Energy efficient geographic routing resilient to location errors

The thesis first analyses the importance sensor placement has in a large scale WSN application using geographic routing. A simulation-based topological study is made
for a forest fire prevention application using both deterministically and randomly placed nodes. Sensor deployment can be projectile, from the network edge, made
through manual scattering or by air release. Results reveal the impact of sensor distribution, density or destination location on the routing component.
Furthermore, geographic routing analysis focuses on location information assumptions. Because all methods of localisation are imprecise, it is necessary to consider
the use of estimated coordinates instead of the real ones and to first model the location errors as normally distributed. A more realistic evaluation of the routing
component requires the use of positioning simulations, considering received signal strength (RSS) and time of arrival (ToA) ranging for localisation (both modelled
in this thesis using the linear least square method (LLS) and maximum likelihood (ML) based Levenberg Marquardt (LM) method). Routing behaviour is analysed in terms of throughput, path lengths, energy consumption and failure causes. The energy expenditure of the two ranging methods is also analysed.
Efficient routing solutions for large scale WSNs are explored to cope with location error. A novel, low-complexity, error-resilient geographic routing method is proposed, namely the conditioned mean square error ratio (CMSER) algorithm. CMSER is compared to other progress only forwarding methods. A modified version of the algorithm is proposed to further increase energy efficiency and simulation results also confirm this. Furthermore, because CMSER is designed to make use of the Rice distribution (a statistical assumption valid only when the x and y coordinates of a node have the same location error variance) the precision of this approach is investigated. Although the routing behaviour is not severely affected by this simplifying assumption, because the variance of the errors can be very different in reality, a non-Rician version of the algorithm is proposed, which provides similar results under correct assumptions.