Artificial Learning Algorithms for Partial Discharge Classification

Abstract

Partial discharges (PDs) are localized dielectric breakdowns that may occur within insulating systems under high electric stress, without leading to total failure. Even if, in alternating current (ac) systems, the PD recognition is a fully explored field, accurate identification of PD types under direct current (dc) conditions remains an open problem due to the complexity and variability of the discharge phenomena. In this study, a data-driven classification framework is proposed, employing artificial learning (AL) techniques to distinguish experimentally acquired PD signals corresponding to three typical defect types. The approach compares conventional machine learning (ML) models—such as decision trees (DTs) and discriminant analysis (DA)—with a deep neural architecture based on long short-term memory (LSTM) networks capable of capturing temporal patterns and nonlinear dynamics in PD waveforms. All models were trained and tested on a large dataset of experimental measurements, with hyperparameter optimization used to enhance classification accuracy while preserving computational efficiency. The experimental results demonstrate the potential of AL-based methods also for reliable PD recognition in dc, opening perspectives for condition monitoring, failure prevention, and intelligent diagnostics in high-voltage insulation systems.