Identifying and Locating Earthquake-Induced Damage in a High-Rise Using Neural Operators

Abstract

Rapid response after earthquakes is vital to mitigate the effects of catastrophic structural failures and to save lives. In particular, critical infrastructure relies on accurate yet low-latency damage detection to facilitate timely responses. Although traditional techniques rely on hand-crafted feature extraction and data interpretation to identify the presence of damage within a structure, deep learning has resulted in newer methods that combine feature extraction and data interpretation in one. However, these models often require exorbitant amounts of data to train, are often computationally expensive, and are difficult to cross-validate with established theory due to their black box nature. In this paper, we present our current workflow that addresses the aforementioned challenges: (i) constructing representative finite-element models of a 15-story steel frame building with simulated damage pattern scenarios in the form of weld fractures, (ii) generating a synthetic waveform dataset for earthquake-induced dynamic response using a well-known dynamic analysis simulator, OpenSees, for the finite-element models and (iii) implementing a deep learning model that performs damage identification in addition to damage localization, i.e., locating the damage. To construct the dataset, we apply recorded earthquake ground motions to the 15-story building model modified to include small-scale damage scenarios imposed at beam-column connections. We then compute accelerations at each of the four corners within the structure for each individual modified model and earthquake input. Our deep learning model consists of a neural operator in conjunction with a machine learning mechanism called self-attention. To our knowledge, our work is the first work exploring the use of neural operators for earthquake-specific vibration-based damage identification. In addition, we present an analysis of our model performance and model fit, and compare our model to a subset of machine learning
techniques.

Additional details