Definition
The fraction of anomaly alerts correctly attributed to their true source (instrument fault versus actual asset degradation) across all investigated alerts in the period.
The Measurement Gap
Predictive models cannot distinguish whether an anomaly signal originated from a failing machine or a failing instrument, and no traditional metric measures that distinction. F1, precision, recall, and AUC evaluate whether a prediction was correct but never ask whether the source was correctly identified. The confusion matrix has a hidden column nobody built: "right anomaly detection, wrong attribution." MTBF, OEE, and PM Compliance are silent on attribution. The failure mode is the sensor-driven false alarm that consumes investigation time and erodes technician trust: reliability engineers report wasting 30–50% of investigation effort chasing instrument ghosts. Sensor drift can also corrupt model training baselines, producing miscalibrated confidence on every downstream prediction.
Decision Supported:
When an anomaly fires, is the right next action a calibration work order on the instrument or a corrective intervention on the asset?
Formula
Plain language: For every anomaly alert in the period, classify it as sensor-driven or asset-driven (using cross-residual analysis on correlated sensor bundles), then compare that classification against what investigation actually found. The ratio of correct classifications to total investigations is the SADD score.
Notation:
```
SADD = (Correctly Attributed Alerts) / (Total Investigated Alerts)
Correctly Attributed = (Sensor alerts confirmed as sensor faults)
+ (Asset alerts confirmed as asset degradation)
```
Variables:
| Variable | Definition | Unit |
|---|---|---|
| Correctly Attributed Alerts | Alerts whose pre-investigation classification (sensor vs. asset) was confirmed by the investigation outcome | count |
| Total Investigated Alerts | All anomaly alerts that produced an investigation work order with a recorded outcome | count |
| Cross-residual | Difference between a sensor reading and what its correlated peer bundle predicts based on physical relationships | unit-relative |
| Correlated sensor bundle | Set of sensors monitoring physically related phenomena on the same asset (e.g., vibration, temperature, current) | grouping |
Proposed formula, pending validation pilots.