| 摘要: | 矽鋼是一種含矽的特種鋼，具有磁性，是電工領域廣泛使用的一種軟磁合金，故又名電工鋼，矽鋼板並以其良好的磁性能成為電機尤其是電機定轉子的優良鐵芯材料。 鑑於製造業(尤其是電動車業)對矽鋼片的需求增長，亟需要矽鋼板切割工藝和精度有所增進。近期由於脈衝雷射切割之非接觸性、靈活性、小批量生產的可用性以及與各種裝配線直接集成的能力，已成為一種實用的矽鋼板加工方法。但是使用雷射切割矽鋼板的切口寬度、切口波紋度和熱影響區 (HAZ) 等產品品質主要取決於功率、脈衝重複率和切割速度的設定，且受雷射本身的性質和材料特性的影響，獲得最佳的切割參數非常具有挑戰性。由於工件的許多特性都可以在振動信號中表示，振動信號可以提供一種有用的替代方法來預測用雷射切割時的切口寬度。出於這些原因，本研究旨在研究利用從振動信號中提取的特徵，再結合人工智慧(AI)模型，以預測無取向矽鋼板脈衝雷射切割切口寬度。本論文的主要分析包括四個主要部分：首先，先討論三種不同類型的AI 的預測模型，即機器學習 (ML)、深度學習 (DL) 和集成學習 (EL)，主要目的是得到較高的預測正確定；其次，採用不同的輸入特徵策略進行探討，包含原始時域振動信號和從小波變換技術中提取的特徵；第三，分析最佳的基礎小波選擇和最佳超參數的策略，並採用雷射掃描儀和 X-Y 工作台等兩種激光切割動作機制，以評估從實驗模型和 EL 模型獲得的切口寬度域測結果。由結果中顯示，將提取振動特徵作為輸入與基於 AI 的預測模型相結合，可以預測切口寬度至一定之預測正確度。如使用原始時域振動信號和從小波變換中選擇的特徵作為輸入特徵，DL 模型的預測精度分別為平均百分比誤差 (MAPE) 6.00% 和 5.75%；通過使用從最佳基礎小波中選擇的特徵作為輸入，ML 模型的預測為MAPE 1.69% 的；若使用原始時域振動信號結合 EL 模型並考慮兩種切割運作的預測精度，XY 工作台切割機制的 MAPE 為 5.50%，激光掃描儀切割機制的 MAPE 為 6.98%。最後本論文之總結為：此研究可以用於脈衝雷射切割矽鋼板的切口寬度預測，並且為未來開發基於人工智能的實時預測模型奠定基礎。;Non-oriented electrical steel sheet, emerging as an excellent core material for electrical machinery especially stators and rotors of electric motors due to its good magnetic properties. The need to improve electrical steel sheet cutting processes and accuracy while maintaining a flexible process and lower costs are unavoidable given the manufacturing industry′s explosive growth and demand for electric vehicles (EV). Due to its non-contact nature, flexibility, availability for small-batch production, and ability to be directly integrated with a variety of assembly lines, laser cutting has emerged as a promising method of processing electrical steel sheets. However, the product quality such as kerf width, kerf waviness, and heat-affected zone (HAZ) of cutting electrical steel sheets using laser cutting mainly depends on the optimal setup of laser power, pulse repetition rate, and cutting speed. Choosing the values for the optimal laser cutting parameters is very challenging since affected by the nature of the laser itself and the characteristics of the materials. As many characteristics of a workpiece can be represented in the vibration signals, which may provide a useful alternative judgment to predict kerf width when cutting with a pulsed laser. For these reasons, this research aimed to investigate the potential of employing extracted features from vibration signals combined with artificial intelligence (AI) based predictive models for kerf width prediction of pulsed laser cutting of non-oriented electrical steel sheets. The analysis consisted of four primary parts. Firstly, three different kinds of AI-based predictive models have been explored i.e. machine learning (ML), deep learning (DL), and ensemble learning (EL). Every predictive model has advantages and drawbacks, the main objective of exploring and comparing them is to manage and achieve the highest possible prediction accuracy. Secondly, two strategies from preprocessing the input features into the DL predictive models were considered i.e. raw time domain vibration signals and the extracted features from the wavelet transformation technic. Thirdly, the optimum base wavelet selection and strategies to select the optimal hyperparameters were explored in several notable ML models. Lastly, two laser cutting mechanisms such as the laser scanner and the X-Y table were investigated and compared to assess the kerf width quality obtained from the experimental and EL models. The results indicate that the choice of vibration-extracted features as the input to the AI-based predictive models can provide acceptable prediction accuracy for predicting the kerf width. The prediction accuracy for DL models by using raw time domain vibration signals and selected features from wavelet transformation as the input features yield 6.00% and 5.75% of mean average percentages error (MAPE), respectively. The prediction accuracy for ML models by using selected features from optimal base wavelet as the input features yields 1.69% of MAPE. Meanwhile, the prediction accuracy by using raw time domain vibration signals combined with EL models and considering two types of laser cutting movements yields 5.50% MAPE for the XY-table cutting mechanism and 6.98% MAPE for the laser scanner cutting mechanism. In general, this study lays the groundwork for future research into developing a real-time AI-based predictive model for kerf width prediction in pulsed laser cutting of non-oriented electrical steel sheets. |
