Neural network based image segmentation for spatter extraction during laser-based powder bed fusion processing

Abstract

In situ monitoring of spatter signatures is often employed to improve product quality during laser-based powder bed fusion (LPBF). This paper describes a novel neural network (NN) based image segmentation method for spatter extraction with a simple labeling process and high accuracy results. Use of a 290–1100 nm waveband high-speed camera allowed capturing images with more complete spatter signatures and a more complex background compared with previous LPBF studies. Conventional image segmentation approaches are inadequate to perform spatter extraction because of the complex background. The proposed NN-based image segmentation method split images into a block grid and segmented each block using a parallel convolutional neural network (CNN) and a thresholding neural network (TNN), which permitted extracting 80.48% of spatters in only 70 ms. Furthermore, the ability to extract spatters connected to a molten pool distinguishes the proposed NN-based image segmentation method from conventional image segmentation approaches.