On the inspection station of sterile injectables, the real challenge was never to see, but to tell apart: a drifting glass fragment and a rising air bubble look almost identical to conventional machine vision.
Sterile injectables enter the bloodstream directly, and pharmacopeias enforce near-zero tolerance for visible particles. A sterile injectables plant ran 100% inspection on conventional light-inspection machines, but was trapped by a fundamental contradiction: to never miss a glass fragment, fiber, or metal fleck, sensitivity was pushed to the extreme, and as a result countless normal bubbles and liquid ripples were misjudged as particles, with good units ruthlessly rejected.
The line inspects hundreds of thousands of units per day, and false-reject losses plus re-check labor stayed stubbornly high. Worse, genuine particle-defect samples were extremely scarce: even a mature line could barely gather a few dozen real particle samples in a month, so the conventional supervised approach of feeding massive defect sets simply did not work.
The DaoAI Approach: Re-inspection + Few-shot
We deployed the DaoAI AI-AOI vision system as a deep-learning re-inspection stage downstream of the existing light-inspection machines: legacy equipment performs high-sensitivity first-pass screening, and every vial flagged as abnormal is re-judged by AI-AOI, which uses deep learning to separate the motion trajectory and morphology of rising bubbles from drifting particles. To address scarce defects, APDT positive-sample/few-shot learning models the task from only a handful of real particles.
- High-sensitivity legacy screening plus AI-AOI deep-learning re-inspection, a two-layer gate
- APDT few-shot modeling: go live with dozens of real particle samples, no massive defect set required
- Bubble-versus-particle discrimination by motion and morphology, suppressing false rejects at the root
- Full data trail: inspection parameters and verdicts are GMP-auditable and traceable
Telling particles from bubbles takes not a brighter lamp, but a model that better understands motion and morphology.
Three months after go-live, the plant's real particle catch rose by about 70% over the original light inspection, false rejects fell by about 60%, and re-check labor was nearly halved. The image, model version, and threshold for every verdict are fully archived, passing GMP validation audits smoothly.