UV-DOAS Combined with Spectral Projection Decoupling Neural Network (SPDNN): An Online System for the Simultaneous Detection of Ammonia and Isoprene at Sub-ppb Levels

Ultraviolet Differential Optical Absorption Spectroscopy (UV-DOAS) provides a promising way for the detection of ammonia (NH₃) and isoprene, which serve as biomarkers for kidney disease and lung cancer, respectively. However, the parts per billion (ppb) level concentrations and the spectral overlapping hinder simultaneous detection. This study presents an optical sensor based on UV-DOAS and Spectral Projection Decoupling Neural Network (SPDNN), which enables the simultaneous detection of breath NH₃ and isoprene for the first time. First, the differential absorption spectra of NH₃ and isoprene are obtained using UV-DOAS, and the effect of interfering components on the spectra is studied. Subsequently, a spectral projection method is proposed to decouple the overlapping spectra of NH₃ and isoprene, as well as to filter out the interfering components. Specifically, the interference filtering and overlapping spectra are decoupled by constructing the feature spaces for NH₃ and isoprene, and then projecting the mixture spectra into the two spaces. Finally, two convolutional neural networks are constructed to invert the concentration of the decoupled spectra. The experimental results demonstrate that the SPDNN-based sensor achieves accurate detection of mixed components at the ppb level under various interfering components. Allan's variance indicates that the sensor achieves detection limits of 1.23 ppb for NH₃ and 0.12 ppb for isoprene. Standard addition experiments based on human breath demonstrate that the sensor's can accurately detect NH₃ and isoprene simultaneously. More importantly, the SPDNN offers a novel approach for the simultaneous detection of multiple components with broadband absorption spectra.