One alert in twenty is real. That is the whole problem.
A system that cries wolf too often stops being read. It is the best-documented mechanism in industrial safety, and it is exactly what happens to over-instrumented villas.
Aviation calls it alarm fatigue. Industrial control rooms formalised the problem in the 1990s, after incidents in which operators ignored a genuine alarm buried among hundreds of inconsequential ones.
The same phenomenon happens, on a small scale, in villas fitted with sensors.
I. What a typical installation produces
Take a properly equipped villa: door contacts, motion detectors, water sensors, temperature probes, cameras, connected gate, pool alarm.
Over a normal winter month it generates something like:
- Four outdoor motion notifications — a cat, a branch moving in the wind, a change of light at dusk.
- Two router disconnections, self-resolved.
- One low battery on a detector.
- One pool alarm triggered by wind.
- One low-temperature alert in the garage, which is an unheated space.
- One notification that the gate stayed open for two minutes, because the gardener came.
Zero significant events. Nine notifications.
Repeat for four months. By February the owner has turned notifications off, or sweeps them away unread. The day the plant room sensor reports water, that alert arrives in the same queue as the previous eight.
II. The five sources of noise, and their remedies
Default thresholds. A sensor out of the box is set for an occupied house. A temperature probe alerting below 5 °C in an unheated garage will alert all winter. Configuration must be done room by room, against what is normal for this house in winter, not what is normal in general.
No correlation. A motion detection on its own means nothing. Motion and a door opening and no alarm disarm within thirty seconds is an event. Systems that treat each sensor in isolation mechanically produce noise; those that cross-reference produce information.
No hierarchy. Not all alerts are equal. A low battery is not a water leak. A well-designed system distinguishes at least three levels and does not use the same channel for each.
Missing context. Water consumption detected on a Tuesday in January is an anomaly. The same consumption on the day the gardener waters, or the day after recharging the system, is not. A system blind to the house's calendar invents anomalies.
Cheap hardware. Some very low-cost sensors produce false positives by design: infrared detectors sensitive to thermal variation, misaligned magnetic contacts, pool alarms triggered by wind. A sensor's price should be judged by its false-positive rate, not its datasheet.
III. The alerting architecture we apply
We grade every signal at one of three levels, and each level has its channel, its recipient and its response time.
Level 1 — Information. Low battery, brief loss of connection, gate opened by an expected visitor, temperature variation within a normal range. Handling: recorded in the house log, no notification. Consultable, never pushed.
Level 2 — Anomaly. Temperature falling abnormally, unexplained water consumption, a sensor silent for more than twenty-four hours, a confirmed power cut. Handling: internal notification, human verification within twenty-four hours, owner informed only after verification.
Level 3 — Event. Water detected on the floor, correlated intrusion, prolonged power cut during a freeze, structural alert. Handling: immediate action — automatic shut-off, attendance on site — then a call to the owner. In that order.
That last point is the heart of it. The owner is told what has been done, not what has just gone off. It is the difference between receiving a problem and receiving a report.
IV. What that changes in practice
A house supervised this way typically generates two to four communications a year to the owner, plus the periodic report.
This is not a lower level of surveillance: it is the same volume of signals, filtered upstream. The sorting is done by someone whose job it is, with knowledge of the house — who knows the garage has been at 4 °C every winter since forever, and that the third bedroom has had a temperamental door contact since installation.
That kind of knowledge cannot be configured. It is acquired over one or two winters, and it is what separates supervision from a dashboard.
Frequently asked questions
Why does my home automation system send so many useless notifications?
Usually because thresholds are left at factory defaults calibrated for an occupied house, and because each sensor is treated in isolation. Room-by-room configuration against what is normal in that house in winter typically cuts the noise by more than eighty per cent.
Is it better to have fewer, well-configured sensors or many sensors?
Fewer, well configured. Each extra sensor adds noise; only signal processing adds information. A dense but badly tuned installation is less protective than a light one where every alert means something.
How do I avoid alert fatigue on a connected house?
By grading signals into distinct levels with a different channel for each, and by notifying the owner only after human verification for anything not requiring immediate action. Maintenance information should be logged, never pushed.
Is a sensor that stops communicating detected?
Only if the system supervises the state of its own components, which is not universal. A silent sensor looks exactly like a sensor detecting nothing: it is one of the most dangerous blind spots in an installation, and it should raise an anomaly after twenty-four hours of silence.
Should I be notified immediately of every alert?
No. For events requiring action, it is better that the action is under way before the call, and that the owner receives a report rather than an alarm. For everything else, periodic consolidation is enough and preserves the attention value of genuine alerts.
Updated August 7, 2026