深度學習導向之光譜分析於電漿診斷與氣體檢測之研究

Abstract

近年來，深度學習技術在電漿光譜分析領域展現出強大潛力。然而，現有應用仍面臨多項挑戰，包括光譜儀解析度受限、模型無法提供預測信心水準，以及針對不同資料場景 (domain) 需重新訓練模型等問題。本論文針對上述挑戰，提出一系列具體且有效的解決方案，進一步推進深度學習於電漿光譜分析的應用範疇。 首先，為克服光譜儀硬體解析度不足的限制，本研究開發了Spec-GAN架構，用以將低解析度 (low-resolution, LR) 光譜轉換為高解析度 (high-resolution, HR) 光譜。實驗以氮氣電漿的second positive emission system放光作為示範。訓練資料使用SpecairTM模擬軟體產生之高解析度與低解析度光譜配對，涵蓋300至1200 K轉動溫度 (Trot) 與2000至6500 K振動溫度 (Tvib) 範圍。測試資料則取自低壓與大氣電漿之電漿光譜，以驗證模型在僅使用低解析度光譜儀獲得光譜時產生高解析度光譜的可行性。訓練過程中導入特徵匹配技術 (feature matching) 以緩解訓練不穩定問題，並利用鑑別器分數的分佈作為監控訓練過程的初步依據。結果顯示，模擬高解析度光譜與生成高解析度光譜之間的加權決定係數 (R̅²) 均大於0.9999。進一步比較生成高解析度光譜與實驗高解析度光譜所擬合出的轉動溫度與振動溫度，其誤差多數低於5%。整體結果表明，當無法取得高解析度光譜儀時，本研究提出的Spec-GAN架構可作為一種高效的替代手段以獲取高解析度光譜。 其次，針對濃度預測與消除概念偏移 (concept shift)，本研究提出MTL-VOCNet架構，其為結合分類與迴歸任務的多任務深度學習架構，針對揮發性有機物電漿光譜進行識別與濃度預測，並引入特徵調變模塊以消除不同物質間的概念偏移。實驗中以三個自製微電漿產生裝置 (microplasma generation device, MGD) 收集全新的測試光譜資料，模擬高變異性且無標記的實際應用情境。結果顯示，MTL-VOCNet在未見過之MGD資料中達到近乎100% 的分類準確率，並在濃度迴歸任務中，透過卷積Transformer迴歸架構實現平均絕對百分比誤差僅0.68%，遠優於卷積神經迴歸架構之2.88%、單任務卷積神經迴歸模型之8.6% 以及檢量線方法之13至20% 誤差表現。在預測未知濃度測試中，MTL-VOCNet維持5.5% 左右的誤差，相較於檢量線方法12.2% 左右誤差，展現優異泛化性能。進一步以均勻流形近似與投影 (uniform manifold approximation and projection, UMAP) 與基於遮蔽法之拉曼光譜特徵擷取 (occlusion-based Raman spectra feature extraction, ORSFE) 技術證實，特徵調變模塊可有效消除不同揮發性有機物類別間的特徵差異。整體而言，本研究展示MTL-VOCNet在電漿光譜分析上的高度準確性與消除概念偏移之潛力，為未來電漿診斷與檢測應用提供重要技術基礎。 最後，針對模型預測信心問題，本研究建構了Bayes-CNN架構，其基於電漿光譜之揮發性有機物分類框架，並特別強調預測之信心水準與分佈外資料 (out-of-distribution, OOD) 偵測之能力。為提升模型在不同微電漿產生裝置 (MGD) 間的泛化能力，本研究結合了參數遷移學習 (parameter transfer learning, PTL)，使Bayes-CNN在五組未曾見過的MGD資料上達到超過96% 的分類準確率與0.9604的平均F1分數。透過將單一資料點多次推論所得的標準差作為信心閾值，模型分類準確率可進一步提升至100%，但僅有76% 的資料被視為可靠資料，展現出預測可靠性與資料涵蓋率間的彈性取捨能力。Bayes-CNN亦能有效識別分佈外資料，其預測機率約33至70%，顯著較低，且標準差約0.05至0.25，顯著較高。此外，透過ORSFE熱圖分析展現模型可解釋性，結果顯示Bayes-CNN對物理上具意義之放光波段，如CH、C2與Ar具有高度關注。綜合而言，本研究展示了Bayes-CNN在複雜電漿光譜分析中，作為一種穩健、具信心感知 (confidence-aware) 且可解釋性的分析工具之潛力。

In recent years, deep learning has demonstrated great potential in the field of plasma optical emission spectroscopy (OES). However, current applications still encounter several challenges, including limited spectrometer resolution, the absence of confidence estimation in model predictions, and the need to retrain models for different domains. This dissertation addresses these challenges by proposing a series of specific and effective solutions to further advance the application of deep learning in plasma OES. First, to overcome the hardware limitation of spectral resolution, this study has developed the Spec-GAN architecture, which transforms low-resolution (LR) spectra into high-resolution (HR) spectra. Plasma emissions with second positive system of nitrogen are used for demonstration. The training set consists of paired HR and LR spectra generated by the simulation software, SpecairTM, covering rotational temperatures (Trot) and vibrational temperatures (Tvib) ranging from 300 to 1200 K and 2000 to 6500 K, respectively. The testing set is acquired from both low-pressure and atmospheric-pressure plasmas to evaluate the ability of Spec-GAN to generate HR spectra from experimental LR spectra. To stabilize the training procedure, feature matching is introduced, and the distribution of discriminator scores is used to monitor the training progress. The results show that the weighted coefficient of determination (R̅²) between simulated and generated HR spectra exceeds 0.9999. Additionally, when fitting the to extract, the fitting errors for Trot and Tvib between generated HR spectra and experimental HR spectra are mostly below 5%. These results demonstrate that Spec-GAN is an effective alternative for obtaining HR spectra when high-resolution spectrometers are not available. Second, to improve the performance of concentration prediction and eliminate concept shift across domains, this study proposes the MTL-VOCNet architecture, a multi-task learning (MTL) model combining classification and regression for plasma OES of volatile organic compounds (VOCs). A feature modulation block (FMB) is introduced to eliminate concept shift between VOC types. Experimental data are collected using three entirely unseen microplasma generation devices (MGDs) to simulate highly variable and unlabeled real-world scenarios. The results show that MTL-VOCNet achieves nearly 100% classification accuracy on unseen MGDs. For regression tasks, the architecture of convolutional transformer network (CTN) reaches a mean absolute percentage error (MAPE) of only 0.68%, significantly outperforming the architecture of convolutional neural network (CNN) with 2.88%, the single-task regression model with 8.6%, and traditional calibration curve methods with 13 to 20%. Under scenarios involving unknown concentrations, MTL-VOCNet maintains an error of around 5.5%, compared to 12.2% for the calibration curve approach, demonstrating excellent generalization. Further analysis using Uniform Manifold Approximation and Projection (UMAP) and occlusion-based Raman spectra feature extraction (ORSFE) confirms that the FMB effectively reduces variability among different VOCs. Overall, this study demonstrates that MTL-VOCNet delivers highly accurate plasma spectral analysis while effectively eliminating concept shift, providing a robust technical foundation for future applications in plasma diagnostics and sensing. Lastly, to address the issue of confidence estimation in model predictions, this study construct a Bayes-CNN architecture for VOC classification based on plasma OES. Bayes-CNN emphasizes confidence estimation and the ability to detect out-of-distribution (OOD) data. To enhance generalization across different MGDs, parameter transfer learning (PTL) is incorporated, enabling Bayes-CNN to achieve over 96% classification accuracy and an average F1 score of 0.9604 on five unseen MGDs. By using the standard deviation of multiple inferences for a single data point as a confidence threshold, the classification accuracy further increases to 100%, though only 76% of the data is considered reliable, reflecting the trade-off between prediction reliability and reliable data coverage. Bayes-CNN also effectively identifies OOD data, with prediction probabilities ranging from 33 to 70% and standard deviations between 0.05 and 0.25. Furthermore, model interpretability is demonstrated using ORSFE heatmaps, revealing that Bayes-CNN pays close attention to physically meaningful emission bands such as CH, C2, and Ar. Overall, the study highlights Bayes-CNN as a robust, confidence-aware, and interpretable tool for complex plasma spectral analysis.