Sigmoid-based refined composite multiscale fuzzy entropy and t-SNE based fault diagnosis approach for rolling bearing

Multiscale fuzzy entropy (MFE) has been taken as a prevalent tool for complexity measure of time series. However, the entropy estimation of MFE is sensitive to the preset parameters and will cause an inaccurate and even undefined entropy when the data is not long enough. In this paper the Sigmoid-based refined composite multiscale fuzzy entropy (SRCMFE) is proposed to improve the performance of MFE for complexity measure of short time series. The proposed SRCMFE method is compared with MFE and Sigmoid-based multiscale fuzzy entropy by analyzing Gaussian white noise and 1/f noise to verify its effectiveness and the analysis result indicate that SRCMFE method holds a better distinguish capacity and robustness and can reflect more complexity information of time series. Then SRCMFE is used to the dynamical complexity analysis of mechanical vibration signals and based on that a new fault diagnosis approach for rolling bearing is put forward by combining SRCMFE with t-distributed stochastic neighbor embedding (t-SNE) for feature dimension and the recently proposed variable predictive models based class discrimination (VPMCD) method for fault pattern recognition. In the proposed method, firstly, SRCMFE is utilized to extract the complexity characteristic related with fault information from vibration signals of rolling bearing. Then the feature dimension reduction method named t-SNE is adopt to obtain a low dimensional manifold features of rolling bearing. Next, the VPMCD method is employed for multi-fault classifier construction to fulfill an intelligent fault diagnosis according to the intrinsic inner relationships hidden in the fault features. Finally, the proposed fault diagnosis approach is applied to the experimental data analysis of rolling bearing and is compared with the same kinds of fault diagnosis methods through experiment data analysis. The analysis and comparison results indicate that the proposed method is very effective in distinguishing the fault categories of rolling bearings and can get a higher identifying rate than the contrast methods.