Application of artificial intelligence algorithms and low-cost sensors to estimate respirable dust in the workplace

Ta-Yuan Chang; Guan-Yu Chen; Jing-Jie Chen; Li-Hao Young; Li-Te Chang

doi:10.1016/j.envint.2023.108317

Application of artificial intelligence algorithms and low-cost sensors to estimate respirable dust in the workplace

Environ Int. 2023 Dec:182:108317. doi: 10.1016/j.envint.2023.108317. Epub 2023 Nov 8.

Authors

Ta-Yuan Chang¹, Guan-Yu Chen², Jing-Jie Chen², Li-Hao Young², Li-Te Chang³

Affiliations

¹ Department of Occupational Safety and Health, College of Public Health, China Medical University, Taichung, Taiwan. Electronic address: tychang@mail.cmu.edu.tw.
² Department of Occupational Safety and Health, College of Public Health, China Medical University, Taichung, Taiwan.
³ Department of Environmental Engineering and Science, Feng Chia University, Taichung, Taiwan.

PMID: 37963425
DOI: 10.1016/j.envint.2023.108317

Abstract

The Internet of Things (IoT) and low-cost sensor technology have become common tools for environmental exposure monitoring; however, their application in measuring respirable dust (RD) in the workplace remains limited. This study aimed to develop a predictive model for RD using artificial intelligence (AI) algorithms and low-cost sensors and subsequently assess its validity using a standard sampling approach. Various low-cost sensors were combined into an RD sensor module and mounted on a portable aerosol monitor (GRIMM 11-D) for two weeks. AI algorithms were used to capture data per minute over 14 days to establish predictive RD models. The best-fitting model was validated using an aluminum cyclone equipped with an air pump and polytetrafluoroethylene filters to sample the 8-hour RD for 5 days at an aircraft manufacturing company. This module was continuously monitored for two weeks to evaluate its stability. The RD concentration measured by GRIMM 11-D in a general outdoor environment over two weeks was 28.1 ± 16.1 μg/m³ (range: 2.4-85.3 μg/m³). Among the various established models, random forest regression was observed to have the best prediction capacity (R² = 0.97 and root mean square error = 2.82 μg/m³) in comparison to the other 19 methods. Field-based validation revealed that the predicted RD concentration (35.9 ± 4.1 μg/m³, range: 32.7-42.9 μg/m³) closely approximated the results obtained by the traditional method (38.1 ± 8.9 μg/m³, range: 28.1-52.5 μg/m³), and a strong positive Spearman correlation was observed between the two (r_s = 0.70). The average bias was -2.2 μg/m³ and the precision was 5.8 μg/m³, resulting in an accuracy of 6.2 μg/m³ (94.2 %). Data completeness was 99.7 % during the continuous two-week monitoring period. The developed sensor module of RD exhibited excellent predictive performance and good data stability that can be applied to exposure assessments in occupational epidemiological studies.

Keywords: Artificial intelligence; Internet of things; Low-cost sensor; Random forest method; Respirable dust.

MeSH terms

Air Pollutants, Occupational*
Artificial Intelligence
Dust / analysis
Environmental Exposure
Environmental Monitoring / methods
Inhalation Exposure / analysis
Occupational Exposure* / analysis
Workplace

Substances

Dust
Air Pollutants, Occupational