Hyperspectral-driven PSO-SVM model and optimized CNN-LSTM-Attention fusion network for qualitative and quantitative non-destructive detection of adulteration in strong-aroma Baijiu.

Baijiu adulteration practices, driven by profit motives, seriously endanger consumer health and disrupt the market. This study combined hyperspectral imaging with deep learning for adulteration detection. In the classification of authentic and adulterated samples, PSO-SVM achieved 97.62 ± 1.15 % accuracy through optimized spectral preprocessing. For quantitative prediction, a novel fusion network called Ghost-LSTM-Scaled Dot-Product Attention (GLSNet) was proposed, demonstrating significantly better predictive performance (R = 0.9569 ± 0.0145) than traditional Partial Least Squares Regression (PLSR) and other deep learning models: Convolutional Neural Networks (CNN) and CNN-LSTM networks (CLNet), while improving inference efficiency by 3.55 times compared to PLSR. GLSNet performed well on external validation sets and visualized adulteration distribution through heatmaps. The research shows that hyperspectral imaging combined with deep learning enables rapid and accurate detection of Baijiu adulteration, providing support for quality control and market regulation.