Deep Learning-Based 3D Reconstruction for Defect Detection in Shipbuilding Sub-Assemblies.

Area of Science:

• Industrial manufacturing
• Computer vision
• Machine learning

Background:

• Overshooting defects in shipbuilding subassemblies compromise structural integrity and safety.
• Accurate defect detection is crucial for quality control in industrial settings.

Purpose of the Study:

• To develop and evaluate unsupervised learning methods for automatic detection of overshooting defects in shipbuilding subassemblies.
• To compare the performance of four state-of-the-art autoencoder architectures for defect identification.

Main Methods:

• Utilized reconstruction-based unsupervised learning on 3D point clouds.
• Implemented and compared Variational Autoencoder (VAE), FoldingNet, Dynamic Graph CNN (DGCNN) autoencoder, and PointNet++ autoencoder architectures.
• Employed Isolation Forest on reconstruction errors for anomaly detection.

Main Results:

• Reconstruction-based anomaly detection on 3D point clouds is a viable strategy for industrial defect identification.
• The study highlights the importance of selecting architectures balancing performance, geometric stability, and computational cost.
• Detection performance was analyzed concerning the contamination parameter.

Conclusions:

• Unsupervised learning offers a robust approach to identifying overshooting defects in complex industrial components.
• The choice of autoencoder architecture significantly impacts the effectiveness and efficiency of defect detection.
• This methodology supports enhanced quality control and safety in shipbuilding.