Identification of boundary conditions of railway bridges using artificial neural networks

Abstract

This article presents a study that aims to identify the boundary conditions of a railway bridge using system identification and artificial neural networks. Vibrations generated by three different train types recorded during a 24-h long measurement campaign is used to identify the modal frequencies and mode shapes of a single-span 50 m long railway bridge. Frequency Domain Decomposition and Stochastic Subspace Identification with Covariance methods were used to identify the modal properties from the recorded vibrations and the effect of the used Operational Modal Analysis on the identified modal properties was evaluated. An initial finite-element (FE) model based on the design drawings was not able to replicate the observed dynamic behavior of the bridge. Using a sensitivity analysis, the key parameters of the finite-element model that impact the vibration frequencies of the bridge was determined. 300 finite-element models were created by changing the values of these key parameters within their effective range and were used to identify the relationship between these parameters and the vibration frequencies using Artificial neural networks (ANNs). Leveraging this relationship, the values of the FE model parameters that minimizes the error between the measured and computed frequencies was determined. As a result, the mean error between the computed and the identified vibration frequencies was reduced from 27.3% for the initial model to 3.0% for the updated model. The study indicates that boundary conditions are among the most influential parameter on the dynamic behavior of bridges and can deviate significantly from the simplistic models generally used in the FE models.