An ultra-sensitive optical sensor system for sulfur dioxide and nitric oxide based on improved UV-DOAS combined with the domain-transformed spectral decoupling convolutional neural network (DTSD-CNN).

Simultaneous measurements of trace sulfur dioxide (SO2) and nitrogen monoxide (NO) play an indispensable role in industrial emission reduction and combustion efficiency optimization. Ultraviolet differential optical absorption spectroscopy (UV-DOAS), a highly promising technique for detecting trace gases, is capable of simultaneously measuring SO2 and NO. However, baseline drift caused by NO and spectral overlap between different gases pose significant challenges to the accurate detection of trace gases. We report an ultra-sensitive optical sensor that combines improved UV-DOAS with a domain-transformed spectral decoupling convolutional neural network (DTSD-CNN). First, an improved UV-DOAS based on segmented fitting is introduced to efficiently mitigate baseline drift in mixed gases induced by NO. We then propose a DTSD based on spectral intensity mapping, which helps obtain decoupled SO2 and NO spectral signals with a high signal-to-noise ratio. Finally, the CNN model trained on single-component spectra achieves accurate quantification of trace SO2 and NO, circumventing dataset construction challenges associated with mixed gases. The results indicate that the DTSD-CNN achieved a coefficient of determination (R2) of 0.9999 for both SO2 (25.3-1995.2 ppb) and NO (9.8-1992.4 ppb), with mean absolute errors (MAE) of 4.0 ppb and 4.2 ppb, respectively. Allan variance analysis shows that the minimum detection limits for SO2 and NO are 1.1 ppb and 2.2 ppb, respectively. Thus, the proposed sensor can be considered an effective system for SO2 and NO detection.