Journal of Theoretical and Applied Information Technology

© 2005 - 2010 JATIT. All rights reserved.

www.jatit.org

NEURO-FUZZY APPROACH FOR FAULT LOCATION AND

DIAGNOSIS USING ONLINE LEARNING SYSTEM
1
P. CHANDRA SEKHAR, 2B.V. SANKER RAM, 3K.S. SARMA
1
Associate Professor, Department of Electrical and Electronics Engineering, M.G.I.T.,
Hyderabad, India-500075
2
Professor, Department of Electrical and Electronics Engineering, JNTUH College of Engineering,
Hyderabad, India-500085
3
Professor,Department of Electrical and Electronics Engineering, Vardhaman College of Engineering,
Hyderabad, India-501218

ABSTRACT

This paper outlines a hybrid approach of neuro-fuzzy based learning and classification approach based on
the online learning systems. The effect of fault diagnosis for the suggested fault location tool is evaluated
over the conventional fault diagnosis based approaches. The method of fault location based on the
conventional offline neuro controller approach is compared with the suggested hybrid approach for learning
and convergent time evaluation for distributed systems.

Keywords : Neuro-Fuzzy Approach, Fault Location, Fault Diagnosis, Online Learning System, System
Performance

1. INTRODUCTION Another important aspect of a power

quality study is coordination between the power
In a power system, faults, dynamic system behavior and equipment performance. It is
operations, or nonlinear loads often cause various desired that the response of the sensitive equipment
types of power quality disturbances such as voltage during the event be explained and correlated to
sags, voltage swells, switching transients, impulses, specific features of the event, so that either the
notches, flickers, harmonics, etc. (1,2). On the other system behavior or the equipment operating
hand, the increased use of sensitive electronic characteristics can be tuned for improved ride-
circuitry by industrial and residential customer, as through ability or immunity of the equipment to
well as the progress of utility deregulation and specific events (3). It has been noted that the
competition have imposed greater demand on the activities of detecting and classifying of power
quality of power. Consequently, the studies aimed quality events, characterizing and locating events,
at detecting and analyzing as well as eliminating or studying equipment sensitivity, and modeling of the
minimizing the effects of power quality system and equipment are closely related and
disturbances on industrial and customer loads have interdependent. Hence it is natural and desirable
assumed greater importance. that the data processing and analysis as well as
modeling and simulation of the system and
One critical aspect of power quality equipment be studied in one unified framework. In
studies is the ability to perform automatic power this paper a new neuro-fuzzy expert system for
quality monitoring and data analysis. Usually, detection and classification of various types of
utilities install power quality meters or digital fault faults events is developed.
recorders at certain locations so that various power
quality events can be recorded and stored in the Fuzzy-neural networks have been
form of sampled data for further analysis. Efficient proposed as a knowledge engineering technique
and prompt detection, classification, and and used for various application domains by
characterization of the events as well as further authors including Yamakawa and Uchino (1),
identification of the location of these events Uchikawa and Furuhashi (2), and others (3, 4, 5, 6, 9).
facilitate maintenance and control of the system, The fuzzy-neural approach proposed by these
and improve system stability and reliability. authors have been successfully used for learning

48
 Journal of Theoretical and Applied Information Technology

© 2005 - 2010 JATIT. All rights reserved.

www.jatit.org

and tuning fuzzy rules as well as solving sized generation units. These smaller units are also
classification, prediction and control problems. less complex, less expensive and more rapidly
Some recent publications suggest methods for constructed, allowing more rapid adaptation to the
training fuzzy neural in order to adjust to new or requirements of expanding industries.Therefore,
dynamically changing data and situations (4, 5). interconnection of power systems yields technical,
economical and environmental benefits, as excess
This paper introduces a new architecture capacity generated by hydro resources from one
for fuzzy-neural , called FZ-NN, which stands for region may augment load supplied by fossil-fuelled
Fuzzy-Neural Network, and investigates some generation in another.
learning and adaptation strategies associated with it.
In addition, the use of rule insertion and rule For interconnections to operate as
extraction algorithms are investigated. intended, transmission systems must have adequate
capacity or load rating to transmit the quantity of
2. FAULT DIAGNOSIS IN POWER SYSTEM power intended. If the existing transmission system
does not have adequate ratings for the predicted
Power supply systems comprise three power transfer, additional lines may be constructed
broad categories – generation, transmission and in parallel with existing transmission lines, or the
distribution. Electricity is generated at power transmission system may be upgraded to a higher
stations from various natural resources such as coal, operating voltage. This is not always possible due
natural gas, hydro and wind, in large quantities to environmental, economic and time factors, and
where economies of scale can be achieved. These public opinion. In a conventional system the
generating stations are proximate to the fuel processing of currents and voltages are carried out
resource, often a considerable distance from major to protect the devices and to make a suitable and
load centres. Therefore, large quantities of fast control on the faults happening in the
electricity are transported at high voltages via the transmission line. For the analysis of the suggested
transmission network to strategically located bulk controlling this paper introduces a controlling
supply substations, and from there to smaller operation of fault detection and controlling based
substations where the supply is further reduced for on a hybrid modeling of neural network and fuzzy
distribution. Modern power supply systems have logic. A system architecture for the suggested
evolved from separated utilities to large approach is as shown below:
interconnected systems, with many generating
Frequency
stations and load centres being interconnected Inputs: Data resolution
through power transmission lines. With the high Voltage wave capturing feature
degree of system interconnection, there is increased forms descriptor
power exchange over larger distances at higher
system voltage levels. Operation and expansion of
power systems impacts society and several factors Fuzzy-Neuro Outputs:
must be considered. modeling for detection Detection and
and classification classified type
Technical and economic factors concern of a disturbance
the improvement of existing equipment and design
Figure 1. Suggested operational flow model for the fault
of new equipment, which could offer economies. diagnosis system
Social and demographic factors deal with the
tendency of the power industry to influence social The application module “Detection and
and political processes, including setting of Classification” automatically detects and classifies
industrial enterprises and distribution of working the type of the disturbance captured in the recorded
population.These concerns have resulted in the or simulated waveforms. The types of disturbances
emergence of distributed generation, the current include the voltage sag, swell, outage, harmonic,
trend of interconnected smaller sized generating notch, flicker, impulse and switching transient.
units such as Kogan Creek gas fired power station After the disturbance is detected and classified, the
as opposed to the traditional coal fired large module “Waveform Characterization” further
capacity stations such as Tarong and Stanwell. This processes the waveform. Eight different sub-
trend is emerging as a consequence of increased modules corresponding to the eight types of events
Greenhouse gas issue significance, leading to have been designed. The software automatically
consideration of alternate energy sources such as selects the appropriate sub-module for computing
solar, wind and wave that operate with smaller parameters pertinent to the event. Then, one may

49
 Journal of Theoretical and Applied Information Technology

© 2005 - 2010 JATIT. All rights reserved.

www.jatit.org

proceed to the module “Equipment Sensitivity 3. FUZZY-NEURO MODELING

Study” for evaluating how various waveform
features affect the behavior of the equipment during The FZ-NN model is designed to be used
the event. Finally, the module “Event Location” in a distributed, and learning-based environment.
aims at accurately pinpointing the location of the The architecture provides learning from data and
event occurrence. Presently, we are focusing on approximate reasoning, as well as fuzzy rule
locating faults that caused the sag event using the extraction and insertion. It allows for the
waveforms recorded by a limited number of DFRs combination of both data and rules into one system,
and related power system data. Short-circuit studies thus producing the synergistic benefits associated
are employed to obtain an optimal fault location with the two sources. In addition, it allows for
estimate subject to a defined performance criterion several methods of adaptation (adaptive learning in
by iteratively posing faults in the system, running a dynamically changing environment). FZ-NN uses
simulations, and comparing the simulated a Multi-Layer Perceptron (MLP) network and a
waveforms with the recorded waveforms. The modified back propagation training algorithm. The
detection and classification problem consists of two general FZ-NN architecture consists of 5 layers of
steps. The first step is feature extraction, during neurons with partial feed forward connections. It is
which the distinct and dominant features (or an adaptable fuzzy-neural where the membership
patterns) of various events are selected and functions of the fuzzy predicates, as well as the
obtained using appropriate techniques. fuzzy rules inserted before training or adaptation,
may adapt and change according to new data. The
The second step is called decision-making, interface unit for the suggested design is as briefed
during which an inference engine to determine the below :
types of the events further processes the extracted
features. Appropriately chosen features are essential 3.1. Input layers
for both simplifying the Decision-Making System
(DMS) and improving the correct identification rate The input layer of neurons represents the
of the system. For decision-making, a neural input variables as crisp values. These values are fed
network based system was presented(4). The author to the condition element layer, which performs
suggests using the time-delay network to capture fuzzification. This is implemented using three point
the temporal features of the input signals. One triangular membership functions with centers
drawback of using neural network is the difficulty represented as the weights into this condition
of the training process. Some authors have element layer. The triangles are completed with the
proposed using the fuzzy logic to model the minimum and maximum points attached to adjacent
uncertainties of the training error so that the centers, or shouldered in the case of the first and
learning rate can be finely tuned to improve the last membership functions. The triangular
convergence of the system (5). However this method membership functions are allowed to be non-
still belongs to the category of neural network and symmetrical and any input value will belong to a
does not utilize fuzzy logic to model the maximum of two membership functions with
uncertainties of the input patterns. degrees differing from zero. These membership
degrees for any given input will always sum up to
Fuzzy logic based DMS is well suited to one, ensuring that some rules will be given the
solve the realworld problems. It bridges the opportunity to fire for all points in the input space.
quantitative and qualitative considerations. It has This center based membership approach taken by
found wide applications in the areas of load FZ-NN avoids the problems of uncovered regions
forecasting, harmonic tracking, power metering, in the input space that can exist with more flexible
etc. (11, 12). In this work, application of fuzzy logic membership representation strategies. These do not
techniques to the detection and classification of always limit centers and widths in such a way as to
power quality events is explored. For the analysis ensure complete coverage.
of the suggested controlling this paper introduces a
controlling operation of fault detection and While algorithms could be formulated and
controlling based on a hybrid modeling of neural used in such cases to force the memberships to
network and fuzzy logic. cover the input space, the simple center-based
approach taken by FZ-NN seems both more
efficient and more natural, with fewer arbitrary
restrictions. It should be noted that there are no

50
 Journal of Theoretical and Applied Information Technology

© 2005 - 2010 JATIT. All rights reserved.

www.jatit.org

“bias” connections necessary for this representation 2.19722 would ensure that a rule node would
in FZ-NN. The weights from the input to condition provide activation values from 0.1 to 0.9 when the
element layers of neurons can take values in [0,1] net input values are between -1 and +1. These
only since the data are assumed to be normalised to values may be desirable as part of the architecture’s
this range. This normalisation is normally carried “fuzziness.” The semantic meaning of the
out as part of the FZ-NN pre-processing operations, activation of a node is that it represents the degree
and can be performed and reverse transparently in to which input data matches the antecedent
applications. component of the associated fuzzy rule. However
the synergistic nature of rules in a fuzzy-neural
Initially the membership functions are architecture must be remembered when interpreting
spaced equally over the weight space, although if such rules. The connection weights from the
any expert knowledge is available this can be used condition element layer (also called the
for initialisation. In order to maintain the semantic membership functions layer) to the rule layer
meaningfulness of the memberships contained in represent semantically the degrees of importance of
this layer of connections some restrictions are the corresponding condition elements for the
placed on adaptation. When adaptation is taking activation of this node.
place the centers are limited,remain within equally
sized partitions of the weight space. This avoids The values of the connection weights to
problems with violating the semantic ordering of and from the rule layer can be limited during
membership function labels. Therefore, under the training to be within a certain interval, say [-1,1],
FZ-NN architecture labels can be attached to thus introducing non-linearity into the synaptic
weights when the network is constructed and these weights. This option mimics a biologically
will remain valid for the lifetime of the network. plausible phenomenon (10) but should be
For example, a membership function weight implemented in accordance with an appropriate
representing low will always have a center less than gain factor for the activation function. For example,
medium, which will always be less than high. if the interval is [-1,1] a suitable value for the gain
factor may be 2.19722 as described above. The
The condition element layer of neurons is weight limitation would ensure that inputs into the
potentially expandable during the adaptation phase rules remain within [-1,1] (since the input
with more nodes representing more membership membership functions are all between 0 and 1) and
functions for the input variables. Simple activation the gain factor would only allow the rules to output
functions are used in the condition element nodes to values in [0.1,0.9]. This further enhances the
perform fuzzification. An important aspect of this meaningfulness of the rules and weight saturation
layer is that different inputs can have differing will not occur. As an example of the problems of
numbers of membership functions. The same rule interpretation with unrestricted weights, it is
principle applies to the output membership difficult to interpret a rule that has input weights
functions. This allows for very different types of that are very high values, without some form of
inputs to be used together. As a simple example, normalization. With this weight limiting option, the
temperature may be divided into seven different necessity for such normalization is removed.
membership functions representing the range from
cold to hot, while holiday (which is a binary 3.3. Output layers
variable to indicate whether it is a public holiday or
not) can be represented using just two, for yes and In the action element layer, a node
no. represents a fuzzy label from the fuzzy quantization
space of an output variable, for example small,
3.2. Rule layer medium, or large for the output variable “required
change in the velocity.” The activation of the node
In the rule layer each node represents a represents the degree to which this membership
single fuzzy rule. The layer is also expandable, in function is supported by the current data used to
that nodes can be added to represent more rules as recall the FZ-NN. So this is the level to which the
the network adapts. The activation function is the membership function for this fuzzy linguistic label
sigmoidal logistic function with a variable gain is “cut” according to the current facts. The
coefficient g (a default value of 1 is used giving the connections from the rule layer to the action
standard sigmoidal activation function). For the element layer represent conceptually the confidence
gain coefficient large values will make it close to factors of the corresponding rules when inferring
the hard limited thresholding function. A value of fuzzy output values. They are subject to constraints

51
 Journal of Theoretical and Applied Information Technology

© 2005 - 2010 JATIT. All rights reserved.

www.jatit.org

that require them to remain in specified intervals as main departure being partitions, much of the large
for the previous layer with the same advantages of body of theory regarding such networks is still
semantic interpretability. The activation function applicable. For those results not immediately
for the nodes of this layer is the sigmoidal logistic applicable for FZ-NN the modifications are made
function with the same (variable) gain factor as in much simpler, given FZ-NN’s natural structure and
the previous layer. Again, this gain factor should be algorithm.
adjusted appropriately,given the size of the weight
boundary. The output layer performs a modified Here the fuzzification layer and the
center of gravity defuzzification. Singletons, defuzzification layer change their input connections
representing centers of triangular membership based on simple and intuitive formulae. These
functions, as it was the case of the input variables, changes reflect the concepts represented by the
are attached to the connections from the action to layers and must satisfy the restrictions
the output layer. Linear activation functions were (partitioning) imposed on the membership functions
used. For example, a small, medium and large can (the movements of their centres cannot take them
be represented as connection weights having values out of the membership’s partition). The same
of 0, 0.5 and 1.0 respectively from the output range principles apply for the two layers, but different
of [0,1] if normalized outputs are considered. formulae are used to calculate the change of the
Adapting the output membership functions would weights. Figure 3 shows the initial membership
mean moving the centers, but the requirement that functions of a variable x (either input or output) and
the membership degrees to which a particular the membership functions after adaptation. The
output value belongs to the various fuzzy labels amount of change has been exaggerated in order to
must always sum up to one, is always satisfied. For demonstrate the concept involved. In the normal
each center, there is a constraining band (partition) course of training changes to membership functions
where this value can move to. This is used in the are limited to small, gradual movements, with the
same way as the input membership function centers majority of weight changes occurring with the
restrictions are. More than one output variable can weights into and out of the rules.
be used in a FZ-NN structure and the different
output variables can have different numbers of
membership functions.

Figure 3. Initial membership functions of a variable x in

the FZ-NN and the membership functions after
Figure 2. A FZ-NN structure for two initial fuzzy rules
adaptation (dotted lines).
Rule 1: IF x1 is A1 (DI1,1) and x2 is B1 (DI2,1) 4. SYSTEM MODELING
THEN y is C1 (CF1).
Rule 2: IF x1 is A2 (DI1,2) and x2 is B2 (DI2,2) The suggested FZ-NN has a flexible
THEN y is C2 (CF2), where Ds are architecture, which allows for different training and
degrees of importance attached to the adaptation strategies to be tested before the most
condition elements and CFs are confidence suitable is selected for a certain application. Some
factors attached to the consequent parts of of the issues involved in this adaptation are
the rules. discussed below:
One of the advantages of the FZ-NN 4.1. Initialization
architecture is that it manages to provide a fuzzy
logic system without having to unnecessarily Uniformly distributed triangular
extend the traditional MLP. Since standard transfer membership function can be used as initial values
functions, linear and sigmoidal, are used along with for the input variables, and uniformly distributed
a slightly modified back-propagation algorithm, the

52
 Journal of Theoretical and Applied Information Technology

© 2005 - 2010 JATIT. All rights reserved.

www.jatit.org

singletons can be used as initial values for the to exhibit long-term trends with occasional
output variables. These are the defaults that are departures from that trend. However except in
used in the absence of other information. unusual circumstances, such as a large financial
market crash, the long-term pattern will eventually
4.2.Membership function insertion be restored. The fuzzy condition making is as
suggested below:
If some expert knowledge is available then
this can be used to initialize the memberships, or The core of the rule set of the implemented
atleast initialize those for which knowledge exists fuzzy expert system is illustrated as follows:
with the remained being initialized using the default
method. 1) Detection: For detection, one rule follows.
Rule 1: If THDn is A2 or PSn is B2 or Vn is C3 or Vn
4.3.Rule insertion is C1 then DETECT=1

If initial set of rules is available, it is used 2) Classification: Fifteen rules follows.

for initialization of the FZ-NN structure through Rule 1: Vn1 is A4 and Nn is F1 and OSn is G1 then
rules insertion mode. The rules are represented as IMPULSE=1
weights as well as inserting the existence of a rule, Rule 2: Vn is A1 or Vn1 is A1 then OUTAGE=1
the relative importance of that rule and its Rule 3: Vn is A6 or Vn1 is A6 then SWELL=1
sensitivity to input variables can be provided. Rule 4: Vn is A5 and PSn is C1 and PSn1 is C1 and
EWn1 is D1 and {TSn1 is H2 or [TSn1 is H4
4.4.Training & TSn2 is H1]} then SWELL=1
Rule 5: Vn1 is A5 and {PSn is C2 or PSn1 is C2}
The FZ-NN can be accomplished either for then SWELL=1
the inner two rule weight layers, in which case the Rule 6: Vn1 is A2 then SAG=1
system adapts its fuzzy rules but does not adapt the Rule 7: Vn1 is A3 and {PSn is C2 or PSn1 is C2}
membership functions, or for the four weight then SAG=1
layers, in which case the system adapts both the Rule 8: Vn1 is A3 and {PSn is C1 and PSn1 is C1}
rules and the membership functions. The only and {THDn1 is B1 or [THDn1 is B2 and
difference between these two options is that the OSn1 is G4]} then SAG=1
connections in the fuzzification and defuzzification Rule 9: Vn1 is A3 and PSn is C1 and PSn1 is C1
layers are “frozen” in the former case and they are and OSn is G2 and THDn1 is B2 and
subject to change in the latter case. THDn2 is B2 and THDn3 is B2 then
NOTCH=1
4.5.Aggressive Rule 10: Vn1 is A3 and Nn is F2 and OSn is G2 then
NOTCH=1
A section of new data is used for further Rule 11: Vn1 is A4 and PSn is C1 and PSn1 is C1 and
training and adaptation without using any of the THDn is B3 and THDn3 is B1 and {OSn is
old, previously used, data. G4 or OSn1 is G4} then TRANSIENT=1
Rule 12:Vn1 is A4 and TSn1 is H3 and TSn2 is H3
4.6.Conservative and TSn3 is H3 and OSn1 is G4 then
HARMONIC=1
New data is added to the entire old data Rule 13:THDn1 is B4 and THDn2 is B4 and THDn3
and training is performed on the entire set. is B4 and OSn2 is G4 then HARMONIC=1
Rule 14:TSn1 is H4 and TSn2 is H4 and TSn3 is H4
Obviously, these concepts of and OSn2 is G4 then HARMONIC=1
aggressiveness and conservatism in training are Rule 15:If RNisK1 then FLICKER=1
fuzzy. In fact, some compromise of using the new
data with a percentage of old data tends to be most In the above rules, Ai, Bi, Ci, Di, Fi, Gi, Hi,
efficient. The amount of old data retained depends and Ki are the membership functions for the input
on operating requirements (since for on-line patterns, and the following trapezoidal and
adaptation using a large data set may not be triangular functions are used.
feasible), the stationary of the relationships, and the
length of time that changes tend to persist for until
returning, if at all, to the original relationships. An
example here is that of the stock market. This tends

53
 Journal of Theoretical and Applied Information Technology

© 2005 - 2010 JATIT. All rights reserved.

www.jatit.org

5. SIMULATION RESULTS

A number of power quality events of

various types have been simulated and
corresponding waveforms are obtained. The
following eight distinct features inherent to
different types of power quality events have been
extracted: the Fundamental Component (Vn), Phase
Angle Shift (δn), Total Harmonic Distortion
Figure 5. The passed query values for
(THDn), Number of Peaks of the Wavelet the developed electrical system
Coefficients (Nn), Energy of the Wavelet
Coefficients (EWn), Oscillation Number of the
Missing Voltage (OSn), Lower Harmonic Distortion
(TSn), and Oscillation Number of the rms
Variations ( RN ). The formulae for computing
these features are given as follows:
Vn = 2absVn(⎡⎣1⎤⎦) / N
αn = angle(Vn ⎡⎣1⎤⎦) − angle(V1[1])
int(N/2)
Figure 6. Passed three phase pulse as observed at
∑ { }
2
abs(Vn ⎡⎣k ⎤⎦) the generation side
THDn =
k =2
V1 ⎣⎡1⎦⎤
Nn = peak(abs(WC5 ))
le
EWn = ∑ abs ( WC [k])
k =1
n

OSn = root V5miss ( )

10

∑ {abs(V }
2
n
TSn = ⎡⎣k ⎤⎦)
k =2
V1 ⎣⎡1⎦⎤
Figure 7. The distorted phase observed in
(
RN = root V srms − mean(V srms ) ) the transmission line due to fault introduced

For the evaluation of the suggested work a

distribution system with the DFR units interfaced at
the load side is developed as shown in figure 4.
Bus2 Bus3
Feeder1 DFR1
Es Load1
Figure 8. The discrete feature coefficient observed for
Feeder2 Bus4
~ the three current pulses after distortion
Load2
Error Voltage
300

200
Feeder3 Bus5 DFR2 100

Bus1 Load3
0
Amplitude

-100
Fault
-200

-300

-400
Figure 4. Distributed system considered for -500
the electrical system 0 100 200 300 400 500
Time
600 700 800 900 1000

Figure 9. Effective phase voltage observed for

the developed system with sag effects

54
 Journal of Theoretical and Applied Information Technology

© 2005 - 2010 JATIT. All rights reserved.

www.jatit.org

250

200
REFERENCES
150

100
[1] Yamakawa, T., Kusanagi, H., Uchino, E. and
50

0
Miki, T., “A new Effective Algorithm for Neo
-50 Fuzzy Neuron Model”, Proceedings of Fifth
100 IFSA World Congress, (1993) 1017-1020.
150
[2] Hashiyama, T., Furuhashi, T., Uchikawa, Y.,
“A Decision Making Model Using a Fuzzy
200

250

Neural Network”, Proceedings of the 2nd

0 100 200 300 400 500 600 700 800 900 1000

Figure 10. The effective phase observation obtained International Conference on Fuzzy Logic &
after the reference model comparison Neural Networks, Iizuka, Japan, (1992) 1057-
250
Unbalance Voltage
1060.
200 [3] Kasabov, N., “Learning fuzzy rules and
150 approximate reasoning in Neuro-Fuzzy hybrid
100
systems”, Fuzzy Sets and Systems, 82 (2),
Amplitude

50
1996, pp.1-19
0
[4] Kasabov, N., Foundations of Neural Networks,
-50
Fuzzy Systems and Knowledge Engineering,
The MIT Press, CA, MA, 1996
-100

-150
0 100 200 300 400 500
Time
600 700 800 900 1000
[5] Kasabov, N., “Investigating the adaptation and
Figure 11. The compensated voltage observation after forgetting in fuzzy neural networks by using
application of Neuro modeling for the compensation the method of training and zeroing”,
250
Load Voltage
Proceedings of the International Conference
200
on Neural Networks ICNN'96, IEEE Press,
Washington DC, June 3-6, 1996
150

100

50
[6] Hauptmann, W., Heesche, K., “A Neural Net
Topology for Bidirectional Fuzzy-Neuro
0

-50

-100
Transformation”, Proceedings of the FUZZ-
IEEE/IFES, Yokohama, Japan, (IEEE, 1995),
-150

-200

-250
0 100 200 300 400 500 600 700 800 900 1000 1511-1518.
Figure 12. The compensated load voltage observed after [7] Goldberg, D., Genetic Algorithms is Search,
the correction applied to the developed system Optimization and Machine Learning , Addison
Wesley, 1989
6. CONCLUSIONS [8] Mang, G. Lan, H., Zhang, L., “A Genetic-
based method of Generating Fuzzy Rules and
In this paper a fuzzy neural network Membership Functions by Learning from
architecture for digital fault processing in Examples”, Proceedings of International
distributed power system is introduced. The Conference on Neural Information Processing
adaptive learning algorithms used along with the (ICONIP ’95) Volume One, pp 335 - 338.
rule extraction and rule insertion techniques show 1995.
that this is a promising approach to building [9] Gray, A. R., Kasabov, N., “Round-Trip System
adaptive intelligent information processing systems, Engineering in Neuro-Fuzzy Hybrid Systems”,
which is suitable for fault diagnosis, is developed. Journal of Intelligent and Fuzzy Systems, 1996.
The developed system is evaluated over a fault [10]Carpenter, G., “Distributed Learning,
current signal extracted with the resolution feature Recognition, and Prediction by ART and
description and passed to fuzzy neuro model for the ARTMAP Neural Networks”, Proceedings of
classification and detection of fault condition in the ICNN’96 , IEEE Press, Volume “Plenary Panel
observing power system. The effects of harmonic, and Special Sessions”, pp 244-249.
power disturbances were studied and were observed
to be improved by the suggestive approach.

55