Date of Award

Spring 5-15-2016

Author's School

School of Engineering & Applied Science

Author's Department

Electrical & Systems Engineering

Degree Name

Doctor of Philosophy (PhD)

Degree Type

Dissertation

Abstract

Wireless sensor networks provide useful information for various applications but pose challenges in scalable information processing and network maintenance. This dissertation focuses on statistical methods for distributed information fusion and sensor synchronization for target tracking in wireless sensor networks.

We perform target tracking using particle filtering. For scalability, we extend centralized particle filtering to distributed particle filtering via distributed fusion of local estimates provided by individual sensors. We derive a distributed fusion rule from Bayes' theorem and implement it via average consensus. We approximate each local estimate as a Gaussian mixture and develop a sampling-based approach to the nonlinear fusion of Gaussian mixtures. By using the sampling-based approach in the fusion of Gaussian mixtures, we do not require each Gaussian mixture to have a uniform number of mixture components, and thus give each sensor the flexibility to adaptively learn a Gaussian mixture model with the optimal number of mixture components, based on its local information. Given such flexibility, we develop an adaptive method for Gaussian mixture fitting through a combination of hierarchical clustering and the expectation-maximization algorithm. Using numerical examples, we show that the proposed distributed particle filtering algorithm improves the accuracy and communication efficiency of distributed target tracking, and that the proposed adaptive Gaussian mixture learning method improves the accuracy and computational efficiency of distributed target tracking.

We also consider the synchronization problem of a wireless sensor network. When sensors in a network are not synchronized, we model their relative clock offsets as unknown parameters in a state-space model that connects sensor observations to target state transition. We formulate the synchronization problem as a joint state and parameter estimation problem and solve it via the expectation-maximization algorithm to find the maximum likelihood solution for the unknown parameters, without knowledge of the target states. We also study the performance of the expectation-maximization algorithm under the Monte Carlo approximations used by particle filtering in target tracking. Numerical examples show that the proposed synchronization method converges to the ground truth, and that sensor synchronization significantly improves the accuracy of target tracking.

Language

English (en)

Chair

Arye Nehorai

Committee Members

Arye Nehorai, Martin Arthur, Nan Lin, Hiro Mukai,

Comments

Permanent URL: https://doi.org/10.7936/K7FB5175

Share

COinS