The intelligent classification algorithm is applied well in partial discharge pattern recognition, but it needs to extract features manually, so there are problems of feature loss and low recognition efficiency. In this paper, the traditional convolutional neural network was improved by multi-layer feature fusion, and it was used for partial discharge pattern recognition. The pre-processed PRPD map was used as input to automatically extract the map features and performed deep and shallow features fusion to prevent features Lost, and finally output the classification result.. In addition, the algorithm also improved the pooling strategy of traditional CNN, and used the maximum two-mean pooling to further preserve the effective features of the graph. The experimental results show that compared with the traditional mode of extracting statistical features manually and then input into the classifier, feature fusion CNN has higher recognition accuracy and less time.