Reseach On Recognition Of Oil-Paper Insulation Thermal Aging Stages Based On Feature Of Partial Discharge

Oil-paper insulation is one of the most important forms in oil immersed transformers and thermal aging is the main reason for the deterioration caused by oil-paper insulation. By signals of partial discharge contains a wealth of information on insulation aging, this paper focuses on features which used for characterize the degree of thermal aging of oil-paper insulation from the partial discharge signal, and focuses on condition recognition methods which can recognition thermal aging stage of oil-paper insulation. The researches lay a foundation for condition monitoring and life prediction and evaluation of transformer.Firstly, according to the CIGRE Method II, an experimental model was devised to stimulate the flat cavity inside oil-impregnated insulation paper, and design its thermal accelerated aging experiments in the high-pressure laboratory. Based on method of pulse current, collected discharge signals of samples at 7 thermal aging stages in 0-21 days. By measured the degree of polymerization at different thermal aging stages to verify the rationality of the node division in aging stages. Filter the noise of measured partial discharge signals by using db8 wavelet soft threshold.Secondly, constructs two kinds of partial discharge features which contain aging information. One is structure three-dimensional statistical feature matrix, then SVD the matrix and filter out first 15 odd bits singular value as the φ-q-n 3D Pattern-SVD vector. Another is EMD the partial discharge signals and relies on the correlation coefficient screened 14 IMF components to structure features matrix and SVD the matrix get 14 singular values as feature of EMD-SVD. Show variation of two type features with a deeper level of thermal aging by figures and tables.Finally, with the above two kind of features which used for characterize the degree of thermal aging compared pattern recognition capabilities of Support vector machine(SVM) and random forest. Classification results show that the Random Forest’s recognition accuracy is higher than SVM. Based on this, used random forest as the classifier and recognize two kinds of signal features of oil-impregnated insulation paper in 7 thermal aging stages of 0-21 days. The result shown that the aging stage recognition based on φ-q-n Pattern-SVD features the recognition accuracy is just about 60% with a relatively low stability, and recognition accuracy of 90% with a strong stability based on EMD-SVD features. It is concluded that EMD-SVD feature of PD signals has higher discrimination than feature of φ-q-n Pattern-SVD wherein oil-paper insulation thermal aging stage recognition.