关键词: 近红外光谱；特种乳；掺水鉴别；卷积神经网络

Abstract: To address the increasingly severe issue of water adulteration in the non-bovine milks market, this study proposed a rapid, non-destructive method for identifying multiple types of adulterated non-bovine milks based on portable nearinfrared spectroscopy combined with convolutional neural networks (CNN). Spectral data were collected for four types of non-bovine milks, camel, horse, donkey, and yak milks, as well as samples of these milks diluted with water at varying concentrations (2%–80%, V/V). We compared the effects of standard normal variate transformation (SNV) and Savitzky- Golay filter smoothing preprocessing on the performance of partial least squares discriminant analysis (PLS-DA) and CNN models. The results showed that non-bovine milks exhibited significant characteristic absorption of moisture at 6 882 and 10 200 cm–1. The CNN model with SNV preprocessing achieved the best performance, with a classification accuracy of 98.71% on the test set, and its precision and recall were superior to those of traditional PLS-DA model. This model not only accurately distinguished between milks from different species but also exhibited high sensitivity in detecting water adulteration as low as 2%. The present study demonstrates that the discrimination system based on a portable near-infrared spectrometer and CNN can effectively reduce matrix interference in minor species milks.

Key words: near-infrared spectroscopy; non-bovine milks; water adulteration identification; convolutional neural network