A new method for automatic Multiple Partial Discharge Classification

A new wavelet based feature parameter have been developed to represent the characteristics of PD activities, i.e. the wavelet decomposition energy of PD pulses measured from non-conventional ultra wide bandwidth PD sensors such as capacitive couplers (CC) or high frequency current transformers (HFCT). The generated feature vectors can contain different dimensions depending on the length of recorded pulses. These high dimensional feature vectors can then be processed using Principal Component Analysis (PCA) to map the data into a three dimensional space whilst the first three most significant components representing the feature vector are preserved. In the three dimensional mapped space, an automatic Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is then applied to classify the data cluster(s) produced by the PCA. As the procedure is undertaken in a three dimensional space, the obtained clustering results can be easily assessed. The classified PD sub-data sets are then reconstructed in the time domain as phase-resolved patterns to facilitate PD source type identification. The proposed approach has been successfully applied to PD data measured from electrical machines and power cables where measurements were undertaken in different laboratories