Abstract "Fragility analysis is currently used to develop a probabilistic seismic demand model with an assumed lognormal distribution and then determine the probability of exceeding certain seismic demand thresholds for various states of damage. Due to the complexity in probability calculations, the seismic demand is often defined by one intensity measure of the earthquake ground motion such as peak ground acceleration, and its lognormal distribution has been repeatedly demonstrated inaccurate as the level of damage increases.
This study aims to develop artificial neural networks (ANNs) for a near real-time evaluation of the regional structural damage of a highway bridge network after a catastrophic earthquake. Bridge responses to the earthquake are treated as the earthquake-induced ground motion classifiers for structural damage states. The input and output layers of an ANN represent intensity measures of a ground motion and their corresponding damage state, respectively. To achieve this objective, the scope of work includes: (1) Select representative bridges along a major highway, (2) Collect and organize a big data set of ground motions, (3) Model the representative bridges and evaluate their damage states based on a damage index under the ground motions through time history analysis, (4) Label the ground motions with corresponding damage states and develop a balanced set of training and test data, (5) Train the ANN with the training dataset and evaluate the overall accuracy of damage prediction using unseen test dataset, (6) Optimize the ANN architecture for robust and accurate performance by ranking the importance of various intensity measures, comparing two structural damage indices, and considering varying numbers of hidden layers and neurons, and (7) Evaluate the performance of the ANNs for existing bridges along an emergency designated route with practical considerations of three intensity measures availed in the AASHTO Guide Specifications for LRFD Seismic Bridge Design."
Description This study aims to develop artificial neural networks (ANNs) for a near real-time evaluation of the regional structural damage of a highway bridge network after a catastrophic earthquake. Bridge responses to the earthquake are treated as the earthquake-induced ground motion classifiers for structural damage states. The input and output layers of an ANN represent intensity measures of a ground motion and their corresponding damage state, respectively. To achieve this objective, the scope of work includes: (1) Select representative bridges along a major highway, (2) Collect and organize a big data set of ground motions, (3) Model the representative bridges and evaluate their damage states based on a damage index under the ground motions through time history analysis, (4) Label the ground motions with corresponding damage states and develop a balanced set of training and test data, (5) Train the ANN with the training dataset and evaluate the overall accuracy of damage prediction using unseen test dataset, (6) Optimize the ANN architecture for robust and accurate performance by ranking the importance of various intensity measures, comparing two structural damage indices, and considering varying numbers of hidden layers and neurons, and (7) Evaluate the performance of the ANNs for existing bridges along an emergency designated route with practical considerations of three intensity measures availed in the AASHTO Guide Specifications for LRFD Seismic Bridge Design.
Impacts/Benefits Once successfully developed, the collected big dataset, design and performance data of the proposed ANNs will be disseminated through professional meetings (e.g., TRB annual meetings), national/international conferences, and archival journals. In particular, the testbed study with bridges along a major highway based on the AASHTO Specifications will be presented to transportation agencies.
Deliverables
Download the Final Report