Study On Distributed Particle Filter Algorithm Based On Consensus

Distributed Particle Filter(DPF) are a powerful and versatile approach to decentralized state estimation in wireless sensor network(WSN), they are especially suited to large-scale, nonlinear and non-Gaussian distributed estimated systems. However, it has certain limitations on the bandwidth of communication, the resource of nodes, the dynamic network topology and the ability of communication and so on. In order to reduce influences of the links failure or intermittent link that exist in the network for estimating the target, the deep study of distributed particle filter algorithm has a great significance.In view of non-linear, non-Gaussian tracking application in sensor networks, the paper proposes a consensus/fusion based distributed implementation of the particle filter(CF/DPF). It runs two particle filters(PFs) at each node, they are respectively local filter which comes from the distributed implementation of the particle filter and the fusion filter which computes the global filtering distribution. The paper approximates the product of the local probability density functions with Gaussian distribution, and the average consensus algorithm is used to calculate the parameters of the Gaussian distribution, then realize the target state estimation. this algorithm and the tracking performance of centralized particle filtering will be finally compared With Monte Carlo simulation, they show that the algorithm has a better filtering performance.Due to the convergence rate of consensus is very crucial in the above algorithm, so the paper focuses on probability-based the weight optimization method of consensus. The method introduces the weights optimization problem in consensus algorithms for spatially correlated random topologies, it chooses the consensus mean-square error(MSE) convergence rate as the optimization criterion and expresses this rate as a function of the link formation probabilities, the link formation spatial correlations and the consensus weights. Because the MSE convergence rate is a convex, non-smooth function of the weight for the symmetric random networks, the paper gives the closed form and sub-gradient algorithm solution to solve the problem. and the optimization method is compared with other weight selection method by the simulation, results show that the weight design has a significant performance gain.In conclusion, based on the intermittent network links in the network, probability-based the weight optimization method of consensus is combined with consensus-based distributed particle filter algorithm, by overcoming the delay problem of convergence of the consensus algorithm that the links failure lead to, thus improving the tracking effect of consensus based distributed particle filter algorithm. it is compared with the centralized particle filter and consensus-based distributed particle filter algorithm by the simulation analysis, simulations verify that the algorithm is close to the optimal tracking performance.