MIR, NIR, fluorescence and Raman spectroscopy combined with chemometric modelling, imaging and deep learning to predict the functional properties of bovine milk and reduce subjectivity of reference methods.

Prediction of milk functional properties and milk quality would facilitate the adoption of an agile approach in powder and cheese manufacture, by allowing decreased reliance on wet chemistry reference methods. Using spectroscopy is inexpensive, rapid, generates little chemical waste and does not require extensive training. Predictive models for the heat stability, pH, Titratable Acidity (TA), Freezing Point Depression (FPD), zeta potential, osmolality, density and gel onset of raw bovine milk were created using MIR, NIR, fluorescence and Raman spectroscopy with Partial Least Squares (PLS) and non-linear regression models in Chapter 3. The dataset used is from seven dairy processors in Ireland. Data fusion of spectral modalities only improved model performance in the case of FPD. Using the spectra combined with selected compositional properties was used to predict the properties, improving the performance of models for heat stability and TA. Chapter 4 extracts clusters or ‘seasons’ identified in the dataset used in Chapter 3 and Chapter 5. Then, seasonal models are created in order to potentially improve the performance of the models in Chapter 3. For heat stability and TA, model performance was improved using seasonal modelling. Chapter 5 focuses on a particular quality parameter, Milk Ethanol Stability (MES). MES is a pass/fail test, therefore a binary classification model is used instead of a regression model. Traditional Partial Least Squares Discriminant Analysis (PLS-DA) was compared with other methods, including non-linear methods, to investigate which technique results in the greatest accuracy. Chapter 6 attempts to address and solve the problem of lack of objectivity present in the current oil bath method for determining heat stability. Reducing the subjectivity of the oil bath measurement would help to lower error in the heat stability predictive models in Chapter 3 and Chapter 4. In this research, an Automated Machine Learning (AutoML) with transfer learning is used to predict whether videos of reconstituted Skim Milk Powder (SMP) from the dairy industry are not coagulated, or coagulated during the heat stability reference measurement. Then, a custom model with hyperparameter search ranges is used to predict SMP coagulation time. Both models achieved a high accuracy, with the custom model performing slightly better yet taking longer to train, costing significantly more and requiring knowledge of hyperparameter tuning. Chapter 7 compares the accuracy of three calibration transfer methods: cubic spline interpolation, Piecewise Direct Standardization (PDS), and Artificial Neural Networks (ANNs) for calibration transfer between spectra from a benchtop and two handheld instruments using SMP and Full Fat Milk Powder (FMP) samples from industry. Spectra from an independent handheld instrument of the same specifications were inputted into the transfer models created using the benchtop and the first handheld instrument to investigate if the calibration transfer can be used across two handheld instruments of the same type. Regression models (for protein and fat) and classification models (for SMP and FMP) were created. Spectra transformed using PDS achieved the lowest Root Mean Square Error of Prediction (RMSEP) in the benchtop base models for the original handheld instrument, and spectra transformed using a simple ANN yielded the lowest RMSEP for the independent handheld instrument. This work can improve dairy processing efficiency.