Definition
Percentage of AI-suppressed "nuisance" alarms that were followed by a process upset on the same unit within 72 hours: detecting whether the alarm rationalization algorithm is removing early-warning precursor signals along with genuine nuisance alarms.
The Measurement Gap
Alarm rationalization is a defensible response to alarm flood: operators cannot manage thousands of alarms per shift, and AI suppression of repeat nuisance signals is a measurable workload improvement. What the alarm management metrics do not measure is whether the suppression is removing precursors: the leading edge of upsets that look like noise until they are not. Alarm rate, alarm flood frequency, and operator response time all improve when nuisance alarms are suppressed. Time-in-flood goes down. Alarm-per-operator-per-hour drops. The metric set rewards aggressive rationalization. None of these metrics asks whether the suppressed signals included the early indicators that experienced operators recognize as warning patterns. A precursor that gets suppressed because it correlated with prior false alarms is silent until the upset it would have warned about arrives. The plant cannot see it from the alarm dashboard, and the alarm management program cannot diagnose itself.
Decision Supported:
Is the AI alarm rationalization program improving operator workload by removing noise, or degrading the operator's ability to detect developing upsets by suppressing the early signals that precede them?
Formula
Plain language: Of all the alarms the AI suppressed in the period, count how many were followed by an upset on the same unit within 72 hours. Divide by total suppressed alarms, multiply by one hundred. Track both the percentage and the absolute count.
Notation:
PSSR = ( Suppressed_alarms_followed_by_upset / Total_suppressed_alarms ) × 100
Variables:
| Variable | Definition | Unit |
|---|---|---|
| Suppressed_alarms_followed_by_upset | AI-suppressed alarms where a process upset occurred on the same process unit within 72 hours | count |
| Total_suppressed_alarms | All alarms suppressed by the AI rationalization system in the measurement period | count |
| Upset | Pre-specified definition: safety system activation, emergency shutdown, manual operator intervention for abnormal condition, or production loss exceeding a defined threshold | event |
Report PSSR percentage alongside absolute counts. A 2% PSSR with 50,000 suppressed alarms per month is 1,000 potential precursors per month.
Proposed formula, pending validation pilots.