From Volume to Mass: Transforming Volatile Organic Compound Detection with Photoionization Detectors and Machine Learning

Le résumé fourni par la source

(1) Objective: Volatile organic compounds (VOCs) monitoring in industrial parks is crucial for environmental regulation and public health protection. However, current techniques face challenges related to cost and real-time performance. This study aims to develop a dynamic calibration framework for accurate real-time conversion of VOCs volume fractions (nmol mol−1) to mass concentrations (μg m−3) in industrial environments, addressing the limitations of conventional monitoring methods such as high costs and delayed response times. (2) Methods: By innovatively integrating photoionization detector (PID) with machine learning, we developed a robust conversion model utilizing PID signals, meteorological data, and a random forest’s (RF) algorithm. The system’s performance was rigorously evaluated against standard gas chromatography-flame ionization detectors (GC-FID) measurements. (3) Results: The proposed framework demonstrated superior performance, achieving a coefficient of determination (R2) of 0.81, root mean squared error (RMSE) of 48.23 μg m−3, symmetric mean absolute percentage error (SMAPE) of 62.47%, and a normalized RMSE (RMSEnorm) of 2.07%, outperforming conventional methods. This framework not only achieved minute-level response times but also reduced costs to just 10% of those associated with GC-FID methods. Additionally, the model exhibited strong cross-site robustness with R2 values ranging from 0.68 to 0.69, although its accuracy was somewhat reduced for high-concentration samples (>1500 μg m−3), where the mean absolute percentage error (MAPE) was 17.8%. The inclusion of SMAPE and RMSEnorm provides a more nuanced understanding of the model’s performance, particularly in the context of skewed or heteroscedastic data distributions, thereby offering a more comprehensive assessment of the framework’s effectiveness. (4) Conclusions: The framework’s innovative combination of PID’s real-time capability and RF’s nonlinear modeling achieves accurate mass concentration conversion (R2 = 0.81) while maintaining a 95% faster response and 90% cost reduction compared to GC-FID systems. Compared with traditional single-coefficient PID calibration, this framework significantly improves accuracy and adaptability under dynamic industrial conditions. Future work will apply transfer learning to improve high-concentration detection for pollution tracing and environmental governance in industrial parks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.