A Data-driven Seismic Damage Assessment Framework of Regional Highway Bridges
Skip to main content
eScholarship
Open Access Publications from the University of California

A Data-driven Seismic Damage Assessment Framework of Regional Highway Bridges

  • Author(s): Wang, Dong
  • Advisor(s): Zhang, Jian
  • et al.
Abstract

Recent earthquake disasters have demonstrated the seismic vulnerability of highway bridge systems. Rapid seismic assessment of regional highway bridges is critical to help reduce severe loss of life and property. However, measurement of the regional scale system performance faces the challenge of dealing with the large uncertainty in structural properties and spatial characteristics. Traditionally, the numerical modeling approaches are established to simulate nonlinear response for each highway bridge across a regional portfolio. This process is largely limited by accuracy of model and computational effort. Especially some key structural component parameters are almost impossible to be retrieved for some ancient bridges. An alternative data-driven framework is developed to predict seismic responses or damage level of bridges using machine learning techniques. The proposed hierarchically structured framework enables a customized application in different scenarios. Firstly, the typical modeling technique for reinforcement concrete highway bridges is introduced using specific elements for different components. However, the modeling procedures are material-level parameter dependent and time consuming. The nonlinear analysis convergence is also a frustrating problem for numerical simulations. Due to these realistic limitations, a simple, fast and robust numerical model which can be developed with only component-level information needs to be adopted. It’s shown that the bridge bent representation can be simplified as a single degree of freedom system. The force-displacement relationship of the bridge can be roughly approximated by a bilinear curve. So a simplified 2D bilinear model is adopted for highway bridges throughout the study. Secondly, the statistical distributions for selected bridge input parameters can be derived based on the regional bridge inventory. Then an iterative process by sampling and filtering input parameters can be used to generate as many bridges as possible candidates for a specific region. The proposed bridge models and selected historical ground motions will be utilized to develop a seismic response prediction model using machine learning for instrumented highway bridges. This study investigates the optimal features to represent the highway bridge and ground motion. Different regression models are applied for near-fault motions and far-field motions and similar performance can be achieved, which significantly outperformed the traditional methods. Finally, to predict the seismic response of the non-instrumented highway bridges whose ground motion information is missing, the kriging interpolation model is implemented first. Then graph network is exploited to improve the performance. Different rules are evaluated for constructing an undirected graph for the highway bridges in an active seismic region. Subsequently, the Node2vec model is conducted to extract the embedding for each node and a graph neural network is implemented to predict the seismic response. Furthermore, vast amounts of text description data from online social platforms can be used to help detect the potential severely damaged bridges rapidly once an earthquake happens. A Convolution Neural Network classification model is implemented to evaluate the overall damage level distribution based on the collected text data. GloVe model is used to generate the word vector as its distributed representation.

Main Content
Current View