Study On The Features And The Classification Accuracy Of Partial Discharge Considering Detection Distance

Partial discharge (PD) happens due to insulation defect in the insulation system of power equipment under the rated voltage, along with the development of PD, the insulation system is degraded and the residual life of insulation system is reduced.There are three common and effective methods to detect PD, including impulse current detection, ultrasound detection and ultra-high frequency (UHF) detection. UHF method is based on monitoring the electromagnetic waveform (EM) induced by PD in detect. Currently; UHF method is applied to classification of PD, diagnosis of residual life of insulation, location of PD, et.al. Recently, UHF has been used widely in GIS, while rarely used in transformer, and moreover, the features would change along with the propagation distance, so in this paper, based on the typical PD patterns in transformer, the variation of electromagnetic waveform features and the classification accuracy of partial discharge considering detection distances are researched.Firstly, by the meaning of studying the principle of UHF signals and the detection method, an experimental system considering detection distance is set up, and the method of feature selection is proposed, which can extract features in time-domain and frequency-domain from PD EM. Skewers and kurtosis of the shape of the EM wave and the energy vector in frequency domain, which is obtained by combing the wavelet packet decomposition and singular value decomposition, are proposed as features of EM wave induced by PD.Secondly, because the attenuation in different frequency components along with transporting distance is different, some features of EM wave will change monitored at different distance. In this thesis, the changing tendency of features, including features in time-domain and frequency-domain of EM wave and statistic parameters of EM PD pattern, along with monitoring distance has been analyzed, the result shows that the features are changed along with the monitoring distance. According to the relationship between statistical parameters of UHF signal and detection distance, multivariate nonlinear regression method is established, thus the location error is less than 40 cm.Finally, the classification accuracy of PD pattern along with monitoring distance has been analyzed. The result shows that if the PD classification model trains the model using the features obtained from the signals only in one monitoring distance, such as 40cm, 160cm and etal, the accuracy is very low, below 55%, while if the train features for classification model are obtained from the whole data monitored in different distance, so the accuracy cab be above 85%.So, in this paper, a very clear conclusion is proposed that if the classification result is content, the train features should be assembled from enough data monitored in different distance, because the features are changed along with the transporting distance.