Recognition of partial discharge patterns

Abstract

This paper aims to provide a robust data mining framework for partial discharge (PD) pattern recognition, specifically to classify the PD signals based on their shapes. The framework contains feature extraction (FE), feature selection (FS), unsupervised clustering analysis and clustering result validation. In the process of FE, Principal Component Analysis (PCA) is shown to be the suitable dimension reduction technique by extracting the majority of the variation in the original data sets. We show that singular value decomposition (SVD) can provide additional insight to understand the results of PCA which are often difficult to interpret. By comparing the patterns of the PD pulses and the Normalised Autocorrelation Functions (NACFs) of the pulses after applying SPCA, the PD pulses are chosen to be the features for cluster analysis. In the process of cluster analysis, the need for cluster validation in unsupervised learning is discussed. Experimental results provide evidence that using several indexes gives greater confidence in choosing the appropriate unsupervised clustering algorithm and determining the correct number of clusters.