A graph neural network integrating physical prior knowledge for defect monitoring in laser powder bed fusion

Abstract

The complex physical interactions between the laser and the powder during the laser powder bed fusion (LPBF) process significantly affect the consistency and stability of the component quality. Existing online monitoring technologies predominantly employ Convolutional Neural Networks (CNNs) to achieve defect monitoring within a single time window, which struggle to capture the complex coupling relationships and synergies caused by spatial knowledge in additive manufacturing. This study presents a graph structure construction algorithm that integrates physical priors to reflect heat transfer effects, aiming to explicitly model spatial structural information and utilizing Graph Convolutional Networks (GCNs) to capture acoustic information in the adjacent spatial regions of the melt pool across the melt pool. The algorithm utilizes spatial prior knowledge to construct a graph structure that corresponds to the spatial relationships of real components. Furthermore, the graph structure is established utilizing two indicators that possess significant physical meanings: PatchSize and LinkMode. PatchSize refers to the quantity of melt channels and the length of the single melt channels incorporated within the graph structure, while LinkMode signifies the mode of heat transfer occurring between the melt pool and its surrounding area. Experimental results indicate that, in comparison to non-graph structures and traditional graph structures, the method enhances accuracy by an average of 3.56 % and 2.42 % on acoustic datasets with different porosity levels respectively. Finally, this study explores the impact of different physical knowledge on GCNs by changing the graph construction indicators, providing new solutions to improve the reproducibility and quality stability of LPBF technology.