During standard material calibration, each individual sensor is screened for quality assurance by testing its response to known concentrations of standard gases of criteria air pollutants (SO2, NO2, CO, and O3). Then, the selected sensors are assembled into “sensor node” to measure the multiple pollutants simultaneously. These sensor nodes are then put in a control chamber to perform simulated environmental calibration. Standard gases and particulate matters are injected into the chamber simultaneously to simulate a wide range of ambient conditions by controlling temperature and humidity. Machine learning and neural-networking algorithms are applied to characterize sensor response. Next, the sensor nodes are installed outdoor with a Federal Reference Method (FEM) monitor in close vicinity to conduct in-field combined supervision calibration. Since the real ambient atmosphere is more complicated than the controlled chamber conditions, the FEM data is used to train the algorithms for improved sensor response. In regions without FEM nearby, the transfer calibration is implemented using mobile or portable equipment to optimize the calibration parameters.
The result shows that (1) after standard material and simulated environmental calibration, the correlation between sensors and FEM measurements increased from 0.4-0.6 to over 0.95; (2) after adaptive learning through in-field combined supervision and/or transfer calibration, the correlation between sensor and FEM improved from 0.6-0.75 to over 0.85. Over 10,000 sensor nodes (over 60,000 single sensor) calibrated through this four-stage calibration system have been successfully deployed in more than 80 cities across China and are currently being used in air quality monitoring for environmental management, research, and consulting.
![](images/flag.png)
![](images/blue_ribbon.png)