3D-ADCS: Untrained 3D Anomaly Detection for Complex Manufacturing Surfaces

Abstract Surface quality inspection of manufacturing parts through 3D point cloud data has increasingly gained focus in recent years. Accurate 3D anomaly detection on manufacturing parts is challenging due to their complex shapes, the scarcity of anomaly-free samples, and the irregular distribution of scanned points. This article proposes a novel untrained anomaly detection methodology integrated with domain knowledge, i.e., inherent manufacturing characteristics of real products, to precisely detect anomalies using 3D point cloud data for complex manufacturing surfaces. Specifically, the manufacturing parts such as many axis-symmetric products tend to be component-wise surfaces, i.e., consisting of basic and simple components, and the cut profiles of these parts exhibit similarity. On the basis of the domain knowledge, we propose to segment complex surfaces into simple components and model the similar profiles as low-rank representations, thus enabling the modeling of manufacturing surfaces. Finally, combined low-rank representations and assumed sparsity of anomaly, we introduce robust principal component analysis (RPCA) as a unified formula for surface anomaly detection. Extensive numerical experiments across various types of parts have shown that our method delivers promising results, surpassing existing benchmarks.