Content uploaded by Samaher Al-Janabi
Author content
All content in this area was uploaded by Samaher Al-Janabi on Oct 21, 2017
Content may be subject to copyright.
Design and evaluation of a hybrid system for detection and prediction
of faults in electrical transformers
Samaher Al-Janabi
a,
⇑
, Sarvesh Rawat
b
, Ahmed Patel
c,d
, Ibrahim Al-Shourbaji
e
a
Department of Information Networks, Faculty of Information Technology, University of Babylon, Babylon 00964, Iraq
b
School of Electronics and Electrical Engineering (SELECT), VIT University, Vellore 632014, India
c
School of Computer Science, Centre of Software Technology and Management (SOFTAM), Faculty of Information Science and Technology (FTSM),
Universiti Kebangsaan Malaysia, 43600 UKM Bangi, Selangor Darul Ehsan, Malaysia
d
School of Computing and Information Systems, Faculty of Science, Engineering and Computing, Kingston University, Kingston Upon Thames KT1 2EE, United Kingdom
e
Computer Network Department, Computer Science and Information System College, Jazan University, Jazan, Saudi Arabia
article info
Article history:
Received 17 May 2014
Received in revised form 4 November 2014
Accepted 1 December 2014
Keywords:
Dissolved Gas-in-oil Analysis (DGA)
Electrical transformer
Fault detection
Fault prediction
Genetic algorithm
Neural network
abstract
Transformers are the vital parts of an electrical grid system. A faulty transformer can destabilize the elec-
trical supply along with the other devices of the transmission system. Due to its significant role in the
system, a transformer has to be free from faults and irregularities. Dissolved Gas-in-oil Analysis (DGA)
is a method that helps in diagnosing the faults present in an electrical transformer. This paper proposes
a hybrid system based on Genetic Neural Computing (GNC) for analyzing and interpreting the data
derived from the concentration of the dissolved gases. It is further analyzed and clustered into four sub-
sets according to the standard C57.104 defined by IEEE using genetic algorithm (GA). The clustered data is
fed to the neural network that is used to predict the different types of faults present in the transformers.
The hybrid system generates the necessary decision rules to assist the system’s operator in identifying
the exact fault in the transformer and its fault status. This analysis would then be helpful in performing
the required maintenance check and plan for repairs.
Ó2014 Elsevier Ltd. All rights reserved.
Introduction
A transformer is one of the most crucial element of an Electrical
Power Transmission System (EPTS). A fault in the transformer can
introduce major problems for the consumers as well as for the
maintenance engineers. Many incidents have taken place in the
past few years that greatly disrupted the electrical transmission
system. One such catastrophe occurred in New Jersey, USA, in
December 2013, where, approximately 12,000 people lost their
power supply due to a fault in the transformer [10]. Another major
incident took place on February 2014 in Stamford, USA, where a
transformer caught fire rendering more than 1000 people without
light for days [20]. In the year 2000, a disastrous loss was reported
at another power plant, where a $86 million US dollars business
was interrupted due to a faulty transformer [12].
There is an urgent need of a prefailure analysis and protection
system that can protect the transformers from any kind of
liabilities. Analysis of the transformer’s dielectric oil is the classical
and reliable method used for checking the irregularities present in
the transformers by using the Dissolve Gas-in-oil Analysis (DGA)
method. Several gases are generated during the normal operation
of a transformer. The ratio and concentration of certain gases facil-
itate the operator in the detection and prediction of the indiscre-
tion and problems that exists in the transformers. The main
gases responsible for the faults are methane (CH
4
), acetylene
(C
2
H
2
), ethane (C
2
H
6
), and ethylene (C
2
H
4
)[13]. Problems like
corona discharge, overheating, and arcing in the transformers are
easily detected by DGA.
There are several methods available to analyze the faults, such as
the (i) International Electro technical Commission (IEC) ratio
method, (ii) Rogers ratio method, (iii) Doernenburg method, (iv)
Duval triangle method, and the Key gas method. The first three
methods do not give any sort of quantitative indication of the fault.
In many cases, where multiple faults occur, gases produced from
different types of faults are mixed up, creating confusing ratios
among the various components of the gases. For our analysis, we
will follow the IEEE standard C57.104, based on the Total Dissolve
Concentration of Gases (TDCG) and the Key gas method. It measures
the concentration of each fault gas produced in the transformer
http://dx.doi.org/10.1016/j.ijepes.2014.12.005
0142-0615/Ó2014 Elsevier Ltd. All rights reserved.
⇑
Corresponding author.
E-mail addresses: Samaher@itnet.uobabylon.edu.iq (S. Al-Janabi), sss.
sarvesh888@gmail.com (S. Rawat), whinchat2010@gmail.com (A. Patel), i_shurbaji@
yahoo.com (I. Al-Shourbaji).
Electrical Power and Energy Systems 67 (2015) 324–335
Contents lists available at ScienceDirect
Electrical Power and Energy Systems
journal homepage: www.elsevier.com/locate/ijepes
during a fault. In this method, the individual concentration of each
gas is measured rather than the ratio which is the basic principle of
this method. The use of DGA in the transformer is widely accepted
for analyzing and spotting the faults as it can diagnose the degrada-
tion of the transformer and can estimate its life efficiency [16].In
addition, it can appraise the internal situation of the transformer
and plays a crucial part of the maintenance checking and testing
system.
Soft computing is a consortium of methodologies that works
synergistically and provides, in one form or another, flexible infor-
mation processing capability for handling real-life ambiguous situ-
ations. It aims to exploit the tolerance for imprecision, uncertainty,
approximate reasoning, and partial truth in order to achieve trac-
tability, robustness, and low-cost solutions. The guiding principle
is to devise methods of computation that leads to an acceptable
solution. Several methods have been devised for using Artificial
Intelligence (AI) and Soft Computing (SC) for more advanced and
accurate diagnosis of transformers [4,17]. In 2012, Souahlia et al.
used fuzzy logic, Support Vector Machine (SVM) and Neural Net-
works (NN) for fault diagnosis in the transformers [18]. Way back
in 1997, Huang et al. showed the use of fuzzy logic for diagnosing
the faults in the transformer [22]. A set of induced rules was gen-
erated from a quantitative data using a fuzzy set based learning
algorithm [15]. But the membership function used in fuzzy is not
suitable for representing the boundary value conditions [5,6].In
2005 Ganyun et al. used SVM for identifying the faults in the trans-
formers [19]. It provides a three layered classifier for classifying the
state of the transformer. Although it showed a good reliability and
is suitable for online fault diagnosis, but the selection of the exact
kernel function and the optimization of parameters to make a SVM
classifier is a typical problem. The main problem with all these
methods is that they are mostly suitable for a transformer having
a single fault or any dominating fault. There is no application
focusing on the prediction of faults and real trend analysis.
There are several problems associated with an electrical trans-
former, such as, overloading, overvoltage, overheating and other
factors that ultimately lead to a permanent failure. As such, there
is a major need of monitoring the parameters associated with the
transformer to prevent it from shutting down. Therefore, there is
an acute need of new technologies which can monitor the supply
systems more effectively to prevent them from unexpected and
unconditional failures. Soft Computing (SC) hybridization is an asso-
ciation of computing methodologies centering on Fuzzy Logic (FL),
Neural Computing (NC), Genetic Computing (GC), Probabilistic com-
puting (PC) and their hybridization [1–3]. Collectively, these meth-
odologies provide a foundation for the conception, design and
deployment of the intelligent systems. The basic idea underlying
SC is that its constituent methodologies are, for the most part,
complementary rather than competitive. The complementarity of
the constituents of soft computing implies that their effectiveness
may be enhanced by using them in combination rather than isola-
tion. At this juncture, the most visible systems of this combined
type are the neuro-fuzzy systems. Less visible, but potentially of
equal importance are the fuzzy-genetic systems. Each of the con-
stituents of soft computing has a set of capabilities to offer. In
the case of fuzzy logic, it is the machinery for dealing with impre-
cision, information granulation and computing with words. For
this purpose, the principal tools are provided by the fuzzy logic
center on the use of linguistic variables and the calculation of fuzzy
based ‘‘if-then’’ rules. In the case of genetic computing, the princi-
pal tool is a systematized random search. The most known meth-
ods of hybridization of these tools are (i) Neural-Fuzzy
Computing, (ii) Fuzzy Genetic Computing, (iii) Genetic-Neural
Computing (iv) and Neuro-Genetic-Fuzzy Computing.
In this work, we have used Genetic-Neural Computing using
DGA analysis, where the challenge is to build a practical neural
network choosing the right architecture and the right learning
parameters to find the faults present in the transformers [13].
We know that the Multilayer Perceptron (MLP) with one hidden
layer, using the sigmoid transfer function, could perform any map-
ping from a set of inputs to the desired outputs. Unfortunately, this
tells us nothing about the learning parameters, the necessary num-
ber of neurons, or whether any additional layers would be benefi-
cial. It is, however, possible to use a genetic algorithm to optimize
the network design. A suitable cost function might combine the
root mean square error with the duration of training [2]. Super-
vised training of a neural network involves adjusting its weights
until the output patterns are obtained for a range of input patterns.
They must be as close as possible to the desired patterns. The dif-
ferent network topologies use different training algorithms for
achieving this weight adjustment, typically through back-propaga-
tion or errors. However, it is also possible to use GA for training the
network. This can be achieved by allowing each gene to represent a
network weight so that a complete set of network weights is
mapped onto an individual chromosome. Each chromosome can
be evaluated by testing a neural network with the corresponding
weights against a series of test patterns. A fitness value can be
assigned according to the error so that the weights represented
by the fittest generated individual corresponds to a trained neural
network [3–5]. The most crucial part of using neural network in
our system lies in the fact that it can learn and update its knowl-
edge whenever it is required [8,9]. It offers a far superior perfor-
mance than the other systems due to the non-linear mapping
property of the neurons. Following this model, the operator will
be able to conduct prefailure analysis and plan for the required
maintenance checks.
The rest of the paper is structured as follows: Section ‘Cause of
gas formation’ presents the cause of gas formation. Section ‘Need of
a hybrid system’ presents the main tools used in the hybrid system,
while in Section ‘Main stages of the suggested hybrid system’, the
suggested hybrid system that contains various stages are
explained. Section ‘Experiment’ shows the experiments. Finally,
the conclusion of the paper is presented in Section ‘Conclusion’.
Cause of gas formation
The main and the most profound cause of gas formation in the
transformer is thermal heating and electrical discharges. It decom-
poses the oil into different gases like CO, CO
2
,C
2
H
2
,C
2
H
4
,C
2
H
6
,H
2
,
and CH
4
. The cellulose and the minerals present in the transformer
oil decompose to produce these gases as shown in Fig. 1. The
decomposition of cellulose produces carbon oxides, methane and
some hydrogen. The rate of production of these gases abruptly
increases with the increase in temperature and volume of the
material present in the oil.
Beta fluid and mineral oil consist of a variety of hydrocarbon
molecules. They decompose into active hydrogen atoms and
Fig. 1. Composition of the gases evolved during a normal functioning of a
transformer.
S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335 325
fragments of hydrocarbons which combine to form new molecules.
The further rearrangement and decomposition of molecules lead to
the formation of other gases like acetylene and ethylene. The con-
centration of these gases is analyzed by the DGA covered in the
next section. It has to be monitored on a regular basis so that the
inconsistencies in the transformer can be scrutinized properly
[14].Table 1 shows the principle gas evolved during the thermal
and electrical decomposition of the beta fluid and cellulose.
Need of a hybrid system
The conventional methods, like the IEC ratio method, Rogers
ratio, Doernenburg method and the Key gas method highly depend
on human expertise and skills of the operator. The operator has to
thoroughly inspect the concentration of the gases. He is required to
compare the output results from the different methods to derive a
conclusion. So, a huge expertise is needed for the operator to ana-
lyze the results and avoid the conflicts. Sometimes, the possible
number of different combinations of codes exceeds the fault types.
Thus, the traditional DGA methods do not offer any absolute or
objective type of result. AI based fault diagnosis can become an
additional asset here. The aim of the proposed system is to draw
the conclusions for the system’s operator by analyzing the state
of the transformer, so that he can take further steps and can plan
for maintenance [11]. NN and GA have been widely used in solving
many real time problems [9]. The whole system is adaptive in nat-
ure. NN can successfully reveal the explicit relationship between
the non-linear input–output data. It can find the patterns from
the input training data and can increase its learning and adaptabil-
ity for the new set of obtained data. The adopted method is more
effective and acclimative as compared to the conventional method
of fault diagnosis. It can produce more efficient results showing
better performance than the other methods. The proposed network
following the least error function, can exclaim the best possible
guess about the functionality of the transformer under a given con-
dition. The most significant advantage of using this method is that
it eliminates the boundary type problems which results in the ‘‘No
Decision’’ type cases that are mostly found in conventional meth-
ods. The system can autonomically directly self-learn from the
input variables and update itself according to its necessity.
Fig. 2 shows the basic steps that are followed in the proposed
system. There are 4 basic steps that are involved in the whole pro-
cess. The first step includes the analysis of the transformer oil and
finding the concentration of the different gases present in it [21].
The second step features the data pre-processing unit and the
use of GA for clustering the concentration of the different gases.
These gases are clustered on the basis of four conditions of the
standard C57.104 defined by IEEE [7]. In the third step. ANN is used
to predict the value of the fault using the derived clusters of GA.
Finally, the decision rules are generated for the system’s operator
that are inspected and analyzed by using different statistical
techniques.
Tools used in the hybrid system
This section discusses the main tools that are used for building
the hybrid system.
A. Dissolved Gas-In-Oil Analysis
DGA is one of the most important diagnostic tests performed on
the transformer oil in order to determine the state of the power
transformer [15]. We can also detect very low concentration levels
of the harmful gases [14].Fig. 3 shows the process of DGA that is
used for analyzing the concentration of the gases.
This technique involves the stripping of gases from transformer
oil and infusing them into a gas chromatograph. A sample of the oil
is taken using a gas tight syringe of appropriate capacity. This syr-
inge is capable of taking a sample of the oil from the main stream
point of the transformer. It is stored in a dark enclosure to prevent
the oxidation of gases. The next phase includes the extraction of
gases from the sample. In the final step, the sample is subjected
to gas chromatography. This is used for separating the different
constituents of the gases from a mixture. Fig. 4 shows the whole
process involved in the gas chromatography.
The use of DGA in the transformer is widely accepted for ana-
lyzing and spotting the faults as it can diagnose the degradation
of the transformer and can estimate its life expectancy. In addition,
it can appraise the internal situation of the transformer and is a
crucial part of the maintenance checking and testing system.
B. Genetic algorithms
Genetic algorithms (GAs) are a heuristic approach used to find
approximate solutions for the problems that are difficult to solve
by applying the principles of evolutionary biology to computer sci-
ence. Genetic algorithms use biologically-derived techniques such
as inheritance, mutation, natural selection, and recombination (or
crossover). Genetic algorithms are a particular class of evolutionary
algorithms.
GAs are typically implemented as a computer simulation in
which a population of abstract representations (called chromo-
somes) of candidate solutions (called individuals) to an optimization
problem evolving towards better solutions. Traditionally, solutions
are represented in binary as strings of 0s and 1s, but different enco-
dings are also possible. The evolution starts from a population of
completely random individuals and happens in generations. In
each generation, the fitness of the whole population is evaluated,
multiple individuals are stochastically selected from the current
population based on their fitness and modified mutated or recom-
bined to form a new population, which becomes current in the
next iteration of the algorithm.
Main stages of the suggested hybrid system
Soft computing methodologies have been applied to handle the
different challenges posed by a database. The main constituents of
soft computing, in this paper, include Detection, GA and NN. Each
of them contributes a distinct methodology to address the prob-
lems in its domain. This is done in a cooperative, rather than a
competitive, manner. The result is a more intelligent and robust
system providing a human-interpretable, low cost, approximate
solution, as compared to the traditional techniques.
Stage 1: fault detection
Every transformer generates certain gases during its operation.
The generation of the combustible gases is a result of various fac-
tors like overheating, corona discharge and dielectric problems.
These associated abnormalities are termed as faults. For example,
when cellulose is overly heated it produces hydrogen (H
2
), meth-
ane (CH
4
), carbon dioxide (CO
2
) and carbon monoxide (CO). Gases
like ethane (C
2
H
6
), acetylene (C
2
H
2
), and ethylene (C
2
H
4
) are pro-
duced in beta fluid by internal faults. The presence of these gases
indicates the occurrence of one or more combination of these
Table 1
Principal gas evolved during a fault.
Decomposition Thermal Electrical
Fault Overheating
of oil
Overheating of
cellulose
Corona
discharge
Arcing
Principle Gas Ethylene Carbon
monoxide
Hydrogen Acetylene
326 S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335
(electrical, corona or thermal) faults. The concentration of all the
gases is determined by the gas chromatography [21]. The whole
analysis results in categorizing the fault as either a thermal fault
or an electrical fault. It is further classified according to the high
and low intensity of the faults:
Thermal faults generally produce gases of low molecular weight
like H
2
,CH
4
and small quantities of other compounds having
higher molecular weight, namely acetylene, comprising of all
the mineral oils and beta fluid. On the other hand, thermal
decomposition of cellulose produces carbon dioxide (CO
2
) and
carbon monoxide (CO).
Electrical faults of low intensity such as intermittent arcing and
partial discharge, mainly produce hydrogen (H
2
) along with small
quantities of acetylene (C
2
H
2
) and methane (CH
4
). The concentra-
tion increases with respect to the intensity of the discharge.
In the case of electrical faults of high intensity or arcing, a large
amount of acetylene becomes predominant in the system. The
temperature of the system exceeds 700 C.
By measuring the concentration of the gases, we can identify
the kind of fault involved, as shown in Table 2.
Stage 2: pre-processing of the gas database
Fault diagnosis is generally considered as a boundary set
problem as the dataset consists of many inconsistencies. In this
scenario, training a neural network is very difficult. As such,
there is a huge need of pre-processing the data before feeding
it to the NN. The extracted database from the above stage
is pre-processed using a Linear Transformation method as
follows:
Stage 1:Extraction Stage
DGA chamber
Stage 2: Preprocessing Stage
Stage 3:Processing Stage
Stage 4: Decision Stage
Transformer Sampling of Oil Extraction of
Gas
Gas
Chromatograph
Preprocessing
the Data
Clustering of
Data using GA
Gas
Chromatograph
Training Neural
Network
Prediction
of Faults
Decision
Rules
Operator
Fig. 2. Proposed hybrid architecture for fault diagnosis.
Sampling of Oil Extracon of
Gases
Gas
Chromatography
Analysis of
Chromatograph
Concentraon of
Gases
Fig. 3. Steps followed in finding the concentration of the gases.
S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335 327
Here;L
0
¼½ðLminÞ=ðmax minÞ ðmax
0
min
0
Þþmin :
where min is the old minimum value, min
0
is new minimum value,
max is the old maximum value and max
0
is the new maximum
value.
Stage 3: genetic algorithm for clustering the database according to
standard C57.104 defined by IEEE
In this step, GA is applied to find the number of clusters existing
in the Gas database (i.e. find the best seed for each cluster and the
number of pixels on it). Before this, we need to determine the
parameters of GA, such as the population size, minimum number
of cluster, selection, and the crossover methods. Fig. 5 shows the
flowchart of GA for clustering the Gas Database.
Stage 3.1: representation (encoding of solution)
The chromosomes are made up of list pointers. If the pointer at
any gene is not null, that means there is a supposed center. This
center is drawn randomly from the data set. On the other hand,
gene (pointer) with null mean, has had no center encoded in it.
The value of Kis assumed to lie in the range [K
min
;K
max
], where
K
min
is chosen to be 2 unless specified otherwise. The length of a
string is taken to be K
max
, where each individual gene position rep-
resents either a pointer to the actual center or a null.
Stage 3.2: population initialization
For each string iin the population (i=1,......,P, where Pis the
size of the population), a random number Ki in the range
[K
min
–K
max
] is generated. This string is assumed to encode the
centers (each center represents a weight of node of Back-Propaga-
tion Neural Network) (BPNN) of Ki clusters. For initializing these
centers, Kid points are chosen on the basis of the four conditions
from the dataset. These points are distributed randomly in the
chromosome.
Stage 3.3: fitness computation [23]
The fitness of a chromosome is computed using the Davies–
Bouldin index. This index is a function of the ratio of the sum of
within-cluster scatter to between-cluster separation. The scatter
within C
i
, the ith cluster, is computed as:
S
i;q
¼1
jC
i
jX
x2C
i
kxz
i
k
q
2
!
1=q
where z
i
is the centroid of C
i
, and is defined as:
Z
i
¼1=n
i
X
x2C
i
x
and n
i
is the cardinality of C
i
(i.e., the number of points in cluster C
i
).
The distance between cluster C
i
and C
j
is defined as:
d
ij;t
¼X
p
s¼1
jz
is
z
js
j
t
"#
1=t
¼kz
i
z
j
k
t
Specifically, S
i,q
used in this article, is the average Euclidean distance
of the vectors in class ito the centroid of class i. While d
ij,t
is the
Minkowski distance of order tbetween the centroids that character-
ize clusters iand j(i.e., in this work, we use t= 4). Subsequently, we
compute:
R
i;qt
¼max
j;j–i
s
i;q
þs
j;q
d
ij;t
The Davies–Bouldin (DB) index is then defined as:
DB ¼1
KX
k
i¼1
R
i;qt
The objective is to minimize the DB index for achieving proper clus-
tering. The fitness function for chromosome jis defined as 1/DB
j
.
Fig. 5 shows the flowchart of the GA method used for clustering
the gases database.
Fig. 4. Gas chromatography for DGA analysis.
Table 2
Categorization of fault gases.
Corona Pyrolysis Arching
Oil Cellulose Oil Cellulose
Low temperature High temperature Low temperature High temperature H
2
C
2
H
2
(CH
4
C
2
H
6
C
2
H
4
)
H
2
H
2
CO CO
2
CH
4
C
2
H6 C
2
H
4
H
2
(CH
4
,C
2
H
6
)CO
2
(CO) CO (CO
2
)
328 S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335
Stage 4: applying the Back-Propagation Neural Network (BPNN) to
predict the fault values
The following main steps are executed to train the BPNN [24]:
Step 4.1: Input initial values to learning rate (
g
0
), maximum
acceptable error to network (E
max
), maximum
number of epochs to learning network (E
pochmax
),
momentum rate (
a
).
Step 4.2: Put the network error value (MSE) equal to zero and
current training pattern error equal to one and
determine the learning rate value.
Step 4.3: Compute the hidden neurons activity by unipolar
sigmoid function, with k= 1, according to the
equation below:
h
k
¼fX
ns
i¼1
s
i
:
v
ik
!
where k¼1;2;......;n
h
:
Fig. 5. Flowchart of genetic algorithm for clustering.
S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335 329
START
Network error gives zero value and learning rate determines the epoch
Pass training pattern from hidden layer to the output layer and compute activity for each node.
Compute output nodes error of the pattern
Determine the cost function value
STOP
Yes
No
Pass training pattern across input layer to hidden layer and compute activity for each node.
Compute hidden nodes error of the pattern
Adjust weights between hidden layer and output layer
Adjust weights between input layer and hidden layer
Is training
pattern pass
completed ?
Is termination
criterion
achieved ?
No
Yes
Input initial values of network parameters: learning
rate, momentum rate, number of epochs
Fig. 6. Flowchart of BPNN for forecasting the fault value [24].
Fig. 7. The concentration of all the gases present in the transformer.
330 S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335
Step 4.4: Compute output neuron activity according to the fol-
lowing function:
o
j
¼fX
nh
k¼1
h
k
:w
kj
!
where j¼1;2;......;n
o
:
Step 4.5: Compute error signal value to output neurons of pat-
tern p according to the following equation:
d
j
¼ðd
j
o
j
Þ:
fðnet
j
Þ
we can find the derivative of function as follows:
fðnet
j
Þ¼ 1
1þexpðnet
j
Þ
fðnet
j
Þ¼o
j
:ð1o
j
Þ;where j¼1;2;...;n
o
;
Step 4.6: Compute the error signal value in hidden neurons
which depends on the output neurons error:
d
k
¼X
no
j¼1
d
j
:w
kj
:
f net
k
ðÞ;where k¼1;2;...;n
h
fðnet
k
Þ¼h
k
:ð1h
k
Þ
Step 4.7: Adjust weights between the hidden layer and the out-
put layer. To do this, error back propagation algorithm
uses a negative first derivative of the cost function
ratio to weight as follows:
D
w
kj
¼
g
o
:@E
@w
kj
¼
g
o
:
@0:5X
no
j¼1
d
j
o
j
2
@w
kj
;o
j
¼fðnet
j
Þ
¼
g
o
:
@0:5X
no
j¼1
d
j
fðnet
j
Þ
2
@w
kj
;net
j
¼X
nh
k¼1
w
kj
:h
k
¼
g
0
:ðd
j
o
j
Þ@fðnet
j
Þ
@w
kj
¼
g
o
:ðd
j
o
j
Þ@fðnet
j
Þ
@net
j
:@net
j
@w
kj
¼
g
o
:ðd
j
o
j
Þ:
fðnet
j
Þ:@net
j
@w
kj
¼
g
o
:ðd
j
o
j
Þ:
fðnet
j
Þ:h
k
¼
g
o
:d
j
:h
k
The adjustment equations:
D
w
ðtþ1Þ
kj
¼
g
:d
j
:h
k
þ
a
:
D
w
ðtÞ
kj
;
w
ðtþ1Þ
kj
¼w
ðtÞ
kj
þ
D
w
ðtþ1Þ
kj
where k=1,2,...,n
h
and j=1,2,...,n
o
, and
a
is the momentum rate
which is:
D
w
ðtÞ
kj
: that represent the difference between the current weight
and the prior weight.
Step 4.8: Adjust weights between the input layer and the hidden
layer as follows:
Dv
ik
¼
g
o
:@E
@V
ik
¼
g
o
:
@0:5X
no
j¼1
d
j
o
j
2
@
v
ik
¼
g
o
:X
no
j¼1
ðd
j
o
j
Þ@fðnet
j
Þ
@
v
ik
¼
g
o
:X
no
j¼1
ðd
j
o
j
Þ@fðnet
j
Þ
@net
j
:@net
j
@
v
ik
¼
g
o
:X
no
j¼1
ðd
j
o
j
Þ:
f net
j
:@net
j
@
v
ik
¼
g
o
:X
no
j¼1
d
j
:@net
j
@h
k
:@h
k
@
v
ik
¼
g
o
:X
no
j¼1
d
j
:w
kj
@h
k
@net
k
:@net
k
@
v
ik
;where net
k
¼X
ns
i¼1
v
ik
:s
i
¼
g
o
:X
no
j¼1
d
j
:w
kj
:fðnet
k
Þ:@net
k
@
v
ik
¼
g
o
:X
no
j¼1
d
j
:w
kj
:fðnet
k
Þ:s
i
¼
g
o
:X
no
j¼1
d
j
:w
kj
:h
k
ð1h
k
Þ:s
i
;where
fðnet
k
Þ¼h
k
:ð1h
k
Þ
¼
g
o
:d
k
:s
i
;where d
k
¼X
no
j¼1
d
j
:w
kj
:h
k
ð1h
k
Þ
The adjustment equations are:
Dv
ðtþ1Þ
ik
¼
g
o
:d
k
:s
i
þ
a
:
Dv
ðtÞ
ik
;
v
ðtþ1Þ
ik
¼
v
tðÞ
ik
þ
Dv
ðtþ1Þ
ik
where k=1, 2, ...,n
h
and i=1, 2, ...,n
s
, and
a
is the momentum
rate:
D
v
ðtÞ
ik
: represent the difference between the current weight and
the prior weight.
Step 4.9: Increase the value pby one to input the next pattern in
the learning process. If it does not reach to the maxi-
mum number of training the patterns then return to
step 3 to train the network on that pattern else trans-
form to step 10.
Step 4.10: After completing the input to all training patterns of
the network, compute the cost function value that is
represented by the mean square error:
MSE ¼1
2X
P
p¼1
X
no
j¼1
d
p
j
o
p
j
2
Step 4.11: In this step, the termination criterion is tested. This
condition is valid if the total error value of the net-
work becomes less than the expected error of it
(E
max
), or the current Epoch value (t) is bigger than
the maximum number of learning epochs (E
pochmax
).
Else, return to step 2.
Fig. 6 explains the flowchart of BPNN for forecasting/predicting
the fault values.
Stage 5: decision making process: rule generation
After verification of one of the stopping criteria to the BPNN
algorithm, such as the verified cost function condition or exceeding
the number of epochs to the maximum number of learning epochs
without reaching a network error to a value less than the required
value, we can say that the BPNN is complete.
S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335 331
If the cost function condition is verified, this means that the
network can train itself on the input pattern (i.e., the network
is successful in the training process). While, if the second condi-
tion is verified (i.e., the network does not reach to an acceptable
error and exceeds the number of epochs), this means that the
network fails in the training process and recognition of the input
pattern.
In this work, we provide discovered knowledge which has a cer-
tain predictive power. The basic idea is to predict the value of the
fault based on the previously observed data. In this context, we
want the discovered knowledge to have a high predictive accuracy
rate. The discovered knowledge has to be comprehensible for the
user. This is necessary whenever the predicted knowledge is to
be used for supporting a decision to be made by a user [6]. Knowl-
edge comprehensibility can be achieved by using high-level knowl-
edge representations. A popular one, in the context of making a
decision, is a set of:
IF-THEN (Prediction) rules, where each rule is of the form:
IF <some_conditions_are_satisfied> THEN
<its_belong_to_certain_class>
As a result, prediction rules,(if-then) have been widely used to
represent knowledge and they have the advantage of being easily
interpreted by human experts because of their modularity.
Experiment
In our system, we have analyzed the individual concentration of
the gases and the value of the Total Dissolved Combustible Gas
(TDCG), which is measured in parts per million (ppm) using the
Key gas method. In this method, four level criteria have been devel-
oped to categorize the faults and risks involved in the functioning
Fig. 8. Associated faults of the transformer.
Fig. 9. Pre-processed data.
332 S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335
of the transformer defined by the IEEE standard C57.104. The four
conditions are:
1. If TDCG is below 720 ppm, the transformer is working in a safe
state.
2. If TDCG lies in the range 721–1920 ppm, then it is working in a
slightly deviated condition. Further investigation is required if
any individual gas is found to be exceeding its specified level.
3. If TDCG lies in the range 1921–4630 ppm, it indicates
that decomposition is of high level. In such a scenario,
immediate action should be taken and any gas exceeding
its normal concentration should be investigated right
away.
4. If TDCG is greater than 4630 ppm, it suggests that there is
excessive decomposition of cellulose and oil. The transformer
will fail if it is allowed to work further.
Fig. 10. Results of BPNN.
Fig. 11. Results of the predicted stage.
S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335 333
The concentration of all the gases present in the transformer
used for the experiment is shown in Fig. 7. We have taken 80 dif-
ferent fault samples that are gathered from different sources and
publications [7,9,18].
Fig. 8 shows the associated faults that are present in the trans-
former that are classified according to the standard IEEE C57-104.
After acquiring the data, it is pre-processed and normalized for
further investigation. The following example shows how these
values are computed by considering the old data that ranges from
[0–100] to transform it to a more appropriate range [5–10]:
L
0
¼½ðL0Þ=ð100 0Þ ð10 5Þþ5
L
0
¼½L=1005þ5
L
0
¼ðL=20Þþ5
Let L¼0 Then L
0
¼5
If L¼10 Then L
0
¼ð1=2Þþ5¼ð1þ10Þ=2¼5:5:
Fig. 12. Comparison between the predicted and actual values of the faults.
Rule 1:IF (H2 IS 2182.35666) AND (CH4 IS 1553.25349) AND (C2H2 IS 98.33467) AND (C2H4
IS 241.75233) AND (C2H6 IS 180.36833) AND (CO IS 1542.62683) AND (C2O IS 11948.74998) THEN Fault
is 4.
Rule 2: IF (H2 is 1999.99997) AND (CH4 is 1500.97699) AND (C2H2 is 99.703) AND (C2H4 is 243.371)
AND (C2H6 is 185) AND (CO is 1555.00001) AND (C2O is 10999.9999) THEN Fault is 4.
Rule 3:IF (H2 is between ( 1820.63904 - 3000 ) ) AND (CH4 is between ( 1100.63199 - 1750 ) ) AND (C2H2 is
between ( 85.904 - 100.3355 ) ) AND (C2H4 is between ( 210.672 - 260.281 ) ) AND (C2H6 is between ( 155 -
190 ) ) AND (CO is between ( 1450.76099 - 1000 ) ) AND (C2O is between ( 10500.73012 - 7000.02 ) ) THEN
Fault is 4.
Rule 4:IF (H2 IS 3000) AND (CH4 IS 1750) AND (C2H2 IS 100.3355) AND (C2H4 IS 260.281) AND (C2H6
iIs 190) AND (CO IS 1000) AND (C2O IS 7000.02) THEN Fault is 2.
Rule 5:IF (H2 is 0) AND (CH4 is 0) AND (C2H2 is 0) AND (C2H4 is 0) AND (C2H6 is 0) AND (CO is 0)
AND (C2O is 0) THEN Fault is 2.
Rule 6:IF (CH4 IS 3100.16102) AND (C2H2 IS 183.1565) AND (C2H4 IS 457.74825) AND (C2H6 IS 322.648)
AND (CO IS 1875.19476) AND (C2O IS 12900.38761) THEN Fault is 3.
Rule 7:IF (H2 is 5800.00001) AND (CH4 is 3300.58904) AND (C2H2 is 188) AND (C2H4 is 500) AND (C2H6
is 350) AND (CO is 1950.77902) AND (C2O is 13000.00007) THEN Fault is 3.
Rule 8:IF (CH4 is between ( 2600.05502 - 1750 ) ) AND (C2H2 is between ( 165 - 100.3355 ) ) AND (C2H4 is
between ( 350.431 - 260.281 ) ) AND (C2H6 is between ( 260 - 190 ) ) AND (CO is between ( 1750 - 1000 ) )
AND (C2O is between ( 12000.92019 - 7000.02 ) ) THEN Fault is 3.
Rule 9:IF (H2 IS 608.50147) AND (CH4 IS 481.21928) AND (C2H2 IS 43.85818) AND (C2H4 IS 84.96748)
AND (C2H6 IS 80.10002) THEN Fault is 1.
Rule 10:IF (H2 is 556.732) AND (CH4 is 265.00002) AND (C2H2 is 43.102) AND (C2H4 is 84.771) AND
(C2H6 is 78) AND (CO is 478.742) AND (C2O is 3650.00009) THEN Fault is 1
Rule 11:IF (H2 is between ( 49.99995 - 3000 ) ) AND (CH4 is between ( 20.15202 - 1750 ) ) AND (C2H2 is
between ( 15.893 - 100.3355 ) ) AND (C2H4 is between ( 12.29 - 260.281 ) ) AND (C2H6 is between ( 15 - 190 ))
THEN Fault is 1.
Fig. 13. Prediction rules generated by the hybrid system.
Table 3
Different fault cases and maintenance schedule for the operator.
Faults Condition 1 Condition 2 Condition 3 Condition 4
TDCG level (ppm) <720 721–1920 1921–4630 P4630
Sample interval according to TDCG rate >30 10–30 <10 >30 10–30 <10 >30 10–30 <10 >30 10–30 <10
Monthly Quarterly Annual Monthly Monthly Quarterly Weekly Weekly Monthly Daily Daily Weekly
State of transformer Normal level Abnormal level Highly abnormal level Very highly abnormal level
334 S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335
Fig. 9 explains the results of normalization of all fields in a given
database. It has been scaled to the range [0:1].
After pre-processing the data, GA is used for clustering it. The
cluster seeds based on the above four conditions are given as:
The population size is 50 that is used for training the network,
and the chromosome size is 7 which represents the different gases
(H
2
,CH
4
,C
2
H
2
,C
2
H
4
,C
2
H
6
, CO and CO
2
). The maximum number of
expected clusters are 8 and the minimum number of expected
clusters are 2. The limit of the generation count is 50, and the num-
ber of detected clusters are 4. The DBi index is found to be
0.3646815551.
In the next stage, BPNN is used for predicting the values of the
associated faults. It consists of three layers; input layer contains
seven neurons, hidden layer contains six neurons and output layer
contains a single neuron. The ratio is shown as (7:6:1). The associ-
ated parameters are found to be as learning factor = 0.5, momen-
tum factor = 0.6, max accept errors = 0.05, max number of
iterations = 100. The BPNN is trained in 43 epochs and means
quare error is found to be 0004987; the result of prediction is
based on the testing stage as shown in Fig. 10.
In Stage 5, the predicted rules are generated that are shown in
Fig. 11.
Finally, Fig. 12 shows the comparison between the predicted
values of the faults obtained by the proposed system with the
actual faults. The Yaxis of the figure indicates the different catego-
ries of the experienced faults. While, the Xaxis indicates the differ-
ent 30 samples that are used in testing the proposed hybrid
system. The blue bar shows the actual values of the faults and
the red bar shows the predicted values of the associated faults. It
is clearly evident from the bar chart that the trained network has
achieved an output of high accuracy.
The irregularities present in the electrical transformers are pre-
dicted from the concentration of the unusual gases in the trans-
formers as per the rules generated in Fig. 13. Different
combinations of the concentration of gases define different cases
of the faults. These faults are divided into 4 different categories
as discussed in Section ‘Need of a hybrid system’. We have used
genetic neuron computing as the soft computing technique for the
analysis and prediction of the associated faults in the electrical
transformer.
A transformer is a pivotal part of the electrical power supply.
The maintenance of a transformer is a major issue for the opera-
tors. A fault detection inference engine is proposed in this paper
using AI techniques. Table 3 shows the different fault cases and
the state of the transformer. It helps the operator to determine
the required sample interval for DGA analysis and plan for the
maintenance. It gives a clear advanced idea to the operator about
the potential problems in the transformer. This estimation can help
him in the early planning and scheduling of the maintenance
activity [13,14].
Conclusion
The aim of this paper was to propose a hybrid system that could
be used for detection and prediction of the faults present in a trans-
former via soft computing methodologies, which involved neural
networks, genetic algorithms, and their hybridization. Every trans-
former generated certain types of gases during its operation. The
concentration of these gases were analyzed and classified into dif-
ferent groups. GA was used for clustering the input concentration
into four different fault conditions, according to the C57.104 stan-
dard defined by IEEE. BPNN was used to predict the faults present
in the transformer through generating decision rules for the oper-
ator. It strived to provide a low cost solution, thereby speeding up
the whole process. This system proved as robust in analyzing the
faults and issuing the maintenance check plans. Using this system,
the operator would be able to forecast and make more intelligent
and accurate decisions. For our future studies, we would in visage
to extend this work to implement it in a real life situation. The
effect of other failures caused due to mechanical disturbances
and other natural factors would also be analyzed and explored.
These additional features like recovery voltage, visual inspection
test, winding displacement and the partial discharge test would
be taken into account for a more efficient analysis.
References
[1] Konar. Artificial inelegant and soft computing: behavioral and cognitive of the
human brain. Florida: CRC Press; 2000.
[2] Mitra S, Mitra P, Pal SK. Data mining in soft computing framework: a survey.
IEEE Trans Neural Networks 2002;13(1).
[3] Fidelis MV, Lopes HS, Freitas AA. Discovering comprehensible classification
rules with a genetic algorithm. Congress on evolutionary computation – 2000
(CEC-2000). La Jolla, CA (USA); July 2000, p. 805–10.
[4] Santos R, Nievola JC, Freitas AA. Extracting comprehensible rules from neural
networks via genetic algorithms. In: 2000 IEEE Symp. on combinations of
evolutionary computation and neural networks (ECNN-2000). San Antonio, TX
(USA); May 2000, p. 130–9.
[5] Pal SK, Mitra S, Mitra P. Rough fuzzy MLP: modular evolution, rule generation
and evaluation. IEEE Trans Knowledge Data Eng 2001.
[6] Freitas AA. Book review: data mining using grammar-based genetic
programming and applications. Gen Program Evolvable Mach
2001;2(2):197–9.
[7] Duval M. Dissolved gas analysis: it can save your transformer. IEEE Electric
Insulat Magaz 1989;5(6):22–7.
[8] Yanming T, Zheng Q. DGA based insulation diagnosis of power transformer via
ANN. Proceedings of the sixth international conference on properties and
applications of dielectric materials, vol. 1; June 2000, p. 133–6.
[9] Castro A, Miranda V. Knowledge discovery in neural networks with application
to transformer failure diagnosis. IEEE Trans Power Syst 2005;20:717–24.
[10] Biermann J. Transformer Explosion Knocks Out Power In Northern N.J. 2013,
16 December 2013. <http://newyork.cbslocal.com/2013/12/16/transformer-
failure-knocks-out-power-in-northern-n-j/>.
[11] Bartley WH. Analysis of transformer failures. <http://www.bplglobal.net/eng/
knowledge-center/download.aspx?id=191>.
[12] James R, Bartley W. IEEE C57.140, IEEE guide for the evaluation and
reconditioning of liquid immersed power transformers, Draft 2003, 9 March
2003. <http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=04168522>.
[13] Rogers RR. IEEE and IEC codes to interpret incipient faults in transformers
using gas in oil analysis. IEEE Trans Electric Insulat 1978;13(5):349–54.
[14] DSI Ventures Inc. Dissolved gas analysis guide for transformers filled with beta
fluid, 3 January 2013. <http://dsiventures.com/2013/02/03/dissolved-gas-
analysis-guide-for-transformers-filled-with-beta-fluid/>.
[15] Abu-Siada A, Arshad M, Islam S. Fuzzy logic approach to identify transformer
criticality using dissolved gas analysis. In: Power and energy society general
meeting, 2010 IEEE. IEEE; 2010. p. 1–5.
[16] Snow T, McLarnon M. The implementation of continuous online Dissolved Gas
Analysis (DGA) monitoring for all transmission and distribution substations.
In: Electrical insulation (ISEI), conference record of the 2010 IEEE international
symposium on. IEEE; 2010. p. 1–4.
[17] Lin E, Ling JM, Huang CL. An expert system for transformer fault diagnosis and
maintenance using dissolved gas analysis. IEEE Trans Power Deliv
1993;8(1):231–8.
[18] Souahlia S, Bacha K, Chaari A. Artificial intelligence tools aided-decision for
power transformer fault diagnosis. Int J Comput Appl 2012;38.
[19] Ganyun LV, Haozhong C, Haibao Z, Lixin D. Fault diagnosis of power
transformer based on multi-layer SVM classifier. Electric Power Syst Res
2005;74(1):1–7.
[20] Gross M. Blown transformer causes power outage in Stamford. 2014. <http://
connecticut.news12.com/news/blown-transformer -causes-power-outage-in-
stamford-1.7185580>.
[21] Ucdavis.edu, ‘‘Gas Chromatography’’, <http://chemwiki.ucdavis.
edu/A naly tica l_Ch emistry /Ins trum enta l_A naly sis/ Chro matogra phy/ Gas_
Chromatography>.
[22] Huang YC, Yang HT, Huang CL. Developing a new transformer diagnosis system
through evolutionary fuzzy logic. IEEE Trans Power Deliv 1997;12(2):761–7.
[23] Kaghed N, Abbas T, Hussein Ali S. Design and Implementation of Classification
System for Satellite Images based on Soft Computing Techniques. In: IEEE
Information and Communication Technologies, 2006. ICTTA ’06. 2nd, vol. 1. p.
430–6. http://dx.doi.org/10.1109/ICTTA.2006.1684408.
[24] Kaghed N, Abbas T, Hussein Ali S. Designing a software for classifying objects
for air photos & satellite images using soft computing, Thesis of Master in
computer Science, Science College, Babylon University, Iraq, 2005.
S. Al-Janabi et al. / Electrical Power and Energy Systems 67 (2015) 324–335 335