A Novel Uncertainty-Aware Pairwise Ranking Method for Remaining Useful Life Prediction

Accurate and consistent remaining useful life prediction is crucial for proactive maintenance, yet traditional regression methods often suffer from prediction lag and nonmonotonicity due to noise and the non-linear nature of degradation. To address this, a novel uncertainty-aware pairwise ranking method is proposed, which strategically integrates probabilistic modeling with structural constraints. Specifically, a Gaussian regression head is used to estimate remaining useful life as a distribution. This predicted variance is then leveraged by probabilistic pairwise ranking to enforce the required monotonic degradation trend. Crucially, an uncertainty attention mechanism is introduced, which dynamically weights the rank constraint using uncertainty, effectively mitigating the detrimental effects of noisy samples and stabilizing the learning process. Evaluated on the C-MAPSS dataset, the proposed method achieved significant performance improvements over the baseline. The results confirm that the proposed method successfully achieves an optimal balance between predictive fidelity and physical consistency, providing a robust solution for remaining useful life estimation.