Intelligent GIS Partial Discharge Diagnosis with Limited Samples via Conditional Data Augmentation and Modern CNNs

Abstract

To address the challenges of defect sample scarcity and reliance on manual feature extraction in partial discharge (PD) diagnosis of Gas-Insulated Switchgear (GIS), this study proposes a small-sample intelligent diagnostic method that integrates a Conditional Generative Adversarial Network (CGAN) with the ConvNext architecture. Under conditional constraints, the CGAN is employed to perform data augmentation on PRPD spectrograms, effectively mitigating issues of sample imbalance and overfitting. Subsequently, the ConvNext network is introduced to extract multi-scale texture and morphological features from PRPD data, enabling efficient end-to-end classification. The model’s performance is validated using a 220 kV GIS typical discharge dataset, and the influence of key architectural parameters on diagnostic performance is further investigated. Experimental results show that the proposed method achieves remarkable improvements in small-sample classification accuracy — with data augmentation enhancing accuracy by more than 5%, and the optimal configuration achieving diagnostic precision exceeding 97% — demonstrating strong potential for practical engineering applications.