DEVELOPMENT OF A MACHINE LEARNING-BASED STRESS SPECTRUM ESTIMATION TECHNIQUE FOR FATIGUE MONITORING

초록

Load monitoring of aircraft is essential for safe operation. By monitoring loads occurring on the aircraft during operation, fatigue life can be predicted, and structural damage and defects can be detected in advance to ensure the structural safety of the aircraft. For this, a study is conducted in this paper to accurately estimate loads based on flight parameters recorded during flight tests. Flight tests were previously conducted with load monitoring sensors attached to the aircraft to obtain flight parameters and sensor data. MLR (Multiple Linear Regression) and ANN (Artificial Neural Network) regression models are applied to predict load monitoring sensor data using the acquired flight parameters. The regression performance of the two models is quantitatively evaluated using RMSE (Root Mean Squared Error) and adjusted R-squared (Adj. R²). As a result, the mean Adj. R² values for all targets were 0.9751 for ANN and 0.8858 for MLR, and the mean RMSE values were 0.0601 for ANN and 0.2043 for MLR. This indicates that the regression performance of ANN is higher than that of MLR. Additionally, the trained MLR and ANN models are tested using new flight test data. The performance of ANN is also higher than that of MLR for this new flight test data, confirming that the generalization performance of ANN is significantly superior. Through the correlation coefficients and variance inflation factors of the flight parameters, it is confirmed that multicollinearity exists among the flight parameters. Consequently, the ANN regression model is more suitable than the MLR model for load monitoring using flight test data. © 2024, International Council of the Aeronautical Sciences. All rights reserved.

키워드