•  
  •  
 

Keywords

partial discharge; feature map; convolutional neural network; multi-attention mechanism; defect identification

Abstract

The partial discharge (PD) phenomenon is an early indicator of insulation degradation in power equipment, and its frequent occurrence leads to severe failures. Traditional PD detection methods are susceptible to noise interference in complex environments, and their detection accuracy is low. To solve the problem of accurately identifying the PD of typical defects in cable joints, an identification method based on a feature map and a multi-attention mechanism with a convolutional neural network (CNN) is proposed. First, a PD feature matrix map is established, and the features of the time domain, spatial domain, and channel domain are organically combined and visualized, which enables the model to more comprehensively capture the multidimensional information of PD signals. Second, an improved CNN model is constructed by combining the channel attention (CA), spatial attention (SA), and temporal attention (TA) mechanisms. This effectively enhances the perception ability of the model for key features and improves detection accuracy and robustness. Finally, simulation experiments are conducted to analyze the accuracy and effectiveness of the model, and it is compared with other models. The research results indicate that the comprehensive accuracy of the model in PD defect identification reaches 98.89%. The effectiveness of the multi-attention mechanism is verified in an ablation experiment. After the temporal attention mechanism is removed, the identification accuracies of the surface discharge and air gap discharge categories decrease to 97.09% and 91.28%, respectively. Compared with the BP neural network, support vector machine, and random forest models, the performance of this model is more outstanding in all aspects.

DOI

10.19781/j.issn.1673-9140.2026.02.029

First Page

325

Last Page

337

Download

Share

COinS