Defect detection of gear parts in virtual manufacturing

Area of Science:

• Engineering
• Computer Science
• Materials Science

Background:

• Gears are crucial components in virtual manufacturing and digital twins.
• Acquiring images of gear tooth defects is challenging due to their non-convex geometry.

Purpose of the Study:

• To propose a deep learning network for effective gear defect detection.
• To address the challenges of acquiring and analyzing gear defect data.

Main Methods:

• A 3D gear dataset with 10,000 instances was created, classifying defects into fracture, pitting, glue, and wear.
• A novel Combinational Convolution Block was developed within the Gear-PCNet++ network for enhanced feature extraction.
• The network utilizes point cloud representations for defect analysis.

Main Results:

• The proposed Gear-PCNet++ method demonstrated superior recognition results for gear defects.
• The approach achieved higher efficiency and practicability compared to existing methods.
• Effective extraction of local gear information and complex topology was achieved.

Conclusions:

• Deep learning on point cloud data offers a viable solution for gear defect detection.
• The novel Combinational Convolution Block significantly improves defect identification.
• The developed method enhances the reliability of virtual manufacturing systems.