On the Development of Distributed Estimation
Techniques for Wireless Sensor Networks

Abstract

Wireless sensor networks (WSNs) have lately witnessed tremendous demand, as evidenced by the increasing number of day-to-day applications. The sensor nodes aim at estimating the parameters of their corresponding adaptive filters to achieve the desired response for the event of interest. Some of the burning issues related to linear parameter estimation in WSNs have been addressed in this thesis mainly focusing on reduction of communication overhead and latency, and robustness to noise. The first issue deals with the high communication overhead and latency in distributed parameter estimation techniques such as diffusion least mean squares (DLMS) and incremental least mean squares (ILMS) algorithms. Subsequently the poor performance demonstrated by these distributed techniques in presence of impulsive noise has been dealt separately. The issue of source localization i.e. estimation of source bearing in WSNs, where the existing decentralized algorithms fail to perform satisfactorily, has been resolved in this thesis. Further the same issue has been dealt separately independent of nodal connectivity in WSNs.
This thesis proposes two algorithms namely the block diffusion least mean squares (BDLMS) and block incremental least mean squares (BILMS) algorithms for reducing the communication overhead in WSNs. The theoretical and simulation studies demonstrate that BDLMS and BILMS algorithms provide the same performances as that of DLMS and ILMS, but with significant reduction in communication overheads per node. The latency also reduces by a factor as high as the block-size used in the proposed algorithms.
With an aim to develop robustness towards impulsive noise, this thesis proposes three robust distributed algorithms i.e. saturation nonlinearity incremental LMS (SNILMS), saturation nonlinearity diffusion LMS (SNDLMS) and Wilcoxon norm diffusion LMS (WNDLMS) algorithms. The steady-state analysis of SNILMS algorithm is carried out based on spatial-temporal energy conservation principle. The
theoretical and simulation results show that these algorithms are robust to impulsive noise. The SNDLMS algorithm is found to provide better performance than
SNILMS and WNDLMS algorithms.
In order to develop a distributed source localization technique, a novel diffusion maximum likelihood (ML) bearing estimation algorithm is proposed in this thesis which needs less communication overhead than the centralized algorithms. After forming a random array with its neighbours, each sensor node estimates the source bearing by optimizing the ML function locally using a diffusion particle swarm optimization algorithm. The simulation results show that the proposed algorithm performs better than the centralized multiple signal classification (MUSIC) algorithm in terms of probability of resolution and root mean square error. Further, in order to make the proposed algorithm independent of nodal connectivity, a distributed in-cluster bearing estimation technique is proposed. Each cluster of sensors
estimates the source bearing by optimizing the ML function locally in cooperation with other clusters. The simulation results demonstrate improved performance of the proposed method in comparison to the centralized and decentralized MUSIC algorithms, and the distributed in-network algorithm