Prediction of weld quality in pulsed current GMAW process using artificial neural network

Abstract

Weld joint dimensions and weld metal mechanical properties are important quality characteristics of any welded joint. The success of building these characteristics in any welding situation depends on proper selection-cum-optimisation of welding process parameters. Such optimisation is critical in the pulsed current gas metal arc welding process (GMAW-P), as the heat input here is closely dictated by a host of additional pulse parameters in comparison to the conventional gas metal arc welding process. Neural network based models are excellent alternatives in such situations where a large number of input conditions govern certain outputs in a manner that is often difficult to adjudge a priori. Six individual prediction models developed using neural network methodology are presented here to estimate ultimate tensile strength, elongation, impact toughness, weld bead width, weld reinforcement height and penetration of the final weld joint as a function of four pulse parameters, e.g. peak current, base current, pulse on time and pulse frequency. The experimental data employed here are for GMAW-P welding of extruded sections of high strength Al–Zn–Mg alloy (7005). In each case, a committee of different possible network architectures is used, including the final optimum network, to assess the uncertainty in estimation. The neural network models developed here could estimate all the outputs except penetration fairly accurately.