在銑削製程中，當刀具發生老化或崩刃時，常會造成加工精度變異。目前許多研究均著重於透過蒐集足夠樣本以建立精度推估模型，然在現實加工環境中，為獲得足夠的機台稼動率，實難以中斷加工以蒐集足夠樣本。如此將造成樣本蒐集的困難，從而拉長模型可上線時間，更不適於少量多樣的加工型態。
本研究提出一可因應侷限樣本的加工品質推估方法。首先，利用不同感測源的特徵萃取機制，應用長短期記憶(Long Short Term Memory, LSTM)網路以推估刀具劣化的三階段曲線，可監控因刀具老化而產生的加工品質變異。其次，基於加工所蒐集的侷限樣本，透過生成對抗網路(Generative Adversarial Nets, GAN)可生成刀具劣化漸變過程的仿製樣本，作為建立LSTM模型所需之初期樣本，以減少實際樣本蒐集時間，從而縮短模型可上線時間。
研究結果顯示，在減少大部分的實際樣本下，基於實際感測源的侷限樣本與GAN所產生的仿製樣本，本研究所提出之LSTM初期模型，估測面粗度Ra的平均絕對誤差率(MAPE)可在20%以內；而當建模樣本全部為全實際樣本時，MAPE則可小於5%。換言之，藉由初期侷限樣本可快速建立推估模型，當加工過程有樣本可補充時，再逐步更新模型，如此將能利於少量多樣加工型態的品質監控。

In the milling process, variation in machining accuracy may cause from the tool is aging or chipping. At present, many researches focused on collecting sufficient samples to build the accuracy estimation models. However, in a real machining environment, collecting sufficient samples by interrupting machining is hard to maintain sufficient machine productivity. It will cause difficulties in sample collection and lengthen the time for modeling, and is infeasible for small and diverse production.
This study proposes a method for estimating processing quality with limited samples. First, a feature extraction mechanism with various sensing sources and a LSTM (Long Short Term Memory) network are applied to estimate the three-stage of tool degradation, so that quality variation in processing due to tool aging can be monitored. Secondly, based on the limited samples collected from machining and the GAN (Generative Adversarial Nets), the required samples can be generated to build the initial LSTM model. Hence, the time for collecting sample and the time to model release can be reduced.
The results show that with reducing half of the actual samples, the MAPE (Mean Absolute Percentage Error) of the surface roughness estimation is within 20% when applying the initial LSTM model with the limited samples collected from the actual sensors and the fake samples generated by GAN. In addition, the MAPE can be less than 5% when the model is remodeled with full actual samples. In other words, the time to model release can be efficiently reduced due to the time for collecting sample being reduced by applying the initial model. The model can be updated gradually when actual samples are available, which can be utilized to for small and diverse production.

摘要 i
Abtract ii
誌謝 iii
目錄 iv
表目錄 vii
圖目錄 ix
第1章 緒論 1
1.1 研究背景 1
1.2 研究目的 3
1.3 研究架構 4
第2章 理論方法 5
2.1 訊號來源 5
2.1.1 加速規 5
2.1.2 電流感測器 6
2.2 特徵計算方法 7
2.2.1 小波包分解 7
2.3 關鍵特徵挑選方法 8
2.3.1 LASSO (Least absolute shrinkage and selection operator) 8
2.4 加工品質估測方法 9
2.4.1 類神經網路(Neural Network) 9
2.5 仿製樣本生成方法 10
第3章 系統開發 11
3.1 系統架構 11
3.1.1 系統概念 11
3.1.2 系統流程 12
3.2 感測與製程資料蒐集 14
3.2.1 感測器安裝 14
3.2.2 加工製程 15
3.3 有效資料擷取 16
3.4 特徵計算 17
3.4.1 各項感測器所計算之小波包 17
3.4.2 特徵趨勢 19
3.5 類神經模型建立 22
3.5.1 NN模型(Neural Network) 22
3.5.2 LSTM模型(Long-Short-Term-Menory) 23
3.6 仿製樣本生成 24
第4章 案例研究與分析 25
4.1 實驗設計 25
4.1.1 實驗參數 25
4.2 各項感測器之加工品質估測 27
4.2.1 實驗一 27
4.2.2 實驗二 43
4.2.3 感測器估測結果比較 56
4.3 仿製樣本生成 57
4.3.1 基於GAN生成仿製樣本 57
4.4 仿製樣本可靠度驗證 72
4.4.1 實驗一仿製樣本可靠度驗證 73
4.4.2 實驗二仿製樣本可靠度驗證 76
4.4.3 仿製樣本可靠度比較 79
第5章 結論與未來展望 80
5.1 結論 80
5.2 未來展望 81
參考文獻 82

[1]B. J. Chen, X. F. Chen, B. Li, Z. J. He, H. R. Cao, and G. G. Cai, “Reliability estimation for cutting tools based on logistic regression model using vibration signals,” Mech. Syst. Signal Process. vol. 25, No. 7, pp. 2526-2537, 2011.
[2]C. A. Suprock and J. S. Nichols, “A low cost wireless high bandwidth transmitter for sensor-integrated metal cutting tools and process monitoring,” Int J Mechatron Manuf Syst, Vol.2, No. 4, pp. 441–454, 2009.
[3]N. Ghosh, Y.B. Ravi, A. Patra, S. Mukhopadhyay, S. Paul, A.R. Mohanty, A.B. Chattopadhyay, “Estimation of tool wear during CNC milling using neural network-based sensor fusion,” Mech Syst Signal Process Vol. 21, pp. 466–479, 2007.
[4]S. S. Abuthakeer, P. V. Mohanram and G. Mohan Kumar, “Prediction and Control of Cutting Tool Vibration Cnc Lathe with Anova and Ann,” International Journal of Lean Thinking, Vol. 2, No. 1, pp. 1-23, 2011.
[5]J. Herwan, S. Kano, R. Oleg, H. Sawada, and M. Watanabe, “Comparing Vibration Sensor Positions in CNC Turning for a Feasible Application in Smart Manufacturing System,” Int. J. of Automation Technology, vol. 12, No. 3, 2018.
[6]D. R. Salgado and F. J. Alonso, “An approach based on current and sound signals for in-process tool wear monitoring,” Int J Mach Tools Manuf, Vol. 47, No. 14, pp. 2140–2152, 2007.
[7]Z. R. Wang, J. Wang, and Y. R. Wang, “An intelligent diagnosis scheme based on generative adversarial learning deep neural networks and its application to planetary gearbox fault pattern recognition,” Neurocomputing, Vol. 310, pp. 213-222, 2018.
[8]P. Xu, R. Du, and Z. B. Zhang. “Pvredicting pipeline leakage in petrochemical system through GAN and LSTM,” Knowledge-Based Systems, Vol. 175, 2019.
[9]Fanny and T. W. Cenggoro, “Deep learning for imbalance data classification using class expert generative adversarial network,” Procedia Computer Science, Vol. 135, pp. 60–67, 2018.
[10]K. Zhang, G.Zhong, J. Dong, S. Wang, “Stock Market Prediction Based on Generative Adversarial Network,” IIKI, Vol. 147, 2019.
[11]J. Chen, Y. Xie, K. Wang, Z. H. Wang, G. Lahoti, Z. H. Wang, G. Lahoti, C. Zhang, M. A. Vannan, B. Wang, and Z. Qian, “Generative Invertible Networks (GIN): Pathophysiology-Interpretable Feature Mapping and Virtual Patient Generation,” Medical Image Computing and Computer Assisted Intervention- MICCAI, Vol. 14, pp. 537-545, 2018.
[12]L. Zou, J. Xia, X. Wang, and G. Hu, “Optimization of die profile for improving die life in the hot extrusion process,” Journal of Materials Processing Technology, Vol. 142, pp. 659-664, 2003.
[13]I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning (Adaptive Computation and Machine Learning series), MIT Press, 2016.
[14]生成對抗網路(GAN)，GeeksforGeeks， https://www.geeksforgeeks.org/generative-adversarial-network-gan/
[15]生成對抗網路(GAN)，SlideShare, https://www.slideshare.net/yenlung/gan-90396897
[16]R. C. Gonzalez, R. E. Woods, Digital Image Processing, 2nd 2002, ISBN 0-20-118075-8。
[17]M. Stéphane, A Wavelet Tour of Signal Processing, 3rd 2009, ISBN 978-0-12-374370-1。