24 / 2023-04-19 10:07:23

Multi-scenario validation and assessment of a particulate matter sensor monitor optimized by machine learning

Air pollution; Exposure assessment; Environmental monitoring; Light scattering method

Objective: To evaluate the performance of sensor-based PM_2.5 and PM₁₀ monitors under a variety of typical indoor and outdoor emission scenarios, and to optimize its validation against standard measurements via machine learning.

Method: Parallel measurements and comparisons of PM_2.5 and PM₁₀ were carried out between sensor monitors and standard instruments in indoor and outdoor environment, respectively.  The validations were compared in a variety of multivariate machining learning models adjusting for temperature, relative humidity (RH), sampling months and PM_2.5/PM₁₀ ratios. The performance of the sensor is ultimately evaluated using an optimal model.

Results: Measurements of particles in indoor environment ranged in 0.8-370.7μg/m³ for PM_2.5 and 1.9-465.2μg/m³ for PM₁₀, while in outdoor environment, they were in 1.0-211.0μg/m³ and 0.0-493.0μg/m³ respectively. Compared with other machine learning methods, the random forest method showed the best fitting for PM_2.5 and PM₁₀ validation, either in indoor (R² of 0.97 and 0.91 and RMSE (Root-Mean-Square Error) of 1.91μg/m³ and 4.56μg/m³, respectively) or outdoor environment (R² of 0.90 and 0.80 and RMSE of 5.61μg/m³ and 17.54μg/m³, respectively). By comparing the validated measurements to the reference data, it showed that the correlation coefficients (r) of PM_2.5 and PM₁₀  were greater than or equal to 0.99, and the intra-device variability(IDV) was less than or equal to 3.31μg/m³ both in indoor and outdoor scenario. Comparison with the standard instruments, the lowest mean relative error(MRE) was indoor PM_2.5(9.80%). 

Conclusion: The random forest method provided the best validation models for sensor-based measurements after adjusting for temperature, RH and PM_2.5/PM₁₀. This sensor monitor could meet the requirements for a real-time instrument of PM_2.5 and PM₁₀ in both indoor and outdoor environment (at least for a year) with high accuracy, low variability high intra device consistency.