Abstract: Aimed at the problems of complicated bearing working environment and low diagnostic performance under variable working conditions, a bearing fault diagnosis method based on improved convolutional capsule network is proposed. The Inception structure and the channel and space double attention modules were used to replace the single-layer convolution kernel structure in the capsule network, and the multi-scale key information acquisition of data was performed. Through the structure of the capsule network, vector neurons were constructed. Under the feature transfer mode of the dynamic routing algorithm, combined with the optimized loss function, the fault diagnosis was completed. In order to verify the diagnostic effect of the model, experiments were carried out on the bearing data set of Case Western Reserve University under single and variable working conditions. The results analysis show that this method can effectively diagnose faults.