AbstractNon-Destructive evaluation of in-service power

transformer condition is necessary for avoiding catastrophic failures.

Dissolved Gas Analysis (DGA) is one of the important methods.
Traditional, statistical and intelligent DGA approaches have been
adopted for accurate classification of incipient fault sources.
Unfortunately, there are not often enough faulty patterns required for
sufficient training of intelligent systems. By bootstrapping the
shortcoming is expected to be alleviated and algorithms with better
classification success rates to be obtained. In this paper the
performance of an artificial neural network, K-Nearest Neighbour
and support vector machine methods using bootstrapped data are
detailed and shown that while the success rate of the ANN algorithms
improves remarkably, the outcome of the others do not benefit so
much from the provided enlarged data space. For assessment, two
databases are employed: IEC TC10 and a dataset collected from
reported data in papers. High average test success rate well exhibits
the remarkable outcome.

KeywordsDissolved gas analysis, Transformer incipient fault,
Artificial Neural Network, Support Vector Machine (SVM), K-
Nearest Neighbor (KNN)
I. INTRODUCTION
OWER transformers are always under the impact of
electrical, mechanical, thermal and environmental stresses
that degrade their insulation quality. To avoid the power
failure, periodically monitoring of the conditions of
transformers is necessary. Results of early detection of fault
are large savings in operation and maintenance costs and
preventing any premature breakdown/failure.
There are routine maintenance procedure for power
transformers such as dissolved gas analysis (DGA), moisture
analysis in transformer oil [1, 2], oil breakdown voltage test,
the tan (delta) test, resistivity test, acidity test, sludge test,
interfacial tension test and partial discharge (PD) acoustic
emission sensing. Among these methods, DGA is an effective
one for the early detection of incipient faults [3].
It is well known that overheating, arcing, partial discharge,
winding circulating currents, and continuous sparking are the
main factors in deteriorating transformer condition. These
phenomena develop certain dissolved gaseous in the insulation
oil. The gases include hydrocarbons such as: methane (CH4),
ethane (C2H6), ethylene (C2H4), acetylene (C2H2) and others
such as: hydrogen (H2), carbon dioxide (CO2), and etc.


S. Qaedi is with Shahed University, Tehran, Iran (e-mail:
[email protected]).
S. Seyedtabaii is with Shahed University, Tehran, Iran (e-mail:
[email protected]).

The gases are extracted from the oil under high vacuum and
analyzed by Gas Chromatograph, to get each gas concentration
separately. By interpretation of the gas contents, the
developing faults in the power transformers can be diagnosed.
Many diagnostic criteria have been developed to establish
relationships between the gases and the fault conditions, which
some are obvious and some are not (hidden relationships). The
gas concentrations, generation rates, specific gas ratios, and
the total combustible gas are important parameters for
interpreting the result of DGA. But the actual diagnosis must
also consider other information of transformer such as size,
volume of oil, type of transformer etc. Therefore, much of
diagnostic relies on experts to interpret the results correctly.
To automatize the procedure of power transformer fault
classification, several algorithms have been studied. Presently,
the conventional ratio methods, statistical schemes and
artificial-intelligence (AI) methods are the major interpreting
approaches. The conventional ratio methods mainly include
Rogers Ratios [4], Duval Triangle [5], and IEC Ratios [6].
Since conventional ratios boundaries are sharp, they are
unable to provide interpretation for every possible
combination of ratio values [7].
The Artificial Neural Network methods have also been used
to explore the nonlinear and complex relation between the
gases concentration and the type of faults. Multilayer back
propagation (MLP) [8], [9], self-organizing map network [10],
Adaptive Back-propagation learning algorithm [11] and
Extension NN [12, 13] are among them. ANN training suffers
from trapping in local minima; therefore evolutionary training
algorithms have got deserved attention in this field [7], [14],
[15]. Other methods that have been investigated are wavelet
decomposition [15], SVM [15], KNN [15] and fuzzy learning
vector quantization network [16], [12].
In this paper SVM, KNN and ANN are considered for fault
classification. The data sets are IEC TC10 and database 1, a
collection of data gathered from many research papers. It is
observed that the number of faulty patterns for each class is
not equal and enough for well training of intelligent schemes.
As a remedy, the bootstrapping technique is employed to
equalize the number of faulty patterns for each class of the
transformer condition. The results show that the method based
on ANN takes more advantage of the bootstrapped input and
renders remarkable improvement in the performance while the
other two methods do not show such advances.
In section 2 the three classification methods are briefly
reviewed. Section 3 discusses data preparation and pre-
S. Qaedi, and S. Seyedtabaii
Improvement in Power Transformer Intelligent
Dissolved Gas Analysis Method
P
World Academy of Science, Engineering and Technology 61 2012
1144


processing. Simulations are detailed in section 4 and finally,
conclusion comes in section 5.
II. CLASSIFICATION METHODS
A. Intelligent Method: Neural network
What makes artificial neural net algorithms valuable is that
they can be taught to perform a particular task, such as
recognizing patterns inherent in an incoming data set, curve
fitting and data clustering.
In order for the artificial neural net to carry out a useful
task, one must connect the neurons in a particular
configuration, set the weights, and choose the layer functions.
Fig. 1 shows a Multilayer Perceptron NN having I inputs, a
hidden layer of J nodes and an output layer of K nodes.
Weights w
ij
and w
jk
link the corresponding nodes to each other.
In this network, all links are from left to right without jumps.
Layers also have an extra node called bias where its value is
often set to -1. Neurons of each layer have a specific type of
function F that is mainly nonlinear such as tansig (tangent-
sigmoid), logsig (log-sigmoid) and so on.
Output of each neuron is calculated by,
|

\
|
=

i
i ij i j
o w F o
where j and i represent the index of the current and the
preceding layer neurons, respectively [17].
In pattern recognition and cure fitting, network is trained
based on a set of inputs and a set of desired outputs. This is
called supervised learning against unsupervised learning that
no desired output is introduced. Error backpropagation based
on gradient descent method of optimization was a scheme that
gave the NN idea a push. Other training methodologies are
evolutionary methods, simulated annealing and so on.
B. Statistical methods
1) K-NN classifier
Among the various methods of supervised statistical pattern
recognition, the K-Nearest Neighbor (K-NN) rule achieves
consistently high performance, without a priori assumptions
about the distributions from which the training examples are
drawn. It is part of supervised learning that has been used in
many applications in the field of statistical pattern recognition
and many others. KNN categorizes objects into K class. K is a
positive integer that is chosen appropriately [18]. It is usual to
use the Euclidean distance, though other distance measures
such as the Manhattan distance could in principle be used
instead. Algorithm works as below:
1. Determine K=the problem dependent number of
neighbors, beforehand.
2. Calculate the distance between the query-instance and all
the training samples.
3. Sort the distances for all the training samples and
determine the nearest neighbor based on the Kth minimum
distance.
4. Since this is a supervised learning, get all the categories of
the training data for the sorted value which fall under K.
5. Use the majority of nearest neighbors as the prediction
value.
2) SVM classifier
Support vector machine is recognized as one of the standard
tools for machine learning and data mining, which is based on
advances in statistical learning theory. Originally developed to
solve binary classification problems, SVM determines a
number of support vectors from training samples and converts
them into a feature space using various kernel functions,
among which the most commonly used are Gaussian Radial
Basis Function (RBF), polynomial, etc. [19]. Thus, by solving
a quadratic optimization problem, the optimal separating
hyper-plane with a maximal margin between the two classes is
defined.
For the purpose of multi-category classification, various
different binary classification methods are implemented, such
as one-against-all, one-against-one and binary tree. Binary
tree needs K-1 binary SVM for a K class problem while
one against all requires K(K-1)/2 binary SVM [20].
III. DGA DATA PREPARATION AND PRE-PROCESSING
A. Test Data
In this investigation two databases are used. The first is the
well known IEC TC10 and the other is a dataset, called
database 1 formed from a collection of data taken from various
research papers. The IEC TC10 and database 1 contain 151
and 273 samples of various fault cases, respectively. Each
sample comprises H
2
, CH
4
, C
2
H
6
, C
2
H
4
and C2H2 gas
concentration value and its respective fault classes. A set of 5
commonly used gas ratios is chosen as fault indicator and
algorithm input as follows,
(

=
2 2
6 2
2
2 2
6 2
4 2
2
4
4 2
2 2
H C
H C
H
H C
H C
H C
H
CH
H C
H C
x
(1)
The fault types have been assigned to the samples by the
diagnostic tools and industry experts. The transformer
condition is categorized in six classes: Normal Condition (NC-
class 1), Partial Discharges (PD-class 2), Low Energy
Discharge (LED-class 3), High Energy Discharge (HED- class
Fig. 1 A MLP neural network configuration
World Academy of Science, Engineering and Technology 61 2012
1145


4), Thermal Faults<700C (TF1- class 5) and Thermal
Faults>700C (TF2- class6).
B. DGA data Bootstrapping
Since, in practice, the number of samples per fault type is
not equal, it degrades the training and validation ability of
classifiers. Therefore, the collected DGA data are firstly pre-
processed by bootstrap to be equalized in the sample number
for each fault type.
Bootstrap was first introduced by Efron [21] as a computer
intensive re-sampling technique that draws a large number of
re-samples from initial data repeatedly. This is designed to
obtain reliable standard errors, confidence intervals, and other
measures of uncertainty in cases when the initial sample
number is not sufficient for accurate analysis by other
statistical techniques. Because resampling is conducted in a
random order, bootstrap assumes no particular distribution for
the available input data, which gives more applicability with
respect to other classical statistical methods [21].
In this case the number of data for each fault case is raised
to 100 samples, extending the overall database sample to the
amount of 600 samples. This is done for both data bases. In
case that algorithm needs training and test sets, 70% are
allocated for training and the remaining 30% are devoted to
the test procedure. It should be noticed, that the assembled
training and testing datasets are independent of each other and
are employed to confirm the reliability and efficiency of the
proposed ANN classifier.
IV. CLASSIFICATION METHODS AND SIMULATION RESULTS
The proposed method is a hybrid one, integrating bootstrap
with ANN, K-NN and SVM classification methods. In the case
of ANN a Multilayer Perceptron NN and in the case of SVM,
a one against all multiclass SVM classifier is employed.
A. K-NN Classification
First, a K-NN classifier is used for classification of the
transformer condition class. The K closest neighbors are found
from the training dataset by calculating the Euclidean distance
between the examined point and the training samples. The
classification performance of KNN is listed in Table I, where
the number of neighbors, K, is a parameter. As the result
indicates the classification success rate of about 95% for IEC
TC10 can be achieved with K=6. In this situation the success
rate for the database 1 is as low as 83.33% which is not so
good. With increase in K the algorithm outcome is seen to be
deteriorated.
B. SVM Classification
For SVM classification, a One-against-all strategy is
adopted. The results of tests have been depicted in Table II
and III. The results show very low 51.67% success rate for
IEC TC10 and 59.9% for the database 1. Choice of SVM
parameter, , affects marginally the outcome of the algorithm;
however, by no means somehow acceptable results are
obtained.
C. ANN Classification
For ANN classification, a three-layer MLP (multilayer
perception) structure with input, hidden and output layers is
employed as the classifier for the transformer fault
classification. The results have been depicted in Table IV and
V for each of the databases.



TABLE III
CLASSIFICATION ACCURACY OF THE SVM METOD FOR THE DATABASE 1
VERSUS SVM PARAMETER













C=250 C=2500
0.0001 44.79 46.35
0.0005 46.35 45.31
0.001 50.00 45.31
0.005 40.10 48.96
0.01 45.83 43.75
0.1 39.58 39.06
1 47.92 45.83
10 44.27 52.08
100 54.69 49.48
1000 58.33 60.42
2000 59.90 57.29
TABLE I
TEST CLASSIFICATION ACCURACY OF THE K-NN METHOD FOR THE TWO
DATABASES VERSUS K
K Classification
Accuracy %
TC10
Classification
Accuracy %
Database 1
6 95 80.53
10 88.89 78.65
15 90 83.33
20 83.89 79.69
40 81.67 79.17
60 77.78 73.96

TABLE II
CLASSIFICATION ACCURACY OF THE SVM METHOD FOR IEC TC 10
VERSUS SVM PARAMETER
C=250 C=2500
0.0001 47.78 38.33
0.0005 38.89 42.78
0.001 49.44 43.89
0.005 43.33 42.78
0.01 42.78 45.00
0.1 44.44 46.11
1 45.56 42.22
10 40.00 43.33
100 48.89 43.89
1000 50.56 51.11
2000 51.67 46.67
World Academy of Science, Engineering and Technology 61 2012
1146


The number of hidden layer neurons is left as a parameter to
change from 3 to 20. Increase in the number of neurons
exhibits no improvement meaning that a small size network is
adequate for the task.
For each row of the Table IV and V, several experiments
are conducted and the average has been listed. The best
accuracy is 100% that is higher than those obtained using
KNN and SVM.
V. CONCLUSION
In this paper, three well-known classifiers: ANN, K-NN and
SVM are used for DGA and power transformer incipient fault
classification. Often due to the unequal samples for each fault
classes, the training of the algorithms does not proceed well
and test classification does not succeed appropriately. To
manage the difficulty, the bootstrapping technique is employed
to equalize the number of samples for each class. Categorizing
the fault cases based on the bootstrapped preprocessed DGA
data using ANN shows a remarkable improvement, however
the treatment does nothing special for the two other statistical
classifiers. The error in the achieved accuracy reaches less
than 2% that a small size fast converging network can yield.
ACKNOWLEDGMENT
This work has been partially supported by the Research
Department of Shahed University, Tehran, Iran.
REFERENCES
[1] T.K. Saha, Review of modern diagnostic techniques for assessing
insulation condition in aged transformers, IEEE Trans. Dielect. Electr.
Insul., vol. 10, no. 5, pp. 903917, Oct. 2003.
[2] V.G. Arakelian, The longway to the automatic chromatographic
analysis of gases dissolved in insulating oil, IEEE Elect. Insul. Mag.,
vol. 20, no. 6, pp. 825, Nov./Dec. 2004.
[3] IEEE Guide for the Detection and Determination of Generated Gases in
Oil-Immersed Transformers and their Relation to the Serviceability
Equipment, ANSI/IEEE C57.104-1978.
[4] R.R. Roger, IEEE and IEC Codes to Interpret Incipient Faults in
Transformers Using Gas in Oil Analysis, IEEE Transactions on
Electrical Insulation, vol. 13, no. 5, pp. 348-354, 1978.
[5] M. Duval, Dissolved gas analysis: It can save your transformer, IEEE
Electrical Insulation Magazine, vol. 5, no .6, pp. 22-27, 1989.
[6] T. O. Rose, Mineral Insulating Oil in Transformers, IEEE Electrical
Insulation Magazine, vol. 14, No. 3, May/June, pp. 6-28, 1998.
[7] X. Hao and S. Cai-xin, Artificial Immune Network Classification
Algorithm for Fault Diagnosis of Power Transformer, IEEE TRANS.
ON POWER DELIVERY, vol. 22, no. 2, APRIL 2007, pp. 930-935.
[8] Z. Wang, Y. Liu, and P. J. Griffin, A combined ANN and expert
system tool for transformer fault diagnosis, IEEE Trans. Power Del.,
vol. 13, no. 4, pp. 12241229, Oct. 1998.
[9] A. Akbari, A. Setayeshmehr, H. Borsi and E. Gockenbach, Intelligent
Agent-Based System Using Dissolved Gas Analysis to Detect Incipient
Faults in Power Transformers, IEEE Electrical Insulation Magazine,
Vol. 26, No. 6, pp.27-40, November/December, 2010.
[10] K.F. Thang, R.K. Aggarwal, A. J. McGrail, and D. G. Esp, Analysis of
power transformer dissolved gas data using the self-organizing map,
IEEE Trans. Power Del., vol. 18, no. 4, pp. 12411248, Oct. 2003.
[11] N.K. Patel and R.K. Khubchandani, "ANN Based Power Transformer
Fault Diagnosis," IE (I) Journal. EL, vol 85, pp. 60-63, June 2004.
[12] D.V.S.S. Siva Sarma and G.N.S. Kalyani, "Application of AI
Techniques for Nondestructive Evaluation pf Power Transformers Using
DGA," IJIESP, Vol. 2, no. 1, 2007.
[13] M.H. Wang, Extension Neural Network for power transformer
incipient fault diagnosis, IEE Proceedeings, Generation,
Transmission, Distribution, vol.150, No.6, Nov 2003.
[14] A. Shintemirov, W. Tang, and Q. H. Wu, Power Transformer Fault
Classification Based on Dissolved Gas Analysis by Implementing
Bootstrap and Genetic Programming, IEEE Trans. On System, Man,
and Cybernetics-Part C: APPLICATIONS AND REVIEWS, vol. 39, no.
1, pp. 69-79, Jan 2009.
[15] W. Chen, C. Pan, Y. Yun, and Y. Liu, Wavelet Networks in Power
Transformers Diagnosis Using Dissolved Gas Analysis, IEEE
Transactions on Power Delivery, Vol. 24, no, pp. 187-194 , Jan. 2009.
[16] H.T. Yang, C.C. Liao, and J.H. Chou, Fuzzy learning vector
quantization networks for power transformer condition assessment,
IEEE Trans. Dielect. Electr. Insul., vol. 8, no. 1, pp. 143149, Feb.
2001.
[17] S. Haykin, Neural Networks: A Comprehensive Foundation (2nd
Edition). Prentice Hall, 1998.
[18] M. Dong, D.K. Xu, M.H. Li, et al. Fault Diagnosis Model for Power
Transformer Based on Statistical Learning Theory and Dissolved Gas
Analysis, in proc. IEEE International Symposium on Electrical
Insulation, USA, 2004, pp.85-88.
[19] C.W. Hsu, C.J. Lin, A Comparison of Methods for Multiclass Support
Vector Machines, IEEE Transactions on Neural Networks, vol 13, no
2, 2002 .
[20] S.F. Yuan and F.L. Chu, "Support vector machines based fault diagnosis
for turbo-pump rotor," Mechanical Systems and Signal Processing, vol
20, no 4, pp. 939952, 2006.
[21] R. Wehrens, H. Putter, and L.M.C. Buydens, The bootstrap: a
tutorial, Chemometrics and Intelligent Laboratory Systems, vol 54, pp.
3552, 2000.

TABLE IV
FAULT CLASSIFICATION ACCURACY (%) OF THE ANN METHOD VERSUS
THE NNUMBER OF HIDDEN LAYER NEURONS APPLIED TO THE IEC TC10
neuron
s
validatio
n
trainin
g
test
3 99.50 100 98.33
4 100 100 100
6 99 100 96.67
8 99.5 100 98.33
10 99.83 100 99.44
12 99 100 96.66
15 99.33 99.52 98.88
20 100 100 100

TABLE V
TEST CLASSIFICATION ACCURACY (%) OF THE ANN VERSUS THE
NUMBER OF HODDEN LAYER NEURONS APPLIED TO THE DATABASE 1
neuron
s
validatio
n
trainin
g
test
3 99.42 100 98.09
4 99.28 100 97.61
6 98.71 99.79 96.19
8 99.14 100 97.33
10 99.14 99.59 98.09
12 99.42 100 98.09
15 99 99.59 97.61
20 98.85 100 96.19

World Academy of Science, Engineering and Technology 61 2012
1147