Detection of water adulteration levels in milk using near-infrared spectroscopy combined with chemometrics.

Milk is a nutrient-rich food, and water adulteration in milk can reduce its quality and increase food safety risks. Nondestructive and efficient detection of milk adulteration levels is crucial to addressing this issue. This study employed a portable near-infrared spectrometer to measure and analyze the absorbance of milk samples within the wavelength range of 900 to 1,800 nm. Based on the original spectra, models such as soft independent modeling of class analogy (SIMCA), naive Bayes, k-nearest neighbors, and support vector machine were constructed for the discrimination of water-adulterated milk. Preprocessing methods including Savitzky-Golay convolutional smoothing, Savitzky-Golay filtered derivative, multiplicative scatter correction, standard normal variate, vector normalization (VN), and min-max normalization were applied to the original spectra. Additionally, models such as one-dimensional convolutional neural network, partial least squares regression, and support vector regression (SVR) were established for the prediction of water content in milk. The results showed that the SIMCA model achieved 100% discrimination accuracy = 1, sensitivity = 1, specificity = 1, precision = 1, F-score = 1 in the analysis of water-adulterated milk. In the prediction of milk water content, the VN-SVR model performed the best (coefficient of determination for the training set R = 0.9996, root mean square error for the training set = 0.0477%, coefficient of determination for the test set R = 0.9992, root mean square error for the test set [RMSEP] = 0.066%), and also exhibited the best performance in model validation (R = 0.9760, RMSEP = 0.2554%). This study provides theoretical references and technical support for dairy product quality control and food safety regulation.