Abstract: Gas path fault diagnosis is an essential component of aero-engine health management systems, with fault feature extraction being its key aspect. In recent years, with the advancement of deep learning techniques, gas path fault feature extraction methods based on graph neural networks have attracted considerable attention. However, conventional graph neural networks can only capture linear weighted aggregation relationships among nodes while neglecting the nonlinear coupling relationships prevalent in engine gas path systems, resulting in insufficient cross-condition diagnostic accuracy and interpretability. To address this issue, a gas path fault diagnosis method based on nonlinear correlation mining is proposed. This method achieves interpretable extraction of gas path fault features through an original nonlinear correlation mining layer and an improved graph convolutional layer. The performance of the proposed method was validated using full-lifecycle simulation data encompassing 500 flight sorties. The proposed method achieved zero false alarms throughout the entire operational period, with a detection rate of 88.99% and an isolation rate of 98.61%, significantly outperforming comparative methods based on convolutional, graph convolutional, and graph attention networks in terms of convergence and diagnostic accuracy.

Key words: aero-engine, gas path diagnosis, feature extraction, graph neural network, interpretability