Prediction of Residual Stresses by Radial Basis Neural Network in HSLA-65 Steel Weldments

Abstract

This paper investigates the residual stress fields in the vicinity of weld bead in HSLA-65 steel weldments using a neural network. This study consists of two cases: (i) <em>the experimental analysis </em>was carried out on the measurement of residual stresses by XRD technique. Many different specimens that were subjected to different conditions were studied. The values and distributions of residual stresses occurring in welding of HSLA-65 plate under various conditions were determined. (ii) <em>The mathematical modeling analysis </em>has proposed the use of radial basis (RB) NN to determine the residual stresses based on the welding conditions. The input of RBNN are welding current, welding voltage, welding heat input, travel speed of welding, wire feed speed and distance from weld. The best fitting training data set was obtained with 18 neurons in the hidden layer, which made it possible to predict residual stresses with accuracy of at least as good as the experimental error, over the whole experimental range. After training, it was found that the regression values<em> (R</em><em>²</em>) are 0.999664 and 0.999322 for newrbe and newrb functions respectively. Similarly, these values for testing data are 0.999425 and 0.998505, respectively. Based on the verification errors, it was shown that the radial basis function of neural network with newrbe function is superior in this particular case, and has the average error of 7.70% in predicting the residual stresses in HSLA-65. This method is conceptually straightforward, and it is also applicable to other type of welding for practical purposes.