Abstract: In order to fully express and utilize the connection of physical characteristics of GIS defects to guide fault diagnosis and improve the reliability and interpretability of GIS partial discharge fault diagnosis methods, we propose a neural supervision decision tree (NSDT) algorithm. This algorithm aims to achieve interpretable fault diagnosis under high accuracy of GIS partial discharge. The final linear layer of the convolutional neural network is replaced with a hierarchical structure, the induced layer is constructed, the node representation vector is optimized using the tree supervised loss function, and the hierarchical loss is suppressed by adjusting the softmax function. The supervised layer is constructed, and the output of the induced layer is supervised and corrected by the original neural network output vector to improve the recognition accuracy while retaining the interpretability. The performance of the NSDT method is tested by collecting partial discharge data of three typical single-source defects from the tip of the guide rod, the tip of the shell, and the high voltage end surface through the 110 kV multi-perception GIS platform, and three dual-source defects are composed. The test results show that the NSDT multi-level neural decision structure can accurately identify GIS partial discharge types, and it has higher identification reliability, richer decision information, and better interpretability compared with traditional deep neural network identification methods such as convolutional neural networks.