Topological Issues in Sensor Networks

Recent years have seen considerable interest in sensor networks. In this paper we study networks of chemical or similar sensors detecting a plume of contaminant, such as that in Figure (i), although our analysis also applies to sensors of other kinds. In order to optimise power consumption, we assume that sensors are normally quiescent, but can “wake” each other if contaminant is detected. Consequently, a wave of activation spreads through the sensor network, much like the spread of infection through a population. In addition, exchanging information with other nodes allows the network to compensate for sensor detection errors. We describe a Java-based simulation of such a network, which we use to study the performance of different sensor network topologies. In the model, the contaminant plume is simulated with a probability density function of chemical concentration, compatible with the RichardsonObukhov theory of turbulent mixing. Sensor nodes are simulated with a simple agent-based model incorporating message-passing. Of the six networks examined, the best-performing networks were a square grid and a network with short-range random links. These two networks most effectively compensated for sensor errors, while minimising the overall power consumption, as a result of not “waking up” unnecessary nodes. In general, random links in a sensor network appear to be effective, as long as the distance between linked nodes is small compared to the size of the contaminant plume. The wave of activation which spreads through the network differs from traditional models of the spread of infection through a population, in that initial growth in the number of recently activated sensors is approximately linear, followed by an exponential decay. The linear phase corresponds to an expanding circle of activated nodes within the contaminant plume, while the decay phase occurs when the wave passes beyond the plume. Further work will be conducted to model this process in more detail.