Prediction and optimization of geometrical quality for pulsed laser cutting of non-oriented electrical steel sheet. (May 2022)

• Record Type: Journal Article
• Authors: Rohman, Muhamad Nur; Ho, Jeng-Rong; Tung, Pi-Cheng; Tsui, Hai-Ping; Lin, Chih-Kuang
• Is Part Of: Optics & laser technology. Volume 149(2022)
• Publication Date: 2022-05
• Subjects: Laser cutting -- Non-oriented electrical steel -- Roundness -- Kerf width -- Deep neural network -- Genetic algorithm; Optics -- Periodicals; Lasers -- Periodicals; Electronic journals; 621.366
• DOI: 10.1016/j.optlastec.2022.107847
• Languages: English
• ISSNs: 0030-3992

Abstract

Graphical abstract: Highlights: Cutting of electrical steel sheet is performed using pulsed fiber laser. Laser power, pulse frequency, and cutting speed are important process parameters. DNN models are developed for predicting laser cutting quality of electrical steel. DNN and GA models are reliable and effective for predicting laser cutting quality. Optimal combination of process parameters generates the best cutting quality. Abstract: Prediction and optimization of geometrical qualities, namely roundness of circular cut and kerf width of square cut, for pulsed laser cutting of non-oriented electrical steel are performed using deep neural network (DNN) and genetic algorithm (GA). Analyses using random forest method and response surface method show that laser power, pulse frequency, and cutting speed significantly affect the cutting qualities and are properly used as input variables in the prediction models. A real-coded GA is employed to determine the optimal DNN architecture, and the final DNN models are obtained through pre-training and fine-tuning processes. A binary-coded GA is utilized to determine the optimal combination of process parameters for generating the optimum geometrical qualities. The developed DNN-GA models show great ability in prediction of the roundness and the kerf width, as demonstrated by a very low mean absolute percentage error (≤2.60%) and a very high absolute fraction of variation (≥0.9972) for training, validation, and testing datasets. In addition, the performance of the DNN-GA models is evaluated by means of nine statistical criteria in comparison with other artificial intelligence based models, namely random vector functional link network (RVFL) and support vector machine for regression (SVR) integrated with equilibrium optimizer (EO) and grey wolf optimizer (GWO). The results indicate that the performance of the DNN-GA models is better than that of the RVFL-EO, SVR-EO, RVFL-GWO, and SVR-GWO models. The predicted optimal geometrical qualities of the DNN-GA models are verified by validation experiments in which a combination of the smallest roundness and kerf width is generated.