Abstract: Aiming to realize edge computing on transformer UHF partial discharge monitoring, a method based on MobileNet is proposed. The various PRPD patterns collected by transformer UHF partial discharge monitoring systems are pre-processed and data-enhanced to form PRPD pattern datasets. The weights in the pre-trained MobileNetV1 model are transferred to the new task of partial discharge, and the network structure and weights of the model are fine-tuned. The new model after migration is trained, the model with the highest recognition rate is used as the test model, and the PRPD patterns in the test set are tested. The experiment shows that the recognition accuracy of the algorithm is more than 95%, and is not influenced by the phase shifting of PRPD.