System Identification via Unscented Kalman Filtering and Model Class Selection

Abstract

Identifying the mechanical properties of civil structures is required for life-cycle assessment. Kalman filters are exploited for this goal, enabling the online update of a numerical model, acting as the digital twin of the structure, and quantifying the uncertainty of the estimated properties. As uncertainty about model formulation is usually disregarded in the identification, model class evidence has been recently formulated to compare different parametrizations of the properties of the monitored structure through a metric, allowing selection of the most plausible one. When dealing with parameter estimation, predominantly model evidence is deployed in batch Bayesian estimation. Here, the formulation of model class evidence is proposed for the unscented Kalman filter, which allows online calculation of model class evidence for a system without the need to compute the mapping gradient in time. This formulation was inspired by the model class evidence developed for the extended Kalman filter. Numerical results related to shear buildings are presented to validate the metric, showing the impact of under- and over-parametrizations on identification.