Accurately identifying the onset of bearing degradation is critical for reliable remaining useful life (RUL) prediction. Existing data-driven prognostic models commonly assume a fixed or uniformly defined change point throughout the operational lifetime, which may obscure the actual transition from healthy to degrading behaviour in vibration signals. To address this limitation, this study proposes a dual-stage long short-term memory (LSTM) framework that integrates vibration change detection with bearing health assessment for RUL estimation. In the first stage, an LSTM-based vibration change detection (VCD) module is employed to identify the onset of degradation. In the second stage, a separate LSTM model predicts a bearing health index (BHI), which is subsequently mapped to RUL. The proposed framework is validated using a newly developed run-to-failure vibration dataset collected from a dedicated bearing–shaft misalignment simulator operating under angular misalignment conditions. To investigate representation effects, vibration signals are analysed in both time and frequency domains. Furthermore, the proposed method is compared against three baseline LSTM configurations and two gated recurrent unit (GRU) models. Experimental results demonstrate that frequency-domain features, when combined with the dual-stage LSTM architecture, provide the most accurate and stable prognostic performance, outperforming all baseline models. Specifically, the frequency-based dual-stage LSTM model achieves an RMSE of 0.0067, an RPE of 3.698, and a scoring function value of 0.0002. These findings highlight the advantages of frequency-domain analysis in enhancing change-point localization and health index modelling, ultimately improving the accuracy and robustness of bearing RUL prediction under misalignment-induced degradation. • A novel dual-stage LSTM framework for bearing RUL prediction which integrating vibration change-point detection and bearing health index. • A novel, proprietary bearing vibration dataset under naturally occurring angular shaft misalignment. • A new dedicated test rig was developed to produce this dataset under realistic operational conditions. • A comparative analysis of bearing vibration data in both the time and frequency domains using single and dual LSTM architectures • Frequency-domain features combined with the two-stage LSTM architecture, yield the most accurate and stable performance for bearing RUL prediction.
Atmaji et al. (Fri,) studied this question.