| Abstract Scope |
Visual monitoring of Wire Arc Additive Manufacturing (WAAM) using deep learning presents an effective solution for real-time anomaly detection by leveraging the continuous stream of image data produced during fabrication. A major challenge in this domain is label inaccuracy. Labels are assigned post-process through segment-level measurements of the workpiece, resulting in identical labels for groups of consecutive image frames. Since anomalies may only occur within a fraction of a measured segment, the limited measurement granularity often leads to inaccurate annotations and contaminates the anomaly class with normal samples, ultimately reducing model performance. While various deep learning-based methods have been explored for anomaly detection in WAAM, few have addressed this specific challenge of label inaccuracy. To mitigate this issue, we propose a self-correcting framework that iteratively refines inaccurate labels and learns compact latent representations. Our approach alternates between two key processes: unsupervised feature learning using a β-TC Variational Autoencoder (β-TC VAE), which captures meaningful latent representations from the dataset, and probabilistic label refinement using a Gaussian Mixture Model (GMM), which models the distribution of normal features to reassign likely-normal samples within the anomaly class. This iterative cycle jointly improves label accuracy and feature discriminability. Experimental results demonstrate that the proposed approach outperforms both fully supervised and unsupervised baselines, particularly in scenarios with severe label inaccuracy. By reducing label contamination and enhancing representation learning, the approach significantly improves the robustness and accuracy of real-time anomaly detection in WAAM. |