We propose an energy-efficient framework, called SAF, for approximate querying and clustering of nodes in a sensor network. SAF uses simple time series forecasting models to predict sensor readings. The idea is to build these local models at each node, transmit them to the root of the network (the "sink"), and use them to approximately answer user queries. Our approach dramatically reduces communication relative to previous approaches for querying sensor networks by exploiting properties of these local models, since each sensor communicates with the sink only when its local model varies due to changes in the underlying data distribution. In our experimental results performed on a trace of real data, we observed on average about 150 message transmissions from each sensor over a week (including the learning phase) to correctly predict temperatures to within 0.5C.SAF also provides a mechanism to detect data similarities between nodes and organize nodes into clusters at the sink at no additional communication cost. This is again achieved by exploiting properties of our local time series models, and by means of a novel definition of data similarity between nodes that is based not on raw data but on the prediction values. Our clustering algorithm is both very efficient and provably optimal in the number of clusters. Our clusters have several interesting features: first, they can capture similarity between far away nodes that are not geographically adjacent; second, cluster membership to variations in sensors' local models; third, nodes within a cluster are not required to track the membership of other nodes in the cluster. We present a number of simulation-based experimental results that demonstrate these properties of SAF.
Mendeley saves you time finding and organizing research
Choose a citation style from the tabs below