Adaptive random sensor selection for field reconstruction in wireless sensor networks

Wireless sensor networks (WSNs) allow for the sampling of a physical phenomenon over long periods of time and across extended geographical areas [1]. Once reported to a central collecting unit, the samples may be used to reconstruct the developing of the physical phenomenon of interest -- also referred to as signal or sensor field -- in both time and space. Work in information theory [3] shows that a reliable signal reconstruction is possible if a sufficiently large number of nodes sample the signal at sufficiently close time and space intervals. Clearly, the achievable quality of the reconstruction can be maximized by letting the highest possible number of nodes collect and report samples. However, on typical sensor nodes, sensing and communication modules require the largest amount of energy and their continuous use can rapidly deplete node batteries [1]. Limiting the number of nodes actively participating in sensing and communication is thus the most effective way to increase the lifetime of both single sensor nodes and the network as a whole [2, 5]. Sensor selection algorithms can be used to schedule individual sensing activity in order to balance the accuracy of the reconstruction with energy consumption. In real WSNs deployments, the irregular spatial distribution of the nodes typically produces nonuniform sampling geometries that and reconstruction techniques able to deal with scattered samples must thus be used. In this context, the ACT reconstruction algorithm [3, Chapter 6] is one of the most computationally efficient and robust techniques known in literature and appears as a perfect fit to perform field reconstruction in WSNs. In particular, the ACT can deal with both very irregular sampling geometries and presence of noise in the data. However, the more the sampling geometry resembles a uniform grid, the better the performance of the ACT. With these considerations in mind, we investigate sensor selection strategies able to generate, given the constraints of the physical network topology, sampling geometries providing limited number of samples but still enabling the ACT to work properly. To this scope, we resort to random sensor selection strategies [2] and propose an adaptive method to determine, in a distributed fashion, the probability of activation of single sensor nodes. Our preliminary experimental results show that our approach succeeds in making the ACT able to reconstruct the sensor field with good accuracy, thereby using a lower number of sensors with respect to other random sensor selection strategies.