People are becoming increasingly interested in mobile air quality sensornetwork applications. By eliminating the inaccuracies caused by spatial andtemporal heterogeneity of pollutant distributions, this method shows greatpotential for atmospheric research. However, systems based on low-cost airquality sensors often suffer from sensor noise and drift. For the sensingsystems to operate stably and reliably in real-world applications, thoseproblems must be addressed. In this work, we exploit the correlation ofdifferent types of sensors caused by cross sensitivity to help identify andcorrect the outlier readings. By employing a Bayesian network based system,we are able to recover the erroneous readings and recalibrate the driftedsensors simultaneously. Our method improves upon the state-of-art Bayesianbelief network techniques by incorporating the virtual evidence and adjustingthe sensor calibration functions recursively.Specifically, we have (1) designed a system based on the Bayesian belief network todetect and recover the abnormal readings, (2) developed methods to update thesensor calibration functions infield without requirement of ground truth,and (3) extended the Bayesian network with virtual evidence for infieldsensor recalibration. To validate our technique, we have tested our techniquewith metal oxide sensors measuring NO, CO, and O in a real-worlddeployment. Compared with the existing Bayesian belief network techniques,results based on our experiment setup demonstrate that our system can reduceerror by 34.1 % and recover 4 times more data on average.
展开▼
