Nieuwe sensorconcepten en gegevensverwerkingstechnieken voor online kwaliteitscontrole van melk

The rapid growth of the global population, expected to reach 9.7 billion by 2050, drives an urgent demand for sustainable, safe, and high-quality food systems. The dairy sector, valued at &euro;157 billion in the EU27 in 2023, faces complex challenges, including disease management, regulatory demands, and environmental concerns. To address these issues and maintain economic sustainability, the industry increasingly relies on technology-enabled precision livestock farming (PLF) data. This dissertation focuses on developing novel sensor concepts and data processing techniques to enable real-time (or 'on-line') milk composition analysis. Accurate monitoring of key milk components—fat, protein, and lactose—is essential for cow health, feeding optimization, and high-quality dairy products, all of which are critical for improving production efficiency and sustainability in dairy farming. At the core of this research is the application of long-wave near-infrared (LW-NIR) spectroscopy, a non-destructive method supporting on-line milk analysis, coupled with Partial Least Squares Regression (PLSR) to create robust milk composition prediction models. Initial development of LW-NIR transmittance at-line sensors demonstrated promising accuracy, with root-mean-square errors of prediction (RMSEP) of 0.08 %, 0.11 %, and 0.09 % for fat, protein, and lactose (wt/wt), respectively, in a real-time approach. These RMSEP values meet the International Committee of Animal Recordings (ICAR) guidelines for on-farm (< 0.20 %) and in-line (< 0.25 %) milk analyzers, approaching laboratory standards (< 0.10 %). However, external variations, particularly temperature fluctuations, posed a challenge, necessitating robust pre- and post-processing techniques to improve prediction model reliability. To address these interferences several spectral transformation and orthogonalization techniques were evaluated, including Piecewise Direct Standardization (PDS), Continuous PDS (CPDS), External Parameter Orthogonalization (EPO), and Dynamic Orthogonal Projection (DOP). Among these, DOP emerged as the most effective approach, significantly reducing temperature-induced variability with minimal control points, making it ideal for on-line applications. Specifically, DOP reduced the temperature dependence of RMSEP values by 80 % for fat and 95 % for protein and lactose compared to PLSR models calibrated at a single temperature. In the most extreme cases across the temperature range evaluated, DOP decreased RMSEP for fat from 0.92 % to 0.25 %, for protein from 0.74 % to 0.17 %, and for lactose from 1.07 % to 0.16 %. These results highlight the enhanced robustness and transferability of DOP, particularly for in-line applications, where the sensor is embedded within production for continuous monitoring. To further enhance milk composition analysis and move closer to in-line applications, spatially-resolved spectroscopy (SRS) was introduced. SRS captures light interactions at varying distances from the illumination point, providing more detailed information on milk serum composition. This approach identified optimal illumination-to-detection distances, which displayed improved accuracy for predicting milk components, with RMSEP values of 0.09 % for fat, 0.13 % for protein, and 0.12 % for lactose. However, challenges such as significant light loss at longer distances underscored the need for developing a more robust and efficient SRS setup with an improved signal-to-noise ratio (SNR) and light loss compensation. The use of SRS introduces multiple spectral sources, each providing complementary information on milk composition. To capitalize on the predictive potential of combining different data sources, a novel multi-block data fusion technique was presented with PROSAC-SO-PLS, which integrates the predictive capabilities of different raw and pre-processed spectral data blocks. This approach led to reductions in RMSEP ranging from 9 to 23 % across two datasets, confirming its effectiveness. Specifically for SRS, while there was no significant improvement in fat predictions, the RMSEP for protein decreased from 0.13 % to 0.10 %, and for lactose from 0.12 % to 0.10 %. These results also validated the optimal selection of illumination-to-detection distances, paving the way for the development of a more reliable and accurate SRS sensor for in-line milk composition analysis. Based on the insights gained from all previous research, an SRS interactance fiber-optic probe was developed to leverage the potential of SRS for milk composition analysis. This probe incorporates three concentric rings, with progressively increased light injection as the distance between illumination and detection increases, effectively compensating for light loss and improving SNR. This new design addressed the limitations of earlier models, reducing RMSEP from 0.09 % to 0.07 % for fat, from 0.10 % to 0.05 % for protein, and from 0.10 % to 0.07 % for lactose, compared to the initial SRS approach, both using the PROSAC-SO-PLS algorithm. Optimal illumination-to-detection distances were identified in the SRS probe—3.2 mm and 1.6 mm for fat, 3.2 mm and 4.8 mm for protein, and 4.8 mm and 3.2 mm for lactose—confirming that increased light interaction at longer distances significantly enhances lactose detection. As the sector undergoes significant transformation due to global market shifts, regulatory changes, and evolving societal expectations, integrating advanced sensor technologies like the SRS interactance fiber-optic probe into automatic milking systems (AMS) for real-time milk composition analysis presents a substantial opportunity. Early disease detection—such as mastitis and ketosis—could save farmers significant costs, with average failure costs per lactating animal estimated at &euro;120 for mastitis and &euro;33 for ketosis annually, totaling approximately &euro;51 million in Flanders and &euro;3.2 billion across the EU27. A 10 % reduction in these costs could translate to a market potential of &euro;5 million in Flanders and &euro;326 million in the EU27. This sensor technology could also support sustainable practices by enhancing feed optimization, dairy product differentiation, and contributing to the shift of the industry towards high-value products, aligning with the EU27 dairy market growth estimated to be &euro;178 billion by 2032. In conclusion, this research advances on-line milk composition analysis through the development of innovative sensor technologies and data processing techniques, with the use of LW-NIR spectroscopy and SRS. By integrating robust pre-processing and model-building methods, (DOP in an on-farm context) and PROSAC-SO-PLS for multi-block data, significant improvements in prediction accuracy and reliability were achieved. The latest version of the SRS interactance fiber-optic probe further enhances measurement precision by mitigating light loss and optimizing the illumination-to-detection distances for milk composition prediction. These advancements offer substantial benefits for disease management, production efficiency, and sustainability in the dairy sector, contributing to a more economically viable and sustainable future for dairy farming.