Multi-stage Inspection of Laser WeldingDefects using Machine Learning

As welding processes become faster and components consist of many more
welds compared to previous applications, there is a need for fast but still precise
quality inspection. The aim of this paper is to compare already existing
approaches, namely single-sensor systems (SSS) and multi-sensor systems
(MSS) with a proposed cascaded system (CS). The introduced CS is
characterized by the fact that not all available data are analyzed, but only
cleverly selected ones. The different approaches consisting of neural networks
are compared in terms of their accuracy and computational effort. The data are
recorded from scratch and include two common sensor systems for quality
control, namely a photodiode (PD) and a high-speed camera (HSC). As a
result, when the CS makes half of the final decisions based on a SSS with
PD signals and the other half based on a SSS with HSC images, the estimated
computational effort is reduced by almost 50% compared to the SSS with HSC
images as input. At the same time, the accuracy decreases only by 0.25% to
95.96%. Additionally, based on the CS, a general cascaded system (GCS) for
quality inspection is proposed.