Semi-supervised deep learning framework for milk analysis using NIR spectrometers

Deep learning DL models of NIR spectral data outperforms traditional chemometrics algorithms specially when analyzing complicated materials spectra with overlapping bands. The wide spread of portable miniaturized spectrometers allows the collection of larger datasets which is necessary to build robust DL models. However, with the high cost of chemical referencing most of the collected samples are unreferenced (unsupervised). In this paper, a semi-supervised DL algorithm is proposed to provide a robust scalable model across a wider sample space and sensor space. Two cow milk datasets were collected and measured with 14 Neospectra spectrometers. The proposed algorithm is used to predict milk fat content and water adulteration ratio in milk. Results show that with a reduced referenced (supervised) dataset of only 35% of the milk samples and 50% of the spectrometer units augmented with the remaining unsupervised dataset we can predict milk fat content with R2 = 0.95 and RMSE = 0.22 and milk water adulteration with R2 = 0.8 and RMSE = 0.12. • A novel semi-supervised deep learning regression model framework is proposed. • Two cow milk datasets are collected using portable NIR spectrometers. • Unsupervised samples augmentation reduces sensor variations and sample variations effect on model performance. • Milk fat content and water adulteration ratio are predicted using deep neural networks.