Академический Документы
Профессиональный Документы
Культура Документы
, V. M. Becerra
, I. Khoo
, J.M.F.Calado
i
=
i
+
i
N
j
j
+
j
for i = 1, 2, , N (3)
where
i
represents the weight associated with each
fault mode, represents the result of the optimization
of the cost function (9), p represents the result of the
probabilistic approach, and represent the degree of
condence in each one of the approaches.
The values of the weights help to isolate a fault, hence
this solution performs the FDI function. If a certain weight
is equal or very close to one (then all other weights are
2007 Medterranean Conference on Contro and
Automaton, |uy 27 - 29, 2007, Athens - Greece
T14-009
very close to zero), that indicates that the real process
is best represented by the model corresponding to that
weight. In such case, it can be said that a particular model
is being red, which indicates that the plant is likely to
be working in a particular (faulty or normal) regime.
1) Optimization: In this module the operating regime
(including faulty regimes) is identied. Dene a vector
Y
p
H
with the last H samples of the plant output as
follows:
Y
p
=
_
_
y(k)
y(k 1)
.
.
.
y(k H + 1)
_
_
(4)
Dene the following model data matrix:
Y
m
=
_
_
y
1
(k) . . . y
N
(k)
y
1
(k 1) . . . y
N
(k 1)
.
.
.
.
.
.
.
.
.
y
1
(k H + 1) . . . y
N
(k H + 1)
_
_
where k is an integer index that denotes present time,
the row dimension H represents the number of past
samples that are considered, and the column dimension
N is the number of models. Matrix Y
m
contains the last
H samples predicted by each of the N models in the
model bank. Notice that the elements of Y
m
are one-
step-ahead model predictions of the output, such that
y
i
(k j) = y
i
(k j|k j 1), j = 0, . . . , H 1.
Dene a vector of weighting factors =
[
1
, . . . ,
N
]
T
which satisfy:
N
i=1
i
= 1
0
i
1 i = 1, , N
(5)
The purpose of these weighting factors is to quantify the
degree of activation of each model in the model bank.
Using the individual models from the model bank, and
considering H past output samples from all models, it is
possible to dene a weighted historical output vector of
the model bank, Y
w
H
that considers the weighted
outputs of the individual models:
Y
w
=
N
i=1
Y
m,i
i
= Y
m
(6)
where Y
m,i
represents the past H samples of the i-th
model output, and consists of the i-th column of matrix
Y
m
.
It is now possible to dene the residuals
H
between the weighted historical output of the model bank
Y
w
and the historical plant output Y
p
, as follows:
= Y
p
Y
w
= Y
p
Y
m
(7)
and, using these residuals, it is possible to dene a cost
function to be minimized:
(Y
p
, Y
m
, ) =
T
= (Y
p
Y
m
)
T
(Y
p
Y
m
) (8)
Given the plant data Y
p
and the model data Y
m
, this
cost function is minimized with respect to the weight
vector subject to the constraints given in Equation
(10). The solution of the optimization problem gives an
optimum vector of weighting factors , which indicate
what combination of the model set best represents the
real process under the current operating conditions. This
optimization problem can be solved at every sampling
instant using quadratic programming [12], with the model
data matrix Y
m
and the plant data vector Y
p
being updated
with new values using the receding data window concept.
When experimenting this solution in a real process it was
clear that in some situations the fault diagnosis could
be improved. One solution proposed in [14] to solve
this problem is to change the cost function making sure
that the penalty is not only on the residuals described
by (7), but also on the weighted sum of the square of
the difference between the most recent plant output y(k)
and the corresponding prediction y
i
(k) from each of the
models:
(Y
p
, Y
m
, ) = (Y
p
Y
m
)
T
(Y
p
Y
m
)
+
N
i=1
i
(y(k) y
i
(k))
2
(9)
where > 0 is a scalar parameter that denes the
importance of the second term of the cost function and
N
denes a vector with the level of condence in
each one of the models.
2) Probabilistic approach: In order to achieve a higher
level of condence in the FDI mechanism, a probabilis-
tic approach was introduced. Markovian Jump Systems
(MJS) [15],[16], [17], are used as a switching mecha-
nism between a series of models that represent different
operating regimes of the system. One can consider a
system similar to the one represented in (1). Suppose now,
that the system faults could be represented by a Markov
chain, taking values in: s(k) M = {1, 2, , N}. The
switching can be calculated using initial and transition
probabilities,
P{s
i
(0)} =
i
(0)
P{s
i
(k)|s
i
(k 1)} = P{s
i
(k)|s
i
(k 1), Y
k1
p
} =
ij
for all i, j = 1, 2, , N (10)
where
i
represents the probability of each model,
ij
represents the transition probability matrix and Y
k1
p
represents the measurement obtained from the system
until sample time (k-1). The probability transition matrix
represents the transition probability between different
operating regimes. The
ij
represents the probability of
transition from mode i to mode j. Using a minimum mean-
square error (MMSE) estimation it is possible to obtain
the optimal mode for a certain operating regime [17].
Because several numerical problems can occur (exponen-
tial computation and memory growth) using this optimal
estimation, a sub-optimal method known as the interacting
multiple model (IMM) has been proposed [16]. Although,
this sub-optimal method solves the numerical problems
has a major drawback, the fact that the transition matrix
has to be known in advance. In this sense the results
achieved with this method are very dependent of the initial
guess of the transition matrix. In order to make this
2007 Medterranean Conference on Contro and
Automaton, |uy 27 - 29, 2007, Athens - Greece
T14-009
method more robust to the initial transition matrix, it has
been proposed that this matrix can be updated at every
sample time using the measurement information received
from the process. In this work two different algorithms are
tested to update the transition matrix. Both algorithms are
based on a simple idea:
At every sampling time the mode probability is
updated and the mode likelihood functions (just like
in IMM algorithm or generalized pseudo-Bayesian
algorithm [16]):
i
(k) P{s
i
(k)|
(k 1), Y
(k1)
p
}
L
i
(k) p[Y
p
(k)|s
i
(k),
(k 1), Y
(k1)
p
] (11)
The transition probability matrix is updated using
L
j
(k)}p[
1
j|Y
k1
p
] (12)
for j = 1, 2, , N (13)
where:
j
(k) =
i
(k 1)
(k 1)
(k 1)L(k)
(14)
Using this approach the update of the transition prob-
ability matrix can be done on-line using two different
methods:
Quasi-Bayesian Algorithm (QBA)- in this algorithm
it is dened that:
ij
(k) =
1
k +
i
(0)
ij
(k) (15)
where,
ij(k) =
ij
(k1)+
ij
(k 1)g
ij
(k)
N
j=1
ij
(k 1)g
ij
(k)
(16)
and:
g
ij
(k) = 1 +
i
(k)[L
i
(k)
i
(k 1)L(k)] (17)
Numerical-Integration Algorithm (NIA) - Using the
non decoupled probability density function version
of (12):
p[
|Z
k
] =
(k 1)
L(k)
(k 1)
(k 1)L(k)
p[
|Z
k1
]
(18)
Using t
pm
= 1, 2, , V transition probability ma-
trixes it is possible to:
p
(t
pm
)
(k) =
(k 1)
(t
pm
)
L(k)
(k 1)
(k 1)L(k)
p
(t
pm
)
(k 1)
(k) =
1
N
t
pm
=1
t
pm
p
t
pm
(k) (19)
See [17] in order to see a complete description of both
algorithms.
0 500 1000 1500 2000 2500 3000 3500 4000
1
2
3
4
5
0 500 1000 1500 2000 2500 3000 3500 4000
0
10
20
30
0 2000 4000 6000 8000 10000 12000 14000 16000 18000
0
1
2
3
4
5
Time (s)
Signal Fault
Fig. 2. Faults introduced in the system.
III. RESULTS
In order to test the proposed FDI mechanism an HVAC
system is used. Using models of the different system
components under different fault conditions, a simulation
is created to evaluate the performance of the FDI, before
the real experiments are performed. After the simulation
test the proposed method was applied to the real HVAC
system, using real faults introduced in the HVAC compo-
nents.
A. Simulation Results
The different fault regimes in the actuator are modelled
using neural network ARX models [13] and were built in
order to cover the operating regime of the heater bank
actuator fault. Sensor faults were modelled using four
different operating regimes and a nominal model. The four
different regimes were computed to cover the domain of
the output of the temperature sensor of the heater bank.
The following simulation tests were conducted:
Actuator Fault - in this case an incipient fault was
introduced in the system (see Fig. 2 a)). The fault
was introduced in the 1000 second of the simulation.
Fig. 3 presents three different results with three
different strategies in the FDI mechanism. The rst
result presented is obtained with an FDI mechanism
that uses only (9) without the probabilistic approach.
Fig. 3 b) and 3 c) represent the results obtained using
the cost function (9), algorithm QBA and algorithm
NIA, respectively. The sample time used in this
simulations was 10 seconds and the data window
used in the optimization of the cost function is of 40
samples.
Sensor Fault - the sensor fault introduced was also
a incipient fault, and was introduced in the 1000
second of the simulation (Fig. 2 b)). Also, in this
case three different approaches were used to test
the different algorithms. In Fig. 4 is presented the
2007 Medterranean Conference on Contro and
Automaton, |uy 27 - 29, 2007, Athens - Greece
T14-009
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.5
1
FDI without Probability Analysis
W
e
i
g
h
t
s
Time(s)
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.5
1
FDI with Probability Analysis QBA
Time(s)
W
e
i
g
h
t
s
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.5
1
FDI with Probability Analysis NIA
Time(s)
W
e
i
g
h
t
s
Process Model Fault Model 1 Fault Model 2 Fault Model 3
Fig. 3. Actuator incipient faults in simulation
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.5
1
FDI without Probability Analysis
W
e
i
g
h
t
s
Time(s)
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.5
1
FDI with Probability Analysis QBA
Time(s)
W
e
i
g
h
t
s
0 500 1000 1500 2000 2500 3000 3500 4000
0
0.5
1
FDI with Probability Analysis NIA
Time(s)
W
e
i
g
h
t
s
P. Model F. Model 1 F. Model 2 F. Model 3 F. Model 4
Fig. 4. Sensor incipient faults in simulation
different results, with and without the probabilistic
approach. In this case the sample time is also of 10
seconds and the cost function is calculated using 400
seconds of data.
B. Experimental Setup
A heater bank (part of the HVAC system) was used to
test the FDI algorithm. The heater bank is an electric
heater with a control action between 0 to 4 V. The
temperature is measured before and after the heater. The
air supply ow of the air handling unit passes through
the heater and is delivered to three different rooms. The
heater has a complex nonlinear behavior if the air ow in
the inlet is changed. A supply fan is used to distribute
Fig. 5. Experimental setup of the developed architecture.
the hot air ow to the rooms, and a terminal unit is
used to control the ow entering the room. Fig. 5 shows
a schematic of the process used to test the algorithms.
The fan works normally at a constant speed, but it is
possible to change it by means of a 0 to 10 V signal. In
this work, two temperature sensors reading the inlet air
temperature and outlet air temperature in the heater are
used. The instrumentation used in the system is connected
by a LON Network and all the signals are acquired by
a Personal Computer with Microsoft Windows
R
and
are available to the operating system by a Dynamic
Data Exchange (DDE) Server. Matlab
R
, Simulink
R
, and
the Optimisation Toolbox is used to implement the FDI
mechanism.
Two types of faults were simulated. A combination of
incipient fault and abrupt faults were tested. The fault
introduced represent usual problems in heaters, such as
material settling in the heater surface. The faults are
simulated a power losses in the heater. The multiple
model approach was used using three different models
to cover the domain of the fault plus one model for
the normal conditions. Neural network ARX models [13]
were employed. (see Table I). The faults were introduced
in the system during a period of 17000 seconds and
within that period three faults were tested: two of them
incipient faults and one abrupt fault. The structure of the
faults was built in order to simulate the degradation of
this kind of systems. The degree of condence in each
one of the methods was = 1 = 0.8. Regarding
the probabilistic approach and NIA algorithm the initial
number of transition probability matrix were 50, and the
initial guess of the transition matrix for algorithm QBA
was,
_
_
_
_
0.9 0.08 0.01 0.01
0.02 0.9 0.04 0.04
0.01 0.03 0.9 0.06
0.01 0.01 0.08 0.9
_
_
_
_
(20)
and the initial probability vector was:
_
0.9 0.05 0.0025 0.0025
(21)
Fig. 6 shows the results obtained with the different
approaches. Observing Fig. 6 it is possible to see that in
different situations the probabilistic approach can improve
the detection and isolation mechanism. One of the aspects
that is more relevant is the less variability in the models
identied by the weights. This will introduce less false
alarms in the system and will make control recongu-
ration more accurate. In the period between 12000 and
16000 seconds the fault is constant and it is clear that the
methods with the probabilistic approach are more accurate
2007 Medterranean Conference on Contro and
Automaton, |uy 27 - 29, 2007, Athens - Greece
T14-009
TABLE I
NEURAL NETWORK MODEL.
Weights
Neuron 1 Input to Hidden Layer 0.0404 -0.0821
0.0000 -0.9310
Hidden to Output Layer 35.2373
Neuron 2 Input to Hidden Layer 0.0404 -0.0821
0.0000 -0.9310
Hidden to Output Layer 35.2373
0 2000 4000 6000 8000 10000 12000 14000 16000 18000
0
0.5
1
FDI without Probability Analysis
W
e
i
g
h
t
s
Time(s)
0 2000 4000 6000 8000 10000 12000 14000 16000 18000
0
0.5
1
FDI with Probability Analysis QBA
Time(s)
W
e
i
g
h
t
s
0 2000 4000 6000 8000 10000 12000 14000 16000 18000
0
0.5
1
FDI with Probability Analysis NIA
Time(s)
W
e
i
g
h
t
s
Process Model Fault Model Fault Model 2 Fault Model 3
Fig. 6. FDI results in a heater bank of a HVAC system
in the isolation of the fault. With the abrupt change
introduced at 10000 seconds, the probabilistic approach
(QBA) achieves a fast fault identication. Nevertheless,
the probabilistic approach can result in worst results when
the fault is not clearly ring a certain mode of operation.
Between 4000 and 6000 seconds the fault introduced in
the system is between two different fault models, and in
this case the probabilistic approach has some difculties
to obtain a suitable combination of weights. This kind
of behavior can be modied by changing the degree of
condence of the two techniques, or even by making some
small changes in the initial transition probability matrix.
IV. CONCLUSIONS
A hybrid fault diagnosis and isolation scheme has
been proposed and applied to a heating unit of a Heat-
ing, Ventilation Air Conditioning system. The hybrid
technique combines probabilistic and optimisation based
methods. The results indicate that the introduction of the
probabilistic approach leads to better fault identication.
REFERENCES
[1] J. M. C. S. M. J. G. C. Mendes, J. M. F. Calado and J. M.
G. S. da Costa, Industrial actuator diagnosis using hierarchical
fuzzy neural networks, in European Control Conference, Porto,
Portugal., 2001, pp. 27232728.
[2] S. K. H. Yoshida and Y. Morita, Online fault detection and
diagnosis in vav air handling unit by rarx, Energy and Buildings,
vol. 33, pp. 391401, 2001.
[3] R. J. Patton, in IFAC Symp. on SAFEPROCESS, Hull, UK.
[4] T. A. Johansen and M.-S. Roderick, Multiple-Model Approaches
to Modelling and Control. Taylor & Francis, 1997.
[5] I. K. P .M. Silva, V. M. Becerra and J. M. F. Calado, A multiple-
model approach to model complex nonlinear systems. in 7th
Portuguese Conference on Automatic Control, September 2006.
[6] J. Wang and H. Hu, Vibration-based fault diagnosis of pump using
fuzzy technique, Measurement.
[7] G. G. P. Amann, J.M. Perronne and P. Frank, Identication of
fuzzy relational models for fault detection, Control Engineering
Practice, vol. 9, pp. 555562, 2001.
[8] J. Q. Zhang and Y. Yan, A wavelet-based approach to abrupt
fault detection and diagnosis of sensors, IEEE TRANSACTIONS
ON INSTRUMENTATION AND MEASUREMENT, vol. 50, no. 5,
pp. 13891396, 2001.
[9] Y. Chen and M. Lee, Neural networks-based scheme for system
failure detection and diagnosis, Mathematics and Computers in
Simulation 58 (2002) 101109, vol. 58, pp. 101109, 2002.
[10] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control
- Analysis and Design. John Wiley & Sons, 1996.
[11] S. Kanev and M. Verhagen, Controller reconguration for non-
linear systems, Control Engineering Practice, vol. 8, pp. 1223
1235, 2000.
[12] K. P. C. Edwin and H. Z. Stanislaw, An Introduction to Optimiza-
tion. John Wiley & Sons, 1996.
[13] M. Norgaard and O. Ravn and N. K. Poulsen and L. K. Hansen
Neural networks for modelling and control of dynamic systems
Springer, 2000.
[14] I. K. P .M. Silva, V. M. Becerra and J. M. F. Calado, Multiple-
model fault tolerant control of terminal units of hvac systems, in
Internaional Symposium on Industrial Electronics, July 2006.
[15] Y. Boers and H. Driessen, A multiple model multiple hypothesis
lter for markovian switching systems, Automatica, no. 41, pp.
709716, 2005.
[16] Y.M.Zhang and J.Jiang, Integrated active fault-tolerant control us-
ing imm approach, IEE Transactions on Aerospace and Electronic
Systems, vol. 37, no. 4, pp. 12211235, 2001.
[17] V. P. Jilkov and X. R. Li, Online bayesian estimation of transition
probabilities for markovian jump systems, IEEE TRANSACTIONS
ON SIGNAL PROCESSING, VOL. 52, NO. 6, JUNE 2004, vol. 56,
no. 6, June 2004.
2007 Medterranean Conference on Contro and
Automaton, |uy 27 - 29, 2007, Athens - Greece
T14-009