Latent Diffusion Neural Fields for zero-shot and uncertainty-aware bearing remaining useful life prediction

Remaining useful life (RUL) prediction for rolling bearings remains challenging under nonstationary degradation, sparse sensing, missing observations, and operating-condition shifts. Existing methods are predominantly formulated as deterministic regression on fixed observation grids, which limits their ability to handle irregular measurements, multimodal degradation evolution, and test-time adaptation. To address these limitations, this study reformulates bearing prognostics as conditional generation and posterior inference over a continuous time-frequency degradation field, and proposes a latent diffusion neural field (LDNF) framework for zero-shot and uncertainty-aware RUL prediction. A SIREN-based coordinate neural field is used to represent vibration responses continuously across time, frequency, and channel coordinates, while FiLM conditioning injects operating-state information to improve cross-condition consistency. On this basis, a latent diffusion model is trained to capture distributions of degradation trajectories in latent space rather than a single deterministic future path. During inference, posterior-guided sampling assimilates sparse, incomplete, or corrupted observations without retraining, enabling zero-shot adaptation to new bearings and observation patterns. Decoded future fields are mapped to monotonic health-index trajectories, and RUL is estimated through a first-passage-time criterion, yielding both point predictions and credible intervals. Experiments on the PRONOSTIA benchmark show that the proposed framework consistently outperforms representative statistical, machine-learning, and deep-sequence baselines in terms of point prediction accuracy and probabilistic quality. In addition, it provides a unified capability for super-resolution reconstruction, cross-channel synthesis, missing-data restoration, and calibrated uncertainty quantification, demonstrating strong potential for risk-aware maintenance under realistic monitoring constraints.