Aeroengine Remaining Useful Life Prediction Via Integrating Gated and Hybrid Attention Transformer

To address the limitations in aero-engine remaining useful life (RUL) prediction where existing deep learning models use fixed degradation thresholds without considering individual engine variations or intelligent parameter selection, exhibit low prediction accuracy, fail to capture both short-term and long-term dependencies in time series, demonstrate insufficient feature extraction, and show insensitivity of attention mechanisms to local features, this study develops a P-PCA-based performance evaluation metric to determine initial and linear degradation thresholds, employs a dual-GRU model with dynamic programming change-point detection to identify and retain degradation-phase data for intelligent performance assessment, and introduces the Hybrid Attention-GRU (HAG) model which incorporates a Transformer encoder backbone with multi-head attention for temporal feature extraction and hybrid self-attention mechanisms, achieving an average 15% reduction in mean squared error and 46% decrease in scoring function values compared to other advanced models.