In view of the high equipment cost and long training time required by common deep learning methods for fault diagnosis, this paper proposes a bearing fault classification method based on the Inception-ResNet model. By using the Inception network's parallel structure, the network learns features of different scales, resizing structures are introduced to reduce degradation caused by network deepening, and three-dimensional convolution is added to allow information between different channels to blend. In order to verify the performance of this method, case Western Reserve University data set and IMS data set were used for verification, and compared with the traditional shallow learning method and deep learning method, experiments were conducted. The results show that, compared with other methods, the method presented in this paper not only has excellent diagnostic ability, but also is better in terms of resource utilization and training efficiency.