Polarity is a key safety item in automotive electronics, yet legacy AOI floods polarity checks with false calls, and operators spend their time clearing false reversals. This plant fed its accumulated false-call images to DaoAI for few-shot learning so the model learned true reversal from glare and color shift.
Automotive electronics has zero tolerance for reversed polarity; the orientation of diodes, tantalum caps, and electrolytic caps is a key safety characteristic. But polarity marks rely on silk and body color bands, and legacy AOI readily misreads silk glare, batch color shift, and slight rotation as reversal, so false calls at the polarity station run high for years. Operators wear themselves out clearing false reversals one by one, which dilutes vigilance toward true reversals—an unacceptable risk on automotive parts.
Rather than train a large model from scratch, the plant used the 164 historical polarity false-call images already accumulated on the line, building a targeted model with DaoAI few-shot learning—teaching the model precisely the hard cases that 'look reversed but are really glare or color shift.' Once live, inference runs under one second to match inline takt; polarity accuracy is about 99%, true reversals are caught reliably, and false reversals from glare and color shift are filtered out at the semantic layer.
Key practices
- Use the 164 historical false-call images already on the line—no mass labeling
- Few-shot learning targets the 'glare/color-shift pseudo-reversal' hard cases
- Under-1-second single inference fits inline takt
- Semantic false-call filtering separates true reversal from mark glare
On automotive parts, you can neither miss a reversal nor cry wolf every day.
After go-live, false-reversal alarms at the polarity station fell sharply and the manual effort to clear false calls dropped markedly, while polarity accuracy held near 99% and true reversals stayed reliably caught—meeting automotive-electronics control requirements for this critical characteristic.