Physical structure and graph learning for manufacturing defect detection

Abstract

Manufacturing inspection combines measurements taken on irregular geometries with decisions about conformity, rework, and further examination. Graph neural networks can represent these measurements, but geometric connectivity, physical consistency, and defect evidence are different forms of information. This review develops a framework for separating them. It examines graph construction, anomaly reconstruction, explicit physical residuals, and the consequences of smoothing for localized faults. Analytical examples quantify signal attenuation, the effect of defect prevalence on alarm precision, and the distinction between node and part decisions. A proposed evaluation protocol links these calculations to manufacturing inspection through controlled topology changes, grouped validation, sensor uncertainty, and independently confirmed defect labels.