Performance of unscented Kalman filter for model updating with experimental data

Abstract The unscented Kalman filter (UKF) is one of the most widely used algorithms for identifying and updating numerical model parameters. When updating from experimental data, the UKF performs well but it is sensitive to the selection of the initial algorithm variables and vulnerable to the influence of measurement noise. Furthermore, the ability to capture new behavioral features, such as hardening at large displacements, is a challenge. To achieve a more robust algorithm that can learn new features, techniques from the constrained UKF and the adaptive UKF are combined with an additional weighting on learning based on the magnitude of the input. The weighted adaptive constrained unscented Kalman filter (WACUKF) updates the numerical model with adaptive calculation of measurement noise and assigns a different learning rate for measurement data through the weighting function. To confine parameter values in the ranges with physical sense, the WACUKF adopts the sigma points projecting strategy for constrained parameters. Data from two experiments conducted on different structural components, in different laboratories, with different setups are used to validate the effectiveness of the WACUKF and compare against the performance of the UKF. The results demonstrate that the WACUKF can capture and preserve new features during model updating and is more robust and accurate than the UKF.