Detection Transformer-Based Deep Learning for Multisource Partial Discharge Recognition in High-Voltage Rotating Machine Insulation

Abstract

Partial discharge (PD) is a critical early indicator of insulation degradation in high-voltage rotating machines. Minor undetected defects during manufacturing, combined with operational stresses, often trigger PD activity that accelerates insulation aging. The elongated structure of stator windings increases the possibility of multiple PD sources occurring simultaneously, each generating a distinct phase-resolved PD (PRPD) pattern. However, when multiple defects are present, these patterns overlap, leading to challenges such as occlusion, nesting, and the multisource PD horizontal flattening effect (MPD-HFE), where weaker discharges are visually suppressed, which challenge conventional classifiers due to signal occlusion and distortion effects. This work proposes a detection transformer (DETR)-based deep learning approach, enhanced with flattening augmentation to replicate dominant pattern distortions. The method enables simultaneous detection and classification of overlapping PRPD patterns by utilizing DETR’s global self-attention to mitigate discharge dominance and feature suppression. Evaluated on a multisource PRPD dataset comprising six defect types and 63-class combinations, the proposed model achieved 97.4% accuracy on single-source PD defects and 90.9% on multisource cases, with an overall accuracy of 93.7%.