淡江大學覺生紀念圖書館 (TKU Library)
進階搜尋


下載電子全文限經由淡江IP使用) 
系統識別號 U0002-0909202014315600
中文論文名稱 基於遷移式學習之智慧製造數據分析與品質預測
英文論文名稱 Intelligent Manufacturing Data Analysis and Quality Prediction based on Transfer Learning
校院名稱 淡江大學
系所名稱(中) 資訊工程學系碩士班
系所名稱(英) Department of Computer Science and Information Engineering
學年度 108
學期 2
出版年 109
研究生中文姓名 陳昱名
研究生英文姓名 Yu-Ming Chen
學號 607410452
學位類別 碩士
語文別 中文
第二語文別 英文
口試日期 2020-06-12
論文頁數 54頁
口試委員 指導教授-張志勇
共同指導教授-郭經華
委員-游國忠
委員-廖文華
中文關鍵字 人工智慧  遷移式學習  數據不平衡  數據分析  長短期記憶模型 
英文關鍵字 Artificial intelligence  Transfer learning  Data imbalance  Data analysis  Long-term and short-term memory models(LSTM) 
學科別分類 學科別應用科學資訊工程
中文摘要 近年來,製造業一般而言均有能力透過自動化的機器進行大量生產,而產品的品質會直接影響到公司營業額,但在生產的過程中會遭遇一困難點,工廠無法提前預知產品的品質好壞,只能等待製造完成後經由檢測才能得知是否有瑕疵。因此,良率問題會是製造業者的致命傷,不僅增加不必要的成本,更是企業信譽的風險因子。
而今工業4.0的發展日漸蓬勃,大部分的工廠機具都加裝感測器使得在生產過程中,也能透過感測器回傳的各種數據即時監控內部狀況。但只有專業老師傅能透過數據找出與品質之間的相關聯,而新進的員工則需要花費大量時間才能學會老師傅的經驗,工廠更需要針對營運成本做出精準的評估。
因此,本論文基於人工智慧之技術,擬發展數據分析與品質預測系統的設計與實作,對於染整業的染程數據進行分析並對其找出與品質之間的關聯性,最後透過染程數據提前進行品質預測,以協助操作機器的人員,判斷當下該如何進行調整,避免整批材料報廢,以達到節省時間、降低染整廠的生產成本及增加良率的目標。
本論文所提出的「基於遷移式學習之智慧製造數據分析與品質預測」大致可分為二大類別,一為收集整理並分析數據,二為基於遷移式學習的採樣應用。其一,染程數據通常為時間序列資料,且每一筆資料的序列長度皆不等長,所以我們需要對數據先行格式整理,並對其進行分析與品質之間關聯;其二,普遍工廠中的數據大部分為正常品,只有少部分為瑕疵品,數據不平衡會影響所訓練出來的模型表現不佳。雖然,新、舊工廠都存有數據不平衡等問題,但舊工廠勝在資料龐大。因此,我們會先透過採樣法中的過採樣技術,產生瑕疵品數據,以平衡資料間數量。其後會再以欠採樣去除類別之間過於相似的數據,而改善數據不平衡所帶來的問題。但新工廠的數據相對於舊工廠來說還是較為少量,所以透過遷移式學習將舊工廠的知識遷移至新工廠的LSTM品質預測模型上,進一步改良新工廠預測模型的準確率以及召回率。
染整工廠使用本論文所設計之系統,可以提前預測產品品質,透過該品質回饋可使現場工作人員提前對於產品進行補救,以節省生產時間、降低瑕疵品與材料報廢所帶來的成本考量。
英文摘要 In recent years, the manufacturing industry is generally capable of mass production through automated machines, and the quality of products will directly affect the company’s turnover, but there will be a difficulty in the production process. The factory cannot predict the quality of the product in advance. Good or bad, you can only know whether there are defects after the manufacturing is completed and tested. Therefore, the yield problem will be a fatal injury to manufacturers, not only increasing unnecessary costs, but also a risk factor for corporate reputation.
Now that the development of Industry 4.0 is booming, most of the factory machines are equipped with sensors so that the internal conditions can be monitored in real time through various data returned by the sensors during the production process. However, only professional masters can find the correlation with quality through data, and new employees need to spend a lot of time to learn the experience of masters, and factories need to make accurate assessments of operating costs.
Therefore, based on the technology of artificial intelligence, this paper intends to develop the design and implementation of a data analysis and quality prediction system. It analyzes the dyeing process data of the dyeing and finishing industry and finds the relationship between it and quality. Finally, through dyeing The process data predicts the quality in advance to assist the operators of the machine to determine how to make adjustments at the moment to avoid scrapping the entire batch of materials to achieve the goal of saving time, reducing the production cost of the dyeing and finishing plant and increasing the yield.
The "Intelligent Manufacturing Data Analysis and Quality Prediction Based on Transfer Learning" proposed in this paper can be roughly divided into two categories, one is collecting and analyzing data, and the other is sampling applications based on transfer learning. First, the dyeing process data is usually time series data, and the sequence length of each data is not the same length, so we need to organize the data in advance, and analyze it to correlate with quality; second, in general factories Most of the data are normal products, and only a small part are defective products. Unbalanced data will affect the poor performance of the trained model. Although both the new and old factories have problems such as data imbalance, the old factory is better than the huge amount of data. Therefore, we will first generate defective product data through the oversampling technique in the sampling method to balance the number of data. Later, under-sampling will be used to remove data that is too similar between categories to improve the problems caused by data imbalance. However, the data of the new factory is still relatively small compared to the old factory, so the knowledge of the old factory is transferred to the new factory's LSTM quality prediction model through migration learning to further improve the accuracy and recall rate of the new factory's prediction model.
Dyeing and finishing factories can predict product quality in advance by using the system designed in this paper. Through this quality feedback, on-site staff can remedy the products in advance to save production time and reduce cost considerations caused by defective products and material scrap.
論文目次 目錄 VI
圖目錄 VII
表目錄 VIII
第一章、簡介 1
第二章、相關研究 5
第三章、背景知識 8
3.1 LSTM技術 8
3.2 遷移式學習技術 9
3.3 採樣技術 11
第四章、系統架構 13
4.1 環境與問題描述 13
4.2 系統架構 15
第五章、系統實作 29
第六章、實驗分析 33
第七章、結論 37
參考文獻 38
附錄-英文論文 39
圖 1:一般遞歸神經網路與長短期記憶神經網路比較圖 9
圖 2:遷移式學習概念圖 11
圖 3:染整系統流程圖 15
圖 4:系統架構圖 16
圖 5:預訓練階段架構圖 17
圖 6:影響染布品質的遠近因素示意圖 18
圖 7:每一工單數據序列長度不等長示意圖 19
圖 8:LSTM模型架構示意圖 20
圖 9:牢度品質預測階段架構圖 21
圖 10:遷移式學習模型微調示意圖 22
圖 11:基於遷移式學習採樣架構示意圖一 23
圖 12:SMOTE產生不合格品質圖 24
圖 13:LDA欠採樣方式對數據進篩選並剔除圖 25
圖 14:基於遷移式學習採樣架構示意圖二 26
圖 15:基於遷移式學習採樣架構示意圖三 27
圖 16:使用階段架構圖 28
圖 17:舊工廠染程數據示意圖 29
圖 18:舊工廠染程數據特徵提取圖 30
圖 19:透過LDA映射染程數據圖 31
圖 20:遷移式學習模型微調圖 32
圖 21:使用流程圖 32
圖 22:Precision和Recall效能比較圖 35
圖 23:Accuracy效能比較圖 36
表 1:相關研究功能比較表 7
表 2:混淆矩陣表 34
參考文獻 [1] Maximilian Christ, Andreas W. Kempa-Liehr and Michael Feindt, "Distributed and parallel time series feature extraction for industrial big data applications", arXiv preprint arXiv:1610.07717.
[2] B. D. Fulcher and N. S. Jones, "Highly Comparative Feature-Based Time-Series Classification," in IEEE Transactions on Knowledge and Data Engineering, vol. 26, no. 12, pp. 3026-3037, 1 Dec. 2014.
[3] Y. Wang, S. Zhu and C. Li, "Research on Multistep Time Series Prediction Based on LSTM," 2019 3rd International Conference on Electronic Information Technology and Computer Engineering (EITCE), Xiamen, China, 2019, pp. 1155-1159.
[4] Y. Liu, Z. Su, H. Li and Y. Zhang, "An LSTM based classification method for time series trend forecasting," 2019 14th IEEE Conference on Industrial Electronics and Applications (ICIEA), Xi'an, China, 2019, pp. 402-406.
[5] C. Liu, X. Wang, K. Wu, J. Tan, F. Li and W. Liu, "Oversampling for Imbalanced Time Series Classification Based on Generative Adversarial Networks," 2018 IEEE 4th International Conference on Computer and Communications (ICCC), Chengdu, China, 2018, pp. 1104-1108.
[6] M. Y. Arafat, S. Hoque, S. Xu and D. M. Farid, "An Under-Sampling Method with Support Vectors in Multi-class Imbalanced Data Classification," 2019 13th International Conference on Software, Knowledge, Information Management and Applications (SKIMA), Island of Ulkulhas, Maldives, 2019, pp. 1-6.
[7] S. J. Pan and Q. Yang, "A Survey on Transfer Learning," in IEEE Transactions on Knowledge and Data Engineering, Vol. 22, No. 10, pp. 1345-1359, Oct. 2010.
[8] N. V. Chawla, K. W. Bowyer, L. O. Hall and W. P. Kegelmeyer, "SMOTE: Synthetic minority over-sampling technique", J. Artif. Intell. Res., vol. 16, no. 1, pp. 321-357, 2002.
論文使用權限
  • 同意紙本無償授權給館內讀者為學術之目的重製使用,於2020-09-10公開。
  • 同意授權瀏覽/列印電子全文服務,於2020-09-10起公開。


  • 若您有任何疑問,請與我們聯絡!
    圖書館: 請來電 (02)2621-5656 轉 2487 或 來信