FCTAN: A frequency-constrained temporal attention network for rolling bearing remaining useful life prediction

Accurate remaining useful life (RUL) prediction of rolling bearings is crucial for condition-based maintenance of rotating machinery. However, vibration-based degradation modeling remains challenging due to strong noise interference, multi-scale temporal dynamics, and the difficulty of jointly capturing local degradation patterns and long-term dependency information. To address these issues, this paper proposes a novel Frequency-Constrained Temporal Attention Network (FCTAN) for bearing RUL prediction. Specifically, a frequency-constrained degradation representation module is first introduced to transform the vibration-derived degradation feature sequence into the frequency domain and perform learnable spectral re-weighting, thereby enhancing degradation-sensitive spectral responses and suppressing redundant spectral information. Then, a temporal feature expansion and contrastive normalization (ContraNorm) mechanism is employed to project the input sequence into a higher-dimensional latent space while reducing feature redundancy and stabilizing representation learning. Furthermore, a multi-scale Inception1D module is designed to capture local temporal patterns at different resolutions, and a Transformer encoder is adopted to model long-range temporal dependencies across the entire degradation trajectory. Finally, an attention-based temporal aggregation mechanism is applied to adaptively emphasize degradation-critical stages and generate a robust global representation for RUL regression. Experimental results on two public bearing degradation datasets demonstrate that the proposed FCTAN achieves superior prediction accuracy and stability compared with several state-of-the-art methods, validating its effectiveness for bearing RUL prediction under complex and noisy operating conditions.