利用機器學習解析有機物電漿放射光譜之應用

Abstract

微電漿系統的發展為實時現地的氣體偵測帶來一種新的可能性，利用電漿光譜進行氣體偵測同時具有高選擇性、響應速度快、高靈敏度等等優點，然而如何定性辨別揮發性有機物之電漿放射光譜仍然存在挑戰。 機器學習為一實現人工智慧之技術，藉由統計理論結合最適化問題，從大量的資料中分析複雜資料之行為及規律性，最終整理規則以達分類或回歸之目的，依據訓練時資料是否有標記可將機器學習分為非監督式學習及監督式學習。 本研究設計一種可攜式微電漿產生裝置可應用於實時現地之氣體檢測，此裝置為一介電層放電型微電漿，由5V直流電源結合自製升壓模組驅動，本裝置以市售銅箔基板做為原料，運用碳粉轉印技術及濕式蝕刻技術製備而成，具有低成本、製備容易、可客製化等等優點，我們使用此裝置結合光學偵測裝置建立一可連續式收取大量光譜資料的平台，結果顯示，此平台能夠收取及分析乙醇光譜達半定量之成果，並可於長時間下高效率連續收取高質量且穩定之光譜，而後透過機器學習技術定性辨別不同揮發性有機物之電漿放射光譜。本研究測試了三種不同的機器學習演算法包括主成分分析(Principal component analysis, PCA)、線性判別分析(Linear discriminant analysis, LDA)及支持向量機(Support vector machine, SVM)，其中PCA為非監督式學習，LDA及SVM為監督式學習。 本研究成果指出使用適當的機器學習演算法能夠有效提升光譜的分析能力。PCA結果顯示，PCA能夠使訓練資料分群，但不適用於測試資料，其原因為PCA主要辨識出各濃度下自身之變異結果，而非不同有機物造成之光譜差異。LDA結果顯示，LDA訓練出之模型能夠100%正確分辨同一次實驗同一張MGD的測試資料，但因其演算法之限制使其容易產生過適情形，而無法通用於不同天不同張MGD的測試資料。SVM的分析結果顯示，SVM能夠有效提升我們對於光譜的分析能力，發現光譜偏移的異常情形，而在適當選取訓練資料或有正確的資料前處理的情形下，能夠100%正確分辨不同有機物之電漿放射光譜。

The development of microplasma enables real-time monitoring of the gaseous species using plasma emission spectroscopy. Quantification and identification of volatile organic compounds (VOCs) using optical emission of plasmas remain challenging. Machine learning (ML) is a technology to realize artificial intelligence. It using statistic models combine with optimal problems to perform specific tasks. Supervised ML examines data sets with labels while unsupervised ML handles data sets without labels. In this work, we presented the design and test of a portable microplasma generation device (MGD) for gaseous composition analysis. The MGD is a dielectric-barrier-discharge-type (DBD) microplasma and is low-cost and easy-to-fabricate. It is fabricated using the toner transfer method, which allows for making MGD in a low cost and flexible manner. This MGD is driven using a high-voltage module that is powered by 5-V power supply. We develop a platform which can acquired high quality and stable plasma emission spectroscopy efficiently by this device. Experiment results indicate that this platform can acquire and analyze the spectra of ethanol to achieve semi-quantification. Machine learning (ML) algorithms are used on the spectral analysis for discrimination of VOCs. Algorithms examined include principal component analysis (PCA), linear discriminant analysis (LDA), and support vector machine (SVM). PCA is an unsupervised ML algorithm while LDA and SVM are supervised ones. Preliminary studies show that machine learning with proper algorithm has great potential for spectroscopy analysis. It is found that PCA is unable to discriminate VOCs tested most likely due to the rather large variance of emission spectra. LDA are able to classify methanol-containing and ethanol-containing ambient even with different concentrations only when the spectra are acquired with a single MGD within the same day. LDA, however, is not able to properly classify different VOCs when spectra are acquired using multiple MGDs or within multiple days, due to the limitation of this algorithm. SVM is shown that is able to enhance the ability to analyze spectra. Furthermore, it can classify different VOCs ambient even when spectra are acquired using multiple MGDs or within multiple days while used proper training data.