Fifty signals are one incident, not fifty tickets.
A failing pump does not announce itself once. It arrives as a dozen symptoms across three systems, and only one of them is the cause.
A real failure arrives as a crowd.
One coolant distribution unit loses differential pressure. Within ninety seconds the rack inlet sensors downstream of it report high, the hall's return temperature climbs, a humidity alarm trips on the swing, and the DCIM environmental view lights up across two rows.
Every one of those signals is real. None of them is the problem. A system that treats each as its own event produces a queue of tickets and no understanding — and the queue is worst exactly when the incident is worst.
Correlation is not deduplication.
Grouping alarms by time window or asset tag collapses noise, but it cannot tell you what is at risk. That requires knowing what is connected to what, and what remains if this component is gone.
- Topology
- Which loops, CDUs, CRAHs and racks sit downstream of the failing component.
- Redundancy state
- Whether N+1 is intact, already degraded, or about to be consumed by this failure.
- Thermal margin
- How much time the remaining capacity buys before the load is affected.
- Operational risk
- What the affected load actually is, and what commitment covers it.
The decision is a record, not a notification.
A notification says something happened. A decision record says what was observed, what it implies, how confident that inference is, and what response is authorised — so a human can approve it or overrule it on the evidence.
Coolant flow restriction upstream of CDU-2E
- Observed
- CDU-2E Δp 0.8 → 0.3 bar over 90 s · RACK-14/16 inlet +6.2 K · hall return +2.1 K
- Inferred
- Coolant flow restriction upstream of CDU-2E. Rack symptoms are downstream, not independent faults.
- Confidence
- Medium — consistent with two prior events on this loop; no flow meter on the affected branch.
- At risk
- Loop 2E · 14 racks · N+1 already consumed by scheduled work on CDU-2F.
- Margin
- ≈ 11 minutes of thermal headroom at current load before inlet exceeds class limit.
- Response
- Qualified on-site engineer with CDU authorisation · MOP-114 attached · escort not required.
- Approval
- Requires named human approval before dispatch. The system does not act on this alone.
- Closes when
- Δp holds ≥ 0.7 bar for 30 min AND rack inlet returns to baseline AND the engineer confirms the cause found.
It reads the plant. It never drives it.
The connection to your BMS, DCIM and controls is read-only, and it can be run against a historian or inside a DMZ. It takes no control authority, cannot change a setpoint, and cannot dispatch anyone without a human approving the decision first.
Where this sits next to the tools that already do part of it.
Those categories are not failures — each does its job. The gap is that all three assume somebody has already worked out what the condition means. That step is the one nobody owns.
- Alert transport
- Delivers the signal to a person or a queue. Does not interpret it.
- Alarm management
- Suppresses, groups and prioritises alarms. Works on the signal, not the physical chain.
- Alarm-to-work-order
- Classifies an alarm and opens a ticket. The decision about what the condition means is still human.
- This layer
- Correlates through topology, computes what is at risk, and produces an authorised decision to approve.
How is alarm correlation different from alarm suppression?
Suppression reduces how many alarms reach a human, usually by time window, priority or asset tag. Correlation through topology works out that the fifty signals describe one physical condition with one cause, and which of them is the cause. Suppression makes the queue shorter; correlation makes it a single incident with an owner.
What does it need to read from our systems?
Alarms and point state from the BMS or BAS, assets and environmental history from DCIM, and equipment relationships to build the cooling chain. All read-only. In a design-partner engagement this can run against a historian or a replay of past alarms rather than a live connection.
Can it decide without a human?
No. It produces the decision and the evidence behind it; a person approves before anything is dispatched. The scope in which it may act at all is explicitly agreed in advance, and physical work always belongs to a human.