Predicting Residual Weld Stress Distribution with an Adaptive Neuro-Fuzzy Inference System

Abstract

This work is an investigation into the applicability of the adaptive neuro-fuzzy inference system (ANFIS), a machine learning technique, to develop a model of the relation of residual stress distribution in a single weld bead-on-plate part to weld heat input and distance from the center of the weld line. Residual stress distributions required to train the ANFIS model were obtained through thermal elastic-plastic finite element analysis. Appropriate conditions for training the ANFIS model were investigated by evaluating the prediction error of the ANFIS model developed under various conditions. Afterward, residual stress distributions obtained by the developed ANFIS model trained under the appropriate conditions were compared with those obtained through thermal elastic-plastic finite element analysis. Discrepancies between the residual stresses obtained through the ANFIS model and thermal elastic-plastic finite element analysis were smaller than ±40 MPa in all regions. The results suggest that the ANFIS modeling had the ability to learn and generalize residual weld stress distributions in single weld bead-on-plate parts.