Use of Adaptive Estimation as a Predictor to Evaluate Sensor Reading Anomalies in Wireless Sensor Networks

Wireless Sensor Networks (WSNs) are deployed in the field to collect some physical data and report the data to a management entity. The very context of WSN usage implies autonomous operation. Consequently fault tolerant attributes are of high interest in WSNs, and therefore an active area of research. Healthy sensor nodes can occasionally report readings that are outside of the expected range. These out of range data, which are invalid and should be discarded, can be due to noise in the environment, a temporary power fluctuation, or any other transient malfunction. The question of how do we know if a sensor is reporting abnormal data, as opposed to the sensor being healthy and simply reporting accurate data that happens to differ in characteristics from the previous patterns is of particular interest in this paper. In this paper we present the underlying theory of our approach, the implementation of this algorithm, as well as simulation results that confirm auto-regressive models as a reliable signal behavior predictor.