Abstract: In high-voltage substation, the accuracy of gas insulated switchgear (GIS) partial discharge diagnosis is restricted by the number of labeled samples. In order to solve the problem that unlabeled data cannot be used in conventional partial discharge diagnosis methods, and the difference between training samples and samples to be tested cannot be overcome, a GIS partial discharge diagnosis method based on deformable convolution and self-supervised contrastive learning is proposed in this paper. First, the feature extraction network is trained by comparing the similarity and difference between the unlabeled data samples to obtain the feature representation of the input data. Then, the classifier is trained by the labeled data to learn the defect categories represented by the features of different partial discharge data. Finally, the samples to be tested are input into the model to achieve partial discharge diagnosis. In order to further improve the perception ability of the model in the process of feature extraction, a deformable convolutional neural network and a spatial transform module are introduced to enhance the adaptability of the convolutional check feature map. The results show that self-supervised contrastive learning can make full use of unlabeled data to achieve efficient feature capture. In the case of insufficient labeled data, the model pre-trained by unlabeled data can improve the PD diagnosis accuracy by 9.34% on average. The self-supervised contrastive learning method proposed in this paper can provide a new solution for the partial discharge diagnosis.