網宇實體系統會在雲端建立預測用的數值模型，再利用由工廠自動收集的資料訓練該數
值模型，並且持續追隨實體的情況更新模型。為了避免異常資料混入正常資料中導致更新後
之數值模型資料汙染，進而造成的預測不準確。因此需要異常資料過濾器將資料進行過濾，
以維護網宇實體系統的性能。
本研究之過濾器由自動編碼器與分類器構成。先使用正常資料訓練前端的自動編碼器，
使其能夠初步辨識正常資料與異常資料。接著在後端建立用於分辨資料正確性之分類器，由
支援向量機、隨機森林、K-近鄰演算法與集成式學習構成。分類器的輸入資料包含自動編碼
器之輸入與輸出，也可以加入兩者差異的量化指標。最後，分類器會將資料區分成正確或錯
誤二類，只有正常資料可以用於更新網宇實體系統之數值模型，避免異常資料影響其性能。
本研究使用二個製程案例對研究方法進行驗證及調整；第一個案例是使用文獻上的雷射
金屬沉積製程優化案例，第二個案例是使用 3D 生物列印之製程優化案例。先評估資料汙染
程度對數值模型的影響，再利用本研究發展之方法過濾異常資料，並且分析分類的情況；其
中真陽性與偽陽性是影響資料汙染的關鍵指標。結果顯示本研究分類的正確率大於 94.7％，
錯誤率小於 3.3％，所以本研究提出之方法可確實避免網宇實體模型受到資料汙染的影響。

The cyber-physical system of the manufacturing industry will build a numerical model for
prediction in the cloud, and then will train the numerical model with the data collected automatically
from the factory, and continuously update the model. To avoid the inaccurate predictions caused by
the data poisoning of the updated numerical model due to the mixing of the wrong data with the
correct data, an error data filter is needed to classify the data and maintain the performance of the
cyber-physical system.
The filter in this study consists of Auto-Encoder and a classifier. First, use the correct data to
train the front-end Auto-Encoder so that it can initially identify correct and incorrect data. Then, the
classifier is built at the back-end to distinguish the correctness of the data, which consists of Support
Vector Machines, Random Forest, a K-Nearest Neighbor and Ensemble Learning. The input data of
the classifier includes the input and output of the Auto-Encoder, and can also include quantifiers of
the difference between them. Finally, the classifier classifies the data into correct or incorrect
categories. Only correct data can be used for updating the numerical model of the cyber-physical
system.
In this study, two process cases are used to validate and adapt the research method; the first case
is Laser Direct Metal Deposition, and the second case is Modeling and Parameter Optimization of 3D
Printing Process with Bio-material. The impact of data poisoning on the numerical model was
evaluated, and then the method developed in this study was used to filter out the wrong data and
analyze the classification; true positive and false positive were the key indicator of data poisoning.
The results show that the classification accuracy rate of this study is greater than 94.7%, and the error
rate is less than 3.3%. Therefore, the method proposed in this study can indeed avoid the influence of
data poisoning on the cyber-physical model.

1. Y. Liu, and X. Xu, Industry 4.0 and cloud manufacturing: A comparative analysis, Journal of
Manufacturing Science and Engineering, vol. 139, March, 2017.
2. F. Qiao, J. Liu, and Y. Ma, Industrial big-data-driven and CPS-based adaptive production
scheduling for smart manufacturing, International Journal of Production Research, 2020, doi:
10.1080/00207543.2020.18364.
3. J. Lee et al., A cyber-physical systems architecture for industry 4.0-based manufacturing
systems, Manufacturing Letters, vol. 3, pp. 18-23, January, 2015.
4. J. Lee et al., Service innovation and smart analytics for industry 4.0 and big data environment,
Procedia CIRP. vol. 16, pp. 3-8, 2014
5. Z. Li, A. Barenji, and G. Huang, Toward a blockchain cloud manufacturing system as a peer
to peer distributed network platform, Roboitcs and Computer-Integrated Manufacturing,
vol.54, pp. 133-144, December 2018.
6. L. Monostori et al., Cyber-physical systems in manufacturing, CIRP Annals, vol. 65, issue. 2,
pp. 621-641, 2016.
7. Y. Lu, Industry 4.0: A survey on technologies, applications and open research issues, Journal
of Industrial Information Integration, vol. 6, pp. 1-10, June 2017.
8. X. Xu, From cloud computing to cloud manufacturing, Robotics and Computer-Integrated
Manufacturing, vol. 28, issue. 1, pp. 75-86, February 2012.
9. A. Sethi, and S. Sethi, Flexibility in manufacturing: a survey, The International Journal of
Flexible Manufacturing Systems. 2, pp. 289-328, 1990.
10. D. Zissis, and D. Lekkas, Addressing cloud computing security issues, Future Generation
Computer Systems, vol. 28, issue. 3, pp. 583-592, March 2012.
11. M. Swan, Blockchain: Blueprint for a new economy, 2015.
12. J. Lee et al, A blockchain enabled Cyber-Physical System architecture for Industry 4.0
manufacturing systems, Manufacturing Letters, vol. 20, pp. 34-39, April 2019.
13. T. Chiba et al., A Countermeasure Method Using Poisonous Data Against Poisoning Attacks
on IoT Machine Learning, International Journal of Semantic Computuing, vol. 15, no. 2, pp.
215-240, 2021.
14. 賀碩, 不準確標註下的多標記學習算法研究, 碩士, 西南大學, 2020.
15. S. Goldwasser et al., Planting undetectable backdoors in machine learning models,
arXiv:2204.06974, vol. 1, April 2022.
16. J. Liu et al., Semi-supervised anomaly detection with dual prototypes autoencoder for
industrial surface inspection, Optics and Lasers in Engineering, vol. 136, January 2021.
17. S. Omar, A. Ngadi, H. Jebur, Machine Learning Techniques for Anomaly Detection: An
Overview, International Journal of Computer Applications, vol.79, pp. 33-41, October 2013.
18. M. Adkisson et al. Autoencoder-based Anomaly Detection in Smart Farming Ecosystem, presented at the 2021 IEEE International Conference on Big Data (Big Data), 2021
19. D. Liao et al. Anomaly detection for semiconductor tools using stacked autoencoder learning,
presented at the 2018 International Symposium on Semiconductor Manufacturing (ISSM),
2018.
20. M-A. Lutz et al., Evaluation of Anomaly Detection of an Autoencoder Based on Maintenace
Information and Scada-Data, Energies, val. 13, 2020.
21. 陳品元, 基於機器學習方法之半導體製造異常檢測研究, 碩士, 科技管理研究所, 國立陽
明交通大學, 新竹市, p. 64, 2021.
22. R. Zolfaghari, Diagnosis of diabetes in female population of pima indian heritage with
ensemble of bp neural network and svm. IJCEM International Journal of Computational
Engineering and Management, vol.15, issue. 4, pp. 2230-7893, July 2012.
23. B-A. Tama, F. Rodiyatul, and H. Hermansyah, An early detection method of type-2 diabetes
mellitus in public hospital, Telecommunication Computing Electronics and Control, vol.9,
no.2, pp. 287-294, 2021.
24. R. Ali et al. Prediction of diabetes mellitus based on boosting ensemble modeling, presented
at the International Conference on Ubiquitous Computing and Ambient Intelligence, pp. 25-
28, 2014.
25. E-S. Shaker et al., A comprehensive medical decision–support framework based on a
heterogeneous ensemble classifier for diabetes prediction, Electronics, vol. 8, pp. 635, 2019.
26. SAP. 什 麼 是 機 器 學 習 ？, https://www.sap.com/taiwan/insights/what-is-machinelearning.html.
27. 機器學習的前世今生：一段波瀾壯闊的歷史, https://read01.com/zh-tw/zD73kR.html,
2016/09/12.
28. D. Birla, Basics of Autoencoders, https://medium.com/@birla.deepak26/autoenco ders76bb49ae6a8f, 2019/03/12.
29. Y-R. Tsai, What are Autoencoders. 2019/3/9.
30. 10 程式中. 核模型–支持向量機(SVM), https://ithelp.ithome.com.tw/articles/10270447,
2021/09/23.
31. J. Wang, 隨機森林(Random Forest), https://reurl.cc/gMD96p, 2018/03/26.
32. Pylnvest. K-近鄰演算法 KNN. https://pyecontech.com/2020/04/19/knn/, 2020/04/19.
33. W. Zhiqiang, and L. Jun. A review of object detection based on convolutional neural network.
Presented at the 2017 36th Chinese control conference (CCC). 2017.
34. J. Hung, GAN 簡介, https://ithelp.ithome.com.tw/articles/10226112, 2019/10/09.
35. S. Wu, 3 Best metrics to evaluate Regression Model, https://reurl.cc/XVEoRM, 2020/05/23.
36. SarielWang. 機器學習模型評估指標–confusion matrix, precision, recall, and, F1-score,
https://reurl.cc/bE9MYr, 2020/05/15.
37. F. Caiazzo, and A. Caggiano, Laser direct metal deposition of 2024 Al alloy: trace geometry
prediction via machine learning, Materials, vol. 11, pp. 444, March 2018.
38. 林建沅, 生醫材料 3D 列印製程的建模與參數最佳化, 碩士, 機械工程學系, 國立中央大
學, 桃園市, 2022/06/30.
39. Casting Product Image Data For Quality Inspection, https://reurl.cc/GEA7p3.
40. RobTheMan, Defect Image, https://www.kaggle.com/code/robtheman/defect-image.
41. V. Chandola, A. Banerjee, and V. Kumar, Anomaly detection: A survey, ACM computing
surveys (CSUR), vol. 41, pp. 1-58, 2009.
42. V. Hodge, and J. Austin, A Survey of Outlier Detection Methodologies, Artificial Intelligence
Review, vol. 22, pp. 85-126, 2004.