故事 · 在你怀疑过程之前,先怀疑尺子
Origin Story · Doubt the Ruler Before You Doubt the Process
上世纪汽车业最痛的一课是:工程师辛苦改进一道工序,数据却忽高忽低,怎么调都不见好 —— 后来才发现,抖动的不是工序,是卡尺和操作员。
于是 AIAG 把这件事系统化为 测量系统分析(MSA):任何观测值都是真值 + 测量误差,要先证明尺子值得信,再谈过程能力。
Gauge R&R 用一组零件、几位操作员各测几次,把测量误差拆成两个来源 —— 重复性(Repeatability,怪设备)和再现性(Reproducibility,怪人)。
两个 R 合起来就是 GRR,占总变异越大,说明你的尺子越没资格评判产品。测量系统是一切数据的地基,地基歪了,上面盖什么都白搭。
One of the auto industry's hardest lessons last century: engineers worked themselves ragged improving a process, yet the numbers kept jumping up and down for no obvious reason — until somebody realized it wasn't the process shaking, it was the calipers and the operators.
AIAG codified the response as Measurement Systems Analysis (MSA): every observation equals true value + measurement error, and you must prove the ruler is trustworthy before you start judging capability.
Gauge R&R takes a small set of parts and a handful of operators, has each measure each part several times, then decomposes the measurement error into two sources — Repeatability (blame the gauge) and Reproducibility (blame the people).
Add the two R's and you get GRR. The larger its share of total variation, the less right your ruler has to render a verdict. The measurement system is the foundation under every datum — if the foundation tilts, nothing built on it stands.
1 总变异拆开看:哪一块是尺子在抖? Crack Open the Total Variance — Which Slice Is the Ruler?
%GRR 0%方差是可以相加的:绿=过程、黄=重复性、红=再现性。三块叠成总方差,黄+红的高度占比就是测量误差吃掉的份额。拖动滑块看每一块怎么涨落。 Variances add up: green = part, amber = repeatability, red = reproducibility. Stack the three and you get total variance — the amber+red share is exactly what measurement error eats. Drag the sliders to see each slice grow and shrink.
2 三位操作员各测同几个件:重复 vs 再现,肉眼可见 Three Operators on the Same Five Parts — See Repeatability vs Reproducibility
横轴是 5 个零件,每位操作员用一种颜色各测 3 次。同一操作员同一件的散开就是重复性;不同操作员之间整体的高低偏差就是再现性。理想状态:点紧紧贴在各件真值上。 Five parts along the x-axis; each operator gets a color and measures every part three times. The scatter inside one operator-part cluster is repeatability; the vertical bias between operator clusters is reproducibility. Ideal world: every dot sits exactly on its part's true value.
3 现实里的 Gauge R&R Gauge R&R in the Real World
MSA 是先决条件:做 SPC、过程能力前必须先过 MSA,否则控制图上的波动可能全是尺子在抖。
MSA is a prerequisite: clear MSA before you run SPC or capability studies — otherwise every wiggle on the control chart might just be the gauge.
%GRR 判定线:<10% 接受、10～30% 视情况(看成本与重要度)、>30% 必须整改测量系统。
%GRR acceptance bands: <10% accept, 10–30% judge case-by-case against cost and criticality, >30% you must fix the measurement system before anything else.
EV 高怎么办:重复性差查设备(夹具松动、卡尺磨损、读数分辨力不足)。
High EV — what next: bad repeatability points at the gauge: loose fixtures, worn calipers, or read-out resolution too coarse for the feature.
AV 高怎么办:再现性差查人(操作规范不统一、培训不到位、读数习惯各异)。
High AV — what next: bad reproducibility points at the people: inconsistent procedure, gaps in training, different reading habits between operators.
一句话In One Line
关键在于方差可加:你看到的总变异,其实是过程变异和测量变异在方差尺度上的简单相加。
测量变异里,重复性问"同样的人和件,重测能不能复现"(设备的稳定性),再现性问"换个人量,结果一不一致"(操作的一致性)。
把它们除以总变异,就得到 %GRR —— 一个直白的体检指标:测量系统占了多少话语权。
%GRR 越小,剩给真实过程的"信号"越纯。所以六西格玛永远是:先证明尺子可信,再去改进过程,否则你优化的可能只是噪声。
The whole trick rests on variances adding: the total spread you observe is just part variance plus measurement variance, summed on the variance scale.
Inside the measurement piece, repeatability asks "same operator, same part — can the reading reproduce itself?" (gauge stability), and reproducibility asks "swap the operator — do we still get the same number?" (procedural consistency).
Divide their root-sum-square by the total and you get %Study Variance — a plain checkup score for how much of the conversation the measurement system is hogging.
The lower it goes, the cleaner the signal left for the real process. Six Sigma's rule never changes: prove the ruler before you tune the process, or you are just optimizing noise.
常见误用Common Mistakes
跳过 MSA 直接算过程能力。测量系统未验证时,Cpk 里混着尺子的抖动,改进方向会被带偏。
Jumping straight to capability without MSA. With no MSA evidence, Cpk silently absorbs the gauge's jitter and points your improvement effort the wrong way.
把标准差直接相加。是方差(σ²)相加,不是 σ 相加;σGRR = √(σ²EV + σ²AV)。
Adding standard deviations directly. Variances add, not σ's. The correct combination is σGRR = √(σ²EV + σ²AV).
%GRR 合格就万事大吉。还要看 ndc(可区分类别数)≥ 5,二者是一枚硬币的两面。
Declaring victory the moment %GRR passes. You still need NDC (number of distinct categories) ≥ 5 — %GRR and NDC are two faces of the same coin.