U. Dilthey et J. Heidrich, Using Al-methods for parameter scheduling, quality control and weld geometry determination in GMA-welding, ISIJ INT, 39(10), 1999, pp. 1067-1074
At the ISF-Welding institute at Aachen University neural networks have been
used for some years, for developing effective quality control systems of G
MA welding. Also artificial intelligence methods are applied to recognise t
he quality and to calculate the seam geometry (seam height and thickness).
In addition the neural networks are used as target functions for genetic pr
ogramming in order to find out an optimised welding parameter set.
Primary welding parameters and statistically evaluated transient signals of
the welding process are taken as an input data record for a neural network
.
Since it is very time Consuming to obtain a comprehensive information about
the quality of the whole weld beat by evaluating numerous metallographic i
mages, a software tool which calculates the beat geometry from the data sup
plied by a laser scanner, has been developed.
Recognition rates of neural networks between 90 and 100% have been achieved
for short acid pulsed are processes in online quality control and in off l
ine parameter optimisation. The error in the geometry prediction by the neu
ral network was found to be within the range of 2-12%.