Continuous degradation representation learning for remaining useful life prediction

Abstract Predicting the remaining useful life (RUL) of aero-engines is crucial for aviation safety and maintenance scheduling. Existing deep learning approaches, however, often learn fragmented or discontinuous degradation representations from multivariate sensor data, limiting their ability to accurately model degradation trajectories and estimate RUL. To address this issue, we propose CoDeR, a continuous degradation representation learning framework for aero-engine RUL prediction. CoDeR is designed to learn temporally coherent degradation representations that better capture the progressive degradation process. Specifically, a dual-axis coupled transformer encoder (DCTE) is developed to jointly model temporal dynamics and latent feature interactions along two complementary axes. In addition, we introduce a time-aware contrastive learning (TACL) module that incorporates temporal priors into contrastive learning. The module explicitly regulates relationships among samples at different degradation stages, thereby preserving degradation continuity in the feature space. By jointly optimizing degradation representation learning and RUL regression, CoDeR learns more discriminative and continuous degradation representations, leading to improved prediction accuracy. Experimental results on representative aero-engine benchmark datasets demonstrate the effectiveness and superiority of the proposed framework.