Parity Equations for Fault Detection and Isolation

I. Fagarasan, S.St. Iliescu,

Automatic Control and Industrial Informatics Department,
Automatic Control and Computers Faculty,
University POLITEHNICA of Bucharest,
{ioana, iliescu}@shiva.pub.ro

Abstract- Sensors, actuators or physical components in engi- fault fault fault

neering systems are often subjected to un-permitted or unex- u y

Actuators Process Sensors
pected deviations from usual operating conditions. With a rela-
tively small computing effort and with good results to un- Process model
measured disturbances, the parity equations build for a specific uˆ, yˆ , xˆ faulty
behavior
process are proven to be very useful to provide fault detection Residual
generator
residuals in order to identify the fault.
r (residuals)

normal Test the thresholds s Unified symptoms s Fault f

I. INTRODUCTION behavior overpassing (symptoms) representation identification (fault)
Fault detection Fault isolation
The model-based fault detection methods rely on the concept
of analytical redundancy. In contrast to physical redundancy, Figure 1. FDI model-based general scheme
when measurements from parallel sensors are compared to
One modern solution to generates residuals is the parity
each other, now sensor's measurements are compared to ana-
equations approach. This approach is based on the deviations
lytically computed values of the respective variable. Such
of estimated process states or outputs from measured ones. The
computation use present and/or previous measurements of
parity equations can be derived from the mathematical model
other variables, and the mathematical plant model describing
equation. There exist many approaches for the design of parity
their nominal relationship to measured variable. The idea can
equations. These are based, for example, on discrete state-
be extended to the comparison of two analytically generated
space models [1], time continuous state-space models [7], ob-
quantities, obtained from different sets of variables. In either
servers [4] or transfer function models [5,9].
case, the resulting differences, called residuals, are indicative
of the presence of faults in the system. Here, the estimations are made using a transfer function
In order to avoid the possible loss of the systems perform- model of the nominal process behaviour. The model runs in
ance because of a fault appearance, many research efforts in the parallel to the process, leading to symptoms with different sen-
field of process supervision, fault detection and diagnosis have sitivities to a specific fault.
been made [6,10,11].
Parity equations are a simple and straightforward approach
The generation of residuals needs to be followed by residual
to build residuals that indicate faulty systems. This model-
evaluation, in order to arrive at the detection and isolation deci-
based fault detection method runs a HM model in parallel to the
sion. Because of the presence of noise and model errors, the
process HP itself leading to an output error (Figure 2a):
residuals are never zero, even if there is no fault. Therefore, the
detection decision requires testing the residuals against thresh- ry ( s ) = [H P ( s ) − H M ( s )]u ( s ) (1)
old, obtained empirically or by theoretical considerations.
The objective of fault detection and isolation (FDI) is not Anyway, the process parameters must be known a-priori. In
only to determine some fault presence in the system (fault de- a case of equity between HP(s) and HM(s), the output error for
tection) but also to establish the kind and location of fault (fault additive faults becomes:
isolation).
II. FDI STRATEGY
ry ( s ) = H P ( s ) f u ( s ) + f y ( s ) (2)

A general FDI scheme for a process consists of two levels Another possibility is to generate a polynomial error (Figure
(Figure 1), a symptom generation part and a diagnostic part. In 2b):
the first level, symptoms are generated based on residuals and
these features indicate the state of the process. In the second rx ( s) = AM ( s ) y ( s ) − BM ( s )u ( s )
one, the relations between symptoms and faults are established. HP =HM (3)
To detect and isolate a fault is important to find the significant = BP ( s ) f u ( s ) + AP ( s ) f y ( s )
symptoms, which are robust against noises or disturbances.
Both situations (Figure 2) lead to different time-responses
for additive faults, ry(s) corresponds to output error for parame-
ter estimation and rx(s) to state-equation error [9].

1-4244-2577-8/08/$20.00 ©2008 IEEE

 that each residual responds to a different subset of faults and is
insensitive to the others.
fu fy fu fy
u u y
y
HP(s) HP (s)
III. FAULT AND PROCESS MODELLING
+

- ry A fault is defined as an un-permitted deviation from a normal

- +
HM(s) BM(s) AM(s) behaviour for at least one variable/characteristic of a process.
rx
a) b)
Therefore, a fault could lead to a malfunction or breakdown of
the process. The faults could be classified [10] above their
Figure 2. Parity equation for fault detection: shape (random or not), above their time behaviour (permanent,
a) Output-error method; b) State equation -error method abrupt, intermittent, noise, step or ramp fault) or above the
geographic dimension of the fault (local or global).
Equations (2) and (3) generate residuals and are so called The time behaviour of the fault could be depicted in figure 3.
parity equations. SISO (Single Input Single Output) processes
with intermediary measurements and MIMO (Multi Input
f f
Multi Output) ones offer better conditions to design parity b c f
d
f

a
equations. A residual generator is used for this purpose. PROCESS
Characteristic
f
The residual generator is a linear algorithm acting on the ob- Time Time Time Time
servable variables. Its generic form is:
Figure 3. Time behaviour of the fault: a) permanent or step type; b) positive
ramp; c) non-permanent; d) pulse; e) noise.
r ( s ) = V ( s )u ( s ) + W ( s ) y ( s ) (4)

Taking into account the process modelling, the faults are

where r(s) is the vector of the residuals and V(s), W(s) are
classified according to figure 4. The additive faults modify the
the transfer function matrices.
y variable by adding the influence of f fault and the multiplica-
Considering the real process described by the transfer func-
tive fault acts like a multiplicative factor f for an input variable
tion matrices HP(s), the relation (4) becomes:
u. An additive fault example is sensors offset and a multiplica-
tive one is internal parameters changing of the process.
V ( s )u ( s ) + W ( s ) H P ( s )u ( s ) = 0 (5)

This relation must be satisfied for all u(s) and therefore:

V ( s ) = −W ( s ) H P ( s )
The residual generator could be written as follows:
Figure 4. Fault type: a) additive; b) multiplicative

r ( s ) = W ( s )[ y ( s ) − H P ( s )u ( s )] (6)
A MIMO model was choose in order to offer proper condi-
Ideally, the residuals should remain unaffected by noise, dis- tions to design parity equations. An example of a heat ex-
turbances or model errors. However, these factors lead to make changer model, a MIMO system with m inputs and r outputs is
residuals non-zero, and thus interfere with the fault detection. presented in [8]. For each output a linear local model could be
Therefore the residual generator needs to be designed to have considered taking into account all process input:
minor deflections at these factors. Many efforts in designing
residual generators are going into achieving sufficiently robust BP ( s )
y( s) = u (s ) ;
residual performance. AP ( s )
Another important characteristic of the residual generator is
fault sensitivity of the residuals. This characteristic is evaluated y ( s ) = [ y1 ( s ) ... yr ( s )]T (7)
in accordance with the triggering limit, the value of a particular u ( s ) = [u1 ( s ) ... um ( s )]T
fault, which brings a particular residual to its threshold, consid-
ering that no other faults are present.
The FDI scheme was tested on a heat exchanger case by pre-
In addition to having robust detection properties, the residual
senting several different process and sensor faults (TABLE I),
generator needs to be designed to support the isolation of
using the model library (figure 5).
faults. Isolation always requires a set of residuals. To facilitate
The considered system inputs are: water input temperature
fault isolation, the residual set needs to have distinctive proper-
ties, unique characteristics of particular faults. Residuals sets θLi, water speed wL (or water flow ML), air-wall thermal flow
designed with this objective in mind are referred as structured qWG (or air flow MA) and air input temperature θAi . The system
or directional residuals. Structured residuals are so designed outputs considered of major interest are the water output tem-
perature θLe and secondary, the air output temperature θAe.
  1  0   H1 ( s ) ⋅ H qL ( s )  •
TABLE I r ( s ) = wT ( s) ⋅    ⋅θ Le ( s ) +   ⋅θ Ae ( s) −   ⋅ M A ( s) −
Fault Fault examples  0 1   H 2 ( s) ⋅ H qA ( s) 
F1 - permanent additive step fault on sensor, offset  H ( s) ⋅ H wL ( s)  •  H1 ( s ) ⋅ H θ L ( s )  
− 1  ⋅ M L ( s) −  H ( s ) ⋅ H ( s)  ⋅θ Li ( s) 
to sensor signal (±10%) H ( s ) ⋅ H ( s )
 2 wA   2 θA  
F2 - incipient additive ramp fault on an actuator ( 11 )
F3 - multiplicative fault on process parameters To obtain the decoupled residuals for each measured signal,
a specific condition must be satisfied in order to remove the
dependency by each measurement.
u1=MA
(S51) For example the residual rϑLi is decoupled for the measure-
HqL
ment of the water input temperature sensor θLi by satisfying a
u2=θ Li HwL condition like:
y1=θ Le
(T32)
HθL (T41)  H ( s ) ⋅ Hθ L ( s )  ( 12 )
wT ( s ) ⋅  1  ⋅ θ Li ( s) = 0
Hτ L  H 2 ( s ) ⋅ Hθ A ( s ) 
u3=ML
(F31) The above condition is satisfied if:
HqA
wθTLi (s ) = [ − H 2 ( s) ⋅ Hθ A (s ) H1 ( s) ⋅ Hθ L ( s)] ( 13 )
HwA
y2=θ Ae A solution for the decoupled residual generator elements is
HθA (T52) underlined:
u4=θAi
Hτ A  rϑLe ( s )   0 1 
(T51)   
r
 ϑ Ae ( s )   1 0 
 r (s) =  − H H  1  0
Figure 5. Heat exchanger MIMO model structure.
 MA   2 qA H 1 H qL  •    θ Le ( s ) +   θ Ae ( s ) − (14)
 0 1 
 rM ( s )   − H 2 H wA H 1 H wL    
 L   
The above transfer functions, their gains and time constants  rϑ Li ( s )   − H 2 H θ A H 1 H θ L 
are calculated in [2]. Considering the MIMO transfer function
model for the heat exchanger [2, 3] the next equations systems  H 1 H qL  • H H  • H H  
−  M A ( s ) −  1 wL  M L ( s ) −  1 θ L  θ Li ( s ) 
is obtained, where H1 and H2 take into account the sensor  H 2 H qA   H 2 H wA   H 2 H θ A  
model and channel model between heat exchanger and sensors:
The design residuals relation (14) is necessary for fault de-
 M•  tection and identification. Each residual was designed to be-
 A come independent to a specific measurement. In case that some
θ Le   H 1 0   H qL H wL H ϑ L   •  (8) of the measurements are damaged, the decoupled residual rest
θ  =  0  H ϑ A  
M L
 Ae   H 2   H qA H wA at a low value, instead all the other residuals are affected.
θ 
 Li  The relation (14) is completed with another two that repre-
 
sent the deviation of estimation values from the measured ones
representing the residual build up with an output error method:
IV. RESIDUAL GENERATION
r PϑLe (t ) = ϑˆLe (t ) − ϑLe (t ) ( 15 )
The design of parity equations starts with the mathematical
model of the process for water-air heat exchanger, Figure 5, rϑPAe (t ) = ϑˆAe (t ) − ϑAe (t ) ( 16 )
and equation (8).
To obtain the decoupled residuals for a measured signal, a spe-
Following the definition of the parity equations, for the heat cific condition must be satisfied in order to remove the depend-
exchanger case, these equations could be written using equa- ency of the residual from this signal. The solutions for residual
tion error method as: generator elements are depicted in TABLE II.
• The group of residuals affected or not permits to locate cer-
0 = θ Le ( s) − H1 ( s ) H qL ( s ) M A ( s ) − H1 ( s ) Hθ L ( s )θ Li ( s ) (9) tain faults. The residuals have the property to become zero if
•
any fault exists or be different from zero if a fault appears in
0 = θ Ae ( s) − H 2 (s) H qA ( s) M A ( s) − ( 10 ) measurements or process.
•
− H 2 ( s) H wA ( s) M L (s) − H 2 ( s) Hθ A ( s)θ Li ( s) To identify if a fault has occurred it must be established if at
least one residual exceeds its threshold value that implies that
The residual generator is introduced by: the residual is considered different from zero.
 TABLE II faults the sensitivity is decreasing and only faults greater then
10% could be isolated. Smaller errors could be only detected.
Si Decupling condition: Solution:
gnals TABLE IV. INCIDENCE MATRIX
Sensor’s fault
θ Le 1  T
wθ Le ( s ) = [0 1]
w T ( s )  θ Le ( s ) = 0 F31 S51 T32 T41 T52
(T41) 0 
1 1 0 0 1
rϑLe
0  0 1 1 1 0
θ Ae w T ( s )  θ Ae ( s ) = 0
T
wθ Ae ( s ) = [1 0] rϑAe
1 
(T52) r• 0 0 0 1 0
MA

r• 0 1 1 1 1
MA  H1 H qL  w T
( s ) =  − H 2 H qA H 1 H qL 
(S51) wT ( s )   M A (s) = 0 MA ML

 H 2 H qA  rϑ Li 1 1 0 0 1

M  H 1 H wL  w• T
( s ) = [ − H 2 H wA H1 H wL ] r PϑLe 0 1 1 1 0
L
wT (s)  M L (s) = 0 ML
(F31)  H 2 H wA  1 1 0 0 1
r PϑAe
0 – undeflected residual
θ Li  H1Hθ L  T
wθ Li ( s ) = [ − H 2 Hθ A H1 Hθ L ] 1 – significant residual deflection
wT ( s )   θ Li ( s ) = 0
(T32)  H 2 Hθ A 

To illustrate the above considerations a fault analysis will be

The structured residuals are so designed that one residual re- presented. Considering a fault in temperature sensors θLi, that
appears at moment t=2000s, it could be detected and isolated
sponds to some distinct faults and it is insensible to others.
due to its value. The residuals deflections are depicted in
When a fault occurs some residuals respond some others don’t.
Figure and for a better illustration these results were normal-
The response of residual set is characteristic to each fault and is
ized by its thresholds. This operation allows uniform represen-
so called signature or code of the fault. For these residuals the
tests of exceeding the thresholds (ki) are applied separately for tation of information and implies a unique threshold to be used
for all residuals.
each residual (ri(t)) as follows:
The fault signature is different for deviations between 2-6%,
 0 if ri ( t ) ≤ k i ( 17 )
ε i (t ) =  i = 1..n as seen in TABLE V. The fault can be isolated as sensor θLi
 1 f ri ( t ) > k i (T32) fault (Figure 6), but for faults greater than 3oC one cannot
distinguish between sensor’s offset and bias. Analysing Figure
where ε = [ε1 ... ε n ]' is the fault signature. for temperature θLi (T32) sensor fault the most sensitive residu-
In order to establish the identifiable fault signature the als are rML and rPϑLe, residuals that assure fault detection, fol-
threshold value ki must be settled. This value could be settled lowed by rϑAe and rMA that indicate the fault (fault isolation).
upon statistical or experimental considerations. In the last case
5
this value can take into account the noise effects and the mod-
elling errors too. The optimal selection of the thresholds is
made through a compromise between false alarms and leak of 0

fault detection. The thresholds values established for these re-

siduals, in TABLE III, could cover the noise effects as well as rθLe-5
modelling errors. rθAe
TABLE III. RESIDUAL’S THRESHOLD VALUES rθLi-10
Residual rMA
rϑLe rϑAe r• r• rϑ Li r PϑLe r Pϑ Ae rML
MA ML
Threshold 1,5
o
1,8
o
0,3 0,8 1,1 0,6 1,5 rPθLe
-15

value C C % m3/h o
C o
C o
C P
r θAe
The detection of a fault depends on the most sensitive resid- -20
1000 1500 2000 2500 3000 3500 4000 4500
ual and its isolation on the less sensitive one (but not on the time (sec)
t(s ec )

decoupled residual). To isolate the fault source a set of residu- Figure 6. Unified representation of residuals for a sensor T41 offset (10 oC)
als with different responses for each fault is needed. This prin-
ciple is illustrated by incidence matrix (TABLE IV), each col- Another example for an incipient additive fault in an actua-
umn representing a fault signature. The results presented in •
previous incidence matrix are valuable for investigated faults. tor, M L , is depicted in Figure 7. The early detection of the
This set of residuals is build to be very sensitive to faults in fault will depend on the most sensitive residual (rθLi).
temperature sensors, so a small deviation like 1÷3 oC (1÷6%
from maximal value) could be detected and isolated. For other
 TABLE V. FAULT’S SIGNATURES output. Herein, MIMO processes were considered. For each
Fault in sensor θLi residual it was set a suitable threshold that allows detection or
1 oC 2 oC 3 oC 10 oC k=0.9-1.1 isolation decision for a fault by monitoring the residual deflec-
tions. The residual sensitivity depends on parity equation pa-
Sensor’s offset fault Sensor’s bias fault
rameter and input measured signals. It was underlined that de-
rϑLe 0 0 0 0 0 tection of a fault depends on the most sensible residual but, in
the mean time, the fault isolation depends on the less sensible
rϑAe 0 1 1 1 1 residual (no the decoupled one).
r• 0 0 1 1 1 An improvement of this method could be obtained by con-
MA

r• 1 1 1 1 1 sidering the direction of residuals deflections. An oriented in-

ML cidence matrix is possible to be designed, directional residuals
rϑ Li 0 0 0 0 0 too and these tools could allow detecting and isolating multiple
faults.
r PϑLe 1 1 1 1 1
ACKNOWLEDGMENT
r P
ϑ Ae
0 0 0 0 0
This paper is the result of the authors research in the frame
of the cooperation with TU-Darmstadt. The authors express
their gratitude to Professor Rolf Isermann, Institute of Auto-
2
matic Control, for his guidance and continuous support.
1.5 REFERENCES
1 [1] Chow, E.Y. and Willsky, A.S. (1984) Analytical redondancy and the
design of robust failure detection systems, IEEE Transaction on
0.5 Automatic Control, AC-29, pp.603-614.
rθLe [2] Făgărăşan Ioana, Arsene Patricia, Iliescu S.St. (2002)- Fault detection and
0
rθAe isolation of sensors faults using model based methods, Control Engineer-
ing and Applied Informatics, CEAI, Vol. 4, pp. 37-46, ISSN 1454-8658,
-0.5 rθLi 2002
rMA [3] Făgărăşan, I., Iliescu, S., Barbulea, S., Arsene, P. (2005)- System Model-
-1 rML ling Versus Residual Generation in FDI Methods, The 15th International
rPθLe Conference on Control Systems and Computer Science, CSCS’15, May,
-1.5
rPθAe 2005, Bucharest, Romania, 2005, pg.86-91
[4] Frank, P.M. (1990). Fault diagnosis in dynamic systems using analytical
-2
1500 2000 2500 3000 3500 4000 and knowledge-based redundancy. Automatica, Vol.26, pp.459-474
time (sec)t(s ec ) [5] Gertler, J. (1991). Analytical redundancy methods in fault detection and
Figure 7. Unified representation of residuals for an incipient fault on actua- isolation. IFAC-Symposium SAFEPROCESS, Baden-Baden. Vol.1, pp.
tor F31 started at time t=2000 s. 9-21. Oxford: Pergamon Press.
[6] Gertler, J. (1998). Fault detection and diagnosis in engineering systems,
Marcel Dekker, New York.
V. CONCLUSIONS [7] Höfling, T. (1994) Detection of additive and multiplicative faults - Parity
space v. parameter estimation. Preprints of 2nd IFAC SAFEPROCESS
Symosium, (Helsinki, Finland), Vol. 2., pp. 539-544
A modern approach for FDI in industrial processes was pre- [8] Iliescu S.St., Arsene P., Făgărăşan I., Secuiu I., Barbulea S. (1999). Some
sented in this paper. Herein, 6 measurement signals were used aspects of the modeling and identification of cross-flow heat exchanger,
to generate 7 residuals leading to detect and isolate certain Conference on Control Systems and Computer Science, CSCS '12, Bu-
charest, vol. I, pag. 121-127
faults. This approach is based on analytical relations between [9] Isermann, R. and P. Balle (1996). Trends in the Application of Model
characteristics of analyzed process and on the measured sig- Based Fault Detection and Diagnosis of Technical Processes, Proceed-
nals, also. However, in many processes the sensors already ings of the 13th IFAC World Congress, San Francisco, Vol. N, pp. 1-12
exist for control and supervision purpose, but the analytical [10] Isermann, R. (2006) Fault-Diagnosis Systems, ISBN-10 3-540-24112-4
Springer Berlin Heidelberg New York.
relation between the measured signals is not exploited. In these [11] Patton, R.J., Frank, P., and Chen, J. (2000). Issues of fault diagnosis for
cases the approach can easily improve the process supervision. dinamic systems, Springer New York.

Analytical model’s parameters have been used to generate

the suitable equations to detect and isolate faults. A powerful
method for this purpose seams to be building residuals with
parity equations support. The set of affected and unaffected
residuals points to the fault location. Design of the parity equa-
tions is suitable for processes with more than one measured