|
大綱
|
摘要
近年來智慧製造成為趨勢，製造業中工具機加工的成本日趨重視，過往在切削時切削參數通時是根據加工師傅所訂的，但現今客製化盛行，少量多樣的切削需求成為主流，因此若僅憑藉加工師傅的經驗來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。但超出固定切削條件外的可辨識性仍有待商榷，同時此方面的相關研究並不多。隨著現今加工產品的要求，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(Genaralization)是本研究的目標。本研究嘗試將特原有的切削條件數量減少，以確保本刀具磨耗預測模型能在少量的數據仍有不錯的準確度，藉此提升模型的泛化能力。
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。選擇加速規是由於精準度高、穩定度高、功耗低、結構簡單，不易受雜訊及溫度波動的影響，適合做輕微結構的動態測試(如本研究之微銑削)，相較於動力計安裝於虎鉗更為方便，因此適用本研究之訊號量測，同時主軸電流訊號與切削力有高度正相關，故選擇加速規、電流勾錶做為訊號來源。本研究利用上述特徵以隨機森林(Random
forest)及皮爾森積動差相關係數 (Pearson correlation
coefficient)選取時域、頻域的重要特徵，將其重要特徵置於元學習(Meta
Learning)以建立刀具磨耗預測模型，為提升其泛化能力，將排列好之切削條件依序放入，探討切削條件順序對整個模型準確度之影響。
關鍵字：Meta learning、振動訊號、加速規、刀具磨耗、銑削、深度學習
Abstract
In recent years, the manufacturing industry has paid more and more
attention to the development of models for predicting machine failure
or tool wear. In the past, there are researches studying tool wear
monitoring system based on data analysis. Most of them built tool wear
monitoring model using cutting conditions, for example speed and feed,
according to ones they had already set. By using these models, we can
correctly identify the testing data which correspond to the condition
they set. In contrast, few researches that studying identifying the
data which is not correspond to the condition we set. Therefore,
generalization of models using these methods still needs to be
discussed. In this research, data corresponding to specific cutting
condition is excluded from training data set. Then, these data will be
included to testing data set to test the accuracy of the model. The
generalization of tool wear monitoring model could be test by this
method.
The accelerometer is attached on vise to capture vibration signal, and
the current clamp meter is used to capture spindle current signal.
These signal will be the features of their tool wear condition.
Accelerometer is small, easy to install, relatively low cost and is
has high signal-to-noise ratio. Using vibration signal can monitor the
change of tool status more instantaneously than the cutting force. The
spindle current has highly positive correlation with cutting force. In
this study, vibration and current signal will be used to build a
random forest model to identify the wear status. In order to test the
generalization of the model, some data capture in specific cutting
condition is exclude from training set but include in testing set. The
result is that the accuracies are higher than 80% in almost all
conditions, and the lowest on is about 65%. After standardization, the
power factor will be removed so that features of structure vibration
when feature selecting, and the lowest accuracy can increase to 75%,
in addition, the problem that model fail identified serious wear into
little wear is solved. Therefore, the model can identify tool wear
status more correctly.
Keywords:
vibration
signal,
accelerometers, tool wear, machine learning, milling
|
