• Dual-attention CNN–GRU model for RUL prediction under varying operating conditions • Channel attention weights sensor degradation features, improving robustness to noise • Sequence attention highlights critical degradation stages for interpretability and decisions • Dilated residual convolutions enable multi-scale temporal feature extraction • The proposed model achieves state-of-the-art performance on NASA C-MAPSS FD002 Remaining Useful Life (RUL) prediction in complex industrial systems is challenged by (i) covariate shift induced by heterogeneous operating regimes and (ii) the misalignment between commonly used training objectives and operational risk metrics in Prognostics and Health Management (PHM). To address these challenges, this paper proposes a regime-aware dual-attention Convolutional Neural Network–Gated Recurrent Unit (CNN–GRU) framework for predictive maintenance. Operating regimes are automatically discovered through unsupervised clustering of operating-condition variables, and per-regime normalization is applied to mitigate inter-regime distribution shift without modifying the downstream model architecture. The network integrates channel attention within residual dilated CNN blocks to enhance sensor-level feature salience and sequence attention over GRU hidden states to identify diagnostically critical temporal segments. In addition, a metric-aware training objective is introduced by blending SmoothL1 loss with an asymmetric PHM-score surrogate, explicitly penalizing late RUL underestimation to better align optimization with maintenance risk preferences. The proposed framework is evaluated and validated on the challenging NASA C-MAPSS FD002 dataset, which provides a stringent test for operating-regime variability under a single fault mode, under the PHM-standard last-window protocol, achieving an RMSE of 13.36 and a PHM score of 465, outperforming the next-best baseline (STAR, RMSE = 13.47 , PHM = 784 ) by approximately 0.8% in RMSE and 40.7% in PHM score. Ablation studies confirm that regime-aware normalization, dual-attention modeling, and metric-aware optimization each contribute materially to performance gains. Overall, the proposed pipeline provides a robust, interpretable, and deployable template for multi-condition RUL prediction in real-world predictive maintenance applications.
Aasi et al. (Sun,) studied this question.