Industries / Industrial equipment and services / IoT and industrial automation
Alarms
Alarm management decides what sounds an alarm, how urgent each alarm is and what the operator should do about it, so that warnings stay few enough to act on.
4
Milestones achieved, from 1994 to 2014.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1994
The defining alarm flood
On 24 July 1994 a severe electrical storm caused disturbances across Texaco's Milford Haven refinery in Wales, and about five hours later an explosion injured 26 people on site. The UK Health and Safety Executive found alarms arriving "at the rate of one every two to three seconds"; in the last 10.7 minutes the two operators had to recognise, acknowledge and act on 275 alarms. Of about 2,040 alarms configured, 87% were high priority, and there was "no philosophy for determining what priority an alarm should have."
LessonThe HSE recommended that alarms be limited to the number an operator can monitor, and that plant safety not rely on an operator's response to a control-system alarm.
Credit
UK Health and Safety Executive (investigation)
1999
A guide to alarm rates
EEMUA, a UK association of engineering equipment users, first published its guide to alarm systems, Publication 191, in 1999, and calls it the "globally accepted and leading guide." The UK HSE, which contributed to it, summarizes its benchmarks: no more than one alarm every ten minutes on average in normal operation, no more than ten in the first ten minutes after a major upset, and priorities split about 5% high, 15% medium and 80% low. The fourth edition came out in 2024.
Credit
EEMUA, with the UK HSE
2009
An alarm management standard
ISA's Standards and Practices Board approved ISA-18.2, "Management of Alarm Systems for the Process Industries," in June 2009. Its co-chair Donald Dunn of Aramco Services said, "The committee was motivated by the common experience of incidents when the alarm system failed to prevent an event."
Credit
the ISA18 committee, co-chaired by Nicholas Sands of DuPont and Donald Dunn
2014
The same rules worldwide
The IEC published IEC 62682, its international version of ISA-18.2, on 15 October 2014. EEMUA says the current editions of ISA-18.2 and IEC 62682 are aligned with EEMUA 191.
Credit
IEC
Watch for
An agent for alarm floods
Emerson says its Alarm Insight Agent for the Ovation control system "prioritizes and summarizes alarm context during alarm floods" and acts only with human approval, with 28 more agents planned. The announcement gives no measured results.
Achieved whena customer or an outside party reports measured effects during real floods, such as response times or missed critical alarms.
Claimed by
Emerson, July 2026
Open
01
Draft the alarm rationalization
An AI agent that drafts, for each alarm on a plant's list, the record that alarm rationalization requires: its cause, the consequence of ignoring it, the operator's action, the time available to act and the resulting priority, for alarm engineers to review and approve.
The briefCredit · unclaimed02
Publish a real alarm log
An open, anonymized dataset of at least a year of alarm and event logs from a working plant, with its alarm list, priorities and the operators' notes on what caused each flood.
The briefCredit · unclaimed
Open 01 · the brief
Draft the alarm rationalization
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An AI agent that drafts, for each alarm on a plant's list, the record that alarm rationalization requires: its cause, the consequence of ignoring it, the operator's action, the time available to act and the resulting priority, for alarm engineers to review and approve.
Why it is still open
Rationalization is done in facilitated team workshops, which the consultancy exida lists as "typically 10 days," with the results kept in a master alarm database. A 2022 US patent application says rationalization takes an experienced person about 15 to 20 minutes per alarm, and that a plant may have tens of thousands of alarms. Searches found no peer-reviewed study of AI models drafting these records; a 2024 study used a language model to build a searchable knowledge base from power-grid alarms, not to rationalize them.
Built with what exists
a public alarm list, such as the one defined for the Tennessee Eastman simulation in PER4Mance; AI models that give probabilities for fixed-answer questions such as a priority level (see the Note "A thousand questions, one pass"); a written alarm philosophy for the plant; a review sheet for the engineers.
First step
Write a one-page alarm philosophy for the Tennessee Eastman plant and have a model draft records for its first 20 alarms.
Done looks like
For a list of at least 300 alarms, three alarm engineers rate each draft; at least 80% of priorities match the engineers' agreed choice, and the time per alarm, review included, falls below 5 minutes.
Open 02 · the brief
Publish a real alarm log
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open, anonymized dataset of at least a year of alarm and event logs from a working plant, with its alarm list, priorities and the operators' notes on what caused each flood.
Why it is still open
Alarm-flood methods are tested on simulations. A 2024 study by Gianluca Manca and Alexander Fay of ABB and Helmut Schmidt University used a public Tennessee Eastman dataset, and a 2025 paper that adds a dataset from a simulated fluid catalytic cracking unit says benchmark datasets "remain scarce."
Built with what exists
the alarm export of any control system; a script that replaces tag names and times consistently; an AI model to sort operators' log entries, checked by a person; an open data repository such as Zenodo.
First step
Ask a municipal water plant or a university power plant for one month of alarm logs, and agree in writing what must be removed before publishing.
Done looks like
Twelve months of logs from at least one plant published under an open licence, with EEMUA 191 alarm rates computed and at least five flood episodes labelled with their cause.