An Attention-Based Downsampling GRU Neural Network with Degradation Model for Remaining Useful Life Prediction of Aero-engine

As the core power unit of aircraft, aero-engines require accurate health state evaluation to ensure flight safety and optimize maintenance costs. While deep learning methods have been widely adopted for predicting remaining useful life (RUL), current approaches predominantly focus on multi-sensor correlations and long-term temporal patterns, often neglecting critical variations in degradation information at individual sensor timesteps. Moreover, existing methods rarely explore the potential of multi-scale frequency sampling for richer temporal feature extraction. To address these limitations, this study proposes an Attention-based Downsampling GRU Neural Network (ADGNN). The model innovatively combines two key mechanisms: Degradation-aware attention weighting to dynamically emphasize timesteps with salient degradation signatures; Multi-scale GRU architecture to concurrently capture long-term and short-term temporal dependencies through hierarchical downsampling. Comprehensive experiments on the C-MAPSS benchmark dataset demonstrate ADGNN's superiority over methods, achieving a 7.64% reduction in RMSE. At the same time, it also demonstrated competitive results in terms of scoring. The proposed framework not only advances RUL prediction accuracy but also exhibits robust stability. This study provides an effective intelligent forecasting approach for health management in complex industrial systems and holds substantial significance for improving reliability as well as informing operation and maintenance decisions regarding aero-engines.