Abstract
<title>Abstract</title> <p>Defect detection on machined parts is a critical task in industrial manufacturing, yet it remains challenging due to irregular data representations, geometric variability, and severe class imbalance between defective and non-defective regions. This work proposes a deep learning–based inspection framework for point cloud defect identification that addresses these challenges while maintaining low and predictable inference latency. The approach extends the PointNet architecture by introducing a deterministic canonical alignment preprocessing step based on singular value decomposition (SVD), eliminating the need for learned spatial transformation networks. In addition, a weighted categorical cross-entropy loss is employed to emphasize rare defect points during training. The proposed framework is validated through experimental evaluation on real 3D point cloud datasets acquired from robotic repair scenarios. Results demonstrate improved defect detection accuracy compared to baseline PointNet models, together with reduced inference time, supporting applicability in real-time industrial inspection pipelines.</p>