How many Nelson rules are there?

There are eight Nelson rules. Seven derive from the older Western Electric rules, and Nelson added the fourteen-point alternating rule for over-adjustment. Each rule detects a different non-random pattern: a shift, a trend, excess or reduced variation, or a mixture of populations.

What is the difference between Nelson rules and Western Electric rules?

Seven of the eight Nelson rules come from the Western Electric rules published in 1956. Nelson added the fourteen-point alternating rule and uses nine points for the run rule where Western Electric uses eight. Both share the core signals and the zone-based approach; they differ mainly in details and in how many patterns they name.

What are zones A, B, and C on a control chart?

The zones divide the chart by distance from the centerline in standard deviations. Zone C is within one standard deviation, zone B is between one and two, and zone A is between two and three, with the control limits at three. The zones are mirrored above and below the centerline, and most Nelson rules are stated as counts of points in particular zones.

Should you use all eight Nelson rules at once?

Usually not. Every rule you add raises the false-alarm rate, because each has its own small chance of firing on a stable process. Running all eight makes a healthy process throw frequent false signals, and operators learn to ignore them. Choose the rules that match your process's real failure modes, plus rule 1, which everyone runs.

What does Nelson rule 4 detect?

Nelson rule 4 flags fourteen points in a row alternating up and down. A regular zig-zag pattern usually signals over-adjustment, reacting to normal variation by tweaking the process each time, or two alternating sources feeding the chart, such as sampling from two machines in turn.

Do Nelson rules apply to attribute control charts?

The pattern-based rules are most naturally applied to variables charts with symmetric, roughly normal behavior. On attribute charts like the np-chart or c-chart, the point-beyond-limits rule always applies, but the zone-based run rules should be used with care because attribute distributions are skewed, which changes the probabilities the rules assume.

Nelson Rules for Control Charts

A single point outside the limits is the obvious signal, but a process can go out of control in patterns that never cross a line. The eight Nelson rules are how you catch them, without jumping at every wiggle.

The Nelson rules are eight tests for detecting non-random patterns on a control chart. Each one flags a signal that a special cause has entered the process, a shift, a trend, or a run, that a single point outside the control limits would miss. Lloyd S. Nelson published them in 1984.

They matter because a control chart's power is not just the three-sigma limits, it is the ability to read the pattern of points between the limits. A process can drift, cycle, or split into two populations without any single point ever crossing a line, and if all you watch for is a point outside the limits, you miss the early warning entirely. The Nelson rules turn the chart into a pattern detector. This guide covers what the rules are, all eight of them, how they differ from the older Western Electric rules, and how to use them without drowning in false alarms.

What are the Nelson rules?

The Nelson rules are a set of eight criteria for identifying when a control chart is showing something other than random, in-control variation. They build on the idea that a stable process produces points scattered randomly around the centerline, most of them near it and fewer as you move outward, following a predictable statistical pattern. When the points stop looking random, too many on one side, a steady climb, a cluster far out, an unnatural calm, a special cause is likely at work, and the rules give you objective triggers for saying so instead of eyeballing it.

To apply them, the chart is divided into zones based on distance from the centerline in standard deviations (often called sigma or standard error). Zone C is within one standard deviation of the centerline, zone B is between one and two, and zone A is between two and three. The limits sit at three. Most of the rules are stated in terms of these zones, which is what lets them describe a pattern precisely instead of vaguely. The rules apply on the same control charts used throughout statistical process control.

The Nelson rules read the chart in zones measured in standard deviations from the centerline. Most rules are stated as counts of points falling in particular zones, which lets them describe a pattern precisely.

What are the eight Nelson rules?

There are eight, and each one catches a distinct kind of non-random behavior. Here they are in the order Nelson published them:

One point beyond zone A. A single point more than three standard deviations from the centerline. This is the classic out-of-control signal, a gross error or a large shift.
Nine points in a row on one side of the centerline. A run this long on one side is very unlikely by chance and signals a sustained shift in the process mean.
Six points in a row steadily increasing or decreasing. A monotonic trend points to tool wear, a warming bath, gradual depletion, something changing continuously.
Fourteen points in a row alternating up and down. A regular zig-zag suggests over-adjustment, two alternating sources, or systematic sampling from two streams.
Two of three consecutive points in zone A or beyond, on the same side. Two of three points more than two standard deviations out on one side signals a shift the eye can miss.
Four of five consecutive points in zone B or beyond, on the same side. Four of five points more than one standard deviation out on one side is another shift signal, subtler than rule 5.
Fifteen points in a row in zone C, both sides. An unnaturally calm stretch hugging the centerline suggests reduced variation, often mis-set limits, stratification, or a data problem, not genuine improvement.
Eight points in a row on both sides with none in zone C. Points avoiding the centerline entirely suggest a mixture of two distinct populations feeding the chart.

A few of the eight patterns: a sustained run on one side (a shift), a steady climb (a trend), and points hugging the centerline (unnaturally low variation). Each has a different root cause.

What is the difference between the Nelson and Western Electric rules?

They are close cousins. Seven of Nelson's eight rules come from the older Western Electric rules, the set published in the Western Electric Statistical Quality Control Handbook in 1956. Nelson compiled and refined them into eight tests and added the pattern the earlier set did not cover, the fourteen-point alternating rule for over-adjustment. The other visible difference is in the run lengths: Nelson's shift rule uses nine points on one side of the centerline, while the classic Western Electric run rule uses eight. The two systems agree on the core signals, a point beyond three sigma, two of three in zone A, four of five in zone B, and differ mainly in these details and in how many patterns they name.

In practice, the choice between them matters less than picking a consistent set and applying it the same way every time. Most modern SPC software offers both, and a plant should standardize on one so operators are not reacting to different triggers on different lines. What both share is the underlying purpose: separate the special-cause signals from the ordinary common-cause variation that a stable process always shows.

By the numbers. The eight rules were published by Lloyd S. Nelson in “The Shewhart Control Chart, Tests for Special Causes,” Journal of Quality Technology, vol. 16, no. 4 (1984), pp. 237–239, with a follow-up note in 1985 (Journal of Quality Technology, Nelson 1984). Seven of the eight derive from the Western Electric rules in the 1956 Statistical Quality Control Handbook, and NIST's engineering statistics handbook documents the same zone-based tests for special causes (NIST/SEMATECH e-Handbook, Western Electric rules).

Why do the rules use those specific numbers?

The run lengths are not arbitrary, each is chosen so the test has a small, comparable chance of firing on a stable process by accident. The anchor is rule 1: on an in-control process, a point lands beyond three sigma only about 0.27% of the time, roughly one false alarm in every 370 points. The other rules are tuned to match that order of magnitude. A run of nine points on one side of the centerline, for instance, has a probability of about one in five hundred on a stable process, because each point independently has about a fifty-fifty chance of being above or below the centerline and (1/2) to the ninth power is close to 0.2%. Set the run length shorter and the rule cries wolf; set it longer and it misses real shifts.

This is also why the numbers differ slightly between the Nelson and Western Electric systems and why you should not invent your own. Nelson's nine-point run and Western Electric's eight-point run both aim for the same balance between sensitivity and false alarms; they just land on adjacent integers. The zone rules, two of three in zone A, four of five in zone B, work the same way, trading a few points of evidence for a tighter probability. The lesson for the floor is to take the published rules as given rather than loosening a count because a chart is nagging you, since every adjustment quietly changes the false-alarm rate you are living with.

How do you apply the rules without over-reacting?

By remembering that every rule you add raises the false-alarm rate. Each test has its own small probability of firing on a perfectly stable process by pure chance, and those probabilities add up: run all eight rules at once and a healthy process will throw signals often enough that operators learn to ignore them, which defeats the whole point. This is the same trap as over-tightening any alarm system, too many false positives and people stop trusting the real ones.

The disciplined approach is to choose the rules that match the failure modes your process actually has. Rule 1 (a point beyond three sigma) is universal, everyone runs it. Beyond that, add the trend rule if you have tool wear, the shift rules if you have setup changes, the stratification rules if you feed the chart from multiple heads or streams. Do not run a rule whose pattern your process cannot physically produce. And when a rule does fire, investigate the cause rather than adjusting the process on reflex, reacting to a false alarm by tweaking a stable process (the mistake rule 4 is designed to catch) adds variation instead of removing it. The rules are best paired with a chart matched to the data, such as an individuals and moving range chart for one-at-a-time data or an np-chart for counts of defectives.

Keeping the rules useful on the floor is partly a data problem. When chart signals surface in real time at the line instead of in a report the next morning, an operator can act on a genuine trend while the cause is still present and traceable, and a rash of false alarms becomes visible fast enough to retune which rules you are running. That live feedback is part of what Harmony gives a plant, and it is the shift the team at CLS made when process data moved from next-morning paperwork to something visible during the shift. A Nelson rule only helps if someone sees it fire in time to do something about it.