Recurrent neural networks to model input-output relationships of metal inert gas (MIG) welding process

The mechanical strength of weld-bead is dependent on its geometric parameters like bead height, width and penetration, which depend on input process parameters, namely welding speed, arc voltage, wire feed rate, gas flow rate, nozzle-to-plate distance, torch angle etc. Recurrent neural networks were used for conducting both forward and reverse mappings using three approaches. The first approach dealt with the training of Elman network through updating its connecting weights using a back-propagation algorithm. In second approach, a real-coded genetic algorithm was used along with the back-propagation algorithm to tune the network. The third approach utilised a real-coded genetic algorithm only to optimise the network. In forward mapping, third approach was found to outperform the others, but in reverse mapping, first and second approaches were seen to perform better than the third one. The performances of these approaches were found to be data dependent.