UNI-CAL: A Universal AI-Driven Model for Air Pollutant Sensor Calibration With Domain-Specific Knowledge Inputs

Abstract

Portable Sensor Nodes (PSNs) can supplement geographically sparse government-run static air quality monitoring stations (AQMSs). A PSN typically consists of several low-cost pollution sensors for different air pollutants, which must be calibrated to improve the accuracy of measurements. These sensors can be co-located with the high accuracy monitoring equipment (HAME) at AQMSs for calibration. Existing studies have suggested that different pollution sensors may favor different calibration models; even the same pollution sensors in different PSNs may favor different models. However, it is impractical to co-locate each PSN with HAME due to limited access to AQMSs, making large-scale sensor calibration difficult. This study proposes UNI-CAL for calibrating different pollutants, including nitrogen dioxide (NO₂), ozone (O₃), and particulate matter (PM_2.5 and PM₁₀), based on a novel AI-driven model with residual blocks capturing the complex non-linear interactions of raw measurements plus citywide domain-specific information, including meteorology, background pollution, and temporal characteristics. UNI-CAL further allows transfer calibration, i.e., the calibration of sensors from calibrated ones. UNI-CAL has improved the performance of direct calibration by 3.143% on average compared to the best baseline across all pollutants and PSNs on all evaluation metrics. Moreover, domain-specific information has significantly improved the direct calibration performance of UNI-CAL by 4.852% on average. Furthermore, UNI-CAL has demonstrated a strong capability in transfer calibration and achieved the best performance in most scenarios after incorporating domain-specific information. In the future, one can collect more data covering different environmental conditions and explore advanced semi-supervised learning techniques to improve the consistency, robustness, generalizability, and transferability of the proposed calibration framework.