LPBF-tailored forgetting-resistant incremental learning for defect recognition in open-scenario additive manufacturing

The growing complexity of defect detection in laser powder bed fusion (LPBF) necessitates the development of machine learning models that can adapt to evolving defect categories in open industrial scenarios. Whilst extant methodologies demonstrate considerable efficacy within static environments, they are prone to catastrophic forgetting when incrementally learning new classes, thus severely restricting their practical deployment. This work proposes a novel incremental learning framework, LPBF-Tailored Incremental Learning Framework (LPBF-Tailored ILF), tailored for continuous defect recognition task in open scenarios. First, this study proposes an LPBF-tailored class-balanced augmentation method that mitigates severe class imbalance while preserving the defect formation causality. Second, this study develops a contrastive knowledge transfer framework with memory replay that mitigates catastrophic forgetting, ensures deployment efficiency, and delivers robust defect classification. Third, this study designs a tripartite coordinated incremental learning loss orchestrating sample stability, feature plasticity, and boundary robustness through experience replay, contrast distillation, and separation regularization, to alleviate feature overlap in continual defect inspection. The quantitative and qualitative results on the custom-build dataset indicate that the proposed method exhibits superior single-phase incremental performance with 94.14 % incremental accuracy and 0.59 % forgetting rate. The deployability analysis demonstrates that this study provides a practical paradigm for sustainable defect recognition in real-world additive manufacturing environments.