Abstract:  In order to improve the accuracy of bearing fault diagnosis, the full view feature extraction method of bearing fault and the pattern diagnosis method of expert forest algorithm were proposed. Firstly, in terms of fault feature extraction, parameters from time domain, frequency domain and time-frequency domain were selected as the initial fault feature library. Then the global structural features of the basic fault library were extracted using KPCA, and the local structural features of the basic fault library were extracted using t-SNE algorithm, so as to retain the full view feature parameters that were relatively sensitive to fault modes. In the aspect of fault pattern recognition, expert attributes and expert weights were given to the decision tree, and the concept of expert tree was obtained. Based on the idea of expert tree, an expert forest algorithm was proposed, which solves the problem that the random forest algorithm treats the decision tree indiscriminately. Finally, experiments were used to verify the full-view feature extraction method of bearing faults and the mode diagnosis method based on the expert forest algorithm. The experimental results show that the full view fault features extracted by KPCA + t-SNE are better than the global and local structure features extracted separately; the average diagnosis accuracy of random forest algorithm is 96.14%, and the average fault diagnosis accuracy of expert forest algorithm is 99.48%, which is 3.47% higher than that of random forest algorithm, and verifies the superiority of the proposed fault diagnosis method.

Key words:  bearing fault diagnosis; global structural features; local structural features; initial feature library; expert forest algorithm