The traditional feature extraction methods of partial discharge of Gas Insulated Switchgear (GIS) has the disadvantages of relying on expert experience, high blindness and low recognition accuracy. This paper will convert PRPD data of partial discharge into gray-scale maps, the identify features of which were extracted by the convolutional neural network with the powerful adaptive feature extraction ability. The extracted features were applied to classical classifiers such as SVM, random forest, and BP neural network, to realize the effective integration of deep learning methods and traditional machine learning methods. Experiment results show that the features extracted by this method have higher differentiation degrees, which can effectively improve the accuracy of partial discharge pattern recognition.