Reliable prognosis for complex systems based on spatial-temporal representation modeling and uncertainty-aware dynamic Fusion

Remaining useful life (RUL) prediction for complex systems is essential for ensuring their reliability and operational safety. Deep neural networks have demonstrated strong capabilities in prognostic modeling by learning degradation patterns from multivariate sensor data. However, their performance remains limited due to two key challenges: the difficulty of jointly capturing temporal degradation dynamics and cross-sensor dependencies, and the effective fusion of information from multiple variables. To overcome these challenges, this study develops an intelligent prognostic framework composed of three main modules: (i) a spatial–temporal feature embedding module for multivariable feature encoding; (ii) a Bayesian event-triggered dynamic graph update mechanism to robustly characterize evolving spatial relationships during degradation; and (iii) an uncertainty-aware fusion module for decision-level RUL prediction. Experiments on complex system RUL prediction tasks show that the proposed spatial–temporal network significantly improves prediction accuracy. The Bayesian event-triggered graph update further makes graph evolution more robust and better aligned with degradation progression. In addition, the uncertainty-aware fusion strategy improves uncertainty calibration and interval coverage, allowing a greater proporsion of true RUL values to fall within the predicted confidence bounds. The proposed framework make significant contribution on the provition of increased more reliable RUL prediction and improved support for predictive maintenance.