Comparative Study on CNN-based Bridge Seismic Damage Identification Using Various Features

Abstract

Quick and accurate identification of bridge damage after an earthquake is crucial for emergency decision-making and post-disaster rehabilitation. The maturing technology of deep neural networks (DNN) and the integration of health monitoring systems provide a viable solution for seismic damage identification in bridges. However, how to construct damage features that can efficiently characterize the seismic damage of the bridge and are suitable for the use with DNN needs further investigation. This study focuses on seismic damage identification for a continuous rigid bridge using raw acceleration responses, statistical features, frequency features, and time-frequency features as inputs, with damage states as outputs, employing a deep convolutional neural network (CNN) for pattern classification. Results indicate that all four damage features can identify seismic damage, with time-frequency features achieving the highest accuracy but having a complex construction process. Frequency features also demonstrate high accuracy with simpler construction. Raw acceleration response and statistical features perform poorly, with statistical features deemed unsuitable as damage indicators. Overall, frequency features are recommended as CNN inputs for quick and accurate bridge seismic damage identification.