- 學位論文
深度學習應用於線切割放電加工之變厚度智慧型控制
Application of Deep Learning to the Intelligent Control of Wire Electrical Discharge Machining of Variable Thickness Workpiece
摘要
線切割放電加工應用中，不可避免會遇到工件厚度改變之需求，本文目的在於建立放電加工於變厚度製程中，可兼顧加工效率、穩定性以及槽寬一致性的深度學習智慧型控制系統。過去線切割放電加工於變厚度的控制，多聚焦在加工效率與穩定性的提升，對於同時考量槽寬不一現象影響工件尺寸精度的問題之研究則是較為欠缺的，從切割截面上可以得知加工後真平度不佳，且將可能導致未來工件配合上契合度不良等問題。文中利用深度學習(Deep Learning)藉著把資料透過多個非線性隱藏層的轉換，自動萃取出足以代表資料特性的特徵，亦即從多樣且大量的放電訊號中找出與工件厚度高度相關的隱性特徵函數且具有良好的強健性，不易受到雜訊或是少部份極值的影響，因此可準確得知工件厚度。 第二階段為加工參數的設定，在過去多依賴經驗豐富的操作人員，因此再次利用深度學習的技術，使輸入資料的隱性特徵經過歸一化處理後，讓模型輸出猶如豐富經驗操作員的判定。如此的智慧型控制策略不只可以減少操作人員技術門檻，各厚度槽寬均一變動量小的目標，可以使第二刀精修時的材料移除率更為穩定，減少線痕的產生。 由實驗結果顯示，本文所建立的深度學習工件厚度線上估測系統，可以準確估測出工件厚度，且加工參數智慧型控制系統，其可根據不同的工件厚度調整放電休止時間與輔助放電休止時間，使放電頻率被控制在期望值之內，在工件厚度由薄變厚時，可有效提升加工效率；在工件厚度由厚變薄時，可以改善進給速度過快與放電集中所導致斷線的現象。另一方面，本研究之策略有效兼顧了槽寬一致性，亦即加工不同工件厚度時，其各階級厚度之槽寬平均值與變動量可有效維持在穩定的數值內。
並列摘要
The purpose of this study is to establish a deep learning intelligent control system that can combine processing efficiency, stability and kerf consistency in variable thickness processing. In the past, the wire electrical discharge machining Control for variable thickness workpiece, and focus on improving processing efficiency and stability. The research on the problem that the kerf difference affects the accuracy of the variable thickness workpiece is relatively lacking. The true straightness is not good after processing, and it may lead to error fitting of future workpieces. Deep Learning is used to automatically extract features that are representative of data characteristics by transforming data through multiple nonlinear hidden layers, that is, from a variety of discharge signals that are highly correlated with the thickness of the workpiece. The recessive feature function has strong robustness and is not sensitive to noise or some of extreme values, so the thickness of the workpiece can be accurately known. The second stage is the setting of processing parameters. In the past, it relied on experienced operators. Therefore, the deep learning technique was used again to normalize the hidden features of the input data, so that the model output is like a experience operator to determination what the parameters should be setting. Intelligent control strategy not only reduce the operator's technical threshold, but also achieve a goal of uniform material removal rate during the trim cutting and reduce the occurrence of line marks. The experimental results show that the deep learning workpiece thickness online estimation system established in this study which can accurately estimate the thickness of the workpiece, and the processing parameter intelligent control system can adjust the discharge TOFF and TAOFF according to different workpiece thicknesses. The discharge frequency is controlled within the desired value, and the processing efficiency can be effectively improved when the thickness of the workpiece is thin to thick; when the thickness of the workpiece is thick to thin, the feed rate will be accelerated and the wire is broken by discharge concentration phenomenon. On the other hand, the strategy of this study effectively balances the kerf uniformity, that is, the average value and variation of the kerf of each class thickness can be effectively maintained within a stable value when processing different workpiece thicknesses.