Sei sulla pagina 1di 8

Proceedings of the 10th Mediterranean Conference

on Control and Automation - MED2002


Lisbon, Portugal, July 9-12, 2002.

ADAPTIVE OBSERVER BASED FAULT DIAGNOSIS


APPROACH APPLIED TO A THERMAL PLANT
L. B. Palma*, F. J. Coito*, R. N. Silva*
* Universidade Nova de Lisboa, Faculdade de Ciências e Tecnologia
Departamento de Engenharia Electrotécnica
fax: +351-212948532
e-mail: {LBP,FJVC,RNS}@fct.unl.pt

Keywords: On-line fault detection and diagnosis, Fault diagnosis systems concern with the detection
adaptive observers, closed-loop identification, and diagnosis of faults in engineering systems, such
thermal plants. as machines, vehicles, etc., whether they occur in
the plant or in its measurement and control
instruments. In general, Faults are deviations from
Abstract
the normal behaviour in the plant or its
This paper presents an on-line model-based instrumentation. The faults of interest belong to one
procedure for the detection and diagnosis of sensor of the following categories: additive process faults,
and process faults in plants. The detection and multiplicative process faults, sensor faults, and
diagnosis method proposed is based on parity actuator faults.
equations and on an adaptive observer based Fault detection and diagnosis systems implement
approach. The authors propose the use of dynamic the following tasks [4,5]: (i) Fault detection (FD),
features (static gain and bandwidth) of black-box i.e., the indication that something is going wrong in
(AutoRegressive with eXogenous input - ARX) the monitored system; (ii) Fault isolation (FI), i.e.,
models, as features for fault detection and diagnosis the determination of the exact location of the fault
purposes. The ARX models of the thermal process (the component which is faulty); (iii) Fault analysis
under investigation are obtained by closed-loop (FA) or identification, i.e., the determination of the
recursive identification techniques. Residual magnitude of the fault.
analysis and geometrical tests are then used for
fault detection and diagnosis, respectively. The Most practical systems contain only the fault
proposed procedure has been evaluated using a detection and isolation stages (and are referred to as
benchmark thermal process. FDI systems). Also, in many cases "diagnosis" is
used simply as a synonym to "isolation". Usually,
1 Introduction the FDI activity takes place on-line, in real-time;
the two tasks, detection and isolation, may be
In the last few decades, process monitoring and performed in parallel or sequentially.
fault diagnosis are becoming an ingredient of
modern automatic control systems, due to the The methods of fault detection and diagnosis may
increasing demands for higher performance and be classified into two major groups [4,5,10]:
quality, more cost efficiency, and also to the (i) Model-Free methods. These FDI methods do not
continuous growing of the automation degree of use the mathematical model of the plant. They
technological processes [4]. A traditional way to range from hardware physical redundancy through
improve the process safety and reliability is to signal processing methods (limit-checking, etc).
enhance the quality and robustness of each process
(ii) Model-Based methods. These FDI methods use
component; even so, a fault-free process operation
a mathematical model of the monitored plant, and
cannot be guaranteed.
rely on the concept of analytical redundancy: the results are shown. Finally, the conclusions and
measured values are compared to analytically future work is presented in Section 5.
computed values of the respective variable. This
idea can be extended to the comparison of two 2 Process description and closed-loop
analytically generated quantities, obtained from
different sets of variables. In either case, the model identification
resulting differences, called Residuals, are Thermal plants are those that involve heat transfer
indicative of the presence of faults in the system. from one substance to another, by means of
This formulation is usually implemented as parity conduction, convection or radiation. Steam boilers,
equations, and is more suitable for detection of furnaces, and HVAC (Heating, Ventilating, and
additive faults. Another class of model-based Air-Conditioning) are some of the many examples
methods relies on parameter estimation, and are of this kind of processes in industrial plants.
more suitable for detection of multiplicative faults.
In the field of fault diagnosis there is a clear trend 2.1 Thermal process model description
from the well-established but in their efficiency
limited traditional methods of signal based fault In continuous-time, the nonlinear thermal process
diagnosis, towards the model based approaches under study, the spray-dryer PT-326 depicted in
using analytical and/or qualitative models for Figure 1, can be modelled by a first-order process
residual generation, and modern strategies of with pure time-delay [3,8,9]. Its transfer function is
residual evaluation including the methods of given by Equation (1), where K is the system static
decision making with the aid of fuzzy logic and gain, τ is the time constant, and τ c denotes a pure
neural networks [4]. Clearly the black-box time-delay. With nominal conditions, temperature
modelling approach have shorter development set-point ( 35º C ), air-inlet (damper) position
times than the the methods that require the full ( Ω = 50º ), and position of sensor T1( III ), a set of
parameter was determined: θ ≈ [0.3;0.5s;0.1s ] .
T
physical modelation of the plants.
In general, there are four model-based residual
Ke − sτ c
;θ = [K ,τ ,τ c ]
T
generation methods, somewhat overlapping [4,5,6]: Gc ( s,θ ) =
(i) Kalman filtering, (ii) diagnostics observers, (iii) sτ + 1 (1)
parity (consistency) equations, and (iv) parameter
estimation. To implement FDI practical systems, In the spray-dryer PT-326, depicted in Figure 1, air
different methods must be used in conjunction, in is fanned through a tube and heated at the inlet. The
order to maximize the FDI performance [6]. input u is the power to the heating device
(actuator). The output is the outlet air temperature,
The benchmark thermal process under study (the measured by thermistors (T1 and T2).
spray-dryer PT-326 [3]) has a dynamic behaviour
(nonlinear first-order process with pure time-delay) 18
0
o

Air
I II III B
similar to many industrials plants, such as steam- outlet

boilers, furnaces, HVAC (Heating, Ventilating, and Ω


9 o
0
Air-Conditioning) systems [7], etc. In Section 2, the R

PT-326 is presented and a mathematical model is


Actuator Sensor T1 Sensor T2
derived; the closed-loop control and identification 1o
0

architecture is also described. In Section 3 is Air inlet

explained how parity equations and adaptive


observer based approach are used for the generation Figure 1: Schematics of the modified PT-326.
of features, and applied to the detection and
Since the process linearized dynamics is time-
diagnosis of faults. In Section 4, experimental
varying, a black-box ARX model is used to model
the process, with repeated on-line identifications
[8]. The ARX(na,nb,nk) model can be expressed by (T2 − T0 ) versus Pa . These ARX models can be
Equation (2), where y[k ] and u[k ] are, seen as adaptive observers of the process at each
respectively, the output and the input signals time instant t k .
measured at discrete-time k. The model orders are
na and nb, and nk is a pure-time delay. The on-line identification of the ARX models uses
a recursive least-squares (RLS) algorithm based on
y[k ] + a1 y[k − 1] + ... + a na y[k − na] = the Kalman filter [8].
b1u[k nk ] ... bnb u[k nk nb 1]
− + + − − + (2) ^
ID module y
RARX model
T0
2.2 Digital PID controller PRBS
signal
In the present work, the process was controlled uPS
u uc y
around the nominal conditions using a digital PID Actuator Process Sensor
controller, based on the discrete-time velocity
control algorithm [1]. PID controller parameters e
were tuned using a Ziegler-Nichols approach, PID
controller uPID
giving Kp = 2.4 (proportional gain), Ti = 0.5s R
-
(integral time), and Td = 0.075s (derivative time).
The algorithm was implemented on Matlab® and a
data acquisition board (DAQ) was used to interface Figure 2: Control and identification architecture.
to the process, via dynamic link library (DLL). The
sampling period was 110 ms. 3 Fault Detection and Diagnosis
Approach
2.3 Closed-loop model identification
The proposed approach to FDI relies on two
Typically industrials plants work continuously in residual generation methods: parity (consistency)
closed-loop control mode, therefore the tasks of equations and adaptive observers. Different
model identification and fault detection and methods must be used in order to implement FDI
isolation must be performed on-line without practical systems [6].
opening the loop.
The Figure 2 shows the control architecture. In 3.1 Process and Fault Modelling
order to ensure persistence of excitation conditions
a Pseudo-Random Binary Signal u PS is added to The PT-326 process is modelled by a discrete-time
Linear and Time-Invariant (LTI) system, expressed
the control action u PID ; the amplitude of the PRBS
according to Equation (2), assuming a nominal
signal is chosen to be small enough to minimize its steady-state operation point.
effect on the control error, but large enough to be
persistently exciting [8]. Equation (3) is an input-output model of the
process, in a fault-free situation [4]. The input and
The identification (ID) module receives as inputs output signals are, respectively, u and y , Gu (z )
the heating power applied ( Pa , proportional to the
denotes the transfer function, and ∆y (z ) is a signal
control action u ), and the temperature gap between representing disturbances and model uncertainty
room temperature ( T0 ) and the outlet air that are unknown but bounded.
temperature ( T1 or T2 ). It predicts the temperature
y ( z ) = Gu ( z )u ( z ) + ∆y ( z ) (3)
at the current time instant t k . Two SISO (single-
input single-output) ARX models are used: M S1
models (T1 − T0 ) versus Pa , and M S 2 models
Equation (4) extends the model to include faults, FAULTS
where G f (z ) denotes a known transfer function
Input Output
matrix from the fault f (z ) to the output. Signals Signals
Actuator Process Sensor

y ( z ) = G u ( z )u ( z ) + ∆ y ( z ) + G f ( z ) f ( z ) (4)
Fault Detection
The basic requirements of a residual generator are Mathematical Process model
expressed by Equations (5) and (6).
r ( z ) = 0 ⇐ f ( z ) = 0, ∆y ( z ) = 0 (5) Generation of Features

Features
r ( z) ≠ 0 ⇐ f ( z) ≠ 0 (6) Detection of Deviations

The dynamic system expressed by Equation (7), is a


residual generator whose dynamics is governed by Fault Symptoms

Equation (8). Fault Diagnosis


r ( z ) = y( z ) − Gu ( z )u ( z ) (7) Knowledge Base

r ( z ) = ∆y ( z ) + G f ( z ) f ( z ) (8)
Detected and
Due to the existence of ∆y(z ) , the performance of Computer Isolated Faults

the above system may be poor, and Gu (z ) may be


an unstable system. Ding and Frank [2] introduced Figure 3: Typical model-based FDI architecture.
the construction form of LTI residual generators
expressed by Equation (9), were R(z ) is a post- The proposed FDI methodology is based on parity
filter. equations, and on adaptive observer based approach
[4]. This adaptive observer based approach is an
∧ ∧
r ( z ) = R ( z )( M u ( z ) y ( z ) − N u ( z )u ( z )) = integration and extension of the observer based and
∧ (9) parameter estimation approaches. The core of such
R( z )( y ( z ) − y ( z )) kind of FDI systems consists of an adaptive
The generation of residuals based on observers observer who is used both to improve the
does not need the application of state-space theory, robustness against model uncertainty due to
since it is possible to use output observers designed parameter changes and to detect and identify the
as filters expressed by Equation (9). faults.
Parity (consistency) equations are rearranged direct
3.2 Proposed FDI approach input-output model equations, subjected to a linear
dynamic transformation [5]. The transformed
Figure 3 shows the basic model-based FDI scheme residuals serve for detection and isolation, like
[4,10]. Based on measured input and output signals, Equation (9).
fault detection methods generate features by means
of a mathematical process model; these features The parameter estimation is a natural approach to
may include physical parameters, residuals, mean the detection and isolation of parametric
values, etc. The features are compared to that one (multiplicative) faults [5]. A reference model is
in the nominal (fault-free) case. The respective obtained by first identifying the plant in a fault-free
deviations are compared to thresholds. As a result situation. Then the parameters are identified on-
one gets fault symptoms. The symptoms are finally line. The process needs a continuous excitation in
fed into a knowledge base that maps the symptoms order to estimate the parameters; this may increase
to the respective faults. the regulation error in the case of processes
operating at stationary operating point.
The proper threshold values selection is important e) Static gain residual of model M S1 ,
for detecting faults. The thresholds can usually be rsg1(t k ) = sg1(t k ) − µ sg1 . The difference
determined from statistical properties of the
process. Whenever a measurement exceeds a between the static gain of model M S1 at time
threshold limit a fault is presumed to have instant t k and the nominal value.
occurred. In this work, a three-sigma limit
{μ − 3σ;μ + 3σ } was used as the threshold value [7]. f) Static gain residual of model M S 2 ,
Mean ( µ ) and standard deviation ( σ ) for each rsg 2(t k ) = sg 2(t k ) − µ sg 2 . The difference
feature is computed from a set of test data, obtained between the static gain of model M S 2 at time
from samples acquired between 180 and 360 s of instant t k and the nominal value.
operation (see Figure 4), in the nominal-operating
region (without faults). g) Bandwidth residual of model M S1 ,
In this work, the features (see Figure 3) used for rbw1(t k ) = bw1(t k ) − µ bw1 . The difference
residual generation were obtained from process between the bandwidth of model M S1 at time
variables and dynamic features (static gain and instant t k and the nominal value.
bandwidth) of the ARX models M S1 and M S 2 (see
section 2.3); these features are on-line estimated at The static gain (STG) of the ARX models is
each time instant t k . The computed Residuals are computed according to
the following, and can be observed in Figures 5 and STG = G u ( z = 1) (10)
6: The bandwidth (BW) of the ARX models is
a) Output residual, ry1(t k ) = y1(t k ) − y1m(t k ) (see computed assuming a first-order model, according
Figure 4 and Figure 5). The difference between to Equation (11) where t r is the rise-time.
the measured temperature T1 and the predicted 3 (11)
BW ≅
temperature obtained from ARX model M S1 , at tr
each time instant t k . In the proposed approach the tasks of fault
detection (FD) and fault isolation (FI) occurs
b) Output residual, ry 2(t k ) = y 2(t k ) − y 2m(t k )
simultaneously.
(see Figure 4 and Figure 5). The difference
between the measured temperature T2 and the
3.3 Fault detection algorithm
predicted temperature obtained from ARX
model M S 2 , at each time instant t k . The fault detection algorithm detects a fault if a
certain residual rX at time instant t k exceeds the
c) Heating power residual expressed by
three-sigma limit (see Section 3.2). The mean µ rX
rAP(t k ) = Pa (t k ) − µ Pa . The difference between
of the residual and the standard deviation σ rX were
the heating power Pa at time instant t k , and the
obtained from a test data set acquired in the
nominal (mean) value of Pa obtained from test nominal operating region (without faults).
data set (nominal operating point).
Each fault has a signature (pattern) in the space of
d) Temperature residual, expressed by the parity residuals described by the following five residuals:
equation ry12(t k ) = y1(t k ) − y 2(t k ) − δ 12 . The {rAP; ry12; rsg1; rsg 2; rbw1}. The fault signature
difference between the two sensors ( T1 and T2 ) must be capture on test experiments (see Figure 6);
minus an offset (resulting form the heat loss each fault can be detected and isolated according to
across the ventilation tube). the variations in the residuals. The values on the
pattern vector belong to the range {− 1;0;+1}; "0"
means no fault (the residual belongs to the three-
sigma limit), "-1" means that the residual exceeds platform used is the Matlab® running on a
the lower limit threshold, and "+1" means that the computer based on a Pentium® processor at 400
residual exceeds the upper limit threshold. MHz.
The three real Faults that take place are: (i) a
positive voltage offset at temperature sensor T2
(additive fault on sensor, AFS 2 ); (ii) a 90 degree
rotation of temperature sensor T1, in order to
simulate a clogging (multiplicative fault on sensor,
MFS1 ); (iii) an increase in the air flow rate
(multiplicative fault on process, MFP ). The time
sequence was that AFS 2 , MFS1 , and MFP (see
Figure 5).
In order to detect the additive fault AFS 2 , the
signature used is {x;−1; x;+1; x}, were x means any
value in the range {− 1;0;+1}.
In order to detect a multiplicative fault MFS1 or
MFP , the following signature {+ 1; x;−1; x; x} is
used.

3.4 Fault isolation algorithm


Figure 4: Measured and ARX models outputs.
This algorithm implementation is similar to the
fault detection algorithm. The faults are isolated
4.1 Operating conditions
using the following signatures:
a) Fault AFS 2 : {0;−1;0;+1;0}; Room temperature is around 22.5ºC (see Figure 4),
measured by a NTC thermistor similar to sensor T2.
b) Fault MFS1 : {+ 1; x;−1;0;−1}; Temperature set-point is 35º C , the damper
position is Ω = 50º , and the sensor T1 is at position
c) Fault MFP : {+ 1;0;−1; x;+1}. III. The sensor T2 dynamics is slower than sensor
The additive fault AFS is simple to detect and T1.
2
isolate. On the other way, multiplicative faults The nominal-operation, without faults, goes from
MFS1 and MFP may only be isolated from their 180 s to 360 s, and it is used to define the
effect on the bandwidth residual, rbw1 (see Figure thresholds for the residuals.
6). The amplitude of the PRBS signal is computed in
order to obtain a maximum control error (CE)
4 Experimental results percentage of about 5%, relative to temperature set-
point (spT). Experimental results give
The experiments are performed with the thermal CE
plant working at a stationary operating point, and spT ≅ 4.3% .
additive and multiplicative faults are applied to the
process as described in Section 3.3. Results show The process is modelled by two ARX models M S1
the proposed FDI performance, obtained with and and M of 3rd order, like ARX(3,3,1) as
S2
without low-pass filtering the residuals; the filter
mentioned in Section 2.3.
cut-off frequency is 0.6 rad/s. The software
The sampling time is Ts = 110 ms , giving a rate of 4.3 Fault detection results
computing time versus sampling time of Fault detection results can be observed in signal
Tcpu fdv (fault detection vector, see Figure 6). This
= 0.5 . Figure 4 shows, from top to bottom: the
Ts vector has three values: 0, 1, or 2; with "0" meaning
room temperature ( Te = T0 ), the temperature set-
no fault detected, "1" the detection of an additive
point R and the sensors (T1 and T2) temperatures fault on sensor T2, and "2" the detection of a
y1 and y 2 , the heating power Pa , and the two multiplicative fault on sensor T1 or on the process.
The fault detection algorithm used has some fault
ARX models predicted outputs y1m and y 2m .
isolation properties, and estimates also the duration
of the fault.
4.2 Experimental results
Figure 5 shows, from top to bottom, the output
residuals ry1 and ry 2. Filtered residuals are shown
in solid line, and non-filtered residuals are shown in
dotted line; these output residuals are not used for
FDI purposes. Bellow are plotted the static gain
( sg1 and sg 2 ) of the ARX models ( M S1 and M S 2 ).
At the bottom, is the bandwidth of model M S1 . The
effects of each fault may be easily observed.

Figure 6: Residuals and FDI vectors.


The fdv based on low-pass filtered residuals is
shown in solid line, and in dotted line is shown
fdv based on non filtered residuals; the use of a
low pass filter rejects high frequency signals.
The first fault ( AFS 2 ) occurs at time instant
t k ≅ 384.5 s , and is detected 2 s afterwards; this
delay depends strongly on the cut-off frequency of
the low-pass filter. Its real duration is about 25 s,
and the algorithm indicates a duration of about 27
s. The second fault ( MFS1 ) occurs at time instant
Figure 5: Output residuals and model features. t k ≅ 495 s , and is detected 2 s afterwards. It lasts
about 10 s, and the algorithm indicates a duration
of 14 s. The third fault ( MFP ) occurs at time
instant t k ≅ 730 s , and is detected 2 s afterwards. It analysis of variations on ARX parameters, the
lasts 25 s, and the algorithm indicates 27 s. adaptation of this approach to non stationary
operation, and finally the application of this type of
FDI approach to more complex plants.
4.4 Fault isolation results
The fault isolation results can be observed in signal
fiv (fault isolation vector, see Figure 6). This References
vector can have four values: 0, 1, 2, or 3; "0" means [1] K. Astrom, T. Hagglund. “Automatic Tuning of
no fault isolated, "1" means the isolation of fault PID Controllers”, Instrument Society of America,
AFS 2 , "2" means the isolation of fault MFS1 , and (1988).
"3" means the isolation of fault MFP . The fiv
[2] X. Ding, P. Frank. “Fault Detection via
vector also shows the low-pass filtered and non- Factorization Approach”, Systems Control Letter,
filtered performance. The delay of fiv relative to vol. 14, pp. 431-436, (1990).
fdv is more relevant for the multiplicative fault on
[3] Feedback. “Process Trainer PT-326 / PCS-327”,
sensor T1 ( MFS1 ).
Feedback Instruments Limited, (1985).
[4] P. Frank, S. Ding, T. Marcu. “Model Based
5 Conclusions Fault Diagnosis in Technical Processes”, Trans. of
the Institute of Measurement and Control, vol. 22,
In this paper a parity equations and an adaptive no. 1, (2000).
observer based approach for fault detection and
diagnosis was presented. The performance has been [5] J. Gertler. “Fault Detection and Diagnosis in
evaluated applying the methodology on-line, in Engineering Systems”, Marcel Dekker Inc., (1998).
real-time, to a thermal plant. The FDI approach [6] R. Isermann. “Supervision, Fault-Detection and
presented do not require a detail mechanistic model Fault-Diagnosis Methods - An Introduction”,
of the plant, only the approximate orders and pure Control Eng. Practice, vol. 5, no. 5, pp. 639-652,
time-delay. It was proposed the use of dynamic (1997).
features (static gain, and bandwidth) of ARX
models for FDI purposes. This approach seems [7] W. Lee, C. Park, G. Kelly. “Fault Detection in
adequate for plants that can tolerate a small an Air-Handling Unit Using Residual and
percentage of control error. Recursive Parameter Identification Methods”,
ASHRAE Transactions, vol. 102, no. 1, pp. 528-
Parameter estimation may be more reliable than the 539, (1996).
analytical redundancy methods, but it is also more
demanding in terms of on-line computation and [8] L. Ljung. nd
“System Identification: Theory for the
input excitation requirements. User”, 2 edit., Prentice-Hall, (1999).

Different methods for residual generation must be [9] L. B. Palma, R. N. Silva, F. V. Coito.
used in order to implement FDI practical systems, “Intelligent Fault Diagnosis Methods in Industrial
and the residuals must be low-pass filtered with the Plants - Application to the PT-326” (in
objective of obtain a better FDI performance. Portuguese), VII Jornadas Hispano-Lusas de
Engenharia Electrotécnica, vol. 4, pp. 303-308,
The choice of thresholds is also a crucial problem, Madrid-Spain, (2001).
and must be based on statistical properties of the
signals on the system. [10] L. B. Palma, F. V. Coito, H. Duarte-Ramos.
“Supervision and Control of Industrial Plants – An
Some pointers to future work are the development overview” (in Portuguese), VII Jornadas Hispano-
of more reliable FDI algorithms (possibly using Lusas de Engenharia Electrotécnica, vol. 4, pp.
decision systems based on fuzzy logic), the analysis 165-170, Madrid-Spain, (2001).
of fault detection and isolation performance, the

Potrebbero piacerti anche