See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/352806309

Comparative Dissolved Gas Analysis with Machine Learning and Traditional

Methods

Conference Paper · June 2021

DOI: 10.1109/HORA52670.2021.9461371

CITATIONS READS

8 960

3 authors:

Merve demirci Haluk Gözde

Kafkas University Gazi University
15 PUBLICATIONS 57 CITATIONS 36 PUBLICATIONS 1,224 CITATIONS

SEE PROFILE SEE PROFILE

M. Cengiz Taplamacioglu
Gazi University
81 PUBLICATIONS 2,114 CITATIONS

SEE PROFILE

All content following this page was uploaded by Merve demirci on 21 September 2021.

The user has requested enhancement of the downloaded file.

 Comparative Dissolved Gas Analysis with Machine
Learning and Traditional Methods

Merve DEMIRCI Haluk GOZDE M. Cengiz TAPLAMACIOGLU

Electrical and Electronic Engineering Electrical and Comminication Electrical and Electronic Engineering
Gazi University Engineering Gazi University
Ankara, Turkey National Defense University Ankara, Turkey
[email protected] Ankara, Turkey [email protected]
[email protected]

Abstract— Power transformers are one of the vital TC 10 dataset and the data in the literature, and performed
equipment for power systems. Therefore, in case of outage of the transformer fault diagnosis using the Extreme learning
service, the effects on the system are fatal. The fault diagnosis is machine method, which uses gas ratios and percentages as
of great importance. In this study, interpretation methods of input data, depending on the gas concentrations in this data
dissolved gas analysis used in diagnosis of power transformers set. They showed that the method used has a higher diagnostic
are examined. In the MATLAB GUI, a user interface has been accuracy than Fuzzy logic and ANN [11]. In 2014, Shari et al.
created for traditional fault diagnosis. Fault diagnosis is made used the K-nearest Neighbor method as a classifier and the K-
and the shortcomings of traditional methods have been shown. nearest Neighbor method as a classifier, using the IEC TC 10
Support Vector Machine, K-Nearest Neighbors and Decision
dataset and the data set they obtained from the Malaysian
Tree algorithms are used for diagnosis of machine learning
methods. Between these methods, the Python programming
Utility Company [12]. In 2017, Ma et al. performed fault
language is used for fault diagnosis and fault classification is classification with SVM using three gas ratios from 118 DGA
made to the IEC TC 10 database. Confusion matrices and data from the IEC TC 10 dataset and used Particle Swarm
classification performance measurements of machine learning Optimization for the optimization of SVM parameters. The
methods are obtained. The classification accuracy of the results obtained are compared with SVM and Genetic
methods has been investigated and compared to each other. Algorithm -SVM [13]. In 2019, Zhang et al. used gas ratios as
input data for the SVM, as well as gas concentrations obtained
Keywords—Dissolved gas analysis, traditional methods, SVM, from an IEC TC 10 dataset and a dataset of 117 samples. They
KNN, decision tree, power transformer, Python used Krill Herd Optimization algorithm for optimization of
SVM parameters and compared the results with different
I. INTRODUCTION optimization algorithms [14].
A power transformer is one of the most important and most In this study, DGA method used for diagnosis of power
expensive parts of power systems. Diagnosis of transformers transformers is mentioned, traditional methods and smart
is of great importance in order to ensure the continuity of methods are explained. The application and results of
energy and to avoid costly problems [1,2]. traditional methods in MATLAB GUI are given in detail.
Dissolved gas analysis (DGA) method has been widely Applications and results of smart methods used to eliminate
used in transformer fault diagnosis in recent years. Using the the shortcomings of traditional methods in Python are also
gas concentrations obtained by these methods, the fault given in detail simultaneously.
diagnosis is made with traditional and intelligent methods.
While traditional methods provide fault classification using II. DISSOLVED GAS ANALYSIS
rule bases created using expert knowledge and experience, Gas molecules of different densities are formed during
intelligent methods are used for performing the fault failure in power transformers. The formation temperature and
classification with computer-assisted learning methods [3]. formation energies of these gas molecules are different from
each other. Since the type and amount of gas molecules
Rogers Ratios Method [3], IEC Ratios Method [4],
formed during the fault differ, these gases are can be used for
Doernenburg Ratios Method [3] and Duval triangle method
fault detection. Dissolved gas analysis method is one of the
[3] are widely used in the diagnosis of power transformers.
successful methods used to detect dissolved gases in
These methods make fault classification using different ratios
transformer insulation liquid. With the chromatograph
of gas concentrations and gas percentages [3, 4]. Intelligent
method, the amount of gases in the insulating liquid is
methods are used to overcome the existing shortcomings of
measured and used for fault detection [3]. The measured
these methods. Commonly used intelligent methods are
gases by DGA method and used for transformer fault
artificial neural networks [5], expert systems [6], fuzzy
diagnosis are as follows: Hydrogen, Acetylene, Methane,
methods [7] and machine learning methods [8, 9].
Ethylene, Ethane, Carbon monoxide, Carbon dioxide [3]. The
One of the frequently used data sets in the literature in gas concentrations obtained by DGA are diagnosed by using
diagnosing power systems is the IEC TC 10 dataset. Using this both traditional and intelligent methods.
data set, a fault diagnosis has been made with different
A. Classification of Transformer Fault Types
classification algorithms. Liu et al. used five characteristic gas
ratios from the IEC TC 10 dataset and 115 data, TPC datasets There are six different types of possible faults in
in 2012 and made fault classification with Decision tree [10]. transformers [3, 4]. These:
In 2015, Malik et al. created a data set by combining the IEC

978-1-6654-4058-5/21/$31.00 ©2021 IEEE

 1) Partial Discharge (PD): It occurs as a result of TABLE I. INFORMATION OF TRADITIONAL METHODS
breakdown of the insulation in the transformer.
Traditional Used Gas Fault Types
2) Low Energy Discharge (D1): They are small, regional Method Ratio/Percentages
arcs that occur as a result of aging or loss of insulating Rogers C2H2/C2H4 Unit Normal
properties of oil. Ratios CH4/H2 Low Energy arcing- PD
3) High Energy Discharge (D2): They are large diameter Method C2H4/C2H6 Arcing- High Energy Discharge
arcs caused by the progression of low energy discharge and Low temperature thermal fault
resulting in short circuit failure. Thermal fault (<700°C)
4) Thermal Fault (T1): Thermal failure occurs below 300 Thermal fault (>700°C)
° C due to operating under overload or failure in the cooling Doernenburg CH4/H2 Thermal Fault
system. Ratios C2H2/C2H4 Low Energy Discharge
Method C2H2/CH4 High Energy Discharge
5) Thermal Fault (T2): Thermal failures that occur at a
C2H6/C2H2
temperature between 300 ° C and 700 ° C.
IEC Ratios C2H2/C2H4 No fault
6) Thermal Fault (T3): It is the fault that occurs as a result Method CH4/H2 Thermal Fault(<150°C)
of short circuit currents in the core layers at temperatures C2H4/C2H6 Thermal Fault (150°C- 300°C)
above 700 ° C. Thermal Fault (300°C- 700°C)
B. Traditional Methods Thermal Fault (>700°C)
Traditional methods are based on expert knowledge and Low Energy PD
experience. Statistics of malfunctions occurring over many High Energy PD
years and boundary conditions and graphs created based on Low Energy Discharge
these are used. High Energy Discharge

Commonly used traditional methods: Rogers Ratio Duval %CH4 Partial Discharge (PD)
Method, Doernenburg Ratio Method, IEC Ratio Method and Triangle % C2H4 Low Energy Discharge (D1)
Duval Triangle Method. Information on traditional methods Method % C2H2 High Energy Discharge (D2)
is given in Table I. Thermal Fault(T1) (<300°C)
Thermal Fault (T2) (300°C-
1) Rogers Ratio Method:
700°C)
In order to make fault classification with this method, Thermal Fault (T3) (>700°C)
three ratios of gas densities obtained from DGA analysis are Electrical/ thermal fault (DT)
used to classify transformer conditions in six different
situations [3]. 1) Decision Tree:
2) Doernenburg Ratio Method: It is a method with high diagnostic accuracy, where the
rules and decision process can be obtained. The rule base
In order to make fault classification with this method, four consists of branches and nodes [9]. Nodes represent tags
ratios of gas densities obtained from DGA analysis are used while branches represent classification conditions. The rule
to classify transformer conditions in three different situations base of the decision tree can be obtained by following the
[3]. nodes and branches.
3) IEC Ratio Method: Thanks to the decision tree used in transformer fault
This method makes fault classification in nine different diagnosis, user classification rules and decision mechanisms
situations using three ratios of gas obtained from DGA can be learned. This helps maintenance personnel to diagnose
analysis [4]. the situation of the oil [9]. Data quality is very important in
the Decision tree method, the higher the data quality, the
4) Duval Triangle Method: higher the diagnostic accuracy is obtained.
In order to make fault classification with this method, a 2) K-Nearest Neighbors (KNN):
triangular rule base is formed by using the percentages of KNN is an easy-to-apply and easy-to-understand method
three different gas densities. Transformer condition is that classifies unlabelled data according to the label of the
classified in six different ways [3]. closest data [15].
First, the distance of the unlabeled data to the labeled data
is obtained by the distance measurement method. Next, the
C. Intelligent Methods
number of k neighborhoods is determined and the tag of the
These are computer-aided intelligent methods used to closest neighbors is determined. Finally, the label of the data
increase diagnostic accuracy and overcome the shortcomings with the highest number of nearby parts is taken as the label
of traditional methods. Commonly used intelligent methods: of the data to be classified [15].
Expert systems, Artificial Neural Network, Fuzzy Systems,
Machine Learning Methods.
3) Support Vector Machine (SVM):
In this study, fault classification is made using K-Nearest It is a method that aims to find a suitable hyperplane with
Neighbors and Decision tree methods, which are machine better generalization ability and good performance that will
learning methods, and the results are compared.
maximize the margin between the data and the differential
hyperplane [16, 17].
SVM does not easily fall to the local optimum. It can
easily optimize model structure and parameters [16]. It also
has high diagnostic accuracy. Because of these advantages, it
is one of the most frequently used machine learning methods
in transformer fault diagnosis.
III. APPLICATION FOR TRADITIONAL METHODS
An application is made to diagnose power transformers
using traditional methods from DGA interpretation
techniques in MATLAB GUI. The diagnostic rule base has
been established using the rules in the existing standards for
traditional methods. A user interface has been created where
DGA gas concentrations are entered by the user, gas rates and
percentages are calculated and the type of failure is obtained
and used in diagnosis. Fig. 2. Example of MATLAB GUI

The user interface created is given in the Fig. 1. First, the

DGA data are entered by the user, then the diagnosis is made IV. APPLICATION OF TRANSFORMER FAULT CLASSIFICATION
by creating the gas ratios and percentages required for BY INTELLIGENT METHODS
traditional methods.
A. Dataset
In this study, the IEC TC10 dataset is used. The dataset
consists of 167 DGA gas concentrations [18]. The data set
contains 7 gas ratios. These are: hydrogen, methane, ethane,
ethylene, acetylene, carbon monoxide and carbon dioxide
Transformer failure status have been classified as four
different situations. These are: Partial Discharge, Arc
Discharge, Thermal Fault and Normal State. As shown in the
data set in the Fig. 3, there is 5.38% PD fault, 20.35% thermal
fault, 44.31% arching fault, 29.94% Normal state.

Fig. 1. MATLAB GUI User Interface

Traditional methods may fail to diagnose in some cases.

On the rule bases created for traditional methods, a faulty
definition has not been made for any cases. In some cases,
traditional methods may produce results that are incompatible
with the real situation or with each other. In this application, Fig. 3. Fault percentages in IEC TC 10 dataset
diagnosis has been made to traditional methods and
deficiencies have been shown.
B. Data Pre- processing
In the case where the transformer real fault is low-energy
discharge, the Rogers ratio and IEC ratio methods cannot While the gas concentrations obtained from Dissolved
produce a fault, while Doernenburg has classified it as high- Gas Analysis are at low ppm values in normal transformers,
energy discharge, only the Duval triangle method has been they are very high in case of any malfunction. Since gas
applied the correct fault diagnosis. This situation is shown in concentrations of different values are troublesome for the
the user interface in Fig. 2. operation of statistical learning algorithms, gas
concentrations are normalized and used in classification
problems. The data are sorted by size, using a logarithmic
transformation to put gas concentrations in order of
magnitude [19]. Then, standardization process is applied to
the obtained data.
The steps of the process for data preprocessing are given
in the Fig.4. Where xi is DGA gas concentration, X is the
logarithmic transform of xi , E(X) is the mean value all
examples and is the standard deviation of the X.
Fig. 5. Flow chart of classification methods

xi =DGA gas ( ) Precision= (1)

X=log10(xi) input=
concentration √

Recall= (2)
Fig. 4. Data pre-processing (Log- normalized and standardizing) ∗ ∗
F1 Score= (3)
C. Machine Learning Methods
In the present study, Support Vector Machine, Decision The confusion matrix is used to show actual values and
Tree and k-Nearest Neighbors algorithms are used for predicted values and to see them comparatively. The
transformer fault classification using IEC TC 10 dataset. diagonals of the matrix show the correctly predicted classes,
First, the dataset is prepared for the entry of preprocessed while the non-diagonal components show incorrect
algorithms. Then, the data set is randomly divided into two prediction results.
as training data set and test data set. While 80% of the data The confusion matrices and classification measures of
set is reserved as training data set, 20% of it is used as test the classification methods trained using the IEC TC 10
data set. However, data sets are arranged to include examples database are given in Table II, Table III and Table IV.
from each class. The classification algorithm used is trained
with the training data set, and then the classification Table II contains the confusion matrix and classification
performance of the algorithm is measured with the test data measures of the Decision Tree method. Decision tree method
set. Finally, the classification performances of the algorithms correctly predicted one of the PD failures in the test data set.
are compared. However, the wrong predicted one sample. At the same time,
this method correctly predicted three thermal failures while
The flow chart for the study is given in the Fig. 5. The predicting three different failures. In total, the number of
classification process is completed by applying these steps. correctly predicted faults in the matrix diagonal is 28, while
Classification algorithms are made using Python the number of faults outside the diagonal is 6.
programming language.
TABLE II. CONFUSION MATRICE AND CLASSIFICATION MEASURES OF
DECISION TREE
D. Classification Result
Decision Tree
A confusion matrix is used to demonstrate the Actual Classification
classification ability of classification algorithms. The Class\ measures
Thermal Fault
Arching Fault

Confusion matrix consists of True positives, true negatives,

Normal State

Predicted
false positives, and false negatives [20]. Of these Class
PD

Precision

components, True positives (TP) represent the number of

F1 Score
Recall

class instances that are correctly recognized. True negatives

(TN) represent the number of correctly classified patterns that
does not belong to this class. False Positives (FP) represent
the number of samples incorrectly assigned to the class, while PD 1 0 1 0 1 0,50 0,67
False negatives (FN) represent the number of samples are not Arching 0 13 0 1 0,87 0,93 0,90
recognized as class instances. Using these calculations, the Fault
classification measures, Precision, Recall and F1 score are Thermal 0 2 3 1 0,60 0,50 0,55
calculated. Precision represents the rate at which the data Fault
class tag matches the classifier tag. Recall represents the Normal 0 0 1 11 0,85 0,92 0,88
State
classifier activity for each class tag. F1 Score is the harmonic
mean of precision and recall used in unbalanced classification
problems [20]. Classification measures are given in the (1),
(2) and (3) respectively.
 by the user correspond to commonly used traditional
methods. However, it is shown by example that in some cases
TABLE III. CONFUSION MATRICE AND CLASSIFICATION MEASURES OF a fault diagnosis cannot be made due to the deficiencies that
KNN
exist at the limits of the definition in traditional methods.
KNN
In order to overcome the shortcomings of traditional
Actual Classification
Class\ measures methods, fault diagnosis has been conducted using intelligent

Thermal Fault
Arching Fault methods. Python programming language is used for

Normal State
Predicted
Class classification with intelligent methods. In the classification
PD

Precision
process, the IEC TC 10 dataset is used. The classification

F1 Score
Recall
results are obtained by randomly dividing the data set so that
the train-test ratio is 80-20%.

PD 2 0 0 0 1 1 1
The obtained classification results show that SVM has a
higher performance than other classification methods. The
Arching 0 12 0 2 1 0,86 0,92
Fault obtained results showed that SVM is the most appropriate
Thermal 0 0 3 3 1 0,50 0,67 fault classification method for the dataset used.
Fault
Normal 0 0 0 12 0,71 1 0,83
State
REFERENCES
TABLE IV. CONFUSION MATRICES AND CLASSIFICATION MEASURES [1] Fu Yang, Jin Xi and Lan Zhida, "A neural network approach to power
OF SVM transformer fault diagnosis," Sixth International Conference on
Electrical Machines and Systems, 2003. ICEMS 2003., Beijing, China,
SVM 2003, pp. 351-354 vol.1.
Actual Classification [2] M. Duval, "Dissolved gas analysis: It can save your transformer," in
Class\ measures IEEE Electrical Insulation Magazine, vol. 5, no. 6, pp. 22-27, Nov.-
Thermal Fault
Arching Fault

Normal State

Predicted Dec. 1989.

Class [3] IEEE Guide for the Interpretation of Gases Generated in Mineral Oil-
PD

Precision

F1 Score

Immersed Transformers. IEEE Std C57.104TM- 2019. IEEE Power

Recall

and Energy Society.

[4] IEC 60599 Standard. Mineral Oil-Impregnated Electrical Equipment in
Service–Guide to the Interpretation of Dissolved and Free Gases
Analysis; IEC 60599 Standard: Geneva, Switzerland, 2007.
PD 2 0 0 0 1 1 1
[5] J. L. Guardado, J. L. Naredo, P. Moreno and C. R. Fuerte, "A
Arching 0 14 0 0 0,82 1 0,90 comparative study of neural network efficiency in power transformers
Fault diagnosis using dissolved gas analysis," in IEEE Transactions on
Thermal 0 1 5 0 1 0,83 0,91 Power Delivery, vol. 16, no. 4, pp. 643-647, Oct. 2001.
Fault [6] D. Ma, W. Zhang and W. Yao, "Establish Expert System of
Normal 0 2 0 10 1 0,83 0,91 Transformer Fault Diagnosis Based on Dissolved Gas in Oil," 2013
State International Conference on Information Science and Cloud
Computing Companion, Guangzhou, China, 2013, pp. 681-685.
The classification accuracies of machine learning [7] S. Apte, R. Somalwar and A. Wajirabadkar, "“Incipient Fault
methods are given in Table V. The highest classification Diagnosis of Transformer by DGA Using Fuzzy Logic”," 2018 IEEE
accuracy is 91% and is achieved with SVM. The lowest International Conference on Power Electronics, Drives and Energy
Systems (PEDES), Chennai, India, 2018, pp. 1-5.
classification accuracy is 82% and is obtained with the
[8] J. Li, Q. Zhang, K. Wang, J. Wang, T. Zhou and Y. Zhang, "Optimal
Decision tree. Comparing the results, it is seen that SVM is dissolved gas ratios selected by genetic algorithm for power
the most suitable diagnostic method for the IEC TC10 transformer fault diagnosis based on support vector machine," in IEEE
dataset. Transactions on Dielectrics and Electrical Insulation, vol. 23, no. 2, pp.
1198-1206, April 2016.
TABLE V. ACCURACY OF CLASSIFICATION METHODS [9] H. Xie, J. Chen, P. Zhang, F. Zhang, X. Meng and J. Cheng,
"Knowledge Acquisition for Transformer Condition Assessment Using
Classification Actual Classification Synthetic Minority Over-sampling Technique and Decision Tree
Method Prediction/Test Accuracy (%) Algorithm," 2019 IEEE Innovative Smart Grid Technologies - Asia
samples (ISGT Asia), Chengdu, China, 2019, pp. 4186-4191.
DECISION TREE 28/34 %82 [10] Chih-Hsuan Liu, Tai-Li Chen, Leeh-Ter Yao and Shun-Yuan Wang,
"Using data mining to dissolved gas analysis for power transformer
KNN 29/34 %85 fault diagnosis," 2012 International Conference on Machine Learning
and Cybernetics, 2012, pp. 1952-1957.
SVM 31/34 %91 [11] H. Malik and S. Mishra, "Extreme learning machine based fault
diagnosis of power transformer using IEC TC10 and its related data,"
2015 Annual IEEE India Conference (INDICON), 2015, pp. 1-5.
[12] Z. B. Sahri and R. B. Yusof, “Support vector machine- based fault
V. CONCLUSIONS diagnosis of power transformer using k nearest- neighbor imputed
DGA dataset,” Journal of Computer and Communications, vol. 2, pp..
In this study, applications related to traditional methods 22-31, July 2014.
and intelligent methods used in diagnostics with DGA data in [13] H. Ma, W. Zhang, R. Wu and C. Yang, “Apower transformers fault
transformer fault diagnosis have been made accurately. First, diagnosis model based on tree DGA ratios and PSO optimization
a user interface is created in MATLAB GUI to diagnose with SVM,”IOP Conference Series: Materials Science and Engineering, vol.
traditional methods. Thanks to the interface created, it is easy 339, pp. 15-17, December 2017.
to find out which malfunction the gas concentrations entered
[14] Y. Zhang et al., "A Fault Diagnosis Model of Power Transformers and K-L feature extracting,” 2008 International Conference on
Based on Dissolved Gas Analysis Features Selection and Improved Machine Learning and Cybernetics, Kunming, China, pp 1510-1514
Krill Herd Algorithm Optimized Support Vector Machine," in IEEE [18] M. Duval and A. dePabla, "Interpretation of gas-in-oil analysis using
Access, vol. 7, pp. 102803-102811, 2019. new IEC publication 60599 and IEC TC 10 databases," in IEEE
[15] M. E. A. Senoussaoui, M. Brahami and I. Fofana, “Combining and Electrical Insulation Magazine, vol. 17, no. 2, pp. 31-41, March-April
comparing various machine-learning algorithms to improve dissolved 2001.
gas analysis interpretation,” IET Generation, Transmission & [19] P. Mirowski and Y. LeCun, "Statistical Machine Learning and
Distribution, vol. 12, no. 15, pp. 3673-3679, August 2018. Dissolved Gas Analysis: A Review," in IEEE Transactions on Power
[16] J. Meng, F. Guo, L. Wang, X. Zhou and Y. Fan, “Transformer fault Delivery, vol. 27, no. 4, pp. 1791-1799, Oct. 2012.
diagnosis based on artificial bee colony algorithm and support vector [20] M. Sokolova and G. Lapalme, “A systematic analysis of performance
machine,” 2017 2nd International Seminar on applied Physics, measures for classification tasks,” Information Processing and
Optoelectronics and Photonics (APOP 2017), pp. 94-100. Management, vol. 45, pp. 427-437.
[17] X. Y.Sun, D.H. Liu, J.P. Bian JP, “The study of fault diagnosis model
of DGA for oil-immersed transformer based on SVM active learning

View publication stats