Skip to content
Measurement Systems and Gauge Management | Part 4 of 6

Running a Gauge Study and Reading the Variance Components

How to run a repeatability and reproducibility study with ten parts and three people, and how variance components show whether gauge or method is the cause.

Calibration proves that a gauge reads right on one day, in a laboratory, in trained hands. A study of repeatability and reproducibility proves that the measurement works every day, at the machine, in all hands. The Six Sigma series introduces the idea. This part shows how to run the study so that the result means something, and how to read it.

The run decides the value

The characteristic is the bearing seat of a shaft, 32.000 ±0.010 mm, measured at grinding cell G2 with a micrometer by hand. The rules of the run:

  1. Ten parts from production that cover the spread of the process, with hidden numbers.
  2. Three people who measure this characteristic in daily work.
  3. The gauge, the place and the method of daily work. No extra care.
  4. Each person measures all parts in random order. A second person writes. Nobody sees earlier values.
  5. Three rounds, each in a new order. Two are the minimum.

Ten parts, three people and three rounds give 90 readings.

A first look

The averages of the three operators over the same ten parts were 32.0004, 32.0009 and 32.0023 mm. Operator C read 1.8 µm more than operator A. That alone is nine percent of the tolerance.

Variances add up

The study splits the variation of the 90 readings into its sources. The calculation is a two-way analysis of variance: sums of squared deviations for parts, people and their combination, divided by their degrees of freedom. What remains inside the cells, one person on one part, is repeatability. It needs no table of constants.

Source Share of variance Standard deviation in µm
Repeatability (gauge) 3.3 % 0.80
Reproducibility (people) 5.1 % 1.01
Measurement system 8.4 % 1.29
Part to part 91.6 % 4.25

The shares of variance add up to 100 percent. The standard deviations do not.

The share of the tolerance

The number most plants decide on is the spread of the measurement against the tolerance:

share of tolerance = 6 × standard deviation of the measurement system / tolerance width

Here: 6 × 1.29 µm = 7.7 µm against a tolerance of 20 µm, or 38.6 percent. Common practice reads up to 10 percent as fit, 10 to 30 percent as acceptable for less important characteristics after weighing risk and cost, and more than 30 percent as not fit. Customers may set their own limits.

A second view is the number of groups of parts the measurement can tell apart. Five or more is the usual expectation. Here it is four.

From finding to fix

The larger source tells you where to look. Repeatability points to the gauge and its contact with the part. Reproducibility points to the method. In this study both were large, so training alone would not have been enough. The team had watched the run and seen three things:

  • Operator C closed the micrometer by the thimble, the others by the ratchet.
  • The bearing seat is 18 mm wide and slightly tapered, and each person measured at a different place.
  • Shafts came warm from the grinder, and the micrometer was in a hand for the whole shift.

The fix took the hand out of the measurement: a bench gauge with a V support, a stop for the position, a fixed measuring force and a digital indicator that is set with a master.

The second study

Same ten parts, same three people. Repeatability fell from 0.80 to 0.18 µm, reproducibility from 1.01 to 0.13 µm. The measurement now uses 6.8 percent of the tolerance and tells 25 groups apart.

The effect showed in scrap. A recheck of 50 rejected shafts had found 21 within tolerance. With the bench gauge, rejects for the bearing seat fell from 1.8 to 1.0 percent of 96,000 shafts per year: 768 shafts at EUR 18, or EUR 13,824. The investment of EUR 4,200 paid back in 3.6 months.

A passed study is valid for this gauge, this method, these people and this tolerance. The next part applies the same thinking to checks without numbers.

Newsletter

One Lean idea.
Every two weeks.