In the field of geophysical exploration, fault identification is an important part. However, Due to the fixed identification mode, conventional computing methods cannot adapt to complex fault characteristics and environmental noise; And, the convolution neural network is prone to loss of global information because the slow convergence of global information. For those reasons, the previous fault identification methods cannot achieve higher resolution and accuracy fault identification, such as the identification results are very coarse, the identification results are inaccurate and results are discontinuous, especially in the case of severe environmental noise and complex fault interlacing. So in order to solve these problems, we proposed a deep learning network based on Transformer with improved window mechanism. Starting from the features of the data itself, quickly gather global information, and results on field data show that our network effectively solve the above problems. This work has positive significance for high-precision exploration work.