Laser beam welding is getting more and more popular and is on the way to be a standard production technology like laser cutting. In the case of cutting it is possible to determine easily the quality of the cut (with the eyes). In the case of laser welding it is not possible to look into the seam and often it is even not possible to see the roof of the seam because of the construction. Therefore a process control of the full penetration in laser welding of a workpiece independent of the industrial environment becomes more and more important.
To monitor the degree of full penetration the temporal fluctuations of the welding plasma above the workpiece can be detected. Because the fluctuations will change by changing different process parameter like beam power, focal position, speed and gap, a neural network is used for analyzing the signals.
The neural network used is based on a feed forward structure and trained by a back-propagation method, that means the use of pattern (combination of input vector and desired output vector) and the modification of the intensity of coupling between the neurons. The input vector is the temporal fluctuation of the plasma intensity. The output vector is the determination between no seam and full penetration welding.
With such a training of a neural network it it possible to classify or to detect the degree of full penetration in laser welding with a guarantee better than 98,5% by analyzing the plasma above the workpiece.