ControlEng. Practice,Vol. 4, No. 10, pp.

1361-1369, 1996
Copyright © 1996 Elsevier Science Ltd
Pergamon Printed in Great Britain. All rights reserved
0967-0661/96 $15.00 + 0.00
PII:S0967-0661 (96)00146-3

F A U L T D E T E C T I O N BASED ON ADAPTIVE PARITY E Q U A T I O N S

AND S I N G L E - P A R A M E T E R TRACKING
T. Hiifling and R. Isermann
Institute of Automatic Control, Laboratory of Control Engineering and Process Automation, Darmstadt University of
Technology, Landgraf-Georg-Str. 4, D-64283 Darmstadt, Germany ([email protected])

(Received October 1995; in final form June 1996)

Abstract: Faults which appear in technical processes can often be described as additive or
multiplicative faults with respect to the process model. To perform fast detection of these
faults, continuous-time parity equations are used. Parameter deviations are estimated directly
from residuals, providing information about their size. The combined method enables one
to distinguish between additive and parametric faults. In case of time-variant processes with
slow parameter changes, the coefficients of the parity equations can also be adapted with
respect to the tracked parameters. The problem of persistent process excitation for estimation
is by-passed. The fault-detection scheme is demonstrated on a permanently excited d.c.motor
on a laboratory rig. Its properties, and the experimental results, are discussed.

Keywords: Fault detection, adaptive, parameter estimation, parity space, continuous-time,

d.c.motors.

1. INTRODUCTION Very often, process faults cause changes of the physi-

cal parameters. Continuous-time model descriptions
In the field of model-based fault detection and diag- are used, as their model parameters are directly
nosis, research efforts have progressed significantly in coupled to the physical parameters of the process.
the last twenty years (Willsky, 1976; Isermann, 1984, Then, depending on the complexity of the system
1994; Patton, et al., 1989, Frank, 1990; Gertler, 1991; considered, and the available measurements, the phy-
Patton, 1994; etc.). It has been shown that apparently sical parameters can often be estimated directly. At
different residual generators based on parity space or least the relationship between model parameters and
on observers are closely related. Parameter estimation physical parameters is simpler than in discrete-time
is a somewhat different approach, despite the fact that system~.
the relationship between parameter estimation and the
parity space has recently been worked out analytically In this paper the details of a new fault-detection
(Delmaire, et al., 1994, 1995; Gertler, 1995). Primar- scheme are deduced and described. It consists of
ily, it is designed to track parametric variations of the continuous-time parity equations for fast detection,
process. Therefore, the accessible input and output which is shown in Section 2. Then, the sensitivity of
signals must be processed nonlinearly. An important the residuals with respect to faults is investigated. In
problem is that the input signals must satisfy several Section 4 parameter deviations are estimated directly
conditions concerning process excitation with respect and recursively from residuals. A single-parameter
to the dynamics and the nonlinearities which are to be tracking algorithm is proposed. A procedure to distin-
estimated. guish between additive and multiplicative faults is
given in Section 5. The next section describes how

1361
1362 T. H6fling and R. Isermann

the coefficients of the parity equations can be adapted r_(t) : WY_(t) - W Q U ( t ) . (6)
in the case of slight changes of time-variant proces-
ses. An illustrative example with a d.c.motor finally Under the assumption that in the fault-free case the
demonstrates the capability of the procedure, with parameters do not change, the residual shows the fol-
measured data and experimentally installed faults. lowing properties:
• r(t)'= 0 fault-free case (noise effects, uncertainty)
° r(t) ~: O a fault has occurred.
2. CONTINUOUS-TIME PARITY SPACE
The parity equations (6) are linear equations in the
This section briefly describes the design of a linear observables Y(t) and U(t), including the parameters of
continuous-time parity space (H6fling, 1993, 1996; the matrices A, B and C, see eqs.(l) and (2)i Degrees
H6flingand Pfeufe~; 1994). It is preferred to a discrete of freedom in the design of matrix W can be used to
one in order to get direct and simple relationships generate a structured set of residuals according to
between the parameters of the mathematical model (Gertler and Singer, 1990). Then at least one
and the physical parameters, which are often directly measured quantity has no impact on a specific resid-
related to the faults. In what follows, this is exploited ual. In case of a faulty measurement the decoupled
to get information about the fault location, and its residual remains small, while the others increase their
size. value. This feature helps to locate the fault.

2.1 Generation o f continuous-time parity equations 2.2 State variable filtering

As continuous-time parity equations require time
Based on known ways of theoretically modelling the derivatives of sampled data to be evaluated, the use
structure of a linear mathematical model of order n in of a state variable filter is proposed (Young, 1981;
continuous time, a state-space representation is Peter and Isermann, 1989; Isermann, et al., 1992).
obtained. The idea is to transform the considered process model
,¢(t) = A~(t) + Bu(t) (1) with its measured data by introducing a set of ident-
ical linear filters operating on each of the time signals
y(t) : Cx_(t) (2) into a system which uses filtered input and output
signals instead. This method applied to a single input
with u e R q, x ~ R", 2 e R " and the matrices A, B single output system is shown in Fig. 1 (for multiple-
and C with appropriate dimensions. The continuous input multiple-output systems the appropriate number
time parity space (H6fling, 1993; H6fling and of state variable filters must be used):
Pfeufer, 1994) can be generated similar to the design
of a discrete one (Chow and Willsky, 1984; Patton
U
t process y>
and Chen, 1994; Frank, 1990; Gertler, 1991). Insert-
ing (1) in (2) and differentiating p times (p<n) yields
a set of equations (3) and (4), respectively. The
$
optimal p can be determined by an observability test
according to (Chow and Willsky, 1984).
svf I
model
rl cA2x'°+ CA.C.O O[I1 F F F ) Y F• Y F F

Y(t) = Tx_(t) + O U ( t ) (4) Fig. l SISO-System with state variable filters

with U ~ R u(p+~, Y ~ R "°'+~) and proper matrices T The advantage is that sampled data is low-pass fil-
and Q. Note that Y(t) and U(t) contain the time deri- tered and therefore high.frequencynoiseis suppressed.
vatives of the relevant signals. Equation (4) is multi- Moreover the derivatives of the filtered signals are
plied by a matrix W which is the nullspace of T. That available. To generate the p-th order derivative of a
means W consists of linearly independent row vectors filter output at least a p+l-th order state variable filter
w_4r, and satisfies the condition: shoulh be applied (Fig. 2).
WT = 0. (5)
A pure low-pass filter satisfies the required perform-
With this constraint, eq.(4) becomes independent of ance:
the (unknown) state vector x(t). The parity relations ZF(S) =
fo (7)
r(t) are obtained by: 6w(s) : z(s) fo÷f~s+...<_ls"-l+s"
 Fault Detection Based on Adaptive Parity Equations 1363

For instance the parameters f ~ . . . fo_~are designed as low-pass filter can be applied in order to smooth the
a Butterworth-filter with an adequate cutoff frequency adaptwe threshold.
o~. o~ is chosen so that it is close to the highest pro-
laa = - - i ( - TOs
- u +const. ) (8)
cess eigenfrequency which guarantees that the import-
ant information of the signal is kept and higher fre-
l+T2s l + Tls
quencies according to disturbances are well sup- The complete structure contains four parameters
pressed. which must be determined: A constant factor 4/3 (or
3/2) times the maximum residual value in the fault
The designed state variable filter works analogue, but free case with constant input signals is used. Time
it can be replaced by a digital filter. Let the system constants T/ and T2 are chosen according to the
(AF, BF, CF) be the controllable canonical form of the slowest process eigenvalue or the cut-off frequency of
transfer function G~vt(S), then the state vector contains the state variable filter, respectively. The ratio To/T/
the derivatives of the filter output. Using the z-trans- i s a measure for the uncertainty of the dynamics.
form and the sampling-interval T~ the filter is trans-
ferred into a discrete representation. To provide better
estimation of the derivatives a polynomial approxima-
tion for the continuous-time filter input signal is
suggested (Peter and Isermann. 1989).

Fig.4. Example of a residual with an adaptive thres-

Fig.2. State variable filtering hold

2.3 Adaptive Thresholds Fig.4 illustrates the performance of an adaptive thres-

hold in the fault-free case (t<O.2s) and after the fault
Due to apparent model uncertainties residuals always occurrence (t>O.2s). The adaptive threshold decreases
deviate from zero, dependent on the input signals. the delay of detection and the susceptibility to false
alarms.
const.
Using the z-transform the computation of the thresh-
1 + Tzs[ 1" old can be discretized, which reduces computational
costs in on-line applications, without a loss of per-
process input adaptive threshold
formance.
Fig.3. Generation of an adaptive threshold I~d r0

The use of an adaptive threshold (Clark, 1989; Frank, l a d - z I - e re2 r2

( T D z -TI- - lzu- +
TO e c o rlTO
nSt']
1991), instead of a fixed one, can improve the per-
formance of a fault-detection scheme significantly
with respect to delay of detection and false alarm
rate.
3. SENSITIVITY OF PARITY EQUATIONS
If the model uncertainty is mainly contained in the
dynamics an adaptive threshold driven by the process In this section the sensitivity of a parity equation with
input signal can be used, Fig.3. The threshold consists respect to an additive or a parametric fault is investi-
of a filter with lead-lag behaviour which is driven by gated analytically. Similar work has also been carried
the input signal. This causes the filter output to be out in (Gertler and Kunwer, 1993) but with the objec-
zero for steady-state inputs: otherwise it is a measure tive of decoupling from multiplicative model uncer-
for the dynamic input excitation. In order to rectify tainties. Here, however, the main goal is to show the
this signal the absolute value is computed. If more different effects of the faults. Based on this an
input signals are involved for each input the highpass- advanced fault-detection strategy is proposed below.
filtering and the rectification has to be performed. A
constant value is added due to the effects of measure- Rearranging eq.(6) for a single residual r(t) leads to
ment noise on the residual. Optionally a first-order
1364 T. H0fling and R. lsermann

That means that whether the same parametric fault

r(t) = [_w ~'- _w~" O] [u(o = "O~)'*(O (lO) exceeds the threshold or stays below it depends on
the instantaneous process state (set point, load,
dynamic excitation, etc.). This must be kept in mind
with vector o r ( p ) denoting the model information when considering these faults using parity equations.
with respect to the physical parameters £, and ~(t)
containing the available signals (inputs and outputs). The cNculation deduced above, however, is only valid
If a fault occurs it can mostly be expressed as an if the considered residual r(t) is linear in the access-
additive fault A ~ or a parametric fault A/Z. ible signals ~(t). Otherwise, more-complicated ex-
pressions are obtained. In a more general notation,
1. An additive fault A ~ ~ __0(AE = ~ has the tollow- including a class of nonlinearites, r(t) can be written
ing effect: as

r(t) =-Or(/2) "~yau;t(t) = f9 r(12) "[31/-(t) + A ~ ] = (9 r(/2) . A , r(t) = ,or(12,~(t))'~(t). (16)

(it) Then the change of the residual r(t), due to a fault, is
as o r ( p ) ~ ( t ) = 0 according to eq.(10) describes the
fault-free case. r(t) = [0*r(t)
' o ( / 2 ' ~ -O
) ~,/_ + "or(12'*)) A-0- -(17)

2. A parametric fault shows up as Ap ~ O (A~ = 0_). - ,~(t) a,o(e.,) 612.

These faults shift the underlying differential equation 012
to a new balance:
0 = -oT(2~faul)'~(t ) = "OT(p+A12)'..IJ[(t) . (12) Hence, in the case of nonlinear model components,
J~ the effects of additive faults on the residual also
Using a Taylor-series approximation, where OO(p_)/Op__ depend on the present process state.
represents the Jacobian matrix with respect to the
physical parameters, yields: Note that these considerations are only approxima-
tions in the vicinity of the nominal parameter set.
0 ='O r(12 + A12) "~ (t) =-O r(12) "_~(t) + O(,Or(12) ",(t)) .6/2 +...
012
(13) 4. ESTIMATION OF A SINGLE P A R A M E T E R
Ignoring terms of higher than first order and using USING PARITY RESIDUALS
residual eq.(7) results in
Assume that a parametric fault has appeared. Then
r = - _0_(t)'--o'or(12) .A12 . (14) the residuals are deflected according to eq. (15). In
a12 cases of a small parameter change the relevant thresh-
Both effects can be put in one equation which charac- old of the residual will not necessarily be exceeded,
terizes the residual sensitivity: depending on the present process state.

_ O_Lrzx, _ Oral2 That problem can be solved by a parameter-tracking

0,_ 012 algorithm, whose sensitivity is independent of the
signal magnitudes. Because residuals, as derived in
o(,oR~) . , ) a(o~q2) . , ) (15)
A, A12 Section 2, can also be used as estimation equations, a
o, 012 parameter deviation Ap; from its nominal value can be
0r(12) a , - ~z r o,o~) /~12 estimated directly from a residual deviation (here a
012 single residual r(t) is considered).

The result obtained can be interpreted as follows: a [,O r(~) .,(t) ]

r(t) = - A pj =
apj (18)
1. In the case of an additive fault O__T(p_.)is constant
and the residual change is only __Or(p) A~. Under the r O,O~)
= - , (t)~Apj = - K(t) A pj
assumption that the fault acts as a step function the ~vj
minimum fault size required to exceed the related
threshold, which is mainly determined due to noise Note that at each time instant K(t) is a scalar value. It
and model uncertainty, can be calculated. comprises measured values and known model infor-
mation (pseudo-measurement). Using a time window
2. Parametric (or multiplicative) faults behave differ- of accumulated data (sampling interval To)
ently. The residual deviation is not only settled by the K(t) = [K(t) ... K(t-kTo) ... K(t-UTo)] r (19)
fixed elements of the Jacobian matrix and the fault
size Ap_, but also by the current signal values ~(t).
 Fault Detection Based on Adaptive Parity Equations 1365

r eft) = [r(t)... rff-kTo).., r(t-NTo)] r (20) be considered:

Ap~ can be estimated with a simple least-squares 1. The estimated parameter and the one affected by
approach. As K(t) and ~(t) are vectors, the following the fault are the same. Then the parameter tracking
equation is obtained: algorithm eq.(22) yields the correct value of the
Kr(t) r (t) shifted parameter.
A13j = - [Kr(t) K(t)]-lKr(t) r ( t ) = Kr(t) K(t) -
2. The estimated parameter Apt is not really affected
N by the fault, but the tracking algorithm tries to find a
y~ X(t-kTo) r(t-kTo) correlation between the deviated residual r(t) and the
k=0
(pseudo-) measured signal K(t) which is mapped on
N
Apt. In this case the estimation does not converge. It
E r (t-tro) can vary and increase its standard deviation, but there
k=O
(21) is probably no change in the mean. Then the estimate
can only be taken as a fault indication, like the resid-
An on-line solution to recursively track a parameter ual itself. Another parameter must have changed, or
using a forgetting factor ~,<1 (e.g. ~,=0.99) is also an additive fault must have occurred, which caused
possible. the residual deflection due to a fault-exciting input
sequence, because otherwise the residual would not
Sxr(t) ~, SKr(t- To) - K(t) r(t) (22) have changed. Then a standard parameter estimation
can be launched, taking into account the input signal
@j(O - sx o - x s, x(t-ro) + r2(t)
and all estimatable parameters. In summary, this
The computational burden of this is two additions, approach uses parity equations to estimate a specific
four multiplications (only two when no forgetting parameter by a parameter-estimation method.
factor is used) and one division.

As mentioned earlier, parameter estimation in 5. DISTINCTION B E T W E E N A D D I T I V E A N D

dynamic and nonlinear systems has special require- P A R A M E T R I C FAULTS
ments concerning the process input sequences. To
estimate the dynamic parameters, i.e. the time con- Very often a fault is diagnosed as an additive or
stants of the process, the input sequence must be rich parametric one according to the used fault detection
enough to persistently excite the process in all its methods. When a residual causes a threshold viol-
modes. For estimation of a nonlinearity, a sequence is ation, the user tends to interpret the underlying fault
required whose magnitude covers at best the complete as an additive fault. Otherwise, when parameter esti-
range of the nonlinearity. If these conditions are not mation is used and parametric changes occur, it is
satified the parameter estimation has to be stopped, in usually considered as a parametric fault. Even when
order to prevent divergent parameters. both methods are applied, certain closely related but
physically completely different faults cannot be
This, however, can be by-passed when a residual is distinguished (e.g., parameter p~ with shift and sensor
used which compensates for the dynamics and non- y~ with gain fault). Based on the investigations in
linearities, because the residual can be evaluated inde- Sections 3 and 4 and using structured sets of resid-
pendently of the current input excitation of the pro- uals, a reliable distinction between additive and para-
cess. It only has to be guaranteed that the denomina- metric faults becomes possible.
tor of eq.(21) is now zero. This is always true when
the parameter Pi chosen to be tracked concerns the The procedure is as follows: After the fault occur-
linear process gain in the transfer function. Additional- rence the residual pattern indicates a subset of poss-
ly the process must not be out of operation, which is ible faults. Usually it gives no indication whether the
represented by all measured signals being zero. This fault i; caused by an addive or a parametric fault.
is true when the estimated parameter deviation Apj This ~s also true for complete parameter estimation.
concerns the linear process gain and the process is
not out of operation (which is represented by all Now assume the task is to decide whether parameter
measured signals being zero). Being constantly in a P/ or sensor y~ is faulty. Moreover there are two
steady state makes K(t) non-zero and therefore does residuals e.g. rl and r4 with the following properties.
not lead to trouble with the evaluation of eq.(21). rl=f(plyl) is a function of Pl and Y/ while
This means the process just has to operate to fulfill r4=f(pl)4f(yl) is a function of pj but not of Yr The
the presumptions for the proposed fault-detection distinction can be achieved by estimating the possible
scheme. parametric change Apt from both residuals r / and r4.

In the case of a fault the following scenarios have to If parameter p~ has changed, both estimates yield the
1366 T. HOfling and R. lsermann

same value (Ap/rl)=Apl(r4). If the sensor Yl is meters drift in a known way, the tracked parameter
faulty only the estimated parameter change from r~ change Apj, can be taken as pointer for a look-up
delivers a non-zero value while the estimate from r4 table (in the case of temperature drift it is used as an
remains zero (Apl(rl)~Apl(r4). This gives a clear indi- indirect temperature measurement), Fig. 6. The look-
cation which of two suspected faults is really present up table comprises a priori information about the
in the process (see also the example given in Section complete parameter set, depending o n the current
7.3). value of parameter P i" This, however, requires many
experiments in order to find the parameters of the
table.
6. A D A P T I V E P A R I T Y EQUATIONS
After adaptation it is recommended to reset the para-
In the case of a slow time-variant process the tracking meter-tracking algorithm in order to avoid stability
algorithm described above can be used to perform problems caused by feedback. Using this technique
fault detection according to Fig. 5. Here, for simplic- the performance of a fault-detection scheme can be
ity, only one parity equation and one tracked parame- considerably enhanced.
ter are considered. look-up table:
tracked parameter Pj P, P2
L# faults/-¢ pj = 1 , 5 0

\ 1,30 0,69 0,33 ...

1,40 0,68 0,33 ...

1,60 0,66 0,33 ...

Fig.6. Look-up table for parameter adaptation

7. APPLICATION: D.C. M O T O R

7.1 Grey-box model of the D.C.motor

The process being considered is a permanently

excited d.c.motor (rated power 550W). It has two
inputs: the armature voltage UA and the load torque
M L. Both state variables, the armature current 1A and
the speed o3. are measurable. A completely linear
model can be retrieved from the literature, which is
good for controller design. However many experi-
ments have shown that this model with constant para-
meters does not match the process in the whole oper-
ational range. Therefore, two nonlinearities are
Fig.5. Adaptive parity equations
included so that the model fits the process better.
This is called "grey-box modelling", and is described
If the parity equation exceeds its given threshold, a in (Bohlin, 1994). The resulting first-order differential
fault is detected and an alarm is released. The tracked equations are:
parameter is assumed to be the only time-variant
parameter (e.g., due to temperature drift) or at least LA/A(t) = -RAIA(t)-V co(t)-K,]co(t)]ZA(t)+U~(t~
the one with the main impact on the residual. There-
fore, it is necessary to adapt the coefficients of the
J 6o(t) =qt I A (t) -MF1 co(t) -MFosign(co (t)) - M t (0!24)
parity equations in intervals with respect to the
tracked parameter. But this is only done in case of
small deviations within a bounded range; otherwise, The term K R ]o3(t) l IA(t) compensates for the voltage
an alarm is released. drop at the brushes in combination with a pulse-
width-modulated power supply. Also the speed-fric-
Two mechanisms of adaptation are possible: tion-curve has been measured. For simple parametri-
I. Only the tracked parameter change Api is used to zation two components, viscous and dry friction, in
change the coefficients of the residual. terms of MHO3 and MF0sign(o3), respectively, have
been included in the differential equation. The para-
2. If the time-variance of the process is due to un- meters are identified by least-squares estimation into
modelled influence like temperature drift and all para- continuous time domain. Their nominal values are:
 Fault Detection Based on Adaptive Parity Equations 1367

armature resistance Ra = 1.52 f~ rl(t ) = Lala(t ) + RaIa(t ) + t P o ( t ) - Ua(t )

armature inductance La = 6.82"10 .3 ~ s
magnetic flux q~ = 0.33 V s r2(t ) = Jd)(t) - WIA(t ) + MFlO(t ) + ML(t )
voltage drop factor Kn = 2.21"10 -3 V s / m
inertia constant J = 1.92"10 .3 kg m 2 r3(t ) : JLala(t ) + (LAMFI +JRA)IA(t) +
viscous friction Mpj = 0.36"10 -3 N m s + (1,t t2 +RAMF1)IA(t ) - (26)
dry friction Mr0 = 0.11 N m
JOa(t ) - Me1 U a ( t ) - ~ M L ( t )
Most o f them (Ra, ~ , K B, MFI, Mro) influence the pro-
r4(t ) = JLa(b(t ) + (LAMFI+JRA)(O(t) +
cess gain, and the other two (L A, J) the time con-
stants. The data is measured with a sampling fre- + (W2+RamF1)6o(t) -
quency o f 5 kHz, and state variable filtered by a
fourth-order lowpass-filter with Butterworth character- tY UA(t ) + LAfflt(t ) + RAML(t )
istic and a cut-off frequency of 250 Hz.
Then, if an additive fault occurs, all residuals except
the decoupled one are deflected. This supports the
isolation of actuator and sensor faults in particular.
The scheme of the structured residuals is not touched
by the compensation for the nonlinear voltage drop of
the brushes, as its magnitude is small enough to be
ignored. Two parameters Ra and Met, however,
depend on the present temperature. The behaviour o f
Ra and its effect on residual r / is depicted in Fig. 5.
armature resistance R, Ill] mean and range of deviation of residual r,
(temperature dependence) (fault free, parameters for warm motor)
I l I" 1.55
cold warl/i
2 iiiolOf ~ IllOlOf
1.5 I 0 mill
Fig.7. Block diagram o f the considered d.c.motor
1.45 [ 130mi" 90mia
!111111, - , . m lm

7. 2 Parity equations and residual sensitivities

1.4

1.35 cold wltrtll

-'t | = °
"I ~ [ ] l
....
I I ,2Om.,
I~

1.3 -2" --
5(1 I(X) 151) 0 3tl NI 9tl 12(I
For the detection and isolation of sensor (output) and time Imint time I minl
actuator (input) faults a set of structured parity equa-
Fig.8. Temperature effects on resistance R a and resi-
tions is appropriate (Gertler and Singer, 1990).
dual r I
As the differential equations (20) and (21) are nonlin- Therefore, the use of adaptive parity equations (as
ear the design procedure for a linear parity space proposed in Section 5) is recommended to improve
cannot be applied directly, but defining UA*-KB1 A Ic01 the residual performance.
as voltage input UA and Mr:ign(o~) as load input M L
leads to a linear description. An observability test According to the considerations of Section 3, the
reveals that both outputs (I a and co) can also observe residuals are examined concerning their sensitivity to
each other. This is a precondition for a parity space additive and parametric faults. As r t and r 2 comprise
of full order (here: 2). Then W is chosen such that a all parameters and all signals, it is sufficient to con-
set o f structured residuals is obtained, where residual sider only these two, although rj or r 4 can also be
r / t ) is independent o f ML(t), r2(t ) o f Ua(t), r3(t) of taken. From eq.(15) the following is obtained:
tO(t) and r4(t) o f IA(t) (see also Hrfling and Pfeufer, rl(t ) -- - ALa]a(t ) - ARala(t ) - Aqt ~ ( t )
1994):
LA 0 0
+ LA AiA(t) + R A AIa(t ) + V A co(t) - A UA(t )

0 J 0 r2(t ) = - AJ~b(t) + A ~ I a ( t ) - AMF1 ~ ( t )

W: -ty Me1 (25)
0 o + J A d e ( t ) - ~ A/aft) + MF1 Ard(t) + AML(t )
(27)
. o 13 o J/..
Then, it can easily be seen that in the presence o f
with O~ : v/Z2+RAMF1 residual noise, e.g. of r t with a magnitude o f about 1
= LAMFI+JR a IV] and an armature current of some 3 [A], a resis-
tance change must be at least 0.3 [~2] in order to
The residuals are computed as: deflect the residual significantly. Therefore, the two
linear parameters R a and Met are selected to be
tracked according to eq.(22). X is chosen to be 0.99.
1368 T. H6fling and R. Isermann

mined (here, either A R a or AW) and no statements are

mR.- SI"rl AMF1- S°r2 (28)
possible about other parameters (here ALA). That
SI"I. So~ means that the fault detection is still working, but the
support for a diagnosis is poor. Another interesting
7.3 E x p e r i m e n t a l results fault to examine is an increase in the friction of about
0.2Nm, Fig. 11. Residual r~ shows no reaction, as
Each residual is provided with an adaptive threshold expected. But residual r 2 changes only slightly and
according to Section 2.3. In the following figures hardly violates its threshold. The parameter estimate
each residual is divided by its threshold, so an excess of AMrs, however, significantly indicates the fault.
of 1 or -1, respectively, indicates a threshold viol- This proves the feasibility of the parameter tracking
ation. Faults are frequently caused by the brushes. As as a supplement to fault detection.
the number of pole pairs is two, there are four lesidual I~ estimated parameter deviation AM

brushes and one o f them was lifted off so that there Ill JlNms

was no longer any contact to the commutator. This

fault is a typical parametric fault, with expected
'vl,r, i

effects on the resistance R a and the inductance L a. It

appears suddenly at time t = 0.5s and residual rj, r 3
/i,u
and r 4 display the fault immediately as does the
05 1 L5 05 I 1.5
tracked parameter ARA, Fig. 9. ume Isl rune Isl
normalized residual r, tracked parameter deviation AR^ Fig. l l: Residual r: and parameter AMvs (increased
IQI
friction)

o.: The last experiment demonstrates the capability to

residual divided by
ilS adaptive ~mesh(i
ur'l
i111 I I I i

_it.:
distinguish between additive and multiplicative faults
according to Section 5. Two faults, a parametric fault
(ARA/R A = + 5 4 % ) and a sensor gain fault ( A I A / I a = -
(i.5 11,5
time [s] time Is] 3 5 % ) have been experimentally applied. Their effects
Fig.9. Residual rs and tracked parameter /~ke a (brush on the residuals are almost the same, especially on r/.
fault) Also the parameter estimate of A R A estimated from r s
The rough curve o f ARa also indicates that an addi- is (must be) very similar, 12.
tional parameter must have changed. As the process
dynamics are excited by the input, the second para- residual r, from r, estimated AR. [t~]
1.5
meter could be ALa or AW. Performing the complete
ARA-fault
estimation with
rl(t ) = - A L a J , ( t ) - A R , I , ( t ) - A ~ ( t ) (29) k,~'v,,/'~ o.5
0
leads to smooth estimates of ARA = 1.4 f2, A L A = ~I,-fa
5 . 1 0 .3 £2s and Au/ = 0 Vs. The curves of the 0,2 0.4
-0.5
0.2 0.4
recursively estimated parameters in Fig. 10 show that time [s] time [s]

in the fault-free case the parameters do not deviate, Fig. 12. Residual r~ and parameter estimate ARA for
while their deviations can be estimated convergently both faults
when a parametric fault has occurred. This is the To distinguish between the two faults, r4 (which is
expected and correct behaviour. independent of IA: ~ r g g l a = 0) has to be considered
measured signals parameter deviations
I(X)
with respect to parameter RA:
A r4(t)
4 Or4(t) - J d ) ( t ) + M v l ~ ( t ) + ML(t) ~ - - - (30)
5O
oR. AR.
2
0 Then a change of R a can be estimated directly from
0 ] " A~I' IVsl
residual r~:
-50 +2 I
0.3 0.5 0.7 0,3 0.5
time Isl
0>7 aq(n ARa = Ar4(t) . (31)
time Isl 0R.
Fig. 10. Measurements and parameter deviations
(brush fault) Fig. 13 shows that the distinction between the two
Considering the same fault in the case of constant faults becomes very simple.
input signals the residual rs cannot be deflected by
ALA, because the time derivative of the current IA is Some more faults (concerning bearings, thermal
zero. This also means that ALa cannot be estimated. stress, sensors, etc.) were installed on the d.c.motor.
With constant signals, only one linear parameter con- These experiments confirmed the results described
cerning the static gain of the process can be deter- above.
 Fault Detection Based on Adaptive Parity Equations 1369

~idtud ro ~om r, esamated 0 . . [D] Frank, P.M. (1990). Fault diagnosis in dynamic sys-
1.5
tems using analytical and knowledge-based re-
dundancy, Automatica, 26, 459-474.
0.5
Frank, P.M. (1991). Enhancement of Robustness in
observer-based fault detection, IFAC-Symposium
-2
0 0.2 0.4
-0.5
0.2 0,4
SAFEPROCESS "91, Baden-Baden, I/275-287.
tim* Ill time lsl Gertler, J. and D. Singer (1990). A new structural
Fig.13. Residual r~ and parameter estimate ARA for framework for parity equation-based failure de-
both faults tection and isolation, Automatica, 26, 381-388.
Gertler, J. (1991). Analytical redundancy methods in
8. CONCLUSION failure detection and isolation, IFAC-Symposium
SAFEPROCESS '91, Baden-Baden, I/9-21.
Parity equations and single tracked parameters per- Gertler, J. and M. Kunwer (1993). Optimal residual
form fast fault detection together with low computa- decoupling for robust fault diagnosis, Interna-
tional on-line effort. The problem of process tional Journal of Control, Vol. 61, No. 2, 395-
excitation for parameter estimation is by-passed by 421.
estimating the deviation of only one linear parameter Gertler, J. (1995). Diagnosing parametric faults: From
and clearing the estimation equation from the parameter estimation to parity relations, Ameri-
dynamic and nonlinear components using a parity can Control Conference, Seattle.
residual. Then no supervisory level for the parameter Isermann, R. (1984). Process fault detection based on
tracking is required, and the sensitivity of the fault- modelling and estimation methods: A survey,
detection scheme to parametric faults is no longer Automatica, 20, 387-404.
dependent on the current process state. Slow time- Isermann, R., K.-H. Lachmann and D. Matko (1992).
variant processes can be handled by proper adaptation Adaptive Control Systems, Prentice Hall (Series
of the parity equations. The experimental results, in Systems and Control Engineering), London.
achieved with a d.c.motor, have demonstrated the Isermann, R. (1994). Integration of fault detection and
capability of the proposed procedure. After fault diagnosis methods, IFAC-Symposium
detection a second stage must be implemented to gain SAFEPROCESS '94, Espoo, 11/597-612.
complete model-based information, available for H6fling, Th. (1993). Detection of parameter variations
diagnosis. Then the process is usually driven with a by continuous-time parity equations, 12th IFAC-
specially designed test cycle (i.e., rich enough to World-Congress, Sydney, VII/511-516.
excite all the modes and the nonlinearities of the H6fling, Th. (1996). Methoden zur Fehlererkennung
process) which allows all the model-based symptoms mit Parametersch~tzung und Parit~itsgleichungen,
to be computed on a comparable basis. Finally, fault VDI-Fortschrittsberichte, Reihe 8, Nr. 546, VDI-
diagnosis can be performed by establishing simple Verlag, Diisseldorf, in German.
fuzzy rules, due to the distinct information delivered H6fling, Th. and Th. Pfeufer (1994). Detection of
by parity residuals and parameter estimates. additive and multiplicative faults, IFAC-Sympo-
sium SAFEPROCESS '94, Espoo, 11/539-544.
REFERENCES Patton, R.J., P.M. Frank and R.N. Clark (eds) (1989).
Fault diagnosis in dynamic systems - Theory and
Bohlin, T. (1994). A case study of grey box identifi- applications, Prentice Hall (Control Engineering
cation, Automatica, 30, 307-318. Series).
Chow, E.Y.and A.S. Willsky (1984). Analytical Patton, R.J. (1994). Robust model-based fault diagno-
redundancy and the design of robust detection sis: The state of the art, IFAC-Symposium
systems, IEEE Trans. on Automatic Control, AC- SAFEPROCESS '94, Espoo, 1/1-24.
29, (7), 603-614. Patton, R.J. and J. Chen (1995). Review of parity
Clark, R.N. (1989). State estimation schemes for space approaches to fault diagnosis for areospace
instrument fault detection, in Patton, et a/.(1989), systems, AIAA Journal of Guidance, Control and
21-45. Dynamics, 17, (2), 278-286.
Delmaire, G., J.-P. Cassar, M. Staroswiecki (1994). Peter, K. and R. Isermann (1989). Parameter-adaptive
Comparison of identification and parity space PID-control based on continuous-time process
approaches for failure detection in single input models, IFAC Symposium on Adaptive Control
single output systems, Conference Control Appli- and Signal Processing, Glasgow.
cations, Glasgow. Willsky, A.S. (1976). A survey of design methods for
Delmaire, G., J.-P. Cassar, M. Staroswiecki (1995). failure detection in dynamic systems, Automatica,
Comparison of generalized least square identifica- 12, 601-611.
tion and parity space techniques for FDI purpose Young. P.C. (1981). Parameter estimation for con-
in SISO-systems, 3rd European Control Confer- tinuous time models - a survey. Automatica 17,
ence, Rome. 23 -39.