AI System for Routing Visual Manufacturing Defects for Review

The problem

When camera images, product codes and the defect catalogue are held in different sources, fields have to be matched by hand before a draft defect region and review class can be prepared.

The quality control specialist must compare every suggestion in the defect region and review class with its source record; missing or conflicting evidence must not take the place of a final decision.

If a reflection or new variant that resembles a defect is not separated for review, a sound part may be rejected or a defect may pass through; the record must not progress without a human decision.

What we set out to improve

• Time for the draft defect region and review class to reach the quality control specialist with its supporting evidence
• Outcomes of records separated for human review because a reflection or new variant resembles a defect
• Number of suggestions corrected after identifying the risk that a sound part could be rejected or a defect could pass through

How the system works

1. 01 Receive the camera image, product code and defect catalogue from permitted sources with the record identity
2. 02 Link the source, date and version details to a common work record
3. 03 Validate fields and units
4. 04 Prepare the defect region and review class with their supporting evidence
5. 05 Check the exception: a reflection or new variant may resemble a defect
6. 06 Have the quality control specialist accept, correct or reject the suggestion
7. 07 Write the approved decision and its source to the activity log

Methods we used

• A common, source-identified data schema for the camera image, product code and defect catalogue
• A decision structure that separates deterministic rules from AI suggestions
• A review screen showing the defect region and review class beside the raw evidence
• A mandatory human review queue when a reflection or new variant resembles a defect
• An event identity that prevents duplicate processing

Where people stay involved

Data and security

Camera images, product codes and the defect catalogue should be processed using only the fields required for this decision. Access to the source should be authorised separately from the quality control specialist role; personal data, commercial information and technical secrets should not be passed into the model context unless needed. Reading, suggestions and human decisions should be logged separately.

Who this suits

Frequently asked questions