Near-infrared spectroscopy and machine learning for classification of food powders during a continuous process

In food production environments, the wrong powder material is occasionally loaded onto a production line which impacts food safety, product quality, and production economics. The aim of this study was to assess the potential of using Near Infrared (NIR) spectroscopy combined with Machine Learning to classify food powders under motion conditions. Two NIR sensors with different wavelength ranges were compared and the ML models were tasked with classifying between 25 food powder materials. Eleven different spectra pre-processing methods, three feature selection methods, and five algorithms were investigated to find the optimal ML pipeline. It was found that pre-processing the spectra using autoencoders followed by using support vector machines with the all spectral wavelengths from both sensors was most accurate. The results were improved further using under-sampling and boosting. Overall, this method achieved 99.52, 97.12, 94.08, and 91.68% accuracy for the static, 0.017, 0.036 and 0.068 m s-1 sample speeds. The models were also validated using an independent test sets.