| 摘要: | 為改善現今變頻馬達開發的過程當中，新馬達參數設計模式，一直依賴開發人員的個人設計經驗進行參數設定，造成同樣的馬達可能會因為開發人員的想法不同而有不同的參數組合，並且不斷的微調馬達設計參數以達到最佳化設計，使馬達開發階段的設計參數與測試驗證組合變得相當繁雜，因而導致開發時程一再延宕。
本研究希望透過機器學習之資料探勘的方式，顛覆現今的設計模式，建立馬達參數設計評估的模型，藉由R公司的馬達產品於開發階段所彙整的參數資料，作為研究數據集，選取與參數設計相關的重要特徵，作為建立模型的訓練與測試用，分別使用隨機森林、多輸出(向量)、梯度提升、多層感知器等四種迴歸模型進行預測，再經由網格搜尋與交叉驗證演算法對迴歸模型進行自動調參，透過模型的準確率與迴歸評估指標進行模型選用，找出適合本研究使用的迴歸模型，藉由全因子實驗設計模型並採用歐幾里得距離量測的演算法的預測，在給定目標值後，找出與目標值最接近的馬達設計參數，以縮短馬達設計驗證時程，並可以提供馬達設計參考之方向。
藉由迴歸評估指標結果顯示，以網格搜尋(隨機森林) 迴歸模型所呈現的預測模型的準確率達91%以上，整體解釋力達91%以上，調整後的模型解釋力達90%，因此本研究選取自動調參的隨機森林迴歸模型進行參數選用，透過實驗設計模型並採用歐幾里得距離量測的演算法的預測，平均可減少2到4次的馬達設計驗證次數，換算下來約可省掉40%的工程天數。
另對於網格搜尋(隨機森林) 迴歸模型設定不熟悉的人員，在使用設定上較為簡易，可利用預測模型的評估，協助開發人員提升開發速度，縮短馬達設計與驗證測試時程，以期達成客戶要望。
;To improve the current inverter motor development process, the new motor parameter design mode has always relied on the developer’s personal design experience. As a result, the same motor may have different parameter combinations due to the different ideas of the developer, and constantly fine-tuning the motor design parameters to achieve the optimal design makes the combination of design parameters and test verification in the motor development phase quite complicated, which leads to repeated delays in the development schedule.
This research hopes to subvert the current design model through the data mining method of machine learning and establish a model of motor parameter design evaluation. The parameter data collected during the development stage of R company′s motor products are used as the research data, selection and parameters Design-related important features are used for model training and testing. Four regression models such as random forest, Multi-output, Gradient boosting, and Multi-layer perceptron are used to predict, and then the regression model is analyzed through Grid-search and Cross-validation algorithms. Carry out automatic parameter adjustment and select the model through the accuracy of the model and the regression evaluation index, then find the regression model suitable for this research, use the full factorial experimental design model and Euclidean distance measurement algorithm to predict, to find the motor design parameter that is closest to the target value after the target value is given. To shorten the motor design verification time and provide the direction of reference for the motor design.
The regression evaluation index results show that the accuracy of the prediction model presented by the Grid-search (random forest regression) model is more than 91%, the overall explanatory power is more than 91%, and the adjusted model explanatory power is 90%. Therefore, in this study, a random forest regression model with automatic parameter adjustment was selected for parameter selection. Through the experimental design model and the prediction using the Euclidean distance measurement algorithm, the number of motor design verifications can be reduced on average by 2 to 4times, and the conversion can save about 40% of the engineering days.
In addition, for those who are not familiar with the setting of the automatic tuning random forest regression model, it is easier to set the user setting. The evaluation of the predictive model can be used to help developers increase the development speed, shorten the motor design and verification test timeline, to achieve customer expectations. |
