High-speed visual control of laser welding processes by cellular neural networks (CNN)

Former investigations showed that many errors in laser welding processes are detectable by analyzing the parameters of the keyhole shape and the melt. By performing this analysis in real time, the welding process can be controlled and errors can be eliminated as they occur. The high dynamics of the process require constant image processing frame rates of about 10 kHz. Therefore, we decided to use a CNN based camera architecture allowing a pixel-parallel processing with frame rates of up to 10 kHz. To observe the welding process, the camera is connected to the optics of the welding machine coaxially by a beam splitter. The camera input is filtered to obtain wave lengths of infrared light. The image shows the interaction zone and its environment as seen by the welding beam.