Measurement and indicators

MSA — measurement system analysis

In French: MSA — analyse du système de mesure

Measurement system analysis checks that the instrument and the method are reliable enough for the data to be used to decide. Without it, every analysis rests on figures whose quality is unknown.

What is measured when measuring

Every value recorded mixes two things: the real variation of the part, and the variation introduced by the measurement system itself. If the second is of the same order as the first, the data describes the instrument as much as the process.

Measurement variation breaks down into two parts. Repeatability is the spread obtained when the same person measures the same part several times — that is the instrument. Reproducibility is the gap between several people measuring the same parts — that is the method, the training, the interpretation.

The study that separates them is called gauge R&R. It typically requires about ten parts covering the real range, two or three operators, and two or three measurements per part and per operator — the parts being presented in a shuffled order so that nobody remembers their previous value.

The other properties of a measurement system

Accuracy (bias) is the gap between the mean of the measurements and the true value of a reference standard. An instrument can be very repeatable and systematically offset: it always gives the same answer, and it is wrong.

Linearity is the constancy of that accuracy over the whole range: an instrument accurate in the middle of its range may drift at the extremes.

Stability is its constancy over time, which metrological follow-up monitors through repeated measurements on a reference standard.

The case of attribute inspection

When the check is a judgement — conforming or not, appearance acceptable or not — the same question arises and is handled differently: the same parts are examined several times by several inspectors, and the verdicts are compared with one another and with an established reference.

The results are often worse than those of physical measurements, and they surprise: it is not rare for the same inspector to contradict themselves on a borderline part from one day to the next. That is what justifies physical reference samples — the part that shows the limit of the acceptable is worth more than the sentence that tries to describe it.

The share of variation due to measurement

%R&R = (measurement variation / total variation) × 100

The customary benchmarks: below 10 %, the system is judged acceptable; between 10 and 30 %, acceptable depending on the stakes and the cost of an improvement; above 30 %, unusable for deciding. These thresholds are professional conventions, not normative values — what counts is the decision one wants to take with the measurement.

The common mistakes

Doing the study on parts that are too similar. The measurement share is compared with the total variation: on ten near-identical parts, that variation is tiny and the result will be catastrophic whatever the instrument. The parts must cover the range actually encountered.

Letting the operators see the previous value, or measuring the parts in order. The reproducibility measured then becomes the ability to remember, and the system looks far better than it is.

Considering that a calibrated instrument is validated. Calibration addresses accuracy, not spread nor the gaps between operators — a perfectly calibrated instrument can be unusable in gauge R&R.

On the shop floor — Precision engineering

A characteristic judged unstable for months was the subject of a gauge R&R study: measurement accounted for 42 % of the total variation. The process was in fact consistent; it was the way the part was clamped during inspection that varied. None of the actions launched on the machine could have succeeded.

The other tools illustrated by this trade: machining, mechanical engineering and finishing.

Frequently asked questions

When should a gauge R&R study be done?

Before any data collection intended for a decision: start of the measurement phase of a project, commissioning of an inspection device, or as soon as a process looks unstable with no identified cause. It is the first thing to check when the data tells no coherent story.

How many parts and operators?

The most widespread practice retains ten parts, three operators and three repetitions, i.e. ninety measurements. One can go down to two operators and two repetitions when availability requires it, accepting a less precise estimate. What is not negotiable is the representativeness of the parts.

What should be done if the %R&R is poor?

Look first at the breakdown: poor repeatability points to the instrument or the fixture, poor reproducibility points to the method and the training. The second is often corrected without investment — a precise procedure and half a day of training are frequently enough.

Does this apply to administrative data?

Yes, and it is rarely done. Two people coding the same complaint, two ways of dating the start of a file, two interpretations of a rejection reason: the "measurement" variation there is often enormous, and it explains a good share of inconsistent indicators.