Fault recognition from seismic data is crucial to the seismic data interpretation，but with the expansion of exploration scale，the traditional artificial fault interpretation cannot meet the actual production needs. How to develop a high-precision fault recognition method and improve the operation speed of the method is an urgent problem for those skilled in the art. Therefore，an automatic fault recognition method based on the improved AlexNet model is proposed to treat fault recognition as binary classification of image recognition. First，instead of local response normalization（LRN），batch normalization is used to accelerate the model convergence. Then，the balanced cross entropy loss is introduced to solve the problem of unbalanced height between the fault and the non-fault in seismic data，which makes the model converge in the right direction. Finally，the convolution layer is adopted to replace the full connection layer，which greatly reduces the training parameters and speeds up the training. The prediction results of the theoretical data and actual data of the training model show that the improved AlexNet model fully learns the fault features and has the ability to identify faults from seismic data.