AI Quality Control in Wire Harness Manufacturing
The traditional Automated Optical Inspection (AOI) paradox: set the tolerance high enough to catch micro-defects, and you flag 30% of good parts. Set it low, and a single wire strand slips out of the crimp zone.
The Limits of Rule-Based AOI
In standard crimp force monitoring (CFM) and traditional AOI, rules are rigid. A camera looks for a specific geometric bounding box. If the insulation crimp diameter exceeds 2.4mm ± 0.1mm, it flags a defect.
But real-world harnesses aren't rigid. Lighting changes. Grease smudges on the applicator. Copper strands reflect light differently depending on the angle. This leads to a massive false-reject rate (Type I errors), forcing human operators to manually verify "failed" parts, defeating the purpose of automation.
Common CNN Detection Classes
- Good Crimp (Nominal)
- Insulation in Core Crimp
- Strand Out (Flare)
- High/Low Crimp Height
- Missing Bellmouth
- Brush Length Out of Tolerance
Moving to Convolutional Neural Networks (CNNs)
Instead of hardcoding geometric rules, AI models (specifically CNNs) are trained on thousands of labeled images of both perfect and defective crimps. The model learns the features of a defect, such as the specific shadow cast by a rogue wire strand, rather than just measuring pixels.
Data Acquisition Challenges
To train a robust model, you need bad parts. In a Six Sigma facility, bad parts are rare. Manufacturers must intentionally induce defects—adjusting the press shut height, removing strands, altering strip length—just to capture training data for the AI.
ROI and Next Steps
Implementing AI vision typically costs between $15k-$40k per press for hardware (cameras, edge inference compute) and software licenses. The ROI comes not from firing operators, but from reducing the time spent manually reviewing false rejects and preventing catastrophic field failures (e.g., in automotive or aerospace harnesses).