New to this topic?
We recommend reading these guides first to get the most out of this one:
SPC
Detect Before Defect
UCL/LCL
Control Limits (±3σ)
2
Types of Variation
React
Only to Special Causes

What Is SPC?

Statistical Process Control uses control charts to monitor a process in real time and detect when something has changed. Instead of inspecting parts after they are made (reactive), SPC watches the process as it runs and alerts you when it starts drifting — before defects are produced (proactive).

SPC is the link between process capability (can the process meet spec?) and daily operations (is the process still meeting spec right now?). A capable process that is not monitored will drift. SPC catches the drift.

Two Types of Variation

TypeWhat It IsExamplesCorrect Response
Common CauseNormal, inherent process variation — the system doing what it doesMinor temperature fluctuations, material batch-to-batch differences, normal tool wearDo NOT adjust the process. Only reduce through system improvement (better equipment, tighter material specs, improved method).
Special CauseAbnormal variation from something that changedBroken tool, wrong material loaded, new operator without training, machine setting bumpedStop, investigate, fix. Find what changed and restore or improve.

The Cardinal Sin of SPC

Adjusting the process in response to common cause variation makes it worse, not better. This is called "tampering" — like a golfer overcorrecting after every shot. If the process is stable (only common cause variation), leave it alone and work on system improvement. Only react to special causes.

Tampering, drawn The same 25 readings from the same unchanged process, run through two hands. Left alone, all 25 sit inside the control limits. Corrected after every reading, the same readings spread 41% wider and 2 of them land outside.

The page has no worked measurement example, so every quantity here is in units of the process's own σ — which is how the page itself talks: control limits at ±3σ (the stat grid, and the control chart table's "CL + 3σ" and "CL – 3σ"), zone lines at ±1σ (the two "hugging" rules), and 25 readings, the top of the page's own "collect 20-25 subgroups" baseline window. The 25 readings are a fixed illustrative sample of common cause variation, scaled so their realised standard deviation is 1.00σ; they are the SAME 25 readings in both panels. The left panel follows the page's Common Cause row, "Do NOT adjust the process", so the plotted point is the reading. The right panel follows the page's SPC Theater line, "Operator adjusts process after every reading", in its usual form: the dial moves by the last deviation in the opposite direction, so the point that lands on the chart is reading i minus reading i−1. That subtraction doubles the variance — Var(eᵢ − eᵢ₋₁) = 2σ² for independent readings — so the spread has to grow by √2 = 1.414, whatever the operator's intentions. The drawn sample realises 1.00σ → 1.41σ, a ratio of 1.41 and 41% wider; the range grows from 4.66σ to 6.37σ, points beyond ±1σ from 6 to 11, and changes of side from 10 to 16 of 24 — the alternation the page's "hugging limits" row describes. The two red points are readings 3 and 14. Untouched they are +2.43σ and −2.23σ, comfortably inside the limits; each is thrown out only because the previous reading's correction is still sitting under it. Checked against all five of the page's rules, the untouched lane trips none: 0 points outside the limits, longest run on one side of CL 6 (needs 7), longest trend 5 (needs 7), longest stretch inside ±1σ 5 (needs 15), longest alternating stretch outside ±1σ 2 (needs 8). The tampered lane trips rule 1 twice. The dial meanwhile made 24 adjustments over 26.7σ of total travel, ranging across 4.66σ, on a process whose mean never moved. Every figure in this note is recomputed from the plotted arrays.

The Control Chart

A control chart plots measurements over time with three lines:

LineWhat It IsCalculation
Center Line (CL)Process averageMean of all subgroup averages
Upper Control Limit (UCL)Upper boundary of expected variationCL + 3σ (3 standard deviations above mean)
Lower Control Limit (LCL)Lower boundary of expected variationCL – 3σ (3 standard deviations below mean)
- - - UCL - - -   (Upper Control Limit)
   •   •     •   •    •    •   ✖ ← Special cause!
─── CL ───   (Center Line / Mean)
  •    •   •     •   •    •
- - - LCL - - -   (Lower Control Limit)
Control chart: points within limits = stable process. Point beyond UCL = special cause — investigate immediately.

Control limits are NOT specification limits. Spec limits come from the customer (what they will accept). Control limits come from the process (what it actually does). A process can be in statistical control but out of spec (not capable), or in spec but out of control (unstable). See process capability.

Types of Control Charts

ChartData TypeWhat It MonitorsBest For
X-bar & RVariable (measurements)Subgroup average (X-bar) and range (R)Most common. Dimensions, weights, pressures, cycle times.
X-bar & SVariableSubgroup average and std deviationLarger subgroups (n > 10)
Individuals & MRVariableIndividual readings and moving rangeDestructive testing, slow processes, batch measurements
p-chartAttribute (pass/fail)Proportion defectiveGo/no-go inspection, % defective per lot
c-chartAttribute (count)Number of defects per unitScratches per panel, errors per form

Detecting Special Causes (Rules)

A point beyond UCL or LCL is the most obvious signal. But patterns within the limits can also indicate special causes:

PatternRuleWhat It Suggests
Point beyond limits1 point outside UCL or LCLSomething unusual happened at that moment
Run7+ consecutive points on same side of CLProcess mean has shifted
Trend7+ consecutive points trending up or downGradual drift (tool wear, temperature change)
Hugging center15+ points all within ±1σData may be stratified (e.g., mixing two sources)
Hugging limits8+ points outside ±1σ alternating sidesTwo different processes or conditions mixed

Implementing SPC

Select the critical characteristicDo not SPC everything. Pick the 3-5 features that matter most for quality, safety, or cost. These are typically the features that drive customer complaints or scrap.
Choose the right chart typeVariable data (measurements) → X-bar & R. Attribute data (pass/fail) → p-chart. Individual measurements → Individuals & MR.
Collect baseline dataRun the process under normal conditions and collect 20-25 subgroups. Calculate CL, UCL, LCL. Verify the process is stable (no special causes in baseline) before using the limits for monitoring.
Train operators to chart and reactThe person running the process should be the one plotting and interpreting. Train them on the rules: what is in control, what is not, and what to do for each. See TWI for effective training.
Define the reaction planWhen a special cause is detected: stop, mark the chart, notify the supervisor, investigate using structured problem solving. Document what was found and what was fixed. This creates a learning loop.
✅ SPC Done Right
  • Operators own the chart and understand the rules
  • Special causes are investigated the same shift
  • Charts are updated in real time, not backfilled
  • Control limits recalculated after process improvements
  • Common cause variation reduced through system improvement
❌ SPC Theater
  • Charts filled in at end of shift from memory
  • Out-of-control points with no investigation
  • Operator adjusts process after every reading (tampering)
  • Control limits never updated (same since 2015)
  • Charts exist for the auditor, not for the operator

🎯 Key Takeaway

SPC is your early warning system. It detects process changes before they become defects, tells you when to act (special cause) and when to leave it alone (common cause), and creates a data-driven quality culture on the floor. Start with your top 3 critical features, train operators to own the charts, and investigate every special cause the same shift. Over time, as you eliminate special causes and reduce common cause variation, your process becomes more stable, more capable, and more predictable.

Build a Control Chart

Generate data points to build a live control chart. Then inject a special cause to see what an out-of-control process looks like.

⚡
Try It Yourself
Control Chart Builder
▼
Generate data points to build a control chart. Then inject a special cause to see what an out-of-control process looks like. UCL and LCL are set at ±3σ from the target.
UCLCLLCL565044Click "Generate Data Points" to start
0
Data Points
50.00
Process Mean
0
Out of Control
NO DATA
Status
Ready for the full knowledge check? Test your understanding with guided scenarios and data export.
PROTake the Pro Knowledge Check →
Free forever · Every feature included

Stop reading, start modeling

Model your process flow, run simulations, optimize staffing with TOC math, and test your knowledge with 107 interactive checks — all in one platform.

Open Workbench → Pro Knowledge Check

Take this to a room

The running order

For operators and supervisors on a monitored process. They should leave able to tell common cause from special cause and to say what they will do for each.

7 beats · 11 min
  1. 1

    Two kinds of variation

    Everything on the chart is one of two things, and the correct response to each is the opposite of the other.

    • Common cause: the system doing what it normally does. Temperature drift, batch-to-batch material, normal tool wear.
    • Special cause: something changed. Broken tool, wrong material, uncalibrated setting.
    • Common cause: leave the process alone. Special cause: stop and investigate.

    Ask the room When our last reading looked off, which one was it?

  2. 2

    The cardinal sin

    Adjusting a stable process because a reading looked low makes it worse. That is tampering.

    • Like a golfer overcorrecting after every shot.
    • Each correction adds variation that was not there.
    • If the process is stable, the only route to less variation is system improvement.

    Ask the room Who is allowed to adjust our settings, and on what evidence?

  3. 3

    What the three lines mean

    Centre line is the process average. The limits are plus and minus three sigma of the process, and they are not the specification.

    • UCL and LCL come from the process, not from the drawing.
    • A part can be in control and out of spec, or in spec and out of control.
    • Confusing control limits with spec limits is the most common misreading there is.
  4. 4

    Pick the right chart

    The chart follows the data type. Getting this wrong makes the limits meaningless.

    • Measurements in subgroups: X-bar and R.
    • Individual readings, slow or destructive tests: Individuals and moving range.
    • Pass or fail: p-chart. Counts of defects per unit: c-chart.
  5. 5

    Five patterns worth knowing

    A point outside the limits is the obvious one. The others catch a shift before anything goes out of spec.

    • Seven or more points on one side of centre: the mean has moved.
    • Seven or more trending one way: gradual drift - tool wear, temperature.
    • Fifteen points hugging the centre: probably two sources mixed together.
    • Points hugging the limits, alternating sides: two different processes on one chart.
  6. 6

    The operator owns the chart

    If the person running the process is not the one plotting and reading it, you have a filing system.

    • Plot in real time, not backfilled from memory at end of shift.
    • Train the rules: what is in control, what is not, and what to do.
    • Recalculate the limits after a real process improvement.
  7. 7

    The reaction plan is the point

    A chart without a defined reaction is decoration. Write the plan before you hang the chart.

    • Special cause: stop, mark the chart, notify, investigate the same shift.
    • Document what was found and what was fixed - that is the learning loop.
    • Baseline first: 20 to 25 subgroups under normal conditions before the limits mean anything.

    Ask the room What are our top three critical features, and do any of them have a chart?