Document Type

Journal Article

Publication Date

2023

Abstract

This paper aims to investigate the use of machine learning algorithms in vulnerability assessment of buildings and structures. Traditionally, dynamic performance of buildings under earthquakes is determined by means of non-linear time-history analyses. This method is accurate, but it is known as a time-consuming process and it requires advance knowledge on modelling. As an alternative, responses of building under earthquakes can be obtained using well-trained machine learning models. Nowadays, this is also called a data-driven approach. In the current study, machine learning models for damage class prediction are developed using the building data generated from numerous incremental dynamic analyses. Building models with variety of ground and structural parameters, such as peak ground acceleration, peak ground velocity, aspect ratio of building, member’s size, axial load ratio, etc., are considered in this study as the input parameters. Different past earthquake histories with increasing peak ground acceleration are used in the study. Material non-linearity is modelled using plastic hinge model, while geometric non-linearity is considered using P-delta effects. The effectiveness of typical machine learning models, including ensemble models and deep learning via artificial neural network, is investigated. The results reveal that well-prepared machine learning models are also capable of predicting structural response and damage level with adequate accuracy and minimum computation efforts. The performance of Random Forest and XGBoost is generally better. Other possible applications of machine learning models have been investigated as well in the study.

Source Publication

The 2023 World Congress on Advances in Structural Engineering and Mechanics (ASEM23), Seoul National University, Seoul, Korea, Aug 2023

Download

Share

COinS