元學習用於銑削加工中刀具磨耗偵測

Abstract

由於製造業中工具機加工的成本極為龐大，因此近年來智慧製造成為趨勢。過往在切削加工時參數是根據操作人員的經驗所訂，但現今客製化產品盛行，少量多樣的切削加工成為主要的需求，因此若僅憑藉操作人員的經驗進行加工參數的調整，其人力成本、刀具成本相當可觀。以往研究由工具機擷取出各種加工訊號來建立出刀具磨耗的預測模型，但大多刀具磨耗預測模型是以原先設定好之固定切削條件(如進給、切深、刀具半徑、轉速)所建立，相對來說刀具磨耗預測的準確度也較高。但預測模型對於固定切削條件外的可辨識性仍有待商榷。隨著現今加工客製化，固定的切削條件進行刀具磨耗預測已無法滿足，提升各種切削條件的刀具磨耗預測模型泛化性(Generalization)是本研究的目標。 本研究以電流勾錶擷取主軸電流訊號，並同時將加速規置於虎鉗以擷取振動訊號，最後以這些訊號作為刀具磨耗狀態的特徵進而建立刀具磨耗預測模型。本研究利用9種不同的切削條件建立9個特徵集合，最後以FCNN (全連接層類神經網路)為基底之元學習(Meta Learning)以建立刀具磨耗預測模型。研究結果顯示元學習對特徵集合的輸入有其順序性，越早輸入之特徵集合對最終模型的影響力較小，反之則愈大，因此本研究主要探討9種不同切削條件的輸入順序，找出其最佳化特徵集合的輸入順序，用於建立刀具磨耗預測模型，提升其整體泛化能力，並探討切削條件順序對整個模型準確度之影響。

Smart manufacturing has become a trend in recent years due to the enormous cost of tool processing in manufacturing. In the past, cutting parameters were based on the experience of operators, but now customized products are prevalent, and a small number of different cutting processes become the main requirements. Therefore, if the parameters are adjusted only by the experience of operators, the labor cost and tool cost will be considerable. Previous studies have established tool wear prediction models by extracting various processing signals from tool machines, but most of them are based on fixed cutting conditions (such as feed, cut depth, tool radius, rotation speed) that were originally set up, and the accuracy of tool wear prediction is relatively high. However, the predictive model's identifiability beyond fixed cutting conditions remains to be discussed. With today's customization of cutting, predicting tool wear under fixed cutting conditions can no longer be satisfied. To improve the generalization of tool wear prediction models under various cutting conditions is the goal of this study. In this study, the current signals of the spindle are retrieved by a current clamp meter, and the accelerometer is placed on the vice to capture the vibration signals. Finally, the tool wear prediction model is established based on these signals as the characteristics of the tool wear state. This study uses nine different cutting conditions to set up nine feature sets. Finally, a tool wear prediction model is built using FCNN (Full Connected Layer Neural Network) as the base of Meta Learning. The results show that Meta Learning has a sequential effect on the input of feature sets. The earlier the feature sets are input, the smaller the influence on the final model, and the larger the converse. Therefore, this study mainly explores the input order of nine different cutting conditions, finds out the input order of the optimal feature sets, and is used to build a tool wear prediction model, improve its overall generalization ability, and explore the effect of the order of cutting conditions on the accuracy of the entire model.