<![CDATA[Carbon monoxide (CO) is a toxic, odorless gas commonly present in industrial processes and poses serious risks to occupational safety and health. This study proposes an optimized machine-learning-based approach to predict CO concentration using metal-oxide semiconductor (MOS) sensor arrays. The model was trained and evaluated on a public dataset comprising 650 time-series measurements from 14 thermally modulated MOS sensors, tested across CO concentrations ranging from 0 to 8.9 ppm under dynamic relative humidity (15%–75%). To optimize computational efficiency and mitigate multicollinearity, a multi-method feature selection strategy that combines Random Forest importance, Recursive Feature Elimination (RFE), and Mutual Information (MI) was implemented, successfully isolating sensors R10, R11, and R13 as the most robust predictors. A Random Forest Regression model, optimized via grid search and validated through five-fold cross-validation, was subsequently developed. The proposed framework demonstrated high predictive accuracy, achieving an R² of 0.884, Root Mean Square Error (RMSE) of 2.189 ppm, Mean Absolute Error (MAE) of 1.215 ppm, and Symmetric Mean Absolute Percentage Error (SMAPE) of 34.27%. These results highlight the potential of combining low-cost, feature-optimized MOS sensor arrays with ensemble machine learning for accurate, real-time gas monitoring. The framework provides a computationally efficient decision-support tool for the early detection of hazardous CO levels, contributing to safer process environments.]]>
Copyrights © 2026
