Aiming at the unclear characteristic information of single-phase grounding fault in distribution network and the existing line selection methods are easy to be affected by fault conditions and environmental noise, according to the transient fault characteristics and steady-state fault characteristics of distribution network, an improved deep neural network is proposed for fault line selection. The loss function and learning rate of the deep learning network are optimized to further improve the efficiency and accuracy of line selection. The feasibility of this method is verified by simulation. The results show that the number of training iterations after the improvement is reduced from 86 to 30, the training efficiency is improved by 65.12%, and the accuracy of fault judgment is improved from 95% to 99%, which has good anti-interference ability and certain practical value.