Estimation of Geometry and Properties of Weld Bead Using Artificial Neural Networks

Investigations on prediction or estimation of output responses in welding using artificial neural networks (ANN) have become popular among researchers. In this work, metal active gas (MAG) welding was implemented to join EN-8D medium carbon steel plates together by varying welding current, weld voltage and torch traverse speed as input parameters. Depth of penetration, reinforcement, hardness and bend angle at failure were the responses. Then the input parameters and output parameters are used to train the neural networks in this work. Feed forward network with Levenberg-Marquardt training function is implemented. 3-10-4 model of ANN is used for the prediction of depth of penetration, reinforcement, hardness and bend angle. From the regression chart, it is found that the designed model predicted the results of both replications with quite less error and hence, the effectiveness of the technique.