202
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 1, NO. 3, AUGUST 2005
Identification and Fault Diagnosis of a Simulated
Model of an Industrial Gas Turbine
Silvio Simani, Member, IEEE
Abstract—In this study, a model-based procedure exploiting an-
alytical redundancy for the detection and isolation of faults of a
gas turbine system is presented. The diagnosis scheme is based on
the generation of so-called “residuals” that are errors between es-
timated and measured variables of the process. The work is com-
pleted under both noise-free and noisy conditions. Residual anal-
ysis and statistical tests are used for fault detection and isolation,
respectively. The final section shows how the actual size of each
fault can be estimated using a multilayer perceptron neural net-
work used as a nonlinear function approximator. The proposed
fault detection and isolation tool has been tested on a single-shaft
industrial gas turbine model.
Index
Terms—Fault diagnosis, gas turbines, identification,
Kalman filtering, observers.
I. I
NTRODUCTION
T
HE control devices currently in use to improve the overall
performance of industrial processes involve both sophisti-
cated digital system design techniques and complex hardware
(sensors, actuators, processing units). The complexity means
that the probability of fault occurrence can be significant and an
automatic supervisory control system should be used to detect
and isolate anomalous working conditions as early as possible.
This study is motivated by the challenge of developing di-
agnostic tools for gas turbine systems to indicate the onset of
developing faults and provide a mechanism for predictive main-
tenance requirements. Although the gas turbine is a complex
dynamical system we show that the model-based approach to
fault detection and isolation (FDI) can work very well under
realistic fault conditions.
Two main aspects of the proposed methodology should be
underlined
. Firstly, the system complexity does not indicate a
requirement for a complex physical or thermodynamic model.
This is because for FDI input–output model identification
methods are used, thus obviating the requirement for phys-
ical models. A linear mathematical model (state-space or
input–output descriptions) of the input–output links are, in fact,
obtained by means of identification schemes. Auto Regressive
eXogenous (ARX) models can be used in connection with the
least square method. On the other hand, Errors-In-Variables
(EIV) models are used according to Kalman [1], [2]. In the latter
approach the identification technique is based on the rules of the
Frisch scheme, based on traditional application to the analysis
of economic systems [3]. This approach gives a reliable model
of the plant under investigation, as well as providing variances
Manuscript received September 20, 2004; revised January 18, 2005.
The author is with the Department of Engineering, University of Ferrara,
44100 Ferrara, Italy (e-mail: ssimani@ing.unife.it).
Digital Object Identifier 10.1109/TII.2005.844425
of the input-output noises [4]. Both of these approaches have
been used in this study. Secondly, in this study linear prototypes
for the design of linear output estimators [5]–[10] have been
developed instead of using complicated nonlinear models.
Non-linear modeling is far from a straightforward subject,
even if there is a steady increase in research in this subject.
Regardless of the surge in interest in nonlinear systems, the
linear approach to fault diagnosis is still advantageous in terms
of solution complexity and performance. This is especially true
if so-called robust solutions are sought, where the robustness is
used to minimize the effects of modeling errors.
The problem of fault detection and isolation (FDI) in dy-
namic processes has received great attention during the last two
decades and a wide variety of so-called model-based approaches
have been proposed [11]–[14]. Model-based methods all use
mathematical models of the plant being monitored. However,
the conceptual realization of these models can vary according to
the following approaches: the parity space [15], state estimation
[16]–[21], the fault detection filter [19], [22], [23] and param-
eter identification [11], [16], [18]. In each case, to guarantee that
faults can be detected and isolated (and distinguished one from
another), mathematical models of the process under investiga-
tion are required, either in state space or input-output or transfer
function form. State space descriptions generally provide math-
ematically rigorous tools for system modeling and residual gen-
eration that may be used in fault detection of industrial systems,
both for the noise-free case and the noisy environments. Resid-
uals should then be processed to detect an actual fault condition,
rejecting any false alarms caused by noise or spurious signals.
This work aims to define a comprehensive methodology for
fault diagnosis by using a state estimation approach, in con-
junction with residual processing schemes, including a simple
threshold detection. This work describes how this is achieved
in the noise-free case as well as the noisy case using statistical
analysis tools, when the data are affected by noise. The complete
procedures of model identification, residual generation and fault
identification and isolation have been tested on a single-shaft in-
dustrial gas turbine prototype. The study provides a description
of extensive simulation results.
II. P
LANT
M
ODEL
D
ESCRIPTION
In the following it is assumed that the monitored system, de-
picted in Fig. 1(a), can be described in fault-free condition, by a
linear, discrete-time, time-invariant, dynamic model of the type:
(1)
1551-3203/$20.00 © 2005 IEEE
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
SIMANI: IDENTIFICATION AND FAULT DIAGNOSIS
203
where
is the state vector,
the process
output vector and
the control input vector.
,
, and
are constant matrices of appropriate dimensions obtained by
means of modeling techniques or identification procedures.
Under fault-free conditions, the input and the output link of
the sensors can be described by the following relation:
(2)
In real applications, variables
and
represent noises
which, due to technological reasons, affect sensor behavior.
They are generally described as white, zero-mean, uncorrelated
Gaussian noises. It is assumed that
and
are the only
available measurements from the real process.
The vectors
and
model input and output sensor faults, respectively. The scheme
shown in Fig. 1(a) describes the relations among the actual
sensor inputs
and
, the sensor faults
and
and the sensor outputs
and
.
According again to Fig. 1(a), when a component fault
occurs in the plant described by (1), the dynamic system will
be modeled as
(3)
A fault
may also occur on the regulator in the control
loop. In such a case, under the assumptions that
and
, the link among the output
of the regulator, its
input
and the controller fault
will be modeled as
(4)
where
represents the input–output behavior of the con-
troller.
Usually
,
,
, and
signals are described by
step and ramp functions representing abrupt and incipient faults
(bias or drift), respectively.
Under fault-free assumptions, representations of types (1) and
(2) are known as EIV models.
The design of state observers and Kalman filters requires the
knowledge of a state-space model of the system under investiga-
tion. When classical modeling techniques cannot be used since
the complete physical knowledge of the system is not available
or the model parameters are unknown, a black-box identifica-
tion approach has to be considered.
III. E
QUATION
E
RROR
M
ODELS
Equation Error (EE) models belong to an identification model
class that can be successfully exploited in this study and, in par-
ticular, different equation error models can be extracted from
the data. A specific discrete-time, time-invariant, linear dynamic
model, e.g. ARX or ARMAX (Auto Regressive eXogenous or
Auto Regressive Moving Average eXogenous) [24], [25], can
be selected only inside an assumed family of models.
On the other hand, the Frisch scheme [3] can be also applied
to perform the EIV dynamic system identification, in particular
when noise signals affecting the data have to be estimated [4].
Fig. 1.
(a) Monitored system and (b) the logic diagram of the fault detection
system.
Such a scheme allows to determine the linear discrete system
which has generated the noisy sequences as well as the variances
of the noises
and
corrupting the data [4]. In the ideal
Frisch scheme hypotheses, these signals are assumed zero-mean
white noises, mutually uncorrelated and uncorrelated with every
component of
and
.
In particular in this work, the input–output link will be
mathematically described by performing the identification of a
number of ARX Multiple-Input Single-Output (MISO) models
of the type
(5)
equal to the number
of the output variables has been per-
formed. The order
and the parameters
and
, with
, of the model have to be determined by the iden-
tification approach. The term
takes into account the mod-
eling error, which is due to process noises, parameter variations,
etc.
The next step is the transformation of input-output discrete-
time time-invariant linear models (5) into state-space represen-
tations. The state-space systems obtained by the equation errors
models are useful to design dynamic observers, whilst the ones
coming from the Frisch scheme can be used in order to build
Kalman filters.
It can be proved that a state-space formulation of the input-
output equation error model (5) in fault free conditions, for the
th output
becomes:
(6)
where the matrices
,
,
,
and
are functions of the order and the
and
parameters [26].
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
204
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 1, NO. 3, AUGUST 2005
A. ARX Model Identification
Consider an assumed order for a SISO (Single-Input Single-
Output,
) ARX model (5) and the input-output sequences
observed in the time interval
. If the model
(5) is used to compute predicted output values
in the
times, for a given set of parameters:
(7)
the mean square prediction error
is given by
(8)
By introducing now the following Hankel matrices
and
:
..
.
. ..
..
.
(9)
and
..
.
. ..
..
.
(10)
it follows that
..
.
(11)
It can be proved that the parameter vector minimizing the cost
function (8) is given by
..
.
(12)
where
denotes the pseudo-inverse of the
matrix. The
algorithm gives an estimate
of , which converges asymptoti-
cally to the real parameter of the process that has generated the
data.
To estimate the order
of the ARX process, an integer
and the
matrix of input–output samples given
by
(13)
are considered.
If
in (5) the following properties hold:
for;
for;
(14)
It could be possible to consider the increasing sequence of
matrices:
(15)
where
and to evaluate their singularity. The first
singular matrix
would define the correct order for the model
. Unfortunately, the presence of
in (5) leads
to the nonsingularity of every matrix in (15).
It can be proved that if
is large enough, an estimate of the
standard deviation
of the process
in (5), is given by [27]
(16)
If the following, the quantity
is defined:
(17)
it can be shown that
for
,
for
and
for
. In other words, if
is large
enough, a sequence of decreasing values of
followed by a
stabilization once the correct order is reached, can be noted. The
criterion can be used to evaluate a suitable order or, at least, an
interval of admissible orders for the model before computing its
parameters. It is easy to show that
for an ARX
model with order
and parameters
given by (12).
If the value (17) is expressed as percentage of the standard
deviation of the measured output, the Predicted Per Cent Re-
construction Error criterion (PPCRE) is obtained [28]. The
PPCRE
gives the prediction error of an ARX model of order
without requiring any computation of its parameters and
predictions. The application of the PPCRE criterion consists
in computing an increasing sequence of PPCRE
(or
)
and in selecting the minimal order that, once increased, does
not lead to a significantly better performance. Relation (17) can
also be used in the application of the well-known FPE, AIC and
MDL order estimation criteria [26].
B. EIV Miso System Identification
In this section, the Frisch scheme procedure for the identifi-
cation of dynamic EIV MISO system from input-output
,
(with
) noisy (fault-free) sequences will
be summarized.
A finite sequence of the variables
observed with a constant sampling interval is considered. If dy-
namic linear relations exist among these variables, they can be
described by models of the type
(18)
which describe linear MISO (multiple-input, single output) dis-
crete-time systems whose order is
and whose parameters are
and
.
At first, the following problem is presented.
Problem 1 (Realization):
Given a noiseless input–output se-
quence
generated by a system of type
(18), determine the order
and the parameters
and
of
the system.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
SIMANI: IDENTIFICATION AND FAULT DIAGNOSIS
205
The following vectors and matrices can be defined:
(19)
(20)
(21)
(22)
(23)
(24)
(25)
where
is assumed large enough to solve the problem consid-
ered. Please note that the matrices
and
have the
same number of rows.
The matrix
is partitioned as follows:
..
.
..
.
. ..
..
.
(26)
To solve the realization problem it is possible to consider the
sequence of increasing-dimension matrices:
(27)
testing their singularity. As soon as a singular matrix
is
found then the order
and the parameters
,
describe the dependence relation-
ship of the th vector of
on the remaining ones.
In Problem 1 it has been assumed that
is large enough to
avoid unwanted linear dependence relationships due to limita-
tions in the dimension of the involved vector spaces; this means
. The minimal number of samples must be
therefore equal to
. If a lower number of samples is
available then only a partial realization problem can be solved.
In the noisy case the following identification problem can be
proposed.
Problem 2 (Identification):
Given a noisy input–output
sequence
unequivocally determine, if
possible, the order
and the parameters
and
of a model
(18) of the system which has generated the noiseless sequences
.
Note that in presence of noise the procedure described for the
solution of Problem 1 would obviously be useless since matrices
would always be nonsingular.
In the ideal Frisch scheme it is normally assumed that
(28)
where every noise term
and
is independent of every
other term
and
only and are known. Without loss
of generality, all the variables may be assumed as having null
mean value. Consequently the generic positive definite matrix
associated with the input-output noise-corrupted sequences
may always be expressed as the sum of two terms
(29)
where
(30)
since no correlation has been assumed among the noise samples
at different times. This condition is verified for additive white
noise with variance
and
on the input–output
sequences.
Problem 3:
Given a sequence of increasing-dimension
symmetric positive definite covariance
matrices:
(31)
find, for each , all diagonal nonnegative definite matrices
such that
(32)
It is worth observing now that, unlike the algebraic case, for
each
the noise space is always
, while the parameter
space is
. It can be noted that for each
the solu-
tion set of the previous relation describes, in the first orthant of
the
hyper-plane, a hyper-surface whose con-
cavity faces the origin [4].
Previous results hold for every value of . Since determina-
tion of the system order requires the increasing values of to be
tested, it is relevant to analyze the behavior of the associated
curves when varies. This corresponds to a comparison of the
admissible solution sets for different model orders. In this con-
text the following result can be proved [4].
Theorem 1:
The solution sets of condition
for dif-
ferent values of
are noncrossing curves [4].
It is also important to observe that, since we assume that a
system (18) has generated the noiseless data, for
all the
hyper-surfaces of type
have necessarily at least one
common point, i.e. the point
corresponding
to the true variances
and
of the noise affecting the in-
puts and the output of the system, respectively. The search for
a solution for the identification problem can thus start from the
determination in the noise space of this point.
The following considerations can now be stated. With refer-
ence to the diagonal nonnegative definite matrices:
(33)
the following properties hold.
•
If
the matrices
are positive definite.
•
If
the dimension of the null space of
and,
consequently, the multiplicity of its least eigenvalue, is
equal to
.
•
For
, the matrix
is characterized by a linear
dependence relation among its
vectors
and the coefficients which link the th vector of
to
the remaining ones are the parameters
and
, with
and
, of the system (18) which
has generated the noiseless sequences.
•
For
all linear dependence relations among the
vectors of the matrix
are characterized by the same
coefficients
and
.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
206
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 1, NO. 3, AUGUST 2005
Fig. 2.
Monitored system.
If
models of the type (18) are used to describe the mathe-
matical behavior of a multivariable dynamic system with
in-
puts and
outputs
, the previous identification procedure
must be repeated
times. At every step the identification pro-
cedure must lead to the same values for the input noise variances
.
It is worth noting how this approach cannot be applied
immediately in the identification of real processes, since the
hypotheses on the linearity, finite dimensionality and time
independence of the system and on the additivity and whiteness
of the noise are not usually verified, so that the hyper-surfaces
have no common point for
. The definition
of a suitable criterion of model selection in such cases was
suggested in [29].
IV. R
ESIDUAL
G
ENERATION FOR
FDI
The problem treated in this work regards the detection and
isolation of the faults on the basis of the knowledge of the
measured sequences
and
. The structure of the fault
detection device is depicted in Fig. 1(b). The symptom (or
residual signal,
) generation is implemented by means of
dynamic observers or Kalman filters, driven by
and
,
in order to produce a set of signals,
, from which it will be
possible to isolate faults associated to actuators, components
and sensors. The symptom evaluation refers to a logic device
which processes the redundant signals generated by the first
block in order to estimate and unequivocally identify a fault
occurrence.
This work will present a FDI technique to elaborate a set of
symptoms from which it will be possible to unequivocally de-
tect faults. With reference to Fig. 1(b) the symptom signals are
differences between estimated signals (given by observers or
Kalman filters) and the actual ones supplied by the input and
output sensors.
Moreover, it is assumed that only a single fault may be present
in the actuators, components or input sensors of the plant at any
given time. On the other hand, multiple output sensor faults can
be handled.
V. N
OISE
-F
REE
F
AULT
D
IAGNOSIS
The aim of this study consists in finding a procedure in order
to detect and isolate faults on actuators, components and sen-
sors of single-shaft industrial gas turbine. The model of such
a turbine was developed in SIMULINK environment [30] and
it was supplied by the ALSTOM Power Technology Centre,
Whetstone, U.K. Fig. 2 shows the gas turbine layout as well as
its inputs and outputs.
The time series of data used to identify the models were
generated with a nonlinear dynamic model in SIMULINK
environment and they simulate measurements taken on the
machine with a sampling rate of 0.08 s and without noise
(
and
) due to measurement uncertainty
which, instead, is always present in the real measurement
systems. The nonlinear SIMULINK model of the gas turbine was
validated in steady state conditions against engine measurements
where available, and against the prediction of a more rigorous
steady state gas turbine model where measurements were not
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
SIMANI: IDENTIFICATION AND FAULT DIAGNOSIS
207
TABLE I
T
URBINE
C
ONTROL
I
NPUTS
available. The SIMULINK model variables were found to be
within 5% of the measured and rigorous modeled values. For
the majority of variables the accuracy was within 1%. In the
dynamic case no model validation has been carried out as yet.
Orders and output reconstruction errors of each ARX model
are shown in Table II. The th model (with
and
monitored outputs) is driven by
and
and
gives the prediction of the th output
. The control inputs
are
and they are summarized in Table I.
Table I also reports measurement accuracy that depend on mea-
surement sensors exploited by the turbine prototype.
Even if, according to Fig. 2 the measurements of ambient and
pressure temperature (
and
) are inputs for the turbine, they
were not considered, since they are constant all the times.
The monitored outputs of the turbine prototype are
. Some of them (the most
meaningful that will be considered for the FDI task) are sum-
marized in Table II. Table II shows also the measurement
accuracy, due again to measurement sensors used by the turbine
prototype.
Each model was tested in different operating conditions and
it has always provided an output reconstruction mean square
error (MSE) lower than 0.5%. Moreover, two time series of data
generated by the gas turbine nonlinear model were exploited in
order to validate the ARX models. These models have always
provided in full simulation an output reconstruction error lower
than 1%.
A very effective way of evaluating the adequacy and flexi-
bility of identified models consists, in fact, in their use for per-
forming complete simulations (i.e. using only the initial samples
of the observed outputs) and in comparing the obtained predic-
tions with observed output samples. This procedure, that can
be applied when a single set of data is available, gives the best
results when applied to sequences different from those used to
identify the model. The mean square prediction error between
the observed outputs and the ones obtained by simulation can be
used to compare models with different orders. The reconstruc-
tion errors of each ARX model are summarized in Table III.
In the following, the FDI problem for the case of noise-free
measurements is solved firstly via the implementation of a bank
of output dynamic observers.
VI. S
IMULATED
F
AULT
C
ONDITIONS
Four gradually developing faults that may affect the turbine
prototype can be represented as follows.
1) Compressor contamination (core engine performance de-
terioration),
.
2) Thermocouple sensor fault (output sensor failure),
.
TABLE II
T
URBINE
O
UTPUTS AND
MISO ARX M
ODEL
C
HARACTERISTICS
TABLE III
D
YNAMIC
ARX M
ODEL
V
ALIDATION
3) High Pressure turbine seal damage (core engine perfor-
mance deterioration),
.
4) Fuel actuator friction wear (controller fault),
.
Note that in real industrial applications it is commonplace for
each of the above faults to develop slowly over a period of
months. For the purpose of this simulation—in order to avoid
excessively long duration simulations—the fault development
rate will be increased so that significant effects are present after
one hour. However this is still considerably longer than the du-
ration of the gas turbine dynamics which occur over periods of
seconds—a factor which must be taken account of in any FDI
algorithm design.
In the presence of a fault condition, the challenge for the de-
signer of an FDI algorithm may be summarized as follows.
1) Detect that a fault condition exists with minimum delay
from the initial occurrence of the fault.
2) Identify the nature, magnitude and location of the fault,
again with minimum delay from the initial occurrence of
the fault.
Note that it is desirable to avoid introducing perturbation signals
onto the model variables. In the first instance an FDI design
should be based upon data which is available from the normal
day to day operation of the plant, for example during transient
and over prolonged periods of steady state operations.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
208
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 1, NO. 3, AUGUST 2005
The rate of development and magnitude of faults have been
set to nominal values in this case study. It will be of interest to
know how small the fault parameters can be made whilst still
maintaining good FDI performance.
A. Case 1: Compressor Contamination (Core Engine
Performance Deterioration)
Failure “case 1”, represents fouling of the surfaces of the
compressor blades, this reduces air flow, changes the blade aero-
dynamics and consequently changes the surface roughness. The
failure is modeled as a gradual decrease in mass flow rate for a
given pressure ratio. The fault signal
affects the monitored
system of Fig. 1(a) by means of the SIMULINK submodel rep-
resented in Fig. 3(a). The maximum decrease in mass flow rate
is set nominally at 5% while the fault development rate is set to
(5% decrease of normal flow rate)/hour.
In order to design the component
FDI scheme
(
,
and
), with respect to the
turbine SIMULINK model, the subsystem depicted in Fig. 3(b)
was considered.
The inputs for the subsystem are
,
,
, and
, while
,
, and
are the outputs directly affected by the fault
.
It was experimented with the turbine prototype model that the
most sensitive output signal to a ramp fault
is
output
measurement. It is worthy to note how the shape of
tran-
sient is determined by the input variation, not by the slowly de-
veloping compressor hauling fault
(which is also of very
small magnitude).
signal and
fault have, in fact, very
different magnitudes.
A second order
ARX MISO (
,
)
model was identified with an output reconstruction error
. The parameters of such a model, driven
by
and
signals, are represented by the vector
,
.
The diagnosis of the
signal (linked to the faulty turbine
component) requires the knowledge of the triple
with
and the identification of an ARX model with
two inputs which gives the prediction of the 13th output
.
Because of the noise-free case, in which
,
in
the system (6) the term
was neglected
.
In these noise-free conditions, the detection of a fault re-
garding the compressor was hence performed by using the
classical output observer configuration exploited for the FDI of
output sensor faults, as depicted in Fig. 4(a). The inputs
,
and the output
feed the observer to estimate the
signal
itself and to generate the residual
. The poles
of the output observer for the signal
were chosen near
0.5 according to the minimization of the function
. The
eigenvalues
, in fact, were chosen to maximize
the mean square error of the residual sensitivity
to
a fault and minimize the mean square error of the residual
in fault-free condition,
. The minimum
,
where
is the cost function
(34)
Fig. 3.
(a) Fault “case 1” SIMULINK submodel and (b) the monitored
subsystem.
Fig. 4.
(a) Observer scheme and (b) residual functions.
has to be found. The minimization of the cost function can
be performed by exploiting the Optimization Toolbox of
MATLAB, or even better, the Genetic Algorithm Optimization
Toolbox (GAOT) of MATLAB [31].
Finally, Fig. 4(b) shows the fault-free (solid line) and the
faulty residuals (dotted line).
B. Case 2: Fault Diagnosis of the Output Sensor
Failure “case 2” represents the malfunctioning of a ther-
mocouple in the gas path leading to a slowly increasing or
decreasing reading over time. The considered fault regards
the measurement of the
turbine output variable that mainly
affects the output sensor by means of the SIMULINK model
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
SIMANI: IDENTIFICATION AND FAULT DIAGNOSIS
209
Fig. 5.
(a) SIMULINK fault model of the output sensor and (b) the residual
function
r(t).
depicted in Fig. 5(a). There is no limit placed on the error
magnitude while the fault development rate is set to (5% error
in measuring actual temperature)/hour.
In order to diagnose a single
fault on the th output
sensor (
,
,
and
) when the
measurement noises are negligible (
and,
),
with reference to system (6), the model of the th output observer
has the form:
(35)
where
is the th observer state vector,
represents a
fault on the th output sensor and the triple
is a
minimal state-space representation (completely observable) of
the link among the inputs of the process and its th output
.
In the absence of faults, it can be verified that, for the th
output, the residual
is equal to zero. In the presence of a fault on the th output sensor
the th output residual reaches a value different from zero and
this situation leads to a complete failure diagnosis.
In particular, the diagnosis of the
output sensor (thermo-
couple fault), represented in Fig. 5(a), requires the knowledge
of the triple
with
and therefore the
identification of an ARX model with two inputs which gives
the prediction of the 18th output
. A second order ARX
MISO model (
and
), driven by
and
input signals, was identified. Such a model gives an
output reconstruction error equal to
. The pa-
rameters of the ARX model are described by the vector
,
.
The poles of the output observer, whose scheme is similar to
the one depicted in Fig. 4(a), were chosen near 0.3 in order to
minimize the function
.
An incipient fault (drift) was generated in the output sensor of
the SIMULINK model by adding a ramp function with a slope
of
to the
output signals. Moreover, it was de-
cided to consider a fault during a transient since, in this case,
the residual error due to ARX model approximation is max-
imum and therefore it represents the most critical case. The fault
occurring on the single sensor causes alteration of the sensor
signal and of the residuals given by the output observer using
this signal as input. These residuals of Fig. 5(b) indicate a fault
occurrence when their values are lower or higher than the thresh-
olds fixed in fault-free conditions.
The fault detection logic scheme is obtained by comparing
the fault-free
and faulty
residual
obtained from the difference between the values computed by
the observer related to the output
and the ones given
by the sensor. Obviously, the non zero value of the residual in
Fig. 5(b) is due to the identified ARX model approximation.
The considered drift (ramp fault) starts at the instant
.
Since the observer gives the estimate
of
at the instant
by using measurements available from the instant
to
, a fault occurring at the instant
affects only
.
This change can produce an instantaneous peak in the monitored
residual function [5]. In such a case, the peaks are not due to in-
stantaneous changes in the input signals, e.g. fuel flow
or
butterfly valve position
. Thus, they may be used as incip-
ient detector of anomalous behavior of the output sensors.
C. Case 3: High Pressure Seal Damage (Core Engine
Performance Deterioration)
Failure “case” represents failure
of an HP turbine seal.
This results in a reduction in turbine efficiency. The fault is
modeled as a gradual reduction in turbine efficiency over time.
The maximum decrease in turbine efficiency is set nominally at
5% while the fault development rate is set to (5% reduction of
normal efficiency)/hour.
In order to detect such a fault, an output observer fed by
the inputs
,
and
is designed. Moreover,the
SIMULINK subsystem used to inject the component fault
into the monitored systems is depicted in Fig. 6(a).
Under noise-free conditions, with reference to system (6)
and neglecting
, the output observer was designed
for a third order MISO model which gives a mean square
reconstruction error equal to
and the eigen-
values were chosen near 0.3 to minimize the cost function
. The ARX parameter vector estimated is described
by
,
. The scheme used to generate the redundant
residual regarding the
output signal is equal to the one
designed for the previous fault cases. The fault free and the
faulty residual are also shown in the Fig. 6(b).
D. Case 4: Fuel Actuator Friction Wear
Failure “case 4”,
represents the loss of performance
due to wear of the fuel valve actuator. Again, the SIMULINK
model used to generate the fault signal
is depicted in
Fig. 7(a). As there are no specific actuator dynamics in the cur-
rent model, the wear effect of the valve actuator causing slower
response to demanded flow rates is modeled as a simple first
order lag on the resulting fuel flow. The time constant increases
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
210
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 1, NO. 3, AUGUST 2005
Fig. 6.
(a) Fault SIMULINK subsystem and (b)
p (t) signal residuals.
linearly with the time to represent progressive wear damage to
the actuator.
In order to generate the residual for the diagnosis of the actu-
ator fault
, an output observer scheme was exploited again.
In particular, the inputs of the turbine, the fuel flow,
, the
valve angle,
and the outputs
,
,
,
and
were considered. The speed demand,
, one of the
inputs of the governor and
, the third output of the turbine
were also shown.
For each output, a third order
ARX model with
two inputs and one output (
,
) was identi-
fied. By means of the SIMULINK system (2), a single fault
was simulated and the most sensitive output to a fault
regarding the actuator was determined. The
residual
was the most sensitive, with a
.
The third order ARX parameter are collected in the param-
eter vector
,
.
The observer scheme exploited for generating the residuals
regarding the signal
is fed by the signals
,
,
and
. The signals are therefore used by the dynamic ob-
server to estimate the
signal itself. The effects of the fault
on the symptom signal
is shown in Fig. 7(b).
is the
residual concerning the
output measurement in fault-free
and faulty conditions. Because of the closed-loop configuration
of the subsystem considered in Fig. 7(a), the fault shape can not
be described by using a ramp function. The output observer fed
by inputs
and
was designed to estimate
. The
eigenvalues
were chosen near 0.4 to minimize the cost func-
tion
.
As an example, Fig. 7(b) shows how the fault occurring on
the single sensor causes alteration of the input and output sig-
nals and of the residuals given by the output observer using the
signal as input. These residuals indicate a fault occurrence
Fig. 7.
(a) Actuator fault SIMULINK model and (b) residuals in the presence
of
f (t) fault.
TABLE IV
F
AULT
S
IGNATURE
when their values are lower or higher than the thresholds fixed in
fault-free conditions. Fig. 7(b) shows the fault-free (solid line)
and faulty (dotted line) residual
obtained from the differ-
ence between the values computed by the observer related to
the output
and the ones given by the sensor.
VII. F
AULT
I
SOLABILITY
By performing residual sensitivity analysis, i.e. by selecting
the most sensitive residuals to the faults, the Table IV is ob-
tained, in order to isolate different fault occurring at the same
time.
In order to summarize the FDI capabilities of the presented
schemes, Table IV shows the “fault signatures” in case of a
single fault in each actuator, component and sensor. The resid-
uals which are affected by faults are marked with the presence of
“1“ in the correspondent table entry, while an entry “0“ means
that the fault does not affect the correspondent residual. The
bold face entries in the table represent the residuals affected by
the same faults. The italics “1“s are the distinguishable resid-
uals (they are bigger than a fixed threshold). Note how multiple
faults in actuator, components and sensor can be isolated since
a fault affects only the residual function of the observer driven
by the same output.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
SIMANI: IDENTIFICATION AND FAULT DIAGNOSIS
211
VIII. FDI
IN THE
N
OISY
E
NVIRONMENT
In this section, a FDI technique based on Kalman filters de-
signed in the case of noisy measurements is presented. Such a
design is enhanced by processing the noisy data according to the
Frisch scheme identification method [5], [8], [10]. Moreover,
fault size estimation can be performed by means of different
neural network architectures [8], [10]. In particular, neural net-
works can be used as function approximators to estimate single
sensor fault size. The proposed fault diagnosis tool when mea-
surements are affected by noise is tested on the power plant pre-
sented in the previous sections.
In recent years, neural networks have been exploited success-
fully in pattern recognition as well as function approximation
theory and they have been proposed as a possible technique for
fault diagnosis, too. Neural networks can handle nonlinear be-
havior and partially known process.
The aim of this paragraph is to suggest how artificial neural
networks can be exploited to approximate a large class of func-
tions, for fault diagnosis of an industrial plant. In particular,
the problem of the estimate of the slope of faults concerning
actuators, components and output sensors of an industrial
gas turbine can be solved. Faults modeled by ramp functions
create changes in several residuals obtained by using dynamic
observers (Kalman filters) of the process under examination.
A neural network can be used in order to find the connection
from a particular fault regarding input and output sensors to a
particular residual. Residuals are dependent only on sensors
faults. Therefore, the neural network evaluates patterns of
residuals, uniquely related to particular fault conditions [10].
A. Fault Estimation Device
The fault detection and diagnosis system produces and elabo-
rates a set of residuals from which it will be possible to estimate
the amplitudes of the faults regarding actuators, components and
input-output sensors.
With reference to Fig. 8 the symptom generator is designed
to produce a set of signals which are somehow redundant.
These signals are differences between estimated signals given
by Kalman filters and the actual ones supplied by the sensors.
In order to experiment with learning capabilities of artifi-
cial neural networks, on which the diagnosis device in Fig. 8
is based, a bank of classic Kalman filters are used. The number
of filters is equal to the number
of system outputs, and each
filter is driven by a single output measurement and all the in-
puts of the plant. Because of this configuration, the diagnosis
of faults is indeed very easy, since each output measurement is
directly connected to a single residual generator. The basic prin-
ciple of fault detection by using Kalman filtering is illustrated
in Fig. 9.
With reference to the time-invariant, discrete-time, linear dy-
namic system described by a minimal state-space realization
of the input-output MISO system (6) [26], the th
Kalman filter has the structure [32]:
(36)
Fig. 8.
Logic diagram of the fault detection system.
Fig. 9.
Bank of Kalman filters and NNs for residual generation and estimation.
TABLE V
F
RISCH
S
CHEME
M
ODEL
R
ECONSTRUCTION
E
RRORS
The variable
is the one step-ahead prediction of the
state
,
is the estimate of the th component
of the
output
given by the filter. A Riccati equation is used to com-
pute the time-variant gain
of the filter by means of the knowl-
edge of the covariance matrix of the input vector noise
and
the variance of the th component of the output noise
. It
can be proved that the innovation
is
a white process when all the assumptions regarding the system
and the statistical characteristics of the noises are
completely fulfilled. In particular, the innovation converges to a
steady state solution when the pair
is completely ob-
servable and the pair
is completely reachable.
B. Identification Procedure
The Frisch scheme can be applied to perform the dynamic
system identification of the plant [1]–[4], [33]. Such a scheme
allows to determine the linear discrete dynamic system which
has generated the noisy sequences as well as the variances of
the noises
and
corrupting the data. In the ideal as-
sumptions of the Frisch scheme, these signals are assumed white
noises, mutually uncorrelated and uncorrelated with every com-
ponent of
and
.
Table V summarizes the reconstruction errors concerning the
MISO models in the form (18) with two inputs (
and
)
and each monitored output variable, as output.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
212
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 1, NO. 3, AUGUST 2005
TABLE VI
F
RISCH
2-
ND
O
RDER
M
ODEL
P
ARAMETERS
TABLE VII
F
RISCH
N
OISE
V
ARIANCES
Tables VI and VII collect parameters of second order models
as well as the input and output noises.
On the basis of the data collected in Tables VI and VII, four
Kalman filters with two inputs
and one output
were designed. The detection strategy which is commonly
chosen in connection with Kalman filtering methods for failures
detection, consists in monitoring the innovations
. Because
of the linear property of system (1) and because of the effect of
faults on the system output measurements, any change in mea-
surements due to a fault is reflected in a change in the mean and
in the standard deviation of
.
In particular, since the Kalman filter produces zero-mean and
independent white residuals with the system in normal oper-
ation, a method for failure detection and isolation consists in
testing how much the sequence of innovations has deviated from
the white noise hypothesis. The tests which can be performed
on the innovations are the usual ones for zero-mean and vari-
ance, as cumulative sum algorithms as well as independence, as
-type. If a system abnormality occurs, the statistics of
change, so its comparison with a threshold fixed under no faults
conditions, becomes the detection rule.
In Figs. 10(a), 10(b), (11a), and (11b) the examples of the
turbine FDI performed by using the residual generated by the
Kalman filter with two inputs and one output are shown.
In particular, when the fault
affects the
residual
(“case 1”), Fig. 10(a) depicts fault-free and faulty residuals gen-
erated by the Kalman filter. It is driven by the input sensor signal
,
and the
signal itself. A ramped incipient com-
pressor fault (“case 1”), commencing at
causes a
change in the value of the
residual computed in fault-free
condition, as depicted in Fig. 10(a). It is important to note that,
in order to achieve the maximal fault detection capability, the
residual corresponding to the most sensitive filter to a failure on
the
measurement was selected, in accordance with Table V.
On the other hand, the
incipient ramp fault (“case 2”)
affects the output sensor for the measurement of the
signal
and commencing at the instant
s. Hence, in Fig. 10(b)
fault-free and faulty residuals regarding the
signal obtained
from the difference between the values
computed by the
Kalman filter in (36) with
and the ones measured by the
Fig. 10.
Kalman filter residuals
r(t) for the (a) “Case 1” and (b) “Case 2”.
sensor, are shown. Obviously, the nonzero value of the residual
in fault-free conditions is due to the ARX model approximation
and to the actual measurement noise.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
SIMANI: IDENTIFICATION AND FAULT DIAGNOSIS
213
Fig. 11.
Kalman filter residuals
r(t) for the (a) “case 3” and (b) “case 4”.
Fig. 11(a) shows the behavior of the residual when a ramped
incipient fault
(“case 3”) commences at the instant
. According to Table V, in this case,
is the moni-
tored signal for the FDI of a component of the turbine. Then,
Fig. 11(a) depicts the fault-free residual and its change due to
the fault occurrence, as the previous cases.
Finally, Fig. 11(b) shows the change in the fault-free residual
concerning the
measurement due to a ramped
incipient actuator fault (“case 4”). The
fault commences
at the instant
s. In Fig. 11(b), the fault-free and the faulty
residuals are shown.
Because the nature of the incipient ramp fault
affecting
the regulator into the feedback control loop, the output measure-
ments affected by the fault itself are different from ramp signals.
IX. M
INIMAL
D
ETECTABLE
F
AULTS
Tables VIII and IX summarize the performance of the fault
detection and isolation technique both in the noise-free and
noisy environments. The Table collect the minimal detectable
TABLE VIII
M
INIMUM
D
ETECTABLE
F
AULTS BY
M
ONITORING
R
ESIDUAL
AND
I
NNOVATION
V
ALUES
TABLE IX
M
INIMUM
D
ETECTABLE
F
AULTS BY
M
ONITORING
R
ESIDUAL
AND
I
NNOVATION
V
ALUES
Fig. 12.
Detection delay definition.
fault on the four measurements, in case the residual or inno-
vation value is monitored using a geometrical test and fixed
thresholds. The minimal detectable fault values in Tables VIII
and IX are expressed as percentage of the signal values and are
relative to the case in which the occurrence of a fault must be
detected as soon as possible.
It results that the values of the faults obtained by using geo-
metrical analysis on Kalman filter innovations, collected in Ta-
bles VIII and IX, are different than the ones reported in the same
Table and computed in the noise-free environment exploiting
classical observers. Tables VIII and IX show how faults mod-
eled by ramp functions may not be immediately detected, since
the delay in the corresponding alarm normally depends on fault
mode.
The minimal detectable fault can be found by fixing a detec-
tion delay, defined in Fig. 12. If a delay in detection is toler-
able the amplitude of the minimal detectable fault is lower. The
minimal detectable faults on the various sensors seem to be ad-
equate to the industrial diagnostic applications, by considering
also that the minimal detectable faults can be reduced if a delay
in detection promptness is tolerable.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
214
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 1, NO. 3, AUGUST 2005
X. S
OFTWARE
I
MPLEMENTATION
This section tries to describe shortly how the methods
described in the previous paragraphs could be implemented
in a real-world applications by means of the MATLAB and
SIMULINK software.
In particular, the System Identification Toolbox (of the
MATLAB or SIMULINK environments) provides the fol-
lowing tools.
•
It is possible to define arbitrary parameterizations of the
matrices in the input-output state-space form models used
in the previous sections. This MATLAB implementation
is able also to describe state-space or input-output models
with unknown parameters, i.e., the so-called idgrey
model on-line identification. The procedure, in connec-
tion with PEM (Prediction Error Methods, in MATLAB,
the function pem) techniques can be exploited for on-line
identifying a linear state-space model that is able to
describe the turbine model.
•
Once the parameterization of the state-space or input-
output forms (sspar) is defined, a basis for the cor-
responding realization is automatically selected to give
well-conditioned calculations. An alternative is to specify
an observer canonical form
'
' for the
system matrices. This is a grey-box model, since a struc-
ture of the uncertainty can be defined and canonical form
can be forced for estimating all models of a certain order.
•
The structure estimations can all be combined at the esti-
mation call pem(Data, m, ‘sspar’, ‘can’, ‘z’), which
is the same as set(m, ‘sspar’, ‘can’, ‘z’). A model m
is identified from the available data Data.
The overall procedure implemented in MATLAB and
SIMULINK used for describing the developed FDI method is
described the following.
•
On-line model estimation and identification in MATLAB:
'
'
, where mym-
file
contains the parameter model, in any form, par
the parameters of the state-space or input-output model,
aux
some auxiliary parameters); ‘d’ if the estimation
of the model ms is performed in the continuous-time
domain.
•
Again in MATLAB environment,
'
'
once
the ms state-space structure has been defined, the pem
function is the basic estimation command in the Identifi-
cation Toolbox and covers a variety of situations. data is
always an iddata object that contains the input-output
data. The function parameters Property and Value
can be used to select different estimation methods and
properties.
The design of observers and filters is performed in the
SIMULINK environment. On the other hand, the adaptive filter
functions in the Filter Design and Signal Processing Toolboxes
of MATLAB and SIMULINK implement the overall filter
design, replacing the adaptive algorithm with an appropriate
technique.
•
Least Mean Squares (LMS) algorithms represent the sim-
plest and most easily applied adaptive algorithms, the
Recursive Least Squares (RLS) algorithms represents in-
creased complexity, computational cost, and fidelity. In
performance, RLS approaches the Kalman filter (adap-
tkalman
) in adaptive filtering applications, at somewhat
reduced required throughput in the signal processor.
•
Compared to the LMS algorithm, the RLS approach
offers faster convergence and smaller error with respect
to the unknown system, at the expense of requiring
more computations. In contrast to the least mean squares
algorithm, from which it can be derived, the RLS adaptive
algorithm minimizes the total squared error between the
desired signal and the output from the unknown system.
Within limits, any of the adaptive filter algorithms can be
used to solve an adaptive filter problem by replacing the
adaptive portion of the application with a new algorithm.
•
As the signal into the filter continues, the adaptive filter
coefficients adjust themselves to achieve the desired re-
sult, such as identifying an unknown filter or cancelling
noise in the input signal. The adaptive filter in SIMULINK
can be considered as a box comprising the adaptive filter
and the adaptive RLS algorithm.
Hence, an online observer or filter realization for FDI applica-
tion is based on both the on-line identification of the process
models and of the design of the observer or filter itself. The
whole project is based on the characteristics of the input-output
signals to the observer or filter and a signal which represent the
desired behavior of the filter on its input. The design of the ob-
server or filter does not require any other frequency response
information or specification. To define the self learning process
the filter uses, the adaptive algorithm used to reduce the error
between the output signal and the desired signal has to be ex-
ploited. When the LMS performance criteria has achieved its
minimum value through the iterations of the adapting algorithm,
the adaptive filter is finished and its coefficients have converged
to a solution. When the input-output data change their character-
istics, sometimes called the filter environment, the filter adapts
to the new environment by generating a new set of coefficients
for the new data.
Moreover, once the adaptive observer or filter has been imple-
mented in connection with the on-line grey-box estimation algo-
rithm, the so-called residual signals can be generated. A residual
evaluation logic should be able to detect and isolate any fault oc-
currence in the presence of disturbance, modeling uncertainty
and measurement noise. Due to the particular FDI application,
the use of an adaptive threshold logic for robust FDI has to be
implemented. This task can be accomplished in SIMULINK by
using on-line and recursive estimators.
Finally, in many cases it may be necessary also to estimate
the fault-free thresholds on-line at the same time as the input-
output data is received. In this case, in particular, some deci-
sions on-line have to be made, as in adaptive control, adaptive
filtering, or adaptive prediction. For FDI applications, it is then
necessary to investigate possible time variation in the threshold
properties during the collection of data. Terms like recursive
identification, adaptive parameter estimation, sequential estima-
tion, and on-line algorithms are successfully used for such on-
line/adaptive threshold selection algorithms.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
SIMANI: IDENTIFICATION AND FAULT DIAGNOSIS
215
XI. C
ONCLUSION
This study has described a complete design procedure for FDI
in actuators, components and output sensors of a gas turbine
system. Although this is an application study based on a gas
turbine, the principles and methods used are applicable to al-
most any industrial system with dynamic behavior and with sets
of input-output measurements. The FDI tasks were performed
through the use of a bank of dynamic observers or, when the
measurement noises are not negligible, a bank of Kalman fil-
ters. In this study single faults on components of the system,
i.e., faults in actuators and output sensors, were considered.
This does not mean that multiple (simultaneous) faults are not
possible to isolate using model-based methods. Indeed, earlier
studies have shown that model-based methods for FDI are par-
ticularly suited to the detection and isolation of multiple faults,
when certain modeling and design conditions are satisfied. In a
later study we will consider this issue further.
The proposed method does not require physical knowledge
of the process under observation because the input-output links
are obtained by means of an identification scheme, based on
input-output models derived from the data. ARX models and
EIV models were used based on the least square method and the
scheme due to Frisch, respectively. The latter approach provides
estimates of the variances of the input-output noises, which are
required in the design of the Kalman filters. This identification
approach was applied to a SIMULINK model of a single-shaft
industrial gas turbine. In order to analyze the diagnostic ef-
fectiveness of the FDI system in the presence of changes or
drifts in measurements, faults were generated by means of ramp
functions.
The results obtained indicate that the minimal detectable
faults on the system actuator, component, and output sen-
sors are of interest for the industrial diagnostic applications.
However, since in real industrial applications incipient ramp
faults develop slowly over a long period and in order to avoid
excessively long duration simulations, the fault development
rate was increased so that significant effects were present after
shorter periods. This is a factor that must be taken into account
for FDI performance evaluation.
The main aspect of this work was the use of linear system
identification and modeling methods, although the system con-
sidered is nonlinear. This is considered important to avoid the
complexities that would otherwise be inevitable when nonlinear
models are used. As stated in the Introduction there is certainly
an increasing interest in the research literature in the use of non-
linear methods (nonlinear observers, extended Kalman filters,
fuzzy-logic methods, etc) and it is only a question of time before
these techniques find their way into full application projects.
However, as the feature of system supervision is to monitor the
operation and performance of the system with respect to an ex-
pected point of operation, linear system methods are still very
valid. Deviations from expected system behavior could be used
to monitor system performance changes as well as system com-
ponent malfunctions.
Finally, simulation results have shown that the minimal de-
tectable fault sizes, obtained by using geometrical analysis of
Kalman filter innovations, are smaller than the ones computed
in the noise-free environment, i.e. when compared with resid-
uals computed using dynamic observers. Moreover, if a detec-
tion delay is tolerable the amplitude of the minimal detectable
fault is lower. The minimal detectable faults on the various sen-
sors seem to be adequate for industrial diagnostic applications.
Furthermore, minimal detectable faults can be reduced if larger
detection delays are tolerable.
A
CKNOWLEDGMENT
The author wishes to acknowledge Dr. S. Daley and Dr. A.
Pike of the ALSTOM Power Technology Centre (Whetstone,
Leicester, U.K.) for the helpful discussions and for their valu-
able suggestions. He is also grateful to the management of the
ALSTOM Power Technology Centre for permission to publish
this work. He finally wishes to acknowledge Prof. R. J. Patton of
the University of Hull (U.K.) and Prof. C. Fantuzzi of the Uni-
versity of Modena and Reggio Emilia (Reggio Emilia, Italy) for
the helpful discussions and for their valuable suggestions.
R
EFERENCES
[1] R. E. Kalman, “System identification from noisy data,” in Dynamical
System II
, A. R. Bednarek and L. Cesari, Eds.
New York: Academic,
1982, pp. 135–164.
[2]
, “Nine lectures on identification,” in Lecture Notes on Economics
and Mathematical System
.
Berlin, Germany: Springer-Verlag, 1990.
[3] R. Frisch, Statistical Confluenece Analysis by Means of Complete Re-
gression Systems
.
Oslo, Norway: Economic Inst., Univ. Oslo, 1934.
[4] S. Beghelli, R. P. Guidorzi, and U. Soverini, “The Frisch scheme in dy-
namic system identification,” Automatica, vol. 26, no. 1, pp. 171–176,
1990.
[5] S. Simani, C. Fantuzzi, and S. Beghelli, “Diagnosis techniques for sensor
faults of industrial processes,” IEEE Trans. Control Syst. Technol., vol.
8, no. 5, pp. 848–855, Sep. 2000.
[6] R. Rovatti, C. Fantuzzi, and S. Simani, “High-speed DSP-based imple-
mentation of piecewise-affine and piecewise-quadratic fuzzy systems,”
Signal Process.
, vol. 80, pp. 951–963, Nov. 2000.
[7] S. Simani, C. Fantuzzi, R. Rovatti, and S. Beghelli, “Parameter identifi-
cation for piecewise linear fuzzy models in noisy environment,” Int. J.
Approx. Reason.
, vol. 1–2, pp. 149–167, Nov. 1999.
[8] S. Simani and C. Fantuzzi, “Fault diagnosis in power plant using neural
networks,” Int. J. Inform. Sci., vol. 127, pp. 125–136, Aug. 2000.
[9] C. Fantuzzi, S. Simani, S. Beghelli, and R. Rovatti, “Identification of
piecewise affine models in noisy environment,” Int. J. Control, vol. 75,
no. 18, pp. 1472–1485, 2002.
[10] S. Simani, C. Fantuzzi, and R. J. Patton, Model-Based Fault Diagnosis
in Dynamic Systems Using Identification Techniques
, 1st ed.
London,
U.K.: Springer-Verlag, Nov. 2002.
[11] R. J. Patton, P. M. Frank, and R. N. Clark, Eds., Fault Diagnosis
in Dynamic Systems, Theory and Application
. ser. Control Engi-
neering.
London, U.K.: Prentice-Hall, 1989.
[12] R. Isermann and P. Ballé, “Trends in the application of model-based fault
detection and diagnosis of technical processes,” Control Eng. Pract., vol.
5, no. 5, pp. 709–719, 1997.
[13] J. Chen and R. J. Patton, Robust Model-Based Fault Diagnosis for Dy-
namic Systems
.
Norwell, MA: Kluwer, 1999.
[14] R. J. Patton, P. M. Frank, and R. N. Clark, Eds., Issues of Fault Diagnosis
for Dynamic Systems
.
Berlin, Germany: Springer-Verlag, 2000.
[15] J. Gertler, Fault Detection and Diagnosis in Engineering Sys-
tems
.
New York: Marcel Dekker, 1998.
[16] A. S. Willsky, “A survey of design methods for failure detection in dy-
namic systems,” Automatica, vol. 12, pp. 601–611, Nov. 1976.
[17] R. Isermann, “Process fault detection based on modeling and estimation
methods: a survey,” Automatica, vol. 20, pp. 387–404, Jul. 1984.
[18] M. Baseville, “Detecting changes in signals and systems: a survey,” Au-
tomatica
, vol. 24, no. 3, pp. 309–326, 1988.
[19] P. M. Frank, “Fault diagnosis in dynamic systems using analytical
and knowledge based redundancy: a survey of some new results,”
Automatica
, vol. 26, pp. 459–474, May 1990.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.
216
IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 1, NO. 3, AUGUST 2005
[20] L. Xie and Y. C. Soh, “Robust Kalman filtering for uncertain systems,”
Syst. Control Lett.
, vol. 22, pp. 123–129, 1994.
[21] L. Xie, Y. C. Soh, and C. E. de Souza, “Robust Kalman filtering for
uncertain discrete-time systems,” IEEE Trans. Automat. Contr., vol. 39,
no. 6, pp. 1310–1314, Jun. 1994.
[22] J. Chen, R. J. Patton, and H. Y. Zhang, “Design of unknown input ob-
server and robust fault detection filters,” Int. J. Control, vol. 63, no. 1,
pp. 85–105, 1996.
[23] P. M. Frank and X. Ding, “Survey of robust residual generation and eval-
uation methods in observer-based fault detection system,” J. Proc. Cont.,
vol. 7, no. 6, pp. 403–424, 1997.
[24] I. Leontaritis and S. A. Billings, “Input–output parametric models for
nonlinear systems part I: deterministic nonlinear systems,” Int. J. Con-
trol
, vol. 41, no. 2, pp. 303–328, 1985.
[25]
, “Input-output parametric models for nonlinear systems part II:
stochastic nonlinear systems,” Int. J. Control, vol. 41, no. 2, pp. 329–344,
1985.
[26] T. Söderström and P. Stoica, System Identification.
Englewood Cliffs,
NJ: Prentice-Hall, 1987.
[27] L. Ljung, System Identification: Theory for the User, 2nd ed.
Engle-
wood Cliffs, NJ: Prentice-Hall, 1999.
[28] R. P. Guidorzi and R. Rossi, “Identification of a power plant from normal
operating records,” Automat. Contr. Theory Applicat., vol. 2, pp. 63–67,
Sep. 1974.
[29] S. Beghelli, P. Castaldi, R. P. Guidorzi, and U. Soverini, “A comparison
between different model selection criteria in Frisch scheme identifica-
tion,” Syst. Sci. J., vol. 20, pp. 77–84, Sept. 1994.
[30] MathWorks, SIMULINK: User’s Guide, The MathWorks, Inc., Natick,
MA, Mar. 1992.
[31] MathWorks, Inc., MATLAB User’s Guide, Natick, MA, 1990.
[32] A. H. Jazwinski, Stochastic Processes and Filtering Theory.
New
York: Academic, 1970.
[33] S. Beghelli and U. Soverini, “Identification of linear relations from
noisy data: geometrical aspects,” Syst. Control Lett., vol. 18, no. 5, pp.
339–346, 1992.
Silvio Simani
(M’99) was born in Ferrara, Italy, in
1971. He received the Laurea degree (cum laude) in
electrical engineering in 1996 from the Department
of Engineering, Università degli Studi di Ferrara, and
the Ph.D. degree in information science: automatic
control from the Department of Engineering, Univer-
sity of Ferrara and Modena, Italy.
He has been an Assistant Professor since 1992 and
a Research Associate since 1999, both in the Depart-
ment of Engineering, Università di Ferrara. He is also
a member of the Technical Committee SAFEPRO-
CESS since 2000. His research interests include fault diagnosis of dynamic pro-
cesses, system modeling and identification, and the interaction issues between
identification and fault diagnosis. In particular, his main research topics are in
the design and development of automatic fault detection and isolation proce-
dures regarding linear dynamic systems (industrial processes and power plants),
modeling and identification of linear and nonlinear dynamic models, fuzzy mod-
eling and control of linear and nonlinear algebraic and dynamic systems. He is
a reviewer for many international journals and is an author of several papers and
one book on these topics.
Authorized licensed use limited to: IEEE Xplore. Downloaded on May 10,2010 at 19:02:54 UTC from IEEE Xplore. Restrictions apply.