Optimization of scanning paths in laser forming by means of machine learning

Prediction accuracy from imaginary datasets

Abstract

Laser forming, a kind of non-contact bending process, has low cost and high flexibility because it doesn’t require dies or rollers. Therefore, laser forming is more suitable for manufacturing prototypes and small-batch productions than conventional processes. However, the complexity of deformation mechanism makes it difficult to optimize the process conditions for the final shape. In this study, the feed-forward neural network (FNN) is trained by teaching datasets to build a model predicting the scanning paths required for the final shape of a metal sheet. To collect the datasets for machine learning, a large number of imaginary datasets are prepared instead of the experimental ones. In addition, for verification of the trained FNN with imaginary datasets, the experimental datasets are employed to evaluate the prediction performance by the model. Obtained results show that the imaginary datasets can train the model to have a prediction accuracy as high as the experimental data.