Edge AI wafer defect
This example is the fab inspection story. It is complete on its own. You do not need the plant example.
frame-generator on site-a publishes wafer frames. object-detector on edge-c runs wafer-defect. snapshot-service on ops-b keeps the latest inspection for an HTTP gallery. Subjects are vision.wafers.> and vision.inspections.>.
You can change agent.name if your Edgelet nodes use other names.
Container images are ghcr.io/datasance/pot-edge-patterns/<service>:1.2.0. The model repo is datasance/pot-edge-patterns/models/wafer-defect, tag 1.2.0.
Source files: deploy/steps/13a-model-wafer-defect.yaml and deploy/steps/13b-application-vision-inspection.yaml in pot-edge-patterns.
Flow
| Page | Reader does | Already in place |
|---|---|---|
| OCI registry | Registry row, type: oci, for GHCR | An Edgelet node you will call edge-c |
| Wafer-defect model | kind: Model in the catalog | That registry row |
| Vision user rules | Three NatsUserRule objects | Nothing from the plant |
| Vision account | One in-account NatsAccountRule | The user rules can exist before or with it. The Application must come after both. |
| Vision inspection | Application vision-inspection | Model catalog row, user rules, account rule. Nodes site-a, edge-c, ops-b (names are overridable). |
| Snapshot service | kind: Service for the gallery | The snapshot-service microservice is deployed |
| Verify | Model is on edge-c, inspections arrive, gallery opens | The service |
On vision inspection, spec.models names wafer-defect. Deploying that microservice makes the Controller attach the model to that Edgelet node and pull it. You do not run attach model, and you do not pull the file by hand.
attach model for a node with no workload is documented on AI Model Catalog. This example does not use that command.
Topology
Weights land at /models/wafer-defect/wafer-defect.onnx inside object-detector.
The plant floor is a separate example: Acme Smart Plant.