Predictive representation learning via latent degradation dynamics for remaining useful life estimation

Remaining useful life (RUL) estimation is a key task in prognostics and health management, requiring models that can effectively capture degradation progression from multivariate sensor data. While recent deep learning approaches have shown promising performance, most methods formulate RUL estimation as a direct regression problem or rely on reconstruction-based representation learning, which may not explicitly preserve the progressive nature of system degradation. Since degradation is inherently a cumulative and irreversible process, effective RUL estimation requires representations that capture how system health evolves toward failure rather than merely describing the current operating state. To address this challenge, this paper proposes a predictive representation learning framework that learns latent representations by modeling delta-aware multi-horizon latent degradation dynamics. Building upon the Joint Embedding Predictive Architecture (JEPA), which learns representations through future prediction in the embedding space, the proposed framework reformulates predictive representation learning to explicitly capture degradation evolution in the latent space. To this end, temporal differences are incorporated to capture local degradation variations, while latent delta prediction models degradation as progressive latent transitions rather than independent future states. A multi-horizon prediction strategy further learns these latent degradation dynamics across multiple future horizons, encouraging the learned representations to remain predictive throughout the degradation process. Experiments on the NASA C-MAPSS and N-CMAPSS benchmark datasets demonstrate the effectiveness of the proposed framework. The proposed method achieves the best average performance on the C-MAPSS benchmark and the lowest RMSE on the N-CMAPSS DS02 dataset, consistently outperforming recent representation learning approaches. Additional validation on the PRONOSTIA run-to-failure bearing dataset further demonstrates favorable RUL estimation performance under a substantially different machinery type and sensing modality, supporting the applicability of the proposed framework beyond turbofan-engine degradation. Qualitative analyses further reveal that the learned latent space organizes degradation progression into smooth and interpretable trajectories that closely align with system health evolution. These results demonstrate that explicitly modeling degradation evolution through predictive representation learning yields informative, trajectory-consistent, and interpretable representations for prognostics.