Recent advances in deep learning have greatly promoted remaining useful life (RUL) prediction of rotating machinery, yet challenges remain in effectively capturing multi-scale temporal patterns from noisy vibration signals. Existing models generally struggle to represent both long-term degradation evolution and short-term dynamic fluctuations, and their feature extraction is further hindered by substantial noise and redundancy in raw signals. Moreover, many current frameworks lack coordinated optimization of spatiotemporal information, limiting their prediction reliability under complex operating conditions. To address these limitations, this study proposes a hybrid multi-scale TCN–BiLSTM–SE (MTBS) model that integrates temporal convolutional networks (TCN) and bidirectional long short-term memory (BiLSTM) in a parallel architecture. The model incorporates multi-scale convolutional branches for hierarchical temporal representation, a semi-soft threshold denoising module for adaptive noise suppression, and a channel attention mechanism to highlight informative features. Extensive experiments on the IEEE PHM2012 and XJTU-SY bearing datasets validate the effectiveness of the proposed model. MTBS achieves consistent improvements across RMSE, MAE, and Score on both datasets, demonstrating clear advantages over traditional and recent advanced models. These results demonstrate that MTBS more effectively captures discriminative degradation patterns and substantially enhances prediction accuracy and robustness in end-to-end RUL estimation.
Yang et al. (Wed,) studied this question.