應用機器學習於線切割表面輪廓點之預測

Abstract

線放電表面輪廓是指線放電加工後的形貌，其中包含細微的放電表面，如裂縫及凹坑。本實驗透過顯微鏡觀察加工形貌，並從放電加工的原始波形中擷取出物理物理量，針對單一時刻下放電輪廓點進行預測之研究。 表面輪廓點為放電過程中，在極短時刻下的加工厚度(高度)值，其構形出放電輪廓。表面輪廓點為複雜間隙狀態下加工出之結果，包含線振位移及間隙距離。又同時量測兩者具困難性，本實驗藉由觀察在加工過程中的放電波形，從中分析兩者的關係。 本研究透過從線放電原始波形中擷取多變量物理量，並通過機器學習建立模型。基於過去的研究，本研究擷取放電延遲時間、加工能量等放電物理量，另外，本研究導入了一種新物理量，如:大(中、小)電流發數，並建立放電模型。 為了提高模型的預測能力，針對放電延遲時間及大小電流判別，進行物理量擷取方式改變，並進行模型錯誤率的比較。而模型中的層數也對模型有著重要影響性，模型將透過採用錯誤率較低之物理量擷取方式，與適合的模型層數。 本研究透過本研究之物理量擷取和機器學習，建立原始放電波形與加工表面輪廓點的數學模型，並且實現對表面輪廓點變化的預測。

The surface profile of WEDM refers to the morphology machined by WEDM, which includes subtle discharge surfaces, such as chipping and cracks. In this research, surface was measured by a microscope. By extracting physical variables in the original waveform of WEDM, so as to predict the profile points in short period . The surface profile points are the height of machining during the discharge process whithin short period of time, which form the discharge contour. Profile points are related to wire deflection and gap state, and it is difficult to measure both at the same time, In this experiment, the relationship between the discharge waveform and the surface profile was analyzed by observing both during the processing. This research extract multi-variabled features from the WEDM original waveform, and the model is established by neural network,. In this research , some physical varaables that highly ralated to the behavior of WEDM process will be extracted, such as Ignition delay time、Pulse energy etc. Furthermore, the application of a variable-Big Current count (Middle、Small), is used to build the model. In order to improve prediction ability of model, the extraction method of Ignition delay time and Big current will be compared by learning loss. The number of layers in the model has an important impact on the model as well . The model will be improved by extraction and model layers with low error loss The mathematical model between original WEDM waveform and gap is established, by feature extraction developed by this research and the technique of Machine learning. And the prediction on the gap trend is achieved.