Process Monitoring in Absorber-Free Laser Transmission Welding of Plastics by Using Deep Learning Algorithms

Laser transmission welding offers the advantage of a non-contact and highly precise energy deposition. This enables the fabrication of complex and narrow seam geometries. Particularly in case of high-tech products, joints must be reproducibly manufactured in the smallest dimensions while maintaining high quality. Here, a process control is applied with the help of pyrometric sensors. However, the temperature is only measured indirectly and the signal depends on various factors such as material properties, size of the heat affected zone and thermal properties. Furthermore, the pyrometer only delivers a spatially integrated signal which is why no information can be given on the welding seam contour or the shape of the melt pool. The aim of this work is the analysis of the laser transmission welding process by using Deep Learning algorithms. Image frames are recorded which show the interaction area between the laser beam and material. The image will then be automatically processed by performing semantic segmentation. This allows the estimation of typical areas such as the weld pool or seam geometries. The results show a good agreement between the prediction and the ground truth with intersection over union values 0.92. The extracted geometric information is then used to predict the laser power. Here, good prediction results are achieved for laser powers < 15 W.