Modelling gas metal arc weld geometry using artificial neural network technology

A backpropagation network system for predicting gas metal are (GMA) bead-on -plate weld geometry from current, voltage and wire travel speed is reporte d in this study. Moreover, workpiece thickness is a variable that is taken into consideration because its effect on weld shape is to this point unknow n in practice. The database consists of some ninety six welds (cross-sectio nal weld shapes and corresponding welding parameters). DCEP polarity, C-25 shielding and electrode diameter and extension of 0.9 and 19 mm respectivel y are assumed fixed for this study-consistent with the experimental databas e used to train and test the technology. For the purposes of this investiga tion, weld bead size and shape are defined by bead width, bead height, pene tration and a new parameter, bay length at 22.5 degrees, introduced to mode l the underbead recession that occurs in deeper penetration welds. For pict orial representation, the upper bead is modelled by fitting a parabola to t he bead width and reinforcement height while a combination of parabolas is suggested for the bead shape below the plate surface given the width, penet ration and bay length. Deposit and plate fusion areas are also included. Fi nally, the reverse problem-predicting the welding parameters (current, volt age and travel speed) to achieve a given weld shape-is discussed in terms o f the study. (C) 1999 Canadian Institute of Mining and Metallurgy. Publishe d by Elsevier Science Ltd. All rights reserved.