Nelson Rules in Production, Which Patterns Catch Drift
Pull-quote: “The eight Nelson rules are not eight equal alarms. On a real line, three of them do most of the drift detection, two exist to catch problems in your sampling, and running all eight everywhere mostly buys you false alarms.”
Why this matters
A control chart running only the 3-sigma limit test is a gross-error detector. It catches the broken insert and the wrong fixture, and it sleeps through the failure modes that actually cost money: tool wear, thermal drift, reagent depletion, a die warming up through first shift. Lloyd Nelson codified eight supplementary patterns in 1984 precisely because a lone limit test has almost no power against small sustained shifts. Four decades later the rules are standard equipment in every SPC package, which has produced a new failure mode: teams switch on all eight for every characteristic, drown in alarms, and learn to ignore the chart. Knowing which rule catches which physical problem is the difference between a chart that drives intervention and one that decorates a wall.
The eight patterns, and what they usually mean
| Rule | Pattern | Usual cause on a real line |
|---|---|---|
| 1 | One point beyond 3 sigma | Gross special cause: broken tool, wrong part, measurement blunder |
| 2 | Nine consecutive points on one side of center | Sustained small shift: new material lot, setup offset |
| 3 | Six consecutive points steadily rising or falling | Monotonic trend: tool wear, thermal drift |
| 4 | Fourteen consecutive points alternating up and down | Overadjustment, or two alternating streams sampled as one |
| 5 | Two of three consecutive points beyond 2 sigma, same side | Moderate shift, early warning |
| 6 | Four of five consecutive points beyond 1 sigma, same side | Small shift, the workhorse drift catcher |
| 7 | Fifteen consecutive points within 1 sigma of center | Stratification: mixed streams, or inflated limits |
| 8 | Eight consecutive points beyond 1 sigma, both sides | Mixture: two distinct processes on one chart |
The surprise for most engineers is rule 3. It reads like the designated drift detector, and it is the rule people reach for when they think about tool wear. In practice, drift rides on top of common-cause noise, and noise breaks up monotonic runs. Simulation studies and floor experience agree: the zone tests, rules 5 and 6, and the runs test, rule 2, usually fire well before a drifting process produces a clean six-point staircase. Drift announces itself as a shifted mean long before it produces a tidy trend.
What drift looks like before it looks like a trend
UCL ─────────────────────────────────────────────
+2σ ····························x····x··
+1σ ················x···x··x·······x
CL ──x───x··x───x───────────────────────────────
-1σ ····x·······x····
-2σ ·················
LCL ─────────────────────────────────────────────
├── stable ──┤├────── slow tool wear ──────┤
Rule 6 (four of five beyond 1σ) fires here ▲
Rule 1 has still never fired.The points never approach a control limit. The mean has moved, the zone counts see it, and the limit test is structurally blind to it.
The false-alarm budget
Every rule you add raises detection power and raises the false-alarm rate, and the second effect compounds faster than teams expect. A chart running rule 1 alone signals falsely about once every 370 points when the process is stable. Stack the common runs and zone tests on top and the in-control average run length drops to roughly ninety points. On a high-frequency line that is several phantom alarms per shift, per characteristic, and phantom alarms are how operators learn that the chart cries wolf. The practical discipline is a budget: rules 1, 2, 5, and 6 as the standing set on drift-prone characteristics; rule 3 where wear is physically monotonic; rules 4, 7, and 8 switched on as diagnostics when the chart looks wrong, because they describe your sampling and your limits more than your process.
On the line
A well-built SPC system evaluates Nelson rules live on X̄-R, I-MR, and p-charts and treats rule selection as chart configuration rather than a global switch. An out-of-control alert should arrive with its context: which rule fired, which zone, what the recent points looked like, so the response starts from the probable cause instead of a bare red dot.
Closing
The eight rules are a vocabulary, not a checklist. Rules 5, 6, and 2 catch the small sustained shifts that precede most drift-driven defects. Rule 3 confirms the wear you already suspected. Rules 4, 7, and 8 tell you the chart itself is misconfigured, which is worth knowing before you tune anything else. Map the rules to failure physics you actually have, budget the false alarms, and the control chart goes back to doing its one job: saying the process changed, early enough to act.