A machine learning approach for simulation of multi-stage laser forming process

Abstract

Bending angle obtained in laser forming process is much smaller than that in conventional die pressing or roller bending processes. Therefore, multi-stage forming is necessary to form deep drawn shapes and/or complex shapes by laser forming. However, laser irradiation varies material properties, posture, and stiffness of metal plates, and those influences are accumulated during the consecutive stages of process. Thus, thermo-elastoplastic deformation models in theoretical analysis cannot predict the final shape precisely. Regarding this problem, authors propose a laser forming simulator, which employs the artificial neural network to correlate the process parameters and the deformation of metal plates, in this study. Teaching data for machine learning of the neural network is collected through the multi-stage laser bending experiments with stainless-steel plates and a high-power diode laser. The trained network is used to simulate the plate deformation to demonstrate the feasibility of proposed method. And the influences of network structure on machine learning are investigated, and the influences of conditions are discussed in aspect of prediction accuracy. Trained neural network acquired a relationship between the irradiating conditions and the deformation of plates, and work as a simulator to predict the shape of plates formed by consecutive bent at laser scanning paths. Prediction accuracy of the simulator was same as the accuracy of shape obtained by laser bending experiments.