Self-Supervised Multilabel Melt Pool Anomaly Classification in Powder Bed Fusion Additive Manufacturing

Abstract

Laser powder bed fusion (LPBF) is a key technique in metal additive manufacturing (AM) that enables the fabrication of parts with complex geometries and enhanced mechanical properties. Despite its advantages, LPBF's susceptibility to defects remains a challenge, directly linked to melt pool instability. Real-time melt pool monitoring (MPM) systems offer immediate feedback on melt pool states, which is crucial for identifying potential defects during the LPBF process. Recently, machine learning (ML)-based approaches with these monitoring systems have been introduced for real-time process anomaly detection and classification. However, one major hurdle is the requirement for large volumes of labeled and balanced data for ML model development, which is necessary for accurate process outlier identification and categorization. To address the issues, this article introduces a self-supervised learning-based visual transformer (SiT) for multilabel classification of melt pool anomalies, such as abnormal sizes, irregular shapes, spatters, and plumes. Experiments reveal that the self-supervised approach, with an average F1 score of 0.979, surpasses the performance of the supervised approach, which has an average F1 score of 0.836 across all classification cases, particularly in imbalanced datasets. The SiT model efficiently classifies multiple anomalies simultaneously, without extensive manual labeling, addressing data imbalance and enhancing MPM systems' efficiency and reliability. This advancement marks a significant contribution to improving part quality in LPBF processes.