Abstract: In view of weak feature signal, signal nonlinearity and difficult extraction of multi-scale features in rolling bearing fault classification, a rolling bearing fault diagnosis method based on optimized multi-scale permutation entropy (MPE) and convolutional neural network (CNN) was proposed. The decomposition and reconstruction of bearing signals by improving complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) could achieve the signal noise reduction. The MPE was optimized by particle swarm optimization (PSO), the PSO-MPE feature extraction method was proposed, and the parameter-optimized MPE could extract more critical feature information. The resulted permutation entropy was input into CNN for fault classification and dimensionality reduction visualization analysis. The ICEEMDAN-PSO-MPE-CNN method proposed in this paper was analyzed and compared with CEEMDAN-SVM, ICEEMDAN-PSO-MPE-RNN, ICEEMDAN-PSO-MPE-SVM and other methods longitudinally and cross-sectionally by taking Case Western Reserve University’s open bearing database samples as test objects. The results show that the improved method enjoys higher classification accuracy and efficiency with more obvious classification effect under T-SNE visualization, thus can realize the high precision and high efficiency detection of rolling bearing faults.