Development of a Machine Learning-Based Load Monitoring System for Aircraft Fatigue Life Management

초록

An accurate and reliable load monitoring system is essential for aircraft fatigue life management. To develop such a system, sufficient flight data that reflect real flight characteristics, accurate load equations correlating flight parameters with flight loads, and a method to analyze load equation outputs are required. In this study, flight data were acquired by extensive flight tests across diverse flight regimes, and load equations for acquired flight data were developed using machine learning models, such as XGBoost, multi-layer perceptron (MLP), and sparse Gaussian process regression (SGPR), alongside conventional models for comparison. Load equations by XGBoost, MLP, and SGPR showed superior accuracy to conventional models, with SGPR also capturing uncertainty from vibration and noise. As a method to analyze load equation outputs, SHAP (Shapley Additive exPlanations) was applied to estimate each flight parameter's contribution, and by integrating SHAP results into conventional KDE (Kernel density estimation), an enhanced KDE method was developed. It was confirmed that the enhanced KDE method enabled a reasonable evaluation of regression and generalization performance on new data without target values. The load monitoring system proposed in this study enables the development of accurate load equations and ensures the reliability of load predictions, resulting in more efficient fatigue life management.

키워드