Prediction of Process Parameters for Submerged ARC Welding Process Using Back-Propagation Neural Networks

In submerged arc welding (SAW), selecting appropriate values for process variables is essential in order to control bead size and quality. This paper proposes the development of neural network models for prediction of weld quality in Sub-merged Arc Welding. Initially experimentation has been carried out using Taguchi-based L27 experimental design and proposed neural network models are developed using experimental data. In the present work, the quality characteristics considered are weld bead hardness, weld bead width and width of heat affected zone (HAZ). The essential process input parameters were welding current, arc voltage, welding speed, electrode stick out length and % of reuse of flux. Two dif-ferent back-propagation training algorithms such as Baysian Regularization (BR) and Levenberg Marquardt (LM) have been employed for modeling. Out of 27 dataset 22 dataset is used for the training purpose and remaining 5 dataset is used for testing purpose. Altogether 28 networks has been trained and tested by varying the hidden layer neurons from 2 to 15. . Results ensure best prediction capability of BPNN with BR with least mean absolute percentage prediction error