Adaptive Conformal Interval Estimation for Uncertainty Quantification in RUL Prediction

Abstract

Remaining useful life (RUL) prediction under complex operating conditions often suffers from limited reliability, while traditional point prediction methods are unable to characterize the uncertainty associated with prediction results. To address this issue, this paper proposes an adaptive interval estimation method that combines global constraints with local similarity. Specifically, a Gated Recurrent Unit (GRU) is employed as the backbone prediction model, and the feature distribution of calibration samples in the latent space is utilized to characterize the local state of each test sample. Meanwhile, a uniform floor weight is introduced to mitigate the instability caused by purely local weighting, thereby enabling dynamic interval construction across different degradation stages. Experiments conducted on a full-life harmonic drive dataset demonstrate the effectiveness of the proposed method. The results show that the method maintains satisfactory interval validity while significantly reducing the prediction interval width, and it also exhibits stronger adaptability to varying health states. Overall, the proposed method provides offers more reliable support for RUL prediction under complex operating conditions.