以機器學習建立刀具磨耗量之即時監測模型

Abstract

隨著第四次工業革命的到來，客製化、精密化、快速化儼然已成為工具機的必要功能，因此在加工過程中的刀具細微變化便顯得格外重要，它將會影響到工件的尺寸精度與成本，本研究旨在以無須加裝額外的感測器下，透過馬達電流值，建立刀具磨耗監測模型，以達到即時刀具狀態監控(Tool Condition Monitoring，TCM)的目的。實驗環境是以CNC三軸銑床進行實驗，使用刀具為3刃的捨棄式刀具，以轉速 2,520 (rpm)、進給率 756 (mm/min)、切削深度 0.5 (mm) 對中碳鋼進行槽銑。訊號收集是透過架設於工具機工作台上的動力計量測切削力訊號，安裝於電控箱中的馬達驅動器來量測馬達電流訊號，以及架設於工作台上的顯微鏡來量測刀具磨耗值。訊號分析方面，將所收集到的馬達電流訊號與切削力訊號經過低通濾波器濾波之後，擷取穩定切削時段的訊號來進行分析，將該時段的時域訊號以快速傅立葉轉換得到頻域訊號，並選取頻域訊號中的振幅大小當成機器學習的訓練特徵。磨耗分析方面，在切削固定距離後分別用顯微鏡依序量測每刃的平均刀腹磨耗值，然後再取平均值得到刀具的3刃平均磨耗值，並透過曲線擬合的方式找出刀具磨耗的趨勢方程式，藉此擴增訓練的資料點數。建立模型方面，選用的機器學習演算法為隨機森林，分別建立基於馬達電流與切削力和刀具磨耗間的關係、基於切削力和刀具磨耗間的關係、以及基於馬達電流和刀具磨耗間的關係，訓練過程中透過隨機森林的袋外特徵重要性刪除不重要的特徵，以降低模型的均方根誤差，提高決定係數。最後，基於馬達電流與切削力模型結果之均方根誤差為3.1108，決定係數為0.9831；基於切削力模型結果之均方根誤差為3.0874，決定係數為0.9836；基於馬達電流模型結果之均方根誤差為5.8449，決定係數為0.9427。研究結果證實以馬達電流值來監測刀具磨耗為可行且有效的方法。

With the advent of the fourth industrial revolution, customization, accuracy and speed have become the basic functions of machine tools, so subtle changes in cutting tools during processing become more and more important. This thesis aims to establish a tool wear monitoring model through the motor current value without additional sensors, so as to achieve the purpose of real-time Tool Condition Monitoring (TCM). A CNC three-axis milling machine is used for the experiment. A three-flutes throw-away cutting tool is used to slot the medium carbon steel at the rotating speed of 2,520 (rpm) the feed rate of 756 (mm/min), and the cutting depth of 0.5 (mm). For signal collection, the cutting force signal is measured by the dynamometer mounted on the working table of the machine tool, the motor current signal is measured by the motor driver installed in the electric control panel, and the tool wear is measured by the microscope mounted on the working table of the machine tool. For signal analysis, the collected motor current signal and cutting force signal are filtered by a low-pass filter, and the signal in stable cutting segment is selected for analysis. The Fast Fourier Transform (FFT) is used to convert the signal from time-domain into frequency-domain, and the features for model training are chosen by the amplitude. For tool wear analysis, measuring the average flank wear value of each flute sequentially with a microscope after cutting a specified distance, then take the average to obtain the average wear value of the 3 flutes of the tool. To increase the number of training data, the tool wear tendency equation is obtained by curve fitting on the raw data of tool wear. The Random Forest (RF) algorithm is used to build the model correlating the relationship between the motor current and the cutting force and tool wear, the relationship between the cutting force and tool wear, and the relationship between the motor current and tool wear, respectively. Removing the unimportant features through the function called out-of-bag feature importance to reduce the Root Mean Square Error (RMSE) of the model and improve the coefficient of determination (R^2) of the model. Finally, the RMSE based on the results of the motor current and cutting force model is 3.1108 and the R^2 is 0.9831, the RMSE based on the results of the cutting force model is 3.0874 and the R^2 is 0.9836, the RMSE based on the results of the motor current model is 5.8449 and the R^2 is 0.9427. The results show that monitoring the wear of cutting tool from the motor current is a feasible and effective method.