In this paper the research on transformer（partial discharge PD） pattern recognition work is carried out.The problem of the statistical feature quantity for PD pattern recognition depends on the expert experience and lacks certain generalization.The CNN convolutional neural network mode pattern recognition algorithm lacks the time series feature information of PD map.The PD PRPD pattern pattern recognition algorithm of CNN-LSTM which combines the advantages of CNN network to exploit the local spatial information of PRPD map and the advantages of LSTM long and short memory network to exploit the time series feature information of PD map is proposed.The feature vector output by CNN pooling layer is used as the input of LSTM.The spatial characteristics and timing characteristics are model-recognized by Softmax classifier.The performance of CNN-LSTM is compared with CNN network and LSTM network CNN and CNN-LSTM have 100% recognition ability for PD spectrum stable suspension potential defects but CNN-LSTM has better recognition ability than CNN for metal protrusion defects oil paper air gap defects and creeping discharge defects which shows that the overall recognition performance of CNN-LSTM network is better than CNN and LSTM.