Distributed regression for high-level feature extraction in wireless sensor networks

In sensor networks, energy and communication bandwidth are strongly constrained resources. A key technique to reduce the use of these resources is data aggregation, whereby the data collected by the nodes are combined during the routing. Aggregation is however limited to very few operators. This paper provides a new use of aggregation techniques, by showing that they can be used to extract high-level information by means of regression models. The main idea is to distribute the regression model coefficients in the network, in such a way that the model output is computed by aggregating data along a routing tree. We illustrate the use of the technique for a target tracking task, using the real-world acoustic and seismic data of the SensIT deployment. We show that it provides an effective way to estimate the position of the target while keeping the volume of communication low thanks to aggregation.