Bi-directional Prediction between Weld Penetration and Processing Parameters in Electron Beam Welding Using Artificial Neural Networks

Abstract

The bi-directional prediction between processing parameters and weld penetration benefits electron beam welding (EBW) production by reducing costly trials. An artificial neural network (ANN) model was established for the bi-directional prediction between them in EBW. The main processing parameters consist of accelerating voltage, beam current and welding speed, while weld penetration indicates penetration depth and penetration width of weld. The training and test sets were collected through EBW experiments by using 1Cr18Ni9Ti stainless steel. Two-layer supervised neural networks were used with different number of hidden layer nodes. Comparison between experimental and predicted results show the maximum absolute-value error is 6.6% in forward prediction from the main processing parameters to weld penetration, while that is 23.6% in backward prediction reversely. Combined the higher accurate forward prediction with the easy-use backward prediction in EBW production, a flow chart is proposed for optimizing prediction of processing parameters.