Classification of Partial Discharge Sources in Oil-Pressboard Insulation Through a Statistical Approach Based on Detrended Fluctuation Analysis

• Published in: IEEE transactions on plasma science Vol. 53; no. 5; pp. 1037 - 1045
• Main Authors: Pradeep, Lavanya, Haque, Nasirul, Preetha, P.
• Format: Journal Article
• Language: English
• Published: IEEE 01.05.2025
• Subjects: Corona; Density-based clustering; detrended fluctuation analysis (DFA); Discharges (electric); Electrodes; Fault location; high-frequency current transformer (HFCT); high-voltage insulation; Insulation; median; Oil insulation; Oils; partial discharge (PD); Partial discharges; Power transformer insulation; skewness; Surface discharges
• ISSN: 0093-3813; 1939-9375
• DOI: 10.1109/TPS.2025.3552077

This article introduces a novel approach to characterize the partial discharge (PD) sources present in the oil-pressboard insulation depending on the origin and type of PD sources. Three different types of PD sources, including internal void, corona, and surface discharges, were created using oil-pressboard insulation samples and appropriate electrode setups in the laboratory. PD measurements were performed using a high-frequency current transformer (HFCT) and an oscilloscope with different combinations of the aforementioned defects kept in parallel. PD pulses were extracted from the measured data, and a well-known signal processing algorithm, detrended fluctuation analysis (DFA), was implemented on these pulses. Two statistical features associated with the DFA processed signals, median and skewness, were calculated, and it was found that the PD pulses coming from different sources formed well-defined clusters in the 2-D scatter plot of median versus skewness. The clustering was performed using a clustering algorithm called density-based spatial clustering application with noise (DBSCAN). Later, PD sources were identified by comparing the phase-resolved PD (PRPD) patterns corresponding to each cluster with those obtained under single-defect cases. The developed methodology is easy to implement, does not require complex optimization techniques, and classified PD signals with an accuracy in the range of 93%-97%.