A concentration detection model combining frequency-domain physical priors and deep learning: For SO₂ and NO mixed...

In the 200-230 nm ultraviolet wavelength range, there is a significant overlap in the absorption spectra of sulfur dioxide (SO₂), nitric oxide (NO), and ammonia (NH₃). This overlap significantly complicates the detection of SO₂ and NO concentrations. This study proposes a hybrid concentration detection model that combines frequency-domain physical priors with deep learning. Firstly, by utilizing characteristic differences among the three gases in the frequency domain, a band-pass filtering layer separates the frequency band of each target gas from the mixed spectrum. Secondly, we design a parallel dual-output network structure incorporating an efficient channel attention mechanism. Multi-branch feature extraction and an attention-weighting mechanism enhance the model's ability to extract key features. Latin hypercube sampling is employed to generate diverse concentration combinations, and fine-tuning strategies effectively bridge the distribution gap between synthesized and real data. Under the interference of NH₃, the detection limits of SO₂ and NO are 0.25 ppm and 0.26 ppm, respectively, with corresponding uncertainties of 1.25% and 1.30%. This study provides a novel and effective technical path for solving the problem of spectral interference in multi-component gas analysis.