Pattern recognition based on machine learning identifies oil adulteration and edible oil mixtures

Abstract

Abstract Previous studies have shown that each edible oil type has its own characteristic fatty acid profile; however, no method has yet been described allowing the identification of oil types simply based on this characteristic. Moreover, the fatty acid profile of a specific oil type can be mimicked by a mixture of 2 or more oil types. This has led to fraudulent oil adulteration and intentional mislabeling of edible oils threatening food safety and endangering public health. Here, we present a machine learning method to uncover fatty acid patterns discriminative for ten different plant oil types and their intra-variability. We also describe a supervised end-to-end learning method that can be generalized to oil composition of any given mixtures. Trained on a large number of simulated oil mixtures, independent test dataset validation demonstrates that the model has a 50th percentile absolute error between 1.4–1.8% and a 90th percentile error of 4–5.4% for any 3-way mixtures of the ten oil types. The deep learning model can also be further refined with on-line training. Because oil-producing plants have diverse geographical origins and hence slightly varying fatty acid profiles, an online-training method provides also a way to capture useful knowledge presently unavailable. Our method allows the ability to control product quality, determining the fair price of purchased oils and in-turn allowing health-conscious consumers the future of accurate labeling.