表面粗糙度是端銑加工中重要品質指標之一,一般而言,手工量測浪費了許多時間與成本,如何使製程自動調整參數? 就可以解決時間與成本浪費問題,過去有許多研究開始解決此問題,同時也顯示端銑加工中表面粗糙度受到切削速度、切削深度、進給速率等加工參數的影響。因此,在要求效率與品質的前提下,如何能使製程自我學習,事先掌握加工結果的預測以及加工參數的線上自動調整,確有其必要性。 在過去文獻中只考量端銑製程參數是不夠的,本研究將應力因子列入考量。為此,力量檢知器被使用在監控這不易控制的切削系統中,以增加表面粗糙度控制的準確度。從觀察發現,一個適當的切削力量信號是在XY平面的平均端點力量(Fap_xy軸)和Z軸力量絕對值(Fa_z軸)的交互作用。這兩種力量表現在表面粗糙度中關係密切,證明了切削刀具對粗糙度的影響和切削力與粗糙度之間的關係。實驗中,利用表面粗糙度預測來調整製程參數,其精確度已經達到93%以上之結果。 為了精益求精,本研究以端铣加工之實驗數據,依照主成份分析法轉換加工參數成為類神經網路的訓練來源,透過類神經網路進行訓練與測試,最後利用未轉換數據訓練之結果來驗證本研究之績效,結果發現誤差函數MSE比原始數據小於4倍以上的差距,績效相當顯著。另外發現,若能將 5個變數以主成份分析轉變後輸入比少於5個變數(降維處理) 以主成份分析轉變後輸入之誤差函數MSE更小。
Surface roughness is an important indicator of the quality of machined parts. Commonly, manual technique of direct measurement is utilized to assess surface roughness and part quality, which is found to be very time-consuming and costly. How can the surface roughness prediction and machining parameters automatic control? For that reason, this research develop to solve problems and in the end-milling; the Surface roughness is a common quality characteristic which is affected by factors such as spindle speed, feed rate, and depth of cut. So with the goal of improving efficiency and processing quality, how can we develop a prediction model that will by necessity automatically adjust the parameters? In the past, the research only considers whether the processing parameters are not enough. For that reason, the dynamometer sensor was used to monitor the uncontrolled cutting tool conditions to increase the accuracy of the surface roughness control. An empirical approach was applied to discover the proper cutting force signals, the average resultant peak force in XY plane (Fap) and the absolute average force in the Z direction (Faz). These two forces were employed to represent the uncontrollable cutting tool conditions for surface roughness control. A statistical method was employed to prove that the cutting tools could influence the surface roughness, and obtain the correlation between surface roughness and the cutting force signals. The accuracy of MSE was well in advance of 93 %. However the goal of this research is to continuously keep improving the result. So in order to foster continual improvement, data collected from a manufacturing plant was utilized and treated with principle component analysis (PCA) to develop input variables for input to the neural network. This process is able to train the system to improve its predictive ability. Results provided from the original literature are used for comparison to prove the performance improvements of this research. As a result of this research it was discovered the predictive ability could be improved resulting in a 5 fold reduction of MSE. This is an obvious improvement. Another discovery is that in utilizing PCA developed input variables, input of five variables in comparison to four or less results in a lower MSE.