Application of BP-ART2 hybrid neural network based on fuzzy input in power transformer fault synthesis

1. Introduction
Power transformers are one of the important equipment in the power system and play a vital role in the safe operation of the power system. Good operation and maintenance of transformers, especially fault diagnosis, plays a very important role in improving the reliability of safe operation of the power system.

The emergence and gradual maturity of DGA has brought many conveniences to transformer fault diagnosis. There are many methods to use DGA to judge transformer faults, such as Rogers method, characteristic gas method, three-ratio method, electrical research method, etc. However, these methods themselves have a certain degree of imperfection. Based on DGA alone, the fault cannot be accurately judged and located. Combined with electrical tests, such as measuring DC resistance, insulation resistance, absorption ratio, etc., plus some fault characteristics, such as temperature increase, oil level drop, etc., comprehensive judgment can effectively improve the quality of diagnosis.

BP neural network has strong nonlinear approximation ability, can perform fault pattern recognition, fault severity assessment and fault prediction, and is widely used. However, it has low ability to handle abnormal faults and does not have incremental learning function. ART 2 model is a self-organizing network model that adopts unsupervised competitive learning rules. It does not have the problem of strong dependence of BP algorithm on sample knowledge. It can correctly identify abnormal faults and has a fast recognition speed. However, this model completes the pattern classification task through clustering. It cannot perform fault severity assessment and development trend prediction. Combining BP neural network with ART2 model, integrating supervised algorithm and unsupervised algorithm, using fuzzy quantity as input, to form a new fuzzy neural network to diagnose transformers, can achieve good diagnostic effect [1].

2. Common faults and characteristics of power transformers [3]
There are many kinds of transformer faults. Some common faults and their fault characteristics are as follows:
(1) Poor contact of decomposition switch: large DC resistance difference, characteristic gas contains both H2 and CO, and high CH4 or C2H4 content.
(2) Short circuit between winding turns: large ratio deviation, large DC resistance difference, high H2 and C2H2 content, and CO.
(3) Oil leakage in the on-load tap changer box: too high temperature, high oil level drop rate.
(4) Overheating fault: high CH4 and C2H4 content, and may also contain CO and CO2, and high temperature.
(5) Insulation aging: large dielectric loss tg, too low insulation resistance, and more CO, CO2 and CH4 in the characteristic gas.
(6) Severe moisture: large dielectric loss tg, high moisture content, absorption ratio less than 1.3, too low insulation resistance, and high H2 content in the characteristic gas.
(7) Partial discharge in oil: high H2, C2H2, CH4 and CO content.
(8) Wire break fault: large DC resistance difference, and the highest H2 content.
There are many common transformer faults and many causes of faults. Extract the various fault features and send them to the fault diagnosis model for analysis and synthesis, and finally obtain the fault diagnosis results.

3. Fault diagnosis model - establishment of multiple fuzzy neural networks
When using fuzzy neural networks for transformer fault diagnosis, considering that there are many samples and large data differences in actual applications, using one network is very complex, and has poor convergence and low diagnostic accuracy. Therefore, this paper divides the entire sample into several independent sub-sample sets based on certain characteristic indicators and certain rule combinations, and establishes multiple sub-fuzzy neural networks, as shown in Figure 1.

In the figure, x1, x2, and x3 are the input values ​​of the three-ratio method described in Table 2.
The first fuzzy neural network uses the measured values ​​of characteristic gases such as H2, C2H2, CH4, C2H4, C2H6, CO and CO2 as input to generate a series of outputs. The second fuzzy neural network uses the gas three-ratio method as input to generate a series of outputs. The third fuzzy neural network can use the measured values ​​of electrical tests such as DC resistance, insulation resistance, absorption ratio, polarization index, transformation ratio, dielectric loss tgδ, moisture content, etc. as input. The fourth fuzzy neural network can use the measured values ​​of oil level, oil temperature, etc. as input. After the input is fuzzified, it is sent to the BP neural network. After processing, a series of results are generated, which are sent to the ART2 model and then processed to generate diagnostic results. The outputs include: normal, insulation aging, short circuit between winding turns, poor contact of tap joints, insulation breakdown, severe moisture, partial discharge in oil, oil leakage in the on-load tap changer box, wire breakage, overheating fault, iron core short circuit, solid insulation arc decomposition, etc.
The BP neural network in the hybrid neural network has a three-layer structure as shown in the following figure:

BP1 has 3 layers, with 7 inputs. The 1st to 3rd inputs are H2, total hydrocarbons and C2H2 measurement. The 4th to 7th inputs are the proportions of C2H2, H2, CH4 and C2H4 in total hydrocarbons. There are 20 hidden layers and 6 outputs, which represent general overheating (>500℃), partial discharge, spark discharge, arc discharge and overheating and arc discharge respectively. BP2 also has 3 layers, with 3 inputs, 12 hidden layers and 9 outputs. The meanings of its inputs and outputs are shown in Table 2. BP1 and BP2 have been applied in the field, so their input, output and number of hidden layer neurons are given by experience; due to the limitations of field conditions, the input and output quantities of BP3 and BP4 and the number of hidden layers are determined by the actual measurement data that can be provided on site. The simulation uses the measured values ​​of dielectric loss tgδ, DC resistance, absorption ratio, oil level and moisture as input. The network also adopts a three-layer structure, with input layer, hidden layer and output layer of 3, 10, 6 and 2, 8, 5 respectively. The BP neural network adopts the learning algorithm described in reference 5. Since the BP algorithm has problems such as slow convergence speed and low learning accuracy, this paper adopts methods such as adding momentum factor, unequal weights and semi-random initial solution to solve them in order to speed up the convergence speed. [page]
The structure of the ART2 neural network is shown in Figure 3 [9]:

Adaptive resonance theory ART2 has a fast learning algorithm and does not require a large number of samples. It has great application potential in the field of online fault identification. Figure 3 is a typical single ART2 neural network structure, which is suitable for analog vector input. The network can be divided into two parts: attention subsystem and adjustment subsystem. The former completes the similarity matching and competitive selection of the input vector, and the latter verifies whether the similarity between the input pattern and the long-term memory pattern reaches a satisfactory level, and makes corresponding processing based on the test results, success or reset. The extracted feature vector Ii is input into the F1 layer (comparison layer). In the F1 layer, a stable middle-level pattern u is obtained through vector normalization and nonlinear transformation through iteration, and is sent to the F2 layer (recognition layer) through p. The F2 layer activates the F2 layer candidate pattern (corresponding to the fault type in this article) through competitive selection to obtain the short-term memory of the system. The output of the F2 layer is fed back to the F1 layer after being weighted by long-term memory. The feedback information is sent to the adjustment subsystem together with u to test the similarity between the system's long-term memory pattern and the input pattern. If the similarity test passes, it can be determined that the input pattern belongs to the candidate pattern of the F2 layer, and the weight learning is completed in one step according to the fast learning algorithm; if it fails the test, the F2 layer is forced to reset and select the next output node. If all output nodes cannot pass the matching test, a new output node, that is, another new class, is added.

When applying ART2, attention must be paid to the selection of ρ (similarity measure warning limit, a positive number between 0 and 1). The ρ value determines the interval size of the network's classification of input patterns, which directly affects the classification performance. If ρ is too small, the classification is rough and different fault types cannot be distinguished; if ρ is too large, the classification is too rough, and the same fault type may be classified into different output modes, causing misclassification. There is no specific rule for the selection of ρ, which needs to be adjusted in specific applications. In this paper, ρ is set to 0.5 to achieve a satisfactory classification effect. The ART2 network refers to the learning algorithm described in reference 10.

The transformer fault diagnosis process is a non-stationary, nonlinear random process. In the learning stage, by training a sufficient number of samples, the contact weights and thresholds are adjusted layer by layer until the error reaches the accuracy requirement. During the operation, different test samples are input to perform fault diagnosis pattern recognition, and finally the fault type and possible location of the fault are determined in real time.

4. Knowledge Processing
4.1 Fuzzy Knowledge Representation of Characteristic Gas

4.2 Fuzzy knowledge representation by three ratio method
Referring to Table 2, the use of ascending half normal distribution function can be insensitive to weak values, but sensitive to large values ​​that are sufficient to drown out noise. The distribution function is shown in the above formula.

4.3 Fuzzy processing of electrical test data

[page]
4.3.1 DC resistance [2]
Measuring DC resistance can generally be used to analyze faults such as broken wires, broken or unsoldered wires, short circuits, and poor contact of taps. GB stipulates that the difference between the measured values ​​of each phase (to be converted to the corresponding value at 20°C) should be less than 4% of the average value, and the difference between the measured values ​​between lines should be less than 2% of the average value. The actual difference should be compared with the recorded measured value in the factory test record.
4.3.2 Insulation resistance, absorption ratio and polarization index [2]
Measuring the insulation resistance, absorption ratio and polarization index of the winding can be used as a means to detect insulation breakdown and large-scale moisture failure of the transformer. According to GB regulations, the insulation resistance should not be less than 70% of the factory test value. When the absorption ratio k=R60/R15≥1.3, it is considered that the transformer is not affected by moisture. The state of the polarization index is shown in Table 3.

4.3.3 Dielectric loss tg [3]
Measuring dielectric loss tgδ plays a certain role in judging the overall condition of transformer insulation aging, moisture, etc. Under normal circumstances, tgδ (to be converted to the corresponding value at 20°C) is less than 3% for good, greater than 3% and less than 6% for attention, and greater than 6% for poor.
4.3.4 Moisture [4]
Measuring moisture is mainly used to judge the moisture condition of the transformer. Under normal circumstances, moisture less than 35ppm is good, greater than 35ppm and less than 50ppm for attention, and greater than 50ppm for poor.

5. Simulation
After statistically analyzing the fault examples published in relevant technical publications and corresponding materials on power transformers over the years, the data of 811 faulty transformers with clear conclusions after actual inspection and verification were selected. The distribution of each fault type in the training sample set and the test sample set formed by random selection is shown in Table 4.

According to the model described in this paper, the judgment results are shown in Table 5.
As can be seen from Table 5, the BP-ART2 hybrid neural network based on fuzzy input has a high diagnostic accuracy for power transformer fault diagnosis, and has a significant improvement in the diagnostic accuracy of conductive circuit overheating faults such as taps or leads, and discharge faults involving solid insulation such as turn-to-turn short circuits or lead flashovers. This shows that this method can indeed achieve good diagnostic results for transformer fault diagnosis.

6. Conclusion
The purpose of this paper is to find a new and effective method for power transformer fault diagnosis. To this end, the BP-ART2 hybrid neural network is used for this purpose. This method makes full use of the advantages of the BP neural network and the ART2 model, overcomes their respective shortcomings, and is a new attempt for power transformer fault diagnosis. The simulation results show that the application of this method can achieve good results.