A Recursive Approach to Partially Blind Calibration of a Pollution Sensor Network

Distributed, low-cost sensor networks have become a widely used tool to aid in the interpolation of atmospheric measurements between regulatory grade monitoring stations, referred to as Golden Standards (GS). However, the quality of the data from these sensor networks can be questioned, especially in poorly correlated environments. Sensors can be individually calibrated in a laboratory environment before deployment of the network, but this approach is unfeasible for large networks. To overcome these shortcomings, we propose a novel online, autonomous approach to sensor calibration, by leveraging the ground truth measurements of the GS to calibrate neighboring nodes of the sensor network using a Recursive Least-Squares technique. Our algorithm percolates this calibration through the network such that every connected node will converge towards its ideal linear calibration. The algorithm outperforms a pre-deployment laboratory calibration and provides good tracking of quickly-changing environmental stimulus, which is known to change the inherent sensor calibration. Experimental results show a 45% improvement in estimated measurement error when compared to measurements corrected using a laboratory calibration.