Anomalies without the noise: bands, targets and AI analysis

A sensor that raises the alarm over every little thing gets muted within three days. HydroNode should only interrupt you when there is something worth seeing. This post explains how it decides what is unusual, and the three places where you tune that for your sensor.

Three checks for every reading

Your targetsA minimum and maximum you set per measurement. Active from the first reading.

GuidelinesPublished limits such as the WHO and German UBA values for CO₂, particulate matter or UV. Active from the first reading.

Learned bandThe range this sensor usually shows at this time of day. Needs some history first.

What followsOnly a broken target or guideline can become critical. Something merely unusual stays a warning at most.

How HydroNode learns what is normal

For each measurement HydroNode builds a model from the hourly averages of the last 28 days, with recent days weighing more. A warm afternoon is therefore not flagged just because the night was cool.

Stage	Needs	What changes
First reading	Nothing	Targets and guidelines are checked.
Learned band	20 readings over at least 6 hours	Statistical checks start.
Daily pattern	3 days with good coverage	The time of day is taken into account.
Full pattern	7 days	The daily pattern is fully trusted.

Single outliers do not shift what counts as normal. A new level that stays stable for six hours is accepted, for example after you moved the sensor. A level that breaks a target or guideline is never accepted: air that is always bad stays bad.

One incident per problem

A problem that keeps going stays one incident. New readings update it instead of opening new ones. It is only resolved after at least two good readings and ten minutes. Missing data never resolves anything.

OpenThe problem is confirmed
OngoingStill there, may get worse
RecoveringValues back in range
ResolvedTwo good readings, 10 minutes

Severity can go up, but it never quietly goes down in the background.

Incidents show up in two places: in the anomaly panel on the right of the sensor page, and for all sensors under Anomalies. A click shows the trigger, the rule, the readings around it and what to do. Once you rate an incident, it moves to Reviewed.

The anomaly panel of a sensor page with an open warning and an AI explanation — The anomaly panel: open, watched and resolved incidents, with the latest AI explanation below.

Dial 1: sensitivity

Open the settingsOn the sensor page press Settings, then Anomaly detection on the left.
Pick a levelFrom Very relaxed to Very strict. Below it you see how many anomalies the last seven days produced, and where they came from.
Check the suggestionAfter a busy week HydroNode suggests a calmer level, for example Use Relaxed.

Anomaly detection settings with five sensitivity levels and weekly counts — Sensitivity only widens or narrows the learned band. Your targets and the guidelines always apply.

Level	Band	Good for
Very relaxed	2.0×	Only extreme outliers
Relaxed	1.5×	Clear deviations only
Standard	1.25×	The default, balanced for rooms and gardens
Strict	1.0×	Smaller deviations too
Very strict	0.8×	Every noticeable change

Dial 2: your own targets

You know your tomatoes like 60 to 70 % humidity, or your nutrient solution needs a pH between 5.8 and 6.3? Then tell HydroNode.

Open the sectionSettingsLimits and targets
Set the rangePress Set range next to the measurement and enter the minimum and maximum.
Test it on the pastUnder Try it on past data HydroNode works out how many alerts your range would have raised over the last days. Nothing is sent and nothing changes.

Limits and targets settings per measurement — Without a range of your own it says Learned from your readings. With one, leaving it becomes a target breach.

Dial 3: what reaches your phone

Under SettingsNotifications you choose Off, Important only or Everything per sensor. The choice is yours alone and syncs to all your devices. The app still lists every event, and your automations keep running.

What the AI analysis adds

An AI analysis is a short written explanation of how a sensor is doing right now: a summary, up to six findings and up to three concrete next steps. It explains the result, it does not decide anything. Values, severities and incidents are calculated by HydroNode first with fixed rules, and the AI may not contradict them.

AutomaticOn a fixed schedule, if new readings arrived since the last analysis.

On an incidentWhen an important incident opens, after 60 seconds, so related ones are explained together.

By handAnalyze in the anomaly panel. If nothing changed, you get the last one back.

PrivacyRuns on Amazon Bedrock in Frankfurt. No name, no email, no sensor name, no location.

How many analyses you have left this month is shown under SettingsAnomaly detection next to AI summaries. Every detail is in the AI analysis docs.

When instead of an odd value no value arrives at all: Know when a sensor goes quiet.
React to anomalies in Home Assistant: HydroNode in Home Assistant.
Every incident type and rule: anomaly detection docs.