基于太赫兹超材料和机器学习的挥发性有机物分类

Classification of Volatile Organic Compounds Based on Terahertz Metamaterials and Machine Learning

摘要: 挥发性有机物(Volatile organic compounds, VOCs)作为高速发展的科技时代的副产品，存在来源广、范围大、危害多和处理难的问题。VOCs检测时存在采样方法操作复杂、检测时间长、成本高等问题。挥发性有机物的快速精准检测，对及时采取措施，减少其对危害环境空气健康，提高人类生活质量方面有很大帮助。在这项工作中，选用异丙醇、乙苯、乙酸乙酯三种VOCs的研究对象；通过腔体结构营造密闭环境减少有机物挥发对实验的影响；采用具有Fano共振的超表面芯片用于测量了三种挥发性有机物在土壤中不同微含量时太赫兹查表面相互作用后的透射光谱的频移和强度。此外，根据变化趋势选择不同的拟合函数建立单变量回归模型，表明三种挥发性有机物具有明显的不同特性。研究表明，支持向量机(SVM)对三种VOCs的分辨准确率达到96.7%，而基于主成分分析的高斯混合模型(PCA-GMM)分类可视化算法，对于微量的检测物质，PCA-GMM在分类可视化可实现在95%置信区间内实现有效分离。

Abstract: Volatile organic compounds (VOCs), as by-products of the rapidly developing scientific and technological era, have problems of wide source, wide range, harmful to health and difficult to treatment. VOCs detection has the problems of complex sampling method, long detection time and high cost. Rapid and accurate detection of volatile organic compounds is of great help to take timely measures to reduce its harm to environmental air health and improve the quality of human life. In this work, three VOCs of isopropanol (IPA), ethyl benzene (EB) and ethyl acetate (EA) were selected. The cavity structure creates a closed environment to reduce the influence of organic volatilization on the experiment. A metasurface chip with Fano resonance was used to measure the frequency shift and intensity of transmission spectrum after terahertz surface interaction of three volatile organic compounds with different micro-contents in soil. In addition, the univariate regression model was established by selecting different fitting functions according to the variation trend, which showed that the three volatile organic compounds had obviously different characteristics. The results show that support vector machine (SVM) can distinguish the three VOCs with 96.7% accuracy, and the Gaussian mixture model (PCA-GMM) classification visualization algorithm based on principal component analysis can realize effective separation within 95% confidence interval for trace detected substances in classification visualization.