A lightweight neural model for gas concentration prediction in TDLAS under varying environmental conditions

Tunable diode laser absorption spectroscopy (TDLAS), with its strong selectivity, high sensitivity, and fast response speed, has become a widely used measurement technique of gas concentration. However, according to the gas absorption spectral theory, variations in temperature and pressure can alter the absorption spectral lines, leading to significant measurement deviation. In this study, based on the TDLAS/wavelength modulation spectroscopy (WMS) measurement system, the effects of temperature and pressure variations on concentration measurement and the second harmonic signal are analyzed. To address these challenges, we propose the multilayer perceptron optimized by fungal growth optimization (FGO-MLP), a lightweight deep learning algorithm that adaptively optimizes the network architecture by integrating FGO with an MLP. The model is trained using temperature, pressure, and key second harmonic spectral features as input variables. Experimental results demonstrate that temperature and pressure perturbations significantly alter the harmonic signals, thereby affecting concentration prediction accuracy. Compared with conventional optimization algorithms, the FGO algorithm provides more effective hyperparameter tuning for the MLP model, resulting in improved predictive performance. On the test set, the proposed model achieves a mean absolute percentage error (MAPE) of 0.97% and a coefficient of determination (R²) of 0.99991. The model maintains low computational complexity, with an average inference time of 0.34 μs per sample. Allan-Werle deviation analysis further confirms the robustness and generalization capability of the proposed approach. These findings provide reliable technical support and practical value for achieving efficient, stable, and precise gas concentration measurements in complex environments using TDLAS systems.