Wireless sensor networks (WSNs) are transforming many areas of modern science, from ecology, to civil engineering, and public health. At the same time, experiments with WSN testbeds have exposed a number of shortcomings of the current technologies that hamper the wider adoption of this technology by domain scientists. In particular, inexpensive sensor transducers used as sensors often fail intermittently in complex and unexpected ways. Moreover, given the small number of sensors, their hardware limitations, and the complexity of tracking unexpected phenomena, sensor networks may fail to accurately record punctuated events (e.g. flash floods and hailstorms).
In this project we compare the effectiveness and efficiency of different statistical methods used to detect whether sensors are misbehaving. We investigate both off-line and on-line techniques that can be implemented on resource-limited sensor nodes to detect faults in real time. We also investigate the use of similar techniques for learning the inherent characteristics of the physical phenomena that the network is monitoring. This knowledge can be used to dynamically detect and report the onset of punctuated events and modify the behavior of the network accordingly. We will evaluate the performance of these algorithms using archived data and implement them on the nodes of a wireless sensor network used for soil monitoring. The outcome of this project will be a set of algorithms that can be broadly used to build responsive and efficient wireless sensor networks for sensor-based science.