Cost reference particle filter with multi-probability distribution for nonlinear dynamic systems with unknown noise statistics

Abstract

The lack of particle diversity is a vital problem in particle filtering for nonlinear dynamic systems with unknown noise statistics. In this paper, cost reference particle filter with multi-probability distribution (CRPF_MPD) is introduced to minimize the effects of the lack of particle diversity and to obtain more realistic state estimations. The core of multi-probability distribution is that two subsets are parallelly initialized and resampled by the processes of cost reference particle filter (CRPF) and are respectively updated with different distributions. Information interaction and particle selection form a set of particles with multi-probability distribution to estimate the more realistic states of systems. CRPF_MPD relies on multi-probability distribution, information interaction and particle selection to have excellent performance. As can be seen from the results of simulations, CRPF_MPDs can enhance the accuracy of state estimation and are more robust to different systems. Thus, CRPF_MPDs are available for dynamic systems with unknown noise statistics.