| 摘要: | 本論文旨在研究線切割放電加工SKD11模具鋼時，電火花品質與表面粗糙度間的關係，探討線切割放電加工時正常放電、電弧放電與短路放電，於加工的品質特性如工件寬度、加工時間與表面粗糙度的影響，並利用正常放電次數、電弧放電次數和短路放電次數三種參數，以超參數優化之機器學習模型預測工件之表面粗糙度。
本研究係使用超參數優化之機器學習模型來預測線切割放電加工SKD11模具鋼之表面粗糙度，所使用的模型包括ANN、SVR和XGBoost，其中使用了網格搜索法、貝葉斯優化法和TPE等三種超參數優化法對模型的超參數進行優化，並比較了優化前與優化後的模型，結果顯示表現最佳的Validation MAPE與測試組 MAPE為TPE優化ANN模型，其Validation MAPE為1.86%，而MAPE為1.472%，在新增10組測試組的預測中，結果顯示表現最佳的MAPE為貝葉斯優化與TPE優化ANN模型，其MAPE分別為0.95%與1.01%，綜上所述，TPE優化ANN模型為本研究中之最佳模型。
;In this study, the relationship between electrical discharge spark quality and surface roughness of SKD11 die steel in wire electrical discharge machining (WEDM) was investigated. The effects of normal sparks, arc sparks, and short sparks on quality characteristics such as the workpiece width, machining time, and surface roughness were examined. Using machine learning models optimized with hyperparameter tuning, the surface roughness of a workpiece by employing three input parameters, namely normal spark counts, arc spark counts, and short spark counts was predicted. The multiple hyperparameter optimization techniques were used, including artificial neural networks (ANNs), support vector regression, and extreme gradient boosting, to predict the surface roughness of SKD11 die steel in WEDM. Three hyperparameter optimization methods, namely grid search, Bayesian optimization, and tree-structured Parzen estimation (TPE) were used to optimize the hyperparameters of the models. The models were compared before and after optimization. The results indicated that among all models, the TPE-optimized ANN model exhibited the highest performance in terms of mean absolute percentage error (MAPE) and validation MAPE, with the respective values of 1.472% and 1.86%. Prediction of an additional 10 test groups revealed that, among all models, the Bayesian-optimized ANN model and TPE-optimized ANN model exhibited the highest performance, with a MAPE of 0.95% and 1.01%, respectively. In conclusion, the TPE-optimized ANN model is the optimal model in this study. |
