Definition
Percentage of multi-alarm process upsets where the AI's alarm situation grouping was complete on post-event review (all related alarms grouped, no unrelated alarms included) with under-grouping and over-grouping rates reported separately because under-grouping has higher process-safety consequences.
The Measurement Gap
During a multi-faceted upset, the operator needs a coherent picture: which alarms belong to the same situation, which to a separate one, which are unrelated noise. AI alarm grouping is designed to produce that picture, and when it works it cuts response time substantially. When it does not work, the operator sees fragmented presentations of what should be a single situation, has to mentally reconstruct the relationships during the highest-stress moments of the event, and may delay or misdirect response. Alarm flood metrics, alarm response time, and alarm grouping volume all measure activity rather than fitness: they tell you the grouping system is running, not whether the groupings are right. Post-event review forms typically ask whether response was timely; few ask whether the grouping presented to operators was complete. Until SGC is in place, the AI grouping is governed by whether anyone notices the wrong picture, after the event, in a hot wash.
Decision Supported:
Is the AI alarm grouping giving operators a unified picture of developing situations, or is it presenting fragmented views that force operators to mentally reconstruct the situation during time-critical events?
Formula
Plain language: After each multi-alarm event, the operations or process safety team scores whether the AI's grouping included all related alarms and excluded unrelated ones. Aggregate across events for the SGC percentage. Track under-grouping separately from over-grouping.
Notation:
SGC = ( Correctly_grouped_situations / Total_assessed_situations ) × 100
Under-grouping rate = ( Situations_with_related_alarms_omitted / Total_assessed_situations ) × 100
Over-grouping rate = ( Situations_with_unrelated_alarms_included / Total_assessed_situations ) × 100
Variables:
| Variable | Definition | Unit |
|---|---|---|
| Correctly_grouped_situations | Multi-alarm events where post-event review confirms all related alarms in the group and no unrelated alarms included | count |
| Total_assessed_situations | All multi-alarm events reviewed post-event | count |
| Assessor | Operations or process safety team during post-event review using a structured grouping-quality checklist | n/a |
Track novel-upset SGC separately from familiar-upset SGC: the AI can only group based on prior co-occurrence patterns.
Proposed formula, pending validation pilots.