This study utilized terahertz time-domain spectroscopy（THz-TDS） to conduct spectral detection on six food additives，namely benzoic acid，sorbic acid，xylitol，

L

-alanine，melamine，and Sudan I，obtaining their terahertz experimental spectra in the range of 0.4-2.4 THz. The experimental spectra data were corrected and preprocessed using the Savitzky-Golay（S-G） smoothing method，followed by dimensionality reduction of the spectral data through principal component analysis（PCA）. This paper applied convolutional neural networks（CNN） and long short-term memory（LSTM） networks to construct qualitative classification recognition models for food additives. The results indicated that the classification accuracies of the CNN model and the LSTM model were 93.8% and 92.7%，respectively. To further enhance the classification accuracy and the feature extraction capability of the data，as we

ll as to better integrate information from different modalities，this study introduced the Attention mechanism and established a composite neural network（CNN-LSTM-Attention） model. The experimental results demonstrated a significant improvement in the classification accuracy of the constructed CNN-LSTM-Attention model，reaching 99.5%. To create a more objective and comprehensive evaluation system for the aforementioned three models，this paper employed the F1 score as an evaluation metric. Comparatively，the F1 scores of the CNN and LSTM models were 0.91 and 0.88，respectively，while the CNN-LSTM-Attention model achieved a F1 score of 0.95，significantly outperforming the other two models. Furthermore，this paper applied the three models to the qualitative analysis of food additive mixtures，specifically preparing samples of melamine mixed with

L

-alanine and sorbic acid mixed with benzoic acid. The CNN-LSTM-Attention model demonstrated a distinct advantage in identifying these two mixtures，with a recognition accuracy of 90.0% and an F1 score of 0.87，both superior to the CNN and LSTM models. The results indicate that the CNN-LSTM-Attention model still possesses a clear advantage in the qualitative recognition of food additive mixtures. The findings suggest that，compared to the CNN and LSTM models，the qualitative classification model established using the composite neural network CNN-LSTM-Attention is optimal in terms of accuracy and F1 score. This paper provides theoretical support for the rapid，accurate，and non-destructive detection of food additives，offering significant application value and potential prospects.