Machine learning prediction models for ground motion parameters and seismic damage assessment of buildings at a regional scale

Abstract

This study examines the feasibility of using a machine learning approach for rapid damage assessment of reinforced concrete (RC) buildings after the earthquake. Since the real-world damaged datasets are lacking, have limited access, or are imbalanced, a simulation dataset is prepared by conducting a nonlinear time history analysis. Different machine learning (ML) models are trained considering the structural parameters and ground motion characteristics to predict the RC building damage into five categories: null, slight, moderate, heavy, and collapse. The random forest classifier (RFC) has achieved a higher prediction accuracy on testing and real-world damaged datasets. The structural parameters can be extracted using different means such as Google Earth, Open Street Map, unmanned aerial vehicles, etc. However, recording the ground motion at a closer distance requires the installation of a dense array of sensors which requires a higher cost. For places with no earthquake recording station/device, it is difficult to have ground motion characteristics. For that different ML-based regressor models are developed utilizing past-earthquake information to predict ground motion parameters such as peak ground acceleration and peak ground velocity. The random forest regressor (RFR) achieved better results than other regression models on testing and validation datasets. Furthermore, compared with the results of similar research works, a better result is obtained using RFC and RFR on validation datasets. In the end, these models are utilized to predict the damage categories of RC buildings at Saitama University and Okubo Danchi, Saitama, Japan after an earthquake. This damage information is crucial for government agencies or decision-makers to respond systematically in post-disaster situations.