Wireless Sensor Networks (WSNs), consisting of many small sensing devices working in concert, have the potential to revolutionise every aspect of our lives. Although the technology is still in its infancy offers an unlimited number of possible applications, ranging from military surveillance to environmental monitoring. These WSNs are prone to physical sensor failures due to environmental conditions such bio fouling and an adverse ambient environment, as well as threats that arise from their operation in an open environment. Consequently, reliability and fault-tolerance techniques become a critical aspect of the research associated with WSNs. In mission critical applications, such as the monitoring of enemy troops, unreliable or faulty information produced by WSNs could potentially lead to fatal outcomes. In such applications, it necessary to receive both a correct notification of event occurrences and uncorrupted data. Developing a fault-tolerance system for WSNs is a challenging task. New self-configuration, self-recognition and self-organisation techniques are needed due to unique aspects of the operation of WSNs. Our current understanding of WSNs leads to an immunologically inspired solution to the design of a fault-tolerant network. One of the main roles of the Natural Immune System(NIS) is the recognition of self and the elimination of non-self proteins. Hence, in order to have an immune system equivalent for a sensor network, we must have a clear and stable definition of what constitutes the Self and the Non-Self Sets in a sensor network. This thesis explores two different approaches to modelling, collection and representation of the Self-Set in distributed sensor networks. We approach this problem, of identifying what constitutes the Self-Set in terms of sensor readings, using pattern recognition techniques from the machine learning field that leverages a small number of past observations of sensor nodes. We have chosen Competitive Learning Neural Network (CLNN) for the construction of the Self-Set. We define and evaluate two approaches for the aggregation of the Self-Set across multiple sensors in a WSN. The first approach is the Graph Theory Based Aggregation (GTBA) which consists of two main parts, namely: classification of the sensor readings by means of CLNN, which provides the multimodal view data and GTBA of the CLNN output, which takes intersections of intervals produced by CLNN. In this thesis we define and evaluate two different interpretations of GTBA, namely: Midpoint Intersection (MPI): one that considers the midpoint of intervals. Midpoint Free Intersection (MFI): one that does not take the midpoints into account but assigns the confidence levels to each of the resulted intersections. We evaluated both interpretations on three different types of phenomena and have shown that the second interpretation, MFI, consistently produced more precise representations of the environment under observation. However, MFI produced a very strict representation of the phenomenon, which consequently led to a large number of systems' retraining. Hence, we defined and evaluated a technique which produced a more relaxed representation of the Self-Set and at the same time preserved the finer variation in the phenomenon. The second approach is based on unsupervised learning. We define and evaluate three related unsupervised learning procedures ?? Divergence and Merging (DMP), Suboptimal Clustering (SOC), and Simple Clustering (SC) for the collection of the Self-Set. We explore the design tradeoffs in unsupervised learning schemes with respect to the clustering quality. We implement and evaluate these related unsupervised learning procedures on a realworld data set. The outcome of these experiments show that, out of the three unsupervised learning procedures studied in this thesis, the Suboptimal Clustering procedure appears to be the most suitable for the classification of sensor readings, provided that the amount of free memory is large enough to store and recluster an entire training set. We evaluate aggregation of the Self-Set produced by means of the distributed implementation of the unsupervised learning procedures. The aggregation is based on extended unsupervised learning and we evaluate the possibilities of the autonomous retraining of the system. Our experiments show that, in a naturally slowly changing environment, 40% of nodes reporting deviations is a large enough number to reinitialise the retraining of the system. The final conclusion is that it is possible to have a distributed implementation of the unsupervised procedure that produces an almost identical representation of the environment, which makes unsupervised learning suitable for a large number of sensor network architectures.
| Identifer | oai:union.ndltd.org:ADTP/258167 |
| Date | January 2009 |
| Creators | Bokareva, Tatiana, Computer Science & Engineering, Faculty of Engineering, UNSW |
| Publisher | Publisher:University of New South Wales. Computer Science & Engineering |
| Source Sets | Australiasian Digital Theses Program |
| Language | English |
| Detected Language | English |
| Rights | http://unsworks.unsw.edu.au/copyright, http://unsworks.unsw.edu.au/copyright |
Page generated in 0.013 seconds