| dc.description.abstract | 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.
| en_US |