A threshold alarm fires only when a value crosses a line. An SPC pattern alarm looks for shapes that control-chart rules flag — a run on one side, a steady climb — while the value is still inside the limits. It arrives in the release after 0.1.0.24.
What it catches
An alarm is raised when any of the selected control-chart rules (R1 to R8) matches. Three are on by default.
| Rule | Meaning |
|---|---|
| R1 | One point beyond 3σ from the center |
| R2 | Nine points in a row on one side of the center |
| R3 | Six points in a row steadily rising or falling |
R4 to R8 can be turned on when needed. The matching rule number is kept in the alarm message and history.
Setup
- Choose SPC pattern as the condition of an alarm rule and set the baseline (center and σ) in one of three ways:
- Enter it
- Compute from history — fill it from past records (1 h, 8 h, 24 h, 7 days).
- Learn baseline — after the rule is on, the next N samples (50 by default, 20 to 1000) set the center and σ. No judging happens while learning, and the list shows progress such as "learning 23/50". If the value never changes (σ of 0) it reports a failed learning with the reason. There is no automatic re-learning; Relearn starts over. A learned baseline survives restarts.
- Subgroup size 1 to 10 — the mean of n samples is judged as one point (limits at σ/√n), for catching small shifts in noisy values.
- Sample period defaults to 10 s; each period takes the tag's current value as one point. Shorter periods raise more false alarms.
- Auto clear — the alarm clears itself after 8 samples in a row (default) with no new violation.
- Points are reset and gathered again when the value goes bad or stops updating, and after a restart.
- Notices go out the same way as other alarms (screen, email, Telegram).
Tuning by eye
- Small control chart — the rule editor, the rule list and the alarm history show the last 50 points with center line, ±3σ and violating points.
- False-alarm interval hint — from the selected rules, sample period and subgroup it tells you in advance, for example "may ring about once every 2.5 minutes even in a normal process (about every 150 points)" for R1 to R3 with 1-second samples. Processes whose values follow each other closely can ring more often.
- History trend overlay — the history trend screen can overlay SPC rules on past records to show where they would have matched (computed from the rows the screen received, so it may differ from what the alarm saw).
- The analytics screen also gains rule toggles (R1 to R8) to try the same rules on past data.
Good to know
SPC rules assume a stable process. Signals whose values follow each other closely (a slowly moving temperature, say) can trip R2 or R3 often, so start with the default rules and a 10-second sample and tune while watching for false alarms.