基于多元特征驱动的既有建筑地震损伤机器学习预测方法OA北大核心CSTPCD

Rapid prediction of seismic damage of existing buildings is of paramount importance for post-earthquake emergency response and accelerating rescue and recovery in urban regions.To achieve rapid and quantitative diagnosis of the safety of existing buildings,a multi-feature driven method for rapid evaluation and prediction of seismic damage is proposed.For reinforced concrete structures,through measured dynamic properties in-situ and nonlinear performance parameters obtained from HAZUS manual,a numerical model of existing building is developed.Using the ATC-63 ground motion record sets and a data-driven nonlinear damage index,a large database is generated.The feature engineering is adopted to reveal the correlations between various input features.A random forest machine learning model is employed to predict the structural seismic damage quantitatively based on design features,measured structural features,and ground motion features.The coefficient of determination for the test set is 0.99,and the proportion of samples with a relative error within±20％is 99.23％.Model interpretability analyses reveal the importance of various input features on the output results,with structural modal periods being one of the most critical features.Finally,using the data from existing seismic stations,the model successfully predicts the damage condition of the examined existing buildings under a specific ground motion.The result indicates that the proposed framework of seismic damage prediction for existing building structures exhibits ideal predictive accuracy and efficiency.