Quality Inspection Automation

## Capabilities - Dimensional: ±0.02mm repeatability with good calibration - Surface defects: scratches, contamination, missing material, color shift - OCR: lot code, production date - Assembly presence: missing/wrong part, orientation ## Real numbers On a connector customer's line, defect catch rate is >99.2% with a 0.8% false-reject rate. False-reject is the number that matters — pushing catch rate too high starts rejecting good parts and adds re-check labour. We set the "miss 0.5% vs over-reject 2%" trade-off with the customer; there is no universal answer, it depends on downstream cost. ## Lighting beats algorithms About 80% of failed vision projects are a lighting problem, not an algorithm problem. We usually spend more time on the illumination design than on the detection logic. With stable imaging, most defects separate with classical methods — deep learning is not always needed. ## When we use AI For broken characters or micro-defects on complex textured backgrounds, classical thresholds are hard to tune; only then do we bring in a deep-learning model, and it needs ≥300 real defect samples from the customer. Accuracy is not guaranteed with fewer. Named references available on request. ## Technical Specifications | Item | Range | |---|---| | Inspection speed | up to 1,200 ppm | | Vision | 2D/3D cameras, up to 12 MP; deep-learning defect models | | Checks | Fill level, cap/seal, label, code, foreign matter | | Reject accuracy | 100% verified reject with confirmation sensor | | Data | SPC trends, image archiving, MES upload | ## Typical Line Configuration Vision stations (cap/label/level), X-ray or metal detector, checkweigher, reject verification, central quality dashboard. ## Case Study A bottled-water plant cut customer complaints 73% after deploying cap+label+level vision at 900 BPM with image traceability. ## FAQ **Q: When is deep learning better than rule-based vision?** A: For variable, texture-like defects (scratches, contamination, weld pools) with sufficient sample images; rule-based remains best for measurements and presence checks. **Q: What false-reject rate is typical?** A: Well-tuned systems run below 0.3% false rejects; we tune on your real production samples during FAT and again at SAT. **Q: Can inspection images be stored for traceability?** A: Yes - NG images always, and OK images on a rolling window, tagged with batch and timestamp, exportable to your QMS.