Date of Award

May 2014

Degree Type


Degree Name

Doctor of Philosophy (PhD)


Electrical Engineering and Computer Science


Pramod k. Varshney

Second Advisor

Ruixin Niu


Bayesian estimation, Kalman filtering, Particle filtering, Sensor networks, Target tracking

Subject Categories



Distributed inference arising in sensor networks has been an interesting and promising discipline in recent years. The goal of this dissertation is to investigate several issues related to distributed inference in sensor networks, emphasizing parameter estimation and target tracking with resource-constrainted networks.

To reduce the transmissions between sensors and the fusion center thereby saving bandwidth and energy consumption in sensor networks, a novel methodology, where each local sensor performs a censoring procedure based on the normalized innovation square (NIS), is proposed for the sequential Bayesian estimation problem in this dissertation. In this methodology, each sensor sends only the informative measurements and the fusion center fuses both missing measurements and received ones to yield more accurate inference. The new methodology is derived for both linear and nonlinear dynamic systems, and both scalar and vector measurements. The relationship between the censoring rule based on NIS and the one based on Kullback-Leibler (KL) divergence is investigated.

A probabilistic transmission model over multiple access channels (MACs) is investigated. With this model, a relationship between the sensor management and compressive sensing problems is established, based on which, the sensor management problem becomes a constrained optimization problem, where the goal is to determine the optimal values of probabilities that each sensor should transmit with such that the determinant of the Fisher information matrix (FIM) at any given time step is maximized. The performance of the proposed compressive sensing based sensor management methodology in terms of accuracy of inference is investigated.

For the Bayesian parameter estimation problem, a framework is proposed where quantized observations from local sensors are not directly fused at the fusion center, instead, an additive noise is injected independently to each quantized observation. The injected noise performs as a low-pass filter in the characteristic function (CF) domain, and therefore, is capable of recoverving the original analog data if certain conditions are satisfied. The optimal estimator based on the new framework is derived, so is the performance bound in terms of Fisher information. Moreover, a sub-optimal estimator, namely, linear minimum mean square error estimator (LMMSE) is derived, due to the fact that the proposed framework theoretically justifies the additive noise modeling of the quantization process. The bit allocation problem based on the framework is also investigated.

A source localization problem in a large-scale sensor network is explored. The maximum-likelihood (ML) estimator based on the quantized data from local sensors and its performance bound in terms of Cram\'{e}r-Rao lower bound (CRLB) are derived. Since the number of sensors is large, the law of large numbers (LLN) is utilized to obtain a closed-form version of the performance bound, which clearly shows the dependence of the bound on the sensor density, $i.e.,$ the Fisher information is a linearly increasing function of the sensor density. Error incurred by the LLN approximation is also theoretically analyzed. Furthermore, the design of sub-optimal local sensor quantizers based on the closed-form solution is proposed.

The problem of on-line performance evaluation for state estimation of a moving target is studied. In particular, a compact and efficient recursive conditional Posterior Cram\'{e}r-Rao lower bound (PCRLB) is proposed. This bound provides theoretical justification for a heuristic one proposed by other researchers in this area. Theoretical complexity analysis is provided to show the efficiency of the proposed bound, compared to the existing bound.


Open Access

Included in

Engineering Commons