Deep Learning-Enhanced Dual-Component Gas Sensor Based on Wavelength Modulation Spectroscopy

Considering the challenge of qualitative and quantitative detection for gas mixtures caused by spectral overlap, a deep learning-enhanced dual-component gas sensor based on wavelength modulation spectroscopy (WMS) with the 2f/1f signals is proposed, achieving simultaneous detection of exhaled carbon dioxide (CO₂) and methane (CH₄) concentrations using a single laser. A convolutional neural network (CNN)-based concentration prediction model (CPM) is introduced to address the cross-interference caused by the spectral overlap between gas molecules and to predict the concentration of each gas component accurately. Unlike traditional methods that collect a large number of labeled data from time-consuming experiments, a generative adversarial network (GAN) is used for the data augmentation of 2f/1f spectral signals, effectively addressing the issue of scarce experimental data for model training. The predicted concentrations are linearly fitted against the standard concentrations with high determination coefficients, demonstrating the strong feasibility and reliability of the proposed gas sensor. Allan deviation analysis indicates minimum detection limits of 17.34 ppm for CO₂ and 3.52 ppb for CH₄ at integration times of 112 and 159 s, respectively. Critically, the successful measurement of exhaled CO₂ and CH₄ concentrations using this sensor demonstrates its excellent performance in practical applications. This is a successful attempt to apply deep learning-enhanced WMS to dual-component gas detection in human breath, which provides guidance for simultaneous measurement of multicomponent gases and further paves the way for breath diagnosis.