In this talk I will analyze the problem of efficiently answering approximate queries over a sensor network at the sink. In particular, I will show how time series forecasting can be used to provide substantial reductions in the energy required to answer queries without significantly affecting answer quality.

Our approximate query framework, called SAF, comprises a suite of novel techniques for predicting values sensed at the nodes and for grouping together sensor nodes that produce similar data. It relies on a class of simple time series models, which are cheap to learn and dynamically adapt to variations in the data distribution to accurately predict sensor values and detect outliers and periods of data inconsistencies. SAF dramatically reduces energy consumption relative to previous query frameworks by allowing nodes to periodically turn off their radio and by remarkably reducing the amount of communication. Extensive simulation results performed on a trace of real data have confirmed the advantages mentioned above.

The techniques proposed in SAF are general and can be applied to other problems where trade-offs are preferable. I will briefly show how time series models can be applied to the clock synchronization problem to improve the accuracy and robustness of the clock, and the energy consumption. Our approach leads also to a refinement of the optimality bound for external clock synchronization. Finally, I will show how the techniques proposed in SAF can be applied to detect sensor faults in unknown environment.