起源故事 · 一族"万能整形"变换
Origin Story · A Universal "Shape-Fixing" Family
1964 年,统计学家 George Box 与 David Cox 提出了以二人命名的幂变换族:用一个参数 λ 统一表达开方、取对数、取倒数等一连串变换。
它的妙处在于 —— 让数据自己"说"它该被怎么整形:通过最大似然或令偏度最小,自动选出最能把分布掰正的 λ。
在过程能力分析里这一步至关重要:先把偏态数据变正态,Cpk 公式才站得住脚,否则算出来的能力指数会骗人。
In 1964, statisticians George Box and David Cox introduced the power-transform family that now carries their names: a single parameter λ that unifies square roots, logs, reciprocals and everything between.
The beauty is that the data picks its own shape — by maximum likelihood or by minimising skewness, the method automatically selects the λ that straightens the distribution best.
In capability analysis that step is non-negotiable: turn the skew into a normal shape first, and only then does the Cpk formula hold water. Skip it and the capability number lies to you.
1 上:右偏原貌 · 下:被 λ 掰直后的样子 Top: raw right-skew · Bottom: shape after λ straightens it
原始偏态raw skew2 同一份数据,算出两个 Cpk —— 谁对? Same data, two different Cpks — which one is right?
误算 Cpk(直接套正态)和正确 Cpk(变换后)往往差出一截。右偏数据被硬算时,长尾巴会让缺陷率被低估,Cpk 虚高 —— 这正是质量事故的温床。 The naive Cpk (normal formula forced) and the true Cpk (post-transform) usually disagree by a wide margin. On right-skewed data, that long tail makes the defect rate look smaller than it is and the Cpk look better than it is — this is exactly the soil that quality escapes grow in.
3 现实里的非正态能力分析 Non-normal Capability in the Real World
寿命/失效时间:天生右偏(多数早坏、少数特长寿),直接套正态会严重错估可靠性。
Lifetime / time-to-failure: born right-skewed (most fail early, a few last forever). Force-fit a normal and you badly misjudge reliability.
杂质含量:有自然下限 0、偶有高值尖峰,是典型的右偏数据。
Impurity content: bounded below by zero, with occasional high-side spikes — a textbook right-skew.
循环时间:服务/加工耗时常右偏(大多数快、少数拖很久),适合 Box-Cox 或对数变换。
Cycle time: service and processing times skew right (most are fast, a few drag on). Box-Cox or a log transform fits naturally.
替代路线:除变换外,也可直接拟合非正态分布(如威布尔、对数正态)算百分位法 Cpk。
Alternative route: instead of transforming, fit a non-normal distribution directly (Weibull, lognormal) and compute Cpk via the percentile method (Clements).
一句话
In One Line
Cpk 用 ±3σ 估"几乎所有产品落在哪",这个换算只对正态成立。一旦数据右偏,那条长尾巴里其实藏着远比正态预测更多的超规品,
硬算会让 Cpk 虚高、缺陷率被低估,给人虚假的安全感。Box-Cox 用 y=(xλ−1)/λ 把分布掰回对称(λ=0 时取对数),
在变换后的尺度上算能力,再换算回原始单位。核心原则只有一句:先验证正态性(直方图 + 正态检验),不正态就别直接套 Cpk 公式。
Cpk leans on ±3σ to estimate where "almost every part" lands, and that shorthand only holds for a normal distribution. Once the data skews right, the long tail hides far more out-of-spec parts than a normal would predict;
force the formula anyway and Cpk comes out inflated and the defect rate underestimated — a false sense of safety. Box-Cox uses y = (xλ−1)/λ to bend the distribution back to symmetry (λ=0 collapses to a log), scores capability on the transformed scale, then maps results back to the original units.
One rule above all: verify normality first (histogram plus a formal test). If the data isn't normal, do not plug it into the Cpk formula as-is.
常见误用
Common Mistakes
不看分布就直接算 Cpk。先画直方图 + 做正态检验,确认正态再用标准公式。
Computing Cpk without checking the distribution. Plot the histogram and run a normality test first; only use the textbook formula once normality is confirmed.
对含 0 或负值的数据用 Box-Cox。Box-Cox 要求数据为正;有 0/负值改用 Yeo-Johnson 等变换。
Applying Box-Cox to data containing zeros or negatives. Box-Cox requires strictly positive values. With zeros or negatives, switch to Yeo-Johnson or a shifted transform.
只报变换后的 Cpk,规格限忘了同步变换。数据变换了,LSL/USL 必须用同一 λ 一起变换后再算。
Reporting the post-transform Cpk but forgetting to transform the spec limits. Whatever λ you applied to the data must also be applied to LSL and USL before you compute Cpk.