Blind Calibration by Maximizing Correlation

In large-scale IoT systems, blind calibration problem becomes increasingly prominent for sensor calibration without ground truth reference. Most of the existing blind calibration methods adopt either a handcrafted spatio-temporal model or a specific drift mechanism assumption. However, these assumptions may be over-simplified or introduce inappropriate bias, and therefore lead to great performance degradation in the real-world deployment. In this paper, we present a novel generative framework for blind calibration problems without specific data correlation or drift model assumption. We extract the most informative feature that maximizes correlation between reference data and target data using soft-HGR maximal correlation regression. Therefore, our method can be used in different blind calibration tasks especially where data correlation or drift model is unknown or deviated. Besides, our method can be conveniently augmented with a reliable drift model to further improve performance on specific tasks. We conduct comprehensive evaluations over a three-month real-world air pollution sensing dataset collected in Foshan, China. Results show our method can obtain the best performance compared to previous blind calibration methods in the absence of accurate drift model knowledge.