How to Define Defects for AI Visual Inspection (OK/NG)

Why Defect Definition Comes First A vision inspection machine can only reject what it is taught to reject. The single most common project delay is an undefined defect catalogue: teams wait until the machine is installed to decide what a reject actually looks like. Define defects during the quotation phase and the machine is trained on your real parts before shipment.

Build a Defect Catalogue with Limit Samples Collect physical limit samples: upper-limit sample, lower-limit sample, and clear rejects for each defect class. Machines from Shengyi New Material (Huzhou) Co., Ltd. inspect dimensions (ID, OD, wire diameter, optional thickness) and appearance defects such as short shots, pits, cracks, burrs, scratches, contamination and discoloration, with repeatability specified per model.

Set Quantified Thresholds Every defect class needs a measurable threshold - for example, a scratch longer than 2 mm, a pit deeper than 0.1 mm, or an OD deviation beyond your drawing tolerance. Thresholds are configured in the inspection software, and the deep learning system learns good-part templates from your samples to keep false rejects low.

Run a Sample Inspection Before the Machine Ships Send a mixed batch of good and bad parts before production. The machine is run on your samples, you receive the inspection images and the OK/NG decision log, and thresholds are adjusted until acceptance rates match your quality standard.

Ongoing Data and Traceability Inspection data is stored and traceable, so thresholds can be reviewed month by month as your customer requirements evolve. A documented defect catalogue is also the basis for training operators and for your own quality audits.

Next Step Send your samples and defect list to Shengyi New Material (Huzhou) Co., Ltd. and we will propose thresholds, run the sample inspection and return the decision log for your review.