COS-DeformDeep: Adaptive 2T2D spectral feature extraction method for improving the component identification performance in mixtures based on handheld Raman technology.

Background: Raman spectroscopy is extensively utilized for the analysis of mixture components. Handheld Raman spectrometers, characterized by their compactness and portability, can rapidly acquire on-site spectral data without the need for intricate pretreatment or bulky instrumentation. In comparison to traditional laboratory-grade spectrometers, handheld devices offer distinct advantages. Nevertheless, although the unique spectral fingerprints of different substances facilitate identification, accurately quantifying and analyzing each component in complex mixtures remains a significant challenge.

Results: Therefore, a novel method called COS-DeformDeep is proposed to enhance and extract spectral features in handheld Raman mixture component identification. Firstly, synchronous two-trace two-dimensional correlation spectroscopy (2T2D-COS) is performed on pure components and mixture samples to highlight weak signals in overlapped peaks. Subsequently, deformable convolutions (DCNs) enhance the adaptability of deep learning models towards geometric deformation in the correlation peak region, thereby improving the capability of spectral feature extraction in 2T2D-COS. The proposed method was verified on three mixture datasets. Meanwhile, three substances, Ethanol, Diacetone alcohol, and Histidine, were chosen as the identified components with a volume-weight ratio ranging from 2 % to 20 %. The COS-DeformDeep model achieves the best performance with an average accuracy, precision, recall, and F1 score of 94.97 %, 98.45 %, 92.44 %, and 95.06 % respectively.

Significance: The proposed COS-DeformDeep is a highly efficient method for extracting features from weak spectral signals. By effectively capturing and analyzing the subtle variations in signals, it significantly enhances the recognition accuracy of specific components at low concentrations in mixtures. Moreover, its simplicity and suitability for handheld devices make it accessible to a wide range of users.