Use of neural networks for parameter prediction and quality inspection in TIG welding

Abstract

The potential of using neural networks in TIG welding has been investigated. It has been successfully demonstrated by using a commercial package, BRAIN-MA KER, on an IBM PC that the feed forward multilayer neural network, trained by the back-propagation algorithm with suitable structures, can give satisfactory predictions for both manual and mechanised TIG Welding. A neural network has been trained to predict the welding conditions (procedure) necessary to produce a good quality weld. The types of defect arising by the use of other welding conditions are also predicted. The feasibility of incorporating a neural network into a welding control system is also discussed.