基於微控制器之高壓模組結合微電漿光譜與深度學習於有機氣體檢測之應用

Abstract

常壓微電漿因其體積小、功耗低與製造成本低等優勢，近年來在氣體檢測領域中備受重視，並展現出克服傳統檢測技術限制的潛力。本研究旨在建立一套具備優異操作性與高選擇性的揮發性有機物檢測平台，並結合卷積神經網路以實現高準確度之光譜辨識系統。整體研究可分為三大部分：首先，進行高壓模組之操作特性分析；其次，探討微電漿檢測系統在不同操作參數下的光譜與電訊表現；最後，結合資料增強技術與卷積神經網路 (Convolutional Neural Network, CNN) 模型，建立可有效分辨揮發性有機物 (Volatile Organic Compound, VOC) 的深度學習分類系統。 在檢測平台建立方面，為提升系統之可調性與操作穩定性，本研究採用微控制器產生脈衝訊號以驅動高壓模組，並透過微型電腦協同控制微控制器與光譜儀，以實現光電訊號的同步擷取。藉由快速脈衝調變 (FAST PWM) 技術精準控制脈衝參數，顯著提升了電漿的穩定性；同時，採用間歇式電漿控制模式有效延長了微電漿產生裝置 (MGD) 的使用壽命。以本系統進行VOC檢測時，在定條件模式下觀察三種VOC的光譜特徵發現，其有機特徵峰強度雖隨濃度變化而變動，但三者間存在顯著的重疊區間，造成辨識困難。為提升辨別能力，本研究進一步採用多條件模式，於25組不同脈衝條件下對三種VOC進行分析，結果顯示其相對強度變化趨勢具明顯差異，驗證多條件掃描有助於提升不同VOC間的可辨識性。 在深度學習的部分，本研究使用卷積神經網路對不同VOC之多條件光譜進行分析，透過擷取其不同VOC隨條件變化之差異進行辨別，並透過資料增強技術增加訓練資料量提升模型表現。通過比較定條件及多條件之模型表現，確定了多條件光譜有利於模型有更高的穩健性及泛化能力，並且我們僅需收集5筆多條件光譜即可利用資料增強技術將資料量增加100倍而在三種VOC皆獲得高達0.99的分類表現，此外，透過改變訓練資料的MGD資料，發現，我們僅需要使用3張MGD就可使模型在4張MGD的測試準確率獲得0.9左右的分類表現，顯示了多條件資料結合資料增強技術在分辨VOC電漿光譜的潛力。

Atmospheric pressure microplasmas have received increasing interest in recent years for gas detection applications due to their compact size, low power consumption, and low fabrication cost. These systems show great potential in overcoming the limitations of conventional detection technologies. This study aims to establish a volatile organic compound (VOC) detection platform with excellent operational controllability and high selectivity, integrated with a convolutional neural network (CNN) to enable high-accuracy spectral classification. The research is divided into three main parts: (1) characterization of the high-voltage module’s operating behavior, (2) analysis of the microplasma-based detection system under various operating parameters with respect to its spectral and electrical responses, and (3) development of a deep learning classification model for VOC classification using data augmentation and CNNs.

To enhance the adjustability and operational stability of the detection platform, a microcontroller is employed to generate pulse signals for driving the high-voltage module, while a microcomputer synchronously controls both the microcontroller and spectrometer for simultaneous optical-electrical data acquisition. By applying FAST PWM control, the system achieves precise pulse modulation and improved plasma stability. A burst plasma operation mode is also adopted to significantly extend the lifetime of the MGD. In fixed-condition measurements, the spectral features of three VOC were found to exhibit overlapping intensity ranges despite concentration variation, resulting in poor discriminability. To address this, a multi-condition scanning mode involving 25 different pulse conditions was implemented, revealing distinctive variation trends in relative spectral intensities across VOC, thereby enhancing their differentiability. In the deep learning section, a CNN was trained on multi-condition spectra of different VOC to learn their condition-dependent spectral variations for accurate classification. Data augmentation was employed to increase training data volume and improve model performance. Comparative analysis confirmed that multi-condition spectra contributed to better model robustness and generalization compared to fixed-condition data. With only 5 raw spectra per condition, data augmentation increased the dataset 100-fold and enabled classification accuracies as high as 0.99 across all three VOC. Furthermore, using data from just three MGDs for training was sufficient to achieve approximately 0.90 accuracy on spectra from four unseen MGDs, highlighting the effectiveness of combining multi-condition data and augmentation in enhancing the generalizability of plasma-based VOC classification.