Gas Metal Arc Welding parameters obtainment via Artificial Neural Networks: a study of the influence of resampling in the experimental field on prediction error

Abstract Welding is the most important metal bonding process used in industry, with applications ranging from microelectronic to structural as in oil prospection constructions. Control processes in electric arc welding aim to improve the quality of the generated weld bead and also to increase process repeatability. In addition to the above factors, robotic welding minimizes the need for a highly skilled operator, reduces waste caused by failures during the welding process as well as human losses and environmental damage. Proper prediction of welding parameters can be understood as the path to process repeatability as well as guaranteed weld quality. Artificial Neural Networks (ANN) are mathematical models that resemble biological neural structures with computational capacity acquired through learning and generalization. With the ability to learn by example and to generalize information one can consider a great ally in problem solving. The general objective of this work is to evaluate the effect of replication resampling on artificial neural networks applied in the prediction of robotic Gas Metal Arc Welding (GMAW) welding parameters. Replications of 2, 3, 5, 10, 20 and 100 times were used. Replication of 2 times the number of samples presented the smallest error among the researched ones.