Abstract: Various discharge identification models based on local discharge signal sampling data developed on PC computer or server side with high positive judgment rate are almost all deep models, which are large in volume and raise higher requirements for computer resources, and cannot be run on the local discharge detection device. For this reason, this paper proposes a reinforcement learning-based automatic pruning method for partial discharge depth diagnostic models with a model lightweight deployment scheme. On the server side, deep reinforcement learning is used for intelligent body training, and automatic search is performed by interacting with the original partial discharge diagnostic model, which in turn determines the pruning rate of each layer; and then the filter pruning according to the filter pruning by geometric median (FPGM) method is used to discriminate the importance of the filter and realize parameter pruning. Simulation experimental results show that the method achieves more than 85% parameter compression effect on the lightweight series models MobileNetV1 and V2 as well as the classic series neural networks of ResNet50. The compressed lightweight model is converted into lightweight ONNX format, saved on a portable computer, and implanted into a Raspberry Pi smart terminal through wireless transmission, which can realize the diagnostic experimental simulation of partial discharge on the smart terminal. The test results show that the local discharge diagnostic model which is deployed after pruning using this method has great improvements in performance indexes such as memory occupation, power consumption, and inference time.