Aero-engine remaining useful life prediction using state-involving graph networks with multi-scale perception enhancement

Document Type

Article

Publication Date

7-1-2026

Abstract

Accurate remaining useful life (RUL) prediction for aero-engines is vital for flight safety and maintenance optimization. Despite the success of data-driven models, most existing methods still face two critical challenges: the low signal-to-noise ratio of weak degradation-related indicators and the state-dependent dynamic interactions among them. To bridge these gaps, we propose the multi-scale enhanced state-involving graph network (ME-SiGN). Specifically, a multi-scale enhanced local perception module is first developed to extract robust degradation patterns from noisy signals through multi-path pooling and patch weighting. Building on these enhanced features, we construct a state-involving graph construction with graph memory optimization module, optimized by Top- (Formula presented) (Formula presented) neighbor selection. Then, a graph memory refinement unit is introduced to incorporate a causality-inspired temporal order constraint and improve global topological adaptability across diverse operating conditions. Furthermore, a Graph Attention Feature Extractor, featuring parallel dual-branch convolutions, is designed to fuse spatial interactions with temporal dynamics, effectively capturing both long-term trends and short-term fluctuations. Extensive experiments on the C-MAPSS benchmark demonstrate that ME-SiGN achieves competitive performance compared with 13 recent baselines. Notably, on the challenging FD004 subset, ME-SiGN achieves an root mean square error (RMSE) of 13.49 and a Score of 805, improving predictive performance under complex multi-operating-condition settings. On the other complex FD002 subset, it further attains an RMSE of 12.73 and a Score of 681, demonstrating consistent robustness under multi-operating-condition settings. The reduced Score indicates a more conservative RUL estimation behavior, which is important for reducing over-estimation risk in aviation maintenance. Systematic analysis confirms that ME-SiGN maintains a favorable balance between accuracy, computational complexity, and parameter efficiency.

Publication Title

Engineering Research Express

Link to Full Text

Share

COinS