Every process varies. Two parts are never identical — measure closely enough and the numbers always move. The question that matters is not whether a process varies, but why: is this the normal background wobble of a stable process, or has something actually changed? A control chart exists to answer exactly that question, and getting the answer wrong is expensive in both directions.
Two kinds of variation
Statistical process control starts with a single distinction, and everything else follows from it.
- Common-cause variation is the natural, ever-present noise of a stable process — the sum of many small sources like slight material differences, normal machine play and minor measurement scatter. It is predictable within limits, and it is simply the process being itself.
- Special-cause variation is an assignable event: a tool breaks, a new batch of material arrives, someone changes a setting. It is a signal that something specific has happened, and it is worth investigating.
The entire job of a control chart is to separate signal from noise — to tell you when a change is real and when you are only looking at the usual wobble.
The two costly mistakes
Confuse the two and you lose either way. React to common-cause noise as if it were a special cause — adjusting the machine after every reading — and you actually increase variation. This is Deming's famous funnel lesson: tampering with a stable process makes it worse, not better. Ignore a genuine special cause as if it were noise, and defective product leaves the building while you do nothing. Control limits draw the line between those two errors.
Anatomy of a control chart
A Shewhart control chart plots a process statistic over time against three reference lines:
- The centre line (CL) — the process average of whatever you are plotting.
- The control limits (UCL and LCL) — set at the centre line plus and minus three standard deviations of that statistic. This is the voice of the process.
- Zones A, B and C — each one standard deviation wide, used by the pattern rules to spot subtler shifts.
The simplest signal is one point beyond the control limits — there is only about a 0.27% chance of that happening by luck alone, so it is treated as real.
Control limits are not specification limits
This is the single most important idea in SPC, and the most commonly muddled. Control limits are the voice of the process — calculated from your own data, they tell you what the process actually does. Specification limits are the voice of the customer — set by a drawing or a requirement, they tell you what the process should do. They come from different sources and mean different things, and you should never plot specification limits on a control chart. Comparing the two — process spread against customer tolerance — is a separate exercise called process capability, and it only means something once the chart shows the process is stable. We cover it in Cp vs Cpk vs Ppk.
Why three sigma?
Shewhart chose three-sigma limits as an economic compromise. Tighter limits catch small shifts sooner but cry wolf constantly, so you end up chasing noise. Wider limits stay quiet but let real shifts slip through. Three sigma sits at roughly one false alarm per 370 in-control points — sensitive enough to be useful, calm enough to be trusted. Extra pattern rules can raise sensitivity to small, sustained shifts, but every rule you add also raises the false-alarm rate, so add them deliberately.
A control chart is not a pass/fail gauge. It is a running conversation with your process: it tells you when to act and, just as valuably, when to leave things alone.
Once you can read a chart, the next question is which one to build — and that depends entirely on your data. See Which control chart should I use?. When you would rather have the maths done for you, the SPC Control Chart Toolkit builds every chart automatically as you type your data in.