Methods for wavelet-based autonomous discrimination of multiple partial discharge sources

Abstract

Recent years have seen increased interest in the application of on-line condition monitoring of medium voltage networks as the need to maintain and operate ageing cable networks increases. Detection and analysis of partial discharge (PD) activity is generally used as an indicator of pre-breakdown processes that may be indicative of insulation degradation over time. A significant challenge for on-line monitoring is discrimination between multiple partial discharge sources that will often naturally exist in the data. To discriminate between PD sources each PD signal is represented as a feature vector and a clustering algorithm is used to identify clusters in the resulting feature vector space, often after dimensional reduction. Each cluster identified in the data corresponds to a distinct PD source. In this work a comparison of clustering algorithms and dimensional reduction techniques is performed to identify clusters for a variety of PD data sets, in all cases the feature vector is created using wavelet decomposition energies. The three clustering algorithms used were Density Based Spatial Clustering of Applications with Noise (DBSCAN), Ordering Points to Identify Clustering Structure (OPTICS) and Simple Statistics-based Near Neighbour clustering technique (SSNN). The two dimensional reduction techniques considered were Principal Component Analysis (PCA) and t-Distributed Stochastic Neighbour Embedding (t-SNE). At present the most commonly used combination of dimensional reduction technique and clustering algorithm for PD data are PCA and DBSCAN respectively. From the comparison performed in this work it was found that t-SNE combined with OPTICS or SsNN were the most successful at clustering PD data. For use in practical situations SSNN is preferred over OPTICS as it requires only a single input parameter, which generally be hardcoded, leading to a completely autonomous technique. It is therefore suggested that a combination of t-SNE and SSNN is particularly appropriate for discriminating PD sources.