X̄ and R Control Charts

A control chart is simply a running plot of a process's measurements with two lines drawn on it, marking how far things normally wander; a point outside them says "something changed, go look." The limits look like a jumble of constants, but each is one idea: center line $\pm$ 3 standard errors. For the $\bar x$ chart that's $\bar{\bar x} \pm A_2\bar R$ (the grand average, plus or minus a table constant times the average range, which is a quick stand-in for 3σ/√n). The R chart does the same for the ranges. Look up two constants, multiply, and draw lines.

Every process jitters. Weigh five bags from a coffee filler and you get five different numbers: valve timing, bean density and vibration all nudge each fill. That everyday scatter is common-cause variation, built into the process. Then a nozzle partly clogs and every bag comes out light: a special cause, something that changed and can be found and fixed. A control chart tells the two apart, so operators neither ignore real trouble nor "fix" the noise (adjusting a stable process after every reading only adds variation).

Rational subgroups. Every half hour, measure $n$ consecutive units, usually 4 or 5. Units made a minute apart share the same conditions, so the variation within a subgroup is pure common cause, and a special cause shows up between subgroups. Plot each subgroup's mean $\bar x_i$ on the $\bar x$ chart and its range $R_i$ (largest minus smallest) on the R chart. The gauge has to resolve that variation, or many ranges come out 0 and the R chart breaks (see Measurement Error).

Where the limits come from. If the process is stable with mean $\mu$ and standard deviation $\sigma$, a subgroup mean has standard deviation $\sigma/\sqrt n$, and by the central limit theorem it's nearly normal even when single units aren't. So a stable process keeps $\bar x$ inside $\mu \pm 3\sigma/\sqrt n$ 99.73% of the time. We estimate $\sigma$ from the ranges: for normal data the expected range of $n$ units is a fixed multiple of $\sigma$, $E(R) = d_2\sigma$, where $d_2$ depends only on $n$. Then $\hat\sigma = \bar R/d_2$, and

The range has its own standard deviation, $d_3\sigma$, so the R chart's limits are $\bar R \pm 3d_3\bar R/d_2 = \bar R\,(1 \pm 3d_3/d_2)$, tabulated as $D_4\bar R$ and $D_3\bar R$. For $n \le 6$ the lower one would be negative, so $D_3 = 0$.

Check a row: $3/(2.326\sqrt5) = 0.577$. Read the R chart first: the $\bar x$ limits are built from $\bar R$, so if the spread is out of control they mean nothing.

Two phases. In Phase I, take 20 to 25 subgroups, compute trial limits and look for points outside them. Drop the ones whose cause you find and recompute: the result describes the process when it's stable. In Phase II those fixed limits watch new production.

Run rules. Limits at 3σ are slow to notice a small, steady shift, so most plants add pattern rules. Mark zones at 1σ and 2σ of $\bar x$ (that is, at $\hat\sigma/\sqrt n$ and $2\hat\sigma/\sqrt n$) on each side of the center line. A stable process rarely makes any of these:

1. One point beyond 3σ.
2. Two of three consecutive points beyond 2σ on the same side.
3. Four of five consecutive points beyond 1σ on the same side.
4. Eight points in a row on one side of the center line.

Rule 4 is a coin-flip argument: each point lands above center with probability ½, so eight in a row on one side has probability $2(1/2)^8 \approx 0.008$. Every extra rule catches small shifts sooner and adds a few false alarms, the type I versus type II trade-off from hypothesis testing.

How often it cries wolf. A stable process puts a point outside 3σ limits with probability $2(1 - \Phi(3)) = 0.0027$. Each subgroup is an independent trial, so the wait for the first false alarm is geometric with mean $1/0.0027 \approx 370$ subgroups: the average run length, $\text{ARL}_0$. After the mean shifts by $\delta\sigma$, the UCL sits only $3 - \delta\sqrt n$ standard errors above the new mean, so a subgroup signals with probability about $1 - \Phi(3 - \delta\sqrt n)$, and $\text{ARL} = 1/P$.

A common mistake. It's tempting to draw the spec limits on an $\bar x$ chart, right next to the control limits, so you can see both at once. Don't. Control limits are the voice of the process; spec limits are the voice of the customer. An $\bar x$ chart plots averages, and single parts spread $\sqrt n$ times wider: a point comfortably inside the specs can hide parts that aren't. A process can be perfectly in control and still make scrap, which is the question of Process Capability. And subgroups that mix sources (two machines, four filling heads) inflate $\bar R$, widen the $\bar x$ limits and blind the chart.

Every high-volume plant runs control charts. Tablet presses chart weight and hardness through each batch, semiconductor fabs chart film thickness and line width on hundreds of tools, and machining cells chart critical diameters right at the machine, often on a screen that flags run-rule violations automatically. The same logic watches processes that aren't factories: infection rates in a hospital, wait times in a call center and response times of a web service.