Gaussian Mixture Model-Based Clustering and Classification of Phase-Resolved Partial Discharge Patterns in Cast-Resin Dry-Type Transformer Insulation Materials

Abstract

Identifying the source of partial discharge (PD) in a high-voltage (HV) apparatus is crucial for establishing the factors to be considered during the production phase. Because repair is not feasible, identifying the PD source is even more important, particularly in cast-resin dry-type transformers (CRDT). The assessment of PD endurance for 11 distinct insulation materials utilized in CRDT manufacturing–epoxy resin, three-ply glued polyester, pre-impregnated insulation, glass fiber mesh, polyester adhesive tape, polyester film, polyimide film, rectangular conductors with polyester, round conductors with polyester, round conductors with enamel, and uninsulated foil conductors–seeks to pinpoint fault origins from 24 different phase-resolved PD (PRPD) patterns observed during experiments in a specifically engineered testing environment. The identification of fault origins involved a comprehensive analysis of PRPD pattern classifications by International Council on Large Electric Systems (CIGRE) and ABB, leading to categorization of the six identified types based on the relationships between PRPD patterns and PD sources. To identify fault sources from PRPD patterns, these correlations were considered, and the Gaussian Mixture Model (GMM) was employed for classification. The GMM was favored in this research for its proficiency in identifying PRPD patterns from various concurrent PD sources, its enhanced recognition of ellipsoid-like patterns, and its probabilistic clustering attributes. In the label-based assessment, the proposed program showed strong agreement with expert classification, achieving an accuracy of 94.17% and an F1 score of 88.52%. The findings suggest that the GMM-based methodology is a viable automated instrument for diagnosing PD sources in CRDT insulation materials.