Sensor network data faults and their detection using Bayesian methods

The identification of unreliable sensor network data is important in ensuring the quality of scientific inferences based upon this data. Current sensor network technology for use in the environmental monitoring application frequently delivers faulty data due to a lack of an extensive testing and validation phase. Identification of data faults is difficult due to a lack of understanding of the faults and good models of the phenomenon.Aided by of the large amounts of data currently available, we present a detailed study of the most common sensor faults that occur in deployed sensor networks. We develop a set of features useful in detecting and diagnosing sensor faults in order to systematically define these sensor faults.We then present a system to detect these sensor network data faults. First we introduce a Bayesian maximum a posteriori probability method of selecting a subset of agreeing sensors upon which we model the expected behavior of all other sensors using first-order linear auto-regressive modeling. This method successfully selects sensors that are not faulty, but it has limited success in actual fault detection due to poor modeling of the phenomenon.Then we explore the benefits of using hierarchical Bayesian space-time modeling over linear auto-regressive modeling in sensor network data fault detection. While this approach is more complex, it is much more accurate and more robust to unmodeled dynamics than linear auto-regressive modeling.Finally we pair our method of selecting an agreeing subset of sensors with hierarchical Bayesian space-time modeling to detect faults. While this end-to-end Bayesian system requires a carefully defined model for the phenomenon, it is very capable of detecting sensor faults with a low false detection rate.