Palm Oil Adulteration Detection Using Model Averaging of Machine Learning Classifiers on Simulated Chemical Data

Abstract

Palm oil adulteration poses significant health and economic risks, necessitating accurate detection methods. This study develops a machine learning framework combining KNN, SVM, and Random Forest via weighted model averaging to analyze synthetic FTIR spectra simulating pure and adulterated palm oil. SVM emerged as the top performer (97.3% accuracy), significantly outperforming Random Forest (86.9%) and KNN (85.9%). Principal Component Analysis revealed distinct clustering, with PC1 (63.3% variance) strongly correlate with key adulteration markers like ester C=O (1745 cm?¹) and OH (3300 cm?¹) vibrations. Spectral segmentation identified the 1000–1100 cm?¹ region (C-O stretches) as most critical for detection, enabling a proposed two-stage screening protocol that reduces analysis time by 60% while maintaining >90% accuracy for 5% adulterant concentrations. The synthetic dataset, validated against experimental references, replicated physicochemical trends, including peak broadening in oxidized samples (+20% FWHM) and dye-specific N=O peaks (1520 cm?¹). Model averaging enhanced stability, reducing performance variability to 1.2% versus 3.5–4.8% for individual models. These results highlight SVM’s superiority in handling high-dimensional spectral data and non-linear patterns, while the methodological advances—including noise modeling (SNR = 40 dB) and feature selection—offer practical solutions for portable FTIR devices. The framework supports real-time adulteration screening in resource-limited settings, with implications for food safety regulation and IoT-based quality monitoring in global palm oil supply chains.