Nonlinear Kalman Filtering With Expensive Forward Models Via Measure Change

Abstract

Filtering is a subset of a more general probabilistic estimation scheme for estimating the unobserved parameters from the observed measurements. For nonlinear, high speed applications, the extended Kalman filter (EKF) and the unscented Kalman filter (UKF) are common estimators; however, expensive and strongly nonlinear forward models remain a challenge. In this paper, a novel Kalman filtering algorithm for nonlinear systems is developed, where the numerical approximation is achieved via a change of measure. The accuracy is identical in the linear case and superior in two nonlinear test problems: a challenging 1D benchmarking problem and a 4D structural health monitoring problem. This increase in accuracy is achieved without the need for tuning parameters, rather relying on a more complete approximation of the underlying distributions than the Unscented Transform. In addition, when expensive forward models are used, we achieve a significant reduction in computational cost without resorting to model approximation.