Neural networks for online prediction of quality in gas metal arc welding

Modern welding equipment often features a rather complex operator interface that can make it somewhat difficult, even for an experienced welder, to determine optimum settings for a given welding job. For example, the joint fitup may vary, or some form of unexpected contamination may occur. In addition, in welding tasks for which a procedure has been specified, and in automated welding, changes in welding conditions may mean that some adjustment to the pre-established welding parameters is desirable. With online signal inputs from the welding process, artificial neural networks offer the possibility of providing signals that can be used for control, either indirectly by advising the operator of problems when the system conditions have deviated from satisfactory operation, or by direct feedback control of the welding equipment. This paper reports the development of a prototype which takes arc voltage data as online input, and applies the data to a neural network. The neural network has been trained to output the welding metal transfer mode, and whether the operating regime is satisfactory from the point of view of producing a good quality final weld. Various data preprocessing schemes have been investigated, and it has been found that, with suitable processing, accurate quality prediction is possible for both shortcircuiting and spray metal transfer mode.