Weld parameter prediction using artificial neural network: FN and geometric parameter prediction of austenitic stainless steel welds

Abstract

A welding robotization has been used to improve the weld quality, minimize the process of trial and error, ensure process repeatability, and above all avoid the requirement for a highly qualified professional. The objective of this work is to verify the possibility of welding robot be programmed by the desired characteristics of the weld bead and in the case of use stainless steel also by the amount of ferrite in the weld bead. To that, experimental data were obtained under laboratory conditions, using an industrial robot that made welds with gas metal arc-welding process. Welds were made with different values of voltage, current, and different filler material and the following output parameters were measured from the weld bead: ferrite quantity, width, reinforcement, and penetration. Were used three different austenitic stainless steel welding wires and the same plate material (AISI 304), among other parameters that were kept constant. A fed forward artificial neural network, fully connected and supervised learning, was created from the experimental data. The mean absolute percentage error found to ferrite quantity was 4% and maximum was 17%. To width, penetration, and reinforcement of the weld beads, mean absolute percentage errors were, respectively, 5, 6, and 15% and the maximum 20, 23, and 47%. Artificial neural networks are able to predict the great complexity existing between the welding parameters in this case. This statement was made comparing the results with other methods of ferrite prediction and geometric parameter prediction.