Most events the dashboard sees fit a known pattern: cooking, sleep, cleaning, off-gassing, wildfire infiltration. The fingerprint catalog (see co-movement patterns) covers the common cases. When the data does not match anything in the catalog, the dashboard does not guess. It tags the event as an anomaly and logs it for review.
The criterion is specific. An anomaly requires (1) a sustained deviation greater than 2 standard deviations from the rolling baseline, (2) duration longer than the 5-minute spike threshold, and (3) no fingerprint match in the catalog within a 90% confidence window. Single-parameter spikes that match known scenarios are not anomalies; they are recognized events. Sensor-jitter and known cross-sensitivities (humidity stepping the VOC index, fog inflating PM) are filtered out before the anomaly classifier runs.
When an anomaly is flagged, the dashboard does three things. It displays the multi-parameter trace so you can see the shape. It surfaces any contextual signals (outdoor AQI, weather, time-of-day, recent events in the home) that might explain it. And it logs the event so the AI can compare it against future events; recurring anomalies sometimes turn out to be undocumented patterns specific to your household (a particular HVAC quirk, a daily neighbor activity, a slow seasonal shift).
Anomalies are not alarms. They are observations. Most resolve on their own. Some get tagged by the user ("this was my neighbor's leaf-blower" or "this was a 3D printer running"), which extends the fingerprint catalog for your household. Persistent unexplained anomalies are worth investigating with a walkthrough; they sometimes turn up real problems (a slow gas leak, a hidden moisture source, a worn HVAC filter) that no single parameter would have caught alone.
References
- Sensirion - SEN66 datasheet and VOC index info sensirion.com
- Fonollosa et al. - Metal-oxide gas sensor drift doi.org
- AHAM - CADR program for room air cleaners aham.org
- EPA - AirNow: AQI Basics www.airnow.gov
