Classification and quantification of sesame oil in edible oils and adulterated mixtures using 1H NMR spectroscopy combined with multivariate, machine learning, and deep learning models

초록

Sesame oil is often adulterated with cheaper oils, necessitating accurate authentication and quantification methods. This study investigates the performance of AI-based models using 1H NMR spectral data for edible oil classification and sesame oil quantification in adulterated mixtures. All classification models—PCA-LDA, SVM, and 1D-CNN—achieved 100 % accuracy, with 1D-CNN additionally capturing both lignan and fatty acid signals. For regression, PLSR and SVR models achieved RMSEP values of 1.94 and 1.40 (R2 = 0.998), while the 1D-CNN regression model demonstrated superior performance (RMSEP = 1.03, R2 = 0.999) with broader spectral feature integration. External test samples incorporating previously unused oil types further validated the robustness of the CNN model, which accurately predicted sesame oil content within a 2 % error margin. These findings highlight the potential of explainable deep learning integrated with NMR spectroscopy for reliable detection and quantification of adulterated sesame oil.

키워드