起源故事 · 标准化的统计 Esperanto
Origin Story · Standardization, Statistics' Esperanto
1947 年起,所有概率论教材的第一章都写着这个魔法:任何正态分布 N(μ, σ²) 减去 μ 再除以 σ,
立刻变成标准正态 N(0, 1),所有分位数表 z = 1.96 / 2.58 / 3.00 瞬间通用。这就是标准化(z-transform)——
统计学的"世界语"。
Six Sigma 把它当成度量单位:规格离均值多远?答 "几个 σ"。Cpk 公式里那个 3σ 不是凭空,正是 Z = 3 → 单边超规 0.135 %。 Mikel Harry 把 Z 提到核心地位,整个 6σ 体系就是 Z 体系。 Since 1947, every probability textbook opens with the same trick: take any normal N(μ, σ²), subtract μ, divide by σ, and you instantly have the standard normal N(0, 1). The z = 1.96 / 2.58 / 3.00 quantiles in the back-of-book table now apply to every distribution on Earth. That's standardization (the z-transform) — statistics' Esperanto.
Six Sigma adopted it as a unit of measure: how far is the spec from the mean? Answer: "this many sigmas". The 3σ in the Cpk formula isn't arbitrary — Z = 3 corresponds to 0.135 % one-tail defects. Mikel Harry put Z at the center of the methodology — the entire 6σ system is, at heart, a Z-based system.
Six Sigma 把它当成度量单位:规格离均值多远?答 "几个 σ"。Cpk 公式里那个 3σ 不是凭空,正是 Z = 3 → 单边超规 0.135 %。 Mikel Harry 把 Z 提到核心地位,整个 6σ 体系就是 Z 体系。 Since 1947, every probability textbook opens with the same trick: take any normal N(μ, σ²), subtract μ, divide by σ, and you instantly have the standard normal N(0, 1). The z = 1.96 / 2.58 / 3.00 quantiles in the back-of-book table now apply to every distribution on Earth. That's standardization (the z-transform) — statistics' Esperanto.
Six Sigma adopted it as a unit of measure: how far is the spec from the mean? Answer: "this many sigmas". The 3σ in the Cpk formula isn't arbitrary — Z = 3 corresponds to 0.135 % one-tail defects. Mikel Harry put Z at the center of the methodology — the entire 6σ system is, at heart, a Z-based system.
1 原始钟形:先量距离,再除以 σ Raw Bell: Measure the Distance, Then Divide by σ
Z_min = —蓝箭头 = USL−μ 的物理距离,红箭头 = μ−LSL 的物理距离。除以 σ 之后,两条距离都变成无量纲的 Z 值——可以和任何分布比较。 Blue arrow = physical distance USL − μ; red arrow = physical distance μ − LSL. Divide both by σ and they become dimensionless Z values — directly comparable across any distribution.
2 标准化之后:N(0,1) 上的尾部 ppm After Standardization: Tail ppm on N(0,1)
原始钟形已被 Z 化为 N(0,1)。Z_USL 和 Z_LSL 两条竖线把两侧尾部"超规率"亮出来。Z 每涨 1,尾部 ppm 减少一个数量级。 The raw bell has been Z-transformed into N(0,1). The Z_USL and Z_LSL vertical lines expose the defect rate in each tail. Every +1 on Z drops tail ppm by an order of magnitude.
3 现实里的 Z 分数 Z-score in the Real World
SAT / 高考成绩报告:你的"标准分"就是 Z 分数 + 线性平移。比如 SAT 平均 1050、σ 200,你考 1450 → Z=2.0。高考一分一段表本质是 Z 表。
SAT / college-entrance score reports: your "standardized score" is just a Z-score with a linear shift. SAT mean 1050, σ 200 → a 1450 score gives Z = 2.0. The Chinese gaokao percentile table is, at heart, a Z table.
控制图 UCL/LCL:UCL = μ + 3σ、LCL = μ − 3σ 就是 Z = ±3。这是 Shewhart 1924 年定的"3σ 经验法则"。
Control chart UCL / LCL: UCL = μ + 3σ and LCL = μ − 3σ are nothing more than Z = ±3. That's Shewhart's 1924 "3σ rule of thumb".
化验单"标准差以上 2 个":医院 LDL、血压、BMI 等指标常说"超过 2 个标准差",就是 Z > 2 → 1 / 40 概率。
Lab reports flagging "more than 2 SD above": when a clinic flags LDL, blood pressure, or BMI as "2 SD high", that's Z > 2 — roughly a 1-in-40 occurrence.
半导体良率:每道工序用 Z 换算 ppm,再用 Z_bench 合成(下一页)。Z=6 → ppm ≈ 0.001,对应 6σ 标杆。
Semiconductor yield: each process step converts to ppm via Z, then steps combine through Z_bench (next page). Z = 6 → ppm ≈ 0.001 — the Six Sigma benchmark.
一句话In One Line
Z = "几个标准差"——同一把尺子量所有分布。
· Z = 3 → 单边超规 1350 ppm(Cpk = 1.0)
· Z = 4 → 32 ppm(Cpk ≈ 1.33,IATF 量产门槛)
· Z = 4.5 → 3.4 ppm(含 1.5σ 漂移后的 6σ)
· Z = 6 → 0.001 ppm(短期 6σ)
每涨 1 个 Z,ppm 少一个数量级——这就是 6σ 把目标定那么远的理由。Z 是 Cpk、Z_bench、DPMO、Sigma Level 全家的共同底座。 Z = "how many sigmas" — one ruler for every distribution.
· Z = 3 → 1350 ppm one-tail (Cpk = 1.0)
· Z = 4 → 32 ppm (Cpk ≈ 1.33, the IATF production threshold)
· Z = 4.5 → 3.4 ppm (long-term 6σ after the 1.5σ shift)
· Z = 6 → 0.001 ppm (short-term 6σ)
Every +1 on Z drops ppm by an order of magnitude — that's why 6σ sets the bar so far out. Z is the shared foundation under Cpk, Z_bench, DPMO and Sigma Level.
· Z = 3 → 单边超规 1350 ppm(Cpk = 1.0)
· Z = 4 → 32 ppm(Cpk ≈ 1.33,IATF 量产门槛)
· Z = 4.5 → 3.4 ppm(含 1.5σ 漂移后的 6σ)
· Z = 6 → 0.001 ppm(短期 6σ)
每涨 1 个 Z,ppm 少一个数量级——这就是 6σ 把目标定那么远的理由。Z 是 Cpk、Z_bench、DPMO、Sigma Level 全家的共同底座。 Z = "how many sigmas" — one ruler for every distribution.
· Z = 3 → 1350 ppm one-tail (Cpk = 1.0)
· Z = 4 → 32 ppm (Cpk ≈ 1.33, the IATF production threshold)
· Z = 4.5 → 3.4 ppm (long-term 6σ after the 1.5σ shift)
· Z = 6 → 0.001 ppm (short-term 6σ)
Every +1 on Z drops ppm by an order of magnitude — that's why 6σ sets the bar so far out. Z is the shared foundation under Cpk, Z_bench, DPMO and Sigma Level.
常见误用Common Mistakes
把 Z_USL 直接当 Cpk 报。Cpk = Z_min ÷ 3,要除以 3。Z=3 对应 Cpk=1.0,不是 Cpk=3。
Reporting Z_USL directly as Cpk. Cpk = Z_min ÷ 3, you must divide by 3. Z = 3 maps to Cpk = 1.0, not Cpk = 3.
Z 公式记成 (X − σ) / μ 或 (μ − X) / σ。永远是 (X − μ) / σ。X 在前、μ 在后、除以 σ;顺序错了正负号就反。
Mis-remembering the formula as (X − σ) / μ or (μ − X) / σ. It is always (X − μ) / σ. X first, μ second, divide by σ — flip the order and the sign flips with it.
偏态/重尾分布直接套 Z → ppm。Z → Φ⁻¹ 反查只对正态成立;偏态先做 Box-Cox 或 Johnson 变换,重尾用 Percentile(百分位)法。
Applying Z → ppm to skewed or heavy-tailed data. The Z → Φ⁻¹ lookup is valid for normal data only. Transform skewed data with Box-Cox or Johnson first; for heavy tails use the percentile method.