Machine Learning-Based Finite Element Model Updating Based on Ground Vibration Test Data

This study proposes an efficient finite element model updating technique at the aircraft airframe level. Traditional machine learning-based updating methods suffer from excessive sampling requirements and high computational costs, and these inefficiencies are exacerbated when dealing with high-dimensional data or complex models. To address these challenges, the study introduces two finite element model updating strategies. The first is a deep neural network-based surrogate modeling approach that integrates sensitivity analysis and continual learning. The second is a Bayesian optimization method that leverages uncertainty-based sampling. The proposed techniques aim to minimize the number of simulations required to predict natural frequencies and mode shapes while maintaining high accuracy. The framework was quantitatively validated using ground vibration test data from a tiltrotor aircraft, demonstrating high prediction accuracy for natural frequencies and strong mode shape correlation. In particular, the Bayesian optimization approach achieved comparable precision with over 70% lower computational cost compared to the deep neural network-based method, proving its practical advantages in high-dimensional design parameter spaces. This work presents a generalized and efficient machine learning-based finite element model updating strategy that can be readily applied to structural verification tasks in aerospace and other engineering domains.