Shewhart control charts

The formula

For subgrouped measurements (n parts per sample), the X-bar chart tracks the subgroup averages:

CL  = x̿                 (grand average of the subgroup means)
UCL = x̿ + A₂·R̄
LCL = x̿ − A₂·R̄

Reading: the subgroup means should stay inside a band of ±3 estimated standard errors around the long-run average; A₂ converts the average range R̄ into that band.

The companion R chart watches the spread inside each subgroup:

UCL = D₄·R̄        CL = R̄        LCL = D₃·R̄

Reading: if within-subgroup scatter itself grows or shrinks, the R chart fires before the X-bar limits mean anything.

The constants come from the sampling distribution of the range: A₂ = 3/(d₂·√n) and σ̂ = R̄/d₂ ^[4]. For the classic subgroup of five, d₂ = 2.326 and A₂ = 0.577, with D₃ = 0 and D₄ ≈ 2.11 ^[4].

When parts come one at a time, the individuals (XmR) chart uses the moving range between consecutive points:

UCL/LCL = x̄ ± 2.66·mR̄        (2.66 = 3/1.128, the d₂ for n = 2)  [5]

Reading: same 3-sigma logic, with sigma estimated from how much the process moves point to point.

For attribute data — pass/fail counts and defect counts:

p-chart:  p̄ ± 3·√( p̄·(1 − p̄)/n )     (fraction nonconforming, binomial)
c-chart:  c̄ ± 3·√c̄                    (defects per unit, Poisson)

Reading: same idea, different distribution; the square roots are the binomial and Poisson standard deviations.

Where you meet it

The machining cell. A CNC lathe holding a critical bore diameter, five parts pulled every hour, X-bar and R chart taped to the operator's station. When a point lands outside the limits, the cell stops and someone checks tool wear, offsets, and coolant before the next lot runs.
The cal lab or test bench. An individuals chart on a check standard run through the DAQ chain every morning. The chart is what separates "the instrument drifted" from "the instrument is fine and Tuesday was just Tuesday" — and it catches a slipping calibration weeks before the annual cal cycle would.
End-of-line functional test. A p-chart on the fraction of units failing first-pass test each shift. A climb that stays inside the limits is noise; a point above the UCL sends the failure Pareto to the morning production meeting.
The supplier review board. PPAP packages and source-inspection reports arrive with control charts attached. Knowing how the limits were computed is how you catch the supplier who drew them at the spec limits instead of from the data.

How it works

The core idea is that the limits come from the process, not from the drawing. You collect twenty-plus subgroups while the process runs undisturbed, compute the center line and the 3-sigma limits from that data, and freeze them. The chart then answers one question per point: is this variation consistent with the process we characterized, or has something changed? Shewhart's terms were common causes (the ordinary noise of the process) and assignable causes (something specific and findable — a worn tool, a bad reagent lot, a loose connector) ^[1].

The single most common mistake on the floor is drawing the limits at the tolerance. Spec limits are the voice of the customer; control limits are the voice of the process. A process can sit comfortably in control and still ship scrap, or run wildly out of control while every part happens to be in spec. Control and capability are separate questions — the chart handles the first, and a Cp/Cpk study handles the second.

The second mistake is the opposite failure: adjusting an in-control process every time a point wanders off center. Chasing common-cause noise with knob turns adds variation instead of removing it. The chart's whole economic argument is that it tells you when not to act.

Why 3 sigma and not 2 or 4? For roughly normal data, about 99.73% of points from a stable process land inside 3-sigma limits, so a stable process gives a false alarm about once every 371 points ^[6]. Shewhart chose the multiplier on economic grounds — wide enough that people don't waste shifts chasing ghosts, tight enough that real shifts get caught — and was blunt that the choice was practical judgment, not a theorem ^[1]. To catch smaller shifts sooner, the Western Electric rules add pattern tests: 2 of 3 points beyond 2 sigma on the same side, 4 of 5 beyond 1 sigma, 8 in a row on one side of the center line, and runs trending or alternating. The price is a false alarm about every 92 points instead of 371 ^[6].

Limits of validity worth knowing before you trust a chart:

Independence. The math assumes successive points are independent. Sample a slow thermal process every second, or chart autocorrelated DAQ data, and the moving-range sigma estimate collapses — the limits hug the data and everything looks out of control.
Rational subgroups. Subgroups must capture only short-term variation (consecutive parts, one head, one cavity). Mix four cavities of a mold into one subgroup and the R chart inflates, the X-bar limits go wide, and real shifts hide inside them.
Attribute chart sample size. A p-chart with small n·p̄ is lumpy and its lower limit clamps at zero; you learn nothing about improvement from a chart that can't go down.
Small sustained drifts. A plain Shewhart chart is slow against a 1-sigma drift. That's the job CUSUM and EWMA charts were built for; the Shewhart chart's strength is large shifts and interpretability at a glance.

History

In 1918 Walter Shewhart, a physics Ph.D. out of Berkeley, joined Western Electric's inspection engineering department, the group responsible for the quality of telephone hardware coming out of the Hawthorne Works ^[2]^[3]. Inspection at the time meant sorting bad product after the fact. Shewhart's question was different: how do you tell, from the data, whether a process has actually changed?

On May 16, 1924 he answered it in a memo to his boss, George Edwards — about a page long, a third of it taken up by a hand-drawn diagram that any quality engineer today would recognize as a control chart. Edwards later said that single page set out essentially all the principles of what became process quality control ^[1]^[2]^[3]. When Bell Telephone Laboratories formed in 1925, Shewhart moved over and stayed until retiring in 1956 ^[2]^[3]. He worked the idea into a full theory in Economic Control of Quality of Manufactured Product (1931), still the founding text of statistical process control ^[1]^[2]^[3].

The idea traveled through a younger physicist named W. Edwards Deming, introduced to Shewhart in 1927. Deming edited Shewhart's 1939 lectures into the book Statistical Method from the Viewpoint of Quality Control, and in the summer of 1950 taught statistical process control to hundreds of Japanese engineers and executives ^[2]^[7]. Japanese manufacturers took the charts seriously at a moment when much of American industry had shelved them — a decision both countries' industrial histories spent the next thirty years grading. The Shewhart cycle Deming carried with him became Plan-Do-Study-Act, and Shewhart himself is why the American Society for Quality's top medal carries his name ^[2]^[3].

Related tools

/tools/rms-peak — the standard deviation behind the limits is the RMS of the deviations; same math, different label
/tools/snr-enob — separating signal from noise floor is the DAQ version of separating assignable from common cause
/tools/bearing-life-l10 — the other place engineers make decisions from a statistical population instead of a single measurement

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →