PWM-driven Thermal Excitation-based MOS Sensing with Machine Learning for CO–NO₂ Mixture Identification and Quantification

Accurate discrimination and quantification of CO–NO₂ mixtures remain a major challenge for low-cost metal oxide semiconductor (MOS) sensors due to poor selectivity and competitive adsorption pathways under steady-state operation. This study demonstrates that pulse-width-modulated thermal excitation can transform a dual-channel MiCS-4514 MOS sensor into a dynamic kinetic transducer capable of resolving complex multi-gas environments. A 45-state PWM cycle generates rich thermal fingerprints from the CO-sensitive and NO 2 -sensitive channels, revealing adsorption–reaction–desorption dynamics that are not accessible under fixed-temperature sensing. These transient signatures exhibit clear analyte-dependent structure, visualized through radar fingerprints and quantified via Principal Component Analysis, which achieves strong separability among CO, NO₂, and mixture conditions. Machine learning models further leverage this enhanced feature space: Quadratic Discriminant Analysis and Gradient Boosting reach excellent classification accuracy, while calibrated Gradient-Boosted Decision Tree regression achieves R² > 0.99 for both single-gas and mixed-gas concentration estimation under the tested conditions. These results demonstrate the methodological feasibility of PWM-driven sensing for mixture analysis using inexpensive MOS hardware under controlled laboratory conditions. The findings also reinforce the feasibility of integrating PWM-modulated MOS sensing elements with lightweight machine learning models into compact embedded systems, providing a potential pathway toward unattended air-quality monitoring, subject to future validation of long-term drift stability and environmental robustness. • PWM-driven thermal excitation reveals distinct transient kinetic fingerprints of CO, NO₂, and CO–NO₂ mixture using a dual-channel MOS sensor. • Temperature-modulated sensing signatures significantly enhance feature separability, enabling 100% classification accuracy and R² > 0.99 in mixture regression. • The approach converts low-cost MOS hardware into a scalable, machine-learning-ready platform for real-time monitoring of multi-component gas environments.