Aiming at the problems of poor anti-noise performance and high computational complexity in the traditional fault diagno? sis method based on deep learning， a rolling bearing fault diagnosis method based on optimized deep residual networks （ResNet） is proposed. The one-dimensional vibration of the rolling bearing is transformed into frequency domain by fast Fourier transform （FFT）. The DS-ResNet model is constructed by replacing the standard convolution layer of the traditional deep residual network （ResNet） with the depthwise separable convolution （DSC） which has low computational complexity and the pointwise convolution （PWC） which can enhance the nonlinear expression of the network. The frequency domain signals in various fault states are used as the input of DS-ResNet model for identification and classification. The results show that the recognition accuracy reaches 92.71% in a strong noise environment with a signal-to-noise ratio of ?4 dB， and the average recognition accuracy reaches 90.19% under variable speed conditions， which is higher than other commonly used deep learning diagnosis methods， and the training time of the model takes only 2.16 s per round， which proves that the proposed method has better noise immunity， generalization perfor? mance and faster diagnosis efficiency.