以研磨力和數據模型建構研磨工件表面粗糙度估測

Abstract

本研究承襲實驗室開發的自動化研磨工具，透過研磨過程中六軸力規量測的研磨力資訊建立表面粗糙度預測模型，使用之表面粗糙度指標為Ra值，藉此簡化傳統上研磨工件後人工操作儀器量測表面粗糙度的過程，直接透過模型預測表面粗糙度以判斷研磨品質。 在訓練資料集蒐集部分，首先使用田口法(Taguchi method)最佳化進給速度、研磨輪轉速、研磨正向力與研磨角度等研磨工具的研磨參數，並分析其各自對表面粗糙度的影響。接著根據田口法的結果設計資料蒐集實驗，實驗中使用黃銅板進行研磨，實驗變因包含研磨輪轉速與研磨正向力各三種水準，研磨後之試片量測表面粗糙度後作為表面粗糙度模型的訓練資料使用。 在表面粗糙度預測模型訓練部分，訓練了線性迴歸、深度神經網路(DNN)與卷積神經網路(CNN)的模型架構，並使用原始特徵、統計特徵及FFT特徵對原始力資訊進行特徵提取，另外套用最小值最大值標準化與L1標準化，透過在驗證資料上計算預測Ra值相較實際Ra值的平均絕對百分誤差(MAPE)選擇最適合的模型，結果顯示FFT特徵的DNN模型，在套用L1標準化下驗證資料的MAPE可達3.17%，說明模型能準確的預測Ra值。 在表面粗糙度預測模型測試部分，透過表面粗糙度預測實驗及回磨實驗測試提出之表面粗糙度預測模型在實際研磨時的應用情形，表面粗糙度預測實驗中分別研磨黃銅板與不鏽鋼板，實驗結果顯示黃銅板測試資料上Ra預測值的MAPE可達6.96%，但不鏽鋼板測試資料上Ra預測值的MAPE則僅有19.13%，由於預測模型是透過黃銅版的研磨力資訊訓練，所以預測不鏽鋼板Ra值時由於材質特性的不同，MAPE才會明顯上升。另一方面，回磨實驗中使用模型預測黃銅板的表面粗糙度後，針對Ra值高於預期的區域進行回磨，實驗結果顯示表面粗糙度預測模型能有效預測出Ra值過大的區域，回磨也確實能降低Ra值。

This research inherits the automatic grinding machine developed by the laboratory and constructs the surface roughness prediction models using grinding force measured by a six-axis force sensor, the label used in this research is Ra. The prediction model can directly get Ra of the workpieces, rather than measure it by humans, which is effective to determine the grinding quality. In the training data set collection section, firstly use the Taguchi method to optimize the grinding parameters of the grinding tool such as feed rate, grinding wheel speed, grinding force, and grinding angle, and also analyze their respective effects on surface roughness. Then, the data collection experiment was designed based on the results of the Taguchi method, and the brass plate was used for grinding. The experimental variables include three levels of grinding wheel speed and grinding force, and the specimens are used as training data for the surface roughness model after grinding and measuring the surface roughness. In the surface roughness prediction model training section, Linear regression, deep neural network (DNN), and convolutional neural network (CNN) model architectures were trained, and feature extraction of raw force information using raw features, statistical features, and FFT features are conducted, then apply min-max normalization and L1 normalization to the input data. By selecting the most suitable model by calculating the mean absolute percentage error (MAPE) of the predicted Ra values compared to the actual Ra values on the validation data, the DNN model using FFT features with L1 normalization achieves a MAPE value of 3.17% for the validation data, showing that the model can accurately predict the Ra value. In the section on testing the surface roughness prediction model, the application of the proposed surface roughness prediction model in actual grinding is tested by a surface roughness prediction experiment and regrinding experiment. The surface roughness prediction experiment was conducted by grinding brass plates and stainless steel plates respectively, and the experimental results showed that the MAPE of Ra prediction on brass plate test data could reach 6.96%, but the MAPE of Ra prediction on stainless steel plate test data was only 19.13%, since the prediction model used is based on the abrasive force information of the brass plate, so when predicting the Ra value of the stainless steel plate, the MAPE increases significantly due to the difference in material characteristics. On the other hand, after using the model to predict the surface roughness of the brass plates in the regrinding experiment, the areas with higher than expected Ra values were reground, and the results showed that the surface roughness prediction model was effective in predicting the areas with excessive Ra values, and the regrinding did reduce the Ra values.