Fully distributed state estimation with multiple model approach

Abstract

In this paper, the problem of distributed state estimation using networked local sensors is studied. Our previously proposed algorithm [1], [2] is further extended to the scenario where the underlying model of the state of interest is not known to each agent. Instead the underlying model belongs to a finite set of possible models known to all agents, and switches over time. The switching process follows a homogeneous Markov chain with known transition probabilities. Two algorithms are derived from our previous algorithm by following the frameworks of two well-known multiple model (MM) approaches, namely, the first order generalized pseudo Bayesian and interacting MM approaches. The proposed algorithms have the advantages of being fully distributed and robust against agents not directly sensing the target. More importantly, they require the agents to communicate only once during each sampling interval and hence decrease the burdens in communication. It is also shown for a special case when the underlying model is fixed, all local agents asymptotically identify the true model under certain conditions.