Valutazione della qualità del vino tramite un sistema di Electronic Nose

Abstract

translate it to italian:The evaluation of wine quality has traditionally relied on human sensory panels, a process that is inherently subjective, time-consuming, and expensive. This thesis presents the implementation and validation of a machine learning framework for an electronic nose (e-nose) system designed to objectively assess wine quality. The study, carried out at Vega Research Laboratories SRL in collaboration with the Associazione Italiana Sommelier, utilizes a prototype e-nose, featuring a Raspberry Pi and an array of six MQ-series gas sensors, to analyze a diverse dataset of 90 wine samples. Each sample was also evaluated by certified sommeliers, providing 19 key sensory descriptors. The core contribution of this work lies in the development of the data pipeline and predictive models: machine learning algorithms, including Linear SVM, Multi-Layer Perceptron (MLP), and ensemble methods (Voting and Bagging classifiers), were trained on this combined dataset to predict wine quality. Using robust Leave-One-Out Cross-Validation (LOOCV), the system demonstrated high efficacy, with the Voting Classifier achieving peak performance. Feature analysis revealed that a hybrid approach, integrating sensor data with human-rated features, significantly improved prediction accuracy for complex attributes. This work validates the feasibility of a cost-effective e-nose as a powerful tool for rapid, objective wine analysis, with significant potential for application in quality control within the wine industry.

The evaluation of wine quality has traditionally relied on human sensory panels, a process that is inherently subjective, time-consuming, and expensive. This thesis presents the implementation and validation of a machine learning framework for an electronic nose (e-nose) system designed to objectively assess wine quality. The study, carried out at Vega Research Laboratories SRL in collaboration with the Associazione Italiana Sommelier, utilizes a prototype e-nose, featuring a Raspberry Pi and an array of six MQ-series gas sensors, to analyze a diverse dataset of 90 wine samples. Each sample was also evaluated by certified sommeliers, providing 19 key sensory descriptors. The core contribution of this work lies in the development of the data pipeline and predictive models: machine learning algorithms, including Linear SVM, Multi-Layer Perceptron (MLP), and ensemble methods (Voting and Bagging classifiers), were trained on this combined dataset to predict wine quality. Using robust Leave-One-Out Cross-Validation (LOOCV), the system demonstrated high efficacy, with the Voting Classifier achieving peak performance. Feature analysis revealed that a hybrid approach, integrating sensor data with human-rated features, significantly improved prediction accuracy for complex attributes. This work validates the feasibility of a cost-effective e-nose as a powerful tool for rapid, objective wine analysis, with significant potential for application in quality control within the wine industry.