Prediction of bead geometry in pulsed current gas tungsten arc welding of aluminum using artificial neural networks

Pulsed current GTA welding, where the welding current is periodically varied between two values with a particular frequency, has acquired prominence due to the advantages it offers such as weld metal grain refinement, resistance to solidification cracking and controlled heat input. The relation between the welding process parameters used during welding and the resulting bead geometry is highly non linear and modeling the same by conventional mathematical and regression methods is very difficult. In this study an attempt has been made to predict the bead geometry parameters, given the welding process parameters, using artificial neural networks. In this case, the sets of welding variables to be studied, have been selected using Taguchi's Orthogonal array experiment technique. Using a commercially available software different kinds of networks such as Multi Layer Perceptron (MLP) and Generalised Feed Forward networks with various configurations have been built, trained using the experimental data and tested for their capacity to predict with out significant error. The results have been found to be of good accuracy and are of use In predicting bead geometry. An attempt also has been made to see if there exists a relation between top bead width which is one of the easily measurable parameters even while the welding process is going on, and the depth of penetration which can not be easily measured on line. It is found that the relation between top bead width and depth of penetration can be satisfactorily modeled using neural networks.