A Neural Network Quality Classifier For Tig Welding Without Filler

Abstract

A neural network based on error back propagation learning algorithm has been successfully trained and tested as a quality classifier for TIG welding of stainless steel without filler. The classifer consists of two parallelly connected sub-networks, one for the quality of bead penetration and the other for bead profile. The criterion for the termination of training and the decision rule for the network prediction are self-consistent and are both related to the error tolerance used during training. Three types of borders between the desired classes have been predicted. In contrast to conventional understanding, the accuracy of the classifier can be improved and the size of the borders be reduced by choosing a relatively large error tolerance.