On-line monitoring of the quality of laser welding is of interest for many industrial applications. For photodiodes the monitoring strategy usually aims at observing whether the signal exceeds a threshold. This well known technique is mainly based on empirical values, thus the monitoring has to be trained for each application. For improved understanding of the context between the physics of a welding defect generated and the resulting voltage signal, the experiments were observed by high speed imaging, followed by evaluation and modelling. A commercial system with three detector wavelength windows was studied for nine different industrial welding applications with distinct defects. We here present selected cases for which we try to generalise the findings and to draw conclusions for the applicability of commercial monitoring systems. For example humping was clearly detected by the signal, spatter and crater formation for overlapping Zn-coated steel surprisingly not. An interesting cause for welding defects is contamination of the joint prior to welding by oil, grease or detergents. The corresponding experiments showed unexpected signal responses, which can be partially explained by the analysis. We summarise that improved understanding of the signal causes facilitates to apply monitoring systems in a reliable manner.