The validation of laser welding metallic materials is challenging due to highly dynamic process and the limited accessibility to the weld. The measurement of process emissions and the processing laser beam are one way to record highly dynamic process phenomena. However, these recordings always take place via the surface of the weld. Phenomena on the inside are only implicitly recognizable in the data and require further processing.


To increase the validity of the diagnostic process, the emission data is synchronized with synchrotron data consisting of in situ high speed images based on phase contrast videography. The welding process is transilluminated by synchrotron radiation and recorded during execution providing clear contrasts between solid, liquid, and gaseous material phases. Thus, dynamics of the vapor capillary and the formation of defects such as pores can be recorded with high spatial and temporal resolution of < 5 μm and > 5 kHz. In this work, laser welding of copper Cu-ETP and CuSn6 is investigated at the German Electron Synchrotron DESY.


The synchronization is achieved leveraging a three-stage deep learning approach. A preprocessing Mask-R-CNN, dimensionality reduction PCA/Autoencoders and a final LSTM/Transformer stage provides end-to-end defect detection capabilities. Integrated-Gradients and KernelSHAP allow for extraction of correlations between defects and emission data.


The novel approach of correlating image and sensor data, increases the informative value of the sensor data. It aims to characterize welds based on the sensor data not only according to IO/NIO, but also to provide a quantitative description of the defects in the weld.