Cross-attention multi-scale state space model for remaining useful life prediction of aircraft engines

Health monitoring and remaining useful life (RUL) prediction of aircraft engines are critical for aviation safety and maintenance decision-making. However, existing methods struggle to fully exploit the nonlinear interactive features across multi-sensor signals, limiting their ability to represent global degradation trends. Additionally, the dynamic interplay mechanisms between long-term macroscopic deterioration and short-term local anomaly patterns remain insufficiently captured, compromising the granular expression of features. To address these challenges, we propose CM-Mamba, a cross-attention multi-scale state space model for RUL prediction. Specifically, we first devise a dual-channel multi-scale patching strategy to separately extract global long-range degradation features and local short-term anomaly patterns. Then, a bidirectional state space model (Mamba) with reverse scanning mechanism is employed to capture global degradation trends across sensors and enhance spatiotemporal correlations. Moreover, windowed self-attention is adopted to refine local sensor degradation details, complemented by a cross-attention mechanism to facilitate global–local feature interactions. After fusing multi-scale features, a fully connected network generates RUL predictions. Experiments based on the C-MAPSS dataset demonstrate that this method significantly improves prediction accuracy under complex conditions and multiple fault modes, validating its superiority in cross-variable correlation modeling and multiscale degradation dynamics analysis.