This work presents an investigation towards a new framework for structural health monitoring of buildings and structures subjected to known loads with a limited number of measurements. The study addresses the problem of uncertainty quantification for structural parameters and dynamic response predictions in a given system through Bayesian inference models. An inquiry is made into well-known probability models used in the past and novel machine learning approaches developed recently. Several classes of probability models are explored and tested on applicability to the problem. In this work, emphasis is placed on an investigation of the correlation structure of the prediction errors in order to improve the accuracy of posterior uncertainty levels of the inferred structural parameters and...[ Read more ]

This work presents an investigation towards a new framework for structural health monitoring of buildings and structures subjected to known loads with a limited number of measurements. The study addresses the problem of uncertainty quantification for structural parameters and dynamic response predictions in a given system through Bayesian inference models. An inquiry is made into well-known probability models used in the past and novel machine learning approaches developed recently. Several classes of probability models are explored and tested on applicability to the problem. In this work, emphasis is placed on an investigation of the correlation structure of the prediction errors in order to improve the accuracy of posterior uncertainty levels of the inferred structural parameters and reliably propagate uncertainties into the computational model responses. The Gaussian Process Regression (GPR) approach, taken from a domain of supervised machine learning techniques was studied in detail and showed potential to improve inference on structural parameters with reasonable credible intervals for the studied discrepancy signals. For that reason, a new Bayesian framework equipped with the GPR approach was developed and a novel kernel-covariance function derived from the autoregressive process of the discrepancy signal is introduced. Further on, to enhance the scheme and mitigate the computation burdens, a hierarchical approach to the problem is developed and presented in this work. The proposed methodology further enhances state-of-the-art hierarchical Bayesian methods, offering improvements in the accuracy and robustness of the parameter estimates and predicted responses. Numerical simulations and experimental examples are provided to validate the proposed methods and conclusions on the improvements are drawn. As an outcome, suggested probability models provide a unique insight into the correlation structure of the predictions of the dynamic response of a structural system and realistic uncertainty levels around estimated structural parameters. In addition, the methodologies developed in this thesis have significant potential to be applied to other disciplines of engineering and science, contributing to unravelling the problem of system parameters identification.