Industrial Vision Agent
MODELS · VISION · X-RAY · DEVICE–EDGE
Built on the industrial vision model series and X-ray inspection model, orchestrated by crusbro agent-OS for capture, decisions, review, alerts and QMS/MES writeback. Misses first. Evidence retained. Metrics signable.
Product structure
Domain models are independently acceptable. The station-level vision agent owns recipes, gray-zone handling, dual-sign and system writeback.
| Layer | What | Role |
|---|---|---|
| Runtime | crusbro agent-OS | Permissions, audit, versions, human gates |
| Models | Vision series · X-ray | Detect, segment, measure, film reading |
| Agent | This page | Station loop, review, writeback, iteration |
Industrial vision model series
Built for high reliability, scarce defects and strong traceability. Select by station; share versioning and acceptance rules.
| Line | Typical tasks | Method focus | Acceptance focus |
|---|---|---|---|
| Surface defects | Micro-cracks, coating, tool marks, edge chips | Self-supervised pretrain + segmentation + anomaly fallback | Recall first; FN thresholds per defect class |
| Assembly checks | Missing, wrong part, offset, damage | Detection + multi-view fusion + CAD geometry check | Missing/wrong as red lines; coverage and cycle time |
| Internal NDT | Delamination, porosity, inclusions | Ultrasound / thermography multimodal + segment / anomaly | Agreement vs metallography or CT |
| Connector / cable | Pin retreat, tilt, sheath damage, shield break | Metric learning + few-shot + defect synthesis | Model migration speed; cycle time; first-pass yield |
| Precision metrology | Size / GD&T / profile | 3D point cloud + registration + auto GD&T | Accuracy vs CMM and cycle-time gain |
X-ray inspection model
Digital radiography (DR) for welds, castings and structures. AI first pass + certified dual-sign sampling. Reports are archive-ready.
| Item | Detail |
|---|---|
| Defects | Porosity, slag, lack of fusion / penetration, cracks; graded by size and density |
| Stack | DR enhance → detect / segment → severity grade → PDF/A report |
| Compliance | Dual-sign: AI read + certified reviewer sampling; escalate on disagreement |
| Explainability | Boxes / masks and attention maps archived with the film |
| Acceptance | Agreement with certified readers, FN on critical defects, cycle time, report completeness |
Agent capabilities
- Select camera / lighting / model / thresholds per product and station
- Gray zone → multi-angle re-inspect; still uncertain → human confirm
- Consecutive anomalies → stop-line / changeover prompts
- Write decisions, images and model versions to QMS / MES
- Return corrected samples for retraining and edge hot-update
Constraints
- Misses first: false positives can be reviewed; FN is a red line by class
- Scarce defects: good-part pretrain, anomaly fallback, synthetic augmentation
- Bound evidence: image tied to S/N / lot / station
- Human co-sign: dual-sign retained on high-risk ops
Station protocol / acceptance
Sign per station. Do not replace with a vague accuracy claim.
| Dimension | Example |
|---|---|
| FN / FP | Per-class thresholds; red-line items 100% human review |
| Cycle time | Load + imaging + inference + unload; inference latency agreed separately |
| Agreement | X-ray scenes vs certified readers |
| Changeover | New model / defect adaptation window by scenario |
| Traceability | Image + params + model version + disposition replayable |
Deployment
| Item | Guidance |
|---|---|
| Sensing | Industrial camera / line scan / DR / ultrasound or thermography by model line |
| Edge | IPC or inference box; low-latency on the line |
| Center | Training, version registry, drift monitoring, canary release |
| Systems | QMS / MES / PLC / SPC; alerts and stop-line commands configurable |