Bravo-101 Systems
Test station release checklist
Sixteen points, in the order it pays to do them, each with its acceptance criterion. Print it and take it to the floor.
01Before running a single study
Half of the validations that get redone are redone because of something on this list, not because of the study itself.
- The standard is decided and written down.AIAG MSA 4th edition or VDA 5. They are not interchangeable: they use different constants and the same data can give different verdicts. If the customer didn't specify it, ask before, not after.
- The characteristics to validate are listed.You don't validate «the station»: you validate characteristics. Usually the critical and safety ones, not the 300 measurements in the full sequence.
- Each characteristic's limits match the drawing.If the TestStand sequence has one limit and the engineering drawing has another, every tolerance percentage you calculate will be wrong.
- The instruments have current calibration.With a certificate, a date and the stated uncertainty. A study run with out-of-date equipment can't be submitted.
- The reported resolution is at least one tenth of the tolerance.The 10 % rule. If the tolerance is 0.10 V, the report must carry at least 0.01 V of resolution. Careful: what matters is the resolution that reaches the file, not the one on the instrument's front panel.
- The master part exists and has a certified value.A Type 1 study needs a reference unit with a known, traceable value. Without it there's no way to assess bias.
02The studies, in this order
Order matters. Each study assumes the previous one passed: if you run a Gage R&R on an instrument that doesn't repeat, you'll spend 90 measurements to learn something 25 would have told you.
- Type 1 study — the instrument alone.One master part, one operator, 25 to 50 consecutive measurements. It gives Cg (does it repeat?) and Cgk (does it repeat and is it centered?). Criterion: both ≥ 1.33.
- Bias against the master.The mean of those measurements against the certified value. A constant bias is corrected with an offset; what can't be corrected is the spread.
- Gage R&R — the whole system.10 parts × 3 operators × 3 repeats = 90 measurements, in random order and without the operator seeing the previous result. On a fully automated station, Type 3 (no operators).
- %GRR within criterion.< 10 % acceptable. 10–30 % acceptable with documented justification. > 30 % not acceptable. The justification for the middle range is written down, not assumed.
- ndc ≥ 5.How many distinct steps the system can tell apart within the product's real variation. With fewer than 5 you can't do statistical control with that data.
- Process capability, last.Cp and Cpk on at least 30 units per characteristic —ideally 50—. It's calculated after the Gage R&R, never before: if the measurement system isn't reliable, the Cpk measures noise.
03What gets forgotten and the auditor asks about
- Aborted runs are excluded from the calculation.A step that didn't complete usually reports a value of
0with anAbortedresult. That zero isn't a measurement, and one is enough to inflate the standard deviation hundreds of times and sink the Cpk. Check how many readings actually went into the calculation. - Steps with the same name aren't mixed.If the sequence repeats a step —copied and pasted, or inside a loop— with different limits, grouping them by name lumps together measurements of different things and produces a made-up Cpk. Group them by the step's unique identifier, not its name.
- A file with several runs was counted as several units.TestStand in append mode writes several executions into one XML. If the analysis reads them as one, the study's n is wrong and so is everything else.
- Traceability is complete in the package.Serial number of each unit, operator, station, date and time, sequence version and analysis software version. Without this the study isn't reproducible, and a study that can't be repeated isn't valid evidence.
04Acceptance criteria in one table
| Indicator | What it answers | Criterion |
|---|---|---|
| Cg | Does the instrument repeat? | ≥ 1.33 |
| Cgk | Does it repeat and is it centered? | ≥ 1.33 |
| %GRR | How much of the total is measurement noise? | < 10 % ideal · 10–30 % with justification |
| ndc | How many steps can it tell apart? | ≥ 5 |
| Cp | Does the process fit in the tolerance? | ≥ 1.33 (≥ 1.67 if critical) |
| Cpk | Does it fit and is it centered? | ≥ 1.33 (≥ 1.67 if critical) |
The most expensive mistake. A suspiciously high Cpk —15, 40, 200— is almost never an excellent process. It's insufficient resolution: if every reading comes out identical down to the last digit that reaches the file, the standard deviation drops to almost zero and the index shoots up. Before celebrating, check how many distinct values that column has.