The use of neural networks to characterise problematic arc sounds

Automation of electric arc welding has been at the centre of considerable debate and the subject of much research for several decades. One conclusion drawn from all this effort is that there seems to be no single system that can monitor all of the variables and subsequently, fully control any welding process. To date there has been considerable success in the development of seam tracking systems employing various sensing techniques, good progress has been made in the area of penetration measurement and worthwhile use has been made of the integration of expert systems and modelling software within these control domains. Skilled welders develop their own monitoring and control systems and it has been observed that part of this expertise is the ability to listen subconsciously to the sound of the arc and to alter the electrode position in response to an adverse change in arc noise. Attempts have been made to analyse these sounds using both conventional techniques and more recently expert systems, neither have delivered any usable information. This paper describes a new approach involving the use of neural networks in the identification of sounds which indicate that the welding system is drifting out of control.