Reducing the need for communication in wireless sensor networks using machine learning and planning techniques

Abstract

Wireless sensor networks are commonly used to remotely and automatically monitor environments.
One of the main challenges in wireless sensor networks is to use the limited available energy as efficiently as possible, to ensure longevity of the network. For such networks to survive their intended deployment period no energy may be wasted on inconsequential actions. As communication is one of the most energy-consuming tasks a sensor mote can perform, we propose a set of techniques that allow a base station to form an accurate environmental state estimation from a selected subset of measurements. In this thesis we present a novel methodology that combines three forms of intelligence.
The sensor mote and base station both maintain a neural network-based predictor of the environmental state, which the sensor mote uses as input for different controllers (both handmade and based on Partially Observable Markov Decision Processes) that determine the actions performed by the sensor mote. Armed with the prediction mechanism, a model of the environment, the controller executed by the sensor mote, and the reported measurements, the base station performs computations akin to those commonly used with Hidden Markov Models to form an accurate environmental state estimation.
We apply our techniques to real world data sets and reduce the required number of report operations by over 90% whilst incurring only minimal accuracy penalties.