Study on Building Damage Estimation of Various Deep Learning Models Using Seismic Response Analysis and Verification Using Large-Scale Shaking Table Test Result

Abstract

In recent years, the need for structural health monitoring has been increasing to confirm the safety of the building after the earthquake promptly. This study focuses on developing end-to-end deep learning models for estimating post-earthquake structural damage in buildings, aiming to apply them to a structural health monitoring system. The deep learning model takes story-level acceleration data as input and estimates post-earthquake story-level damage using a five-level scale. The dataset for this approach is created by conducting seismic response analyses on an 18-story steel building using 24,385 earthquake cases. The dataset includes story-level acceleration and story-level damage, each categorized into five levels based on hinge occurrence at each timestamp. Since the dataset consists of a time series with spatial relations at the story level, various architectures of deep learning models can be adopted. In the experiment, we evaluate the damage estimation capabilities of four different architectures: RNN, CNN, CRNN, and Transformer-based models. Initially, the deep learning models are trained and evaluated using training and test datasets, which are subsets of the dataset created by seismic response analysis. The Transformer-based model demonstrates the highest accuracy on the test dataset. Subsequently, the trained models are used to infer the results of a previous study¹⁰⁾ on the collapse process of an 18-story high-rise steel building based on the large-scale shaking table test. The results indicate that the estimations of the Transformer-based model differed from the experimental results, likely due to the influence of the analysis model's consistency range and the seismic level in the seismic response analysis.