A data-driven approach for predicting peak floor response based on visually observed rocking behaviors of freestanding NSCs

Abstract

Accurate prediction of structural responses under earthquakes is crucial for seismic performance evaluation. Traditional methods of obtaining structural response mainly rely on costly structural health monitoring (SHM) systems while paying little attention to the damage state of unanchored non-structural components (NSCs). The surveillance system is commonly equipped in commercial and public buildings, which could be used to capture the response motions and damage states of NSCs during earthquakes. To this end, this study aims to develop a method for inferring peak floor acceleration (PFA) based on the observed seismic response of NSCs. Three computer vision tasks for collecting responses of NSCs with different data ambiguity are considered. For a purpose of the method implementation, three prototype structures with different heights are used in this study. Freestanding NSCs with different geometric properties are considered to be placed in the structures. Under a synthesis of ground motions, the datasets for floor acceleration responses of the structures and dynamic responses of freestanding NSCs are obtained via numerical simulations. This study finds out that regression models for predicting PFA values are untrustworthy due to the weak correlation between PFAs and response quantities of NSCs. Instead of predicting exact PFA values, the potential PFA ranges are considered to be predicted and a list of PFA ranges is determined based on rocking fragility models of freestanding NSCs. Machine learning techniques are employed to build the surrogate models and the results demonstrate the accuracy and interpretability of the PFA range prediction with the $F1‐score$ ranging from 84 % to 94 %.