A degradation-informed RUL prediction method for rolling bearings using RBSU-FI feature selection and a BiLSTM network

Abstract

Rolling bearings are critical components of rotating machinery, and accurate remaining useful life (RUL) prediction is essential for predictive maintenance. However, the nonlinear nature of bearing degradation makes degradation onset difficult to identify, while weakly relevant features may compromise prediction reliability. To address these issues, this article proposes a novel degradation-informed RUL prediction method integrating robust symmetric uncertainty-feature importance (RBSU-FI) feature selection with a bidirectional long short-term memory network and a multi-layer perceptron compression module. First, a novel Lempel-Ziv complexity-enhanced Envelope-Kurtosis-Isometric Mapping health indicator is constructed by integrating complementary degradation characteristics with signal complexity to identify degradation onset. Then, a novel RBSU-FI feature selection indicator is introduced to evaluate feature contributions by considering RUL relevance and feature-set importance. Finally, the selected degradation-sensitive features are fed into the proposed prediction model for RUL prediction. The method is evaluated on the PRONOSTIA and Intelligent Maintenance Systems (IMS) datasets. It achieves an average mean squared error (MSE) of 0.012 and mean absolute error (MAE) of 0.042 on PRONOSTIA, and an MSE of 0.002 and MAE of 0.035 on IMS, outperforming representative feature selection methods and prediction models. These results demonstrate the effectiveness of the proposed method for rolling bearing RUL prediction.