- Measurement System Analysis asks how much of your observed variation comes from the measurement rather than the product.
- Repeatability is one person measuring the same part twice. Reproducibility is two people measuring the same part.
- Under 10% of total variation is generally acceptable, 10 to 30% is conditional, above 30% means your data is largely measurement noise.
A capability study, an SPC chart and a scrap analysis all assume the measurement is trustworthy. If a third of the observed variation is the gauge, every conclusion drawn from that data is partly fiction.
What the two Rs mean
- Repeatability. Equipment variation. The same operator, the same part, the same gauge, measured again. Spread here is the instrument.
- Reproducibility. Appraiser variation. Different operators measuring the same part. Spread here is technique, fixturing or interpretation.
Running a study
- Select around ten parts that span the range of variation you actually see, not ten good ones.
- Use two or three operators who genuinely do this measurement.
- Each measures every part two or three times, in random order, without seeing previous results.
- Label parts so operators cannot identify them, or you measure their memory.
- Analyse to separate part variation, repeatability and reproducibility.
Reading the result
- Under 10% of total variation: acceptable.
- 10 to 30%: may be acceptable depending on the importance of the characteristic and the cost of improving it.
- Over 30%: the measurement system needs work before the data is used for decisions.
- Also check the number of distinct categories, which should be five or more; fewer means the gauge cannot really tell parts apart.
When the result is poor
Look at which component dominates. High repeatability variation points at the instrument or the fixturing: worn gauge, unstable setup, insufficient resolution. High reproducibility points at method: operators locating the part differently, applying different pressure, or interpreting an ambiguous instruction differently. The second is usually cheaper to fix, through a written method and a fixture.
Attribute measurement
For pass and fail judgements, run an attribute agreement analysis instead: several appraisers judging the same set of borderline items twice, compared against a known correct answer. Disagreement on visual standards is extremely common and is usually solved with boundary samples rather than more training.