Pattern Recognition of Development Stage of Creepage Discharge of Oil-Paper Insulation under AC-DC Combined Voltage based on OS-ELM

The recognition of the development stage of creepage discharge of oil-paper insulation under AC-DC combined voltage is the basis of fault monitoring and diagnosis of converter transformers, and there are few related studies. In this paper, the AC-DC combined voltage with a ratio of 1:1 was used to study the development stage recognition method of creepage discharge of oil-paper insulation under the cylinder-plate electrode structure. Firstly, the pulse current method was used to collect the discharge signals in the process of creepage discharge development. Finally, based on the online sequential extreme learning machine (OS-ELM), the above characteristic parameters were used to recognize the development stage of oilpaper insulation creepage discharge. The research results show that when the size of the sample training set in the OS-ELM algorithm is closed to the number of hidden layer neurons, higher recognition accuracy can be obtained, and the type of activation function has less influence on it. Based on the OS-ELM algorithm, the developmental stage of the creepage discharge is recognized. The development process of the creepage discharge is recognized as four stages, which is the same as the trend of the characteristic parameters of the whole creepage discharge development process, and the recognition accuracy is 91.4%. The algorithm has fast rate, high accuracy and batch training data characteristics, which can be widely used in the field of online monitoring and evaluation of electrical equipment status.


INTRODUCTION
The insulating structure and operation conditions of converter transformer are different from those of traditional AC transformer. The valve-side winding of the converter transformer will withstand AC-DC combined voltage and harmonic voltage components. According to CIGRE statistics, the failure rate of converter transformer is about twice that of AC transformer, and a considerable part of insulation failure is caused by creepage discharge at the oilpaper interface [1][2][3]. They have carried out some researches on the creepage discharge characteristics of oil-paper insulation under AC-DC combined voltage by partial discharge measurement and other methods, and have obtained some meaningful research results [1][2][3]. Partial discharge is a discharge phenomenon that can reflect the insulation failure process. It can not only be used as a manifestation of transformer oil-paper insulation performance degradation, but also can determine the severity of discharge development by extracting and analyzing partial discharge signals, and can further effectively evaluate the state of oil-paper insulation. Therefore, the research on the development characteristics of the creepage discharge of oilpaper insulation under AC-DC combined voltage can effectively reflect the condition of converter transformer.
At present, the research on the creepage discharge characteristics of oil-paper insulation under AC-DC combined voltage mainly focuses on two aspects. On the one hand, they investigate the factors affecting creepage discharge, including different voltage types and electric field components, discharge characteristics, aging and other factors. On the other hand, the researcher is mainly to find the characteristic parameters and recognition methods which can effectively characterize the development process of creepage discharge under combined AC-DC voltage [1][2][3]. At present, the techniques for recognition of partial discharge development process mainly include fingerprint recognition technology of discharge gray image technology, SVM (Support Vector Machine, SVM), BPNN (Back Propagation neural network, BPNN) and fuzzy clustering analysis and radar spectra [3]. But, there is little research on the pattern recognition of the development stage of creepage discharge of oil-paper insulation under AC-DC combined voltage. Moreover, the above recognition methods can perform pattern recognition after collecting all the data samples. When new data is added, it is necessary to recalculate all the data again, which is time-consuming and inefficient. Therefore, this paper builds a set of creepage discharge test platform of oil-paper insulation under AC-DC combined voltage. Then, based on the online sequential extreme learning machine (OS-ELM) algorithm, the discharge development stage is identified in the development stage of creepage discharge of oil-paper insulation. Combined with the variation law of the characteristic parameters of the creepage discharge, the recognition method of the development stage of the creepage discharge is established.

A. Sample and electrode model
Karamay 25 # transformer oil was used in the test. Before the test, the gas volume fraction of transformer oil was less than 2% and the micro-water volume fraction was less than 10 -5 through the process of degassing, drying and slag removal, which meet the requirements of GB/T7595. The insulating paper was dried at high temperature (100 °C) for 72 h. Then, the insulating paper was impregnated with the previously treated transformer oil for 72 hours under a vacuum of about 10 Pa and temperature of 60 °C [1].

Fig. 1 Cylinder-plate electrodes configuration
In the test, the cylinder-plate electrode model specified by CIGRE Method II and IEC60243 was used, and oilimmersed insulating paper was placed between the highvoltage electrode and the ground electrode, as shown in Figure 1 [1].

Fig. 2 Schematic diagram of experimental platform
The structure of the test system is shown in Figure 2. The AC power and DC power used for the test were all non-PD equipment. The test circuit design was the same as that in reference [1]. The PD measuring instrument used the pulse current method to measure, and the sampling frequency of the system was 5 MHz, and the PD bandwidth was 0.04 ~ 1 MHz. The AC-DC combined voltage was applied to the tested sample through the protective resistor. When the whole test system was loaded to 70 kV (peak value) without test sample, there is no partial discharge signal, and the sensitivity of the test system was 0.1pc.

C. Experimental method
The ratio of AC-DC combined voltage during the test was selected according to the general criteria and test requirements proposed by IEEE and CIGRE for oil-immersed converter transformer test. It was specified that the ratio of AC-DC combined voltage was 1:1, 1:3, 1:5, 1:7, where the AC was valid value and DC was the average [1][2]. In order to obtain the relatively stable and sustainable development of the creepage discharge signal, this paper selected the AC-DC combined voltage ratio of 1:1 to study the development stage of the creepage discharge of oil-paper insulation. The boost steps of the AC and DC voltages were set to 1 kV and 1 kV, respectively. The constant voltage was started after boosting to 1.1 times the AC discharge inception voltage, wherein the AC voltage was 37 kV and the DC voltage was 37 kV.
In order to accurately record the development process of creepage discharge of oil-paper insulation under the AC-DC combined voltage, a set of data was recorded every 1 min interval from the beginning of constant voltage in the test, and each time was recorded continuously for 1 min until the breakdown occurs. The above test was repeated 12 times. There are 12 groups of experimental data and a total of 1210 sets of data were obtained.

III. PATTERN RECOGNITION METHOD BASED ON OS-ELM
Liang and Huang G B et al propose the OS-ELM algorithm, which is an improved algorithm based on ELM [4].As a single hidden layer feedforward neural network, OS-ELM algorithm is composed of three parts: input layer, single hidden layer and output layer. As shown in Figure 3, the connection weight between input layer and hidden layer and the threshold value of hidden layer neurons are randomly generated, which do not need to be repeatedly iterated and adjusted in the training process. OS-ELM algorithm has the advantages of fast learning rate and good generalization performance [4].

Fig. 3 Structure of OS-ELM
The key question of OS-ELM algorithm is the learning process of output weight of single hidden layer network, which is divided into two steps. The first step is the initial stage, and the output weight β of the single hidden layer is obtained by training a small number of sample data. After the initial stage is completed, the second step is the online learning stage, which uses the updated sample data to update the output weight β obtained in the initial stage.
In the initial stage, there are N sample data ( Xi) , where Xi= [x1, x2, …, xn] T ∈R m , m is the dimension of the sample data set, Ti=[t1, t2,…,tn] T ∈R n , n is the dimension of the output vector. Firstly, the number of hidden layer neurons L is determined, then the connection weight w of input layer and the threshold b of hidden layer neurons are randomly set, then the appropriate activation function g (w, x, b) is selected, and finally the initial output matrix H0 of hidden layer is calculated. According to the theory of ELM, find the minimum β0 that can satisfy ||H0β-T0||, where According to the generalized inverse matrix algorithm, the initial output weight β0 is calculated, and the initial learning stage ends, as shown in formula (3). The second step is the online learning stage. Assuming that there are N1 new sample data entering the model, it is desirable to obtain β1 such that formula (4) can be the minimum.
According to the algorithm of generalized inverse matrix, the output weight β1 is calculated, as shown in formula (6).
The output weight βk+1 can be obtained by using the Woodbury formula and the recursive relationship, as shown in formula (7) The above is the basic principle of the OS-ELM algorithm, and the output weight β of the single hidden layer neural network is adjusted through two stages.

A. Development process of oil-paper insulation creepage discharge under AC-DC combined voltage
Because the partial discharge of the oil-paper insulation under AC-DC combined voltage is different from that under the AC voltage. In this paper, the interval time Δt of adjacent two discharges is used instead of the phase φ in the AC partial discharge. Then the statistical characteristic spectra is formed by combining interval time Δt, the partial discharge magnitude Q and the number of discharges n. Time resolved pulse sequence analysis (TPRSA) was used to study the development of creepage discharge of oil-paper insulation under AC-DC combined voltage [1][2][3].
In this paper, four sets of data in the beginning stage, the middle two stages and the adjacent end stage of one test are selected for statistical analysis. The Q-Δtpre-n threedimensional characteristic spectra is formed (where Δtpre is the time interval between the current discharge and the previous discharge). Figure 4 shows the variation law of the Q-Δtpre-n characteristic spectra of the above four sets of data during the development of the creepage discharge.  As shown in Figure 4(a), the n is small at this stage, and the Δtpre is large, and the distribution is in the range of 0.01-0.06 s. The Q is mainly concentrated at about 400 pC, and the distribution of Q is very scattered. As shown in Figure 4(b), the n at this stage increases slowly, and the Q is mainly distributed in the region below 1300 pC, but Q greater than 2000 pC gradually appear. The Δtpre is gradually shortened and distributed within a range of 0.04 s. As shown in Figure  4(c), the n of various Q at this stage shows a large increase compared to the second stage. They are mainly distributed around 1700 pC, and the Q gradually develops into the range of 2000-3000 pC, and the Δtpre is mainly concentrated in the range of 0.02 s. As shown in Figure 4(d), the n is still increasing gradually. The Q is mainly greater than 2000 PC ， and the Δtpre is mainly concentrated in the range of 0.01 s.

B. Pattern recognition of development process of creepage discharge based on OS-ELM
Although the TPRSA spectra can better reflect the time domain distribution of the development process of the creepage discharge, it is difficult to directly use the spectra for pattern recognition because the spectra contain too much information. n-Q, n-Δt, Δtpre-Q，Qmean-Δtpre four kinds of characteristic spectra, we extract their skewness (Sk), kurtosis (Ku), asymmetry (Asy) and cross-correlation coefficient (Cc) and Weibull distribution parametersαand β. Each set of data extracted 24 characteristic parameters from above characteristic spectra. The OS-ELM algorithm code is written based on MATLAB. The results shown that as the number of hidden layer neurons increases, the training accuracy increases gradually, and tends to be stable when the number of hidden layer neurons is close to the number of training sample. So, Sigmod function is selected as the activation function, and the number of hidden layer neurons is 650. There are 12 groups of experimental data, of which 7 groups were randomly selected as the training sample set to train the OS-ELM. The remaining 5 groups were used as a test sample to identify the development stage of creepage discharge and draw a confusion matrix of the recognition results.

Fig. 5 Pattern recognition result of OS-ELM algorithm
The confusion matrix of the recognition results is shown in Figure 5. It can be seen from Figure 5 that the whole development process of creepage discharge can be identified as four different stages by OS-ELM algorithm. The four stages were defined as initiation (C1), development (C2) acceleration (C3) and critical flashover (C4).
In this table, the total number of each row represents the real state data, each column represents the identified state data. The elements on the main diagonal represent the correct number of the development state of the creepage discharge identified according to the test data, and the non-main diagonal elements represent the number of misidentified. 500 set of data were used as test samples. In the first row, the total number is 120, and the four states are 113,5,2,0, respectively.
These data indicate that the total number of real states should be 120. After the algorithm diagnosis, the number of diagnosed C1 is 113, the number of misdiagnosed as C2 is 5, the number of misdiagnosed as C3 is 2, and the number of misdiagnosed as C4 is 0. The last column is the recall rate, which reflects the ratio of the correct number of samples to the actual number of samples. The last row of elements is the precision rate, which reflects the ratio of the predicted number of correct samples to the number of predicted samples. The position at which the recall rate and the accuracy rate intersect is the recognition accuracy, which reflects the ratio of the correct number to the total number of samples. It shows that the recall rate and precision of creepage discharge development process are both very high (>88%) based on the OS-ELM algorithm, and the sample recognition accuracy rate is 91.4%. The recognition result is the same as the result of dividing according to the variation law of the characteristic parameters of the creepage discharge development process in Figure 5 [1]. According to this method, it can provide theoretical support for the pattern recognition of creepage discharge development stage of oilpaper insulation.

V. CONCLUSION
For the creepage discharge data sample type, the hidden layer activation function type has little effect on the training accuracy. In the OS-ELM algorithm parameter selection process, when the sample set size is close to the number of hidden layer neurons, the training rate can be improved and a higher training accuracy rate can be obtained. Based on the OS-ELM algorithm, the development process of the creepage discharge is recognized as four stages, and the recognition accuracy is 91.4%. The algorithm has the advantage of being able to train samples in batches, and has good engineering application prospects in the field of electrical equipment condition monitoring and evaluation.