Energy Conservation in Sensor Networks using Active Prediction

Inter-sensor data modeling and prediction have recently proven to be very promising techniques in drastically reducing the number of active sensors and thus the overall energy consumption in wireless sensor networks. Using existing and recently proposed inter-sensor data modeling techniques as the enablers, we propose an on-line distributed active prediction algorithm to use the available prediction models to put redundant sensor nodes to sleep. We first start by a more general eligible redundant set representation of the problem with an integer linear programming formulation that does not take network connectivity into account. We then discuss an efficient centralized heuristic to deal with the connectivity issue. The proposed distributed algorithm selects a subset of the sensors that form a connected network. After completion, each sensor is either in the active set or its measurements can be directly predicted by a designated predictor sensor in the active set, within specified tolerance levels. The elected predictor sensor performs the prediction and disseminates it along with its own measurements when necessary. We also show that performing the prediction modeling locally, as opposed to at a fusion center, is better suited for dynamic sensor networks. We evaluate the proposed algorithm using simulated data as well as real experimental data collected from 16 indoor temperature sensors from a large office building. The experimental results from the indoor temperature sensors show that the algorithm can reduce the number of active nodes in this case by 60% to 80% with only a 0.5◦F average tolerance level in the predicted measurements.