Using Ensemble Kalman Filter for Distributed Sensor Fusion

Abstract

Sensor fusion is a very powerful tool in various applications. Contrary to conventional methods using a single sensor, sensor fusions with multiple nodes (multiple sensors) give a state estimate from the distinct perspective of the nodes. By letting each node exchange data with its neighbors, a group of nodes form a sensor network, sensor fusion on each node can achieve more accurate estimation.However, in large-scale applications, a centralized sensor fusion which requires data from all nodes at once is not practical. On the other hand, using distributed or decentralized sensor fusion with conventional consensus algorithms which do not consider cross-covariance terms among nodes is not considerably efficient. We propose Ensemble Kalman Filter (EnKF) storing an estimation as a group of distinct particles to determine correlation between estimations. A cross-covariance can be easily obtained by simply approximating a sample second central cross-moment between estimations. Such a property of EnKF enables us to directly update a current estimation with the estimation from nearby nodes, even when the measurement is delayed. The proposed algorithm is proven to be globally optimal and iteratively stable even if redundancy of the estimation between nodes occurs.