System Identification of a Highway Bridge from Earthquake-Induced Responses Using Neural Networks

Two back-propagation neural networks were applied to identify the stiffness coefficients of a highway bridge directly from its seismic responses without any eigenvalue analysis. An emulator neural network was trained to accurately predict the responses of a model structure that represents an as-built state of the highway bridge. A parameter evaluator neural network was trained to relate the response prediction error by the trained emulator neural network to the change in stiffness coefficients of the model structure, which represents various damage states of the highway bridge. An attempt is made to investigate the effect of various training data sets, or earthquake-induced structural responses, on the performance of a well trained emulator neural network. Furthermore, a new evaluation criterion, named weighted root-mean-square (RMS), is proposed for a more consistent performance measurement of the emulator neural networks that are trained under various earthquake excitations. Extensive parametric studies indicated that the two neural networks are very effective in the stiffness identification of highway bridges. An emulator neural network can be well trained with various seismic responses to give consistent performance measured by the proposed weighted RMS.