In-situ quality inspection based on coaxial melt pool images for laser powder bed fusion with depth graph network guided by prior knowledge

In-situ monitoring is crucial for enhancing process quality control in laser powder bed fusion (LPBF). Currently, data-driven approaches in LPBF in-situ quality monitoring have shown remarkable success. However, existing data-driven methods often lack integration with physical knowledge, leading to the opacity of decision-making processes. Research on LPBF knowledge-data mixed-driven modeling is still relatively scarce. To address this gap, this paper proposes a deep graph network method guided by prior knowledge (MK-DGNet) for in-situ quality inspection based on coaxial melt pool images. In the proposed method, prior knowledge is first extracted based on understanding of melt pool. Then, the fusion module is used to place images and knowledge vectors in the same dimensional space. Finally, a deep graph network architecture is elaborately established, taking graph-formatted data as input to learn deep-layer relationships between nodes and edges. The superiority of MK-DGNet is demonstrated using publicly available NIST datasets and self-built CMPQ dataset. Additionally, explainable artificial intelligence methods are employed to explain the basis of network decisions and the effectiveness of prior knowledge. This research provides new methods and perspectives for addressing quality issues in the LPBF process.