Low-rank adaptive transfer learning based for multi-label defect detection in laser powder bed fusion

Defects in the laser powder bed fusion (L-PBF) process have significantly hindered the broader application of this technology, and consistent quality assurance remains a critical challenge. To address this, real-time monitoring technologies are urgently required to effectively guide the production of high-quality parts. Although deep learning has advanced the intelligent development of powder bed defect detection methods, challenges persist in terms of model generalizability and robustness in complex environments. Additionally, the accurate labeling and recognition of tiny or overlapping defects remain difficult tasks. In this study, we propose a novel method for in-situ monitoring of L-PBF powder bed defects, integrating low-rank adaptive transfer learning with multi-label classification. This method offers a robust solution for in-situ monitoring under complex and variable conditions, achieving high recognition accuracy for composite powder bed defects while maintaining a low training cost. Our approach attains a testing exact match ratio of 93.28 % with significantly fewer training parameters (1.6956M), surpassing full fine-tuning methods. Furthermore, the proposed method demonstrates enhanced robustness in scenarios with limited training samples and complex conditions. When transferred to a selective laser sintering defect dataset, the method achieves a 99.32 % testing exact match ratio within just 10 epochs, illustrating its effectiveness in cross-task and cross-process transfer learning. The proposed method holds promise for efficient powder bed defect identification across different additive manufacturing processes.