ScholarWorks@경상국립대학교

초록

To mitigate blast-induced progressive collapses of building structures, blast damage assessment of main structural elements (e.g., column) is crucial. However, field tests and numerical simulations for evaluating blast resistant performance have been expensive and time-consuming. Due to these limitations, many researchers have developed machine-learning models. The model have been learned from a large amount of experiments and numerical simulation-based dataset, which required expensive computational time. This paper presents a novel machine learning approach trained and tested from a reduced dataset to predict blast resistant performance for RC columns. A multi-step machine learning model integrating two distinct models was established as follows: (1) prediction of column failure modes (shear & flexure failure) utilized as the input to the second model, and (2) prediction of blast-induced damage levels for the RC column. A learning dataset associated with the blast column damage was generated from the finite element simulations validated with the previous experimental results. The numerical simulation-based dataset varies with simple column details (longitudinal and transverse rebar details, and axial loading ratio) and blast loading scenarios (scaled distance). To resolve the limitation of the conventional learning models, the reduced dataset with 200 data points was utilized to develop best-fit models for each column damage level, and their models were combined using four different combination methods: (1) sequential prediction method (method-1), (2) maximum positive probability prediction method (method-2), (3) maximum negative probability prediction method (method-3), and (4) combined method between methods-1 and 2 (method-4). Among them, the combined method has the highest prediction performance. As compared to the general method model trained from a large amount of dataset (703 data), the proposed combination method (method-4) can reduce the data points by 71.5 % and enhance the average of accuracy for each blast damage level by 14.3 %.

키워드