Camera-based spatter detection in laserwelding with a deep learning approach

Continuous quality monitoring is essential for automated
production systems and efficient manufacturing. Laser
welding processes are a key technology for many industrial
applications and must fulfill high-quality requirements. Various
influencing factors can lead to defects in the weld seam,
which impair the functionality and quality of the end product.
Therefore, a reliable quality assurance is a prerequisite for
high product quality in welding processes. An indicator for
an unstable situation in welding processes is the occurrence
of spatter on the component. Thus, the detection of spatter
can serve as a significant signal for defective weld seams. This
article proposes the detection of spatter based on a camera image
taken with an industrial camera, which is usually already
integrated in the laser system. Due to the large variance of
weld seams in image-based analysis, algorithms with a high
degree of generalization are required. Using convolutional
neural networks (CNN) and semantic segmentation the camera
image is analyzed and classified pixel by pixel. The CNN
is trained in a multi-class approach in order to recognize the
weld seam as well as the spatter as result classes. The segmentation
map constitutes the classification result. The results of
the deep learning algorithms are evaluated by different methods
and conclusions about their prediction quality are made.