Remaining useful life estimation based on improved attention mechanism

• Proposes a degradation-augmented attention framework (DA-RUL) for remaining useful life estimation. • Integrates a distance-based similarity kernel to emphasize geometrically significant degradation variations. • Introduces sensor-segment embeddings to capture degradation events spanning multiple time steps. • Adopts the scalar LSTM network for efficient temporal modeling. • Demonstrates superior performance on the C-MAPSS and XJTU-SY with demonstrated robustness. Deep learning models have achieved significant success in remaining useful life (RUL) estimation, a task that is essential for guiding maintenance decisions and improving the reliability of mechanical systems. However, degradation patterns often vary substantially under different operating conditions, and most existing approaches fail to fully exploit this degradation information. In addition, many methods treat all sensor measurements equally, despite the fact that only a subset of signals effectively capture critical degradation events that strongly influence system health. To address these limitations, we propose the Degradation-augmented remaining useful life (DA-RUL) model, which integrates distance-based similarity measures into the self-attention mechanism to emphasize geometrically significant variations in the data. By augmenting sensor sequences in this manner, DA-RUL enhances the model’s ability to focus on critical degradation events, thereby improving predictive accuracy. Furthermore, a scalar-LSTM is employed to capture both temporal dependencies and local feature interactions. The proposed model is evaluated on NASA’s C-MAPSS aero-engine run-to-failure datasets and the XJTU-SY bearing datasets. Experimental results show that DA-RUL not only improves RUL estimation accuracy but also adapts effectively to diverse operating conditions while maintaining low computational complexity. The source code of DA-RUL is available at https://github.com/TCCofWANG/DA-RUL .