Real-time information about milk composition would be very useful for managing the milking process. Mid-infrared spectroscopy, which relies on fundamental modes of molecular vibrations, is routinely used for off-line analysis of milk and the purpose of the present study was to investigate the potential of attenuated total reflectance mid-infrared spectroscopy for real-time analysis of milk in milking lines. The study was conducted with 189 samples from over 70 cows that were collected during an 18 months period. Principal component analysis, wavelets and neural networks were used to develop various models for predicting protein and fat concentration. Although reasonable protein models were obtained for some seasonal sub-datasets (determination errors <~0·15% protein), the models lacked robustness and it was not possible to develop a model suitable for all the data. Determination of fat concentration proved even more problematic and the determination errors remained unacceptably large regardless of the sub-dataset analyzed or of the spectral intervals used. These poor results can be explained by the limited penetration depth of the mid-infrared radiation that causes the spectra to be very sensitive to the presence of fat globules or fat biofilms in the boundary layer that forms at the interface between the milk and the crystal that serves both as radiation waveguide and sensing element. Since manipulations such as homogenisation are not permissible for in-line analysis, these results show that the potential of mid-infrared attenuated total reflectance spectroscopy for in-line milk analysis is indeed quite limited.
Detection of adulterants in grape nectars by attenuated total reflectance Fourier-transform mid-infrared spectroscopy and multivariate classification strategies
