Print defects take endless forms, and enumerating defect samples is nearly impossible. DaoAI APDT flips the approach—learn only good samples and treat any deviation as suspect—so it can go live fast even when samples are scarce.
A home-appliance plant prints brand logos, model silkscreens and safety marks on panels and housings. One line serves multiple brands and regional variants, so plates and text content change often. Print defects are highly divergent: double-print, broken-print, missing strokes, ink smearing, position drift, uneven density—many only a few tenths of a millimeter, easily missed by human eyes under fast takt.
Supervised methods demand ample samples for every defect class, but print defects are sporadic and infinitely varied—impossible to enumerate; and every plate change means rebuilding samples and models, making go-live slow and maintenance heavy. The plant needed an approach insensitive to sample count that could restore inspection capability quickly after each changeover.
DaoAI solution: APDT positive-sample learning defines the standard from 1–20 good samples
DaoAI uses APDT positive-sample learning, establishing a baseline from just 1–20 well-printed good samples; the system automatically learns the normal form of character strokes, logo contours and ink distribution. Any double-print, broken-print or missing stroke that deviates from the good sample is flagged as a defect—no defect samples required in advance.
- APDT goes live on just 1–20 good samples, no defect samples needed, rebuilding the baseline fast after a plate change
- Reliably catches double-print, broken-print, missing strokes, ink smearing and position drift
- Inspection throughput rises over 40% versus manual, with in-takt full inspection no longer reliant on human eyes
- Changeover is zero-code, so line operators can switch baselines themselves
No need to predict what a defect looks like—if it doesn't look like a good sample, it gets stopped.
After go-live the printing station achieves full inspection within takt, and high-frequency defects like double-print and missing strokes are reliably caught before shipment. Changeover shifted from engineer-dependent tuning to fast on-site switching, inspection throughput rose over 40% versus the prior manual process, and brand print consistency improved markedly.