Performance-based probabilistic RUL prediction via particle filtering with neural-guided inference

Accurate prediction of the Remaining Useful Life (RUL) is essential for efficient operations and predictive maintenance of industrial systems. Prognostics is traditionally structured as a two-step procedure comprising estimation of the current system condition and prediction of future behavior and condition. Many data-driven approaches map sensor measurements directly to RUL without explicitly modeling degradation dynamics. We propose a data-driven prognostic framework that preserves this estimation–prediction structure without requiring physics-based state models. In the estimation stage, health indices are inferred from sensor measurements. Degradation parameters are treated probabilistically and inferred within a Bayesian framework, with online sequential updates via particle filtering. In the prediction stage, the stochastic degradation model is evaluated conditioned on the inferred parameter posterior, yielding an RUL distribution as a random variable with a hitting time. An offline-trained neural network serves as a particle-filter (PF) controller to improve adaptability while maintaining principled uncertainty propagation. Evaluation on the N-CMAPSS benchmark demonstrates competitive RUL accuracy and statistically grounded uncertainty across operating conditions, and comparisons indicate the advantages of explicitly separating estimation from prediction in data-driven prognostics.