Data faults in sensor networks must be marked to ensure accurate inferences. We introduce a two phase semi-realtime end-to-end Bayesian fault detection system for sensor networks. The first phase selects a subset of agreeing sensors from which a model of expected behavior is derived. The second phase uses this subset to derive and tag questionable sensor data. To accurately model the data, we use a hierarchical Bayesian space-time (HBST) model, as compared to the linear autoregressive modeling used in previous works. Applying this system to simulated and real world data, results are excellent when the phenomenon is well modeled by the HBST model. It achieves high detection rates and almost zero false detection rates. Results also indicate that in cases of critically low spatial sampling density a more accurate model is required.

We present a new application of hierarchical Bayesian space-time (HBST) modeling: data fault detection in sensor networks primarily used in environmental monitoring situations. To show the effectiveness of HBST modeling, we develop a rudimentary tagging system to mark data that does not fit with given models. Using this, we compare HBST modeling against first order linear autoregressive (AR) modeling, which is a commonly used alternative due to its simplicity. We show that while HBST is more complex, it is much more accurate than linear AR modeling as evidenced in greatly reduced false detection rates while maintaining similar, if not better detection rates. HBST modeling reduces false detection rates 41.5% to 96.5% when paired with our simple fault detection method. We also see that HBST modeling is more robust to model mismatches and unmodeled dynamics than linear AR modeling.