Photoacoustic Spectroscopy-Based Multi-Component Gas Detection Empowered by Machine Learning

To address the low sensitivity, poor selectivity and insufficient real-time response of conventional gas-detection techniques when analyzing complex gas mixtures, this study proposed a precise qualitative–quantitative system for multi-component trace gases that integrates photoacoustic spectroscopy (PAS) with advanced machine-learning algorithms. The setup employs a high-power infrared laser and a high-sensitivity cantilever-type photoacoustic cell. An Adaboost-enhanced improving support vector machine (Adaboost-ISVM) classifier was developed, achieving a classification accuracy of 99.17% for multi-component gases, with a Kappa coefficient of 99% and an AUC value of 99.375% for C₂H₂, NO₂ and SF₆ mixtures, significantly outperforming traditional SVM model. Additionally, to address the impact of temperature on detection results, this study introduced the Dung Beetle Optimizer (DBO) to optimize the Back Propagation (BP) neural network, it reduced the mean NO₂ concentration prediction error to 0.29 ppm over 25–60 °C, superior to traditional BP, GA-BP and SSA-BP. The integrated system offers a robust solution for real-time, reliable trace-gas monitoring in complex industrial environments.