自動化研磨系統之視覺檢測方法與研磨接觸力估測模型

A Visual Inspection Method and a Grinding Contact Force Estimation Model of Automatic Grinding Systems

王右勛(Yu-Hsun Wang)

指導教授 ： 林沛群

國立臺灣大學/工學院/機械工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202100271

摘要

砂帶研磨為經常使用於消除前道加工程序造成之瑕疵與毛邊的精加工製程。砂帶研磨加工機具接觸表面的柔軟、貼合特性，使得砂帶研磨更適合用於多變的曲面工件，如：渦輪葉片與水龍頭。既有的砂帶研磨研究聚焦在使用接觸輪的砂帶研磨上，而不使用接觸輪的砂帶研磨類型(文中簡稱為「自由狀態砂帶研磨」)卻可能因涉及砂帶變形而很少被提及。另一方面，由於研磨加工環境多粉塵、具高分貝噪音，加上整體製程勞力密集、產線加工程序繁複的性質，為減低對人體的傷害與降低人力成本，機器人研磨加工已是取代人工研磨的主要趨勢，如何精細化機械手臂加工至為關鍵。因此本研究主旨為增強機械手臂自動化研磨的能力、減少目前業界中機械手臂研磨加工對人工教點、微調與檢測的依賴。從兩個方面著手，其一是增加研磨系統之視覺檢測功能；其二則是加強研磨系統對於自由狀態砂帶研磨接觸力的預測功能。
本研究中的視覺檢測部分著重於金屬經研磨加工後的局部表面紋理檢測。詳述如何建立「不同砂帶目數」、「不同表面粗糙度」與「不同的砂帶磨耗程度」三個研磨後的紋理影像資料集。使用遷移式學習，利用已預先訓練過的卷積類神經網路模型，訓練三個不同的卷積類神經網路模型，分別進行「判別局部表面影像對應研磨砂帶號數」、「由局部表面影像估測表面粗糙度」以及「判別局部表面影像對應砂帶磨耗程度」三個實驗。結果證實運用遷移式學習可使卷積類神經網路模型快速學習研究中自建資料集的分類與迴歸任務，並亦能分辨出不同類別研磨表面的細微紋路。
本研究亦針對自由狀態砂帶研磨提出一個新的三維模型，以估測自由狀態的砂帶與工件之間的接觸力。此三維模型是以二維幾何模型為基礎疊加得出的結果，而後者的估測力是由砂帶的張力以及工件與砂帶的接觸狀況計算。此估測模型最後整合成一個工研院研發之機械手臂產線模擬器Ezsim的外掛功能。研究中利用不同外型與尺寸的試棒對此功能進行實機測試，結果顯示此模型能夠成功估測研磨正向力。因此本模型可在一些較為簡單的自由狀態砂帶研磨加工中替代昂貴的力規設備，提升產線上調整加工軌跡的效率。

關鍵字

卷積類神經網路 ； 遷移式學習 ； 紋理檢測 ； 砂帶研磨 ； 正向力 ； 機器手臂 ； 模型

並列摘要

Belt grinding is a commonly used finishing process that can reduce defects and burrs created by previous machining procedures. Due to the flexible characteristic of the contact surface of the manufacturing machines, belt grinding is more suitable for machining workpieces with complicated surfaces, such as turbine blades and faucets. Although there have been several studies discussing belt grinding, most of them focused on belt grinding with a contact wheel, while manufacturing processes that apply belt grinding with a free strand of the abrasive belt (also named as “free-form belt grinding” in this study) were rarely explored. In the meantime, due to the dusty and high-decibel noise environment of the grinding process, diverse processing procedures on the production line, and the labor-intensive nature of the whole manufacturing process, robotic grinding has emerged as the main method to replace manual grinding for reducing the potential harm to the human body while lowering labor costs. The key point is to heighten the sophistication of robotic arm processing. Therefore, the subject of this research is managing to enhance the manipulators’ ability of automatic grinding, which can be implemented from two aspects. The first one is adding visual inspection features to the grinding system. The second one is enhancing the ability of predicting the contact force of free-form belt grinding to the grinding system.
The visual inspection section of this research focuses on the local metal surface textures that appear after the grinding process. It details the processes of establishing two image datasets, “Abrasive belts of different mesh numbers” and “Abrasive belts of different degrees of wear.” By utilizing transfer learning, which is based on other pre-trained convolutional neural networks (CNNs) to train the CNN models with three different structures. Then, I used three CNN models to conduct three experiments respectively, “Classifying the mesh number of abrasive belts corresponding to the local surface images”, “Estimating the surface roughness of the local surface images”, and “Classifying the degree of wear of abrasive belts corresponding to the local surface images.” The results show that transfer learning enables the convolutional neural network models to quickly adopt the classification and regression tasks of self-constructed datasets in the research, and confirms that the convolutional neural networks are able to distinguish the fine textures on different grinding surfaces.
This research also proposes a new 3-dimensional (3D) model that estimates the normal force being generated between workpieces and the abrasive belts in free-form belt grinding. The 3D model was constructed using integrated 2D interaction forces between the workpieces and the abrasive belt, and the latter force derived based on the tension force of the abrasive belt and the geometric contact configuration. This estimation model is then integrated into a plug-in function of ITRI-developed robotic arm production line simulator EzSim. The model was experimentally evaluated using spheres, cylinders, and frustums grinding specimens, and the results confirm that the model can successfully predict normal forces. Therefore, this model can replace expensive force sensors in some simple free-form belt grinding processes and improve the efficiency of adjusting the machining trajectory on the production line.

並列關鍵字

convolutional neural networks ； transfer learning ； texture inspection ； belt grinding ； robot ； normal force ； model

參考文獻

[1] S. Yang and W. Li, Surface Finishing Theory and New Technology. Springer-Verlag Berlin Heidelberg, 2018, pp. 37 - 42.

[2] W. B. Rowe, Principles of Modern Grinding Technology. Boston: William Andrew Publishing, 2009, pp. 1-14.

[3] W. König, H. K. Tönshoff, J. Fromlowitz, and P. Dennis, "Belt Grinding," CIRP Annals, vol. 35, no. 2, pp. 487-494, 1986, doi: https://doi.org/10.1016/S0007-8506(07)60197-8.

[4] G. Hammann, Modellierung des Abtragsverhaltens elastischer, robotergeführter Schleifwerkzeuge. Springer-Verlag Berlin Heidelberg, 1998.

[5] L. Ri-xian, "Defects Detection Based on Deep Learning and Transfer Learning," Metallurgical and Mining Industry, no. 7, pp. 312 - 321, 2015.