Granular computing, as an emerging soft computing classification method, provides a theoretical framework for solving complex classification problems based on information granulation and is one of the core technologies for simulating human thinking and solving complex classification problems in the current computational intelligence field. In this paper, we propose a design method of bearing fault diagnosis model based on granular computing: Convolutional Neural Networks-Granular Computing (CNN-GC). The method consists of two main components: fault features extraction and fault types determination. In this case, the bearing fault features are extracted using a convolutional neural network (CNN) with hyperparameter optimization to obtain bearing fault features with different output dimensions; fault types determination is obtained by using the extracted fault features as the input of hypersphere information granule based on granular computing. Compared with existing bearing fault diagnosis models, the CNN-GC model proposed in this paper, which accomplishes the conversion from numerical space to grain space, can obtain more accurate values and better grain size results. The superiority of the CNN-GC model in terms of accuracy and interpretability was demonstrated by the Case Western Reserve University(CWRU) bearing dataset.The experimental results show an accuracy rate of 99.8\(\%\).