• Proposed a three-level spatiotemporal model for defect prediction in laser AM. • Designed path graphs to model intra-, inter-track, and inter-layer dynamics. • Employed multi-layer GAT to capture multilevel dependencies in image sequences. • Performed comprehensive evaluation on models, graphs, and temporal levels. • Demonstrated accuracy, robustness, and interpretability of the method. Online monitoring is crucial for improving the quality of laser additive manufacturing (AM). Owing to the highly nonlinear and strongly coupled multiphysics nature of laser AM, developing complete and accurate physics-based models for in-situ process characterization remains a significant challenge. Consequently, data-driven methods are increasingly being adopted in laser AM monitoring owing to their minimal reliance on prior physical knowledge. However, current data-driven methods primarily focus on the instantaneous spatial characteristics of process state information, whereas the spatiotemporal features embedded in the melt-pool dynamic evolution remain insufficiently considered, potentially limiting the ability to fully capture defect-related patterns and challenging accurate localized defect prediction. To address this problem, we propose a three-level hierarchical spatiotemporal deep graph network (TLHS-DGNet) that explicitly models the dynamic dependencies within intra-, inter-track, and inter-layer melt-pool image sequences. In the proposed network, real-time coaxial melt-pool images are acquired during the laser AM process, and temporally ordered path graphs are constructed at each hierarchical level to encode sequential and topological relationships. Additionally, spatial representations are extracted through convolutional operations, whereas attention-enhanced graph convolutions capture the multilevel spatiotemporal correlations that are critical for defect genesis. Comprehensive experiments on CT-labeled datasets demonstrated that TLHS-DGNet consistently outperforms three representative model categories: spatiotemporal graph networks with varying temporal hierarchies, alternative graph-based methods employing diverse graph construction methods, and graph convolutional architectures. The findings of this study establish an effective framework for localized defect prediction in laser AM, advancing the field toward intelligent, data-driven process control and quality assurance.
Niu et al. (Sun,) studied this question.