GENERATING MACHINE LEARNING DATASETS ON DAMAGE IDENTIFICATION USING FINITE ELEMENT BRIDGE MODEL

Abstract

 Recently, machine learning (ML) has been actively applied to various problems in the civil engineering field. To further facilitate the use of ML in the civil engineering field, it is essential to have appropriate benchmark problems and training datasets with the characteristics of the civil engineering field. Nevertheless, such datasets have not yet been proposed sufficiently. In this study, using the finite element analysis, we propose fundamental datasets (no noise and no missing data) as benchmark problems for damage identification of bridge model, with four levels of difficulty. Then, we input the dataset into a total of 19 ML algorithms to assess the quality of the dataset using the coefficient of determination obtained from those algorithms. As a result of numerical experiments, the following points were found. For cases with a single damaged member, the best coefficient of determination for most of the members exceeded 0.5, resulting in suitable benchmarks with appropriate difficulty. For cases with two damaged members, the best coefficient of determination for almost half of the members was below 0.5, resulting in a challenging benchmark problem from the viewpoint of its high difficulties.