A bearing fault diagnosis model based on CWT-CNN was proposed to solve the problems of slow training convergence speed, low recognition accuracy and poor anti-noise performance of traditional fault diagnosis methods for rolling bearings. The data set was expanded by three vertical random clippings of the time-frequency graph generated by continuous wavelet transform of rolling bearing vibration data, then it was imported into the constructed convolutional neural network with batch normalization and random inactivation for model training, and the trained model was used to realize bearing fault classification. To test the model performance, the Case Western Reserve University bearing dataset was used for detection. The experimental results show that compared with the data set constructed by the conventional method, the data set constructed by the proposed method has a faster convergence speed in the training of the constructed convolutional neural network, the performance of the trained model is also more stable, and the final test is accurate. The final highest test accuracy of this data set is 99.75%, and the accuracy rate of the dataset constructed by the conventional method is 99.67%, which proves the feasibility of the method in constructing the dataset. After adding white Gaussian noise with a signal-to-noise ratio of 6dB to the original data, the highest accuracy rate of the data set constructed by the conventional method still reaches 98.67%, showing the strong anti-noise ability of the bearing fault diagnosis model based on CWT-CNN, which proves the effectiveness and feasibility of the proposed method.