Study Of Denoising White Noise And Pattern Recognition In Online PD Detection Of Middle Voltage Cables

Partial discharge is not only one of most important reason but also manifestation of deterioration of the cable insulation condition. The PD online detection of cables is meaningful to ensure the reliable operation of power systems. This paper carries out the related technology research based on current researches of PD online detection at home and abroad.At the present stage, emphsis and technical problems in online detection are mainly concentrated in de-noising and pattern recognition. In term of interference suppression, this paper mainly studys the wavelet thresholding method for removing white noise.In wavelet thresholding method, the distortion of de-noised signal relates closely with the selection of basic wavelet and threshold. In view of basic wavelet, based on the principle of maximum energy of scale coefficients, the optimal basic wavelet of eash are adaptively selected for decomposion and de-noising. The de-noising result of 4 type of partial discharge signal shows that this method has better de-noising effect than single wavelet de-noising method.In view of thresholding, the signal is decomposed by lifting wavelet and the threshold of each scale is calculated based on the wavelet entropy. Then the wavelet coefficients of each scale are processed by thresholding. Finally the processed wavelet coefficients and scale coefficients are used for reconstructiong to get the de-noised signal. The de-noising result of simulation and measurement signal show that the mehod can suppress the white noise effectively.The detected online PD may come not only from the cable and its joints, termination, but also from the switchgear connected with it. Different PD source is with different consequence and has its own criteria. Therefore it is necessary to recognize different PDs. Based on wavelet decomposition, the eigenvector of partial discharge signal is composed by the energy percentage of wavelet coefficents and scale coefficients. Then the eigenvector is respectively input into BP neural nework and support vector machine classifer for recoginition and the recoginition results show that the feature extraction method based on energy percentage can recognize the partial discharge sgnals accurately and reliably.