Abstract: To solve the problem of low identification accuracy of partial discharge（PD）for electric multiple

units ethylene propylene rubber（EPR）cable terminal defects，an on-board cable terminal defect identification

method based on improved residual network was proposed. Firstly，cable terminals with four typical insulation

defects were made，and the PD detection platform was built to obtain PD signals of different defect states and

establish data sets. Then feature transformation was used to transform PD one-dimensional time series signals into

2D topological feature images to enhance the discriminability of defect categories. Finally，an attention mechanism

was added to ResNet101 model of residual network，and Center and Softmax loss function were combined for

training and recognition classification to further improve accuracy. The test results show that the identification

accuracy of the proposed diagnosis method for PD at cable terminal is 97.3%. Compared with other traditional

defect diagnosis methods，the model has higher identification accuracy and better balance.

      Key words: ethylene propylene rubber（EPR）cable terminal；insulation defects；Gramian angular field（GAF）；residual network；fault diagnosis