電子鼻是近代一項重要發明，運用傳感器和演算法模擬生物嗅覺機制，進而對氣體進行分析和辨識，在工業氣體監控、肺癌檢測等領域都有其重要性。在電子鼻演算法方面，影響氣體辨識率的原因主要有三個，一、特徵萃取，二、分類器，三、外在變因影響。在第一、二點方面，早期大多取氣體反應中的斜率變化或氣體反應前後的變化率作為特徵，並使用 KNN、SVM 等機器學習演算法對氣體進行辨識，而最近幾年開始有研究取氣體反應整體過程作為特徵圖，並使用 GasNet 等深度學習演算法分析氣體。在第三點方面，目前主要解決辦法是使用傳統機器學習對外在變數影響的氣體特徵進行偏移補償和校正。
本論文主要研究方向有兩個，其一是改善在深度學習分析氣體方面的特徵萃取方式和分類器結構；其二是使用跨模態(cross-modal) 深度神經網路透過深度學習改善環境變數及製程變異所造成的感測器飄移問題，本研究使用一個開源氣體資料庫，嘗試了幾種數據前處理方法和深度學習架構(ModLeNet-6、SimVGGNet-10、SimResNet-9) 進行氣體分析和比較後，選擇了其中辨識率最好的簡化版 ResNet (SimResNet-9)，使用跨模態深度神經網路進一步學習外在變因與氣體資料的關係以提升辨識準確率，並與早期和近期的氣體演算法做辨識率比較，以證明本研究方法可行性，最後使用 Cross-modal SimResNet-9 amd MLP 進行資料訓練和分類辨識，達到 95% 的辨識率。

An electronic nose is an important invention in modern times. Sensors and algorithms are used to simulate the biological olfactory mechanism and analyze & identify gases. It is critical for industrial gas monitoring and lung cancer detection. In terms of the electronic nose algorithm, there are three main reasons affecting gas recognition rate: 1. feature extraction, 2. classifier, 3. external variable. In the first and second points, most of the early studies took the variation of the slope in the gas reaction or the variation rate before and after the gas reaction, and then use machine learning algorithms, such as KNN and SVM, to identify the gas. In recent years, research has begun to use the overall process of gas reaction as a feature map and use deep learning algorithms, such as GasNet, to analyze gas data. In the third point, the main solution in recent years is to offset and calibrate the gas features affected by external variables by using Machine Learning.
There are two main research directions in this thesis. One is to improve the data preprocessing method and classifier structure in deep learning of gas analysis; the other is to use Cross-modal Deep Neural Networks to improve the sensor drift problem caused by environmental variables and process variation problems. This study used one open-source gas datasets, tried several data pre-processing methods and deep learning architecture (ModLeNet-6, SimVGGNet-10, SimResNet-9) for gas analysis and comparison, choose the best one(simplified ResNet (SimResNet-9)) to use cross-modal Deep Neural Networks to promote the accuracy of classification further by learning the relationship between external variables and gas data, and compare with early and recent gas algorithms to prove the feasibility of this research method. Finally, this paper use Cross-modal SimResNet-9 and MLP for data training and classification and reach 95% recognition rate.

摘要 i
Abstract ii
誌謝 iii
目錄 I
圖目錄 1
表目錄 3
第一章 緒論 4
1.1 生物嗅覺機制 4
1.2 電子鼻系統介紹 5
1.3 研究背景與動機 6
第二章 文獻回顧 7
2.1 傳統氣體辨識演算法 7
2.1.1 K個最鄰近分類器(K-nearest neighbor classifier, KNN) 7
2.1.2 支持向量機(Support Vector Machine, SVM) 8
2.2 CNN氣體辨識演算法 10
2.2.1 LeNet 10
2.2.2 VGGNet 11
2.2.3 ResNet 11
2.2.4 CNN 架構中常用技巧 14
2.2.5 CNN於氣體分析上的應用範例 16
2.3 外在變因問題 19
2.3.1 製程變異 19
2.3.2 感測器飄移 19
2.3.3 製程變異校準和飄移補償 20
2.4 跨模態深度神經網路(Cross-modal Deep Neural Networks) 20
第三章 實驗數據與軟硬體工具介紹 22
3.1 Gas sensor arrays in open sampling settings Data Set 22
3.2 電腦配置與軟體工具 24
第四章 數據處理程序 25
4.1 數據前處理 (Data Preprocessing) 25
4.1.1 KNN與SVM特徵萃取 26
4.1.2 CNN特徵萃取與圖形生成 27
4.1.3 其他特徵參數 30
4.2 CNN 架構建立 31
4.2.1 調整版LeNet 31
4.2.2 簡化版VGGNet 33
4.2.2 簡化版ResNet 35
4.3 CNN-MLP cross-modal structure 37
第五章 實驗結果與討論 39
5.1 分析結果比較與討論 39
5.2 整體結果討論 45
第六章 結論與未來發展 46
參考文獻 47

[1] 林怡文,“一「嗅」萬千的巧妙,”科學人雜誌, 2012. [Online].Available: http://sa.ylib.com/MagCont.aspx?Unit=easylearn&id=1900.

[2] K. Persaud and G. Dodd, “Analysis of discrimination mechanisms in the mammalian olfactory system using a model nose”, Nature, vol. 299, pp. 352– 355, 1982.

[3] 邱仕文,“智慧電子鼻”Journal of Information and Communication Technology, 2018. [Online]. Available:
https://ictjournal.itri.org.tw/Content/Messagess/contents.aspx?MmmID=654304432070702333&MSID=1001520422536221724

[4] Y. Yin, “Discrimination between Chinese Jing Wine and Counterfeit Using Different Signal Features of an Electronic Nose,” Journal of Sensor Technology, vol. 02, no. 03, pp. 109–115, 2012.

[5] P. C. Jain and R. Kushwaha, “Wireless gas sensor network for detection and monitoring of harmful gases in utility areas and industries,” in 2012 Sixth International Conference on Sensing Technology (ICST), 2012, pp. 642–646.

[6] C. Di Natale, A. Macagnano, E. Martinelli, R. Paolesse, G. D’Arcangelo, C. Roscioni, A. Finazzi-Agrò, and A. D’Amico, “Lung cancer identification by the analysis of breath by means of an array of non-selective gas sensors,”Biosensors and Bioelectronics, vol. 18, no. 10, pp. 1209–1218, Sep. 2003.

[7] Pardo, M., & Sberveglieri, G. “Classification of electronic nose data with support vector machines.” Sensors and Actuators B Chemical, 107,730–737, (2005)..

[8] Estakhroueiyeh, H.R., Rashedi, E.: “Detecting moldy bread using an e-nose and the KNN classifier.” In: 2015 5th International Conference on Computer and Knowledge Engineering (ICCKE), pp. 251–255. IEEE (2015)

[9] Peng, P.; Zhao, X.; Pan, X.; Ye, W. “Gas Classification Using Deep Convolutional Neural Networks.” Sensors 2018, 18, 157.

[10] Wei, G.; Li, G.; Zhao, J.; He, A. “Development of a LeNet-5 Gas Identification CNN Structure for Electronic Noses.” Sensors 2019, 19, 217.

[11] R. O. Duda, P. E. Hart, and D. G. Stork, Pattern classification. Wiley, 2001.

[12] J. C. Burges, “A Tutorial on Support Vector Machines for Pattern Recognition,” Data Min. Knowl. Discov., vol. 2, pp. 121–167, 1998.

[13] Support-vector machine from Wikipedia [Online]. Available:
https://en.wikipedia.org/wiki/Support-vector_machine

[14] LeCun, Y., Bottou, L., & Haffner, P. (1998). “Gradient-based learning applied to document recognition.” Proceedings of the IEEE, 86(11), 2278–2324.

[15] Yugandhar Nanda, “What is the VGG neural network.” [Online]. Available: https://www.quora.com/What-is-the-VGG-neural-network (2018).

[16] K. Simonyan and A. Zisserman. “Very deep convolutional networks for large-scale image recognition.” arXiv preprint arXiv:1409.1556, 2014.

[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015. ”Deep residual learning for image recognition”. arXiv preprint arXiv:1512.03385.

[18] Rupesh Kumar Srivastava, Klaus Greff, and Jurgen Schmidhuber. “Highway networks.” arXiv preprint arXiv:1505.00387, 2015.

[19] X. Glorot, A. Bordes, and Y. Bengio. 2011.“Deep sparse rectifier neural networks.” In AISTATS’ 2011.

[20] Mr. Opengate,“深度學習：使用激勵函數的目的、如何選擇激勵函數 Deep Learning : the role of the activation function”, 2017 [Online]. Available: https://mropengate.blogspot.com/2017/02/deep-learning-role-of-activation.html

[21] Ioffe, S., Szegedy, C. “Batch normalization: accelerating deep network training by reducing internal covariate shift.” arXiv preprint arXiv:1502.03167 (2015)

[22] 郭耀華,“【深度學習】深入理解Batch Normalization 批標準化”, 2018 [Online]. Available: https://www.cnblogs.com/guoyaohua/p/8724433.html

[23] Srivastava, N., Hinton, G., Krizhevsky, A., et al.: “Dropout: a simple way to prevent neural networks from overfitting”. J. Mach. Learn. Res. 15(1), 1929–1958 (2014)

[24] Ying Zhang and Huchuan Lu, “Deep cross-modal projection learning for image-text matching,” in ECCV, 2018, pp. 707–723.

[25] Cangea, C., Velickovic, P., Liò, P., “Xflow: 1d-2d cross-modal deep neural networks for audiovisual classification.” CoRR abs/1709.00572 (2017)

[26] A. Vergara, J. Fonollosa, J. Mahiques, M. Trincavelli, N. Rulkov, R. Huerta, “On the performance of gas sensor arrays in open sampling systems using inhibitory support vector machines”, Sens. Actuators B: Chem. 185 (2013) 462–477

[27] J. E. Shore, R. M. Gray, "Minimum cross-entropy pattern classification and cluster analysis", IEEE Trans. Pattern Anal. Mach. Intell, 1982.

[28] Kingma, Diederik P and Ba, Jimmy Lei. “Adam: A method for stochastic optimization.”, arXiv:preprint arXiv:1412.6980 (2014).