Rolling bearing is an important part of high-tech mechanical equipment and also one of the important fault sources. At present, few bearing fault samples, uneven data distribution and unstable effect of traditional bearing fault identification methods bring great difficulties to fault identification technology. A generalized S-transform method was proposed by combining deep learning correlation technology with bearing fault diagnosis technology, and taking advantage of deep learning model to recognize two-dimensional images. Generalized S-transform is the inheritance and development of wavelet transform and short-time Fourier transform. By transforming one-dimensional bearing fault signal data into two-dimensional time-frequency diagram, the model of Xception network is fine-tuned and the hyperparameters are optimized, and then the two-dimensional time-frequency diagram is input into the improved Xception network to carry out transfer learning. The above experiments were carried out based on rolling bearing data published by Case Western Reserve University, and the recognition rate of fault signals under diff ____________________________________________________________________________________________ *基金项目：国家自然科学基金项目( 61372128)，教育部协同育人项目(202002179030)，南京信息工程大学滨江学院科研与教研项目（2020yng001,JGZDI201902） erent working conditions reached 99.95%. The experimental results prove that the recognition method based on generalized S-transform and transfer learning is real and effective.