A novel approach to bearing prognostics based on impulse-driven measures, improved morphological filter and practical health indicator construction

Abstract

Data-driven predictions for bearing remaining useful lifetime (RUL) exhibit the potential in maintaining safety and reliability in industry. However, it is still a challenge to construct an accurate health indicator (HI) and RUL prediction model, particularly for long-term predictions. In this study, a novel approach is developed for bearing prognosis. First, an improved morphological filter and an adaptive bandpass filter are designed to accurately identify resonant frequency bands of bearings and extract weak impulses from noisy vibrations. Two impulse-driven measures, namely fault frequency amplitude (FFA) and its ratio, are newly defined to optimize parameters for signal processing. FFA is also selected as a sensitive feature for assessing bearing degradation. Second, a practical HI is designed based on multi-domain features and feature selection. The HI generates a smooth and monotonic degradation trend while maintaining sensitivity towards incipient defects. Finally, a hybrid model is constructed based on two popular degradation models to improve prediction accuracy. The results, obtained for wheelset bearings and two open-source bearings, demonstrate that the proposed measures and processing methods can adaptively extract repetitive impulses. Furthermore, the constructed HI and hybrid model perform more stable and accurate RUL predictions for different bearings and prediction steps.