Multi-feature extraction and IGWO-SVM for harmonic reducer fault diagnosis

To solve the problem that the extracted feature information was insufficient and the classification network was easy used to fall into over-fitting when the fault diagnosis of harmonic reducer was carried out. A fault identification method was proposed, which combined improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) for denoising and multi-feature extraction, and improved grey wolf optimization (IGWO) to optimize support vector machine (SVM). Firstly, the vibration signals of multi-condition harmonic reducer with different fault sources were decomposed by ICEEMDAN, and correlation analysis was applied to complete signal reconstruction and realized signal noise reduction processing. Secondly, the time-frequency entropy feature of the data was extracted to reduce the dimension of the data and enrich the feature information of the data. Finally, by improving the convergence factor, proportional weight and population initialization of GWO, the IGWO-SVM model was constructed to classify the data and had completed the fault identification of harmonic reducer. The results show that the average accuracy of the proposed method can reach 91.27%. Compared with the verification set of GMO-SVM, the accuracy of the proposed method is improved from 87.5% to 90%. The proposed method can effectively identify the fault of the multi-operating harmonic reducer, and has strong generalization ability.