Measurement system analysis: finding out how much of your variation is the gauge

Repeatability and reproducibility are different diseases

One person measuring one part several times and getting different answers points at the instrument, the fixture or the part surface. Several people measuring the same part and getting different answers points at method: they are holding it differently, reading it differently, or applying a different amount of force. The distinction matters because the remedies have nothing in common. Poor repeatability is fixed with better equipment, a fixture or a cleaner presentation. Poor reproducibility is fixed by writing the method down and training to it, which is usually far cheaper and is often refused because it feels like a criticism of the inspectors.

Bias, drift and the tails of the range

A measurement can be perfectly consistent and consistently wrong, which no amount of repeated reading will reveal. Comparing against a reference artefact of known value exposes offset. Repeating that comparison over weeks exposes drift. Checking near the top and bottom of the working range exposes systems that are accurate in the middle and poor at the extremes, which is exactly where acceptance decisions get made. A gauge that is fine at nominal and biased near the limits will pass parts that should be rejected, quietly, on the only readings that mattered.

When the measurement swallows the tolerance

The practical result of a study is a comparison between measurement variation and the tolerance being judged. Where the measurement consumes a large slice of that band, the station cannot separate acceptable parts from unacceptable ones and everything built on its data is unreliable, including capability figures and process adjustments. There are only three honest responses: improve the measurement, widen the tolerance if engineering agrees it was never functionally necessary, or accept that some good parts will be scrapped and some bad ones shipped, and price that in. Carrying on while quoting the numbers as though they were sound is the option most often chosen.

Judgement checks can be studied too

Visual and tactile checks are usually excluded from this work because they produce verdicts rather than values, but they can be examined with the same logic. Assemble a set of parts spanning clearly good, clearly bad and genuinely marginal, have several inspectors judge them independently and repeatedly, and compare their answers to each other and to a known truth. The results are typically sobering, particularly around the marginal parts, and they make the case for boundary samples far better than an argument does. The exercise also identifies which inspectors need help and which criteria need rewriting.

Study the system that is actually used

Studies conducted in a quiet room, by the engineer, using carefully cleaned parts, describe a measurement system that does not exist on the floor. To be worth anything the work has to use the real gauge, the real fixture, the people who do the job, parts in their production condition and the conditions of the shift. Selecting parts matters as well: they should span the range the process genuinely produces, because a set of near-identical pieces makes any system look poor while a deliberately wide spread makes a mediocre one look competent.

Frequently asked questions

When should a measurement study be done?

Before a gauge is used for acceptance decisions on a new part, after any change to the gauge, fixture or method, and whenever data starts telling a story nobody can explain. The last of these is the most valuable trigger. Processes that appear to jump about without a cause, or capability that varies between shifts with no process difference, are classic signatures of a measurement problem being mistaken for a manufacturing one.

Can a study be done with only one inspector?

It can, and it will tell you about the equipment while telling you nothing about the differences between people, which is frequently the larger contributor. If only one person will ever perform the check that is a reasonable limitation, but few plants stay in that position for long once holidays and shift cover are considered. Where several people do the work, including them all is the point of the exercise rather than an optional refinement.

What if we cannot improve the measurement system?

Then the decisions made with it need adjusting to reflect what it can do. That may mean guarding the acceptance limits so only parts proven inside are released, accepting a higher rate of good parts being rejected, or moving the acceptance decision to a different characteristic or a different stage. Whichever route is chosen, record it and tell whoever relies on the data, because capability figures and process studies calculated from that gauge carry the same limitation.