Noise signal can’ t be inevitable in the collected signals of rolling bearings which work in a harsh and complex environment. In order to realize rapid characteristic extraction and improve recognition rate, a convolutional neural network (CNN) fault diagnosis method based on principal component analysis (PCA) was proposed. In this method, PCA was introduced to preprocess the signals, and then the processed signals were converted into twodimensional feature images, which were input into CNN model to extract the transformed image features for fault pattern recognition and classification. Fault diagnosis experiment was carried out by using the rolling bearing data set from Case Western Reserve University. The results show that the proposed method is feasible and effective, and meets the application requirements of robustness and realtime.