This project aims to compare and calibrate PM2.5 sensors using the exist-ing BAM-1020 monitors, audit the randomly selected sensors deployed at 14 counties or cities, analyze and apply the sensor data. Final goal is to pro-mote the smart IoT sensor network and the proper application of its data through public education, training workshops and expert advisory meetings to disseminate the applicable range and the factors influencing the readings of the sensors. It was found that there was a great influence of the relative humidity (RH) on PM2.5sensor readings. As RH increased, the sensor readings also in-creased, which justifies the need of calibration. Using BAM-1020 data, PM2.5 sensor data (manufacturer calibration) were calibrated by linear regression with R2 of 0.64(Taoyuan), 0.73(Zhongming), 0.75(Tainan), 0.74(Keelung) and 0.68(Pingtung), respectively. In this research, after multi-variable regression based on PM2.5 concentration and relative humidity, R2 was increased to 0.79(Taoyuan), 0.85(Zhongming) 0.85(Tainan), 0.84(Keelung) and 0.81 (Pingtung), respectively. MNB was decreased substantially to ±10% from +37.8~+57.7% and MNE was decreased to 30% from 46.47~69.8%. An increase in wind speed would decrease the measured value because of the decrease in inlet sampling flowrate and the number of sampled particles. When the sen-sors are used for more than eighteen months, it was found the data quality starts to degrade. But this aging problem needs to be investigated further. To sum up, factors such as PM2.5 concentrations, relative humidity, wind speed, sampling flow and ser-vice duration affect the measured values which should be considered when using sensor data in the future. This project audited the performance of PM2.5 sensors setup in 14 coun-ties or cities in two seasons. Until now, the inspection has been finished and accounted for 16.3% of 3,300 deployed sensors. The quality index of Tai-chung was only 50% since the MNE criteria of 30% is more rigorous as the sensors have been used for more than one year. In comparison, the MNE cri-teria of 50% is used for other counties or cities within the first service year. In the warm season sensor quality index ranged between 77.6% to 96.4%, on-ly the sensor quality of Hsinchu county and Kaohsiung city was lower than 65%. In the cold season, sensor quality ranged between 86.3% to 100%, only that of Miaoli county was lower than 65%. The sensor quality of Taichung city was only 50% in warm season. After using the calibration equation de-veloped in this research, the data quality was improved substantially by over 40% in cold season demonstrating the significance of using the correct cali-bration equation. This research proposed a data divergence statiscal method to judge the abnormal sensors and found that there are 5 and 8 abnormal sensors with very high probability in Da Yuan and Guanyin industry areas, respectively, which account for 13% and 6% in the total deployed number of sensors of 37 and 137, respectively. In addition, by using the ANN based on sensor PM2.5 concentrations, temperature and RH as the input layer, R2 was improved from 0.73 to 0.81 which is still lower than 0.85 obtained by the present the calibration equation. There were over 1000 people attending public education and expert training workshops, which improved the understanding of the deployment status of PM2.5 sensors and the limitation of the sensor data. This hopes to enhance the data quality of PM2.5 sensors, and increase the domestic people’s confidence in sensor data quality.

PM2.5 sensor, low cost air sensor, Air quality monitoring, Environ-mental internet of things, ANN, Multiple variable regression