Application of unsupervised learning methods based on video data for real-time anomaly detection in wire arc additive manufacturing

In the Wire Arc Additive Manufacturing (WAAM) process, ensuring the quality of components is of paramount importance. However, existing defect detection research is predominantly confined to laboratory environments, rendering it inadequate for addressing the practical demands of industrial production. Furthermore, these studies primarily depend on supervised learning, which requires extensive labeled data, while anomalous data are scarce in industrial settings. This scarcity further limits the applicability of supervised learning methodologies. To mitigate this issue, this paper introduces an unsupervised anomaly detection framework based on manufacturing videos captured by industrial cameras. This framework integrates a Vector Quantization Variational Convolutional Autoencoder (VQ-VCAE) with the Isolation Forest algorithm, leveraging the temporal characteristics of anomalies inherent in the additive manufacturing process to significantly enhance detection accuracy. In this study, the defects predominantly detected include spatter and holes. However, the framework is capable of detecting various types of shape deviations and geometric defects in real-world industrial applications. Compared to baseline methods, the proposed approach substantially improves both precision and recall, achieving an F1 score of 0.9307 on the test dataset. Additionally, this framework employs video datasets derived from actual industrial production processes, thereby ensuring its feasibility and effectiveness in real-world scenarios.