Partial Discharge Pattern Classification in High-Voltage Equipment: A Probabilistic Diagnostic Approach

Partial discharge (PD) is a major cause of insulation aging and failure in high-voltage equipment, such as gas-insulated switchgears and transformers. Although phase-resolved partial discharge (PRPD) patterns are widely used for diagnosis, their variability and noise sensitivity limit the reliability of conventional analysis methods. In addition, in many PD diagnostic settings, the availability of reliably labeled PRPD data is limited, which makes robust classification under small-sample conditions particularly challenging. In this study, we propose a probabilistic PRPD classification framework with four distinct contributions: 1) a frequency-weighted multivariate kernel density estimation (KDE) that incorporates discharge-event counts directly into the density estimate; 2) an empirical correlation analysis that quantifies the complementarity among Wasserstein distance (WD), Kullback–Leibler (KL) divergence, and maximum mean discrepancy (MMD); 3) a three-stage hierarchical ensemble classifier that integrates these complementary measures through majority voting, pairwise agreement, and normalized composite scoring; and 4) dual-regime validation on both laboratory and in-service PRPD data. The proposed method was validated using 405 PRPD patterns from four representative defect scenarios in an IEC 60270-compliant test cell. For the 60 held-out target samples evaluated under the present test conditions, the proposed method achieved a macro-F1 score of 1.00 under both large- and small-sample regimes, with no observed misclassification, whereas an image-based CNN benchmark showed degraded stability in the small-sample regime. Applicability to in-service equipment was further assessed using 21 field PRPD datasets collected from operating high-voltage apparatus, where defect-specific separability was preserved under field noise conditions. Correlation analysis showed that WD and KL divergence captured global distributional characteristics, whereas MMD was more sensitive to local variations and noise, which supports their complementarity. The proposed framework separates reference-distribution construction from per-target classification and requires no model retraining for new measurements, so that the per-target computation does not depend directly on the size of the accumulated dataset, supporting its applicability to online monitoring. The results demonstrate that the proposed approach provides a statistically robust framework for PRPD pattern classification and is particularly effective for datasets of limited or moderate size commonly encountered in PD diagnostic and classification studies.