ABSTRACT
Evaluation of structural health via modal characteristics is effective when using regression methods such as machine learning. In certain instances, training data is generated through numerical analysis to ensure a robust dataset. However, this approach may lead to reduced model accuracy when applied to real-world structures, a result of overfitting due to discrepancies between input data in training and prediction. In this paper, to deal with thickness estimation in steel members using local vibration modes sensitive to damage, a model construction method using data augmentation is proposed to suppress overfitting. First, finite element analysis and experimental tests on a steel member are performed to assess the differences in modal characteristics between the training and prediction phases. The constructed thickness estimation model is then applied to actual measurement data, showing that the proposed method outperforms conventional models without data augmentation by over 40%.
