What is SPC in simple terms?

SPC is plotting a process measurement over time on a chart with statistically derived limits, so you can tell whether the process is just showing its normal noise (leave it alone) or something real has changed (investigate). It replaces gut-feel reactions to every number with a consistent rule for when to act.

What is the difference between control limits and specification limits?

Control limits are computed from the process's own data and describe what the process actually does. Specification limits come from the customer or the design and describe what you need. A process can be in control and still out of spec, and vice versa. Comparing the two is the job of process capability indices like Cp and Cpk, not the control chart.

Who invented statistical process control?

Walter A. Shewhart of Bell Telephone Laboratories sketched the first control chart in a one-page memo dated May 16, 1924, and laid out the full method in his 1931 book Economic Control of Quality of Manufactured Product. W. Edwards Deming later spread the method through wartime U.S. industry and postwar Japan.

How many data points do you need to start a control chart?

Common practice is 20-25 subgroups (or roughly 100 individual readings for an individuals chart) collected under normal operating conditions before computing control limits. Fewer points make the limits unreliable; the chart can still be plotted earlier, but treat early limits as trial limits.

Do you need software to do SPC?

No. A pencil chart at the station is legitimate SPC, and paper is often the best way to start because it forces clear thinking about the measurement and the reaction plan. Software matters at scale, when you need to trend dozens of characteristics, compare shifts, and connect chart signals to downtime and material data.

Where should a plant start with SPC?

One characteristic on one process where variation demonstrably costs money: fill weight with expensive giveaway, a machined dimension driving scrap, a seal parameter driving leaks. Prove the full loop (signal, reaction, root cause, fix) on that one chart before rolling out further.

SPC: Statistical Process Control Without the Statistics Degree

Q: What is the difference between common cause and special cause variation?

Common cause variation is the stable background noise always present in a process; you reduce it only by changing the process itself. Special cause variation comes from a specific, findable event, like a bad material lot or a broken fixture, and is addressed by finding and removing that cause. Control charts exist to separate the two.

Plot the point, read the signal, follow the plan. The math was finished in 1931.

Statistical process control (SPC) is a method of monitoring a process with simple time-ordered charts so you can tell the difference between normal variation and a real change worth reacting to. Its core tool is the control chart. Its core payoff is knowing when to act and, just as important, when to leave the process alone.

The name scares people off. In practice, the statistics are baked into the chart the day you set it up; running SPC on the floor is plotting a point, comparing it to two lines, and following a reaction plan. Operators who have never heard the word "sigma" run SPC well every day. What follows is the working version: enough theory to use the tool honestly, none of the derivations.

What is the difference between common cause and special cause variation?

Common cause variation is the routine noise built into the process: dozens of small influences (material lot differences, ambient temperature, normal tool wear, measurement noise) that are always present and produce a stable, predictable spread. Special cause variation comes from something specific and identifiable that is not normally part of the process: a broken fixture, a bad raw material lot, a new untrained operator, a machine setting someone changed.

The distinction decides what you do:

Special cause: investigate the event. Something happened; find it and remove it (or lock it in, if it made things better). This is a job for root cause analysis.
Common cause: do not chase individual points. The only way to improve is to change the process itself: better tooling, tighter incoming material, a redesigned step.

Getting this backwards is expensive in both directions. Treating common cause as special ("the number moved, adjust the machine") adds variation instead of removing it, a mistake W. Edwards Deming called tampering and demonstrated with his funnel experiment. Treating special cause as common ("scrap's always about 4%") means real, findable problems stay in the process forever, quietly feeding your cost of quality.

Common cause variation stays inside the limits with no pattern. Special causes show up as points beyond the limits or non-random runs.

Where did SPC come from?

SPC is one of the oldest tools in modern manufacturing and one of the least changed. In a one-page memo dated May 16, 1924, Walter A. Shewhart of Bell Telephone Laboratories sketched the first control chart, proposing that process data be judged against statistically derived limits rather than gut feel. He developed the method through the 1920s and laid it out fully in his 1931 book Economic Control of Quality of Manufactured Product (ASQ, Walter A. Shewhart).

W. Edwards Deming, Shewhart's colleague and editor, carried the method into wartime U.S. industry and then to postwar Japan, where it became a foundation of the quality movement that reshaped global manufacturing. In 1956, Western Electric's Statistical Quality Control Handbook codified the practical pattern-detection rules still used on charts today. The tool predates computers by decades, which is a useful reminder: SPC is a way of thinking about variation, not a software feature.

How do you implement SPC on a real production floor?

Rolling SPC out plant-wide on day one is the classic failure mode. Start with one characteristic on one process and make it work end to end.

Pick one characteristic that matters. Choose a measurable output tied to real scrap, rework, or customer complaints: fill weight, a critical dimension, seal temperature. If it has never caused pain, do not chart it.
Nail down the measurement first. If two people measuring the same part get meaningfully different numbers, the chart will lie. Agree on the gauge, the method, and the sampling point before collecting a single value.
Collect baseline data. Common practice is 20-25 subgroups (or about 100 individual readings) under normal running conditions before computing limits. Resist judging anything during this phase; you are photographing the process as it is.
Build the right chart and compute limits from the data. Control limits come from the process's own variation, never from the spec limits. Choosing among X̄-R, I-MR, p, and the rest is its own topic; our guide to control charts covers chart selection and reading.
Write the reaction plan before going live. For each signal type, define exactly what the operator does: what to check, whom to call, whether to stop. A chart without a reaction plan is wallpaper.
Train the operators who will run it. Not in statistics; in the three skills that matter: plot honestly, recognize the signals, follow the plan. Give them ownership of the chart, not just the pencil.
Review and improve. Once the process runs stable and predictable, measure whether it is actually good enough for the spec, which is a process capability question, and use what the chart taught you to reduce common cause variation. Recompute limits only when you have deliberately changed the process.

The roadmap. Steps 1-3 are where most SPC efforts are quietly won or lost, before any chart exists.

What should an operator do when the chart signals?

The reaction plan (some plants call it an out-of-control action plan, or OCAP) is the bridge between a statistical signal and a physical action. A good one is short, posted at the station, and written in the order an operator would actually check things.

A reaction plan turns a chart signal into a fixed sequence of actions. Without one, signals become debates.

Two details separate plans that work from plans that laminate well. First, the annotation habit: every signal, every adjustment, every material change gets a note on the chart. Six weeks later, those notes are the difference between a 20-minute root-cause session and archaeology. Second, escalation with teeth: if the cause is not found at the station, product from the signal window is treated as suspect and a non-conformance report or hold follows. A signal that never triggers containment trains everyone to ignore signals.

How do you read the signals without overreacting?

A point outside the control limits is the loudest signal, but not the only one. The 1956 Western Electric handbook added pattern rules that catch smaller shifts: two of three consecutive points far out on the same side, four of five points moderately out on the same side, eight consecutive points on one side of the center line. Each pattern is unlikely enough under stable conditions that it is worth treating as a real change. Our control charts guide covers the rules and chart selection in detail.

The discipline cuts both ways. Adding every rule to every chart multiplies false alarms, and a chart that cries wolf weekly gets ignored by week six. Most floor charts do well with the basic rule (a point beyond the limits) plus one run rule (eight in a row on one side). Add more only when the process history justifies the sensitivity, and adjust nothing when no rule fires. The chart's most common message is "leave it alone," and following that instruction is half the value of SPC.

Stable first, capable second

A control chart answers one question: is the process predictable? It does not answer whether the process is good enough. That second question belongs to capability indices, and the order matters:

Stability means the process shows only common cause variation: no signals, no drifts, no shifts. An unstable process has no single "voice"; any capability number computed from it is fiction that will change next week.
Capability compares the stable process's spread and centering against the specification limits, summarized as Cp and Cpk. A process can be perfectly stable and reliably produce 4% out-of-spec product; stability just means it will do so predictably.

Plants that skip the stability check and jump straight to Cpk reporting end up steering by numbers that wobble with every special cause. Chart first, stabilize, then compute capability from data the chart says is trustworthy.

What makes SPC fail on the floor?

The failure modes repeat across industries, and none of them are statistical:

Charts nobody reacts to. If a signal fires and production continues with no check, no note, and no containment, the chart is decoration. The reaction plan, and management actually backing it when it says stop, is the whole game.
Limits set from specs or wishes. Control limits drawn at the spec limits, or where a manager wants the process to run, detect nothing. Limits come from the process data or they are not control limits.
Charting everything. Fifty charts nobody reads lose to three charts that trigger action. Chart what costs money.
Punishing the signal. If an out-of-control point earns the operator a hard conversation, points stop landing outside the limits, on paper anyway. The chart measures the process, not the person; the moment that changes, the data quietly rots.
Recomputing limits to make problems disappear. Limits get recalculated after a deliberate, verified process change, not whenever the old limits become inconvenient.

Where does SPC pay off first?

SPC earns its keep fastest where these three conditions overlap:

High-cost variation. Processes where scrap, rework, or giveaway is measurably expensive: filling and dosing (overfill is pure margin loss), machining features that decide pass/fail, seal and weld parameters that decide leaks. Check your first pass yield data for the process with the biggest gap.
Adjustment-happy processes. Anywhere operators routinely tweak settings between checks "to keep it centered." A control chart either justifies those adjustments or, more often, shows they are tampering that adds variation. Deming's funnel point, live on your floor.
Customer-required characteristics. Automotive and aerospace customers commonly require SPC evidence on key characteristics as part of PPAP or contract requirements; capability indices like Cpk are computed from the same data, so the chart does double duty.

The historical pedigree here is worth trusting: the method survived a century essentially unchanged because it keeps working (ASQ, Statistical Process Control). The 1924 Shewhart memo, the 1931 book, and the 1956 Western Electric handbook define a tool that has outlived every management fad since.

Does SPC need software?

No, and yes. A pencil chart at the station is fully legitimate SPC, and starting on paper is often the right move because it forces the thinking. The limits of paper show up at scale: nobody can trend fifty pencil charts, compare shifts, or connect a chart signal to the downtime log and the material lot without hours of transcription. That is a data-plumbing problem, not a statistics problem, and it is the one Harmony was built for: digitizing station-level capture so quality checks land as structured, timestamped data that can be charted, searched, and trended plant-wide, alongside the QMS and ERP records you already keep. No rip-and-replace; the chart logic is still Shewhart's.

Start with one chart on one process that hurts. Prove the loop: signal, reaction, cause, fix. Then scale. That order, not the software and not the statistics degree, is what separates plants where SPC works from plants where it is wallpaper.