qJ,,~M(t) + AUA(t).
speed proportional (viscous) friction coefficient. (39)
By connecting the single elements of a
(C) Shut-off dynamic operation.
machine, Fig. 4(a), the overall model of the
By switching off the power supply a special
machine is obtained. If all models can be
dynamic behaviour results with UA(t)=0 for
linearized a state vector of nth order
t -->0 and initial values UA0, IAO, ~M0
x_r (t) = [~p,(t)cp,(t)~p2(t)cp2(t) " " qo.(t)qg.(t) l,
JMCM(t) = - M ~ , - MFItPM(/) -- M,(t). (40)
(32)
Without load it is M~(t)= 0, cf. operation mode
can be defined, and the overall state model of (A). If the voltage U'A(t) after the cut-off switch
the motor and the drive chain becomes can be measured additionally
~(t) = A_.S(t) + _bu(t) + [z(t), (33) qJ(IE)¢M(/) = UA(t) (41)
u(t) = AU(t), z(t) = AMs(t), holds.
(34)
y(t) = cr~(t). (D) Wide range static operation.
The models for different steady-states are
AU(t) is the deviation of the actuator input, e.g. obtained by setting all derivatives of the speed to
a voltage or an actuator position and AMs(t) the zero. Then it follows from equations (36) and
deviation of the load torque. Many of the loads (37) with q~M= toM,
or working processes like pumps, metal cutting
or vehicles are nonlinear, cf. Table 3. They tP(iE)iA - MFx,- MF, tOM= hTt,. (42)
introduce further (nonlinear) dynamic models. RAIA + W(/E)tOM = UA- (43)
Frequently the overall model can be simplified
considerably, as the elasticity of the shafts is By eliminating ]A or tOM one obtains, e.g. the
small in comparison to belts or elastic clutches or characteristic curves tOM= f(h41) of IA = g(M0"
gears. If the elasticity of the belt and the gear Table 4 indicates the identifiable process
then is lumped together (physical model coefficient as dependent on the measured
reduction) a lower order model results. By variables and the operation mode for the DC
further lumping of parameters even second order motor in idle running. Five process coefficients
models may result, in the simplest case as can be determined in operation mode (A), four
in mode (B), (C) or (D).
]ep.(t) + El~.(t) + e~o.(t)
By the combination of experiments in
= K, A U ( t ) - AMs(t). (35) different operation modes, e.g. (A) and (C) or
(A) and (D) or (B) and (D), all process
3.2. Operation modes coefficients can be estimated. If the DC motor
With regard to the practical operation and the drives a working process/load, torque equations
fault detection method by parameter estimation as given in Table 3 have to be inserted for Mj(t).
the following operation modes of the machine Depending on the type of equation, some
are significant. To demonstrate the different coefficients of the DC motor and the working
 Machine fault diagnosis 823

TABLE 4. IDENTIFIABLE PROCESS COEFFICIENTS OF A D C MOTOR IN IDLE RUNNING

( M l ( t ) = 0) IN DEPENDENCE ON THE OPERATING MODE MARKED BY X

Measured Identifiable
Operation signals process coefficients
mode UA IA (PM U'A JM C~) Cvt ~11 LA RA Assumptions

(A) Wide range x x x -- x x x -- x x Wisknown

dynamics
(B) Small range x x x -- x -- x x x x
dynamics
(C) Shut off x x -- x x x -- -- JM is k n o w n
dynamics
(D) Wide range x x x x x x -- x
statics

process may lump together, e.g. the ratio of resulting in a vector equation
inertia. If, however, additional variables can be
~, = q'_0 + ¢. (46)
measured, for example the pressure difference
and flow ratio of a pump, many coefficients of Minimization of the loss function
the working process can also be determined by N
dynamic and static experiments, see Isermann V = ~, e 2 ( k ) = e_rg, (47)
(1984) and Isermann and Freyermuth (1991a, b). k=O
This consideration of the operating modes is yields the well-known least squares (LS)
important for the practical application, because estimation equation
the parameter estimation presupposes changes of
the variables. Machine tools or robots show 0 = [_Wr _WI-'_Wr~', (48)
sufficient excitation of the dynamics during in nonrecursive form, see e.g. Eykhoff (1974)
normal operation. Therefore operating mode and Isermann (1992a). By partitioning the
(A) or (B) holds. For machines without frequent recursive least squares method (RLS) can be
dynamics excitation like in power plants or obtained. For the implementation modified
process industries, modes (D) and (C) can be LS methods, like discrete square root filtering
taken. (DSFI) or UD-factorization, are recommended
to improve the numerical properties, or the
method of total least squares, see Neumann et
4. S Y M P T O M GENERATION BY PARAMETER al. (1988), if the signals are rather noisy. The
ESTIMATION
parameter estimation methods may also be used
4.1. Parameter estimation for dynamic processes for nonlinear processes if the equation error e(t)
From theoretical modeling the model structure is linear in the parameters (see Section 2) or the
is known in the form of a differential equation output error is used together with an iterative
(in continuous time). The measured input and parameter estimation method.
output signals will be sampled and are therefore A certain problem is the determination of the
only available in discrete-time. However, the signal derivatives required in the data vector
parameters are estimated for continuous-time ~r(t) if only the signals u(t) and y(t) can be
models, because of the possibility to interpret measured. Several investigations have shown,
the parameters physically and to calculate the that state variable filters (SVF) (Young (1981))
basic process coefficients. are most suitable. They are a state repre-
Hence, a differential equation of type sentation of the transfer function
equation (2) is considered and written in the
form yF(S) 1
F(S) = y(s) - f o + fls + . . . + f~s~,, oc > n, (49)
2/(0 = ~T(t)19(t) + e(t), (44)
and provide simultaneously the time derivation
(cf. equation (14)), where e(t) is the equation
(without differentiation) and filtering of the
error. Measurements of the input and output
noise. The parameters f~ can be taken from a
signals are made at discrete times t = k T o ,
Butterworth filter design. The cut-off frequencies
k = 0, 1, 2. • • N with T0 the sampling time. The
should be fitted to the largest process eigen-
deviations equation (3) are determined and also
frequencies of interest. More details are given in
the derivatives. Then N + 1 equations are
Peter and Isermann (1989).
obtained
The process parameters t9 depend on physi-
y ( k ) = y)r(k)O + e(k), (45) cally defined process coefficients ~o via algebraic
824 R. ISERMANN

functions _0=f(p). As the knowledge of the show oscillations with a variety of frequencies.
process coefficients p is very well suited for Hence, the application of correlation analysis or
detecting and diagnosing faults it is of interest to Fourier analysis (mostly realized using the Fast
solve the inverse relationship e =f-'(_0) (equa- Fourier Transform (FFT)) is a further source for
tion (16)). In general the process parameters the detection of faults in machines. However, a
are nonlinear algebraic functions of the process correlation function or a frequency spectrum
coefficients ~o. Therefore no general solution in forms a nonparametric representation of the
form of equation (16) can be given. In many signal and is not well suited for direct automatic
cases for models of first or second order a direct supervision. Therefore parametric signal models
solution can be found. If higher order systems should be preferred which directly allow to
are considered, computer algebra may be used. estimate the main frequencies and their ampli-
For a unique determination of ~o an iden- tudes. A new estimation method could be
tifiability condition can be given. It can be derived as follows. The signals are modeled as a
derived from the implicit function superposition of damped sinusoids
q = 0 - C_z = 9, (50)
y ( t ) = ~ yo~e -a' sin (Olil + ~i)" (57)
where i=1

z_ = g ( e ) " (51) The corresponding z-transformed signal

The implicit function theorem states that a b~ + b,iz -1
necessary condition for the solution for ~o in the y ( z ) - - , = , 1 + a,---~z---i+ a---~z-2 6 ( z )
neighbourhood of the solution ~oo is that the
bo + b l z -t + • • • + b 2 ~ _ l z I-2~
functional determinant - a(z)
1 +amz -i + • • • + a2~z - ~
det _Qp:~ 0, (52)
where _QP is the functional matrix
O_ = lae. (53) expresses the signal y(z) by an autoregressive
For more details see Isermann (1992b). moving average (ARMA) process. The
Kronecker &impulse is used to excite the
4.2. Parameter estimation f o r static processes oscillations initially. Hereby the relation be-
If the process model for the steady-states is tween the amplitudes Yoi and parameters b/i are
expressed in the form of equations (1) and (5) rather complicated, cf. z-transform tables.
and an equation error e is introduced then Therefore only the a/parameters are estimated
via a two-step maximum entropy estimator. The
Y ( j ) = ~,~r(j)_0s + e ( j ) , j = 1, 2 . . . . . L, (54) amplitudes yo~ and frequencies toi then can be
holds for L steady-state operation points, or in calculated directly (Neumann, 1991).
vector form The effect of noise can be reduced by
introducing the autocorrelation of y(k):
Y = + ¢. (55)
~yy0r) = -a,q)yy(r - 1) . . . . . a2,,,¢yy(r - 2m)
Applying the LS method yields the parameter
estimates + b~¢66(r) + " " + b ~ , - l ~ a a ( ~ - 2 m + 1)

_0~= [_ W r _ ~ l - ' ~ ' , (56) + On,(~) (59)

cf. e.g. Isermann (1992a). Direct solutions or if the noise n ( k ) is uncorrelated and therefore
recursive solutions of this equation can be ~n,(~)={;0 for~=0
applied as well.
for t:~0"
4.3. Parameter estimation f o r vibration signals After ( 2 m - 1) time steps the signal y ( k ) is in
Especially for machines, vibration analysis has stationary oscillation reducing equation (59) to
always played an important role. In most cases
amplitudes and frequencies within a defined
bandwidth are monitored by using vibration = ~,,(~') + a , ~ , , ( t - 1) + . . . + a2,,.dpyy(r - 2m),
sensors for position, speed and acceleration. A (60)
complete spectral analysis is usually not applied
just for monitoring. In addition to special known as the Yule-Walker equation (Box and
vibration sensors many other signals like DC Jenkins (1970)). A linear equation system for
motor current, position, speed, force, torque different time shifts r taking into account the
 Machine fault diagnosis 825

symmetry4'.(-0 = 4'.(0 detected with reference to normal values as a

basis for the decision of whether symptoms exist
¢,(o) 4',(1) ... 4',(z,n) ] or not. Since the coefficients appear usually as
4',,(1): 4',r(O): ... 4',y(2m- 1) stochastic variables because of noise effects for
example a likelihood ratio test or Bayes decision
L¢,,,i~'n) 4',(2m - 1) 4',(0) can be applied, see Geiger (1985) and Isermann

×
[1]at = , (61)
(1984). Also a run sum test or a two-probe t-test
have shown to be useful, see Tou and Gonzales
(1974) and Goedecke (1986). The statistical

.L decision subdivides in two phases.

Phase 1. Determination of the normal state
(training).
which leads to a solution for the unknown
parameters aj and no. The frequencies of "The mean values and the variances of the
significant oscillation components can be found coefficients, for example, may be estimated."
by solving for the poles of the characteristic Phase H. Detection of changes.
equation
"Changes of mean values and variances of the
A(z) = f i (1 + ~,,z + m2/z2), coefficients in comparison to the normal state
i=1 may be observed."
and then
1 [ -aq~] As a result a significant deviation in these basic
toi = - arc cos (62) statistical quantities of each process coefficient is
to
indicated by a specific symptom Si. Hence the
The signal amplitudes Yoi of the oscillation analytic problem solution concludes by providing
components are obtained by the autocorrelation a combination of different (analytically gen-
function of equation (57): erated) symptoms.
Because much analytical data processing is
,.(,) = Y& cos (,o,0 required to reach this stage of symptom
i=1 2
generation, implementation languages like
= ~ 4'oi cos (oJiO. (63) FORTRAN, PASCAL or C are required.
i=l
5. FAULT DIAGNOSIS BASED ON SYMPTOM
For different time shifts r = ~To one obtains
PROCESSING
[ 4'.(to)1 This section describes a concept which was
,.(2ro)/ developed for the fault diagnosis of machines
and drives and implemented with on-line
4',,,(inTo) J coupled personal computers. Now heuristic
knowledge has to be processed and implementa-
Fcos.,r0 cos-2 ,, . . - cos-. o 1 tion in C or P R O L O G was preferred. The

/ c o s m t o , T0 cosmo)2T0 . . . cosmtomTo/
/ intention is that the complete computer assisted
fault diagnosis (CAFD) can be performed on
personal computers and used as an on-line
engineering expert system.
[4'0,']
5.1. Symptom representation
The fault diagnosis is based on symptoms
(facts) which are available from different
4, sources. The following categories can be defined.
By inversion 4'~ can be. calculated. Then the (a) Analytical symptoms.
signal amplitudes follow from An analytically determined set of symptoms
~, (as discussed in the previous section) is stored
Y0~= V~--4'0~. (65)
in a specific area in the process computer
Further information and applications can be memory. This "symptom buffer" is arranged as
found in Neumann (1991). follows (see Freyermuth (1991a, b)).
(record SYMPTOM Sio of)
4.4. Symptom decision (1) Number of the technical subsystem the
After the process coefficients or signal symptom is related to.
coefficients (like frequencies and amplitudes) are (2) Symptom name expressed as a symbolic
determined, significant changes have to be string.
826 R. ISEaMANN

(3) Numerical mean value of the symptom expression

(provided by the symptom decision module).
IF (condition) THEN (conclusion)
(4) Numerical reference value of the symptom.
(5) Physical unit of the symptom. where the condition part contains facts (symp-
(6) Calculated confidence number (provided by toms) as inputs being associated by Boolean
the symptom decision module). AND and OR connectives. The conclusion part
(7) Time of symptom entry into the buffer. represents a so called "event" as a logical cause
(8) Explanatory text. of these facts. Chaining of the rules now
This dynamic database acts as an interface establishes the causal dependencies of symptoms
between the analytical and the heuristic part of and faults (considered as "basic events") in a
the diagnosis system. hierarchical manner. Thus intermediate events
(b) Heuristic symptoms. E~, k = 1 . . . 1 are introduced. This natural
The second class of facts is represented by a procedure results in the establishment of
set of "heuristic" symptoms 5eh, neither directly fault-symptom trees ("directed graphs") struc-
measurable nor analytically computable, e.g. turing the rules hierarchically, relating symptoms
specific acoustic noise, optical observation or to events and faults by a systematic approach
other facts depending on the empirical knowl- (see Fig. 6).
edge of an experienced human operator who is Hence this methodology of knowledge repre-
consulted by the diagnosis system to provide sentation resembles the well-known description
additional symptoms (see Section 5.3). of fault tree analysis (FI'A) and event tree
(c) Process history and fault statistics. analysis (ETA), see also the review in Lee et al.
A third category of facts depends on the (1985). Arrows indicate causal relationships of
general status of the technical process under the different symptoms and events in the
consideration. This set of symptoms is derived knowledge base. In general the appearance of
from the available information on the process symptoms depends on events which may be
history and the fault statistics. Corresponding caused by other events. The leaves represent the
data from mass storage files (e.g. values from symptoms and the fault is considered as the root
SPC="Statistic Process Control") can be node of each fault-symptom tree. Every basic
treated as symptoms belonging to class (a), while event in the list of known faults ~, j = 1 • • rn has
the specific heuristic knowledge of the main- to be associated with its own tree. The following
tenance staff is considered as part of class (b). advantages are obvious.
• Tree-like structures achieve a clear and easy to
5.2. Knowledge representation understand representation of the heuristic
For the purpose of establishing heuristic knowledge. Furthermore they offer a high
knowledge bases several methods for knowledge degree of flexibility for development and
representation do exist, see e.g. Frost (1986) and modification.
Torasso and Console (1989). As a suitable tool • Additionally temporal reasoning is provided
for a systematic treatment of heuristic knowl- by the proposed evaluation of the fault-
edge (especially in the diagnostic domain) symptom trees, because the edges of the trees
specific rules are applied in order to set up inherently represent causal relationships be-
logical interactions between observed symptoms tween the different nodes.
(effects) and unknown faults (causes) and to The condition part of a rule is structured in
structure the knowledge in a problem-adapted Boolean disjunctive normal form. Applying
manner. A single rule is described by the notation and nomenclature from reliability

o o o o

FIG. 6. Structure of the fault-symptom trees (heuristic knowledge base); Si: symptoms; r: logical equations;
E,: events; Fj: faults.
 Machine fault diagnosis 827

theory for binary systems (Barlow and Proschan mendations for maintenance or repair with the
(1975)) this corresponds to the assignment of a aim of achieving the normal process condition.
Boolean equation for parallel-serial connectors The structured heuristic knowledge base can
be constructed and modified by using the
F(~) = •, = 1 -- H (1 -- H ~i)' (66) knowledge acquisition module. Hereby the most
u \ "(~) / important issues are the deep technical com-
to each single node in the tree, defining the prehension of the considered technical process
logical relationships between specific inputs and skilled human expertise. An interactive
_~e {e, 5e} (set of events and symptoms) and dialogue, including windowing technique, sup-
outputs r/e {e, ~:} (set of events and faults), lu ports entering the fault-symptom trees (Nold et
denotes the number of conjunctions in the al. (1987)). Plausibility checks are built in to
condition part and v(/~) the number of literals guarantee structural consistency of the estab-
per conjunction. lishment knowledge bases.
To cope with uncertain conclusions during the
inference process each node (rule) of the tree is 5.3. Knowledgeprocessing (inference
enabled to calculate an output (event) con- mechanism)
fidence number c(rh) from input confidence To process the knowledge a hybrid reasoning
numbers c(~). Assuming that ~ is statistically approach is devised including forward and
independent the following probability measure backward chaining, as this is the natural
or confidence number can be calculated (Bariow approach of human experts coping with diagnos-
and Proschan (1975)): tic problems. After the selection of a specific
knowledge base a complete combination of all
c(rb) = 1 - ~ ( 1 - ,~c(~i)). (67) rules (all nodes of the fault-symptom trees) is
performed based on the symptoms being
provided by the symptom buffer (Modus
The complementary confidence number is
defined by
Ponens). The inference mechanism calculates
the confidence numbers c(Ek) or t~(Ek) of events
= = 1 - (68) and c(~) or g(Fj) of faults according to equations
(67) or (69). This diagnosis step is performed
During the diagnosis process this confidence
automatically. At this stage all concluded events
number is used for two different purposes:
are displayed to the operator.
• control and improvement of the inference
Now a refinement of diagnosis procedure is
mechanism;
achieved by entering additional information due
• generation of additional information with
to heuristic symptoms not available from the
respect to fault decision to increase the
analytical analysis. Backward chaining (Modus
certainty of diagnostic results.
For resolution of conflicts within the inference
Tollens) is applied by the inference mechanism
selecting the most plausible node according to
process the following meta-rules are applied:
the evaluation result of the meta-rules. This
(1) the conjunction parts of a specific rule are
hypothesis is to be verified or disproved in the
selected prior to the evaluation of
next step, activating an interactive dialogue with
disjunctions;
the process operator to enter the required
(2) conjunctions with at least one known fact
information. Then the inference mechanism
are prioritized;
(3) the rule with the largest predicted confidence restarts forward chaining. The procedure is
number repeated until terminated by the user or no more
= max [c(r/i)l (69) solutions can be given.
During the backward chaining procedure the
is selected, where unknown facts in AND user is able to step through the knowledge base
relations are weighted with 1.0, in O R extracting information about already derived
relations with 0.0 in order to assume a worst conclusions (events or faults). Finally the
case. analysed fault (or if declared during knowledge
Each node of the trees can be equipped with acquisition, an intermediate event) together with
information to support the human process explanatory details available in the knowledge
operator during the diagnosis procedure: base are indicated. Proposals for maintenance
• Events. Each rule is associated with additional and repair, and hints as to possible fault
messages containing questions and explana- consequences (subsequent defects) are given.
tions to assist the system user. Examples for this knowledge based fault
• Faults (basic events). For each fault a diagnosis and its implementation are described
description exists including hints and recom- in Freyermuth (1991a, b).
~ITO 29~4-B
828 R. ISERMANN

6. A P P L I C A T I O N : F A U L T D E T E C T I O N A N D J * f o M ( t ) = q U A ( t ) -- M y , t o M ( t )
DIAGNOSIS OF THE ELECTRICAL FEED DRIVE
OF A MACHINE TOOl. - Ms~ sign toM(t). (71)
The automatic fault detection and diagnosis is Hereby MRj and MF, are the summarized
of very high importance for machine tools, coefficients of of Coulomb and viscous friction of
because this is the prerequisite for the increase the whole feed drive.
of unmanned production and longer lifetime of The overall ratio of inertia related to the
the machine and the tools. motor axis is J * = J , + v-2J2, where J, is the
Machine tools frequently have several feed inertia of the motor and ./2 the inertia related to
drives. For the investigation of model based fault the spindle axis and v=wM/~o,=d2/d, the
diagnosis methods, a special test rig was built transfer factor of the belt drive. The verification
such that faults could be generated artificially of the friction laws is shown in Isermann et al.
(Fig. 7). It consists of a table which is moved on (1990). If the belt elasticity is taken into account,
the slideways by a ball screw spindle. The instead of equation (71) the drive chain can be
spindle is driven by a speed controlled DC motor modelled by
through a toothed belt. With tension screws at
J, ipM(t) = qJ/A(t) -- dB Aqb(t) - CB Aqo(t), (72)
the sockets, the friction of the slideways can be
changed and with a tension screw at the casing J2~bs(t) = VWIA(t) -- vJ, (OM(t)
the tension of the belt can be influenced. The - MF,sqbs(t) - M~s sign ~s(t), (73)
measured variables are only the voltage UA(t),
the current IA(t) of the armature and the speed where
~oM(t) of the DC motor in closed loop. The Aq0(t) = q0M(t) - Vq0s(t), (74)
considered electro-mechanical drive can be
and CB is the stiffness and dB the damping factor
considered also as example for other similar
of the belt, M~,s and MF,S the coefficients of
drives as actuators, presses or other servo
Coulomb and viscous friction of the slideways
systems.
and the bailbearings. For the fault detection it is
now assumed that the faults lead to parameter
6.1. Theoretical modelling
changes, i.e. muitiplicative faults are considered.
Several investigations have shown that the
feed drive system can be simplified to a two mass
6.2. Parameter estimation
system (motor and table) with the main elasticity
Based on equations (70) and (71) and
within the belt (Reil3 (1991)). If the belt
measurement of UA(t), IA(t), tOM(t) following
elasticity is neglected, the motor and the drive
parameters are estimated by the DSFI method,
chain can then be modelled by, compare
Section 4.1:
equations (36) and (37), assuming wide range
dynamic operation a,,, b,,,l=[L,,/q' n,,/'V l/q'],
LAIA(t) = --RA/A(t)-- qJtOM(t) + UA(t), (70) _02
T= [a2, a2,, a2o0] = [ J * / ~ M F , / ~ Mw,/W].~75
)t

Belt tenszonlng
%
screw

HTB I @ DC-Motor Tacho-

[[ I UA, IA generator
Belt ~ F~ q [~ /

__ ~ \ I I I~'qlll Ir----"
[TableI II -, ",", / k
~--- !
"
Ball screwll
II
I I !
I
I/
\
"N~
___ :~pindle ",>.~ ~-~l ===n r _ "~
"l / / / I//// //t ' " ' ----11 i (+) I
.. MTsl -~-- - "s' s , - , Posltion Iii m%
. - , ',: I
/~-~
II -~,~- Shaft L~_J sensor ~ ~'----'- ~
Tensionzng screws

IIIIIIIIIIIIIIIII/II/IIIIIIIIIIIIIIIIIIIIIIIT
FIG. 7. Scheme of the feed drive test rig. DC motor: N = 1.8kW; toothed belt: (240 H 150); I = 609.6mm;
ball screw spindle: 1 = 250 mm, d = 32 ram; table: m = 150 kg.
 Machine fault diagnosis 829

15

>~
t UA(t): Armature voltage
~ IA(t): Armature current
1
t
"-I~ 1.0 t~M(t): Angular veloc,ty

. /

o o
O.

{
-10
0 1 2 3 4 5
Time [s]

FIG. 8. Measured signals of the feed drive.

The process coefficients can be calculated as 6.3. Experimental results

follows: The parameter estimation was performed with
sinusoidal changes of the speed reference
u / = 1/blo RA = ato/bto LA = a , / b l o , variable
J* = a2t/bt MFt = a2o/bto MFo = a2oo/bto. W(t) = 1000 r p m . sin (tOw(t));
(76) tOw = 3.142 1 sec -1.
Using equations (72), (73) and measurement of Figure 8 shows the measured signals at the
IA(t), (PM(t), (ps(t), we can estimate the output of analogue antialiasing Butterworth
following parameters: filters of eighth order with corner frequency
_Or = [a3, a3ol = [dB cB], 100Hz. The sampling time for parameter
estimation was T0 = 6 msec and the derivatives
_04
r = [a,2 a,, a,,o] (77)
were obtained from a digital Butterworth
= [J2/v MFIS/V Mn)s/V], state-variable filter of fourth-order with corner
where the inertia of the motor J, is known. In frequency 50Hz. A considerable information
this case the following process coefficients can be source on the process is the armature current, as
calculated: can be seen from the high-frequency components
da = a31 CB = a~} ,/2 = va42
and the peaks during change of motion direction
(78) due to the friction law. In Fig. 9 a fast
MF~S = va4~ MF~)s= v a 4 1 o . convergence can be seen for the estimated
This concludes the generation of the analytic coefficients of the DC motor. They converge
symptoms. within a time period of 0.5 sec and agree fairly

4O
0
O~
Estimated motor-parameters
m
at ~(t)-2s,ooc 4{
.J
...................... 30
a3 z rQ] 0e"

I1: U
ID -.- ' CA [mH'] • 20
U
c
CVs3
, ,/.. A

I0 E
0
g
E
~- 0 0
1 2 3 4 5
Time Es'l

FIG. 9. Convergence of the process coefficient estimates.

830 R. ]SERMANN

4O% ~5
RA: Resistance

LA: i n d u c t a n c e -:D °~
.c_ 30%
<] ~A: Flux linkage A~A [%3

g
>
2o~ ~ 55
d~

~
"D
Q.
; Io~ 40 E

E ;
~_~ 0% -- -- J~ ................ 25 ~o
\ _ _ \ _ _ ~ /nv rzJ z
\
-I0% !0
0 25 50 75 i00 125 150
Number of samples

FIG. 10. Change in coefficient estimates of the DC motor during an ongoing experiment (start of the cold
engine and too large belt tension). Corresponds to fault ( F t ) .

well with the coefficients from the data sheet (F1) Overheating due to insufficient cooling or
overload.
R A = 1.685 f~; LA = 15.6 mH; (F2) Commutator failure through wear.
(II) Faults of the mechanical system.
= 0.83 Vsec. (F3) Lack of lubrication of the slideways.
(F4) Too large bracing of slideways.
As nominal status the following adjustments
(Fs) Too large tension of the belt.
were defined:
(F6) Defect of the belt.
• screw torque for belt tension MTB = 450 Ncm;
(FT) Additional load by tools, etc.
• screw torque for slideway tension M r s =
In the following examples for faults (F,), (F4)
1500 Ncm;
and (Fs) are shown.
leading to following nominal estimates of the
Figure 10 shows results for enlarging the static
mechanical system:
load by increasing the belt tension to MTB =
650Ncm. The slideway tension was Mrs =
J* = 0.0073 kg m2; MF0 = 1.02 Nm;
MF1 = 0.00485 N m s e c rad- ~. 1500cm. The temperatures given in Fig. 10 are
the casing temperatures
measured with a PT 100 resistance thermometer.
of the DC motor

Several faults were then generated artificially The estimated armature resistance and flux
and by comparison of the estimated coefficients linkage show deviations according to the
with the nominal values and statistical decision increase in the temperature. The inductance
tests for significant changes, the following faults changes only little.
turned out to be detectable (except (F2)): The results of increasing the screw torques for
(I) Faults of the DC motor. the slideway tension from M r s = 1 0 0 0 N c m to

300
MTS Tightening torque on e a c h screw
l
MF Coulomb {r,c~.lon coeff=c,ent /
c 30~ o: 250
<~ MF I " V i s c o u s friction CO ....... '"

-~ zo~ zoo
.- ~ ........... nMFO [%] -~
u
Mx ~....... .----~"

[
o~ 1oo
c~

-i0% 50
0 50 100 150 200

FIG. 11. Change in the friction coefficients (I¢IF,) and (MF,) through stepwise increase of the tension screws
of the slideways (fault (F,)).
 Machine fault diagnosis 831

80% 135
Nr0: Coulomb fr,ctlon coefficient ,~ .

MFI V,scous fr,ct,on coeff,c,ent iv

I\.d-~ - ' y , ~ j,.1._~
.c 60~ 115
<~

95 7E
._~ t)
AMFo [%3 i"" .......t-~.{ ~,. . . . . I XMTB t..a

73
20g
\ L__/ 75
m
I-
z
E
MTB: Torque setting on 55
the belt tensioning screw

-20% 35
0 30 60 90 120 150 180
Number of samDles
FIG. 12. Change in the friction coefficients (l~lRi) and (l~lFi) through stepwise increase of the belt tension
(fault (Fs)).

2750 Ncm in steps of 250 Ncm is shown in Fig. 6.4. Fault diagnosis
11. The Coulomb friction coefficient increases To demonstrate the treatment of analytic,
significantly, whereas the viscous friction heuristic symptoms and the process history as
coefficient hardly changes. However, if the belt described in Section 5, three possible faults,
tension is changed, see Fig. 12, both friction (F3), (F4) and (F6) are considered. Figure 13
coefficients show significant increases. For larger shows the fault-symptom trees. In the following
tensions the viscous friction changes more than S~ means the symptom has occurred and S~ the
the Coulomb friction. This can be explained by symptom has not occurred.
the tension dependent speed proportional torque
of the belt (Erxleben (1984)).
Experimental results for other faults are
described in Geiger (1985), Trawinski and
Isermann (1989) and Isermann and Freyermuth
(1991a, b).
Table 5 summarizes the effects of the faults on
the changes of the process coefficients if the belt
elasticity is neglected and only equations (70)
and (71) are used as a basis for parameter
estimation. For each fault a different pattern of
changes is obtained.
Hence it is possible by simple pattern
(a)
recognition of the estimated process coefficients
to pinpoint even small faults (F,) or (F2), (F3)
through (F7).
The required measurements of voltage,
current and speed of the DC motor are available
in modern machine tools, such that no extra
sensors for, e.g. forces or acceleration are
required.
Ib)
TABI.E5. EFFECTS OF DIFFERENT FAULTS ON THE ESTIMATED
PROCESS COEFFICIENTS. OBTAINED WITH THE MODEL EQUA-
TIONS (70). (71) AND (73). + ''- +4-+: INCREASE
SMALL'' • LARGE, -- . . . . . . " DECREASE SMALL"• "
LARGE,0" NOCHANGE
P, RA LA ~A p, J* MR, MF,
F,
(F,) ++ + -- (F~) 0 ++ --
(c)
(F2) + 0 -- (F,) 0 ++ 0
(F0 - +++ +++ FIG. 13. F a u l t - s y m p t o m trees for .some faults of the
(F6) 0 -- +++ feed-drive: (a) F.: defect of the belt; (b) F3: lack of
(F# + ++ -- lubrication of the slideways; (c) 174: too large bracing of the
slideways.
832 R. ISERMANN

Knowledge representation c(E3) = c(Sa5 ) • c(S:o) • c(S;,m)

Definition of events. = 0.3.1). 1.0= 0
(E~) Increase of solid body friction. such that
(E2) Irregular lubrication of slideways. c(G) = 0 (as also c(E2) = 0 from
(E0 Increase of Coulomb friction. backward-chaining
(E4) Decrease of belt tension. through the operator)
(Es) Decreasc of rotational stiffness.
(E6) Old belt--high tension--high load. C(F4) = 0.
lET) Abnormal motion of carriage. Hence, faults (F3) and (F4) have not happened.
Now as an example a belt defect is considered. Now, Fig. 13(a):
Following symptoms were observed. c(E,) = c(S,,) • c(S,,2) = 1.0.0.7 = 0.7.
Analytic symptoms
(S~,,) Decrease of J* (rotational inertia), c(Es) = c(S,3) • c(S;,4) = 1.0.1.0 = 1.0.
c(S,,) = 0. The heuristic symptoms cannot provide higher
IS;e) Decrease of Mvo (Coulomb friction confidence for E4 in this case, as
cool.), c(S,2) = 0.7.
c(ET) = c(S.,) • c(S.2) • c(S,,)
IS,.0 Decrease of CB (torsional spring stiffn.),
c(S,,s) = 1.0. = 0.3 • 0.5 • 0.2 = 0.03.
(S,,4) Decrease of dB (rotat. damping coef.), Now backward-chaining is performed, assuming
c(S,4) = 1.0. the fault IF6) has occurred, compare Fig. 13(a).
IS,.0 Decrease of Mw (viscous friction coef.), The operator is asked for information about
c(S,s) = {I.7. process (belt) history. The following conclusions
(S,6) Increase of J*, c(S,,6) = 0.4. can be drawn.
(S,7) Increase of Mvo, c(S,,7) = 0. (i) For the belt history:
(S,~) Increase of cB, c(S,~) = 0.
IS,,,) Increase of d , , c(S,,,) = 0. c(E6) = c(S,) • c(Ss) = 1.0.
(S, lo) Increase of Mm, c(S,m) = 0. (ii) The centre distance of pulleys is not
Heuristic symptoms reduced, i.e. the tension screw had not been
IS, i) Charac. acoustic noise present, C(Shl)= loosened. Therefore
0.3.
(S,,2) Test run: sluggish movement of carriage, c(S~) = (1.0 - c(S0) = 1.0.
c(S,2) = 0.5. So the fault IF6) is diagnosed by forward-
(Sh3) Positioning: apparent inaccuracy, c(Sh3)= chaining again with confidence
0.2.
(S,4) Excessive increase in motor temperature, c(F6) = c(E,) • c(Es) • c(E6) • c(Ss)
c ( S . ~ ) = 0. = 0.7. 1.0. 1.0- 1.0= 0.7.
Process history
This means a defect of the belt has occurred, for
(St) Type of belt (ordinary, not reinforced,
example by breaking of some cord threads.
c(S,) --- 1.0.
($2) Number of operating hours: high, c(S2)=
0.5. 7. OTHER APPLICATIONS
(Ss) Initial tension: high, c($3)= 1.0. The theoretical and experimental results for
($4) Load: high, c($4)= 0.9. the fault detection and diagnosis of other
(SO Position of tension screw changed (loos- machines or mechanical processes are briefly
ened) c(Ss) = 0. summarized, see also lsermann et al. (1990).
(S~) Great number of operating hours since last
lubrication. 7.1. DC motor and centrifugal pump
($7) Temperature dependence on grease is The theoretical and measured signals are:
high. voltage, armature current and speed of the DC
(Sx) Temperature of slideways high. motor, input and output pressure and mass flow
(SO High load. of the pump. Four linear and nonlinear models
(Sin) Tension screws for slideways changed. contain nine process coefficients. With step
Knowledge processing changes of the reference variable for the
Forward-chaining leads to, compare Figs 13(b) cascaded speed control nine process coefficients
and (c) could be estimated and 13 artificially generated
faults like affected brushes, insufficient DC
clEf) = c(S,.s) • c(S,,6) • c(Sa7 ) motor cooling, bearing faults, split-ring clear-
= 0.7- 0.6 • 0 = 0 ance gap, affected impeller could be detected
 Machine fault diagnosis 833

(Geiger (1985) and Isermann and Freyermuth 8. CONCLUSIONS

(1991a, b)).
The presented methodology for fault diagnosis
based on parameter estimation, calculation of
7.2. AC motor and centrifugal pump physical coefficients and symptom-fault tree
In order to avoid dynamic changes of the processing has in several applications proven to
speed only measurements during different steady give valuable early information on slow and fast
state operation points and during shut-off are developing faults. The faults may appear in the
made, operation modes (D) and (C), Section 3. components of a process, including actuators and
The measured variables of the AC motor are sensors, in open loop or closed loop operation.
voltage, current and speed. Based on different The described methods can be applied as well
load affected operating points faults like air gap for static as for dynamic linear and nonlinear
changes or high rotor temperature could be process behaviour, depending on the specific
detected. For the pump in addition volume flow operating modes of the process. The fault
and the pumphead were measured and faults like detection and symptom generation by parameter
wear or impeller deposits or broken blades could estimation has the following advantages: it is
be detected (Nold (1987) and Nold and especially suitable for multiplicative faults, only
Isermann (1988)). the model structure is required, several para-
meter changes are uniquely detectable, very
7.3. Industrial robot small changes as well as drastic changes are
detectable, the involved physically-based
The investigated robot has six axes driven by
continuous-time models allow a deep fault
DC motors. The measured signals for each axis
diagnosis, on-line nearly real-time fault diagnosis
are voltage, current and speed. Based on
is possible. Except in using different steady-state
analytical models and heuristic symptoms faults
measurements, an input excitation is required in
like backlash or large friction in the drive chain,
the case of dynamic models.
defective bearings or overload could be diag-
nosed (Freyermuth (1990, 1991a, b)). For fault diagnosis it was shown how, as well,
analytical symptoms as heuristic symptoms in the
form of operator observations or process history
7.4. Grinding machines can be processed by if-then rules of inference
By measuring the feed position, the main based on the cause-event chains of the process.
spindle synchronous motor current and work- By weighting with confidence numbers also
piece diameter the parameters of a grinding uncertain knowledge can be treated. The
process can be estimated. Their changes allow to programmed knowledge processing, allows for-
detect faults like abrasive effects of the grinding ward and backward chaining and the diagnosis
wheel, wrong pairs of grinding wheel and result is given with an explanation.
workpiece, wrong coolant flow. Signal para- The described methods were illustrated by
meters estimation of the motor current, Section using mechanical processes as case studies.
4.3, allowed detection of faults like wrong These included well-defined models as, e.g. DC
grinding wheel dimming, roundness errors of the motors, pumps and robots, but also not well-
workpiece and grinding wheel and chattering defined models as, e.g. grinding, milling or
(Janik and Fuchs (1991)). drilling. However, the parameter estimation
based fault detection can also be applied to
7.5. Machine tools for milling and drilling processes with more complicated models, like
processes in power and chemical engineering
Measurement of the feed drive signals voltage,
and in manufacturing.
current, position and/or of the main drive signals
voltage current and speed make it possible to The results of the fault diagnosis are
estimate several parameters of the drive chain prerequisites for preventive maintenance, main-
and the metal cutting process. The detected tenance on request, reconfiguration in case of
faults are, e.g. beltdrive elasticity, friction of the serious malfunctions or just inspection of
spindle bearing, missing lubrication in the assembled products at the end of a manufactur-
ing line.
slideway, tool wear for milling and drilling (Rei8
(1991), Wanke and Reif~ (1991)).
Further successful fault detections via para-
Acknowledgements--The described results were obtained in
meter estimation were demonstrated for a various research projects funded by the Bundesminister fiir
sawing machine (Neumann (1991)) and hydraulic Forschung und Technologie (BMFT), the Deutsche Forsch-
power steering (Appel (1990); Appel and ungsgemeinschaft (DFG), and ESPRIT-PAQO Project 504.
We would like to thank these institutions for the financial
Isermann (1991)). support and cooperation.
834 R. ISERMANN

The author would like to thank Ms Xiaoshan He and Dr dynamic processes with application to an industrial robot.
lng. Bernd Freyermuth who performed the experiments with Int. J. Control. 55, 1287-1298.
the feed drive test rig. Isermann, R. and F. Freyermuth (1991a). Process fault
diagnosis based on process model knowledge. Part I:
principles. J. ~I Dyn. Syst. Measurement and ('ontrol, 113,
62(I-626.
REFERENCES
lsermann, R. and F. Freyermuth (1991b). Process fault
diagnosis based on process model knowledge. Part case
Appel, W. (1990). Parametersch~itzung an einer hydraulis- study experiments. J. of Dyn. Syst. Mea~s'urement and
then Servolenkung. 9th lASTED Int. Conf. Mod, ldentif. Control, 113, 627-633.
and Contr., 18-21 February 1990, lnnsbruck, Austria. lsermann, R., W. Appel, F. Freyermuth, A. Fuchs, W.
Appel, W. and R. Isermann (1991). Fault detection of a Janik, D. Neumann, Th. Reiss and P. Wanke (1990).
hydraulic power steering by parameter estimation. IFAC Model based fault diagnosis and supervision of machines
Symposium SA FEPROCESS, Baden-Baden, Germany. and drives, l lth IFAC Congress on Automatic Control,
Barlow, R. E. and F. Proschan (1975). Statistical Theory of Tallinn.
Reliability and Life Testing. Holt, Rinehart and Winston, Janik, W. and A. Fuchs (1991). Proce~- and signal-model
New York. based fault detection of the grinding process. IFAC
Box, G. E. P. and G. M. Jenkins (1970). Time Series Symposium SAFEPROCESS, Baden-Baden, Germany.
Analysis. Holden-Day, San Francisco. Lee, W. S., D. L. Grosh, F. A. Tillmann and C. H. Lie
Collacott, R. (1977). Mechanical Fault Diagnosis. Chapman (1985). Fault tree analysis, methods, and applications--a
and Hall, London. review. IEEE Trans. on Reliability, R-34, 194-202.
Erxleben, St. (1984). Untersuchungen zum Betriebsverhalten Milne, R. (1987). Strategies for diagnosis. IEEE Trans. on
yon Riemengetrieben unter Beriicksichtigung des elastis- Syst., Man, and Cvb. SMC-17, 333-339.
chen Materialverhaltens. Dissertation RWTH Aachen. Milne, R. (1991). Integration: the key to second generation
Eykhoff, P. (1974). System Identification. John Wiley, applications. IFA(" Symposium SAFEPROCESS, Baden-
London. Baden, Germany.
Filbert, D. and K. Metzger (1982). Quality test of systems by Neumann, D. (1991). Fault diagnosis of machine-tools by
parameter estimation. Proc. 9th IMEKO-Congress, Berlin, estimation of signal spectra. IFAC Symposium
Germany. SAFEPROCESS, Baden-Baden, Germany.
Fink, P. K. and J. C. Lusth (1987). Expert systems and Neumann, D., R. lsermann and S. Nold (1988). Comparison
diagnostic expertise in the mechanical and electrical of some parameter estimation methods for continuous-
domains. IEEE Trans. Syst., Man and Cyb., SMC-17. time models. 8th IFAC Symposium on Identification and
Frank, P. M. (1990). Fault diagnosis in dynamic systems System Parameter Estimation. Beijing, China.
using analytical and knowledge-ba~d redundancy. Nevins, J. L., D. E. Whitney and A. C. Edsall (1987).
Automatica, 2,6, 159-474. Intelligent systems in manufacturing, lOth IFAC World
Freyermuth, B. (1990). Modellgestiitzte Fehlerdiagnose yon Congress on Automatic Control, Munich.
Industrierobotern mittels Parametersch~tzung. Roboter- Nold, S. (1987). Fault detection in AC drives by process
systeme, 6, 202-210. Springer Verlag, Berlin. parameter estimation. IOth IFAC World Congress,
Freyermuth, B. (1991a). An approach to model based fault Munich.
diagnosis of industrial robots. Proc. IEEE Int. Conf. on Nold, S. and R. Isermann (1988). Model based fault
Robotics and Automation, Sacramento, CA. detection for centrifugal pumps and AC drives, l lth
Freyermuth, B. (1991b). Knowledge based incipient fault IMEKO World Congress, Houston, "IX.
diagnosis of industrial robots. IFA C Symposium Nold, S., R. lsermann and B. Freyermuth (1987). A
SAFEPROCESS, Baden-Baden, Germany. multilevel knowledge-based concept for the supervision of
Frost, R. A. (1986). Introduction to Knowledge Base technical processes. 5th IMEKO ,~vmp. on Techn. Diagr.,
Systems. Collins, London. Paderborn.
Geiger, G. (1985). Technische Fehlerdiagnose mitteis Patton, R. J. and J. Chen (1991). A review of parity
Parametersch~tzung und Fehlerklassifikation am Beispiel approaches to fault diagnosis. IFAC Symposium
einer elektriseh angetriebenen Kreiselpumpe. Dissertation SAFEPROCESS, Baden-Baden, Germany.
TH Darmstadt. VDI- Fortschrittsberichte, Reihe 3, Bd. Patton, R., P. Frank and R. Clark (Ed.) (1989). Fault
91. VDI Verlag, Diisseldorf. Diagnosi~ in Dynamic Systems. Prentice Hall, New York.
Gertler, J. (1988). Survey of model-based failure detection Pau, L. F. (1981). Failure Diagnosis and Performance
and isolation in complex plants. IEEE Control Systems Monitoring. Marcel Dekker, New York.
Magazine. Peter, K.-H. and R. lsermann (1989). Parameter-adaptive
Gertler, J. (1991). Analytical redundancy methods in fault PID-control based on continuous-time process models. 3rd
detection and isolation. IFAC Symposium, IFAC Symposium on Adaptive Control and Signal
SAFEPROCESS, Baden-Baden, Germany. Processing, Glasgow, U.K.
Goedecke, W. (1986). Fehlererkennung an einem thermis- Rault, A., D. Jaume and M. Verge (1984). Industrial process
chen ProzeB mit Methoden der Parametersch~itzung. fault detection and localization. Proc. 9th IFAC Congress,
Dissertation TH Darmstadt. VID-Fortschrittsberichte, Budapest.
Reihe 8, Bd. 1341. VDI-Verlag, Diisseldorf. Rault, A. (1989). Detection and diagnosis system and model
Himmelblau, D. M. (1978). Fault Detection and Diagnosis in based approach. 1FAC Symposium on Advanced Informa-
Chemical and Petrochemical Processes. Elsevier, tion Processing in Automatic Control (AIPAC), Nancy,
Amsterdam. France.
lsermann, R. (1984). Process fault detection based on Reig, Th. (1991). Model based fault diagnosis and
modelling and estimation methods----a survey. Automatica, supervision of the drilling process. IFAC Symposium
20, 387-404. SAFEPROCESS, Baden-Baden, Germany.
lsermann, R. (1988). Wissensbasierte Fehlerdiagnose tech- Schneider-Fresenius, W. (Ed.) (1985). Technische
nischer Prozesse. Automatisierungstechnik, 36. Jahrgang, Fehlerfriihdiagnose-einrichtungen. Oldenbourg, Miinchen.
Heft 11,421-426. Sturm. A. and R. F6rster (1988). Maschinen und
lsermann, R. (1989a). Process fault diagnosis based on anlagendiagnostik. VEB-Verlag Technik, Berlin.
process model knowledge. 1FAC Symposium on Advanced Tor&c,so, P. and L. Console (1989). Diagnostic Problem
Information Processing in Automatic Control. AIPAC, Solving. North Oxford Academic, U.K.
Nancy, France. Tou, J. T. and R. C. Gonzalez (1974). Pattern Recognition
Isermann, R. (1992a). Identifikation Dynamischer Systeme, Principles. Addison-Wesley Publishing Company.
2rid ed. Vois I & II. Springer-Verlag, Berlin. Trawinski, P. and R. lsermann (1989). Model based fault
Isermann, R. (1992b). Estimation of physical parameters for diagnosis of a machine toll feeddrive. 21th. Int. Symp. on
 Machine fault diagnosis 835

Automotive Technology and Automation ISA TA, Wanke, P. and Th. Rei6 (1991). Model based fault diagnosis
Wiesbaden, 8-10 November 1989. and supervision of the main and feed drives of a flexible
Tsafestas, Sp., M. Singh and G. Schmidt (Ed.) (1986). milling center. IFAC Symposium SAFEPROCE&~,
System Fault Diagnostics, Reliability and Related Baden-Baden, Germany.
Knowledge-based Approaches. Vols I & II. D. Reidel, Willsky, A. S. (1976). A survey of design methods for failure
Dordrecht. detection systems. Automatica, 12, 601-611.
Viswanadham, N. and R. Strichander (1987). Fault Young, P. C. (1981). Parameter estimation for continuous-
detection using unknown input observers. Control Theory time models---a survey. Automatica, 17, 23-29.
and Advanced Technology, 3, 91-101.

AUTO ~:4-C