Finding Fault: Anomaly Detection for Embedded Networked Sensing
Sheela D. Nair
Advisor: Mark Hansen
Recent advances in sensor technology, computing, and low-power communications have facilitated the development of embedded networked sensing (ENS). Despite the tremendous promise these systems hold as research tools, their widespread use has been hampered by concerns about data quality. We begin this thesis with a presentation of a taxonomy of the various faults and anomalies that can affect a deployed ENS system. We then develop an approach to fault detection based on “”signatures,”” a construction that was originally developed for fraud detection in telecommunications. The first step in implementing signature-based fault detection involves identifying a collection of features that can be computed easily from the stream of sensor measurements. We extend existing methods from process monitoring to accommodate the spatial and spatio-temporal character of our data. The second step in signature-based detection involves dening the distribution of our features under normal and faulty conditions. Here we take a nonparametric approach, ultimately studying the problem of on-line updating of the distributions involved. Various diculties arising in the on-line updating are identified, and some approaches and suggestions for addresses these are given. Finally, we evaluate the signature-based fault detection approach on data from three ENS deployments to demonstrate the applicability and limitations of our approach.