Enhancing Fourier Transform Near-infrared Spectroscopy with Explainable Ensemble Learning Methods for Detecting Mineral Oil Contamination in Corn Oil

Corn oil is rich in unsaturated fatty acids and antioxidants, which contribute to cardiovascular health. This has led to its widespread use in food processing and cooking. However, during the production, transportation, and storage of corn oil, it can be exposed to mineral oil contamination. Therefore, ensuring the quality and safety of corn oil is crucial. At the same time, there is growing attention on developing rapid and environmentally friendly analytical monitoring tools to screen edible oils for impurities, ensuring their quality. This study introduced an innovative method that combines explainable artificial intelligence with Fourier Transform Near-Infrared Spectroscopy (FT-NIR) to detect mineral oil contaminants in corn oil. Five types of mineral oils were selected as potential pollutants, and spectral data from contaminated and uncontaminated corn oil samples were collected. Partial Least Squares Discriminant Analysis (PLS-DA) and ensemble learning methods, AdaBoost, XGBoost, LightGBM, and CatBoost, were applied to address two qualitative objectives. The results showed that PLS-DA effectively captured the spectral differences between normal and contaminated samples, achieving 100 % classification accuracy. Following this, four classifiers were developed using spectral data selected by Competitive Adaptive Reweighted Sampling (CARS) to identify specific contaminants in corn oil. LightGBM demonstrated the best performance, achieving 100 % accuracy, precision, recall, and F1 score across all contaminant categories. The Shapley Additive Explanations (SHAP) algorithm was also used to enhance model interpretability. This algorithm identified the key spectral wavelengths contributing to the classification of each contaminant category. The findings demonstrate that combining FT-NIR, feature selection, and explainable models provides a fast, accurate, and environmentally friendly method for assessing the quality of corn oil. This approach improves contamination detection and enhances consumer confidence in edible oil products. Future work should focus on extending this method to other edible oil safety applications and integrating it into real-time on-site monitoring systems for edible oil production. • FT-NIR and ensemble learning enable efficient detection of mineral oil in corn oil. • CARS combined with SHAP values identified key features, improving model interpretability. • LightGBM achieved high accuracy classification of pollutants in corn oil (ACC=100 %).