Multi-Source Aero-Engine Fault Diagnosis Using Explainable Boosted Tree with Spatiotemporal Attention and Adaptive...

Area of Science:

• Aerospace Engineering
• Mechanical Engineering
• Artificial Intelligence

Background:

• Aero-engine rotating component faults cause over 60% of failures, with early signs often obscured by noise.
• Traditional single-sensor diagnosis methods exhibit limitations in feature utilization, interpretability, and cross-condition generalization.

Purpose of the Study:

• To develop a robust multi-source fault diagnosis method for aero-engines.
• To enhance early fault detection sensitivity and model interpretability under complex operating conditions.

Main Methods:

• Proposed a novel method integrating an explainable boosted tree with spatiotemporal attention (STA) and adaptive feature selection (AFS).
• Utilized multi-domain data from standard core sensors, extracting heterogeneous features.
• Implemented AFS using mutual information and variance inflation factor, and employed STA for effective feature fusion.
• Employed XGBoost classifier with SHAP values for interpretability and quantitative analysis of uncertainty.

Main Results:

• Achieved 99.2% diagnosis accuracy and 97.5% cross-condition generalization accuracy on a fault simulation test rig.
• Demonstrated high sensitivity to early weak faults and stable uncertainty under complex operating conditions.
• Outperformed conventional diagnostic models in accuracy and generalization.

Conclusions:

• The proposed method offers a reliable and cost-effective solution for aero-engine fault diagnosis.
• It enables early identification of fault signatures, clarifies key indicators, and supports maintenance decision-making.
• The approach requires no additional hardware, featuring lightweight computation and low inference overhead.