Powder bed fusion of thermoplastic polymers is a powder based additive manufacturing process that allows for manufacturing individualized components with high geometric freedom. Despite achieving higher mechanical properties compared to other additive manufacturing processes, statistical variations in part properties and the occurrence of defects cannot be avoided systematically. In this paper, a novel method for the inline assessment of part porosity is proposed in order to detect and to compensate for inherent limitations in the reproducibility of manufactured parts. The proposed approach is based on monitoring the parameter-specific decay of the optical melt pool radiance during the melting process, influenced by a time dependency of optical scattering within the melt pool. The underlying methodology compromises the regression of the time-dependent optical melt pool properties, assessed in visible light using conventional camera technology, and the resulting part properties by means of artificial neural networks. By applying deep residual neural networks for correlating time-resolved optical process properties and the corresponding part porosity, an inline assessment of the spatially resolved part porosity can be achieved. The authors demonstrate the suitability of the proposed approach for the inline porosity assessment of varying part geometries, processing parameters, and material aging states, using Polyamide 12. Consequently, the approach represents a methodological foundation for novel monitoring solutions, the enhanced understanding of parameter-material interactions and the inline-development of novel material systems in powder bed fusion of polymers.
Keywords: deep learning; deep neural network; powder bed fusion; process monitoring; quality control; selective laser sintering.