Diffusion-RarePHM: A Physics-Constrained Conditional Latent World Model for Multi-Future Degradation Forecasting and Rare-Fault Prognostics

Industrial prognostics is constrained less by the abundance of healthy telemetry than by the scarcity, heterogeneity, and operational ambiguity of genuine failure evolution. This study introduces Diffusion-RarePHM, a physics-constrained conditional latent world model that converts heterogeneous monitoring streams into probabilistic, multi-future degradation scenarios. A unified health encoder maps engine and compressor observations into a common mechanism-oriented health interface with asset-specific fitting, while state-dependent dual-domain dynamics couple latent health levels with their increments. Conditioned on current health, equipment structure, prospective duty, environment, maintenance action, and target fault mode, latent diffusion generates an ensemble of complete future trajectories. Rare precursors are synthesized as physically directed departures from normal latent evolution and screened by physical, dynamic, and explicit temporal-predicate semantic criteria. The study draws from PHM2024 and MetroPT3 source archives containing 742,625 and 1,516,948 records, respectively; deterministic PHM2024 subsets and 12,642 chronologically retained MetroPT3 observations are used for the reported experiments. On the PHM2024 record-level held-out diagnosis partition, the model attained 98.23% accuracy, 98.16% macro-F1, and 0.9985 AUROC. Across held-out operating-condition groups, it attained 97.86±0.12% accuracy, 97.77±0.13% macro-F1, and 0.9979±0.0003 AUROC over five training seeds. In the MetroPT3-anchored controlled semi-physical benchmark, reconstruction of prescribed health trajectories achieved nine-scenario averages of 2.250 percentage-point RMSE, 2.294% MAPE, and 90.00% pointwise constructed-reference coverage for the nominal 90% interval. Controlled trajectory experiments examine the generator’s response to prescribed precursors, leakage, compound degradation, maintenance, thermal-load stress, and observation corruption, together with their first-passage and tail-risk functionals. The trajectory evidence establishes mechanism-conditioned generation and reconstruction of controlled degradation scenarios, with accepted-path warning, first-passage, and tail-risk statistics derived under the specified dynamics and interventions.