Control Chart

Thinking Tools3.2 Finding the Past FactsAlso known as Statistical Process Control, Shewhart Chart

Developed by Walter A. Shewhart at Bell Labs (Economic Control of Quality of Manufactured Product, 1931).

Plots a measure over time against limits set from its own stable history, separating routine noise from signals that deserve investigation.

Twenty days of defects at a molding line, against limits set from its own stable month

Defects per day on one injection-molding line. The average and the limits come from last month, when the process ran undisturbed: about 15 defects a day, with routine days falling anywhere from 3 to 27. Points inside the limits are the process's ordinary noise; a point outside them, or a long run on one side of the average, is a signal that something changed. Illustrative.

Defects per day

Upper limit, 26.6: 27Average, 15: 15Lower limit, 3.4: 31: 142: 123: 174: 155: 196: 137: 118: 169: 2910: 1411: 1212: 813: 1514: 1715: 1616: 1817: 1718: 1919: 1620: 18Upper limit, 26.6Average, 15Lower limit, 3.401020301234567891011121314151617181920Upper limit, 26.6: 27Average, 15: 15Lower limit, 3.4: 31: 142: 123: 174: 155: 196: 137: 118: 169: 2910: 1411: 1212: 813: 1514: 1715: 1616: 1817: 1718: 1919: 1620: 18Upper limit, 26.6Average, 15Lower limit, 3.401020301234567891011121314151617181920

The figure under this one asks which of these days, or runs of days, deserve an investigation.

Show the numbers
DayDefects per day
114
212
317
415
519
613
711
816
929
1014
1112
128
1315
1417
1516
1618
1717
1819
1916
2018
Upper limit, 26.627
Average, 1515
Lower limit, 3.43

Which of these deserve an investigation?

Select an item from the chart above, then the box it belongs in: routine variation, to leave alone, or a signal, to investigate. Judge by the limits and the run rule, not by how bad a day looks. Day 9 is placed as an example. Show the sort to compare yours. Illustrative.

Items (select one, then a box)
  • Not placedPlacedRoutine: leave itSignal: investigate
  • Not placedPlacedRoutine: leave itSignal: investigate
  • Not placedPlacedRoutine: leave itSignal: investigate
  • Not placedPlacedRoutine: leave itSignal: investigate
  • Not placedPlacedRoutine: leave itSignal: investigate
  • Not placedPlacedRoutine: leave itSignal: investigate

Days

This month
Routine: leave it
Signal: investigate
Routine variationSignal

What the chart says

Day 9: 29 defects. The highest point of the month.Not placed yet.Day 9: 29 defects. The highest point of the month.Routine: leave it: Inside the limits with no run: the process's ordinary noise. Adjusting the machine for it adds variation; improving the whole process is the only way to shrink it.Day 9: 29 defects. The highest point of the month.Signal: investigate: Outside a limit, or seven in a row on one side of the average: something specific changed. Find it, then remove it or make it standard.Above the upper limit of 26.6. A new lot of resin went into the hopper that morning.Day 12: 8 defects. The lowest point of the month, and the shift lead wants to copy what was done.Not placed yet.Day 12: 8 defects. The lowest point of the month, and the shift lead wants to copy what was done.Routine: leave it: Inside the limits with no run: the process's ordinary noise. Adjusting the machine for it adds variation; improving the whole process is the only way to shrink it.Inside the limits: a quiet day of ordinary noise, with nothing specific to copy.Day 12: 8 defects. The lowest point of the month, and the shift lead wants to copy what was done.Signal: investigate: Outside a limit, or seven in a row on one side of the average: something specific changed. Find it, then remove it or make it standard.Day 5: 19, a weekly high. The highest of the first week, and the supervisor asked for an explanation.Not placed yet.Day 5: 19, a weekly high. The highest of the first week, and the supervisor asked for an explanation.Routine: leave it: Inside the limits with no run: the process's ordinary noise. Adjusting the machine for it adds variation; improving the whole process is the only way to shrink it.Well inside the limits; asking for a reason each high day sends people hunting for causes that are not there.Day 5: 19, a weekly high. The highest of the first week, and the supervisor asked for an explanation.Signal: investigate: Outside a limit, or seven in a row on one side of the average: something specific changed. Find it, then remove it or make it standard.Days 14 to 20. Seven days in a row above the average of 15, none of them above the limit.Not placed yet.Days 14 to 20. Seven days in a row above the average of 15, none of them above the limit.Routine: leave it: Inside the limits with no run: the process's ordinary noise. Adjusting the machine for it adds variation; improving the whole process is the only way to shrink it.Days 14 to 20. Seven days in a row above the average of 15, none of them above the limit.Signal: investigate: Outside a limit, or seven in a row on one side of the average: something specific changed. Find it, then remove it or make it standard.A run of seven on one side is a signal by itself: a new operator started on day 14 with a shorter cooling time.Day 8: up 5 on day 7. A rise of five from the day before, as large as any rise before day 9.Not placed yet.Day 8: up 5 on day 7. A rise of five from the day before, as large as any rise before day 9.Routine: leave it: Inside the limits with no run: the process's ordinary noise. Adjusting the machine for it adds variation; improving the whole process is the only way to shrink it.A jump between two points inside the limits is still routine noise.Day 8: up 5 on day 7. A rise of five from the day before, as large as any rise before day 9.Signal: investigate: Outside a limit, or seven in a row on one side of the average: something specific changed. Find it, then remove it or make it standard.Day 3: over the target. Seventeen defects, above the customer's target of 16 a day.Not placed yet.Day 3: over the target. Seventeen defects, above the customer's target of 16 a day.Routine: leave it: Inside the limits with no run: the process's ordinary noise. Adjusting the machine for it adds variation; improving the whole process is the only way to shrink it.Above the customer's target but inside the limits, so routine. The target sits inside the process's normal range: meeting it takes an improvement to the whole process, and explaining day 3 will not get there.Day 3: over the target. Seventeen defects, above the customer's target of 16 a day.Signal: investigate: Outside a limit, or seven in a row on one side of the average: something specific changed. Find it, then remove it or make it standard.

Your sort matched 0 of 6.Your sort matched 1 of 6.Your sort matched 2 of 6.Your sort matched 3 of 6.Your sort matched 4 of 6.Your sort matched 5 of 6.Your sort matched 6 of 6.

Routine: leave it
Inside the limits with no run: the process's ordinary noise. Adjusting the machine for it adds variation; improving the whole process is the only way to shrink it.
Signal: investigate
Outside a limit, or seven in a row on one side of the average: something specific changed. Find it, then remove it or make it standard.

The sort

  • Day 9: 29 defects: Signal: investigate. Above the upper limit of 26.6. A new lot of resin went into the hopper that morning.
  • Day 12: 8 defects: Routine: leave it. Inside the limits: a quiet day of ordinary noise, with nothing specific to copy.
  • Day 5: 19, a weekly high: Routine: leave it. Well inside the limits; asking for a reason each high day sends people hunting for causes that are not there.
  • Days 14 to 20: Signal: investigate. A run of seven on one side is a signal by itself: a new operator started on day 14 with a shorter cooling time.
  • Day 8: up 5 on day 7: Routine: leave it. A jump between two points inside the limits is still routine noise.
  • Day 3: over the target: Routine: leave it. Above the customer's target but inside the limits, so routine. The target sits inside the process's normal range: meeting it takes an improvement to the whole process, and explaining day 3 will not get there.

An example

An injection-molding plant reviewed defects every morning, and every bad day produced an action: the setter adjusted the temperature, the shift lead retrained someone, the supervisor asked for an explanation in writing. Defects were not falling, and the setters spent an hour a day adjusting.

The quality engineer took last month, when the line had run without changes, and set an average and limits from it: about 15 defects a day, with ordinary days anywhere from 3 to 27. She plotted the current month against them. Most of the days that had triggered action sat comfortably inside the limits, including the weekly highs that had drawn written reports.

Two things stood out. Day 9 was above the upper limit, and the log showed a new resin lot loaded that morning. And from day 14 every day sat above the average, a run of seven: a new operator had started that day with a shorter cooling time. The resin supplier got a quality notice, the operator's settings were corrected, and the morning review stopped acting on days inside the limits.

Why it works

Every process varies from day to day for many small reasons that no one can name or remove one at a time. Treating each high day as a problem sends people looking for a cause that is not there, and adjusting the process for ordinary noise adds variation of its own. Some variation does have a specific cause, though: a bad batch, a changed setting, a new person. That kind should be found and fixed.

The chart separates the two by letting the process set its own boundaries. Limits drawn from a stable period show the range ordinary noise produces. A point outside them, or a run of points on one side of the average, is very unlikely from noise alone, so it signals a specific cause worth investigating. Everything else is the system, and only changing the system will move it.

The thinking behind it

Walter A. Shewhart Walter A. Shewhart, a physicist and statistician at Bell Telephone Laboratories, sketched the first control chart in a 1924 memo and set out its theory in Economic Control of Quality of Manufactured Product (1931). His distinction was between common-cause variation, the steady noise of a process in control, and assignable causes, specific changes that can be found and removed. He also drew specification, production, and inspection as a cycle, which W. Edwards Deming later called the Shewhart cycle, and Deming edited his 1939 lectures, Statistical Method from the Viewpoint of Quality Control. He is often called the grandfather of statistical quality control.

How to run it

  1. Pick one measure. Choose a count or measurement that matters, taken the same way at a regular interval.
  2. Find a stable period. Take twenty or more points from a stretch when the process ran without known changes.
  3. Set the average and limits. Compute the center line and limits about three standard deviations either side, using the formula for the type of data.
  4. Plot every new point. Add each day's value to the chart as it comes in, with the limits fixed.
  5. Apply the signal rules. Treat a point outside the limits, or seven in a row on one side of the average, as a signal.
  6. Investigate signals only. Find what changed for each signal, then remove the cause or make it standard if it helped.
  7. Improve the system for the rest. If routine variation is too wide for the customer, change the process as a whole, then reset the limits.

Where it goes wrong

The limits come from the target. A team draws the customer's specification as the limits and treats every point beyond it as a signal; spot it when the lines on the chart are round numbers. Compute limits from the process's own stable data.

Every point gets a reaction anyway. Managers keep asking for explanations of high days inside the limits, and setters keep adjusting. Agree that points inside the limits get no action, and log the adjustments to show what tampering costs.

The limits never move. After an improvement, the old limits make the new process look permanently in control. Recompute them from a new stable period once a change has held.

When to use something else

  • Anomaly Detection. When there are many measures or no stable period, and the threshold has to be learned from the data.
  • DMAIC. When routine variation is too wide and the whole process needs a structured improvement project.
  • Fishbone Diagram. When a signal has been found and the team needs to list its possible causes.

Where it sits in SOLVE eX

Go to the source

  • Economic Control of Quality of Manufactured Product, 1931

Back to the catalog