Validating sensors in the field via spectral clustering based on their measurement data

In this paper we introduce a spectral-based method for validating sensor nodes in the field via clustering of sensors based on their measurement data. We formalize the notion of peer consistency in measurement data by introducing a notion called ¿sensor indexing¿ and model the problem of identifying bad sensors as a problem of detecting peer inconsistency. Suppose all sensors have peers. Then by examining a certain number of leading eigenvectors of the measurement data matrix, we can identify those bad sensors which are inconsistent to peer sensors in their reported measurements. Further, we show that by deemphasizing or removing measurements obtained from these bad sensors we can improve the performance of sensor-based applications. We have implemented this spectral-based peer validation method and measured its performance by simulation. We report the effectiveness of the method in identifying bad sensors, and demonstrate its use in deriving accurate solutions in a localization application.