In response to the problem of low accuracy in fault diagnosis of rolling bearing vibration signals using traditional convolutional neural network (CNN), a superior bearing fault diagnosis method that combined multi domain information with deep separation convolution (MDIDSC) was proposed. Firstly, the bearing vibration signal was decomposed using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm. Then, a multi-spatial state matrix was constructed using the decomposed intrinsic mode functions (IMF) components, and the multi-spatial state matrix was inputted into the proposed deep separation convolutional model for training the proposed network. At the same time, residual structures were added to the deep separation convolutional model to reuse the feature extracted by the proposed method, and the convolutional kernel was separated in the direction of depth, solving the problem of network degradation in the deep model. Finally, a spatial feature extraction method was proposed to prune the model parameters for improving the efficiency of the proposed method, and an adaptive learning rate annealing method was used to avoid the model falling into local optimization in the process of the gradient optimizing. The experimental results indicate that, through a lot of experimental comparisons between different bearing fault datasets, the proposed method exhibits more excellent performance and outstanding ability in recognizing the bearing fault with a maximum testing accuracy of 100%, and the mean accuracy of 99.07%. At the same time, the maximum and mean loss of the proposed method are respectively 0.1345 and 0.0841 in test dataset, which demonstrates the superiority of the proposed method in diagnosing bearing fault. [ABSTRACT FROM AUTHOR]