An Automated Identification Approach for Partial Discharge Detection Using Density-Based Clustering Without User Inputs

A recently published state-of-the-art density-based clustering technique called ICFSFDP for partial discharge (PD) detection requires various sensitive user-defined input parameters. This article presents a parameter-free clustering technique called PDAutoClust for PD detection. The PDAutoClust algorithm can produce high-quality clusters for PD datasets without requiring user-defined input parameters. PDAutoClust produces high-quality clustering results by utilizing a vein-based density clustering approach. The vein of a cluster is produced by using multivariate kernel density estimation and a unique neighborhood set. We compared the performance of PDAutoClust against ICFSFDP and seven other state-of-the-art density-based and non-density-based clustering techniques by using four PD datasets in terms of adjusted rand index, normalized mutual information, F1-score, and purity. Another contribution of the article is a novel merging technique used with PDAutoClust to merge small non-viable clusters that a clustering technique may produce. PDAutoClust produces the final clusters for a dataset by merging the non-viable clusters that a clustering technique may produce. We also evaluate the performance of PDAutoClust with merging technique versus PDAutoClust without merging technique using four datasets. Simulation results for PDAutoClust with the merging technique show good performance compared to ICFSFDP and seven other state-of-the-art clustering techniques. We also performed an ablation study to demonstrate the importance of the steps involved in PDAutoClust.