起源故事 · 怪过程之前,先怪量具
Origin Story · Before You Blame the Process, Blame the Gauge
上世纪汽车业大规模铺开 SPC 时撞上一个尴尬真相:很多"过程失控"其实是量具在骗人。卡尺磨损、读数视角不同、操作员手法各异,
让同一个零件测出五花八门的数。1980 年代 AIAG(汽车工业行动集团)把测量系统分析(MSA)写进供应商手册,Gauge R&R 成为核心:
任何用数据做决策的环节,第一步是验证测量系统。它的洞见很朴素 —— 你测到的总波动 = 零件真实波动 + 测量系统波动,
如果后者太大,你看到的"过程"其实大半是量具的幻影。
When the auto industry rolled out SPC at scale, it ran into an awkward truth: many "out-of-control processes" were really the gauge lying. Worn calipers, different viewing angles, mixed operator habits —
all of it made the same part read differently every time. In the 1980s, AIAG (the Automotive Industry Action Group) wrote Measurement System Analysis (MSA) into its supplier manual, and Gauge R&R became its centerpiece:
anywhere data drives a decision, step one is to verify the measurement system. The insight is plain — total observed variation = true part variation + measurement-system variation;
if the second piece is too big, the "process" you think you see is mostly a mirage thrown by the gauge.
1 10 个零件 · 3 位测量员 · 各测 3 次 10 Parts · 3 Operators · 3 Trials Each
量具可信2 总波动被拆成三块:信号 vs 噪声 Total Variation Split Three Ways: Signal vs Noise
按方差比例拆解总波动:绿=零件真实波动(你想看的信号),橙=重复性,红=再现性。 橙+红就是测量系统的"噪声"。噪声占比越大,数据越不可信 —— 这正是 %R&R 衡量的东西。 Total variation broken down by variance share: green = true part-to-part variation (the signal you want), orange = repeatability, red = reproducibility. Orange plus red is measurement-system noise. The larger that share, the less trustworthy the data — and that is exactly what %R&R measures.
3 现实里的 Gauge R&R Gauge R&R in the Real World
测量系统验证(MSA):任何六西格玛项目,Measure 阶段第一件事就是确认量具靠谱。
Measurement-system validation (MSA): in any Six Sigma project, the very first task in the Measure phase is confirming the gauge is trustworthy.
新量具验收:采购新卡尺、新检具上线前做 R&R,不达标退回,别让坏量具进产线。
New-gauge acceptance: run R&R on new calipers or fixtures before they hit the line; send failures back rather than letting bad gauges into production.
检验员一致性:再现性大说明操作手法不统一,靠 SOP 培训而非换量具来改善。
Inspector consistency: high reproducibility points to mismatched technique — fix it with SOP training, not by swapping gauges.
数据可信前提:R&R 不过关,后面的 Cpk、控制图、DOE 全建在流沙上。
Prerequisite for trustworthy data: if R&R fails, every downstream Cpk, control chart, and DOE is built on quicksand.
一句话
In One Line
方差是可加的:σ²总 = σ²零件 + σ²重复性 + σ²再现性。我们真正想测量的是零件之间的真实差异,
可量具抖动和操作员偏差会往里掺噪声。%R&R = 测量噪声 ÷ 总波动 —— <10% 可放心、10~30% 看情况、>30% 必须先修量具。
重复性大就修/换量具,再现性大就统一操作手法。记住:数据可信,是一切统计工具的前提;量具不过关,再精妙的分析都是自欺。
Variances add: σ²total = σ²part + σ²repeatability + σ²reproducibility. What we really want to measure are the true differences between parts,
but gauge jitter and operator bias mix in noise. %R&R = measurement noise ÷ total variation — <10% is safe, 10–30% depends on context, >30% means fix the gauge first.
High repeatability? Repair or replace the gauge. High reproducibility? Align the operating method. Remember: trustworthy data is the prerequisite for every statistical tool; if the gauge fails, even the most elegant analysis is self-deception.
常见误用
Common Mistakes
没验量具就直接分析过程。先做 R&R,测量不过关,过程结论全不可信。
Analyzing the process before validating the gauge. Run R&R first — if the measurement fails, every process conclusion downstream is unreliable.
R&R 大就一律换量具。先分清:重复性大才修量具,再现性大要培训统一手法。
Replacing the gauge whenever R&R is high. Diagnose first: high repeatability means fix the gauge; high reproducibility means train operators on a unified method.
只看 %R&R 忽略 ndc。可区分类别数 ndc <5 时,量具分不清零件好坏,同样不合格。
Reading %R&R while ignoring NDC. When the number of distinct categories drops below 5, the gauge cannot separate good parts from bad — it fails just as surely.