Software development for prediction of the weld bead in CMT and pulsed-MAG processes

Abstract

Prediction of weld bead shapes prior to welding has been a challenge for researchers in recent years. The assurance of a weld shape with adequate penetration and dilution, good wetting angles and no undercuts leads to less repairing and reworking and thus, lower cost and time consumed. CMT is a variant of MIG/MAG welding process recently introduced in the market. Software control for bead-shape prediction helps the user to get familiar with this modern technology. In the present paper, two approaches were undertaken for software development. These were applied to CMT and pulsed-MAG welding and the results were compared. The graphical interface was built to help the user choose the welding parameters and get the representation of the weld bead profile. The two approaches used to predict the weld bead profiles are neural networks and interpolation method. The data entries for the two different methods are: welding process (CMT and MAG-pulsed), plate thickness, travel speed and wire feed speed. The output data is the complete welding profile defined by X and Y coordinates of the points on the welding profile. The neural networks were trained with different numbers of neurons, multiple layers and different networks, and welding data bases with different structures were used, in order to find the best fitting neural network. The final results for each method were checked against real profiles and it was concluded that the neural network method was more flexible and the accuracy of each methods was nearly the same.