A Condition-Aware Spatio-Temporal Network for Robust Aero-Engine Remaining Useful Life Prediction

Abstract

Aero-engine remaining useful life (RUL) prediction plays a crucial role in ensuring flight safety and optimizing maintenance strategies. Nonetheless, enormous efforts formulate RUL prognostics as a single-modal temporal modeling paradigm, neglecting the significance and complementarity of the complex nonlinear synergy between operating conditions (OCs) and sensor measurements (SMs). In this work, we construct a novel condition-aware prognostics framework termed synergistic feature interaction (SFI)-spatio-temporal hierarchical graph attention network (STHGAT), where heterogeneous feature interactions are incorporated with an SFI module and multiscale spatio-temporal dynamics are aggregated by an STHGAT. Specifically, an adaptive learning module based on nonlinear primitives and hash-based frequency interactions fast Fourier transform (FFT) is developed to capture the global response of condition-dependent modulations, and the STHGAT further investigates the spatial correlation and temporal evolution of the degradation trajectory from a hierarchical perspective. In this manner, we can strategically determine the dynamic influence of diverse operating regimes and aggregate the foremost informative features spanning from high-resolution transient behaviors to long-range stage-wise trends. Extensive experiments on the commercial modular aero-propulsion system simulation (C-MAPSS) and N-CMAPSS datasets, together with rigorous cross-domain transfer learning tests, demonstrate that our framework not only achieves state-of-the-art performance but also exhibits superior robustness and generalization under unseen OCs. These results underscore the effectiveness of explicitly modeling condition-aware feature interactions and hierarchical temporal patterns for advancing RUL prognostics in nonstationary degradation processes.