Data Collection Strategies for Vision Models

The biggest bottleneck in deploying AI vision systems in manufacturing isn't the algorithm—it's the data. An AI model is only as good as the images it was trained on.

The Imbalanced Data Problem

In a standard wire processing facility, the defect rate might be 1 in 10,000 crimps. If you just set up a camera and wait, you will collect millions of images of perfect crimps and almost no images of defects. When trained on this data, the AI will simply learn to predict "Good" 100% of the time.

Intentional Defect Generation (IDG)

To build a robust dataset, engineers must intentionally manufacture bad parts. This is a controlled process:

• Missing Strands: Manually strip the wire and snip 10%, 20%, and 30% of the strands before crimping.
• High/Low Insulation: Intentionally misalign the strip blades to leave too much or too little insulation.
• Over/Under Crimp: Adjust the shut height on the press to apply too much or too little pressure, capturing images of the resulting terminal deformation.

Synthetic Data Generation

When physical defect generation is too costly or dangerous, manufacturers are turning to synthetic data. Using tools like Unreal Engine or Blender, teams create physically accurate 3D models of wire and terminals, then programmatically generate variations in lighting, background noise, and defect types to bulk up the training dataset.