System hierarchical stage-wise evidence domain adaptation for aero-engine remaining useful life prediction

Accurate remaining useful life (RUL) estimation of aero-engine from multi-sensor measurements is essential for condition monitoring and health management. Existing domain adaptation methods for RUL prediction often suffer from temporal stage misalignment and insufficient adaptation of multi-level degradation information. To address these issues, we propose a system hierarchical stage-wise evidence domain adaptation (SHSEDA) method. We construct a three-level representation hierarchy consisting of sensor, component, and energy-transfer layers to characterize the physical degradation hierarchy. Based on this topology, we develop a hierarchical graph neural network with cross-gated dual-adjacency fusion and cross-layer residual gated projection to capture local degradation patterns and global degradation trends from truncated degradation trajectories. Furthermore, a dual-layer stage-wise evidence alignment strategy divides degradation processes into matched stages to mitigate temporal misalignment and aligns evidence distributions between sensor and energy-transfer layers, enabling hierarchical evidence knowledge transfer across domains. Experiments on the C-MAPSS and N-CMAPSS datasets demonstrate the effectiveness of SHSEDA. On the C-MAPSS dataset, SHSEDA reduces root mean square error (RMSE), asymmetric scoring function (Score), and mean absolute error (MAE) by 8.08%, 21.91%, and 8.03%, respectively, compared with existing domain adaptation methods. These results demonstrate improved accuracy of sensor-based RUL estimation under cross-condition distribution shifts and incomplete target-domain degradation measurements.