Define the inspection decision
Start with one product family, inspection station and decision: accept, reject or send a part for review. Document the defects that matter, their severity, visible evidence and the existing inspection method. A cosmetic mark and a missing safety-critical component require different acceptance criteria. Agree who resolves ambiguous examples and how their decisions become versioned labels.
Computer vision can support surface inspection, assembly checks and other visually observable tasks. It cannot establish a property that the camera and acquisition setup do not capture. Dimensional measurement needs appropriate calibration and validation against the required tolerance. Keep nonvisual tests where they provide necessary evidence.
Make image acquisition repeatable
Record camera position, illumination, exposure, trigger timing and part orientation. Collect samples across shifts, batches, tooling conditions and product variants. Include dirty lenses, glare, motion blur, partial views and absent parts so the system can recognize acquisition failures rather than confidently classify unusable images.
Link images to part and batch identifiers without storing unnecessary personal information. Record whether a label comes from visual review, a downstream test or another measurement. Keep correlated views of the same part together when splitting data, and reserve later batches for evaluation. Otherwise, repeated images can make the test look easier than the next production run.
Choose thresholds with the quality team
Evaluate missed defects and incorrect rejects separately, broken down by defect type and operating condition. Accuracy alone can hide poor performance on rare defects. Report the number of evaluated parts and defects alongside the result, including uncertainty when the sample is small. A controlled test with injected defects is useful evidence, but does not establish the natural production defect rate.
Agree an operating threshold based on the consequences of escape, rework and unnecessary rejection. Provide a review route for uncertain cases and unsupported variants. Compare the pilot with the existing inspection process on the same sample. Do not replace an acceptance criterion with an unsupported promise to catch every defect.
Test the complete line interaction
Measure the interval from capture trigger to available decision, including acquisition, preprocessing, inference, network transfer and controller interaction. Confirm the decision belongs to the correct part at the rejection station. Test dropped frames, delayed decisions, duplicate events and camera or network disconnection. Agree what the line does when the inspection result is missing.
Begin with observation mode: record proposed decisions while the existing process remains responsible for disposition. Review disagreements before enabling a controlled action. Changes to machinery or line controls need the plant’s engineering approval and applicable operating procedures. A model evaluation does not validate the complete machine control system.
Track production conditions and changes
Record the deployed model, label specification, supported variants and acquisition configuration. Monitor unusable-image frequency, review workload, reviewed escapes and false rejects. Investigate changes after a new material, supplier, tooling setting, lighting change or maintenance event.
New training data should be reviewed and tested against the held-out evaluation before release. Provide a rollback procedure and retain the previous inspection method as an agreed fallback. Assign ownership for camera maintenance, label disputes, model release and incident investigation.
Evaluate the business result
Compare inspection labor, review effort, rework, scrap and verified downstream escapes over a defined pilot period. Include equipment, installation, integration, compute and support costs. Separate a measured pilot result from a forecast for another line. Yield gains and payback depend on the baseline and defect mix; a generic percentage is not evidence for a particular factory.
NIST’s manufacturing monitoring workcell description illustrates evaluating product quality together with communications and human interactions. It provides context for a systems-level pilot, rather than a claimed TensorBlue performance result.
For broader maintenance and production monitoring, read the manufacturing AI guide. Explore computer vision development or discuss a scoped inspection pilot.