| 摘要: | 雷射加工技術已成為切割薄型矽鋼片，提供高性能疊片鐵心的替代方案，本研究對使用脈衝光纖雷射進行薄型矽鋼片直線與曲線切割之品質進行探討。首先，藉由隨機森林演算法(RFM)和響應表面演算法(RSM)確認輸入雷射加工參數對切割品質的影響;接著，開發人工智慧(AI)模型，以預測和優化切割品質的各種特性。在預測切割品質方面，本研究開發了更可靠的高效能深度神經網路(DNN)模型。此外，還提出了新的最佳化方法，可以在給定的初始實驗製程參數範圍之外找到最佳的雷射加工參數。
在第一部分中，本研究使用DNN和基因演算法(GA)進行尺寸品質的預測和優化，包括圓形切割的真圓度與方形切割的切口寬度。首先確認三個加工參數，即雷射功率、雷射脈衝頻率和切割速度，是對幾何尺寸品質有很大的影響，從而適當地運用在預測模型的輸入參數中。接著使用GA來選定最優的DNN架構，而最終的DNN模型是透過預先訓練和微調的過程取得。依此建立的DNN模型在真圓度和切口寬度的預測中具有良好的表現，不管在資料訓練集、驗證集及測試集的預測上，皆具有極低的絕對百分比誤差(MAPE)和非常高的絕對變異分數值(R2)。此外，此DNN模型的表現更優於其他基於AI的模型，包括隨機向量功能連結(RVFL)與支援向量迴歸(SVR)。最後，利用DNN-GA結合模型所找出的最佳製程參數，進行實際實驗驗證，確認所產出的真圓度及切口寬度為最小值，即為最佳幾何品質。
本研究的第二部分，係利用DNN與改進的灰狼優化器(I-GWO)，預測並優化在不同環境下雷射切割工件的殘渣，工作環境包括油、酒精及空氣，所考慮的加工品質有真圓度、上表面的殘渣高度、下表面的殘渣高度、上表面的殘渣寬度及下表面的殘渣寬度。所考慮的雷射切割製程參數，即工作環境、雷射功率、脈衝頻率及切割速度，皆對工件上殘渣的形成有相當大的影響，在油中切割所形成的殘渣會比酒精及空氣中來的更少。在模型訓練方面，透過堆疊自動編碼器方法與多目標灰狼最佳化工具相結合，先產生預先訓練的DNN，隨後進行微調以獲得最終的DNN。之後，利用I-GWO找出產生最少殘渣加工品質的最佳加工參數組合。依此所開發的DNN模型效能優於RVFL及SVR的效能。由上述DNN和I-GWO演算法所預測的最佳加工參數經過實驗驗證，確實能產生最少殘渣的最佳加工品質。
在第三部分曲線雷射切割中，實驗在油中進行，考慮雷射功率、雷射脈衝頻率、切割速度及曲率半徑作為可控的輸入加工參數，輸出品質特性包括切口寬度、內熱影響區、外熱影響區及再熔接的部分，所選定的輸入參數確實都會影響輸出品質特性。在DNN模型建構方面，是先使用平衡最佳化工具(EO)訓練及優化，再經過微調求得一個具有五層隱藏層的最終DNN模型。該DNN模型的表現優於淺神經網絡(SNN)、廣義迴歸神經網路(GRNN)及自適應神經模糊推理系統(ANFIS)模型。此外，此部分研究亦改良了EO優化器，並結合上述DNN模型，找出可產出最佳輸出品質特徵的最佳加工參數組合。實驗結果證明本研究所開發模型的有效性及穩健性，確實能產生最佳曲線雷射切割的品質，大幅提升每一個品質指標的質量;同時，與其他研究所開發的模型相比較，本研究所開發的預測及最佳化模型，亦具有相對的優越性。
;The laser machining technique has become a promising alternative for cutting thin electrical steel sheets to provide high-performance laminated cores. In this work, the use of pulsed fiber laser was investigated for straight and curved cutting of thin non-oriented electrical steel sheets. The significant effects of the input parameters on cut quality were confirmed via random forest method (RFM) and response surface method (RSM). Artificial intelligence (AI)-based models were developed for predicting and optimizing various characteristics of cut quality. New methods to obtain deep neural network (DNN) models with reliable performance in predicting cut quality were proposed in this study. In addition, new improved optimization methods were also proposed. Based on the results obtained, the optimal laser process parameters were found beyond the process window of the initially given experiments.
In Part I, prediction and optimization of geometrical qualities, namely roundness of circular cut and kerf width of square cut, were performed using DNN and genetic algorithm (GA). Three process parameters, namely laser power, laser pulse frequency, and cutting speed, were considered to experimentally investigate their effects on geometrical quality. All the process parameters significantly affected the cut quality and were properly used as input variables in the prediction models. A real-coded GA was employed to determine the optimal DNN architecture, and the final DNN models were obtained through pre-training and fine-tuning processes. The developed DNN models showed great ability in prediction of roundness and kerf width, as demonstrated by a very low mean absolute percentage error (MAPE) and a very high absolute fraction of variation (R2) for training, validation, and testing datasets. In addition, the performance of the DNN models were better that that of other AI-based models, namely random vector functional link (RVFL) and support vector regression (SVR). The predicted optimal geometrical qualities of the DNN-GA models were verified by validation experiments in which a combination of the smallest roundness and kerf width was generated.
In Part II, dross formation of laser cutting in different environments, namely oil, alcohol, and air, was predicted and optimized using a DNN and an improved grey wolf optimizer (I-GWO), respectively. Five quality indices were used to define the dross formation, namely roundness, dross height on top side, dross height on bottom side, dross width on top side, and dross width on bottom side. The laser cutting process parameters, namely working environment, laser power, pulse frequency, and cutting speed, had a significant influence on the dross formation. In addition, cutting in oil led to less dross formation than in alcohol and air. A stacked autoencoder method combined with a multi-objective GWO was employed to generate a pre-trained DNN, followed by a fine-tuning process to obtain the final DNN. The I-GWO was used to determine the optimal combination of process parameters for minimum dross formation. The performance of the developed DNN model was higher than that of RVFL and SVR. The predicted optimal process parameters by the DNN and I-GWO algorithms were verified by validation experiments in which the minimum dross formation was generated.
In Part III, the experiments of curved cutting were performed in oil, considering laser power, laser pulse frequency, cutting speed, and curvature radius as the controllable input parameters. The output quality characteristics included kerf width, inner heat affected zone, outer heat affected zone, and rewelded portion. All the input parameters significantly affected the cut quality. A 5-hidden-layer DNN model was obtained by pre-training using an equilibrium optimizer (EO), followed by a fine-tuning process. The performance of the 5-hidden-layer DNN outperformed the shallow neural network (SNN), generalized regression neural network (GRNN), and adaptive neuro-fuzzy inference system (ANFIS) models. A new, modified EO was developed and employed with the DNN to determine the optimal laser process parameters for the optimal cut quality. The results of the validation experiments proved the robustness of the models developed in this study, where the best cut quality was generated and a considerable improvement was found for each quality index. A comparative analysis supported the superiority of the developed models over those in other studies. |
