Geometrical defect detection on additive manufacturing parts with curvature feature and machine learning

The geometrical quality assessment for additive manufacturing (AM) is a great challenge because of the complexity of AM parts and low repeatability of AM processes. Existing defect detection algorithms with 3D data mainly use features comprised of point-to-point distance difference between the design and manufactured objects. This study introduced discrete mean curvature measure, a new curvature feature, to capture macro-level information beyond the distances and incorporated it into the training data for machine learning (ML) algorithms. Five ML models (Bagging of Trees, Gradient Boosting, Random Forest, Linear SVM, and K-Nearest Neighbors) were implemented and compared on both synthetic and experimental data. This new curvature feature significantly improves the defect detection performance and improves the F-measure accuracy to as high as 94% on experimental AM barrel samples. Among the five ML models, Random Forest yields the best performance. A comprehensive and graphical tuning process of two important parameters in this method, the Number of Points in Each Patch and Radius of Curvature Calculation, is developed and can be implemented later by other practitioners.