GIS partial discharge diagnosis method under sample imbalance conditions

• A frequency-domain image-based attention mechanism is proposed for extracting features from PRPS images. This method effectively captures peak and phase features of various partial discharges, offering a new approach for efficiently extracting different features from PRPS. • A new loss function, the logit adjustment class weight focal loss (LCF), has been designed to improve the model’s diagnostic accuracy in the case of imbalanced samples by combining logit adjustment, class weight, and focal loss. • Based on physical simulations, different partial discharge defect models were constructed, and a GIS partial discharge experimental platform was built. The models were tested using experimental signals, and an unbalanced experimental dataset was created based on the actual probability of partial discharge occurrences on-site. The experimental results validate the effectiveness of the proposed method in diagnosing partial discharge categories in GIS equipment. Gas-insulated switchgear (GIS) partial discharge (PD) diagnosis is essential for ensuring the stable operation of modern power systems. Existing PD diagnostic methods are predominantly based on balanced datasets and still have room for improvement in feature extraction and recognition speed. This study proposes a PD diagnostic model that integrates frequency channel attention with the LeViT architecture and trained using the logit adjustment class weight focal loss function (LCF) to enhance the diagnostic performance of GIS PD spectrograms. Firstly, the phase-resolved partial discharge (PRPS) of GIS extracts channel features from low-frequency components through the frequency efficient channel attention (FECA) mechanism. Next, the extracted feature maps are fed into the LeViT model and trained using the LCF loss function to classify PD types. The experimental results demonstrate that the LeViT-FECA-LCF model achieves a recognition accuracy of 99.5%, showing improvements of 1.0%, 1.5%, 1.5%, and 2.5% over EfficientNet-B0, MobileNetV3, ShuffleNetV2, and LeViT, respectively. These improvements are achieved while maintaining a relatively low inference time and computational cost. Ablation studies further confirm that each module contributes positively to the overall performance of the model. Compared with existing methods, the proposed model exhibits superior recognition accuracy. Furthermore, this approach effectively extracts PD PRPS spectrogram features, improving PD identification accuracy and providing new insights for GIS equipment health management.