工業4.0(Industry 4.0)智慧製造(Smart Manufacturing)是近來很熱門的議題。現今全球製造業致力於藉由工業物聯網、大數據分析、虛實融合系統(Cyber Physical System, CPS)等技術實現智慧製造(Smart Manufacturing)，以便能夠在節省生產時間與成本的情況下，提高生產品效能與產品品質。本論文探討虛擬量測(Virtual Metrology, VM)研究，在生產過程尚未完成前或完成後，在不需要或無法實際量測產品的情況下預測產品品質。具體的說，本論文聚焦於線切割放電加工機台加工產品表面粗糙度(Surface Roughness)預測，利用2次回歸方程及深度神經網路方法，在產品加工前透過生產參數預測產品表面粗糙度。另外，本論文利用馬可夫鏈(Markov Chain)配合深度神經網路(Deep Neural Network, DNN)方法，在產品加工完成後，透過生產參數及生產過程機台狀態時序性資料預測產品表面粗糙度。為了處理長度不同的時序性資料，本論文利用馬可夫鏈提取特徵以此將長度歸一化，再透過神經網路來預測產品品質。我們透過全因子實驗方法進行實驗數據蒐集以驗證所提方法的預測準確度，實驗結果顯示，本論文所提預測方法具有良好的平均絕對誤差及誤差率。

Industry 4.0 Smart Manufacturing is a hot topic recently. Today's global manufacturing industry is committed to smart manufacturing through industrial Internet of Things, big data analytics, and Cyber Physical System (CPS) technologies to improve product performance and product quality by saving production time and cost. This paper explores Virtual Metrology (VM) research to predict product quality before or after the production process has not been completed, without the need of product measurement. Specifically, this paper focuses on the Surface Roughness prediction of wire-cut EDM machines, and uses the 2nd order regression and the deep neural network method to predict the surface roughness of the product through production parameters before product processing. In addition, this thesis uses Markov Chain and Deep Neural Network (DNN) method to predict the surface roughness of the product through production parameters and time series data of the production process after the product is processed. In order to deal with the time series data with different lengths, this paper uses the Markov chain extraction feature to normalize the length and then predict the product quality through the neural network. We use the full factor experimental method to collect experimental data to verify the prediction accuracy of the proposed method. The experimental results show that the proposed prediction method has good mean absolute error and error rate.

中文摘要 I
Abstract II
誌謝 III
目錄 IV
圖目錄 VI
表目錄 VIII
一、 緒論 1
1.1. 研究背景與動機 1
1.2. 研究目的與貢獻 2
1.3. 相關文獻探討 3
1.4. 論文架構 3
二、 背景知識與相關文獻探討 4
2.1. 放電加工 4
2.1.1. 放電加工簡介 4
2.1.2. 線切割放電加工 5
2.2. 馬可夫鏈 7
2.2.1. 馬可夫鏈簡介 7
2.2.2. 馬可夫時序預測法(Markov Forecasting Model) 8
2.3. 人工神經網路 10
2.3.1. 人工神經網路簡介 10
2.3.2. 反向傳播演算法(Back-Propagation Algorithm) 16
2.4. 深度學習 18
2.4.1. 深度學習介紹 18
2.4.2. 深度神經網路(Deep Neural-Network) 19
2.5. 相關文獻 20
2.5.1. Design of Experiment 21
2.5.2. Box-Behnken design 22
三、 問題定義與研究 24
3.1. 問題定義 24
3.2. 標籤定義 27
3.3. 資料前處理 28
3.3.1. 取出可用資料 28
3.3.2. 特徵萃取 29
3.3.2.1. 資料長度歸一化 30
3.3.2.2. 資料轉換馬可夫鏈形式 31
3.4. 模型架構 35
3.5. 訓練最佳化 37
四、 實驗與分析 39
4.1. 實驗環境 39
4.1.1. 硬體設備 39
4.1.2. 訓練框架 39
4.2. 實驗結果 42
五、 結論與未來展望 47
參考文獻 48

[1] Virtual Methodology:
https://ejournal.stpi.narl.org.tw/sd/download?source=10210-08.pdf&vlId=E587FCF7-697D-4DC8-9820-EAEE679454BC&nd=1&ds=1
[2] K.H Ho, S.T Newman, S. Rahimifard, R.D Allen. “State of the art in wire electrical discharge machining (WEDM)”. In International Journal of Machine Tools and Manufacture. pp 1247-1259, 2004
[3] K.H. Ho, S.T. Newman. “State of the art electrical discharge machining (EDM)”. In International Journal of Machine Tools and Manufacture. pp 1287-1300, 2003
[4] H. Ozkan, F. Ozkan, S. S.Kozat, “Online Anomaly Detection Under Markov StatisticsWith Controllable Type-I Error”. IEEE Trans. Signal Processing. 64(6), 1435–1445, 2016
[5] U. Esme, A. Sagbas, F. Kahraman. “Prediction of Surface Roughness in Wire Electrical Discharge Machining Using Design of Experiments and Neural Networks”. In Iranian Journal of Science & Technology, Transaction B, Engineering. pp 231-240, 2009
[6] A. Kumar, V. Kumar, J. Kumar. “Prediction of Surface Roughness in Wire Electric Discharge Machining (WEDM) Process based on Response Surface Methodology”. In International Journal of Engineering and Technology, 2012
[7] 類神經網路-感知機的原理及實作:
https://1fly2sky.wordpress.com/2017/02/14/%E9%A1%9E%E7%A5%9E%E7%B6%93%E7%B6%B2%E8%B7%AF-%E6%84%9F%E7%9F%A5%E6%A9%9F%E7%9A%84%E5%8E%9F%E7%90%86%E4%BB%A5%E5%8F%8A%E5%AF%A6%E4%BD%9C/
[8] JavaScript neural network implementation:
https://blog.toright.com/posts/5234/javascript-%E5%AF%A6%E7%8F%BE%E9%A1%9E%E7%A5%9E%E7%B6%93%E7%B6%B2%E8%B7%AF-%E7%80%8F%E8%A6%BD%E5%99%A8-deep-learning-%E5%A5%BD%E6%A3%92%E6%A3%92.html
[9] J. E. Lee, J. R. Jiang, “Time Series Multi-Channel Convolutional Neural Network for Bearing Remaining Useful Life Estimation”, 2018
[10] 入門深度學習
https://medium.com/@syshen/%E5%85%A5%E9%96%80%E6%B7%B1%E5%BA%A6%E5%AD%B8%E7%BF%92-2-d694cad7d1e5
[11] Accutex - GE Series:
https://www.accutex.com.tw/products.htm
[12] Design of Experiment
Ronald A. Fisher. “The Design of Experiments”. England : Macmillan Pub Co. 1935
[13] Electronica - Sprintcut 734:
https://electronicagroup.com/cnc-wirecut-edm/
[14] G. Box, D. Behnken, “Some new three level designs for the study of quantitative variables”, Technometrics, Volume 2, pages 455–475, 1960.
[15] CHMER - Q4025L:
http://www.chmer.com/tw/products-view.php?id=76
[16] Tokyo Seimitsu - Surfcom 130A:
http://www.accretech.com.cn/surfcom.html
[17] D. E. Paul, B. J. T, K. Ronald. A, “Materials and Processes in Manufacturing (9th ed.)”, Wiley, ISBN 0-471-65653-4. 2003
[18] D. Zang, J. Liu and H. Wang. “Markov Chain-Based Feature Extraction for Anomaly Detection in Time Series and Its Industrial Application”. In 2018 Chinese Control And Decision Conference (CCDC). pp1059-1063. 2018
[19] G. Klambauer, T. Unterthiner, A. Mayer and S. Hochreiter. “Self-Normalizing Neural Networks”. arXiv preprint arXiv:1706.02515.
[20] Samuel B. Green. “How Many Subjects Does It Take To Do A Regression Analysis”. Multivariate Behavioral Research. pp 499-510. 2010
[21] D. P. Kingma, J. Ba. “Adam: A method for stochastic optimization”. arXiv preprint arXiv:1412.6980. 2014