準確預測加工時間，對於預測交貨時間以及製造成本非常重要。對於使用電腦數值控制(Computer Numerical Control, CNC)工具機進行小批量生產時，評估數值控制(Numerical Control, NC)程式的循環時間會是製造成本和調度排程的關鍵問題。本研究將提出一系列的方法進行收集資料、前處理輸入資料、以及開發深度學習模型。由於NC程式可被視為一種機器語言，因此本研究使用長短期記憶(Long Short-Term Memory, LSTM)模型做為此深度學習模型的核心。
首先，使用CAM軟體分析產品CAD檔之後建立刀具路徑與產生NC程式。接著使用CNC執行NC程式並取得運行資料，再將NC程式與運行資料經過擬合處理，合併為單節運行時間資料。接著，使用簡單的加工案例對研究方法進行驗證和調整，找出合適的深度學習模型。較佳的模型為一層64節點的LSTM、一層32節點的LSTM、以及一層單節點的Dense；適合的輸入資料為單節之刀具路徑長除以單節之進給速度，且適合的輸出資料為單節運行時間，並在複雜的加工案例進行驗證測試。
本研究提出評估準確度的指標應為總時間誤差率、單節平均誤差值、以及實際結果與預估結果的相關係數。在相關的案例中，加工時間的預測總時間誤差率約在0.01%與0.2%間，單節平均誤差值約在0.005秒與0.041秒間，以及實際結果與預估結果的相關係數約在96%與99.9%間。

關鍵字：循環時間、(Long Short-Term Memory, LSTM)、小批量生產、電腦數值控制(Computer Numerical Control, CNC)、數值控制(Numerical Control, NC)、深度學習。

Accurate prediction of machining time is very important for estimating delivery time and manufacturing costs. When using Computer Numerical Control (CNC) machine for small batch production, evaluating the cycle time of Numerical Control (NC) programs is a key issue for manufacturing costs and scheduling. This study propose a series of approaches for collecting data, pre-processing input data, and developing deep learning models. Since NC programs can be regarded as a kind of machinery language, this study uses the Long Short-Term Memory (LSTM) as the core of the deep learning model.
Firstly, CAM software is applied to analyze product CAD files to create tool paths and generate NC programs. Then a CNC machine runs the NC programs to generate the dynamic data. The NC program and the dynamic data are combined into a set of single block data. After that, regular machining cases are applied to study the deep learning model The best model is one 64-node LSTM layer, one 32-node LSTM layer, and an one-node dense layer as the outputting layer. The suitable input data is the ratio of distance and feed rate, the output data is runtime of the target block, and conduct verification tests on the complex machining cases.
The study proposed the performance indexes to evaluate the performance of our model including the error rate of total cycle time, the average runtime error of single blocks, and the correlation coefficient between the actual results and estimated results. In the related cases, the estimated total time error rate of processing time is between 0.01% and 0.2%, the average error of single block is between 0.005 sec and 0.041 sec, and the correlation coefficient of actual results and estimated results is between 96 % and 99.9 %.

Keywords: Cycle Time, Long Short-Term Memory (LSTM), Small Batch Production, Computer Numerical Control (CNC), Numerical Control (NC), Deep Learning.

[1] C. Ou-Yang, T. S. Lin, “Developing an integrated framework for feature-based early manufacturing cost estimation”, The International Journal of Advanced Manufacturing Technology, 13(9), 618–629, September 1997.
[2] M. Hbaieb, et al, “Time modeling in high-speed machining of mold pocket”, The International Journal of Advanced Manufacturing Technology, 53(1-4), 133-120, July 2010.
[3] R. Othmani, et al, ”Machining time in rough milling”, Materials Technology, 23(3), 169–173, 2008.
[4] B. S. So, et al, “Five-axis machining time estimation algorithm based on machine characteristics”, Journal of Materials Processing Technology, 187-188, 37-40, 2007.
[5] H. Siller, et al, ”Cycle time prediction in high-speed milling operations for sculptured surface finishing”, Journal of Materials Processing Technology, 174(1-3), 355–362. February 2006.
[6] R. T. Coelho, et al, “Mechanistic approach to predict real machining time for milling free-form geometries applying high feed rate”, The International Journal of Advanced Manufacturing Technology, 46(9 12), 1103–1111, July 2009.
[7] E. Y. Heo, et al, “Estimation of NC machining time using NC block distribution for sculptured surface machining”, Robotics and Computer-Integrated Manufacturing 22, 437–446.December 2005.
[8] C. Q. Liu, et al, “A feature-based method for NC machining time estimation”, Robotics and Computer-Integrated Manufacturing, Volume 29, Issue 4, Pages 8-14 , August 2013.
[9] S. Tulsyan, “Prediction and Reduction of Cycle Time for Five-Axis CNC Machine Tools”, THE UNIVERSITY OF BRITISH COLUMBIA, Master, December 2014.
[10] 張斐章、張麗秋，類神經網路導論原理與應用，二版，蒼海圖書，新北市，民國一百零三年。
[11] H. Jahan-Shahi, et al, “Multivalued Fuzzy Sets in Cost/Time Estimation of Flat Plate Processing”, The International Journal of Advanced Manufacturing Technology, 17(10), 751–759. 2001.
[12] T. Saric, et al, “Estimation of Machining Time for CNC Manufacturing Using Neural Computing”, International Journal of Simulation Modelling 15(4):663-675, December 2016.
[13] 黃政豪，「應用類神經網路預測CNC加工中心機加之工時間」，國立台灣科技大學機械工程學系，碩士論文，2018。
[14] J. L. Wang, et al, “Bilateral LSTM: A Two-Dimensional Long Short-Term Memory Model With Multiply Memory Units for Short-Term Cycle Time Forecasting in Re-entrant Manufacturing Systems”, IEEE Transactions on Industrial Informatics, 14(2), 748–758. February 2018.
[15] Toshiaki Otsuki, Osamu Hanaoka, “Numerical controller with machining time prediction unit and machining error prediction unit”, Publication of US8938316B2. 2011.
[16] B. H. Kim, B. K. Choi, “Machining efficiency comparison direction-parallel tool path with contour-parallel tool path”, Computer-Aided Design, 34(2), 89–95. 2002.
[17] H. W. Chiu, C. H. Lee, “Prediction of machining accuracy and surface quality for CNC machine tools using data driven approach”, Advances in Engineering Software Volume 114, Pages 246-257, December 2017.
[18] 周國華，工具機應用設計，初版，翰蘆圖書，臺北市，民國一百零二年。
[19] 陳進郎，數控工具機，第五版，全華圖書，新北市，民國一百零一年。
[20] 智研科技有限公司客戶支援網站。2020年，取自：http://support.intek.com.tw/。
[21] F. Chollet，Deep learning 深度學習必讀：Keras大神帶你用Python實作，葉欣睿譯，旗標，臺北市，2019。
[22] 如何用人工神經網路處理決策邊界問題？這裡有一份簡單的代碼參考。2017/02/22取自：https://zi.media/@yidianzixun/post/ru4HFw。
[23] 梯度下降法。2020年，取自：https://zh.Wikipedia.org/wiki/梯度下降法。
[24] 齊藤康毅，Deep Learning|用Python進行深度學習的基礎理論實作，吳嘉芳譯，碁峰資訊，臺北市，2017。
[25] LSTM_深度學習_股價預測。JT。2018，取自：medium.com/data-scientists-playground/lstm-深度學習-股價預測-cd72af64413a。
[26] 工程師。日常。AI技術團隊。2018年5月17日，取自：https://ai4dt.wordpress.com/2018/05/17/recurrent-neural-network-遞歸神經網絡-rnn-part-1/。
[27] Understanding LSTM Networks。Colah’s blog。2015年8月27日，取自：http://colah.github.io/posts/2015-08-Understanding-LSTMs/。
[28] a51nx Horizontal Machining 4-Axis。MAKINO官網。2020年，取自https://www.makino.com/machine-technology/machines/horizontal-4-axis/a51nx。