Accurate and fast identification of fault types and fault feeders in distribution network is helpful to improve the reliability of power supply. The fault information utilization rate of fault identification and line selection is low, and the ability of classifier to mine deep fault features is insufficient. In this paper, the idea of multi-label and multi- classification is put forward, and the improved ResNet parallel training of two branches is built to realize the grounding fault identification and fault line selection of distribution network at the same time. Wavelet analysis is introduced to decompose all kinds of electrical quantities and construct time-frequency matrix. The primary feature matrix of time-frequency matrix is extracted by frequency division and time division as network input, a multi-branch residual unit structure is designed, which is suitable for the accurate and fast identification of fault types and fault feeders. This unit structure is connected end to end, and two branches ResNet are constructed to realize fault identification and line selection of distribution network at the same time. Analysis of the simulation experiment results shows that, compared with the MLP network and the original ResNet, the improved ResNet is used to complete the fault identification and line selection of the distribution network. The index evaluation results are better, and the proposed method can be verified to have stronger adaptability and fault tolerance.