The evaluation of an oil's oxidation status during industrial production is highly important for monitoring the oil's purity and nutritional value during production, transportation, storage, and cooking. The oil and food industry is seeking a real-time, non-destructive, rapid, robust, and low-cost sensor for nutritional oil's material characterization. Towards this goal, a 1H LF-NMR relaxation sensor application based on the chemical and structural profiling of non-oxidized and oxidized oils was developed. This study dealt with a relatively large-scale oil oxidation database, which included crude data of a 1H LF-NMR relaxation curve, and its reconstruction into T1 and T2 spectral fingerprints, self-diffusion coefficient D, and conventional standard chemical test results. This study used a convolutional neural network (CNN) that was trained to classify T2 relaxation curves into three ordinal classes representing three different oil oxidation levels (non-oxidized, partial oxidation, and high level of oxidation). Supervised learning was used on the T2 signals paired with the ground-truth labels of oxidation values as per conventional chemical lab oxidation tests. The test data results (not used for training) show a high classification accuracy (95%). The proposed AI method integrates a large training set, an LF-NMR sensor, and a machine learning program that meets the requirements of the oil and food industry and can be further developed for other applications.
Keywords: CNN; T2 relaxation curve; chemical standards of oil oxidation; data mining; machine learning; pattern recognition; self-diffusion.