Study on Prediction Method of Mixed Gas Concentration Based on Infrared Spectroscopy and Integrated Learning

Precise prediction of mixed gas concentrations is critical for industrial safety and environmental monitoring. This study introduces an optimization method using Stacking ensemble learning to forecast concentrations via infrared spectroscopy. Data from the NIST spectral database undergo preprocessing and feature extraction to generate spectral samples. Individual models (SVR, RF, DT) were trained and evaluated on test data using MSE and $\mathbf{R}^{\mathbf{2}}$, but exhibit limitations in complex scenarios due to nonlinear spectral superposition and coupling. To overcome this, a Stacking model integrates multiple predictions through a metalearner (GBM), enhancing accuracy and robustness. Compared to RF and traditional neural networks, Stacking significantly outperforms with reduced MSE (0.041%) and increased $\mathbf{R}^{\mathbf{2}}(0.983)$. This approach offers a novel, practical solution for infrared spectroscopy in gas detection. (Abstract)