Surface quality inspection is pivotal in manufacturing, where 3D point cloud data offers unparalleled geometric precision over traditional imaging. This dissertation tackles critical challenges in 3D industrial anomaly detection—data scarcity, geometric complexity, and multimodal heterogeneity—through three novel methodologies, balancing theoretical rigor and practical viability.

First, we propose a Bayesian Network-based Anomaly Detection (BNAD) framework enables untrained anomaly detection on a single sample using global geometric priors. By categorizing unstructured point clouds into nominal, anomalous, and outlier classes via variational EM inference, BNAD achieves robust detection of depression and oscillation mark anomalies on steel slabs on a single sample, eliminating reliance...[ Read more ]

Surface quality inspection is pivotal in manufacturing, where 3D point cloud data offers unparalleled geometric precision over traditional imaging. This dissertation tackles critical challenges in 3D industrial anomaly detection—data scarcity, geometric complexity, and multimodal heterogeneity—through three novel methodologies, balancing theoretical rigor and practical viability.

First, we propose a Bayesian Network-based Anomaly Detection (BNAD) framework enables untrained anomaly detection on a single sample using global geometric priors. By categorizing unstructured point clouds into nominal, anomalous, and outlier classes via variational EM inference, BNAD achieves robust detection of depression and oscillation mark anomalies on steel slabs on a single sample, eliminating reliance on labeled datasets.

Second, we propose a sparse learning paradigm with graph representation (PointSGRADE) framework for untrained anomaly detection on smooth free-form surfaces. It decomposes point clouds into reference geometries, sparse anomalies, and noise via penalized optimization with graph-based smoothness metrics. The Majorization-Minimization algorithm ensures computational efficiency.

Third, we propose Geometry-Guided Score Fusion (G²SF), an unsupervised multimodal anomaly detection framework that integrates RGB and 3D point cloud modalities through anisotropic metric learning. By replacing isotropic Euclidean metrics with direction-aware scaling factors, G²SF learns complex nominal patterns in a data-driven manner. (G²SF) reduces false positives significantly and enhances pixel-level localization accuracy on the MVTec-3D AD benchmark, outperforming state-of-the-art fusion methods.

Validated through various industrial case studies, our hierarchical paradigm - spanning from untrained 3D configurations to unsupervised multimodal environments - offers a scalable architecture for Industry 4.0 quality assurance. Moreover, we systematically evaluates methodological constraints and proposes targeted research trajectories to advance automated quality inspection systems.