In-situ monitoring and detection of lack of fusion defects in laser powder bed fusion complex components using melt pool images

Lack-of-fusion (LoF) defects are among the most critical in laser powder bed fusion (LPBF), as they can substantially degrade the structural integrity and service performance of fabricated components. Although data-driven defect detection based on in-situ monitoring signals has advanced rapidly in recent years, most studies have focused on simple geometries. In contrast, complex components remain insufficiently investigated due to geometry- and scan-strategy-induced variations in melt-pool signals. To address this challenge, this study proposes a detection and evaluation framework for LoF defects in complex LPBF components. LoF defects were intentionally induced in a grid component by locally reducing the laser power, guided by the LoF formation mechanism. Scan vector length was introduced as a key geometric-process variable to characterize geometry- and scan-strategy-induced variations in melt-pool signals, revealing a significant negative correlation with mean melt-pool intensity (Spearman’s correlation coefficient: −0.6536). A Transformer-based unsupervised time-series anomaly detection model was then developed and integrated with intra-layer and inter-layer filtering strategies. These strategies reduced false positives and recovered missed defect regions, increasing the F1 score from 0.549 to 0.900 and the intersection over union (IoU) from 0.378 to 0.818. The low coefficients of variation across different defect-inducing laser powers, 0.95% for F1 score and 1.8% for IoU, further indicate robust performance under varying LoF severity levels. These results demonstrate the effectiveness of incorporating geometry-related scan information and hierarchical filtering for reliable in-situ LoF detection in complex LPBF components.