深度學習結合電漿光譜法於揮發性有機物之研究

Abstract

本研究使用卷積神經網路結合電漿光譜判別氬氣中的 VOC。實驗使用了 10 張 MGD 收取了 3 種不同 VOC 在 5 種濃度下的氬氣光譜共 75k 張。實驗分為四個面相：卷積神經網路之應用及限制、轉移學習對減少資料及提高正確率的效果、集成學習及貝葉斯神經網路對資料可信度的分析，以及可解釋人工智慧之解讀。 在實驗中，首先通過拆分光譜、使用 CBAM、自定義前處理等方式改進卷積神經網路之表現，提高了模型的收斂速度和對光譜偏移的耐受度。在使用 CBAM 協助訓練的狀況下，訓練週期可以由 200 降低到 50，而在使用經自定義前處理方式處理過後的資料訓練後，模型對於偏移光譜的分類正確率由 0.33 提高為 0.97。 第二部分分別比較了隨機起始、參數轉移、REPTILE、以及自我學習之表現，實驗證明轉移學習使用源資料後，可以降低訓練時對於目標資料量的需求並提高測試時的正確率。參數轉移以及 REPTILE 在僅有 225 張光譜的狀況下，分別獲得了高達 0.98 以及 0.95 之正確率，而未使用轉移學習之模型僅有 0.82 之正確率。 在第三部分使用貝葉斯神經網路判別結果可信度的實驗顯示，僅需使用同一個貝葉斯神經網路預測同一張光譜多次，即可藉由計算輸出值之標準差判斷分類結果的可信度，若是標準差極度接近 0，即可判定該分類結果 100%正確。而在最後則分別使用 Grad-CAM 和 ORSFE 來判別卷積神經網路判別之依據，結果顯示氬氣段的放光以及有機段的放光皆對模型的判斷有著重大影響。

This study applies convolutional neural networks (CNN) combined with plasma spectroscopy to discriminate volatile organic compounds (VOCs) in argon gas. The experiment involves collecting 75,000 argon spectra with three different VOCs at five concentrations using 10 microplasma generation devices (MGDs). The study consists of four aspects: the application and limitations of CNN in plasma spectroscopy, the effect of transfer learning on data reduction and accuracy improvement, the analysis of ensemble learning and Bayesian neural networks on data reliability, and the interpretation of explainable artificial intelligence. In the experiment, several approaches are employed to enhance the performance of CNN. These include spectrum splitting, CBAM utilization, and customized preprocessing, which improve the convergence speed of the model and its tolerance to spectral shifts. With CBAM assistance during training, the number of training cycles can be reduced from 200 to 50. Additionally, training the model with preprocessed data using a custom preprocessing method increases the classification accuracy for shifted spectra from 0.33 to 0.97. In the second part, the performance of random initialization, parameter transfer, REPTILE, and self-training is compared. The results demonstrate that transfer learning with source data can reduce the demand for target data during training and improve the accuracy during testing. With parameter transfer and REPTILE, accuracies as high as 0.98 and 0.95, respectively, are achieved with only 225 spectra, while the model without transfer learning achieves an accuracy of 0.82. The experiments in the third part focus on determining the reliability of classification results using Bayesian neural networks. It is found that by predicting the same spectrum multiple times using the same Bayesian neural network and calculating the standard deviation of the output values, the credibility of the classification results can be determined. When the standard deviation is extremely close to zero, the classification result can be considered 100% correct. Finally, Grad-CAM and ORSFE are employed to interpret the basis of CNN's classifications. The results show that both the emissions from the argon segment and the organic segment significantly influence the model's judgment.