基於深度學習與超參數優化於磨削Ti-6Al-4V合金表面粗糙度智能穩健辨識方法之研究

本論文主要建構多種卷積神經網路模型進行磨削Ti-6Al-4V後的表面粗糙度辨識，以取代原有接觸量測方式提升製程效益。首先透過穩健設計進行磨削加工來獲得到各種Ti-6Al-4V合金不同的表面粗糙度樣本，同時量測磨削力、磨削溫度與表面粗糙度並進行製程單一目標與多目標優化，在控制不同CBN砂輪粒度與不同磨削參數下進行磨削，優化結果與基準組比較，在#150的砂輪在磨削力比降低37.68%、磨削溫度降低14.27%以及表面粗糙度降低0.46%，在#200的砂輪在磨削力比降低39.09%、磨削溫度降低11.30%以及表面粗糙度降低30.58%。多目標優化與單一目標最佳化相較，因須兼顧多個品質特性在#150砂輪磨削力比損失46.51%、磨削溫度損失1.52%以及在表面粗糙度損失7.92%品質特性。在#200砂輪磨削力比損失3.88%、磨削溫度損失1.88%以及在表面粗糙度降損失22.14%品質特性。在將磨削後的Ti-6Al-4V合金表面粗糙度進行影像拍攝，且依照國際化標準組織(ISO)所規定的表面粗糙度等級N5、N6以及N7進行影像的分類、表面破損特徵分級與標記。將分類好的影像輸入到AlexNet、VGG-19、GoogleNet以及ResNet-50卷積神經網路模型，同時使用穩健設計尋找出各模型單一最佳化超參數設置進行準確度分析，優化結果與基準組準確度比較下，AlexNet準確度為94.36%提升了3.68%、VGG-19準確度為97.10提升了8.33%、GoogleNet準確度為94.66%在準確度提升了2.54%以及ResNet-50準確度為96.34%提升了4.11%，可以穩健快速提高模型的性能。本論文也將找出能夠兼顧多種卷積神經網路模型的超參數設置，供產業界實際運用此技術時可以快速搜索到相容多種模型超參數設置，使用模糊推論系統來求得多重品質特性(MPCI)，在求得多重品質特性最佳化超參數設置後，各模型平均準確度與單一最佳化的平均準確度相比，AlexNet準確度為96.54%提升2.20%、GoogleNet準確度為97.23%提升0.26%、VGG-19準確度為98.30%降低0.30%以及ResNet-50準確度為98.08%降低0.06%。研究結果顯示不論是在單一目標與兼顧多種模型最佳化均有極高的辨識準確率，此方法可實際運用於產業界磨削Ti-6Al-4V過程中磨削粗糙度品質量測。可以進一步促進基於使用卷積神經網路量測的技術從學術界向工業界的轉移。

關鍵字

Ti-6Al-4V ；磨削；穩健製程設計；表面粗糙度；模糊推論系統；多重品質特性衡量指標；卷積神經網路；超參數

並列摘要

This thesis mainly constructs a variety of convolutional neural network models to identify the surface roughness after grinding Ti-6Al-4V to replace the original contact measurement method to improve the process efficiency. Firstly, various samples of Ti-6Al-4V alloys with different surface quality and surface roughness are obtained by grinding robust process design. Measured the grinding force, grinding temperature , surface roughness, and single-objective and multi-objective optimization of the process to action. Control of different CBN grinding wheel particle sizes and different grinding parameters. Compared with the benchmark group, the optimization results show that the grinding force ratio of the #150 grinding wheel is reduced by 37.68%, the grinding temperature is reduced by 14.27%, and the surface roughness is reduced by 0.46% %, the grinding force ratio of the #200 wheel is reduced by 39.09%, the grinding temperature is reduced by 11.30%, and the surface roughness is reduced by 30.58%. Compared with single-objective optimization, multi-objective optimization must consider multiple quality characteristics. The grinding force ratio of the #150 grinding wheel loses 46.51%, the grinding temperature loss is 1.52%, and the surface roughness loses 7.92% of the quality characteristics. In the #200 wheel, the grinding force ratio lost 3.88%, the grinding temperature lost 1.88%, and the surface roughness lost 22.14% of the quality characteristics. After the grinding of Ti-6Al-4V alloy to take photographed, and the images were classified according to the surface roughness grades N5, N6 and N7 specified by the International Organization for Standardization (ISO), and the surface damage characteristics were classified and note. Input the classified photos into AlexNet, VGG-19, GoogleNet and ResNet-50 convolutional neural network models, and use robust process design to find a single optimal hyperparameter setting for each model for accuracy analysis, and optimize the results and benchmark groups. In terms of accuracy, the accuracy of AlexNet is 94.36% that improve by 3.68%, the accuracy of VGG-19 is 97.10% that improve by 8.33%, the accuracy of GoogleNet is 95.66% that improve by 2.54%, and the accuracy of ResNet-50 is 96.34 that improve by 4.11%. This thesis finds out the hyperparameter settings that can take into account a variety of convolutional neural network models, so that when the industry actually uses this technology, it can quickly search for the hyperparameter settings that are compatible with multiple models, and use the fuzzy inference system to find the multiple quality characteristics (MPCI). After obtaining the hyperparameter settings for the optimization of multiple quality characteristics, the average accuracy of AlexNet is 96.54% that improve by 2.20%, the average accuracy of GoogleNet is 97.23% that improve by 0.26%, the average accuracy of VGG-19 is 98.30% that reduce by 0.30% and the average accuracy of ResNet-50 is 98.08% that reduce by 0.06%. The research results show that the identification accuracy is extremely high in both single target and multiple model optimization. This method can be practically applied to the quality measurement of grinding surface roughness in the process of grinding Ti-6Al-4V. The transfer of techniques based on measurements using convolutional neural networks from academia to industry can be further facilitated.

並列關鍵字

Ti-6Al-4V ； Grinding ； Robust Process Design ； Fuzzy Theory ； Surface Roughness ； Multiple Performance Characteristic Index(MPCI) ； Convolutional Neural Network ； Hyperparameters

參考文獻

[1] D. Handa, S. Kumar, S. Babu Thekkoot Surendran, and V. S. Sooraj, 2021, "Simulation of intermittent grinding for Ti-6Al-4V with segmented wheel," Materials Today: Proceedings, vol. 44, pp. 2537-2542.

Google Scholar

[2] C. Dai, W. Ding, Y. Zhu, J. Xu, and H. Yu, 2018, "Grinding temperature and power consumption in high speed grinding of Inconel 718 nickel-based superalloy with a vitrified CBN wheel," Precision Engineering, vol. 52, pp. 192-200.

Google Scholar

[3] S. V. Nosenko, V. A. Nosenko, and L. L. Kremenetskii, 2014, "Influence of dressing of the wheel on the surface quality of titanium alloy in deep grinding," Russian Engineering Research, vol. 34, no. 10, pp. 632-636.

Google Scholar

[4] D. N. Naik, N. T. Mathew, and L. Vijayaraghavan, 2019, "Wear of Electroplated Super Abrasive CBN Wheel during Grinding of Inconel 718 Super Alloy," Journal of Manufacturing Processes, vol. 43, pp. 1-8.

Google Scholar

[5] C. D. Benkai Li, Wenfeng Ding,Changyong Yang,Changhe Li,Olga Kulik, 2021, "Prediction on grinding force during grinding powder metallurgy nickel-based superalloy FGH96 with electroplated CBN abrasive wheel," Chinese Journal of Aeronautics, vol. 34, no. 8, pp. 65-74.

Google Scholar

主題瀏覽