MLCC yield is already high, so micro-cracks appear rarely and there are never enough defect samples to train a conventional classifier. DaoAI's APDT learns only good parts and treats deviations from normal as suspect, sidestepping the sample problem.
Multilayer ceramic capacitors (MLCC) ship in enormous volume and are tiny individually; micro-cracks on the end faces or body may be only microns wide and, at high yield, appear rarely. Conventional vision needs many defect samples to train, but cracks are inherently uncommon and samples can't be amassed; manual visual checks are limited by speed and consistency and are simply unworkable at thousands of parts per minute.
The plant adopted DaoAI APDT positive-sample learning, establishing a 'normal' criterion from just 1–20 good-part images with no need to pre-collect a large crack set—any fine texture deviating from the good distribution is flagged as suspect. Combined with DaoAI high-resolution imaging, the system resolves at micron scale and can spot hairline end-face cracks; each part is cleared in milliseconds, matching the takt of high-speed incoming inspection and pre-taping checks. When a new capacitor spec is introduced, changeover takes five minutes with zero code—no re-collecting samples, no rewriting the flow.
Why positive-sample learning
- Cracks are rare, defect samples hard to collect → learn good parts only, start on 1–20
- Micron-scale cracks → high-resolution imaging + AI detection of fine texture deviation
- Thousands of parts per minute → millisecond inference per part
- Frequent multi-spec switching → 5-minute zero-code changeover
It doesn't hoard crack photos; it just states clearly what 'good' looks like.
After go-live, the system reliably identifies micron-scale cracks while holding millisecond per-part inspection to meet line takt; because only good parts are needed to model, new-spec introduction is far faster, and the labor and consistency burden on visual-check stations eases in step.