Multi-Task Learning Based Blind Calibration for Low-Cost Air Quality Sensor Deployments.

Guodong Li,Zhiyuan Wu,Xinyu Liu,Yue Wang,Lin Zhang

DOI: https://doi.org/10.1145/3560905.3568099

2022-01-01

Abstract:Air pollution problem has caught much attention globally. In addition to the national air quality monitoring stations deployed by the government, the number of low-cost air quality sensors increases rapidly as a supplement to support fine-grained monitoring. In-field calibration methods are necessary for these low-cost sensor nodes to assure the data quality. However, it is costly to collect enough reference data after deployment to train the in-field calibration model and many sensors even have no synchronized reference in the real application scenarios. To address the above challenge, we propose a multi-task learning based blind calibraiton method for air quality sensors after deployments. Our method introduces not only the reference data of the target location to formulate calibration task, but also reference measurements collected from highly accurate stations already deployed by the government in other geographical locations to formulate prediction task. To utilize the reference measurements which are not in the same location with our target sensors, e.g., in other cities, we combine the proposed calibration task and prediction task under a multi-task learning scheme. The introduced references in other locations alleviate our few-reference challenge. Furthermore, we elaborate on the choices of different tasks to have better effect of the target calibraiton task. Evaluations on the real-world collected datasets show that our proposed algorithm has better calibraiton effect.