Link Quality Estimation
The planning, deployment, and maintenance of wireless sensor networks requires efficient robust link quality estimation. Link quality estimation is also a key element in algorithms for placement, routing, and interference modeling. Current model-based approaches to link quality estimation are overly simplistic or rely heavily on detailed topographic and channel information. Meanwhile, naive data-driven approaches require large amounts of training data and are incapable of predicting link quality between locations where sensors are not already present.
In this paper, we present a data-driven non-parametric probabilistic model that exploits spatial structure to accurately predict both received signal strength (RSS) and packet reception rate (PRR) for both existing sensor pairs and between locations where sensors are not already present. In addition, our model quantifies the uncertainty in its estimates by inducing a predictive distribution over the RSS and PRR. We also investigate active learning techniques which inform measurement selection to minimize the uncertainty in link quality estimates. We demonstrate that these techniques significantly reduce the amount of training data required to obtain accurate predictions.
We extensively evaluate our approach in several separate sensor network deployments. Our results indicate that our model accurately captures the link quality distribution using substantially less training data than conventional data-driven techniques. Finally, we also demonstrate how our approach can be used to improve wireless interference models.