Surface Roughness Prediction Based on Surface Temperature and Tool Vibration Using BP Neural Network on Turning Machine

車削工件基於表面溫度和刀具振動的表面粗糙度BP神經網路預測

王彬(Bin Wang) ； 王丹桂(Dan-Gui Wang) ； 畢麗萍(Li-Ping Bi) ； 王輝(Hui Wang)

《中國機械工程學刊》 41卷5期 (2020/10) Pp. 615-622

摘要

In machining process, the quality of surface finish is an important requirement for many turned workpieces. Thus the controlling of machining conditions is very important for improving surface quality. This paper proposes an in-process monitoring system of surface temperature and tool vibration and discusses on surface roughness prediction based on surface temperature and tool vibration. The authors incorporate a new training scheme to BP (back propagation) neural network, namely reinforced strategy of variable learning rate (RVLR), to predict surface roughness using cutting parameters and performance characteristics, surface temperature and tool vibration. Finally, the paper shows surface roughness prediction results. Compared with SD (steepest descent) update method and traditional strategy of variable learning rate (TVLR), the RVLR method required shorter processing time to converge to the global minimum of least mean squared error. SD update method lead neural network fall into local minimum, 26.18 m^2/min^2. With either the TVLR or the RVLR method, the network was able to avoid settling at the local minimum and reach the global minimum. However, their respective least mean square errors were 9.23 m^2/min^2 and 8.21 m^2/min^2 as the best-record. In addition, the new learning algorithm only required a quarter of the TVLR processing time to reach the ＂stable＂ region. This method would be helpful in selecting cutting parameters and controlling of surface temperature and tool vibration for the required surface quality.

關鍵字

RVLR ； Surface quality ； Surface temperature ； Tool vibration

並列摘要

機械加工中，許多車削工件對表面品質要求很高。因此機械加工條件的控制對改善表面品質顯得非常重要。本文介紹了一個加工過程中表面溫度和刀具振動的監控系統，並且基於表面溫度和刀具振動討論了表面粗糙度的預測。作者本文提出了一個新的BP神經網路學習規則，基於切削參數和切削性能來預測表面粗糙度，被稱為學習率變數加強方法（RVLR）。最後，本文給出表面粗糙度的預測結果。與最速下降法（SD）和學習率變數傳統方法（TVLR）相比，學習率變數加強方法（RVLR）需要更短的處理時間收斂到全域最小值。最速下降法使神經網路模型陷入局部最小值，26.18 m^2/min^2。TVLR或者RVLR都使網路避免陷入局部最小值而達到全域最小值。然而，它們各自的最小均方差為9.23 m^2/min^2和8.21 m^2/min^2。此外，新的學習演算法只需要TVLR處理時間的1/4達到“穩定”的區域。為了達到表面品質要求，可以應用此方法選擇切削參數和控制表面溫度及刀具振動。