Two-Stage joint distillation strategy for unsupervised anomaly detection in wire arc additive manufacturing

In wire arc additive manufacturing (WAAM), real-time defect detection is essential for quality assurance, yet the process is hindered by the scarcity of anomaly samples and the inherent unpredictability of anomaly types. To address the above challenges, this paper proposes an online unsupervised anomaly detection model based on knowledge distillation. A two-stage joint distillation strategy is developed to address the generalization limitations of symmetric teacher-student architectures. The designed model contains both forward distillation and reverse distillation stages, which improves the accuracy of the model in recognizing anomalies. In addition, before verifying the effect of the model, this paper produces a WAAM defect dataset using a welding anomaly detection system. The experimental results show that the average AUROC value for image-level anomaly detection can reach 93.3%. The developed method has practical application prospects in industrial anomaly detection and is of great significance for anomaly feedback in the additive manufacturing process.