Research On Feature Extraction And Pattern Recognition Of Partial Discharge In Oil-paper Insulation

Power transformer is one of the most important electrical equipment in power system. Its safety affects the safe and effective operation of the power grid. Partial discharge is the main reason, which causes the insulation damage of transformers. And the mechanism of different types of partial discharge is various. The degrees of insulation damage are also different. Therefore, partial discharge pattern recognition is an effective method to diagnose the insulation condition of high voltage electrical equipment. This paper analyzed the mechanism of the partial discharge damage. Four kinds of typical defect models were tested to study the various types of discharge pattern. We can extract the statistical characteristic parameters to classify and identify. The main works are as follows:Firstly, according to the CIGRE Method II, four kinds of typical defect models were devised to express different types of transformer fault defects. Partial discharge test system and detection circuit were established in the high voltage laboratory. And at the same time, the method of pulse current was used to detect the partial discharge signals, which provided numerical reference for discharge characteristics of different types. This paper also summaried the different discharge mechanism and waveform characteristics.Secondly, the method of wavelet transform was used to filter the noise of partial discharge signals, which used the wavelet of db8 and decompose into 5 layers. According to the PRPD patterns, thirty statistical characteristic parameters were extracted from the corresponding map spectrum, which were used as the characteristic parameters for different types of partial discharge in oil-paper insulation.Thirdly, PCA(principal component analysis) and KPCA(kernel principal component analysis) were used to reduce dimension of 30 statistical parameters, and then compared the two methods. The results show that nine of comprehensive features were extracted by principal component analysis. However, it had only 6 dimensional data characteristics by using kernel principal component analysis, and the cumulative contribution rate of the first three principal components had reached the purpose of dimension reduction. Therefore, in the analysis of kernel principal component analysis, feature dimension was reduced obviously, and the comprehensive features retained the features of the original data, which provided a theoretical grounding for the following partial discharge recognition by multi classification support vector machine.Fourthly,this paper presented a partial discharge recognition method, which was based on the optimized parameters and multi-classification SVM. In this method, grid search algorithm was adopted to realize the parameter optimization of SVM, and the new comprehensive features were used to act as the classification characteristic. We would find the optimal training SVM model by using the five fold cross validation method. Then, the binary classification of support vector machine was extended to multi-classification, which was used the M-ary classification. Finally, the test data would be input to the optimized parameters SVM for classifying. Compared with non-optimized SVM, this method is an effective and reliable for partial discharge pattern recognition, which realized higher recognition rate and computing speed.