Definition
Proportion of ML-suppressed alarms that occurred on assets which subsequently failed within the detection window. Every suppressed alarm that preceded a real failure was an early warning the noise filter deleted.
The Measurement Gap
SCADA alarm management in renewable energy operations has a well-documented problem: alarm flooding. A single weather event can generate thousands of alarms across a fleet, burying genuine failure signals in noise. Machine learning alarm filters are the modern solution, classifying each alarm as "signal" (pass to operator) or "noise" (suppress). The filters work: operator workload drops, response times improve, and the control room becomes manageable. But there is a measurement gap hiding inside this success.
The gap is that no one measures what gets lost in the suppression. When the ML filter achieves an 85% reduction in alarm volume, the control room celebrates. But that 85% includes both genuine noise (redundant alarms, transient spikes, environmental nuisance triggers) and potentially genuine early warning signals that the ML classified as noise because they did not match its training data's pattern of "signal." A vibration alarm that fires once and then resolves might be noise, or it might be the first symptom of a bearing that will fail in 30 days. The ML filter discards it. If the bearing fails, nobody connects the suppressed alarm to the failure because the suppression decision and the failure event live in different data systems.
ASSLR measures this hidden cost by retrospectively checking whether suppressed alarms preceded actual failures. It does not claim every coincidence is causation, but it flags patterns that deserve investigation and ensures the alarm filter's noise reduction is not achieved by discarding early warning signals that operators in the legacy paradigm would have investigated.
The decision ASSLR supports is whether the ML alarm filter is genuinely improving signal quality or achieving a better noise ratio by discarding signals that would have enabled earlier intervention.
Formula
Of all alarms the ML classified as noise and suppressed, how many occurred on assets that subsequently experienced a failure event within N days? If a suppressed alarm was an early indicator of a failure that later occurred, the filter discarded a signal.
Formula (Notation):
ASSLR = Suppressed_Alarms_Preceding_Failure / Total_Suppressed_Alarms times 100
Variable Definitions:
| Variable | Definition |
|---|---|
| Suppressed_Alarms_Preceding_Failure | count of ML-suppressed alarms on assets that experienced a failure event (unplanned work order or forced outage) within N days of the alarm |
| Total_Suppressed_Alarms | all alarms classified as noise and suppressed by the ML filter during the measurement period |
| N | trailing window. Recommended starting values: 30 days for wind turbines, 14 days for inverters, 90 days for BESS |
Edge Cases and Measurement Boundaries:
Not every suppressed alarm before a failure was a precursor. Some are genuinely noise on an asset that happened to fail for unrelated reasons. Track ASSLR by alarm type to distinguish systematic signal loss (a specific alarm type consistently appears before failures) from coincidental correlation. Normalize by asset alarm rate to avoid biasing ASSLR toward assets that generate high alarm volumes. Require a minimum sample of 50 suppressed alarms per alarm type per quarter before computing type-level ASSLR.
Proposed formula, pending validation pilots.