SPC CONTROL CHARTS · RUNS IN YOUR BROWSER · NOTHING UPLOADED

Has your process changed, or is it just noise?

For machinists, quality engineers, lab and line supervisors, and anyone with a column of measurements or defect counts in time order: paste them, and get a Shewhart control chart with all eight Nelson rules, trial limits you can revise, and Cp, Cpk, Pp and Ppk when you add specification limits. Free, with no account and no install.

Statistics on figures you paste. Not engineering, safety or quality-system advice. The example is real published data from NIST and Sematech.

Your data

Rules

Who this is for

A machine shop charting a critical diameter, a food plant checking fill weights, a lab tracking a control sample, a call centre counting errors per hundred calls, a contractor counting punch-list defects per unit. Statistical software does this well and costs money; spreadsheet templates rarely apply more than one rule. This page applies all eight, tells you which rule fired on which point, and says when the chart itself is the wrong one for your data.

The worked example: real wafers, two answers

Press Load the NIST wafer example: 450 line-width measurements from a semiconductor lithography process (Sematech, published by NIST), five sites on each of three wafers from thirty cassettes. Charted the textbook way, one wafer per subgroup on an X-bar and R chart, 25 of the 90 wafers fall outside the limits. The page then points out that the wafer averages move about 1.9 times as much from wafer to wafer as the spread across one wafer predicts, because each wafer's range measures site-to-site differences, not wafer-to-wafer ones. Switch the chart to subgroup averages as individuals and only the three wafers of cassette 27 fall outside. Which chart is right depends on whether wafer-to-wafer variation is part of normal running; NIST's own case study makes the same point about batch processes.

When to use something else

Autocorrelated data (a tank temperature read every second), very small shifts that matter (use CUSUM or EWMA), many correlated variables (multivariate charts), or a regulated quality system that needs validated software and an audit trail. To ask when a long series changed level or volatility, and how sure that is, without subgroups or limits, try the changepoint detector.

Method and citation

Limits are three standard deviations from the centre line, with the standard deviation estimated from within-subgroup variation: R-bar/d2, S-bar/c4, or the average moving range over d2(2) = 1.128 (Shewhart 1931). The constants d2 and d3 are computed by numerically integrating the distribution of the range of a normal sample (Tippett 1925) and c4 from its gamma-function formula; the test suite checks them against the published tables for subgroups of 2 to 25. The run rules are Nelson's (1984), applied to the location chart; range, S and moving-range charts get rule 1 only. Capability uses the within sigma for Cp and Cpk and the overall sample standard deviation for Pp and Ppk, with Bissell's (1990) approximate lower 95% bound on Cpk.

The moving-range chart deserves a caution: a range of two values is skewed, so about 0.9% of in-control points pass its upper limit against 0.27% for the individuals chart. Rigdon, Cruthis and Champ (1994) found it adds little detection power. Its signals are listed separately here.

@book{shewhart1931,
  author = {Shewhart, Walter A.}, title = {Economic Control of Quality of Manufactured Product},
  publisher = {D. Van Nostrand}, year = {1931}}
@article{nelson1984,
  author = {Nelson, Lloyd S.}, title = {The Shewhart Control Chart: Tests for Special Causes},
  journal = {Journal of Quality Technology}, volume = {16}, number = {4}, pages = {237--239}, year = {1984}}
@article{tippett1925,
  author = {Tippett, L. H. C.}, title = {On the Extreme Individuals and the Range of Samples Taken from a Normal Population},
  journal = {Biometrika}, volume = {17}, number = {3/4}, pages = {364--387}, year = {1925}}
@article{bissell1990,
  author = {Bissell, A. F.}, title = {How Reliable is Your Capability Index?},
  journal = {Journal of the Royal Statistical Society, Series C}, volume = {39}, number = {3}, pages = {331--340}, year = {1990}}
@article{rigdon1994,
  author = {Rigdon, Steven E. and Cruthis, Emma N. and Champ, Charles W.},
  title = {Design Strategies for Individuals and Moving Range Control Charts},
  journal = {Journal of Quality Technology}, volume = {26}, number = {4}, pages = {274--287}, year = {1994}}

Example data: NIST/SEMATECH e-Handbook of Statistical Methods, case study 6.6.1, Lithography Process (file LITHOGRA.DAT, a US government work).

Questions

What is a control chart?

A time-ordered plot of a process measurement with a centre line and upper and lower control limits set three standard deviations from the centre, where the standard deviation is estimated from short-term (within-subgroup) variation. Points outside the limits, or unusual runs inside them, suggest something changed in the process: a special cause worth looking for.

Which control chart should I use?

One measurement at a time: individuals and moving range (I-MR). Small subgroups of measurements, such as five parts an hour: X-bar and R. Subgroups of about ten or more: X-bar and S. Counts of defective units out of a sample: p (any sample size) or np (fixed size). Counts of defects: c (same amount inspected each time) or u (varying amount). The tool guesses from the shape of your data and you can override it.

What are the Nelson rules?

Eight tests for special causes published by Lloyd Nelson in 1984: a point beyond 3 sigma, nine in a row on one side of the centre, six in a row trending, fourteen alternating, two of three beyond 2 sigma, four of five beyond 1 sigma, fifteen hugging the centre, and eight avoiding it. Each extra rule catches small shifts sooner but also raises false alarms; this page lets you switch each one off.

How often does an in-control process raise a false alarm?

More often than people expect. In a seeded simulation of 20,000 perfectly stable processes, a first X-bar and R chart of 25 subgroups of five raised at least one signal 15% of the time with rule 1 alone and 29% with all eight rules. An individuals chart of 100 points did so 63% and 84% of the time, mostly because of the moving-range chart. A single signal is a reason to look, not proof.

What are Cp, Cpk, Pp and Ppk?

Capability indices comparing the specification width to the process spread. Cp and Cpk use the short-term (within-subgroup) standard deviation; Pp and Ppk use the overall standard deviation of all the data. The k versions account for the process not being centred. They only predict anything when the chart is in control.

Is my data uploaded?

No. Everything runs in your browser; nothing you paste leaves the page.

Not decided

Whether a signal matters, what caused it, whether to stop the line, and whether your specification limits are the right ones. Limits here are trial (Phase I) limits from the data you paste; for ongoing monitoring, fix limits from a stable period and chart new data against them. Normality is assumed for the variable charts and for the capability figures; heavily skewed data needs a transformation or another method.