Industrial AI · Test results
How well did the AI recognise the pattern?
See how a model was tested on recorded machine readings, including the cases it got wrong. The detailed method and limits are available below.
A learned classifier you can inspect
Can measured cycle features separate recordings labelled with pump leakage? This baseline learns coefficients from 1,653 training cycles in the public archive and retains every held-out prediction.
Same-rig random holdout, not plant validation. Neighbouring cycles can share operating conditions. The reported result does not establish performance on a new machine, future period or client plant.
Confusion matrix · threshold fixed at 0.5
| Source label ↓ / model → | No leakage | Leakage |
|---|---|---|
| No recorded leakage | 306True negatives | 0False positives |
| Recorded leakage | 15False negatives | 231True positives |
Positive means weak or severe pump leakage in the experiment annotations. These predictions use the entire 60-second cycle.
Ranking behaviour · ROC
AUC describes score ranking. It does not validate probability calibration or choose a production alarm threshold.
Errors remain visible
15 genuine held-out errorsAll 15 misses are weak-leakage-labelled cycles in this split. This observation does not establish a cause for the errors.
| Original row | Recorded class | Leakage score | Model decision | Inspect |
|---|---|---|---|---|
| 390 | Weak recorded leakage | 44.84% | No-leakage class | |
| 421 | Weak recorded leakage | 45.55% | No-leakage class | |
| 1146 | Weak recorded leakage | 43.85% | No-leakage class | |
| 1154 | Weak recorded leakage | 35.36% | No-leakage class | |
| 1156 | Weak recorded leakage | 22.83% | No-leakage class | |
| 1157 | Weak recorded leakage | 42.95% | No-leakage class | |
| 1736 | Weak recorded leakage | 46.94% | No-leakage class | |
| 1739 | Weak recorded leakage | 38.10% | No-leakage class | |
| 1741 | Weak recorded leakage | 41.55% | No-leakage class | |
| 1743 | Weak recorded leakage | 40.83% | No-leakage class | |
| 1747 | Weak recorded leakage | 41.64% | No-leakage class | |
| 1862 | Weak recorded leakage | 48.64% | No-leakage class | |
| 1871 | Weak recorded leakage | 43.38% | No-leakage class | |
| 1876 | Weak recorded leakage | 43.75% | No-leakage class | |
| 1879 | Weak recorded leakage | 43.98% | No-leakage class |
Inspect the review-workload assumptions
What would a review workflow cost?
Review every flagged cycle and keep a full routine check for every unflagged cycle. Account for extra correction work without assuming that a model can clear a machine as safe.
Your operating assumptions
Sample values to edit. These are not client measurements or industry averages.
Correction time and per-item costs
Every observed mistake is charged extra correction effort in this scenario. Search uses recorded misses; the additional cost of its chosen method is entered above.
5.6 hours of capacity released under these assumptions.
- Flags to review
- 418.5
- Routine checks retained
- 581.5
- Misclassifications to resolve
- 27.2
Includes ₹1,000 in extra operating costs. Capacity value is not cash saved or revenue.
Workload break-even: about 509 items / month if the same mix and timings hold.
How the workload is counted
Flags receive review. Every non-flagged cycle keeps the full routine check, including missed leakage labels. Misclassifications also add correction time. This assumes errors are identified during review; the review process itself has not been validated.
Applying this workflow to 552 source examples would require 231 review steps, 321 fallback steps and 15 extra correction steps. Corrections overlap those queues. Projected counts can be fractional and do not describe observed customer cases.
Labor value = minutes × assumed hourly team cost ÷ 60. Net difference subtracts fixed and per-item incremental operating cost. Setup and training costs are excluded unless entered in operating cost. Zero activity still incurs the assumed fixed cost; negative results remain visible.