Application of neural network to detection of arc length, extension length and root gap in robotic welding

Abstract

Full penetration control of the weld pool in a first layer of the one-side multilayer welding is important to obtain a good quality of welding. For this purpose, the authors propose a new method, the switch back welding method, to obtain a stable back bead. A welding torch is not only oscillated in the groove, but also moved backward and forward. Both voltage and current are entered to neutral networks to estimate the wire extension and the arc length. Moreover, the gap and the deviation of the oscillation center from gap center are estimated. Training data are made from experimental results. Performance of the arc sensor is examined by giving testing data to the neural networks.