熱變形位移為影響機台加工精度的主要誤差之一,若能有效的降低其在加工的影響,可提高機台在市場上之競爭力。使用軟體補償方法可在不改變機台之原始設計下,提升加工精度。傳統為以多變數迴歸分析方法或類神經網路建立預測模型,但多變數迴歸分析方法無法完整模擬熱變形位移之非線性特性,而類神經網路雖然具有較佳之預測精度,但當其為動態架構時,因增加N個時間為網路之輸入節點,使網路架構增大N倍,產生輸入至輸出間之傳遞延遲,造成預測誤差。因此,本文綜合多變數迴歸分析方法快速建模特性與神經網路對非線性函數之學習能力,提出多種多變數迴歸分析方法與神經網路之混成建模方法,並採用遺傳基因演算法優化混成模型之網路設計參數,預測工具機熱變形位移。 混成模型為將量測到之溫度,預先以多變數迴歸分析方法求得多個時刻之熱變形位移估計值或模型之狀態變數,經正規化後為神經網路之輸入節點,如此可大量減少網路輸入結點而降低模型之規模、減少建模時間並改善網路輸入與輸出間的傳遞延遲,進而提高預測精度。 使用有限元素分析方法,於機體結構之關鍵熱源處,施加不同大小之熱量,分析其溫度與熱變形位移變化,並以分析之結果數據建立預測模型比較模型之預測精度。最後採用實驗之量測數據建模,預測不同之實驗結果,探討混成模型之預測精度與可行性。
During the cutting process, the thermal deformations are the error to affect the accuracy of the machine tools. Therefore, improvement the precision of machine tools can increase the competitiveness in the markets. The soft compensation method can retain the orange design conditions to improve the accuracy for the machine tools. The multiple regression method and neural network are the tradition methods to compensate the thermal deformations. However, the multiple regression method cannot simulate the nonlinear characters and neural network the dynamic model of neural network will increase the input nodes for N times for static model, it cause the delay form inputs to outputs and make the prediction error. Therefore, this study combination the advantages of fast modeling for multiple regression method and mapping the nonlinear function for neural network to build the hybrid model to predict the thermal deformation for the machine tools. The inputs of hybrid models are the measured temperatures, through the estimation for the thermal deformations as the inputs of neural network. Thus, the input nodes can be reduced and decrement the learning time network size can be reduced and improvement the delay between the inputs to the outputs. The margins of various approaches are determined by an ideal model for, which has the relationships between thermal deformation and temperature distribution according to finite element analysis. Also, the prediction results of a practical grinding machine by using various approaches are compared to investigate the accuracy and feasibility for the hybrid models.