Robust quantitative FTIR gas analysis under spectral overlap using an LSTM-PSO-SVR hybrid framework

Fourier transform infrared (FTIR) spectroscopy plays a critical role in quantitative analysis of multi-component gas mixtures. However, in practical industrial monitoring scenarios, this technique suffers from severe spectral overlap among different gas species, which often degrades the accuracy and robustness of conventional quantitative models. To address these challenges, a hybrid framework integrating long short-term memory networks (LSTM), particle swarm optimization (PSO), and support vector regression (SVR), termed LSTM-PSO-SVR, is proposed for gas concentration prediction under complex spectral overlap conditions. To evaluate the proposed framework, CO and CO₂ mixtures-typical industrial emission components-were selected as target analytes. Focusing on the heavily overlapped absorption region of 2210-2250 cm^-1, a three-layer LSTM network transformed the original 83-dimensional spectral inputs into 128-dimensional deep features to enhance discriminative capability. To improve generalization under limited samples, SVR replaced the fully connected layer as the final regressor, and its hyperparameters were optimized by PSO. Compared with the baseline SVR, the LSTM-PSO-SVR framework reduced the root mean square error (RMSE) for CO₂ from 278.99 ppm to.