臺灣博碩士論文加值系統

當扣件品質不良的原因為沖模問題時，機台操作人員可透過調整相關的螺栓與螺帽以消弭品質變異的情況。而由於進行機台的調整，將造成的沖模訊號的變動，在因缺乏調整紀錄且已建立的品質估測模型無法套用下，常需花費更多時間以更新模型。因此，如何縮短模型更新時間以及時估測生產品質，實為扣件鍛造產業邁向智慧製造的挑戰。
本研究開發一扣件製程線上品質估測系統，可遠端分析扣件製程中的品質，並建立快速分析模型，應用於製程中各個道次的檢測。在特徵萃取上，透過自動編碼器識別訊號以萃取關鍵特徵。在評估樣本差異上，以KL散度與獨立樣本t檢定分析訊號的差異性，並決定是否需更換模型。在減少建模所需時間上，使用遷移式學習的方式凍結歷史模型的權重，降低建模所需的資料量。在偵知品質變異上，將特徵以圖形顯示機台穩定度，可有效掌握目前與建模時期的差異。
研究結果顯示，本系統利用KL散度與獨立樣本t檢定可評估樣本之間的相似程度。當需更新模型時，使用遷移式學習，可部分引用原模型，以循序漸進方式持續更新模型，使診斷能力更穩定。而在失效模式分類上，利用增強式隨機森林方法可診斷多種失效模式的正確率達99%；而品質的估測圖形，可使現場操作人員更容易了解生產過程中是否開始出現變異的情況。

A machine operator can adjust relevant bolts and nuts to eliminate the quality variation causing from mold issues. Thus, the process signal will be changed by the machine adjustment. Lacking the available records of machine adjustment and model quality estimation, it often takes more time to update the model. Therefore, to move toward smart manufacturing, the fastener forging industry is challenging how to shorten the time to model fresh for estimating production quality in a timely manner.
This research develops an online quality estimation system for fastener manufacturing, which can remotely analyze the quality of the fastener manufacturing process and establish a rapid analysis model to be applied to the inspection of each pass in the manufacturing process. In feature extraction, the key features are extracted from signals by an autoencoder and. In assessing sample differences, the methods including KL divergence and t test are used to analyze the differences in signals and to decide whether to refresh the model. In modeling time, the transfer learning is applied to freeze the weights of historical models for reducing the sample for modeling. In quality estimation, a forming stability diagram is presented to indicate the difference between the current and the modeling periods.
The research results show that the proposed system can assess the degree of similarity between samples by using KL divergence and t test. Based on transfer learning, the quality evaluation model can be continuously and gradually updated by adopting partial model to improve diagnosis stability. Also, the diagnosis accuracy rate of various failure modes is nearly 99% after using the enhanced random forest method. Finally, the forming stability diagram can assist users to figure out the variation of a forming process.

摘要 i
Abstract ii
誌謝 iii
目錄 iv
表目錄 vi
圖目錄 vii
第1章 緒論 1
1.1 研究背景 1
1.2 研究目的 5
1.3 研究架構 6
第2章 理論方法 7
2.1 類神經網路模型 7
2.2 萃取特徵方法 10
2.3 樣本檢測比對方法 11
2.3.1 KL散度 11
2.3.2 獨立樣本t檢定 12
2.4 遷移式學習方式 13
2.5 推估成形穩定度方法 15
2.6 失效分類方法 16
2.6.1 隨機森林(Random Forest) 16
2.6.2 增強式隨機森林(XGBoost-RF) 17
第3章 系統方法 19
3.1 系統架構 19
3.1.1 系統概念 19
3.1.2 系統流程 21
3.2 壓感訊號之收集 23
3.3 萃取特徵 24
3.4 更新模型 25
3.5 分類模型建立 27
3.5.1 建立隨機森林模型 27
3.5.2 建立增強式隨機森林模型 28
3.6 估測品質 29
3.7 顯示系統介面 30
第4章 鍛造製程案例分析 32
4.1 壓感資料萃取特徵 32
4.1.1 機台規格與資料收集 32
4.1.2 原始壓感訊號標準化 34
4.1.3 壓感萃取特徵 36
4.2 評估樣本差異 39
4.3 遷移式學習案例分析 43
4.3.1 遷移學習案例1 45
4.3.2 遷移學習案例2 48
4.3.3 模型更新方法 51
4.4 失效模式診斷 55
4.4.1 失效模型建立 56
4.4.2 失效模型套用 59
4.5 品質估測 62
第5章 結論與未來展望 65
5.1 結論 65
5.2 未來展望 66
參考文獻 67

[1]C. Hu, T. Ding, H. Ou, and Z. Zhao, “Effect of tooling surface on friction conditions in cold forging of an aluminum alloy,” Tribology International, Vol. 131, pp. 353-362, 2019.
[2]L. Wang, J. Zhou, J. Duszczyk, and L. Katgerman, “ Friction in aluminium extrusion—Part 1 A review of friction testing techniques for aluminium extrusion,” Tribology International, Vol. 56, pp. 89-98, 2012.
[3]B. Li, Y. Zheng, C. Liu, Q. Li, Z. Zheng, and X. Chen, “ Torsional thermomechanical fatigue behavior of 316LN stainless steel,” Materials Science & Engineering A, Vol. 789, 2020.
[4]M. N. James, D. G. Hattingh, D. Asquith, M. Newby, and P. Doubll, “ Applications of Residual Stress in Combatting Fatigue and Fracture,” Procedia Structural Integrity, Vol. 2, pp. 11-25, 2016.
[5]P. Groche, C. MüllerEmail, and A. Jahn, “ Effects of the Tool Lubrication in Cold Forging”, Tribology Letters, Vol. 53, pp. 599-605, 2014.
[6]R. Matsumoto, N. Kai, Y. Tomita, A. Kajioka, S. Mori, and H. Utsunomiya, “ Characterization of surface profile of shot peened cemented tungsten carbide dies with micro valleys and their lubrication performance in cold forging,” Procedia Engineering, Vol. 207, pp. 1135-1140, 2017.
[7]S. Chander, and V. Chawla, “ Failure of Hot Forging Dies –An Updated Perspective,” Materials Today: Proceedings, Vol 4, pp. 1147-1157, 2017.
[8]王威傑，“開發扣件成形製程之失效預診系統”，國立高雄科技大學電機工程研究所碩士論文，2019。
[9]F. Zhuang1, X. Cheng, P. Luo, S. Jialin Pan, and Q. He, “ Supervised Representation Learning: Transfer Learning with Deep Autoencoders,” Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence, 2015.
[10]C. Fan, Y. Sun, F. Xiao, J. Ma, D. Lee, J. Wang, and Y. C. Tseng, “ Statistical investigations of transfer learning-based methodology for short-term building energy predictions,” Applied Energy, Vol. 262, Article 114499, 2020.
[11]M. Talo, “Automated classification of histopathology images using transfer learning,” Artificial Intelligence In Medicine, Vol. 101, Article 101743, 2019.
[12]N. Das, E. Hussain, and L. B. Mahanta, “ Automated classification of cells into multiple classes in epithelial tissue of oral squamous cell carcinoma using transfer learning and convolutional neural network,” Neural Networks, Vol. 128, pp. 47-60, 2020.
[13]S. Raghu, N. Sriraam, Y. Temel, S. V. Rao, and P. L. Kubben, “ EEG based multi-class seizure type classification using convolutional neural network and transfer learning,” Neural Networks, Vol. 124, pp. 202-212, 2020.
[14]A. Abdalla, H. Cen, L. Wan, R. Rashid, H. Weng, W. Zhoue, and Y. He, “ Fine-tuning convolutional neural network with transfer learning for semantic segmentation of ground-level oilseed rape images in a field with high weed pressure,” Computers and Electronics in Agriculture, Vol. 167, Article 105091, 2019.
[15]Y. Xiao, H. Wang, and B. Liu, “ A New Transfer Learning-based Method for Label Proportions Problem,” Information Sciences, Vol. 541, pp. 391-408, 2020.
[16]I. H. Fong, T. Li, S. Fong, R. K. Wong, and A. J. Tallón-Ballesteros, “ Predicting concentration levels of air pollutants by transfer learning and recurrent neural network,” Knowledge-Based Systems, Vol. 192, Article 105622, 2020.
[17]T. Wen, S. Duan, and W. Jiang, “ Node similarity measuring in complex networks with relative entropy,” Communications in Nonlinear Science and Numerical Simulation, Vol. 78, Article 104867, 2019.
[18]R. Xiao, J. Su, J. jiang, X Lin, and L. Lin, “ SFAD Toward effective anomaly detection based on session feature similarity,” Knowledge-Based Systems, Vol. 165, pp. 149-156, 2019.
[19]謝慶嶧，“智慧製造分析工具箱之開發”，國立高雄科技大學電機工程研究所碩士論文，2019
[20]https://zh.wikipedia.org/wiki/%E4%BA%BA%E5%B7%A5%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C，類神經網路，維基百科。
[21]https://zh.wikipedia.org/wiki/%E6%BF%80%E6%B4%BB%E5%87%BD%E6%95%B0，激活函數，維基百科。
[22]https://www.itread01.com/content/1545664696.html，損失函數，itread01。
[23]https://keras.io/api/layers/initializers/ ，初始化方法，keras文檔
[24]https://slidesplayer.com/slide/11334830/，自動編碼器，SlidesPlayer。
[25]https://www.zhihu.com/question/41979241/answer/219137941，遷移式學習，知乎。
[26]https://bigdatafinance.tw/index.php/tech/methodology/988-transfer-learning，深度學習模型- 遷移式學習(Transfer Learning)概述，bigdatafinance。
[27]http://www.taroballz.com/2018/07/14/ML_decision_tree/，決策數，taroballz。
[28]https://zhuanlan.zhihu.com/p/92837676，XGBoost詳細推導，知乎。
[29]https://www.ccm3s.com/%E6%A8%99%E6%BA%96%E5%8A%A0%E9%95%B7%E5%9E%8B%E8%9E%BA%E7%B5%B2%E6%88%90%E5%9E%8B%E6%A9%9F.html，精湛科技股份有限公司。