Feature Calibration in Sensor Networks

Despite recent theory development, methods of calibration that accurately recover signals from biased sensor readings remain limited in their applicability. Acoustic sensors, for instance, which have been popular in low power wireless sensor networks, are difficult to calibrate in this manner, given their significant hardware variability, large dynamic range, sensitivity to battery power level, and complex spatial/temporal environmental variations. In this paper, we submit that the applicability of calibration is broadened by lifting the calibration problem from the level of sensors to that of sensing applications. We show feasibility of adaptive, easy, and accurate calibration at the level of application-specific features, via an example of recovering the feature of acoustic signal-to-noise ratio (SNR) that is useful in event-detection applications. By easy, we mean there is an efficient, purely local, and stimulus-free procedure for recovering SNR (that compares measured variances for multiple randomly chosen sensitivities, effected via acoustic sensor hardware support); unlike extant calibration methods, the procedure does not need to rely on any synchronization among nodes, long-term correlation between their respective environments, or assumptions about training events. And by accurate, we mean the procedure yields low error in SNR estimation. We provide experimental validation of the difficulty of directly calibrating acoustic signals and the accuracy of our SNR calibration procedure.