Incorporating link correlations in models and algorithms for localization in wireless sensor networks

Download File (3.22M)
Publication Type dissertation
School or College College of Engineering
Department Electrical & Computer Engineering
Author Agrawal, Piyush
Title Incorporating link correlations in models and algorithms for localization in wireless sensor networks
Date 2012-05
Description In wireless sensor networks, knowing the location of the wireless sensors is critical in many remote sensing and location-based applications, from asset tracking, and structural monitoring to geographical routing. For a majority of these applications, received signal strength (RSS)-based localization algorithms are a cost effective and viable solution. However, RSS measurements vary unpredictably because of fading, the shadowing caused by presence of walls and obstacles in the path, and non-isotropic antenna gain patterns, which affect the performance of the RSS-based localization algorithms. This dissertation aims to provide efficient models for the measured RSS and use the lessons learned from these models to develop and evaluate efficient localization algorithms. The first contribution of this dissertation is to model the correlation in shadowing across link pairs. We propose a non-site specific statistical joint path loss model between a set of static nodes. Radio links that are geographically proximate often experience similar environmental shadowing effects and thus have correlated shadowing. Using a large number of multi-hop network measurements in an ensemble of indoor and outdoor environments, we show statistically significant correlations among shadowing experienced on different links in the network. Finally, we analyze multihop paths in three and four node networks using both correlated and independent shadowing models and show that independent shadowing models can underestimate the probability of route failure by a factor of two or greater. Second, we study a special class of algorithms, called kernel-based localization algorithms, that use kernel methods as a tool for learning correlation between the RSS measurements. Kernel methods simplify RSS-based localization algorithms by providing a means to learn the complicated relationship between RSS measurements and position. We present a common mathematical framework for kernel-based localization algorithms to study and compare the performance of four different kernel-based localization algorithms from the literature. We show via simulations and an extensive measurement data set that kernel-based localization algorithms can perform better than model-based algorithms. Results show that kernel methods can achieve an RMSE up to 55% lower than a model-based algorithm. Finally, we propose a novel distance estimator for estimating the distance between two nodes a and b using indirect link measurements, which are the measurements made between a and k, for k ? b and b and k, for k ? a. Traditionally, distance estimators use only direct link measurement, which is the pairwise measurement between the nodes a and b. The results show that the estimator that uses indirect link measurements enables better distance estimation than the estimator that uses direct link measurements.
Type Text
Publisher University of Utah
Subject Kernel methods for localization; Localization; Network distance estimation; Shadowing; Correlations; Statistical channel modeling; Wireless sensor ad hoc networks; Linkers (Computer programs); Wireless sensor networks
Dissertation Institution University of Utah
Dissertation Name Doctor of Philosophy
Language eng
Rights Management Copyright © Piyush Agrawal 2012
Format Medium application/pdf (portable document)
Format Extent 3,570,971 bytes
Identifier us-etd3,84729
Source Original in Marriott Library Special Collections, QA3.5 2012 .A37
ARK ark:/87278/s6q24dzr
ID 194343
setname ir_etd
Reference URL