Bayesian history matching for forward model-driven structural health monitoring

Abstract

Computer models are widely utilised in many structural dynamics applications, however their use depends on calibration to observational data. A complexity in calibrating a computer model is that even when the ‘true’ input parameters to the model are known, there may be model discrepancy caused by the simplification or absence of certain physics. As a consequence the calibration technique employed must incorporate a mechanism for dealing with model discrepancy. Bayesian history matching is a process of using observed data in order to identify and discard areas of the computer model’s parameter space that will result in outputs that are unlikely given the observational data. This is performed using an implausibility metric that encompasses uncertainties associated with observational measurements and model discrepancy. The method employs this metric to identify a non-implausible space (i.e., parameter combinations that are likely to have produced the observed outputs). A maximum a posterior (MAP) approach can be used to obtain the calibrated computer model outputs from the non-implausible space. Model discrepancy between the calibrated computer model and observational data can then be inferred using a Gaussian process (GP) regression model. This paper applies Bayesian history matching in order to calibrate a computer model for forward model-driven structural health monitoring (SHM). Quantitative metrics are used to compare experimental and predictive damage features from the combined Bayesian history matching and GP approach.