In order to realize a fault diagnosis for rolling bearings, a visualization of signals combined with proper orthogonal decomposition (POD) and continuous wavelet transform (CWT) were proposed. Thus, the problem of fault diagnosis is changed into a pattern recognition of images. With an establishment of Convolutional Neural Network (CNN) model, the fault diagnosis is completed and the main contents are as follows. Firstly, various vibration signals of faults were converted into a series of images to be distinguished with Continuous Wavelet Transform (CWT), and Proper Orthogonal Decomposition (POD) is applied to make an analysis and filter. The main principles of images were retained and reconstructed. Secondly, a CNN model is established to realize the pattern recognition of reconstructed ones. A model with two layers of convolutions and two layers of pooling was established. Then the CWT-POD-CNN model is used in two experiments of fault diagnosis successfully. Thirdly, a series of methods was used to make a comparison to prove the advantage of given work. Radial Basis Function (RBF), Back Propagation Neural Network (BPNN) and Support Vector Machines (SVM) were all combined with the same reconstructed samples. Furthermore, the CWT-CNN and CWT-POD-CNN models were also compared. On the data of rolling bearings on internet, an accuracy near to 100% was obtained in both models. As to rolling bearings in a real printing equipment, accuracy of 92.59% is obtained with constructed images and higher than 86.67% without POD. The proposed method is proven to be effective and useful in the analysis of fault signals in actual condition.