Enhancement of GIS Partial Discharge Pattern Recognition Accuracy Based on Domain Adversarial ConvNeXt

Partial discharge (PD) detection plays a critical role in assessing the insulation condition of gas-insulated switchgear (GIS), a critical component in modern high-voltage power systems. While recent advancements in machine learning have facilitated the development of various models for phase resolved partial discharge (PRPD)-based PD recognition in GIS, conventional approaches predominantly utilize datasets collected under identical experimental conditions. This methodology induces similar spatial characteristic, such as discharge density and sampling rate, between training and testing sets, resulting in high accuracy metrics with limited generalization capabilities. To address this issue, this paper proposes DACNeXt, a domain-adaptive convolutional network designed for PRPD pattern recognition in GIS. The proposed model enhances conventional convolutional networks through three critical modifications: optimized feature extraction layers for PRPD characteristics, large-kernel attention mechanisms for discharge pattern recognition, and adversarial training through domain discriminators. The experimental framework employs a 252kV GIS platform to generate multi-condition datasets, with preprocessing procedures ensuring domain shift simulation between source and target distributions. Compared to previously proposed models for PRPD pattern recognition including GrabCut-GoogleNet, CNN, LBP-SVM, and RNet, DACNeXt achieves average accuracy improvements of 13.5%, 5.4%, 11.5%, and 16.2%, respectively. The network effectively identifies critical defect types including floating potential discharge, corona discharge, and surface discharge. These advancements contribute to improved GIS PD pattern recognition and early fault detection, enhancing the reliability and safety of high-voltage power systems.