A new rolling bearing fault diagnosis method based on multiscale permutation entropy and improved support vector machine based binary tree

• Rolling bearing feature extraction from noise-contaminated sensor signals based on LMD and MPE.
• Design of a new hierarchical structures in the SVM-BT, which leads to the significant performance enhancement.
• Real-time pattern classification based on LS and ISVM-BT.

A new bearing vibration feature extraction method based on multiscale permutation entropy (MPE) and improved support vector machine based binary tree (ISVM-BT) is put forward in this paper. Local mean decomposition (LMD), a new self-adaptive time–frequency analysis method, is utilized to decompose the roller bearing vibration signal into a set of product functions (PFs) and then MPE method is used to characterize the complexity of the principal PF component in different scales. After the feature extraction, a new pattern recognition approach called ISVM-BT is introduced to accomplish the fault identification automatically, which has the priority of high recognition accuracy compared with other classifiers. Besides, the Laplacian score (LS) is introduced to refine the fault feature by sorting the scale factors. Finally, the rolling bearing fault diagnosis method based on LMD, MPE, LS and ISVM-BT is proposed and the experimental results indicate the proposed method is effective in identifying the different categories of rolling bearings.

Based on the advantages of LMD, MPE, LS and ISVM-BT, A novel bearing fault feature extraction method can be summarized as follows: (1) The vibration signals are sampled by acceleration sensors at a certain sampling frequency fs under different working conditions. (2) Apply LMD method to preprocess the sensor-based vibration signals to obtain a series of PF components. And then the sensitive PF component containing more significant state information is selected for research. (3) Define the scale factor τ and calculate MPE of the selected PF components under different scales. In the whole paper, we set the data of a PF component with data length N = 2048, scale factor s = 20, the PE values of each coarse grained time series acquired by Eq. (16) is computed with the dimension m = 4 and time delay τ = 1. (4) After calculation of MPE, LS is employed to rank the 20 features according to their importance from low to high score. Then choose the first five important features with least scores to construct the new fault feature vector. (5) The obtained new fault features are fed into fault classifier ISVM-BT for training and testing to fulfill the fault diagnosis automatically. A functional framework of LMD–MPE algorithm is presented in Fig. 4.Figure optionsDownload as PowerPoint slide