What you're looking at
One row per (asset, parameter) that this service has studied.
For each combination we look at the last 30 days of incoming readings,
compute some basic statistics, and decide what kind of
signal it is. That decision determines whether we try to forecast the
series or leave it alone.
The four series kinds
- smooth — values that change gradually over time (temperatures, voltages). Forecast and monitored.
- event_driven — usually quiet, with occasional spikes (e.g. operation counters). Forecast both the calm baseline and the spikes.
- sparse — fewer than 100 samples in the last 30 days. Still forecast, but treat with caution: the model has little to go on.
- digital — text states (
ON/OFF) or signals that only ever take the values 0 and 1. Skipped — there is no meaningful "forecast" for a bit.
Columns explained
- Asset / Parameter — which asset and which parameter on it. The parameter label depends on the asset type (the same
hex_key means different things on different asset types).
- Kind — the classification above. Only smooth / event_driven / sparse rows are forecast.
- Cadence — how often this signal arrives, in milliseconds (typical interval between two readings).
- Samples — how many readings the profiler saw in the lookback window. More is better.
- Mean / Std — average value and its spread (standard deviation). Wider spread = noisier signal.
- MAD — median absolute deviation. Like std but ignores outliers; used internally to detect spikes.
- P01 / P99 — the 1st and 99th percentile. 98 % of readings fall between these. These become the default auto thresholds — see the Thresholds page.
- Spike thr. — for event_driven series, anything above
median + spike_thr counts as a spike.
- Profiled — when this row was last computed. Profiling re-runs every Sunday at
03:00 and on-demand via admin.
How to use this page
Use it to confirm which signals are being forecast and which
aren't. If a parameter you expect to see anomalies on is labelled
sparse or digital,
that explains why no forecast/anomaly is being generated.
If Samples is very low compared to a 30-day window, the upstream
feed is probably dropping data.