2023 Volume 79 Issue 11 Article ID: 23-00112
Structural health can be evaluated by modal characteristics, in which regression method such as machine learning is effective. There are cases where training data is generated by numerical analysis to acquire a sufficient amount of data. However, in this approach the accuracy of the model will decrease when using input data obtained from an actual structure. This is due to overfitting caused by the difference between input data in training and in predicting. In this paper, in dealing with thickness estimation on a steel member using local vibration modes which are sensitive to damages, model construction method using data augmentation is proposed to suppress the overfitting. First, finite element analysis and measurement test on a steel member are performed to evaluate the difference between the modal characteristics in training and in predicting. Second, using datasets generated by numerical analysis the performance of the models is evaluated according to the level of the difference of input data. Finally, thickness estimation model constructed by the proposed method is applied to the measurement data. It is proven that the proposed model performs better by more than 40 % compared with the result by the conventional model without data augmentation.
