RaDDA-RUL: Ranking-Driven and Degradation-Aware Domain Adaptation for Cross-Machine Remaining Useful Life Prediction

Cross-machine remaining useful life (RUL) prediction from vibration signals is challenging because structural resonance, sensor installation, operating conditions, and acquisition protocols introduce substantial distribution shifts between training and deployment domains. Conventional unsupervised domain adaptation (UDA) based on global alignment may associate samples at different degradation stages, producing progress-inconsistent representations. This study proposes RaDDA-RUL, a ranking-driven and degradation-aware domain adaptation framework. A SincNet-based temporal convolutional network (TCN) backbone extracts frequency-selective and multi-scale temporal representations. For the unlabeled target domain, a Bradley–Terry pairwise ranking regularizer uses temporal order to regularize predicted normalized lifetime progress, which is subsequently incorporated into progress-conditioned adversarial adaptation. Quantile regression provides prediction intervals. The evaluation follows an offline sample-transductive UDA setting in which complete unlabeled target trajectories are available during adaptation, while target RUL values are used only for post-adaptation evaluation. Across six transfer tasks constructed from the PHM2012 and XJTU-SY datasets, RaDDA-RUL achieves a lower root mean square error (RMSE) and mean absolute error (MAE) than the compared methods, including an average RMSE reduction of 13.1% relative to domain invariant and consistent ordinal representation learning (DICORL). These results support degradation-order-aware cross-machine RUL adaptation under the evaluated offline protocol.