Remaining useful life prediction of rolling bearings via wavelet feature fusion and LSTM networks

Abstract Bearings play a critical role in rotating machinery, making it essential to monitor their operational states to ensure machine reliability and safety. Performance degradation assessment is a key aspect of condition based maintenance (CBM) and predictive maintenance strategies. Predicting the Remaining Useful Life (RUL) of a bearing is crucial for effective health management and ensuring the safety and reliability of rotating machinery systems. With stringent demands on performance, cost, and safety, there is a growing need to develop and enhance techniques for predicting RUL. To enhance the accuracy of RUL prediction, this study presents a data-driven approach for RUL prediction, demonstrating superior performance compared to existing methods. The novelty of the proposed model lies in the fusion of entropy-based statistical features with energy-based wavelet features, a combination that remains underexplored in the context of RUL estimation. This feature integration strategy effectively captures both the statistical complexity and spectral energy distribution of the vibration signals, leading to incremental improvements in prediction accuracy, as validated by our experimental results. Specifically, employing two types of features from wavelet transform, first localized energy distribution through reconstructed wavelet packets and second statistical descriptors from multi-level wavelet decomposition, enabling comprehensive signal characterisation for degradation tracking along. Furthermore, a long short-term memory (LSTM) is employed to effectively capture and analyze fault propagation dynamics for improved bearing perfor- mance degradation assessment. To validate the effectiveness of the proposed approach, experimental evaluations are performed using two widely recognized run-to-failure bearing datasets: PRONOSTIA and XJTU-SY.