Premium
A deep learning approach to rapid regional post‐event seismic damage assessment using time‐frequency distributions of ground motions
Author(s) -
Lu Xinzheng,
Xu Yongjia,
Tian Yuan,
Cetiner Barbaros,
Taciroglu Ertugrul
Publication year - 2021
Publication title -
earthquake engineering and structural dynamics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 2.218
H-Index - 127
eISSN - 1096-9845
pISSN - 0098-8847
DOI - 10.1002/eqe.3415
Subject(s) - fragility , damages , computer science , event (particle physics) , convolutional neural network , seismology , reliability (semiconductor) , time domain , geology , artificial intelligence , power (physics) , chemistry , physics , quantum mechanics , political science , law , computer vision
Every year, earthquakes result in severe economic losses and a significant number of casualties worldwide. In limiting the losses that occur after these extreme events, timely and accurate assessment of seismic damages and mobilizing proportionate post‐event relief efforts play crucial roles. Traditional on‐site investigation generally results in prolonged evaluation windows. Several computational alternatives exist that show promise in addressing the downsides of the traditional approach. Damage estimates based on pre‐computed fragility libraries can provide near‐real time seismic damage quantification, but at present, they are coarse and involve considerable uncertainties. Estimates based on nonlinear time‐history analyses simulate the seismic response in greater detail, yet due to the computation and data requirements, their use at the regional scale is challenging. Given this perspective, herein, a rapid regional post‐event seismic damage assessment procedure based on convolutional neural network (CNN) is proposed. In this approach, an inventory of buildings, anticipated ground motion datasets, and corresponding damage levels for a region are brought together into a scenario bank. The time‐frequency distribution graphs of the ground motions, which serve as detailed visual representations of their frequency‐domain as well as time‐domain features, are generated. These data are then used to train CNN models, which could predict the damage states. The proposed methodology is verified through two numerical studies—one for an individual building, and the other, regional case, involving the buildings in the Tsinghua University campus. The results confirm that the proposed method offers prediction results with sufficient accuracy in near real‐time.