Bayesian inference of dense structural response using vision-based measurements

Abstract

As-built structures typically behave differently than their analytical models because of uncertainties in material properties, boundary conditions, loading scenarios, or other modeling assumptions. Therefore, to perform structural health monitoring effectively, analytical models need to be calibrated or updated to match the measured responses from as-built structures. Bayesian model updating provides a rigorous framework to integrate data and parameter uncertainty and has been applied successfully in numerous application settings employing traditional sensors that are often sparsely distributed. Computer vision has shown tremendous success for the measurement of static and dynamic displacements of structures. Vision-based measurements have the potential to be transformative for model updating of large-scale civil infrastructure, as measurements can be obtained for dense array of points within the camera frame using a single device. However, model updating using vision-based measurements to date has been limited to the use of a few points selected manually, leading to the investigation of simple structures or their global behaviors (e.g. first few modal responses); developing a systematic and highly automated procedure for estimating structural displacement and its uncertainty densely, registering the measured data to the finite element model, and then updating the model based on all the available data remains a challenge. This study proposes Bayesian inference using dense vision-based measurements to estimate the localized response of the entire structure. Finite element models are incorporated into the three-dimensional displacement measurement and registration process (model-informed approach), and then the Markov chain Monte Carlo method is applied to enable the Bayesian inference for large and complex civil infrastructure. The proposed method is applied to a laboratory-scale three-dimensional steel truss to perform inference of the overall system response based on the a posteriori knowledge. The results demonstrate the adequacy of performing Bayesian inference of structural response using vision-based measurements.