Case studyChemicalsDetect anomalies

71% fewer false alarms by replacing thresholds with one anomaly model.

A specialty chemicals plant retires hand-tuned alarm bands and lets the model speak.

Talk to an expertDownload 2-page PDF
Anomaly detectionOperator workflow
ALERTS · last 24h4 events
  • 14:22HIGHCMP_042 residual +3.2σ
  • 11:08MEDDRY_B drift trend
  • 07:41LOWBLR_01 efficiency dip
  • 02:15HIGHPUMP_07 cavitation pattern
At a glancetimeframes and caveats below
−71%
false alarms
3 months
Audited reactor unit
< 7 min
MTTD on real upsets
Same period
2 events; small sample
+18
Net Promoter from operators
Internal survey
n=12; pre/post
6 weeks
pilot length
Single unit
Existing Historian only
Challenge

What was happening before.

True upsets were buried in nuisance alarms. The HSE team flagged alarm fatigue as a top-three risk before the project began.

Constraints
  • Cannot remove safety-instrumented alarms; only nuisance process alarms are in scope.
  • Every replaced alarm has a recorded justification with sign-off.
  • Operator acknowledges every model alert with a one-click outcome label.
Solution

What we built, on what data.

Data
  • › Historian tags across the reactor unit and feed train
  • › Alarm history: source tag, time, ack, comment
  • › Shift logs
  1. 01
    Audit the alarm history
    Identified 38 candidate tags responsible for over 80% of acknowledged-and-cleared events.
  2. 02
    Train a single anomaly model
    Multivariate model on the reactor unit, validated against historical real upsets.
  3. 03
    Replace nuisance alarms one block at a time
    Each retirement was reviewed and signed off by operations and HSE.
  4. 04
    Operator label loop
    Every model alert is labelled true/false at close-out; labels feed the next retrain.
Adoption

Who uses it, and where it fits in the day.

Used continuously by the control-room operator. HSE reviews the labelled alert log weekly. Process engineering owns model retrains.

Modules in use
Anomaly detectionOperator workflow
Results

Quantified, with timeframe and caveats.

  • R1False-alarm rate down 71% on the audited unit.
  • R2Mean time to detect on real upset events stayed under 7 minutes.
  • R3Operator survey showed an 18-point improvement in alert trust.
ALERTS · last 24h4 events
  • 14:22HIGHCMP_042 residual +3.2σ
  • 11:08MEDDRY_B drift trend
  • 07:41LOWBLR_01 efficiency dip
  • 02:15HIGHPUMP_07 cavitation pattern

Operational outcome (proof rule): False-alarm rate on the audited reactor unit fell 71% in three months; mean time to detect on the two real upset events stayed under 7 minutes.

Voice from the floor
"When the alarm goes now, the crew actually walks to the cell. That hasn't been true here in a long time."
Shift Supervisor · Control room · Specialty Chemicals Co.
Also in the room

"Alarm fatigue was a top HSE risk for us. The model didn't make the alarms smarter; it let us delete the ones that never said anything."

Plant Manager · Operations
Replication

Relevant for teams trying to…

Chemicals and process teams with alarm-flood problems on continuous units.

Solution: Detect anomalies →Industry: Chemicals →
Lessons
  • Don't add a new alert layer on top of the old one — replace and retire. Two layers is worse than one.
  • Get HSE into the retirement loop on week one. They're the gate, not a reviewer at the end.
  • Label every alert. The model gets worse without it.
Get started

Talk to an expert about a similar use case.

Tell us your unit and your KPI. We'll come back with what data you'd need and a realistic timeline based on Specialty Chemicals Co.'s rollout.

Book a demoDownload 2-page PDF