Monitoring
Every signal here moves when the process changes, and the process changes far more often than the model does. A drift alarm is a reason to look, not a verdict on the model — and nothing on this page ever deploys anything by itself.
Open alerts
Each one names the action and the owner, because a red number nobody owns is a number people learn to scroll past
Nothing raised
Every monitored signal is inside its limits.
ST-01
within limits
Trigger: above 0.20 sustained for two hours
p = 0.014
Trigger: above 0.60 — the windows have become separable
Comparing 200 recent parts against 200 earlier ones. A plant configures a fixed reference window instead — usually the period the thresholds were fitted over.
ST-02
score distribution has shifted: PSI 0.261 against a 200-part reference
Trigger: above 0.20 sustained for two hours
p = 0.005
Trigger: above 0.60 — the windows have become separable
Comparing 200 recent parts against 200 earlier ones. A plant configures a fixed reference window instead — usually the period the thresholds were fitted over.
The monitoring table
Every signal the platform watches, its trigger, what it means someone should do, and who that someone is
| Signal | Measured by | Trigger | Action | Owner | Retrain? |
|---|---|---|---|---|---|
| Confidence distribution shift | Histogram of defect score against the reference window | PSI > 0.2 for 2 hours | Investigate; sample parts for review | Ml owner | candidate |
| Embedding drift | Embeddings of production images against the golden set | domain classifier AUC > 0.6 | Review; candidate retrain | Ml owner | candidate |
| Reject rate spc | p-chart per station and SKU | Western Electric rules | Alert the quality engineer; check the line first, the model second | Quality engineer | — |
| Hold rate | Held parts against review capacity | above the agreed review capacity for a shift | Review thresholds against capacity; check for a process change | Quality engineer | — |
| Camera lighting | Mean intensity, clipping percentage and blur metric per frame | outside the recipe limits | Raise a maintenance ticket: clean the lens, replace the light | Maintenance | — |
| Operator override rate | Overrides as a share of holds | above 20% for a shift | Review the labels and the threshold | Quality engineer | candidate |
| Confirmed escape | A defect confirmed downstream on a part that passed | any | Root-cause review; add to the golden set; candidate retrain | Quality owner | candidate |
| Deadline misses | Results arriving after the divert deadline | any sustained rate | Check image transfer and thermal throttling before the model | On call | — |
| Buffer fill | Evidence buffer fill | above 80% | Check connectivity and the upload backlog | Ops | — |
| Upload lag | Age of the oldest un-uploaded record | above 1 hour | Check the broker and the object store | Ops | — |
| Selftest failed | Golden-image self-test on the active bundle | any failure | Hold every part; roll back to the previous bundle slot | Ml owner | — |