At present， many rolling bearing fault diagnosis methods based on convolutional networks have the disadvantages of poor diagnosis effect and poor generalization ability under the influence of noise signals and load variations. Aiming at these prob? lems， an improved convolutional capsule network fault diagnosis method of rolling bearing under variable operating conditions is proposed. This method designs a multi-scale asymmetric convolution module， in which asymmetric convolution layers of different scales to extract features from the input data to maximize the extraction of feature information in the data and reduce the number of parameters effectively. In this module， the channel attention mechanism is introduced to better extract useful channel features and improve the feature extraction ability of the method in this paper. By improving the fully connected layer in the network to the fully connected layer of the capsule， the capsule can avoid the loss of characteristic information in the space in the process of outputting vector feature information. Case Western Reserve University bearing dataset and Southeast University gearbox dataset are used to verify the diagnostic performance of the proposed method and compare with other deep learning methods. The experimental results show that the proposed method has a better generalization and performance.