One alarm per ten minutes is considered acceptable during steady-state operation by EEMUA 191, the process industry's nearest equivalent to a consensus alarm-management standard. Discussions with process operations teams at UK chemical and energy sites over the past year found abnormal-condition alarm rates typically running at 5-15x that level. The alarm system is performing as designed. The issue is that operators cannot both manage a complex process unit and respond meaningfully to 150 alarms per hour.
How nuisance alarms differ from false alarms
High alarm volume does not always mean the alarms are technically false. Many correctly identify a process value beyond its configured limit. In practice, the useful distinction is between alarms carrying actionable diagnostic information and signal noise: chattering sensors, cascades caused by one upstream event, and limits set too close to normal process variation.
A level alarm near a control valve may chatter hundreds of times per shift without signaling a real abnormal condition. One upstream pressure drop can cascade into a dozen related alarms within seconds, with none adding information once the source is known. Standing alarms active for days or weeks form a constant background that operators learn to disregard, effectively disabling alarm coverage for those measurement points.
Alarm rationalization programs, defined by IEC 62682, the international alarm-management standard, have addressed this distinction for decades. For AI-based methods, the question is whether machine learning can make the work faster and more continuous than periodic manual reviews.
When basic threshold filters have no answer
The usual engineering answer to alarm floods is to raise setpoints, introduce time delays, and use dead-bands. These measures help, but all share one limitation: they lack awareness of the wider process. A high heat-exchanger outlet temperature matters during steady state, yet may be expected and temporary during a feed composition change. A static filter that hides the alarm in the latter case must know that the composition has changed.
Static filtering cannot resolve this issue. Alarm assessment needs context: which operating mode is active, what changed during the last 30 minutes, and what historical data shows about the alarm's behaviour in comparable conditions. Once that context is required, machine learning becomes genuinely useful rather than simply fashionable.
What AI-based filtering can really add
AI can add two capabilities to alarm management that manual rationalization cannot deliver continuously at scale. First, it can evaluate alarm quality near real time from historical response data. An alarm that fired 60 times over the past month and received no recorded operator action in more than 50 instances differs measurably from one that consistently prompts a response. This does not dictate the remedy, but it gives the instrumentation team documented evidence to prioritize rationalization work.
Second, AI can find alarm relationships that expose cascade patterns. When alarm A precedes alarms B, C, and D by two to eight seconds in 80% of cases in which all four occur, the evidence strongly suggests that B, C, and D are downstream effects of what A detects. Operators should see A as the diagnostic signal, while B, C, and D should be suppressed or grouped beneath A for that event type. Manual review may find these links eventually, but usually after an incident rather than continuously during operation.
In a pilot at a continuous process site, Offset classified roughly 35% of overall alarm volume as noise candidates from 90 days of historical alarm and response data. This did not mean 35% of alarms were automatically suppressed. Instead, 35% were flagged for the site instrumentation engineer, with response-rate data attached, allowing the team to decide which setpoints and dead-band settings to review first. The operations manager said they had long known that much of their alarm volume was noise, but lacked a systematic, evidence-based way to demonstrate that to engineering.
When operators acclimate to alarm floods
High-frequency alarm environments produce a well-established behavioural pattern. Repeated alarm floods teach operators to discount alarms as a category. That response is understandable: when 90% of alarms on a day need no action, it is rational to expect the next one will also need none. The difficulty is that this learned response does not reliably change when a genuine abnormal condition appears.
Process safety research repeatedly finds that alarm floods before major incidents contain those incidents' diagnostic signals, often minutes before a trip, hidden among noise that operators have learned to treat as background. Lowering the false alarm rate is therefore more than an efficiency matter. It also concerns safety. An operator unconditioned to alarm noise responds differently to an urgent alarm.
Limits of automated filtering
We do not propose letting AI choose which alarms to suppress without human review. Automated suppression creates the reverse risk: a model trained on historical cases may remove a genuinely diagnostic alarm because it was usually noise in that data. Processes evolve, and a correlation dependable in historical records may fail after equipment changes or feedstock changes.
The sounder approach is to use AI as a continuous audit of alarm quality, presenting evidence for human decisions. The instrumentation team chooses which alarms to rationalize and sets the relevant setpoints and dead-bands. AI supplies response rates over time, alarm correlations, and operating-mode context. This arrangement is easier to defend for process safety and regulatory purposes, and changes are more likely to endure because engineers understand the reason for each one.
Why alarm rationalization efforts stall
Manual alarm rationalization requires substantial resources and loses relevance quickly. A site completing a detailed review in year one may find that equipment modifications, control-loop changes, and process improvements have weakened its value within three years. The constraint is not a lack of commitment, but insufficient time and staffing to study alarm history, connect it with incident data, and make defensible decisions on an ongoing basis.
AI's practical role in alarm management is not replacing engineering judgment about nuisance alarms. It is performing the data analysis that limits how often that judgment can be used. Continuous alarm-quality monitoring lets a site make rationalization decisions monthly rather than annually. Alarm-system quality follows review frequency, while review frequency depends on analysis cost. Lowering that cost is the concrete contribution.