Advanced black tea fermentation grade identification: Terahertz time-domain spectroscopy coupled with self-attention mechanism guided deep learning

ABSTRACT Fermentation is essential in black tea production, markedly influencing its flavor profile. In this study, terahertz time-domain spectroscopy (THz-TDS) was integrated with deep learning methodologies, specifically a convolutional neural network coupled with self-attention mechanism (CNN-SAM), to enable precise classification of black tea fermentation grades. Principal component analysis (PCA) revealed distinct spatial clustering of samples across different fermentation levels within three-dimensional feature space, albeit with partial inter-class overlap. The variance contributions of the first three principal components were 46.13% (PC1), 23.67% (PC2), and 8.27% (PC3), respectively. The CNN-SAM architecture achieved 100% classification accuracy, substantially outperforming conventional deep learning models (CNN, LSTM, and CNN-LSTM). Through the self-attention mechanism, the model identified 0.70-0.74 THz and 1.51-1.53 THz as discriminative spectral regions, thereby conferring enhanced interpretability. This THz-TDS–deep learning paradigm establishes a robust, non-destructive analytical framework for black tea fermentation quality assessment, offering significant potential for industrial deployment and processing standardization.