AI fault labelling
A walkthrough of the system I built to label a manufacturer’s free-text fault records, from expert labelling to human review of new predictions.
Demo · All data is made up and nothing is sent to a server.
The demo simulates embedding and model training. A simple keyword algorithm produces predictions and measures accuracy on sample records.
FAULT CLASSIFICATION
SHIFT B · 14:32
TIMEOPERATOR NOTELABEL
06:12sensor not reading st4, reset done-
06:40conv. belt jammed, op stepped in, 10 min stopMECHANICAL
07:05pneumatic hose burst, losing air-
07:31material arrived with burrs, rejected-
08:02cut size drifting by 0.4mmOTHER
08:47panel fuse blew, whole line stopped-
09:15box feeder blocked at the inlet-
09:58barcode scanner missed, 3 items passedOTHER
10:21robot offset drifted, part touching edge-
10:44oil leak under the press-
11:30motor drive fault, breaker cycled-
12:06raw material batch 2231 wrong thickness-
+ 12,468
| OPERATOR NOTE | Expert label |
|---|---|
| 06:12 · LINE-2 · ST-04sensor not reading st4, reset done[0.31 -0.58 0.12 …] | Sensor fault |
| 07:05 · LINE-3 · ST-06pneumatic hose burst, losing air[-0.26 0.74 0.48 …] | Hydraulic / air leak |
| 07:31 · LINE-2 · ST-01material arrived with burrs, rejected[0.55 0.19 -0.64 …] | Material quality |
| 08:47 · LINE-3 · ST-03panel fuse blew, whole line stopped[-0.49 0.63 -0.17 …] | Electrical |
| 09:15 · LINE-2 · ST-05box feeder blocked at the inlet[0.68 -0.39 0.43 …] | Mechanical jam |
| 10:21 · LINE-3 · ST-02robot offset drifted, part touching edge[0.24 0.51 -0.32 …] | Setup / calibration |
Incoming fault record
Model processing
Encoding the note
Assigned category
Records routed by category
- Mechanical jam0
- Sensor fault0
- Electrical0
- Material quality0
- Setup / calibration0
- Hydraulic / air leak0
Queue 4
New record
AI suggestion
The model is not sure, pick by hand.
Most similar past records
1 · 2 · 3 to pick, Enter to approve
Queue is empty. All sample records have been reviewed.
What to try
- Watch the expert labels appear and see the notes turn into numeric vectors through embedding.
- Follow the classifier as it trains on labelled examples and routes new records to categories.
- Track how many predictions match the expert labels in the classification scene.
- At the review desk, pick a suggestion with 1, 2, 3 and approve with Enter, or correct it with another category.