Title:

Optimised localisation in wireless sensor networks

Wireless sensor networks (WSNs) comprise of tens, hundreds or thousands of low powered, low cost wireless nodes, capable of sensing environmental data such as humidity and temperature. Other than these sensing abilities, these nodes are also able to locate themselves. Different techniques can be found in literature to localise wireless nodes in WSNs. These localisation algorithms are based on the distance estimates between the nodes, the angle estimates between the nodes or hybrid schemes. In the context of range based algorithms, two prime techniques based on the time of arrival (ToA) and the received signal strength (RSS) are commonly used. On the other hand, angle based approach is based on the angle of arrival (AoA) of the signal. A hybrid approach is sometimes used to localise wireless nodes. Hybrid algorithms are more accurate than range and angle based algorithms because of additional observations. Modern WSNs consist of a small group of highly resourced wireless nodes with known locations called anchor nodes (ANs) and a large group of low resourced wireless nodes known as the target nodes (TNs). The ANs can locate themselves through GPS or they may have a predetermined location given to them during network deployment. Based on these known locations and the range/angle estimates, the TNs are localised. Since hybrid algorithms (a combination of RSS, ToA and AoA) are more accurate than other algorithms, a major portion of this thesis will focus on these approaches. Two prime hybrid signal models are discussed: i) The AoARSS hybrid model and ii) the AoAToA hybrid signal model. A hybrid AoAToA model is first studied and is further improved by making the model unbiased and by developing a new weighted linear least squares algorithm for AoAToA signal (WLLSAoAToA) that capitalise on the covariance matrix of the incoming signal. A similar approach is taken in deriving a WLLS algorithm for AoARSS signal (WLLSAoARSS). Moreover expressions of theoretical mean square error (MSE) of the location estimate for both signal models are derived. Performances of both signal models are further improved by designing an optimum anchor selection (OAS) criterion for AoAToA signal model and a two step optimum anchor selection (TSOAS) criterion for AoARSS signal model. To bound the performance of WLLS algorithms linear Cramer Rao bounds (LCRB) are derived for both models, which will be referred to as LCRBAoAToA and LCRBAoARSS, for AoAToA and AoARSS signal models, respectively. These hybrid localisation schemes are taken one step further and a cooperative version of these algorithms (LLSCoop) is designed. The cooperation between the TNs significantly improves the accuracy of final estimates. However this comes at a cost that not only the ANs but the TNs must also be able to estimate AoA and ToA/RSS simultaneously. Thus another version of the same cooperative model is designed (LLSCoopX) which eliminates the necessity of simultaneous anglerange estimation by TNs. A third version of cooperative model is also proposed (LLSOptCoop) that capitalises the covariance matrix of incoming signal for performance improvement. Moreover complexity analysis is done for all three versions of the cooperative schemes and is compared with its non cooperative counterparts. In order to extract the distance estimate from the RSS the correct knowledge of pathloss exponent (PLE) is required. In most of the studies this PLE is assumed to be accurately known, also the same and fixed PLE value is used for all communication links. This is an oversimplification of real conditions. Thus error analysis of location estimates with incorrect PLE assumptions for LLS technique is done in their respective chapters. Moreover a mobile TN and an unknown PLE vector is considered which is changing continuously due to the motion of TN. Thus the PLE vector is first estimated using the generalized pattern search (GenPS) followed by the tracking of TN via the Kalman filter (KF) and the particle filter (PF). The performance comparison in terms of root mean square error (RMSE) is also done for KF, extended Kalman filter (EKF) and PF.
