ABSTRACT Remaining useful life (RUL) prediction of rolling bearings is essential for intelligent maintenance and equipment health management. However, in practical monitoring scenarios, vibration signals are often affected by noise, operating condition fluctuations, and variations in degradation patterns, which hinder effective feature extraction and temporal modeling. To address these challenges, we propose an adaptive multi‐scale feature alignment and fusion (AMFAF) model. AMFAF first introduces an adaptive denoising module (ADM) that dynamically suppresses multi‐source noise according to the characteristics of the input signals, thereby enhancing the discriminability of raw degradation information. Then, the multi‐scale feature extraction module (MFEM) employs convolutional kernels of different receptive fields to capture degradation evolution features at local, medium, and global scales. On this basis, a feature alignment and semantic fusion (FASF) module is designed to align and fuse diverse degradation modes through neighborhood feature aggregation and semantic consistency constraints, yielding robust temporal representations. Finally, a composite loss function jointly optimizes prediction errors and alignment constraints, further improving the model's generalization and stability. Experiments on the XJTU‐SY and PHM 2012 bearing degradation datasets demonstrate that AMFAF outperforms the compared baseline models in terms of MAE, RMSE, and R 2 , indicating improved prediction accuracy and robustness under the evaluated operating conditions.
Wu et al. (Mon,) studied this question.