| Recent advances in sensor design and miniaturization have provided the opportunity for the creation of large distributed wireless sensor networks. There has been significant progress in combining sensing, processing, data storage and communications capabilities, in network self organization, and in optimizing communication architectures. In contrast to most other network applications, wireless sensor networks face a number of special challenges and constraints resulting from (1) lack of hardwired connections (no external power sources, low communications bandwidths, higher communication error rates), (2) small physical size (small onboard energy supply, small antennas/acoustic transducers, small low energy sensors) and (3) elevated sensor node failure rates.; One of the key remaining challenges is in the area of inference and information fusion (aggregating/filtering/interpreting the sensor data into useful high level knowledge). Many authors have advocated the use of local distributed inference and fusion algorithms such as the iterative message-passing belief propagation algorithms employed on probabilistic graphical models (Bayesian Networks and Markov Random Fields). However, little research has been performed to assess the performance of these algorithms under the special constraints imposed by wireless sensor network applications.; This dissertation reports on a study investigating issues associated with application of these algorithms to realistic wireless sensor networks configurations. This research has produced results delineating the performance and limitations including communications requirements, energy resource requirements and the impacts of different topologies and architectures such as hierarchical/non-hierarchical topologies, centralized or distributed processing, localized (in local node clusters) or full network models, and node cluster size. |
