Stacked long and short-term memory (SLSTM) - assisted terahertz spectroscopy combined with permutation importance for...

Mislabeling of low-value red wines as high-value ones is common, which seriously undermines consumer rights and interests. However, traditional sensory and chemical analysis methods have limitations, which highlights the need for novel detection techniques. To address above issues, terahertz time-domain spectroscopy (THz-TDS) combined with deep learning (DL) was employed to distinguish different red wine varieties quickly and non-destructively, contributing to correctly identifying red wine labels. Compared with the other models, the stacked long and short-term memory (SLSTM) model based on the first derivative (1-st der) spectra performed the best (Precision: 85.72 %, Recall: 85.61 %, F1-score: 85.59 %, Accuracy: 85.61 %). In addition, feature selection (FS) was used to explore the feasibility of improving model accuracy and reducing prediction time by eliminating redundant frequencies. Compared to full frequency, the 1-st der-SLSTM model based on permutation importance (PI) performed slightly lower (Precision: 84.42 %, Recall: 84.10 %, F1-score: 84.14 %, Accuracy: 84.18 %), but the prediction time was reduced by 2 s. Therefore, different models can be selected based on different detection needs by weighing accuracy and prediction time. In conclusion, the current research demonstrates that the SLSTM-assisted THz-TDS technology provides a novel approach for fast, accurate and non-destructive for fast, accurate and non-destructive discrimination of red wine labels, facilitating the maintenance of market discipline.