Power Transformer Fault Diagnosis Based on Dissolved Gas Analysis by Artificial Neural Network

Souahlia Seifeddine 1, Bacha Khmais 2, Chaari Abdelkader 3

1C3S, 5 Taha Hussein street – Tunis, e-mail: seifeddine.souahlia@gmail.com

2C3S, 5 Taha Hussein street – Tunis, e-mail: khmais-bacha@voila.fr 2C3S, 5 Taha Hussein street – Tunis, e-mail: abdelkader.chaari@yahoo.fr

ABSTRACT

This paper presents an intelligent fault classification approach to power transformer dissolved gas analysis (DGA). Artificial neural network (ANN) is powerful for the problem with small sampling and high dimension. ANN is applied to establish the power transformers faults classification and to choose the most appropriate gas signature between the DGA traditional methods and a novel extension method. The experimental data from Tunisian Company of Electricity and Gas (STEG) is used to illustrate the performance of proposed ANN models. Then, the MLP and RBF classifier are trained with the training samples. Finally, the normal state and the six fault types of transformers are identified by the trained classifier. In comparison to the results obtained from the ANN, the proposed DGA method has been shown to possess superior performance in identifying the transformer fault type. The test results indicate that the ANN approach can significantly improve the diagnosis accuracies for power transformer fault classification.

Index Terms— Dissolved gas analysis, Multi-Layer Perceptron, Radial Basis Function, transformer fault diagnosis.

1. INTRODUCTION

Power transformers are important equipments in power systems. Any fault in the power transformer may lead to the interruption of the power supply and accordingly. So it is of vital importance to detect the incipient fault of the transformer as early as possible. Diagnosis of potential faults concealed inside power transformers is the key of ensuring stable electrical power supply to consumers.

Dissolved gas analysis (DGA) has been widely recognized as an effective diagnostic technique for power transformers faults detection. The analysis of specific dissolved gas concentrations in insulation oil of a transformer gives the knowledge about a transformer state, and therefore, allows taking the necessary preventive actions [1]. In the past years, various fault diagnosis techniques have been proposed, including the conventional key gas method [2], ratio method [3] and [4], and graphical representation method [5].

However, the identification of the faulted location by the traditional method is not always an easy task due to the variability of gas data and operational natures.

Recently, artificial intelligence techniques have been extensively used with the purpose of developing more accurate diagnostic tools based on DGA data. A. Shintemirov, W. Tang and Q. H. Wu proposed the Genetic Programming (GP) method for transformer fault detection.

The fuzzy logic [6] is used with three and four digit

codes containing the fault information are created based on the fuzzy logic to achieve better result. The method is applied to three transformers to diagnose the fault by analyzing the dissolved oil based on fuzzy logic.

The back propagation (BP)-based artificial neural nets [2] can identify complicated relationships among dissolved gas contents in transformer oil and corresponding fault types, the BP determines the optimal connection weights and bias terms to achieve the most accurate diagnosis model for DGA.

The present paper is aimed at applying ANN to automation of decision on the power transformers state and chooses the most appropriate gas signature between the DGA traditional methods and a novel method.

This paper consists of five sections. Section 2 illustrates principles of the faults types and DGA methods. Section 3 presents the regression arithmetic of MLP and RBF NN. Section 4 presents our power transformers fault diagnosis based on MLP and RBF and discusses the experimental results. Finally, Section 5 provides some important conclusions that we have drawn from this study.

2. DISSOLVED GAS IN THE TRANSFORMER OIL

2.1. Transformer fault types

Dissolved gas analysis (DGA) is a sensitive and reliable

technique to identify the power transformers faults. By using this technique, it is possible to discriminate fault in a great variety of oil-filled equipment. IEC Publication 60599

[7] provides a coded list of faults detectable by DGA. Table 1 tabulates the fault types and the codes

addressed in this paper.

Table 1. Fault Type used in Analysis.

Fault Type Code Partial discharge PD

Low energy discharge D1 High energy discharge D2

Thermal faults T <300 ° C T1 Thermal faults 300 <T< 700 ºC T2

Thermal faults T > 700 ºC T3 2.2. DGA interpretation methods

Many interpretative methods based on DGA to

detect the incipient fault nature have been reported. In this paper, three of the DGA methods were studied:

- Gas key method; - IEC Ratios method; - The graphical representation method. 2.2.1. Key gas method In this key gas method, we need five key gas

concentrations: hydrogen (H2), methane (CH4), acetylene (C2H2), ethylene (C2H4) and ethane (C2H6) available for consistent interpretation of the fault.

IEC 60599 standard establishes an interpretation by which five gases H2, CH4, C2H2, C2H4 and C2H6 can be used to detect different types of faults.

Table 2 shows the diagnostic interpretations applying various key gas concentrations.

Table 2. Interpretation gas dissolved in the oil.

Gas Detected Interpretation Oxygen (O2) Transformer seal fault

Oxide and Dioxide Carbon (CO and CO2)

Cellulose decomposition

Hydrogen (H2) Electric discharge (corona effect, low partial

di h )Acetylene (C2H2) Electric fault (arc, spark)Ethylene (C2H4) Thermal fault

(overheating local)

Ethane (C2H6) Secondary indicator of thermal fault

Methane(CH4) Secondary indicator of an arc or serious overheating

The ppm concentration typical values range

observed in power transformers according to IEC 60599

are given in Table 3:

Table 3. Concentration typical values observed in transformers.

H2 CH4 C2H6 C2H4 C2H2 CO CO260-150

40-110

50-90 60- 280

3-50 540-900

5100-13000

2.2.2. IEC ratios method The IEC Ratios method utilizes five gases H2, CH4,

C2H2, C2H4 and C2H6. These gases are used to produce a three gas ratios: C2H2 / C2H4, CH4 / H2 and C2H4 / C2H6.

Table 4 shows the IEC standard for interpreting fault types and gives the values for the three key-gas ratios corresponding to the suggested fault diagnosis. When key-gas ratios exceed specific limits, incipient faults can be expected in the transformer.

Table 4. Diagnosis using the ratio method (IEC 599).

Fault type

C2H2 / C2H4

CH4 / H2 C2H4 / C2H6

PD < 0.1 < 0.1 < 0.2 D1 > 1 0.1 – 0.5 > 1 D2 0.6 - 2.5 0.1 - 1 > 2 T1 < 0.1 > 1 < 1 T2 < 0.1 > 1 1 - 4 T3 < 0.1 > 1 > 4

2.2.3. The graphical representation method The graphical representation method using Duval’s

triangle is described in Appendix B of IEC 60599:1999 standard.

The concentrations (in ppm) of CH4, C2H2 and C2H4 are expressed as a percentage of the total (CH4 + C2H4 + C2H2) and define a point (%CH4, %C2H4, %C2H2) in a coordinate system represented as a triangular diagram (Figure 1), which has been subdivided in different zones. Each zone is related to a certain type of fault.

Zone DT in Figure 1 corresponds to mixtures of thermal and electrical faults.

Figure 1. Coordinates and fault zones of the

Triangle. We can translate the Figure 1 in a painting, that

gives the limits of each fault are summarized in Table 5.

Table 5. Graphical representation method zone limits.

PD 98 % CH4

100 % CH4

D1 23 % C2H4

13 % C2H2 100 % C2H2

D2 23 % C2H4

13 % C2H2 38 % C2H4

29 % C2H2

T1 4 % C2H2

10 % C2H4

T2 4 % C2H2

10 % C2H4

T3 15 % C2H2

50 % C2H4 100 % C2H4

3. BASIC CONCEPT OF ANN

The neural network technique is used to recognize and classify complex fault patterns without much knowledge about the process, the used trials or the fault patterns themselves. A neural network consists of many simple neurons which are connected with each other.

The principal neural network will use for the classification are: Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF).

3.1. Multi-Layer Perceptron

MLP is a network organized in layers. A layer is a

uniform neurons group without connection with each other makes a transformation vector.

The architecture of the MLP is composed by an input layer, a variable number of hidden layers and by an output layer, fully connected among them. In particular is outlined in Figure 2, a three-level fully

connected network using a sigmoid output function, has been considered, because it is known that this number of levels allows building decision regions of any shape.

Figure 2. The multi-layer perceptron architecture.

3.2. Radial Basis Function RBF nets belong to the group of kernel function nets

that utilize simple kernel functions, distributed in different neighborhoods of the input space, whose responses are essentially local in nature. The architecture consists of one hidden and one output layer. This shallow architecture has great advantage in terms of computing speed compared to multiple hidden layer nets.

Each hidden node in an RBF net represents one of the kernel functions. An output node simply computes the weighted sum of the hidden node outputs. A kernel function is a local function and the range of its effect is determined by its center and width. Its output is high when the input is close to the center and it decreases rapidly to zero as the input’s distance from the center increases. The Gaussian function is a popular kernel function and will be used in this algorithm [8].

4. TRANSFORMERS FAULTS CLASSIFICATION BASED ON ANN

4.1. DGA training and testing data

This study employed dissolved gas content training and

testing data in power transformer oil from chemistry laboratory of the Goulette central of Tunisian Company of Electricity and Gas (STEG). The data is divided into two data sets: the training data set (160 samples) and the testing data set (30 samples). The extracted DGA data contain not only the: five key gas concentrations, three ratios and three relative percentages but also the diagnosis results from on-site inspections. The training data sets have been evaluated using various methods DGA and the corresponding judgments related to seven classes have been provided, normal unit (70 samples), partial discharge (5 samples), low energy discharge (10 samples), high energy discharge (25 samples), low temperature overheating (10 samples), middle temperature overheating (15 samples) and high temperature overheating (25 samples).

13

DT

T1

50 38

PD

%C2H4 T2

D1 15

%CH4

D2

%C2H2

23

29T3

10

The transformer characteristics used for the taking of mineral oil samples are given in Table 6. From January 2008 to December 2010, the transformer was internally examined.

Table 6. Transformer characteristics used for DGA

samples.

No. 1 2 3 4 5 6 P

(MVA) 150 4 2 42.

5 25 40

V (KV) 235/ 15.5

15.5/6.8

6.6/0.4

230/11

150/30

150/30

No. 7 8 9 10 11 12 P

(MVA) 50 40 100 15 30 20

V (KV) 150/90

90/ 30

225/90

90/ 30

90/ 30

90/ 30

The ANN faults classification is performed using

traditional DGA methods and the proposed DGA method as gas signature.

4.2. Proposed ANN tools implementation An ANN-based power transformer fault diagnostic system includes input features, network topology, fault outputs as well as training patterns. In the current study, we used:

• Two types of neural networks: the MLP and RBF are used for transformers faults classification.

• Four types inputs data: five key gas (key gas method), three ratios (ratios method), three relatives percentages(graphical representation method) and the combined ratios and relatives percentages (combination ratios and graphical representation method) are chosen as input features;

• 3 binary outputs in order to minimizing the neurons number in output layer. The output codification is presented in the following:

001 : no fault (normal working) ; 010 : partial discharge fault (DP) ; 011 : low energy discharge fault (D1) ; 100 : high energy discharge fault (D2) ; 101 : low temperature overheating (T1) ; 110 : middle temperature overheating (T2) ; 111: high temperature overheating (T3).

4.2.1. MLP design and training

At first, we will classify the faults by MLP network.

After the test, with several parameters, we obtained the appropriate MLP architecture with the minimal error rate. Then, the optimal parameters are utilized to train the MLP model. So, we used:

• Transfer function tangent sigmoid for 4 layers which is given in Figure 3:

Figure 3. Tangent sigmoid transfer function.

• Four layers: an input layer, two hidden layers, and an output layer are shown in Figure 4:

Figure 4. MLP neural network training.

• The neurons and iterations numbers are tabulated in Table 7:

Table 7. Neurons and iterations numbers for MLP.

Gas signature Neurons number Epoch

number Input layer

hidden layers

output layer

Key gas 5 4 3 400 Ratios 3 3 3 400

Graphical representation

3 3 3 250

Combined ratios and graphical

representation

6 5 3 300

As Shown in Figure 5, a mean square error (MSE) curve

of training look very smooth without any oscillation and reached the desired goal at epoch 300.

-1

+1

+b

W

+b

W

+ b

W

+b

W

Input layer Hidden layer Output layer

0 50 100 150 200 250 30010

-3

10-2

10-1

100

300 Epochs

Tra

inin

g-B

lue

Performance is 0.00285051, Goal is 0

Figure 5. Error performance of MLP network training.

4.2.2. RBF design and training The parameters of RBF model are optimized by

many tests. The adjusted parameters with maximal classification accuracy are selected as the most appropriate parameters using:

o Mean squared error goal = 0.01 o Spread of radial basis functions = 1 o Maximum number of neurons = 40 o Number of neurons to add between displays = 1

4.3. Numerical tests and discussion

To demonstrate the effectiveness of the proposed

ANN fault diagnosis technique, various power transformers DGA results are tested. The detailed gas data are shown in APPENDIX A, where the AFC expresses the actual fault type, and MLPNN and RBFNN are the classification results by the MLP and RBF neural networks, respectively using combination of ratios and graphical representation method as input data.

The performance of MLP network is analyzed in terms false alarm rate and non-detection rate for four gas signature methods as can be seen in Table 8:

Table 8. The MLP classification performance.

Gas signature False

alarm rate (%)

Non-detection rate (%)

Key gas 0 30 Ratios 0 26.7

Graphical representation

0 23.3

Combined ratios and graphical

representation

0 20

The actual result indicates the superiority of the

combination of ratios and graphical method for the MLP network classification.

The performance of RBF network is analyzed in terms false alarm rate and non-detection rate for four gas signature methods as shown in Table 9:

Table 9. The RBF classification performance.

Gas signature False alarm rate

(%)

Non-detection rate (%)

Key gas 0 26.7 Ratios 0 20

Graphical representation

0 23.3

Combined ratios and graphical

representation

0 13.3

From Table 9, the actual result indicates that

combination of ratios and graphical representation has higher diagnostic accuracy than three other methods, as a gas signature input for the RBF network classification.

According to test results of the MLP and RBF networks, we conclude that the combination ratios and graphical representation method present the best performance; it can be used as a neural network input vector.

To select the appropriate neural network between the MLP and RBF networks, we compare the false alarm rate and non-detection rate given in Table 10:

Table 10. False alarm rate and non-detection rate of

MLP and RBF networks.

Neural networks

False alarm rate (%)

Non-detection rate (%)

MLP 0 20 RBF 0 13.3 The test samples examining by two neural network

shows that the types of faults detected by the RBF network are almost identical to the real fault types, the other network MLP didn't have to correct the faults. The table above shows that RBF network has more excellent performance than MLP network. In this case, we adopt the RBF network.

This justifies the importance of the approximate the choice of neural network parameters (the network number, number of neurons for each network, type of transfer functions and number of epoch).

To conclude, faults classification by ANN techniques can be achieved by the RBF NN using our proposed combination ratios and graphical representation method.

5. CONCLUSIONS

In this paper, the artificial neural network is

implemented for the automation of decision on the power transformers state and to choose the most appropriate gas signature as input data between the traditional DGA methods which are key gas, graphical representation and ratios methods, and our proposed combination ratios and graphical representation method using the artificial neural network classification of power transformers faults.

The effectiveness of ANN diagnosis has been analyzed with MLP and RBF NN. The real data sets are used to investigate its feasibility in forecasting the DGA methods in power transformer oil.

According to test results, it is found that the RBF NN has a better performance than the MLP and the proposed combination ratios and graphical representation method has than three other methods as a gas signature, in diagnosis accuracy. The accuracy of the RBF for faults diagnosis is comparable to conventional methods due to their great facilities for study.

The proposed method can be applied to online diagnosis of incipient faults in transformers. The experimental results reveal the potential of the proposed approach for forecasting the DGA method in power transformer oil.

6. REFERENCES [1] A. Shintemirov, W. Tang and Q. H. Wu, ‘‘Power

Transformer Fault Classification Based on Dissolved Gas Analysis by Implementing Bootstrap and Genetic Programming’’, IEEE transactions on systems, man, and cybernetics—part c: applications and reviews, VOL. 39, NO. 1, pp. 69–79, January 2009.

[2] SUN Yan-jing, ZHANG Shen, MIAO Chang-xin,

LI Jing-meng, “Improved BP Neural Network for Transformer Fault Diagnosis”, Journal of China University of Mining & Technology, Vol.17 No.1, pp. 138–142, Mars 2007.

[3] W. H. Tang, J. Y. Goulermas, Q. H.Wu, Z. J.

Richardson, and J. Fitch, “A Probabilistic Classifier for Transformer Dissolved Gas Analysis With a Particle Swarm Optimizer”, IEEE transactions on power delivery, VOL. 23, NO. 2, pp. 751–759, April 2008.

[4] Chin-Pao Hung, Mang-Hui Wang, “Diagnosis of

incipient faults in power transformers using CMAC neural network approach”, Electric Power Systems Research 71, pp. 235–244, 2004.

[5] Michel Duval, “A Review of Faults Detectable by Gas-in-Oil Analysis in Transformers”, IEEE Electrical Insulation Magazine, Vol.18, No. 3, pp. 8–17, June 2002.

[6] Rahmatollah Hooshmand and Mahdi Banejad,

“Application of Fuzzy Logic in Fault diagnosis in Transformers using Dissolved Gas based on Different Standards”, World Academy of Science, Engineering and Technology 17, pp. 157–161, 2006.

[7] Standard IEC 60599, ‘‘Guide for the interpretation of

dissolved gas analysis and gas-free’’, 2007. [8] Z. Moravej, D.N. Vishwakarma, S.P. Singh,

‘‘Application of radial basis function neural network for differential relaying of a power transformer’’, Computers and Electrical Engineering 29, pp. 421–434, 2003.

APPENDIX A. Tested gas data of transformer and diagnosis by ANN No. H2 CH4 C2H6 C2H4 C2H2 AFC ANN diagnosis

MLPNN RBFNN 1 19.3 103 159 19 0.6 T1 T3 No Fault 2 497 230 51 151 122 D2 D2 D2 3 615 200 42 102 68 D2 D2 D2 4 594 230 44 130 102 D2 D2 D2 5 27 30 23 2.4 0.1 No Fault No Fault No Fault 6 38 55 22 28 0.1 T2 No Fault No Fault 7 30 22 14 4.1 0.1 No Fault No Fault No Fault 8 23 63 54 10 0.3 T1 D1 No Fault 9 2.9 2 1.5 0.3 0.1 No Fault No Fault No Fault

10 4 99 82 4.2 0.1 T1 D1 T2 11 56 286 96 928 7 T3 T3 T3 12 78 161 86 353 10 T3 T3 T3 13 21 34 5 47 62 D2 D2 D2 14 50 100 51 305 9 T3 T3 T3 15 120 17 32 23 4 No Fault No Fault No Fault 16 172 335 171 812 37.8 T3 T3 T3 17 181 262 41 28 0.1 T1 No Fault T1 18 27 90 42 63 0.2 T2 T2 T2 19 160 130 33 96 0.1 No Fault No Fault No Fault 20 180 175 75 50 4 No Fault No Fault No Fault 21 127 107 11 154 224 D2 D2 D2 22 32.4 5.5 1.4 12.6 13.2 D2 D2 D2 23 345 112.3 27.5 51.5 58.8 D1 D1 D1 24 980 73 58 1.2 0.1 PD D1 PD 25 74 142.2 41.8 324.9 5.3 T3 T3 T3 26 76 140 40.8 317 5.2 T3 T3 T3 27 30.4 117 44.2 138 0.1 T2 T2 T2 28 30.8 149 47.9 146 0.1 T2 T2 T2 29 27 136 46.9 131 0.1 T2 T2 T2 30 1607 615 80 916 1294 D2 D2 D2