Multi-level data fusion and machine learning for detecting defective cocoa liquors in quality control

Artificial Intelligence is revolutionising quality control processes in the food industry by providing more accurate and efficient methods for defect detection and classification. Traditional quality control relies on human inspectors, whose evaluations are often subjective and inconsistent. In contrast, AI models offer a systematic and objective approach to food quality assessment. This study investigates the application of AI to identify cocoa liquors presenting off-flavours, employing data from Liquid Chromatography- Ultraviolet/Diode Array Detector (UV/DAD) and Gas Chromatography coupled with Mass Spectrometry. Different Machine Learning and Deep Learning models, including K-Nearest Neighbors, Linear Discriminant Analysis, Logistic Regression, Partial Least Squares Discriminant Analysis, Random Forest, Support Vector Machines, and Multi-Layer Perceptron, were employed for classification of good and defected samples. Furthermore, different Data Fusion (DF) strategies, Low-Level DF (LLDF), Mid-Level DF (MLDF), and High-Level DF (HLDF), were explored to enhance model performance and the trade-off between model accuracy and interpretability was critically assessed. Among the tested approaches, HLDF demonstrated the highest accuracy (93%) in distinguishing defected cocoa samples. The results highlight the potential of AI models to improve food quality control, reduce human error, and streamline industrial processes, making them a promising tool for the future of the food industry. • GC-MS and LC-UV/DAD data were processed using machine and deep learning. • AI models classified off-flavour cocoa liquor with up to 93% accuracy. • High-Level Data Fusion (HLDF) provided the most accurate classification results. • The trade-off between model accuracy and interpretability was critically assessed.