Setting Sensor Alarm Thresholds People Will Not Ignore

A newly installed monitoring system usually starts with enthusiasm: every parameter gets an alarm and everyone receives notifications. Two weeks later, the team's phones are full of mostly unimportant messages, and people start ignoring them. When a truly important alarm arrives, it drowns with the rest. This is known as alarm fatigue.

The key lies in how sensor alarm thresholds are set. This article covers practical principles to keep alarms few, accurate, and always acted on.

The basic rule: every alarm needs an action

Before creating an alarm, answer one question: what should the person who receives it do? If the answer is unclear, what you need may be a report or a chart, not an alarm. Alarms that demand no action only train people to ignore notifications.

Two levels: warning and critical

• Warning: a value is starting to leave its normal range. It needs attention, but there is still time. Suited to the responsible team during working hours.
• Critical: a value is in a dangerous zone or about to cause damage. It needs immediate action and may justify waking someone at night.

This split tells the team what can wait and what cannot. The optimal range between them is also useful as a reference on the dashboard.

Setting threshold values from data

1. Collect normal data first for a few weeks before adding alarms, so each parameter's normal range is visible.
2. Use technical limits from machine specifications or process standards for the critical level.
3. Place warning in between the normal range and the critical limit, with enough margin to act.
4. Differ per device when conditions differ. Two similar machines at different sites can have different normal operating temperatures.

Techniques to cut false alarms

• Time delay. Trigger only when the value stays out of range for several minutes, not on a single momentary spike.
• Hysteresis. Once an alarm is active, the value must return far enough into the normal range before the alarm clears. This stops alarms flickering when a value sits right at the limit.
• Operating state. Zero current on a machine that is deliberately switched off is not an alarm. Combine alarms with operating status where possible.
• Rate of change. For some parameters, a rise that is too fast matters more than the absolute value.

Who receives what

• Send alarms to people who can act, not to everyone.
• Choose the channel by level: in-app notification or email for warnings, instant messaging for critical.
• Set up escalation: if a critical alarm is not answered within a set time, send it to a supervisor or backup.
• Ask recipients to acknowledge alarms, so it is clear which ones are still open.

Review alarms regularly

Spend time each month on the alarm list: which fire most often, which are always ignored, and which never fired when a problem actually happened. Alarms that fire often but are never acted on need a new threshold or removal. Alarms that stayed silent during a problem need tightening. This small cycle makes the alarm system sharper over time.

For patterns that fixed thresholds struggle to catch, a data-driven approach can help, as covered in machine anomaly detection with machine learning.

Example: a motor temperature alarm

An illustrative setup for a pump motor:

• A few weeks of data show a normal casing temperature of 45 to 60 °C while running.
• Specifications and technician experience set 80 °C as the limit that must not be crossed. This becomes the critical level.
• The warning level sits at 70 °C, leaving room to check cooling or load before the limit.
• The alarm only activates if the temperature stays above the limit for five minutes, and clears once it drops below 65 °C.
• Warnings go to the shift technician, while critical alarms also go to the maintenance head if not answered within 15 minutes.

The figures are illustrative. What matters is the structure: a normal range from data, a critical limit from specifications, a warning with room to act, a time delay, hysteresis, and clear recipients.

Frequently asked questions

How many alarms is reasonable?

There is no fixed number. A simple test: every alarm sent should get a response. If many are ignored, there are too many.

What is the difference between warning and critical?

A warning flags a deviation that needs attention. Critical signals a dangerous condition that needs immediate action.

When should we use anomaly detection instead of fixed thresholds?

Fixed thresholds suit clear limits. Anomaly detection helps catch unusual patterns that never cross any fixed limit.

Alarms on the INCLUDE platform

On the INCLUDE platform, optimal, warning, and critical thresholds can be set per parameter and per device. Alarms can be acknowledged and resolved, and more complex logic can be built with rule chains, including JavaScript nodes. Notification channels depend on the plan, from in-app notifications and email to Telegram, as shown on the pricing page.