Multi-Task Learning Network-Based Condition Assessment of GIS Partial Discharge: Diagnosis and Severity Evaluation

Current methods for assessing the state of partial discharge (PD) in gas-insulated switchgear (GIS) often treat the diagnosis and severity evaluation tasks independently, which limits model performance and overlooks the role of defect types in the PD development stages. To address these issues, this paper proposes a novel multi-task network for the condition assessment of PD in GIS. Firstly, a lightweight attention combined with an expert network module is introduced to explore the correlations between different tasks while retaining their unique features. Secondly, we incorporate a convolutional neural network (CNN)-Transformer module to capture both local and global features of GIS PD signals. Finally, we implement a dynamic weight averaging method for the multi-task network, which adaptively adjusts the loss weights via the variation rate of the losses of each sub-task. We conducted experiments on two GIS PD datasets with different insulating gases. Experimental results demonstrate that the multi-task network significantly improves the performance of each sub-task in GIS condition assessment, advancing the development stage evaluation of PD for various insulation defects. Furthermore, experiments on two types of insulating gases validate the generalization and versatility of the proposed method.