MODELING OF SUBMERGED-ARC WELD BEADS USING SELF-ADAPTIVE OFFSET NEUTRAL NETWORKS

The non-linear relationship between the five geometric descriptors (he ight, width, penetration, fused and deposited areas) of a bead and the welding parameters (current, voltage and welding speed) of submerged are welding (SAW) has been modelled using neutral networks. A comparat ive study between multi-output networks and single-output networks, ea ch modelling one geometric descriptor, has shown the advantages of sin gle-output networks. The structure of a conventional feed-forward mult i-layer perceptron network with a single output is modified to accommo date an offset layer which offsets the inputs. This network, known as the self-adaptive offset network (SAON). has definite advantages over conventional multi-layer perceptron networks. Altogether, 21 single-ou tput neutral networks have been trained for the four types of SAW weld s investigated. These networks have achieved good agreement with the t raining data and have satisfactory generalisation. Over-training is av oided by using on-line checking facilities of the software. (C) 1997 E lsevier Science S.A.