Application of neural network to arc sensor

Abstract

Full penetration control of the weld pool in the first layer of a single side multilayer weldment is important to obtain a good quality weld. For this purpose, a new method, the switchback welding method, is proposed to achieve a stable back bead. A welding torch not only weaves along the groove, but also moves back and forth. Also, a neural network (NN) arc sensor is proposed that estimates the wire extension and the arc length by using measurements of both voltage and current. Moreover, from the output of the NN, the gap and the error (deviation) of the oscillation centre of the torch from the groove centre are estimated. Training data are constructed from experimental results, and performance of the NN arc sensor is examined using test data. Seam tracking is carried out via the output of the NN arc sensors: a good tracking result is obtained.