Rapid Discrimination and Quantification of Adulterants in High-Value Yak Milk Using Electronic Nose and Machine Learning

Yak milk is rich in nutrients and holds significant economic value. However, its premium quality and market value are frequently undermined by economically motivated adulteration practices, posing risks to consumer trust and industry sustainability. This study focused on detecting yak milk adulterated with varying proportions of water, Holstein milk, and cattle yak milk, aiming to develop a rapid, accurate, and non-invasive detection method by integrating electronic nose (E-nose) data with machine learning algorithms and multivariate statistical analysis. The results demonstrated that both principal component analysis (PCA) and canonical discriminant analysis (CDA) could clearly distinguish between different types of raw milk. For adulterated samples, both methods effectively differentiated water-adulterated yak milk. Validated via tenfold cross-validation for robustness, the random forest (RF) model yielded an overall area under the curve (AUC) > 0.99, with perfect classification (AUC = 1.00) for the water-adulterated group. The multilayer perceptron (MLP) attained an average accuracy above 96% across all three adulterated categories, reaching as high as 99.8% for yak milk adulterated with Holstein milk. In quantitative prediction, multiple linear regression (MLR) demonstrated strong capability for predicting water content (determination coefficient (Rc2) = 0.9888), whereas MLP achieved the highest Rc2 (0.9998) for Holstein milk-adulterated yak milk. These findings demonstrated that E-nose coupled with CDA provides superior qualitative discrimination of adulterated yak milk, and MLP exhibited exceptional performance in quantifying adulterant concentrations. This study provided a rapid, low-cost, and non-destructive analytical solution, with significant potential for on-site market surveillance of high-value specialty dairy products like yak milk.