Porosity Detection in Powder Bed Fusion Additive Manufacturing with Convolutional Neural Networks

Abstract

The sector of Additive Manufacturing is growing continuously in recent years, creating a wide range of applications such as medical devices and spacecraft parts. As the industry has high demands on the quality of these printed parts, a proper process monitoring is needed to ensure reliable parts while reducing costs. This approach focuses on the Powder Bed Fusion technology and adds an additional laser microphone to monitor the print process in-situ. Multiple defects can occur while printing, one of these is porosity. Since porosity has a strong influence on part stability, there must be no deviations here. To detect the level of porosity early on, a 2D Convolutional Neural Network was trained on in-situ audio recordings. Within an easy to use tweakable pipeline mel spectrograms were generated and fed into the neural network for classification of the porosity level. A F1-Score of 98,5% proves the concept that porosity defects of printed parts can indeed be effectively detected within production by neural networks fed with audio spectrograms. Porosity can thus be directly derived during the printing process itself, saving costs and material as a porous print can be stopped early and a x-ray after the print is done is not necessary anymore. This approach proves that integrated sensors in the printing process can deliver a huge benefit to the additive manufacturing in production.