On the effects of modelling errors and uncertainties in structural damage identification

For structural response prediction, structural health monitoring and structural control, identification of an accurate and representative model of the target structural system is essential. Modelling errors and uncertainties resulting from the simplification of complex structures, discretization of structural systems, inadequate theories for certain structural behaviours and inaccurate assignment of model parameters are inevitable. To address these issues, structural model identification and updating techniques based on measured system data have been developed. When several classes of models are available to represent structural behaviours, making a technical and appropriate choice is nontrivial. In this research, an MCMCS-based Bayesian probabilistic technique is employed to fully consider the uncertainties in updating model parameters, and to select the most plausible model class from available ones using the Bayesian model class selection. Detection of ballast damage of an in-situ sleeper is employed to demonstrate the proposed methodology. Impact hammer test was carried out on a sleeper of an indoor track panel, with damage simulated in the ballast at the right two-third section to obtain in-situ vibration measurement. The proposed damage identification methodology is then applied to predict the location and severity of the ballast damage. Since ballast has been proven to possess non-linear stress-strain characteristics, the effects of modelling errors and uncertainties on the predicted structural response and identification accuracy is then demonstrated by modelling the ballast using three different classes of models: (1) the linear elastic model, (2) the non-linear modified Ludwik’s model, and (3) the non-linear Prager model. Finally, the Bayesian model class selection is used to identify the optimal ballast model.