This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
I hereby declare that I am the sole author of this thesis. This is a true copy of the thesis, including any
required final revisions, as accepted by my examiners.
I understand that my thesis may be made electronically available to the public.
iii
Abstract
In recent years, the trade-off between quality and cost of power system components has become a
matter of interest for many utilities. The widespread use of costly electricity networks either in
residential or industrial areas has encouraged service providers to find a proper strategy that will
minimize the overall life-cycle cost while keeping components in good working condition. The power
transformer, which represents approximately 60% of the overall cost of the network, is ranked as one
of the most important and expensive components. However, the transformer's sudden failure puts the
system in a serious or critical condition which in most cases causes catastrophic loss to both utilities
and customers. Significant attention has been given to monitoring and diagnostic techniques that
observe any abnormal behaviour, assess the transformer's condition, and therefore minimize the
probability of unplanned outage. Yet, applying many various monitoring tests is not always
applicable due to the following factors: some tests require the unit to be taken out from service for
testing, insufficient availability of man power, and significant cost of applying all the tests. Thus,
there is a vital demand for an intelligent method of minimizing the number of monitoring tests
without losing much information about the transformer's actual condition.
In this research, data mining techniques have been employed to evaluate the transformer's state
through intelligent selection criteria that determines the optimal number of monitoring tests in cost-
effectiveness. Feature selection technique based on ranker search method has been used to rank the
monitoring tests (features) in a priority sequence from their individual evaluation, and to select the
most inductive tests that provide the most information about the unit's condition. When the measured
data from monitoring tests is collected and prepared, a diagnostic technique is applied to assess the
condition of the transformer. In this regard, Support Vector Machine (SVM) has been utilized to
perform this task due to its robust classification accuracy. SVM is first applied to the full number of
tests, and then the number of monitoring tests is reduced by one after each classification process using
the feature selection algorithm. The selected number of monitoring tests has shown the best possible
accuracy the classifier can reach over the whole number of tests. Radial Basis Function (RBF)
classifier has been used in the classification process for results comparison purposes. This proposed
work contributes towards finding an intelligent method of evaluating the transformer state as well as
minimizing the number of tests without losing much information about the unit's actual condition.
Therefore, this method facilitates deciding a wise course of action regarding the transformer: either
maintain, repair, or replace.
iv
Acknowledgements
I would like to take this opportunity to express my deep gratitude to my supervisor Dr.Ramadan
Elshatshat. During my master's, he guided me wisely and precisely to put me on the right path. I
greatly got benefited from his experience and positive comments.
I also would like to thank my friends in Waterloo specifically the Libyan community for offering help
and support all the times. And special thanks to the closest friends Elfaitoy, Nabil and Haithm.
I would like to thank all UW staff and members.
Finally, I would like to thank my family, relatives and all my friends in Libya for their support and
encourage.
v
Dedication
This thesis is dedicated to my lovely parents who are raising their hands every day for asking Allah
to help and protect me.
To: my lovely wife for her patience and support all the times.
To: my brothers and sisters for encourage and advice.
vi
Table of Contents AUTHOR'S DECLARATION ............................................................................................................... ii
Abstract ................................................................................................................................................. iii
Acknowledgements ............................................................................................................................... iv
Dedication .............................................................................................................................................. v
Table of Contents .................................................................................................................................. vi
List of Figures ....................................................................................................................................... ix
List of Tables ......................................................................................................................................... x
1.1 General ......................................................................................................................................... 1
cooling, furan, excitation current, core-to-ground, frequency response analysis, main tank condition,
43
connectors, and gaskets. This ranking provides a clear indication of the validity of the feature
selection technique since the most common and important test has been ranked first [2]. The physical
meaning of DGA test being ranked first is of high importance, because this routine test provides
information about the actual condition of a transformer. The concentration of the dissolved gases
when performing DGA test reflects the severity and type of the fault occurring inside the transformer.
While in-service, transformers are exposed to thermal and electrical stress which occasionally results
in variation in DGA test conditions, as opposed to other stable tests such as core-to-ground or turns
ratio. Therefore, the algorithm selects the most inductive tests that vary between transformers in their
levels and affect the output (condition) as in the case of DGA, infra-red, load history, etc.
5.5 Classification Process
First of all, the whole data set of the seventy units and their labels are prepared and adjusted for use in
the classifier. The entries are the 19 features (tests) and the labels (conditions) of each unit for the
whole data set. The classification process begins with all the features and then reduces each feature
step-by-step starting with the least inductive using feature selection techniques. At each step, the
classifier is run to see the correctly classified instances and identify at which number of attributes the
least percentage of error is achieved.
The flowchart in Figure 5.3 summarizes the feature selection and classification process and shows the
classification of tests using SVM and RBF. First, data is collected by employing monitoring tests and
feedback from experts, followed by the preparation and preprocessing of data as explained
previously. When the data is ready for classification, the third step begins selecting and ranking the
best tests (features) based on their individual evaluation. Next, the training and testing process using
k-fold cross-validation is performed. After training the classifier and adjusting its parameters, the
classification operation is conducted and the results are saved for comparison with new results. The
second cycle starts by deleting the least important feature and then retrains the classifier with the new
number of features. Each time, a new subset of data is generated using k-fold cross validation. This
process continues until the number of tests reaches one which is indicated by k=1, meaning that the
deletion process has been stopped. Thus, the accuracy of selected tests is printed out as the best
results of the classifier.
44
Collecting Data
K=1
Select the best features and rank them using ranker search
method. K(tests)=19
Train and test the classifier to optimize
its parameters
Perform classification SVM,RBF
Output the best accuracy and selected
features(tests)
Data Preparation
and Preprocessing
START
END
K=K-1
Print the output of the classifier(accuracy)
No
Yes
Figure 5.3 Classification process using SVM and RBF
45
5.6 Simulation Results
The SVM algorithm has been developed and tested in Matlab environment. The following figure
shows the performance of SVM classifier over each number of features (tests), where each point over
the number of tests represents the classifier accuracy at each cycle or loop as explained in the
flowchart. Note that these results represent the performance of SVM for a combination of test data
generated using k-fold. The trend is evident of the relationship between the accuracy of the classifier
and the number of tests.
Figure 5.4 SVM Performance at each number of attributes
As shown in Figure 5.2, the x-axis represents the number of tests from one to 19 and the y-axis
represents the correctly classified instances at each set of tests. It was observed that for the first three
features the classifier performance achieved accuracy of 84.28%, which can be considered a good
0 2 4 6 8 10 12 14 16 18 2078
80
82
84
86
88
90
92
94
Number of Tests
Cor
rect
ly C
lass
ified
Inst
ance
s %
46
start for the classifier. However, as the number of attributes increased the classification process
improved; hence, the correctly classified instances increased as illustrated at four and five attributes.
At six features it was observed that the classifier provided the best performance (accuracy 92.85%),
which indicates the best number and type of monitoring tests that can be utilized to indicate
transformer condition. However, as the number of tests increased further, the classification accuracy
became less, as in seven, nine, and ten features having the same accuracy. Finally, the classification
accuracy decreased until it reached 16 features with accuracy of 80% and continued with the same
performance until completion.
In data mining, when the dimensionality of the data increases, many techniques of data analysis and
classification problems become more difficult. Moreover, a high number of attributes results in lower
classification accuracy. On the other hand, when the number of the training data is very small, the
created model in the case of a supervised learning technique will be less reliable [44]. Therefore,
feature selection techniques have shown their capability in improving the classification performance
when many features are used. The basic idea of feature subset selection is to eliminate the redundant
or irrelevant attributes from the data as they can lead to a reduction of the classifier performance [45].
In order to have better visualization and to observe the trend of the data points, a curve fitting
technique has been used to find the best curve that passes through these points. Curve fitting
procedures have been implemented using Matlab software as a powerful technique for dealing with
such a problem. This technique helps in identifying the margin and the required tests that should be
used in the assessment process. It was observed that the classification accuracy improved as the
number of tests decreased until feature nine, at which point the classification accuracy decreased.
This trend can be interpreted as the non-inductive or irrelevant features being considered noisy
information that reduces classification accuracy. Hence, utilizing the top ranked features obtained by
the feature selection algorithm improves the classifier performance by ignoring the attributes of less
information. More information regarding the classification accuracy is given by the classifier in Weka
software at each test. In each run, full information is provided about the classifier including
correctly/incorrectly classified instances, average error, accuracy, and most importantly the confusion
matrix as explained in detail in the following tables and paragraphs. Table 5.3 displays one of the
performance indices that provide the best results (six features). This case, which can be considered
optimal in terms of selecting the best number of tests, gives utilities the type and number of
monitoring tests to perform in order to assess transformer condition. Some terms shown in the tables
47
are explained clearly in [46]. These measurements are useful for comparing classifiers. For example,
Kappa statistic measures the agreement of prediction with the true class -1.0 signifying complete
agreement.
Table 5.3 SVM Performance at Six Features
Classifier Performance Results
Correctly classified instances 65 = 92.8571 %
Incorrectly classified instances 5 = 7.1429 %
Kappa statistic 0.8932
Mean absolute error 0.2446
Root mean squared error 0.3234
Relative absolute error 89.7888 %
Root relative squared error 88.1441 %
Total number of instances 70
Other terms shown in Table 5.4 such as True Position (TP), False Position (FP), Precision, and F-
measure are explained further in [46]. These terms are critical in understanding the classification
accuracy where the classified and misclassified examples are stated as a percentage for each
individual class. The following are definitions of terms used in Table 5.4 [46]:
True Positive (TP) rate: the proportion of examples which were classified as class x among
all examples which truly have class x
False Positive (FP) rate: the proportion of examples which were classified as class x but
belong to a different class
Precision: the proportion of the examples which truly have class x among all those which
were classified as class x
F-Measure: a combined measure for precision and recall defined as
2 ∗ 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∗ 𝑅𝑒𝑐𝑎𝑙𝑙/(𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙)
48
Table 5.4 SVM Detailed Accuracy by Class
TP Rate FP Rate Precision Recall F-measure Roc-Area Class
1 0 1 1 1 1 A
0.939 0.054 0.939 0.939 0.939 0.949 B
0.867 0.036 0.867 0.867 0.867 0.96 C
1 0.015 0.8 1 0.889 0.992 D
0 0 0 0 0 0.435 E
0.929 0.034 0.917 0.929 0.922 0.959 Weight Avg
In the field of artificial intelligence, the success of a classifier is measured by the confusion matrix,
which in some cases is called the contingency table. The confusion matrix is a specific table design
which permits the visualization of classification accuracy. In this problem, the correctly classified and
misclassified instances are shown in the confusion matrix in Table 5.5.
Table 5.5 SVM Confusion Matrix
a b c d e Classified
as
17 0 0 0 0 a=A
0 31 2 0 0 b=B
0 2 13 0 0 c=C
0 0 0 4 0 d=D
0 0 0 1 0 e=E
49
From the confusion matrix it is evident that the 17 class A transformers have been classified correctly
and there are no misclassified instances. For the 33 class B transformers, there are 31 units correctly
classified as B and two units classified as C, which indicates two misclassified instances.
Transformers with class C that are represented by 15 units have two class B units while the other 13
units are classified correctly as C class. Finally, transformers with lower conditions such as C are
classified correctly for four units and the one failed E class transformer is classified as D. Overall, it
was observed that five of 70 transformers were incorrectly classified, which represents a 7.14% error
rate. Therefore, the small number of misclassified instances can be considered a good indicator for the
validity of the classifier performance.
5.7 Comparison With RBF
Radial Basis Function (RBF) classifier has been applied to the same data set following the same exact
strategy as SVM in terms of data size and feature selection algorithm. RBF is used as a benchmark to
validate the performance of the proposed algorithm.
It is clear from Figure 5.5 that the best performance of the classifier is given at ten, 14, and 15 tests
with classification accuracy of approximately 90%. In addition, the procedures of applying SVM are
followed. The performance of the classifier is improved by starting with 19 tests and then reducing
the number of the tests step-by-step. The accuracy increases until it reaches its maximum at 14 and 15
features, at which point the classification accuracy starts decreasing. A sudden drop occurred at six
features, in contrast with SVM where at six features the best performance was achieved. After this,
the performance improves with small variation until it reaches the first attribute.
Comparing these results with SVM classifier gives a clear indication of each classifier's performance
in terms of the minimum number of features that can be used to achieve better accuracy. Comparing
these results with SVM classifier gives a clear indication of each classifier's performance in terms of
the minimum number of features that can be used to achieve better accuracy.
50
Figure 5.5 RBF Performance at each number of attributes
Table 5.6 RBF Performance at Ten Features
Correctly classified instances 63 = 90%
Incorrectly classified instances 7 = 10%
Kappa statistic 0.8498
Mean absolute error 0.0438
Root mean squared error 0.1947
Relative absolute error 16.0664 %
Root relative squared error 53.0805 %
Total number of instances 70
0 2 4 6 8 10 12 14 16 18 2078
80
82
84
86
88
90
92
94
Number of Tests
Cor
rect
ly C
lass
ified
Inst
ance
s %
51
RBF performance parameters including TP Rate, FP Rate, Precision, Recall, and F-Measure are
shown in detailed accuracy by class in Table 5.7. Comparing this table with SMV is very beneficial in
distinguishing between the performances of the classifiers.
Table 5.7 Detailed Accuracy by Class RBF
TP
Rate FP Rate Precision Recall
F -
measure
Roc-
Area Class
1 0 1 1 1 1 A
0.97 0.054 0.941 0.97 0.955 0.947 B
0.8 0.036 0.857 0.8 0.828 0.863 C
0.5 0.03 0.5 0.5 0.5 0.792 D
0 0.014 0 0 0 0.79 E
0.9 0.035 0.899 0.9 0.899 0.927 Weight
Avg
As previously explained, the confusion matrix is most important. The classified and misclassified
instances are shown in Table 5.8 in which transformers with condition A are classified correctly, one
instance of B is misclassified, three instances of C are misclassified, two instances of D are
misclassified, and finally one instance of E is misclassified. When comparing the confusion matrix
from SVM and RBF, it was observed that both matrices correctly classified transformers with
condition (class) A. In B classes, SVM misclassified two B class transformers as C class, whereas
RBF misclassified one B class unit as C class. For C class, SVM demonstrated better performance by
misclassifying two units as B class, while RBF misclassified three units. Moreover, SVM correctly
classified all D class transformers and misclassified the only transformer with E class to D class,
whereas RBF misclassified two instances of D class and the one instance of E class.
52
Table 5.8 RBF Confusion Matrix
a b c d e Classified
as
17 0 0 0 0 a=A
0 32 1 0 0 b=B
0 2 12 1 0 c=C
0 0 1 2 1 d=D
0 0 0 1 0 e=E
5.8 Results and Discussion
This chapter investigates the proper use of data mining techniques in the condition assessment of
power transformers. Diagnostic techniques have been used to perform this task through two
classifiers: Support Vector Machine (SVM) and Radial Basis Function (RBF). Firstly, the 19
monitoring tests have been ranked based on their individual evaluation using the ranker search from
the attribute selection filter in Weka Software. Secondly, the features (tests) were reduced by one
each time, and then the classifiers were executed on the new set of data to identify the classification
accuracy. This strategy will allow utilities to minimize the number of tests to a level that properly
indicates the transformer condition without losing much information. The performance of each
classifier has been presented and compared.
In many cases, conducting all the monitoring tests on power transformers is not a cost-effective action
for effective asset management. Therefore, a good method of meeting utilities satisfaction is to rank
these tests by priority sequence in terms of their effectiveness in transformer condition, as well as
select a limited number of these tests to improve classifier performance. SVM achieved better
performance in predicting transformer condition with the first six tests when its accuracy reached
92.85%. This implies that new transformers can be classified reliably and the risk of misclassifying is
very small. While RBF classifier has also achieved good performance, it is not comparable with SVM
in terms of accuracy and number of tests. RBF achieved the best performance with the first ranked
53
ten, 14, and 15 tests, in which classifier performance reached 90%. A comparison of the two
classifiers in terms of their performance on the measured data is shown in Table 5.9.
Table 5.9 Performance Comparison (SVM and RBF)
Comparison Terms SVM RBF
Number of instances 70 70
Accuracy % 92.85% 90%
Number of tests at best performance 6 10, 14, 15
Correctly classified instances 65 63
Overall, applying the methodology of this research facilitates quick assessment of transformer
condition and reliable decisions without performing all tests. There are many beneficial applications
including the minimization of monitoring costs, time, and work that is required to perform these tests.
This model will allow utilities to assess their transformers properly with a limited number of specific
monitoring tests, and create reliable plans regarding proper transformer action.
54
Chapter 6 Conclusion
6.1 Thesis Conclusion
Power transformers are a vital and essential component in any electric network. Indeed, transformer
cost varies between thousands and millions of dollars depending on the design and size of the unit.
Therefore, any failure leads to unplanned outage or early breakdown before the designed lifespan
expires, which costs utilities catastrophic loss. Hence, asset management practices have been widely
utilized in the last decades to minimize overall life cycle cost and maintain cost-effectiveness. Such a
practice ensures operating the system in an efficient and reliable way to meet the providers’ and
customers’ satisfaction. Moreover, an effective asset management procedure needs many practices
and work starts from design until disposal. In this research, condition assessment of power
transformers, which is a part of asset management, has been investigated and a case study from in-
service transformers has been conducted to develop a model that helps in identifying transformer
condition from measured data.
A literature review is introduced which describes transformer importance and its contribution to the
electricity sector. Various maintenance strategies are discussed as a part of asset management. In
addition, a survey is conducted of monitoring and diagnostic techniques and of what has been done
thus far in the area of transformer condition assessment. Common monitoring methods used to assess
transformer state as well as the limits and variations from the unit nameplate interpreted based on
national standards such as ASTM/IEEE and IEC have been investigated. Diagnostic techniques such
as SVM and RBF as well as feature selection algorithms are explained for their employment in the
assessment task.
Applying many monitoring techniques may be economically unreliable, require a lot of skilled work,
and have many challenges and drawbacks from application. Some of these challenges can be
summarized as electricity interruption in the case of off-line tests, time consumption, and high cost.
Hence, the proposed model in this research employs data mining techniques to find a proper method
of minimizing these challenges while proving reliability as a valid model. SVM classifier is used to
predict transformer condition or health index among five classes A, B, C, D, and E from optimal to
least optimal, respectively. The 19 tests (features) from the measured data have been prepared and
processed to be used in the classification process. When the data are prepared and scaled, the feature
55
selection based on ranker search method has been used to rank and select the most inductive tests
(features) in a priority sequence. The algorithm ranked the tests based on their individual evaluation
and their strong impact on the transformer’s condition. For instance, DGA, infra-red, and load history
tests were selected first because they are the most common and important tests used by utilities to
evaluate the state of the transformer. The full number of tests has been used to train and test the
classifier (SVM) using k-fold cross-validation in which each feature (test) is used in the training and
testing at least once, and then the tests have been reduced by eliminating the least inductive feature
(test) after each classification process. Moreover, after each eliminating process, the classifier is
retrained for the new number of tests. The number of tests for which the classifier gives best
performance is selected as the recommended number of tests that reflect the transformer's actual
condition. SVM reached its maximum performance (92.85%) at six tests, which are DGA, infra-red,
load history, bushing condition, power factor, and DGA of LTC. In fact, these tests are the most
common tests conducted by utilities and manufacturers because of the useful information they
provide about the transformer's state. For instance, DGA provides information about faults inside a
transformer and their causes. The thermography test identifies hot spot areas and overloading
conditions. Loading history helps in determining how many times the unit has been operated under
overloading conditions and their impact on the insulation system. Power factor is a routine test used
to evaluate capacitive insulation condition between windings and compartments as well as assess
bushing condition. Finally, DGA of LTC is used as the main test to assess tap changer condition. The
criteria followed for selecting the specific number of tests prove the applicability of the model in
predicting the transformer's actual condition with the least number of monitoring techniques. This
process results in a cost-effective solution to assess the power transformers. The RBF model has been
built in the same way and compared with SVM in terms of accuracy and number of tests. Although its
classification accuracy was good (90%), this percentage was achieved at ten, 14, and 15 tests.
However, RBF model is not comparable with SVM because of the high number of monitoring tests it
needs to achieve this performance.
Considerable attention has been given to transformer condition assessment because of its direct
impact on planning and budget as well as reliability and risk assessment. Furthermore, operating
transformers in acceptable condition even after their designed lifespan has been considered to meet
service providers' and customers’ satisfaction.
56
6.2 Future Work
A similar model using artificial intelligence tools could be designed to predict the transformer's future
condition. Historical data from one unit represents different types of tests that could be utilized to
train the model in order to use it on other units of the same size. Assessing the future condition of the
transformer would provide utilities with more time to develop an effective plan and minimize the
probability of unplanned outages through awareness of the deterioration rate or expected faults of the
transformer.
57
References
[1] N. Dominelli, "Equipment health rating of power transformers," 2004, pp. 163-168. [2] A. Jahromi, R. Piercy, S. Cress, and W. Fan, "An approach to power transformer asset
management using health index," Electrical Insulation Magazine, IEEE, vol. 25, pp. 20-34, 2009.
[3] B. GORGAN, P. NOTINGHER, L. BADICU, and G. TANASESCU, "CALCULATION OF POWER TRANSFORMERS HEALTH INDEXES," health, vol. 602, p. 929481.
[4] Z. Xiang and E. Gockenbach, "Asset-Management of Transformers Based on Condition Monitoring and Standard Diagnosis [Feature Article]," Electrical Insulation Magazine, IEEE, vol. 24, pp. 26-40, 2008.
[5] "IEEE Recommended Practice for Installation, Application, Operation, and Maintenance of Dry-Type General Purpose Distribution and Power Transformers," ANSI/IEEE Std C57.94-1982, p. 3, 1982.
[6] J. Spacek, "Maintenance strategies of power equipments with a brief view to condition monitoring of power transformers," 2008, pp. 16-19.
[8] J. McCalley, Y. Jiang, V. Honavar, J. Pathak, M. Kezunovic, S. Natti, C. Singh, and P. Jirutitijaroen, "Automated integration of condition monitoring with an optimized maintenance scheduler for circuit breakers and power transformers," Power Systems Engineering Research Center, 2006.
[9] X. Zhang and E. Gockenbach, "Asset-Management of Transformers Based on Condition Monitoring and Standard Diagnosis [Feature Article]," Electrical Insulation Magazine, IEEE, vol. 24, pp. 26-40, 2008.
[10] A. Emsley and G. Stevens, "Review of chemical indicators of degradation of cellulosic electrical paper insulation in oil-filled transformers," 1994, pp. 324-334.
[11] J. Singh, Y. R. Sood, and R. K. Jarial, "Condition monitoring of power transformers-bibliography survey," Electrical Insulation Magazine, IEEE, vol. 24, pp. 11-25, 2008.
[12] T. K. Saha, "Review of modern diagnostic techniques for assessing insulation condition in aged transformers," Dielectrics and Electrical Insulation, IEEE Transactions on, vol. 10, pp. 903-917, 2003.
[13] P. J. Baird, H. Herman, G. C. Stevens, and P. N. Jarman, "Spectroscopic measurement and analysis of water and oil in transformer insulating paper," Dielectrics and Electrical Insulation, IEEE Transactions on, vol. 13, pp. 293-308, 2006.
[14] B. H. Ward, "A survey of new techniques in insulation monitoring of power transformers," IEEE Electrical Insulation Magazine, vol. 17, pp. 16-23, 2001.
[15] D. Allan and C. Jones, "Thermal–oxidative stability and oil-paper partition coefficients of selected model furan compounds at practical temperatures–9 th International Symposium on high voltage engineering," 1995.
[16] A. M. Emsley and G. C. Stevens, "A reassessment of the low temperature thermal degradation of cellulose," in Dielectric Materials, Measurements and Applications, 1992., Sixth International Conference on, 1992, pp. 229-232.
[17] Y. Z. Yang Ghazali, M. A. Talib, and H. Ahmad Rosli, "TNB experience in condition assessment and life management of distribution power transformers," in Electricity Distribution - Part 1, 2009. CIRED 2009. 20th International Conference and Exhibition on, 2009, pp. 1-4.
[18] F. R. Barbosa, O. M. Almeida, A. P. S. Braga, M. A. B. Amora, and S. J. M. Cartaxo, "Application of an artificial neural network in the use of physicochemical properties as a low cost proxy of power transformers DGA data," Dielectrics and Electrical Insulation, IEEE Transactions on, vol. 19, pp. 239-246, 2012.
[19] B. Németh, S. Laboncz, I. Kiss, and G. Csépes, "Transformer condition analyzing expert system using fuzzy neural system," 2010, pp. 1-5.
[20] W. Feng-Jiao, Z. Guan-Jun, W. Shi-Qiang, X. Hao, W. Da, and L. Min, "Research on condition assessment method of intelligent power transformer," in Electrical Engineering and Informatics (ICEEI), 2011 International Conference on, 2011, pp. 1-4.
[21] R. Liao, H. Zheng, S. Grzybowski, L. Yang, Y. Zhang, and Y. Liao, "An integrated decision-making model for condition assessment of power transformers using fuzzy approach and evidential reasoning," Power Delivery, IEEE Transactions on, vol. 26, pp. 1111-1118, 2011.
[22] M. Hui, T. K. Saha, and C. Ekanayake, "Statistical learning techniques and their applications for condition assessment of power transformer," Dielectrics and Electrical Insulation, IEEE Transactions on, vol. 19, pp. 481-489, 2012.
[23] L. Jinling and W. Mijia, "Condition assessment for power transformer based on improved particle swarm optimization and Support Vector Machine," in Critical Infrastructure (CRIS), 2010 5th International Conference on, 2010, pp. 1-6.
[24] P. Dash and S. Samantaray, "An accurate fault classification algorithm using a minimal radial basis function neural network," Engineering intelligent systems, vol. 4, pp. 205-210, 2004.
[25] W. Weihua, "License Plate Recognition Algorithm Based on Radial Basis Function Neural Networks," in Intelligent Ubiquitous Computing and Education, 2009 International Symposium on, 2009, pp. 38-41.
[26] H. G. Suykens J.A.K., Advances in learning theory vol. 190 NATO-ASI: IOS Press, 2003, 2003.
[27] G. F. Lin and L. H. Chen, "A non-linear rainfall-runoff model using radial basis function network," Journal of Hydrology, vol. 289, pp. 1-8, 2004.
[28] I. Guyon and A. Elisseeff, "An introduction to variable and feature selection," The Journal of Machine Learning Research, vol. 3, pp. 1157-1182, 2003.
[29] T. M. Mitchell, "Machine learning," ed: McGraw-Hill New York:, 1997. [30] "IEEE Guide for the Interpretation of Gases Generated in Oil-Immersed Transformers," IEEE
Std C57.104-1991, p. 0_1, 1992. [31] M. Duval, "Dissolved gas analysis: It can save your transformer," Electrical Insulation
Magazine, IEEE, vol. 5, pp. 22-27, 1989. [32] "IEEE Guide for Acceptance and Maintenance of Insulating Oil in Equipment," IEEE Std
C57.106-2006 (Revision of IEEE Std C57.106-2002), pp. 1-36, 2006. [33] M. Wang, A. Vandermaar, and K. Srivastava, "Review of condition assessment of power
transformers in service," Electrical Insulation Magazine, IEEE, vol. 18, pp. 12-25, 2002. [34] G. Duke, "Predictive maintenance a case study in infrared thermography," Electrical
Maintenance, pp. 11-12, 1998. [35] I. Arifianto and B. Cahyono, "Power transformer cooling system optimization," in Properties
and Applications of Dielectric Materials, 2009. ICPADM 2009. IEEE 9th International Conference on the, 2009, pp. 57-59.
59
[36] R. Yadav, S. Kumar, A. Venkatasami, A. M. Lobo, and A. M. Wagle, "Condition based maintenance of power transformer: A case study," in Condition Monitoring and Diagnosis, 2008. CMD 2008. International Conference on, 2008, pp. 502-504.
[37] M. A. Franchek and D. J. Woodcock, "Life-Cycle considerations of loading transformers above nameplate rating," in 65th Annual International Conference of Doble Clients, 1998.
[38] A. Setayeshmehr, A. Akbari, H. Borsi, and E. Gockenbach, "On-line monitoring and diagnoses of power transformer bushings," Dielectrics and Electrical Insulation, IEEE Transactions on, vol. 13, pp. 608-615, 2006.
[39] K. Jan, "Operating Damages of Bushings in Power Transformers," TRANSACTIONS ON ELECTRICAL ENGINEERING, p. 89, 2012.
[40] A. Shintemirov, W. Tang, and Q. Wu, "Transformer winding condition assessment using frequency response analysis and evidential reasoning," Electric Power Applications, IET, vol. 4, pp. 198-212, 2010.
[41] M. De Nigris, R. Passaglia, R. Berti, L. Bergonzi, and R. Maggi, "Application of modern techniques for the condition assessment of power transformers," in Cigré Session, 2004, pp. A2-207.
[42] G. J. McLachlan, K. A. Do, and C. Ambroise, Analyzing microarray gene expression data vol. 422: Wiley-Interscience, 2004.
[43] http://www.cs.waikato.ac.nz/ml/weka/index.html. [44] A. G. K. Janecek, W. N. Gansterer, M. Demel, and G. F. Ecker, "On the relationship between
feature selection and classification accuracy," in JMLR: Workshop and Conference Proceedings, 2008, pp. 90-105.
[45] A. L. Blum and P. Langley, "Selection of relevant features and examples in machine learning," Artificial intelligence, vol. 97, pp. 245-271, 1997.
[46] R. R. Bouckaert, E. Frank, M. Hall, R. Kirkby, P. Reutemann, A. Seewald, and D. Scuse, "WEKA Manual for Version 3-7-4," 2011.