Zero-Shot Component Inference of Unknown Gas Mixtures

The identification of unknown mixed odors poses a new challenge for electronic-nose (e-nose) technology. This often occurs in real-world settings due to the unpredictable combinations of specific components in the air. As far as we know, previous e-nose machine learning models cannot deal with this problem. To address this issue, we have developed a zero-shot odor inference (ZOI) model identifying the components of an unknown mixed odor without any prior knowledge. It allows full-category reasoning based on limited types of training samples. ZOI utilizes subspace learning and classifier ensemble techniques to infer possible unseen mixed odors based on limited seen odors. The subspace learning part decomposes single component information from unknown gas-mixture responses, while the classifier ensemble part performs a series of binary classifications to identify the components in an unknown odor mixture. We have conducted experiments using a dataset containing 1-component, 2-component, and 3-component odors collected with a designed e-nose system, and the results demonstrate ZOI's excellent performance in inferring unknown mixed odors. When compared to other methods, ZOI with support vector machine achieved the highest average accuracy of 88.40% in recognizing seen odors and the lowest average Hamming distance of 0.286 in inferring unseen mixtures. In conclusion, the ZOI method represents a successful cognitive intelligence model for analyzing unknown mixed odors.