Portable Raman spectroscopy combined with machine learning for adulteration detection and brand classification in soy sauce

Soy sauce, a traditional fermented condiment, is increasingly subject to adulteration, posing serious threats to food safety and consumer trust. To overcome the limitations of current detection techniques, particularly their insufficient sensitivity and limited ability to generalize across brands, this study proposes a multimodal analytical framework integrating portable Raman spectroscopy with machine learning algorithms. Raman spectra (320–1750 cm−1) were systematically acquired from six commercial soy sauce brands, encompassing samples adulterated with hydrolyzed vegetable protein (HVP) at varying concentrations. The Partial Least Squares Discriminant Analysis (PLS-DA) model was applied for brand-specific adulteration detection, utilizing mean centering and latent variable optimization to enhance sensitivity to HVP-related spectral features. The PLS-DA model demonstrated strong performance, outperforming principal component analysis-linear discriminant analysis (PCA-LDA) and support vector machine (SVM) models in overall accuracy and F1-score across both training and testing sets. To enable cross-brand classification, a custom one-dimensional convolutional neural network (1D-CNN) architecture incorporating adaptive pooling and dual-dropout regularization was developed, facilitating robust brand traceability. Validation on an independent test set confirmed that the 1D-CNN achieved metrics exceeding 99.20% for accuracy, precision, recall, and F1-score. This work establishes a rapid, nondestructive, and machine learning-driven approach for on-site detection, providing a scalable framework for real-time quality assurance across diverse food matrices.