Modified Hierarchical Clustering Algorithm for Partial Discharge Separation

Abstract

To date, one of the main tools for evaluating the reliability of an insulation system is the continuous monitoring of those phenomena which, by interacting with the elements of the system, can induce aging processes or failures. For power grids, a signal that identifies possible aging or improper use of the component is Partial Discharge (PD) activity. Generally, the evaluation of the PD phenomenon is carried out through a two-step procedure: measurement and data analysis. To optimize the PD analysis process, increasingly sophisticated PD separation/classification algorithms are needed. Especially for the measurements carried out in HVDC systems for which the absence of a phase reference makes more difficult to identify the different types of discharge. The purpose of this article is to investigate the possibility of optimizing the input data to a hierarchical clustering algorithm in order to obtain a subdivision of the dataset more faithful to the real behavior of the phenomena. Specifically, the proposed approach is based on the use of the cross-correlation matrix to carry out the clustering operation. This matrix replaces the matrix of the distances among the points distributed in the map used for the representation of the data. Results show that with this modification it is possible to separate phenomena that present partially or completely overlapping patterns. Moreover, the algorithm turns out to be automatic and does not require the choice of references or thresholds to define the similarity among pulses.