You've been asked to deliver the study that releases a station. Acronyms show up (MSA, Cg, Cgk, Gage R&R, ANOVA, ndc, Cpk) and nobody explained which question each one answers.
This guide builds them from scratch. Every idea is drawn before it's named, and no term is used before it's been explained. The figures move: drag the controls and watch what happens.
You plug a board into the fixture. The station measures 4.107 V. The limit is 3.8 to 4.2. It passes. Next.
But ask yourself an uncomfortable question: if you unplug the board and plug it back in without touching anything else, will it read 4.107 again?
No. It will read 4.109, then 4.104, then 4.111. And none of those is “the right one”. A single measurement isn't data: it's data missing its context. The context is how much it moves when you repeat it.
Press the button and keep piling up measurements of the same board.
Two things show up on their own as the dots pile up. First, they cluster around a center: that's the mean, and it's your best estimate of the true value. Second, the cloud has a width. And that width is what this whole guide is about.
We need a number that describes how wide that cloud is. We could use “the largest minus the smallest”, but that depends on a single unlucky point: one odd measurement ruins the figure.
What's used instead is the standard deviation, written with the Greek letter sigma, σ. Think of it as the typical distance from a point to the center. Not the largest, not the smallest: the typical one.
It has a property that makes it useful as a mental rule. If the cloud has the bell shape it almost always has:
That's where the number you'll see everywhere comes from: 6σ —three on each side— is “practically everything your system can spit out”. When an MSA document says 6σ, it means the full width of the cloud.
Now the central idea of the whole subject, and it comes with an analogy worth remembering.
You're taking a photo with a shaky hand. The photo comes out blurry for two different reasons: because what you're photographing is moving, and because your hand shakes. Looking at the photo you can't separate one from the other: you only see the total blur.
Measuring is exactly that. The boards coming out of production really are different from each other —that's the real product variation— and on top of that your station shakes. What you record is the sum.
But widths don't add up the way you'd expect. They add in quadrature: like the legs of a right triangle. A product with a spread of 3 measured by equipment with a noise of 4 doesn't give 7, it gives 5.
Move both controls and watch the third curve, the one you actually see in your reports:
If the gray curve is much wider than the blue one, you'll reject good boards that the noise pushed outside the limit, and you'll accept bad boards that the noise pushed inside. The second one is what reaches the customer and turns into a field complaint.
Back to the camera. How do you measure only your hand's shake, without the scene? Easy: you photograph something that doesn't move. Whatever blur is left is yours.
That's exactly the Type 1 study, the first one you run and the cheapest: a single master part, measured 25 to 50 times in a row, by a single operator. Since the part doesn't change, the whole cloud belongs to the station.
Now we have the width of the noise. What do we compare it against to say whether it's acceptable? Against the tolerance: the width the design allows. If the limit runs from 3.8 to 4.2 V, the tolerance is 0.4 V.
The convention —and it's only a convention, agreed on by the industry— says that the instrument is allowed to take up at most 20% of the tolerance. That's the noise budget.
And with that you can already read the Cg formula without being scared of it:
Cg = (20% of the tolerance) ÷ (6σ of the noise)
In words: how many times your instrument's cloud fits inside the budget it's been given. If it fits exactly once, Cg = 1. The industry asks for it to fit with room to spare: Cg ≥ 1.33.
Move the noise and watch the orange cloud grow until it spills over the blue band:
And the second control, what does it do? It moves the center of the cloud. An instrument can be very repeatable and be off: it always reads 30 mV high. That's called bias, and Cg doesn't see it, because Cg only measures width.
Cgk is Cg penalized for bias. It's the honest version: an extremely consistent instrument that consistently lies isn't much use.
Bias is measured against the certified value of a calibrated master part, not against the center of the tolerance. Many people use the center because it's what they have at hand — and that only gives the same result if the nominal is exactly centered and the master sits exactly on the nominal. With asymmetric limits, the Cgk comes out wrong and the study can't be defended.
You passed Type 1. Your instrument repeats well and is centered. Are you done? No, and here comes the part almost nobody explains well.
Think of a bathroom scale that reads in whole kilograms. It's repeatable: step on it five times and it reads 70 five times. It's well calibrated. And still it's completely useless for weighing letters, because every letter will read zero.
Being repeatable isn't the same as being able to tell things apart. And telling apart is what you need: your station exists to separate good boards from bad ones.
Here's the same scene of ten real boards, measured by systems with different noise. Turn up the noise and watch the moment the bars stop being separate:
The number that sums that up is called ndc, for number of distinct categories: how many different steps your system can resolve within the product's real spread.
With ndc = 2 you can only say “big or small”. With ndc = 1 you can't say anything. The acceptance rule is ndc ≥ 5, which is the minimum for doing statistical control with that data.
The other number is %GRR: what percentage of everything you see is measurement noise rather than product. It's the same idea seen the other way around, and it's the one they'll look at first in the review.
One source of variation is still missing, one Type 1 can't see, because Type 1 has a single operator.
On a test station the operator doesn't turn a knob, but they do seat the board in the fixture, close the press and plug in the harnesses. If John presses harder than Mary, John's measurements come out different. That's real and it has to be measured.
The two R's in Gage R&R are exactly those two sources:
Repeatability — the variation when nothing changes: same board, same operator, measurements in a row. It's the equipment. You'll also see it written as EV, for equipment variation.
Reproducibility — the variation that shows up when who measures changes. It's the operator. It shows up as AV, for appraiser variation.
To separate them you need an experiment where all three sources are present and organized: 10 boards, 3 operators, 3 repeats each. Ninety measurements laid out on a grid.
And now: ANOVA —analysis of variance— is simply the procedure for sharing out the blame by looking at that grid. It isn't magic and you don't need advanced statistics to understand it. It's a three-step idea, and each step looks at a different place on the same grid.
Press each button:
Each cell has three small bars: they're the three times that operator measured that board. The three buttons above point to where each component comes from.
That's all. The three buckets add up to 100% of what you observed, and the report tells you what percentage each one took. If the equipment and the operator together take less than 10%, you have an acceptable measurement system.
There's an older method called averages and ranges that does a similar split with pencil arithmetic. It gives similar results, but it misses a case that does matter: when a specific operator struggles with a specific board. ANOVA does see it, and that's why it's the recommended method today.
If your station measures on its own and nobody handles anything —an in-line test with automatic loading— you can run a Type 3, which is the same study without the operator row. It splits the blame into two buckets instead of three. Many customers accept it; check yours before designing the experiment, because it decides how many people you need.
| Indicator | Criterion | What it means |
|---|---|---|
| Cg, Cgk | ≥ 1.33 | The instrument's noise fits comfortably in the 20% of tolerance budget, and it isn't off-center. Type 1 study. |
| %GRR | < 10% | The measurement system contributes less than 10% of what you observe. Acceptable without conditions. |
| %GRR | 10 – 30% | Acceptable with documented justification, depending on how critical the characteristic is and what improving the equipment would cost. |
| %GRR | > 30% | Not acceptable. The system has to be fixed before it's used to make quality decisions. |
| ndc | ≥ 5 | It resolves at least five steps within the product's spread. |
These two show up in every meeting and are almost always confused with MSA. They aren't MSA. MSA evaluates your instrument; Cp and Cpk evaluate the product. And they only mean something after the instrument has passed, because otherwise you don't know how much of that spread is yours.
Cp asks whether the process would fit in the tolerance: it compares the allowed width against the process width. But it doesn't look at where it sits.
Cpk does. It measures the distance from the mean to the nearest limit. That's why it's always less than or equal to Cp.
Move the process center without changing its width and watch: Cp doesn't budge, Cpk collapses.
If you get Cpk = 20, the most likely explanation isn't a perfect process: it's that your instrument doesn't have enough resolution. When every reading comes out identical down to the last reported digit, σ drops to almost zero and the Cpk shoots up. It's rounding disguised as excellence, and an audit catches it fast.
This is the mistake that sinks the most studies. The ten must cover the real production range, not be ten nearly identical golden boards. %GRR compares measurement noise against product spread: if you don't give it product spread, it comes out sky-high and the ndc drops — not because your station is bad, but because the experiment was set up wrong.
Runs go in random order and without the operator seeing their previous results. If they measure the ten in the same order three times in a row, any station warm-up gets into the data disguised as board variation.
A station doesn't have one measurement: it has dozens or hundreds. A control board with 58 numeric measurements needs 58 complete studies, each with its variance split, its verdict and its charts. Times the 90 runs in the design, that's more than five thousand values to group by board, operator and repeat.
That's what turns a conceptually simple analysis into a week of work — and what leads many stations to be released with the study of a single “representative” characteristic, which is exactly the first thing an auditor questions.
What an auditor looks for in the release package, in the order they review it:
None of the above is hard to understand. What's hard is doing it hundreds of times without making a mistake, and repeating all of it every time the sequence is modified or a fixture is changed.
TestStand already stores everything you need: every measured value, its limits, the unit's serial number, the operator, the station and the time, inside each run's XML reports. The manual work is in getting that out of the files and reorganizing it — not in the statistics.
It reads NI TestStand ATML reports, rebuilds each run as an independent unit and calculates the indicators per characteristic with the standard you choose.
What it doesn't do yet: Gage R&R with ANOVA is in development. The app already reads each run's serial number, operator and station —which is what's needed to build the study grid—; the variance split and its report are still missing. We'd rather tell you before you buy it.
Frequently asked questions
The sixteen points an auditor checks in a station release package, with the acceptance criteria for each one. No sign-up and no email: open it, print it and use it.
Open the checklistTechnical guides on test data analysis: linearity and stability, attribute agreement, and how to read a TestStand report. One every so often, when there's something to say.
No spam. You can unsubscribe from any email.
