Control Limit
How Control Limits Separate Signal From Noise
Walter Shewhart introduced control limits at Bell Telephone Laboratories in 1924 to solve a costly problem: factory operators were constantly adjusting equipment in response to ordinary variation, which made output worse rather than better. His insight was that every process has an inherent, predictable amount of random scatter (common-cause variation) and that meaningful intervention should only follow evidence of a genuine shift (special-cause variation). Control limits draw the line between the two. A subgroup statistic that lands inside the limits, with no nonrandom pattern, indicates a process that is in statistical control and should be left alone. A point outside the limits, or a recognized run pattern inside them, is the signal that something assignable has changed and warrants investigation.
The defining feature is that control limits are voice-of-the-process boundaries, derived entirely from the measured data and the subgroup sampling scheme. They are not targets, tolerances, or contractual requirements. On an RF production line measuring connector return loss, insertion loss, or amplifier gain, the chart is built by collecting rational subgroups (often four or five consecutive units), plotting the subgroup mean and range over time, and computing the limits from the average range. Because the limits move only when the underlying process variation changes, they provide a stable reference against which drift, tool wear, fixture loosening, and material lot changes become visible long before parts go out of specification.
A frequent and expensive error is confusing control limits with specification limits. A process can be perfectly in control yet incapable, producing scrap, if its natural spread is wider than the tolerance. Conversely a capable process can drift out of control while every individual part still passes inspection. The chart catches the drift early so that corrective action precedes any rejects. This is why control charts pair naturally with capability studies and with Cpk reporting on a supplier scorecard.
Control Limit Equations
UCL = μ + 3σ LCL = μ − 3σ
X-bar chart (from average range):
UCLX̄ = X̄̄ + A2 × R̄ LCLX̄ = X̄̄ − A2 × R̄
Range chart:
UCLR = D4 × R̄ LCLR = D3 × R̄
Sigma estimate: σ̂ = R̄ / d2
Where X̄̄ = grand mean, R̄ = average subgroup range, and A2, D3, D4, d2 are subgroup-size constants. For n = 5: A2 ≈ 0.577, D3 = 0, D4 ≈ 2.114, d2 ≈ 2.326. Example: tuned WR-15 filter, X̄̄ = 0.40 dB, R̄ = 0.042 dB → UCLX̄ ≈ 0.424 dB, LCLX̄ ≈ 0.376 dB.
Control Limits vs. Specification Limits
| Attribute | Control Limit (UCL / LCL) | Specification Limit (USL / LSL) |
|---|---|---|
| Source | Calculated from process data (voice of the process) | Set by drawing, datasheet, or customer (voice of the customer) |
| Question answered | Is the process stable and predictable? | Is this individual part acceptable? |
| Typical placement | Mean ± 3σ of the plotted statistic | One- or two-sided tolerance, e.g. insertion loss 0.40 dB max, or gain 30 ± 0.5 dB |
| Applies to | Subgroup means, ranges, or individuals over time | Each individual unit at inspection |
| Changes when | Process variation itself changes | Engineering or contract requirement changes |
| Goes on the chart? | Yes, as horizontal control lines | No, never drawn on a Shewhart control chart |
Frequently Asked Questions
What is the difference between a control limit and a specification limit?
Control limits describe what a process actually does; specification limits describe what the customer requires. Control limits come from the process data itself, placed at the mean ± 3σ of the plotted statistic, and answer whether the process is stable. Specification limits (LSL and USL) come from the drawing or datasheet, such as a 0.30 dB maximum insertion-loss spec, and answer whether a part is acceptable. A process can be in control yet still make out-of-spec parts if its spread is too wide, which is why Cp and Cpk compare the two.
How do you calculate the upper and lower control limits for an X-bar and R chart?
For the X-bar chart, UCL and LCL = X̄̄ ± A2 × R̄, where X̄̄ is the grand average, R̄ is the average subgroup range, and A2 depends on subgroup size (A2 = 0.577 for n = 5). For the R chart, UCL = D4 × R̄ and LCL = D3 × R̄ (D3 = 0, D4 = 2.114 for n = 5). These constants fold in the d2 bias factor that turns average range into a sigma estimate, so the limits land at about ± 3σ of the subgroup mean.
Why are control limits set at three sigma rather than two sigma?
Three-sigma limits balance the two error types Shewhart studied. At ± 3σ, a stable process gives a false alarm only about 0.27% of the time, roughly 1 in 370 subgroups. Tightening to two sigma raises false alarms to about 4.6%, flooding operators with phantom signals and eroding trust in the chart. Three sigma keeps false alarms rare while still catching real shifts of about 1.5σ within a few subgroups, especially when paired with Western Electric run rules.
What do you do when a point falls outside a control limit on an RF production line?
A point beyond a limit signals special-cause variation, so you investigate and remove the assignable cause rather than adjust toward target. On an RF line that means checking the network-analyzer calibration, fixture torque and connector wear, the cable lot or solder reflow profile, and operator changeover. Quarantine the subgroup and root-cause it (often with an 8D or fishbone) before resuming. Reacting to common-cause noise by tweaking setpoints is the over-adjustment mistake Deming's funnel experiment shows increases variation.