SpecRecFormer: Deep Learning-Driven Adaptive Component Identification of PAH Mixtures Based on Single-Component Raman Spectra

The identification of components in mixed spectra is a fundamental challenge in spectral analysis, complicated by factors such as spectral peak overlap due to structural similarities, shifts in characteristic peaks from molecular interactions, and interferences caused by matrix effects. While deep learning offers robust feature extraction capabilities and notable advantages in addressing these challenges, it still faces significant obstacles, including the limited availability of labeled spectral data for effective training and the difficulty of applying fixed-threshold predictive models to spectra containing uncertain components. This paper established a deep learning model, SpecRecFormer, for the rapid identification of individual components in mixed polycyclic aromatic hydrocarbons (PAHs) based on their Raman spectra. The model integrates a dual-channel convolutional neural network (CNN) for local feature extraction with a Transformer module for global representation. It is trained on a reference database composed of single-component spectra, with simulated mixed spectra generated through data augmentation to expand and diversify the training set. This architecture enables the model to evaluate the similarity between unknown mixed spectra and known single-component references. To further enhance recognition accuracy, an adaptive threshold strategy is introduced, dynamically adjusting decision thresholds based on spectral characteristics to retain only components exceeding the threshold as candidate predictions. Experimental results demonstrate that with training data derived from only four single-component reference spectra, the model generalizes effectively to three real-world PAH data sets, achieving accuracies of 93.75%, 89.21%, and 93.63%, respectively, significantly outperforming conventional neural network models. These findings present an innovative and highly effective approach to mixed spectral analysis, with substantial potential for advancing applications in environmental science and chemical analysis.