Model Identification for Wireless Sensor Networks

In many lifetime enhancement strategies for wireless sensor networks (WSNs) it is often necessary to identify the statistical model of the underlying physical field. We consider the problem of in-situ inference as an exemplary application and propose an in-situ model estimation algorithm that works in tandem with a parametric distributed filtering procedure. We demonstrate, via averaged-gradient analysis and simulations, that the resulting adaptive filter is stable, robust and, importantly, fully scalable. It compares favorably with kernel-regression inference, and typically significantly outperforms the latter when the spatio-temporal variations in the natural field are relatively rapid.