Fault describes the boundary position of the stratum, thus the discontinuity of the reflection layer in seismic image can be used as the main basis for fault interpretation. The strong nonlinear nature of deep neural networks can be used as a powerful tool to express the discontinuous features of seismic data. Fault interpretation can be regarded as a pixel-wise binary classification problem, and deep learning methods are used to model and solve the problem. An end-to-end three-dimensional fault automatic identification method is presented based on a deep learning network. Firstly, multiple sets of 3D seismic volume are synthesized by convolution of wavelets and reflection coefficients for deep network to learn fault characteristics. Then a network is built for training, and applied to field seismic data after the network training is completed. Due to the residual module can improve the generalization performance of the network, the proposed method to incorporates the residual block structure into the U-Net to improve the network model's fault identification performance on field data. The input of the trained network is the post-stack 3D seismic data, and the output is 3D data volume with same dimension, where each output value is the probability of the fault at the corresponding position of the input 3D seismic data. Field data example tests show that this method can effectively identify faults in seismic data, and at the same time further improve the generalization performance on field data.