Accurate quantitative gas sensing through UV absorption spectra using a hybrid variational mode decomposition–CNN approach

Ultraviolet absorption spectroscopy, recognized for its rapid, nondestructive, and noninvasive nature, has been widely employed in fields such as environmental monitoring and chemical analysis. However, the accuracy of quantitative spectral analysis remains challenged by data distortions arising from various factors in the instrumental transmission process, including particle scattering, spectral drift, excessive denoising, and fitting errors. To overcome these issues, we propose a specific regression approach that integrates variational mode decomposition (VMD) with a convolutional neural network (CNN) for absorption spectra, termed VMD-CNN. The method first decomposes the spectrum into a series of intrinsic mode functions, from which mode components associated with gas absorption characteristics are extracted. A one-dimensional CNN is then trained on a large dataset of spectra with known concentrations to establish a quantitative inversion model. Experiments conducted on NO and SO2 absorption spectra within the 200 to 226 nm range demonstrate the effectiveness of the proposed approach. Comparative analyses with conventional techniques—including nonlinear least squares, partial least squares, and fast Fourier transform amplitude methods—before and after noise reduction and background elimination show that traditional methods heavily depend on data quality and may produce distorted results under low-concentration conditions due to over-optimization. By contrast, the proposed VMD-CNN model, which requires no preprocessing, consistently achieves the lowest error metrics across broad concentration ranges, specifically yielding root mean square error as low as 12.0 for NO and 2.0 for SO2, with mean absolute percentage error maintained 1.8% and 2.35%, respectively.