Build the solution around the production requirement.
Define the inspection task
A statement such as “check quality” is too broad. The project should name each feature, the smallest acceptable defect, product presentation, colour and surface variation, cycle rate and false-reject tolerance. Representative good and bad samples are used to prove the method.
Controlled presentation and lighting
The robot can present components to a fixed camera or place them into an inspection fixture. Stable position, background and lighting reduce variation and allow simpler, more dependable image processing. Multiple views may be required for hidden or three-dimensional features.
Cavity tracking and reject routing
Multi-cavity parts can be kept in a known order from the EOAT through inspection and downstream placement. Results can identify recurring cavity patterns. Rejected product is routed to a secure location with confirmation that it did not rejoin good production.
AI vision and conventional tools
AI-based vision can help where natural variation makes fixed rules difficult, while conventional measurement and pattern tools remain strong for well-defined features. The choice follows evidence from samples; “AI” is not a substitute for an achievable inspection specification.
Specification checklist
| Defect definition | Feature, size, location, severity and acceptance boundary |
|---|---|
| Image conditions | Product position, background, lighting, lens, views and exposure |
| Data | Pass/fail, measurements, cavity identity, images and retention period |
| Cell response | Accept route, confirmed reject route, alarm and recovery logic |
Frequently asked questions
Can vision find flash and short shots?
Vision can detect many visible examples if the defect is large and consistent enough under controlled imaging. Trials with representative samples confirm capability.
Can the system track individual mould cavities?
Yes, if components remain in known positions or are otherwise identified through take-out and inspection.
Is AI vision always better?
No. AI methods can help with variable appearances, but conventional tools can be simpler and more transparent for stable measurements and features.