How many ways can a capsule go wrong? Dents, deformation, powder leaks, double caps, color shift, blurred print, and so on, nearly a hundred classes; and a human staring all day will eventually tire into a miss.
Hard and soft capsules present a wide variety of appearance defects, and a solid-dose plant catalogued as many as 80 defect classes, spanning dented shells, depressions, bubbles, deformation, powder leaks, double caps, length variation, color shift, and print defects. For years this step relied heavily on manual visual inspection.
The problems with manual inspection are systemic: too many defect classes, hard-to-unify standards, operators tiring quickly under prolonged staring, escapes and false calls fluctuating by shift, and no traceable record of verdicts. To make automation cover so many defect classes and follow frequent product changeovers, conventional custom-vision projects are too long, too costly, and too rigid to adapt.
Handing Training Back to the Operator
The core advantage of DaoAI AI-AOI is zero-code self-training: line operators, without programming or calling in engineers, can label samples, train, and iterate models through a graphical interface, covering all 80 capsule defect classes in one system. Paired with 5-minute zero-code changeover, a product-spec switch is completed on the spot by the operator, with no line stoppage waiting on development.
- A single AI-AOI system covers 80 capsule appearance defect classes
- Operator zero-code self-training: graphical labeling, training, and iteration in one place
- 5-minute zero-code changeover, comfortable even for high-mix small-batch
- Fully recorded verdicts, replacing manual inspection with traceability
Truly sustainable automation hands the power to train and to change over back to the operators who know the product best.
After go-live, the plant upgraded capsule appearance inspection from manual to full AI-AOI inspection, unified coverage across 80 defect classes, eradicated the fatigue-driven escape problem, improved detection consistency and capacity together, and shifted operators from repetitive staring to model maintenance and exception handling.