Domain knowledge-enhanced dual-stream graph joint learning network for aeroengine remaining useful life prediction

Prediction methods integrating domain knowledge with deep learning show great promise for remaining useful life (RUL) prediction of aeroengines. Nevertheless, neglecting the decoupling modeling of multi-level domain knowledge such as component-level physical topology and parameter-level essential dependencies has resulted in insufficient and incorrect utilization of valuable degradation information, limiting the further improvement of prediction accuracy and stability. Therefore, a novel domain knowledge-enhanced dual-stream graph joint learning network (DK-DGJLN) is proposed. First, the physical essential dependencies among components and sensor parameters are extracted through in-depth analysis of aeroengine’s overall structure and working characteristics. Then, a component-level static graph is built to capture physical topology with high fidelity, while a parameter-level dynamic graph is constructed to model dynamic dependencies and degradation evolution. Their decoupling modeling enables faithful embedding of multi-level domain knowledge. Finally, a DGJLN is designed to extract rich and complementary spatial-temporal degradation features from both graphs, with a self-attention-based feature alignment module aggregating them for accurate RUL prediction. Extensive comparative experiments on the C-MAPSS dataset against 19 SOTA methods demonstrate that DK-DGJLN significantly improves prediction performance, especially on FD002 and FD004, where the RMSE values of 12.93 and 13.57 are markedly lower than those of the best comparative methods.