Escolar Documentos
Profissional Documentos
Cultura Documentos
SYSTEMS: A SURVEY
Abstract: Recent approaches to fault detection and isolation for dynamic systems using methods
of integrating quantitative and qualitative model information, based upon soft computing (SC)
methods are surveyed. In this study, the use of SC methods is considered an important extension to
the quantitative model-based approach for residual generation in FDI. When quantitative models
are not readily available, a correctly trained neural network (NN) can be used as a non-linear
dynamic model of the system. However, the neural network does not easily provide insight into
model behaviour; the model is explicit rather than implicit in form. This main difficulty can be
overcome using qualitative modelling or rule-based inference methods. For example, fuzzy logic
can be used together with state space models or neural networks to enhance FDI diagnostic
reasoning capabilities. The paper discusses the properties of several methods of combining
quantitative and qualitative system information and their practical value for fault diagnosis of real
process systems.
Controller
schemes. Faults
• Parity Relations: They are based either on a Input Plant Output
use of the fact that component faults of a Fig. 2 Neural networks scheme for FDI
dynamic system can be thought of as reflected in
Whilst this configuration can be well trained on
the physical parameters as for example friction,
numerical data once that the output is known,
mass velocity resistance. Faults are detected by
symbolic knowledge from experts cannot easily be
estimating parameters of non-parametric models.
incorporated.
The main assumption made when using the above
The mathematical model used in traditional FDI
methods is that a precise mathematical model of the
plant is required. This makes quantitative model- methods can be very sensitive to modelling errors,
based approaches very difficult to use in real parameter variation, noise and disturbance.
However, no mathematical model of the system is
systems, since any un-modelled dynamics can affect
needed to implement a neural network. Online
the performance of the FDI scheme. A way to
training makes it possible to change the FDI system
overcome this, is to design robust algorithms where
easily when changes are made in the physical
the effects of disturbances, on the residual, are
minimised and the sensitivity to faults maximised. process, control system or parameters. A suitably
A lot of approaches had been developed including trained neural network can generalise when
presented with inputs not appearing in the training
unknown input observers and eigenstructure
data. Neural networks have the ability to make
assignment observers (Patton et al 2000; Chen &
intelligent decisions in cases of noisy or corrupted
Patton, 1999), frequency domain techniques for
data. They also have a highly parallel structure,
robust FDI filters such as H ∞ (Edelmeyer et al., which is expected to achieve a higher degree of
1997) and the minimisation of multi-objective fault-tolerance then conventional schemes (Hunt et
functions (Chen, Patton & Liu, 1997). al., 1992). Neural networks can operate
simultaneously on qualitative and quantitative data
3. FDI VIA NEURAL NETWORKS and they are readily applicable to multivariable
To overcome some of the difficulties of using systems. Neural networks can also be applied for
mathematical models, and make FDI algorithms process condition monitoring, where the focus is on
more applicable to real systems, the neural network small irreversible changes in the process which
can be used to both generate residuals and isolate develop into bigger faults. Yin (1993) demonstrated
faults (Chen & Patton, 1999). A neural network is a the application of MLP and Kohonen self-organising
processing system that consists of a number of feature map (KFM) to predictive maintenance or
highly interconnected units called neurons. The condition-based maintenance of electrical drives,
neurons are interconnected by a large number of particularly induction motors. The first method
‘weighted links’. Each neuron can be considered as utilises supervised learning and in the second the
a mathematical function that maps the input and learning is unsupervised.
output space with several inputs. The inputs are
connected to either the inputs of the system or the 3.1 Application Studies
outputs of the other neurons in the system. The Neural networks have been successfully applied to
output of one neuron effects the outputs of other many applications including fault diagnosis of non-
neurons and all neurons connected together can linear dynamic systems (Wang, Brown & Harris,
perform complex processes. 1994; Dong & McAvoy, 1996). Multi-layer
perceptron (MLP) networks are applied to detect
Indeed, one of the main features of neural networks leakages in electro-hydraulic cylinder drive in a
is their ability to learn from examples. Hence, they fluid power system (Watton & Pham, 1997). They
can be trained to represent relationships between showed that maintenance information can be
past values of residual data (generated by another obtained from the monitored data using the neural
neural network) and those identified with some network instead of a human operator. Crowther et
known fault conditions. The configuration used by al. (1998), showed in an application of a neural
Chen & Patton (1999) involved a multi-layer feed- network to fault diagnosis of hydraulic actuators,
forward network configuration (Fig. 2). that experimental faults can be diagnosed using
neural networks trained only on simulation data.
Neural networks are applied to detect the internal
leakage in the control valves and motor faults in
process plants (Sharif & Grosvenor, 1998).
Kuhlmann et al (1999) presented the principle of the location and occurrence time of the faults. Chen and
Device-Specific ANN (DS-ANN) approach to fault Patton (1999) also showed that a single neural
diagnosis. The basic principle of the DS-ANN network can be used for both stages simultaneously,
approach is that neural networks are trained for with increased training time and complexity. Fault
dealing with certain basic groups or electrical isolation requires that the training data are available
devices (e.g. lines, transformers, busbars etc). for all expected faults in terms of residual values or
Weerasinghe et al (1998) investigated the system measurements.
application of a single neural-network for the
The neural network can be used for classification in
diagnosis of non-catastrophic faults in an industrial
conjunction with other residual generating methods
nuclear processing plant operating at different
e.g. non-linear observers. Neural networks have
points. Data-conditioning methods are investigated
been successfully applied in state and parameter
to facilitate fault classification, and to reduce
based FDI schemes (Han & Frank , 1997; Ficola et
network complexity. Maki & Loparo (1997)
al 1997). Han and Frank (1997) proposed a
presented detection and diagnosis of faults in
parameter estimation based FDI using neural
industrial processes that require observing multiple
networks in which physical parameters are estimated
data simultaneously. The main feature of this
by applying the neural network universal
approach is that the fault detection occurs during
approximation property applied to the measured I/O
transient periods of process operation. Vemuri &
data. The deviations from normal values are then
Polycarpou (1997) investigated the problem of fault
used for fault diagnosis. It is assumed that the
diagnosis in rigid-link robotic manipulators with
faults in the monitored process can be described as
modelling uncertainties. A learning architecture
changes in the parameter vector and the nominal
with sigmoidal neural networks was used to monitor
parameter values are known in advance or can be
the robotic system for any off-nominal behaviour
estimated online (e.g. via recursive least-squares
due to faults. Butler et al (1997) discussed the use
method). A linear model of the system can be
of a new neural network supervised clustering
described as:
method to perform fault diagnosis for power
distribution networks. The neural network proposed X (t ) = A(θ (t )) X (t ) + B (θ (t ))U (t ) + ξ (t ) (2)
performs fault type classification, faulted feeder and Y (t ) = C (θ (t )) X (t ) + D(θ (t ))U (t ) + η (t )
faulted phase identification, and fault impedance
or
estimation for grounded and ungrounded distribution
networks. ⎛ d ( n )Y d ( n −1)Y dY ⎞
⎜ n , n −1
,..., ,⎟
Neural networks have also been applied to the
dt
Y (t ) = H ⎜ ( m )
dt dt ⎟θ (t ) + ε (t ) (3)
⎜ d Y d ( m −1)Y ⎟
problem of joint faults in robots, using pattern ⎜ m , ,..., U ⎟
recognition. The joint-backlash of robots is ⎝ dt dt m −1 ⎠
diagnosed by monitoring its vibration response
during normal operation (Pan et al. 1998). James where X ∈ Rn, Y ∈ Rm , θ ∈ Rp are the state, output,
and Yu (1995) used a neural network for the model parameter vectors and ξ ∈ Rn and η ∈ Rm
condition monitoring and fault diagnosis of a high- are the process and measurement noise signals.
pressure air compressor valve. The neural network- A( θ ), B( θ ), C( θ ), D( θ ) and H( ) are known
based FDI scheme can also show when further function matrices. (1) is suitable for FDI with state
increases in fault levels might be likely, thus giving estimation whilst (2) is suitable for parameter
the operator time to take necessary action estimation based FDI. Faults in the process can be
(Boucherma, 1995). described as the change in the parameter vector
Dynamical neural networks (Korbicz et al 1998) are ∆θ = θˆ − θ o , where θˆ and θ o are the estimated
applied to on-line fault detection of power systems,
and nominal parameter vectors.
aircraft systems, chemical plants and nuclear
reactors which are highly non-linear complex. Neural networks have been successfully applied to
the on-line FDI of industrial processes. Fuente &
3.2 Strategies for Fault Diagnosis
Vega (1999) used a neural network for FDI of a
It is clear that neural networks can be applied to
biotechnological process. Real data from
fault diagnosis using different approaches. Pattern
experiments on the plant were used together with
recognition approach and residual generation-
on-line estimation. The back-propagation algorithm
decision making are the most common ones. The
was used to analyse the frequency content of some
second approach is generally more suitable for
fault-indicating signals derived from the
dynamic systems and comprises residual generation
identification step. This gave rise to correct
and decision making stages in just the manner
detection and isolation of each fault. Gomm (1998)
outlined in section 2 of this paper. In the first stage
described an adaptive neural network that
the residual vector r is determined in order to
continually monitors and improves its performance
characterise each fault. Ideally, the neural network
on-line as new fault information becomes available.
models identify all classes of system behaviour. The
New nodes are automatically added to the network
second stage, decision making or classification
to accommodate novel process faults after detection,
processes the residual vector r to determine the
and on-line adaptation is achieved using recursive generalise well when presented with data outside the
linear algorithms to train selected network training I/O space. MLP and Fuzzy Logic based
parameters. systems on the other hand tend to generalise in a
better way. On-line training should be used to
3.3 Taxonomy of Neural Networks
update such networks (Wilson, 1998).
There is a large number of neural network
architectures in use. Feed-forward, recurrent, If some unknown fault conditions appear, the neural
Radial Basis Function (RBF), fuzzy, B-spline, classifier is no longer valid because it is not trained
dynamic, competitive and probabilistic neural to classify this type of fault. Adaptive training
networks are among the most frequently used algorithms should be used with systems requiring
structures for fault diagnosis. Li et al (1999) on-line training.
described a method to diagnose the most frequent
It is not usually possible to acquire all the faulty data
faults of a screw compressor and assess magnitude
for neural network training. Thus unsupervised
of these faults by tracking changes in the compressor
training, which uses a Kohonen network and the
dynamics. Yu et al (1999) investigated semi-
Counter-propagation (CPN) network (Dalmi et al.,
independent neural model, based on an RBF
1999), is necessary in order to classify the faults not
network, to generate enhanced residuals for
known a priori. A combined artificial neural
diagnosing sensor faults in a reactor. Narendra et al
network and expert system tool (ANNEPS) is
(1998) also used the RBF network architecture for
developed (Wang, 1998) for transformer fault
fault diagnosis in a HVDC system. A new pre-
diagnosis using dissolved gas-in-oil analysis (DGA).
classifier was proposed which consists of an
ANNEPS takes advantage of the inherent positive
adaptive filter (to track the proportional values of the
features of each method and offers a further
fundamental and average components of the sensed
refinement of present techniques.
system variables), and a signal conditioner which
uses an expert Knowledge Base (KB) to aid the pre-
3.5 Hybrid neural networks
classification of the signal.
Neural network-based FDI methods usually require
Other networks used in recent applications include pre-processing or signal conditioning algorithms to
Dynamic Back-propagation networks (Narendra, reduce the effect of noise and disturbance and to
1996) and Cerebellar Model Articulation Controller enhance the fault features. Many other techniques
(CMAC) Network (Leonhardt et al., 1995; Brown & have been combined with neural networks, including
Harris, 1994b). Each of these architectures offers fuzzy logic, genetic algorithms and adaptive
different characteristics to suit distinct applications. modelling etc. Aminian et al (2000) developed an
Recent research focuses on networks that can analog-circuit fault diagnostic system based on
optimise their structure during training. Ren and back-propagation neural networks using wavelet
Chen (1999) proposed a new type of neural network decomposition, principal component analysis, and
in which the dynamical error feedback is used to data normalisation as pre-processors. The proposed
modify the inputs of the network. system has the capability to detect and identify
faulty components in an analog electronic circuit by
3.4 Design issues of applying neural networks for
analysing its impulse response. Pantelelis et al
fault diagnosis
(2000) developed simple finite element (FE) models
As discussed earlier, neural networks provide
of a turbocharger (rotor, foundation and
essentially a “black box” signal processing structure,
hydrodynamic bearings) combined with neural
which do not show rules governing their operation
networks and identification methods and vibration
and there is no visibility as to their real behaviour
data obtained from real machines towards the
(from an input-output point of view). This does not
automatic fault diagnosis. Liu(1999) used an
enable the user to understand the system and predict
extended Kalman filter (EKF) and neural network
its behaviour in uncertain situations. On the other
classifier for FDI. Network inputs are the process
hand, B-spline networks can be used to extract and
I/O data, such as pressure and temperature,
include some heuristic knowledge about the system.
parameters estimated by EKF, and state values
The training time required for a specific application calculated by dynamic equations, whilst network
and the complexity of the training algorithm present outputs are process fault scenarios. Zhao et al
further limitations. The earlier back-propagation (1998) presented a multidimensional wavelet (MW)
algorithm used to train MLPs; requires an excessive with its rigorously proven approximation theorems.
training time and is generally an off-line method for Taking the new wavelet function as the activation
training. RBF networks are capable of on-line function in its hidden units, a new type of wavenet
adaptive training if required (Wilson, 1998) but use called multidimensional non-orthogonal non-product
large numbers of neurons if the I/O space is large. wavelet-sigmoid basis function neural network
To accelerate convergence, state variables with (WSBFN) model is proposed for dynamic fault
additional terms can be used in training (Watton & diagnosis. Based on the heuristic learning rules
Pham, 1997). presented by Zhao et al (1998), a new set of
heuristic learning rules is presented for determining
Neural networks which use neurons as membership
the topology of WSBFNs. Izhao et al (1997)
functions, e.g. RBF and B-spline networks, do not
proposed the wavelet-sigmoid basis neural network approach for fault diagnosis, using fuzzy sets. The
(WSBN) with expert system (ES) for dynamic fault system allows not only the fault location, but
diagnosis (DFD). Ye & Zhao (1996) proposed. A identification of the fault type. Currents and
hybrid intelligent system which integrates neural voltages are analysed using the fault phase-
networks with a procedural decision-making impedance and fuzzy sets. Dexter & Benouarets
algorithm to implement hypothesis-test cycles of used a set of fuzzy reference models which describe
system fault diagnosis. faulty and fault-free operation, and a classifier based
on fuzzy matching for fault diagnosis. The
3.6 Multiple fault detection
reference models are obtained off-line from
Neural networks have been found to give more
simulation data. A fuzzy model which describes the
information with regard to multiple-fault conditions
actual plant behaviour is identified on-line from
than some other methods (steady-state position error,
normal operating data and compared with each of
time series analysis). Some neural networks are
the reference models. Mechefske (1998) applied
applied to diagnose multiple faults in the processes
fuzzy logic techniques to classify frequency spectra
but generally it is much more difficult to diagnose
representing various rolling element bearing faults.
such faults in a process because the training data
The frequency spectra representing a number of
needed becomes very large. Maidon et al (1999)
different fault conditions are processed using a
developed a technique for diagnosing multiple faults
variety of fuzzy set shapes.
in analog circuits from their impulse response
function using a fault dictionary. A technique is Although fuzzy systems theory is often applied to
described (Ogg et al 1998) for diagnosing multiple industrial process, the applications often do not work
faults in analogue circuits from their impulse well. Sometimes fuzzy logic designs are completed
response function using multi-layer perceptrons, in without mathematical rigour. The main tasks of
terms of a specific example. A Dirac impulse input finding appropriate membership functions and fuzzy
to the circuit was simulated, and time domain rules are often determined simply by “trial and
features of the output response were classified by a error”. The rules can be obtained by means of
system of two multi-layer perceptrons to produce optimisation methods. LMI optimisation has been
accurate numerical fault values. used in order to design an optimal Takagi-Sugeno
(T-S) observer based on a relaxed stability condition
6.´ FAULT DIAGNOSIS VIA FUZZY (Patton et al., 1998). Another main approach to
LOGIC obtain the number, position and type of rules is to
apply adaptive and learning algorithms to fuzzy
Since Zadeh (1965) introduced the theory of the systems or to apply neural networks to learn the
fuzzy sets – manipulating data that were not precise, parameters of the fuzzy system.
but rather “fuzzy” and since the work of Mamdani
(1974), industrial application studies using fuzzy 6.4 Fuzzy Decision-Making
logic controllers have reached a major position in The advantage of the fuzzy approach is that it
systems engineering. supports, in a natural way, the direct integration of
the human operator into the fault detection and
Fuzzy Systems are useful in any situation in which supervision process. By avoiding an incorrect
the measurements taken are imprecise or their decision that can cause false alarms the aim of the
interpretation depends strongly on the context or on FDI decision making (for fault diagnosis) is to
human opinion. decide whether and where the fault in the system has
4.1 Application studies occurred (Kuipel & Frank (1997). Fuzzy decision-
Application areas include: the process industry, making objectives are very similar to expert systems
electromechanical systems, traffic and avionics and supervisory control. Expert Systems are used
control and biomedical systems etc. to simulate the problem-solving and decision-
making processes of a human expert within a
Evsukoff et al (2000) proposed a decision support relatively narrow domain. This is done using
system dedicated to fault detection and isolation special computer packages along with knowledge,
from a human-machine co-operation point of view. information and databases (Ford, 1991; Tzafestas,
Yang & Liao (1999) proposed an adaptive fuzzy 1989).
system for incipient fault recognition through an
evolution enhanced design approach. Complying Formulation of decision-making
with the practical gas records and associated fault A decision can be formulated by a set of variables
causes as much as possible, a fuzzy reasoning (sets, relations and functions) termed a quintuple (S,
algorithm is presented to establish a preliminary st, C, m, dc) (Verbruggen et al., 1999;
fuzzy diagnosis system. In this system, an Kaymak,1998). By using available information S is
evolutionary optimisation algorithm is further relied the possible actions where a selection of this set is
on to fine-tune the membership functions of the if- performed. St are the set of uncontrolled variables
then inference rules. Lu et al (1998) described a by the decision maker cannot of the environment but
fuzzy diagnostic model that contains a fast fuzzy they must be included in the decision making
rule generation algorithm and a priority rule based process. C is the set of consequences, which must
inference engine. Insfran et al (1999) proposed an be including into a multi-criteria decision-making
scheme. Uncertainties resulting from the various procedures to try to extract learning rules in
identification procedure and inherent uncertainties of combination with other techniques (e.g. using neural
the system are included, in part, in the consequences. networks) (Füssel, 1997).
M is the relation used to obtain the decision-making
solutions by mapping the space S x st into the set 6.4 Fuzzy Techniques in FDI
consequences as S x st → C. The decision-maker In recent years the application of fuzzy logic to
has, a priori, aims and objectives in a preference model-based fault diagnosis approaches has gained
ordering. They are taken into account in dc as a increasing attention in both fundamental research
decision function dc: C →ℜ. and application. Symptoms can be generated using
observers based on the estimation of the output from
A number of fuzzy decision-making methods for the system. The first methods used fuzzy set theory
control have been applied for more than 2 decades. to express cause-effect relations in expert systems.
For example, the formulation by Bellman and Zadeh The key idea of model-based methods is the
(1970). For this approach, there is no distinction generation of signals, termed residuals. These are
between aims and constraints; both are included in usually generated using mathematical methods
the membership functions. (based on state observers, parameter estimation or
6.4 Fuzzy Clustering and Fuzzy Identification parity equations). The models correspond to the
To identify complex non-linear systems it is monitored system (Chen & Patton, 1999). Residuals
common to obtain partitions of the available data, are signals representing inconsistencies between the
each partition or subset is approximated by a simple model and the actual system being monitored, but
model. The data can be quantitative or qualitative the deviation between the model and the plant is
information or a mixture of both. Clustering influenced not only by the presence of the fault but
algorithms are not only used for classification and also modelling uncertainty. One solution is for the
pattern recognition to construct fuzzy models but observer and controller parameters to be tuned via
also for the simplification and optimisation in estimation from the real system for fault isolation
modelling. and threshold adaptation (Schneider & Frank, 1994).
The introduction of fuzzy logic can improve the
Isoc (1998a, 1998b) used quasi-linear fuzzy models decision-making, and in turn will provide reliable
based on the Sugeno approach (from experimental and sufficient FDI, suitable for real industrial
measurement data according to the Box-Jenkins data applications. However, difficulties arise in the
sets). These were compared with the real system training of the algorithm in the inference mechanism
data sets and then with models obtained using other where knowledge is hidden in large amounts of data
identification techniques. Various identification and knowledge is embedded in trained neural
techniques to develop fuzzy models were used: for networks (Chen et al., 1997). A fuzzy feed-forward
example Mendel-Wang fuzzy reference sets (Wang neural network (FNN) is applied to extract rules
& Mendel, 1992) were used. The results obtained from an existing data base. Frank et al. (Frank &
were of good quality because a more natural inter- Kuipel, 1993; Frank 1993; Frank 1994a; Frank
dependence between the data set and extracted fuzzy 1996; Schneider & Frank, 1996; Frank & Köppen-
sets was defined. Seliger 1997) use fuzzy logic for residual evaluation.
The fuzzy approach is becoming a powerful This can be an important way of taking into account
alternative to using artificial expert systems and may modelling uncertainty at decision-making rather than
gain more practical importance in the future. The during residual generator design. By applying a
non-linear system can be identified using a fuzzy fuzzy rule-based approach the fault decision process
multiple-model description of the real system in can be made robust to the uncertainties so that false
parallel and a series model or any combination and missed alarm rates can be minimised.
(series-parallel) (Ballé et al., 1997) and Considering supervisory control (Linkens et al.,
consequently a number of models are identified. 1993, Frank & Kuipel, 1993) with tasks such as
Lopez-Toribio et al (1999) used a different system management, process monitoring,
approach. The identification of locally linear identification, fault detection, diagnosis and adaptive
models using the Takagi-Sugeno (TS) fuzzy capability reduces at lower level the models for
modelling strategy is solved using a convex developing simpler structures for observers and
optimisation technique involving linear matrix controllers using TS fuzzy models.
inequalities (LMI) in order to find the optimum set
of fuzzy models (Ai, Bi, Ci, Di, for I=1,r). This 5. INTEGRATION OF QUALITATIVE
approach has been successfully applied to a real- AND QUANTITATIVE FDI
time induction motor drive test rig. METHODS
The issue that remains a challenge is to obtain not Recent research has focused on the integration of
only a number of multiple linear models but also the symbolic and quantitative knowledge through a
minimum number of models, which describe the neuro-fuzzy system. The aim is to combine the
non-linear system. This optimisation is difficult neural network learning ability with the explicit
because the identification method using fuzzy logic knowledge representation of fuzzy-logic. The
depends on a large number of variables. There are application engineer can therefore extract, from the
data, a high level language description of the system. Altug et al (1999) presented two neural fuzzy
Heuristic plant knowledge can also be included. (NN/FZ) inference systems, namely, Fuzzy Adaptive
Learning Control/Decision Network (FALCON) and
Some other tools like evolutionary algorithms,
Adaptive Network Based Fuzzy Inference System
genetic algorithms or probabilistic reasoning can
(ANFIS), with applications to induction motor fault
also be combined with the above, to enhance the
detection/diagnosis problems. Aggarwal et al
parameter tuning or to deal with the uncertainty in
(1999) addressed the problems of fault diagnosis in
order to establish the desired intelligent system.
complex multi-circuit transmission systems, in
6.4 Combining Neural Networks with Fuzzy particular those arising due to mutual coupling
Logic between the two parallel circuits under different
There are several possible methods of combining fault conditions. The problems are compounded by
neural networks with the fuzzy logic, the advantages the fact that this mutual coupling is highly variable
and disadvantages of which depend on the specific in nature. In this respect, the soft computing
application. provides the ability to classify the abnormal
phase/phases by identifying different patterns of the
Farag et al. (1998) proposed a five layer fuzzy
associated voltages and currents. A Fuzzy ARTmap
neural network (Fig. 3) in which the parameter
(Adaptive Resonance Theory) neural network is
identification of the fuzzy model comprises three
employed and is found to be well-suited for solving
phases. In the first phase, initial parameters are
the complex fault classification problem under
found using the Kohonen self- organising feature
various system and fault conditions. Jota et al
map algorithm. The second phase consists of
(1998) proposed a combinatorial intelligent system
finding the linguistic rules. In the third phase, a
based on neuro-fuzzy, neuro-expert and fuzzy-expert
genetic algorithm known as “multi-resolutional
algorithms can be successfully applied in the
dynamic genetic agorithm (MRD-GA) is used to
detection of a number of faults in a range of
tune the membership functions.
equipments. Pfeufer & Ayoubi (1997) applied a
neuro-fuzzy-structure to the classification of faults,
Output
Layer 5: Output neurons based on symptoms generated by identifying a
mathematical model. Ozyurt & Kandel (1996)
Layer 4: Membership functions
describing output fuzzy
presented a hybrid diagnostic methodology for fault
linguistic variables diagnosis based on a hierarchical multi-layer
Layer 3: Fuzzy rule base. Each
neuron acts as single fuzzy
perceptron-elliptical neural network structure and a
control rule fuzzy expert system. The introduced hybrid system
Layer 2: Membership functions
describing input fuzzy linguistic is noise tolerant, easy to train and maintain and also
variables
reliable under changing process conditions.
Layer 1: Input neurons
Some recent research focuses on neural networks,
Input for example B-spline neural networks, which can be
used to extract the qualitative knowledge of the
Fig. 3 Fuzzy Neural Networks system. The B-spline network is used to classify
(Farag et al., 1998) faults in the process (Patton et al 1999). The faults
Li & Elbestawi (1996), proposed a multiple are assumed to be a priori known, and their
principal component (MPC) fuzzy neural network corresponding data available to the designer. The
for clustering (unsupervised classification) which network will then have as many outputs as classes of
employs fuzzification of the interconnections of a behaviour (Fig 5). Hence, for a system with 2
conventional neural network. This method is used classes of faults, the output of the network will be a
for automated tool condition monitoring in vector of dimension 3; this includes the models
machining and is based on Li and Elbestawi’s fuzzy associated with the two faults as well as that
neural networks (Fig. 4) in which fuzzy membership corresponding to the Healthy one. This method
functions are used for decision making and the offers the extraction of fuzzy rules in addition to
interconnection in the network. FDI.
Outputs
Output Layer
B1 Fault#1
Σ
B2
Fuzzy Logic Connections
y (t ) B3
. Fault#2
Fuzzy Connections at Neurons Hidden Layer Σ
.
.
Fuzzy Logic Connections . Fault#M
u (t ) . Σ
Input Layer
.
. Healthy
Inputs Bp Σ
Fig. 4 Fuzzy Neural Network
(Li & Elbestawi, 1996) Fig. 5 Network-Architecture used for
fault-isolation
As outlined earlier, there are many applications for µ i ( w ( t )) is the grade of membership of the
neuro-fuzzy based FDI. However, successful
premise variable, ω(t), or the tensor product of grade
implementation of neuro-fuzzy FDI methods
depends heavily on prior knowledge of the system of memberships, if ω(t) is a vector. The
and the training data. membership grade function µ i (ω( t )) satisfies the
following constraints
6.´ NON-LINEAR FDI VIA FUZZY r
OBSERVERS ∑ µ (ω(t )) = 1
i
6.4 Takagi and Sugeno Fuzzy Models: A i =1