Digital detection, grouping and classification of partial discharge signals at DC voltage

In this paper, a digital dc PD pulse detection system with bandwidth of 10 kHz - 40 MHz is introduced, which was developed using some artificial intelligence methodologies. Focus is made on digital detection, grouping and classification of random pulse signals generated by PD phenomena at dc voltage. Digital detection is developed only resorting to a band-pass filter, a high-speed digitizer (100 MS/s) and a PC with data processing software. Grouping is realized with feature extraction of pulse waveshapes using equivalent time-frequency method (ETFM), making the 2D parameters plane or 3D parameters space, then using the unsupervised clustering Fuzzy C-Means (FCM) method to achieve fast separation for pulse sequence. And classification resorts to least square support vector machine (LS-SVM) based on a fingerprint, which is derivate form 2D histograms of basic parameters, the discharge magnitude q and the time between discharges Deltat of each sub-group. Field application is made for typical defects of oil-paper insulation under dc voltage. At last, several methods to improve separability of the grouping technique are also given for some special cases, including threshold value grouping, marginal coordinates grouping based on 2D parameters plane and grouping using ETFM preprocessed by wavelet denosing. Experimental results show that the dc PD detection system developed with artificial intelligence methodologies is practical and effective.