MFGNet-Gear: A Synthetic 3D Dataset for Geometric Defect Detection in Gears

This mirror contains mesh files only (.ply format, ~11 GB).

For the full dataset including point clouds (.txt, ~30 GB), download from: Deep Blue Data — DOI: 10.7302/qrdj-n812

For code, documentation, and data generation scripts, see the GitHub repository.

Dataset Overview

Property	Value
Total parts	24,000
Gear designs	12 (T20–T40 series)
Quality classes	4 (G0, P0, W0, R0)
Parts per class	500
Mesh format	.ply (polygon mesh) — hosted here
Point cloud format	.txt (x,y,z comma-separated) — available on Deep Blue Data

Quality classes:

Label	Class	Description
G0	Good / nominal	No defect
P0	Pitting	Surface fatigue damage
W0	Tooth wear	Material loss due to friction
R0	Root breakage	Fracture at tooth root

Repository Structure

MFGNet-Gear/
└── mesh_ply/   ← PLY polygon mesh files, organized by gear design

File Naming Convention

T20ID10G0_00001.ply   → design T20ID10, good part, index 1
T30ID30R0_00412.ply   → design T30ID30, tooth root breakage, index 412

Citation

@article{mei2024deep,
  title={Deep learning of 3D point clouds for detecting geometric defects in gears},
  author={Mei, Ruo-Syuan and Conway, Christopher H and Bimrose, Miles V and King, William P and Shao, Chenhui},
  journal={Manufacturing Letters},
  volume={41},
  pages={1324--1333},
  year={2024},
  publisher={Elsevier}
}

A dataset descriptor paper is in preparation.