A machine learning-based electronic nose system using numerous low-cost gas sensors for real-time alcoholic beverage classification.
Journal:
Analytical methods : advancing methods and applications
PMID:
39158403
Abstract
This study introduces numerous low-cost gas sensors and a real-time alcoholic beverage classification system based on machine learning. Dogs possess a superior sense of smell compared to humans due to having 30 times more olfactory receptors and three times more olfactory receptor types than humans. Thus, in odor classification, the number of olfactory receptors is a more influential factor than the number of receptor types. From this perspective, this study proposes a system that utilizes distinctive data patterns resulting from heterogeneous responses among numerous low-cost homogeneous MOS-based sensors with poor gas selectivity. To evaluate the performance of the proposed system, learning data were gathered using three alcoholic beverage groups including different aged whiskeys, Korean soju with 99% same compositions, and white wines made from the Sauvignon blanc variety, sourced from various countries. The electronic nose system was developed to classify alcoholic samples measured using 30 gas sensors in real time. The samples were injected into a gas chamber for 60 seconds, followed by a 60-second injection of clean air. After preprocessing the time-series data into four distinct datasets, the data were analyzed using a machine learning algorithm, and the classification results were compared. The results showed a high classification accuracy of over 99%, and it was observed that classification performance varied depending on data preprocessing. As the number of gas sensors increased, the prediction accuracy improved, reaching up to 99.83 ± 0.21%. These experimental results indicated that the proposed electronic nose system's classification performance was comparable to that of commercial electronic nose systems. Additionally, the implementation of an alcoholic beverage classification system based on a pretrained LDA model demonstrated the feasibility of real-time classification using the proposed system.