A Synthetic Feature Processing Method for Remaining Useful Life Prediction of Rolling Bearings

Abstract

In the context of industrial big data, the data-driven remaining useful life prediction for rolling bearings has been greatly developed. Aimed at the shortcomings of feature selection, feature fusion, and health state segment, this article proposes a synthetic feature processing method for remaining useful life prediction of rolling bearings. First, a double-layer feature selection method is proposed to screen the feature subset from the candidate features in order to lift the degradation trend sensibility and eliminate redundancy. Second, the health indicator is constructed by a comprehensive method based on the adaptive feature fusion method and auto association kernel regression model. Third, the bottom-up algorithm is utilized to divide the health states of health indicator and obtain the first prediction time point. In this way, the remaining useful life prediction error brought by random-state dividing can be avoided. Finally, in the case of dynamic working conditions and multiple failure modes, the remaining useful life prediction model is built by long short-term memory neural network and piecewise linear fitting health indicator to map the relationship between fitting health indicator and remaining useful life. The proposed piecewise linear fitting health indicator is compared with the original health indicator and original vibration signal. The experimental results demonstrate the more robust and accurate prediction effectiveness of the method.