Examination committee: Dr. Badri Nath (Chair), Dr. Michael Littman, Dr. Richard Martin, Dr. Muthu Muthukrishnan

Abstract:

Sensor networks have become an important source of information with numerous real life applications such as habitat monitoring and tracking. Detrimental actions are usually taken based upon the sensed information, and therefore, the quality, reliability and timeliness are extremely important issues in such applications. Unfortunately, data collected from wireless sensor networks is subject to several problems and sources of errors. Specifically, there are many factors which contribute to the existence of these problems such as the imprecision, loss and transience in wireless sensor networks, at least in their current form, as well as the current technology and the quality of the used cheap wireless sensors. This implies that these networks must operate with imperfect or incomplete information. Therefore, online cleaning of sensor data in real time, before any decision-making, is crucial.

In this proposal we introduce probabilistic, efficient, and scalable approaches for handling several data quality problems in wireless sensor networks. We present a framework for cleaning and querying of noisy sensors. We introduce a Bayesian approach for reducing the uncertainty in an online fashion. Using a statistical approach, we introduce several algorithms for answering a wide range of traditional database queries over uncertain noisy sensor readings. We also present context-aware sensors. We propose a technique for modeling and learning statistical contextual information in sensor networks that is based on Bayesian classifiers. Statistical contextual information encodes the spatio-temporal dependencies that highly exist among the sensors. It enables the sensors to locally predict their current readings based on their own past readings and the current readings of their neighbors. We discuss applications of context-awareness in discovering outliers and detection of faulty sensors, approximation of missing values, and in-network sampling. Finally, we present preliminary evaluations of our proposed work, and highlight our major future work directions and exciting research paths in this area.