Deep learning based spectral overlap separation for dual gases CO2/H2O detection with a 2004 nm laser source

Abstract

Carbon dioxide (CO₂) and water vapor (H₂O) dual gas detection is realized within the same wavelength band by one laser source, giving low system complexity and high measurement accuracy. A neural network based on deep learning is developed to solve the cross-interference problem from the spectral overlap between CO₂ and H₂O molecules. Absorption-induced 2nd harmonic model is established by simulation based on the theory of wavelength modulation spectroscopy (WMS) and calibrated by experimental data, then the calibrated model is used to conveniently generate a large dataset for training and testing the neural network. Verified by experiments, excellent linearity of R² $\approx$ 1 is achieved between the predicted concentrations and the true values for both CO₂ and H₂O. The relative error is evaluated to be $<$ 0.3 % for CO₂ and $<$ 0.9 % for H₂O by measuring 15 groups of CO₂/H₂O mixture. This study is a successful combination of WMS and deep learning to provide both capabilities of high-sensitive detection and spectral overlap separation, giving a promising direction for simultaneous measurement of multi-gases.