故事 · Motorola、Bill Smith 与那 1.5 个 σ Origin Story · Motorola, Bill Smith and Those 1.5 Sigmas
1980 年代，Motorola 的工程师 Bill Smith 推动了六西格玛运动。团队统计大量过程后发现一个规律： 短期内看着很稳的过程，长期跑下来均值会因刀具磨损、温漂、批次差异等缓慢漂移约 1.5σ。 为了让“能力指标”反映客户长期感受到的真实质量，他们约定：把短期 σ 水平减去 1.5 再查表。 于是理论上的 6σ 过程，长期按 4.5σ 计，恰好对应 3.4 DPMO —— 这个数字从此成了六西格玛的图腾。 它简洁、好记，却也因“1.5 凭什么是固定值”而争论至今 —— 真实漂移因过程而异，1.5 只是一个保守的经验通约。 In the 1980s, Motorola engineer Bill Smith launched the Six Sigma movement. After surveying many processes, the team spotted a pattern: processes that looked rock-steady in the short run actually drifted in the long run — tool wear, temperature swings, batch-to-batch variation pushing the mean by roughly 1.5σ. To make capability metrics reflect what customers actually experienced over months and years, Motorola adopted a rule of thumb: subtract 1.5 from the short-term sigma level before looking up the defect rate. A textbook 6σ process becomes a long-term 4.5σ process, which lands neatly on 3.4 DPMO — and that number became Six Sigma's totem. It's clean, memorable — and to this day, hotly debated, because real drift varies by process. The 1.5 is just a deliberately conservative empirical convention.

1 居中的蓝峰 vs 漂移 1.5σ 的橙峰 Centered Blue Peak vs the 1.5σ-Shifted Orange Peak

σ 6.0

2 每升一个 σ，缺陷数断崖式下跌 Every Extra Sigma Drops Defects Off a Cliff

长期 DPMO 随 σ 水平呈指数级下降（纵轴对数刻度）：3σ ≈ 6.7 万、4σ ≈ 6210、5σ ≈ 233、6σ ≈ 3.4。每多一个 σ，缺陷几乎缩小一个数量级 —— 这就是为什么从“四个 9”冲到“六个 9”如此艰难又如此值钱。 Long-term DPMO falls exponentially with sigma level (note the log y-axis): 3σ ≈ 66,800; 4σ ≈ 6,210; 5σ ≈ 233; 6σ ≈ 3.4. Each extra sigma shrinks defects by roughly an order of magnitude — which is why climbing from "four nines" to "six nines" is so brutally hard, and so commercially valuable.

3 现实里的 DPMO 与西格玛水平 DPMO & Sigma Level in the Real World

DPMO 通用：以“机会”为分母，能横向比较订单录入、焊点、装配等完全不同的流程。 DPMO travels well: with "opportunities" as the denominator, you can compare order entry, solder joints, and final assembly on the same scale.
3.4 DPMO：六西格玛的目标值，对应长期 4.5σ；它是“百万分之 3.4”，不是“零缺陷”。 3.4 DPMO: Six Sigma's target, equivalent to a long-term 4.5σ. It's "3.4 per million" — not "zero defects".
长期漂移：1.5σ 是经验通约，真实漂移因过程而异，须用控制图实测而非盲目套用。 Long-term drift: 1.5σ is an empirical convention. Actual drift varies by process — measure it with control charts, don't just paste it on.
良率换算：DPMO 良率 σ 水平三者可互查，是项目“说同一种语言”的公共度量。 Yield conversions: DPMO, yield and sigma level all interconvert — the shared lingua franca that lets a Six Sigma project speak one language.
一句话In One Line
σ 水平回答的是“规格离均值有几个标准差”。居中时，6σ 的尾部只有十亿分之二，近乎完美。 可 Motorola 不信“永远居中”这件事 —— 长期总会漂。于是他们把分布整体平移 1.5σ， 再去数越界的尾巴：6 − 1.5 = 4.5σ 的尾部 ≈ 3.4 DPMO。所以这个图腾数字的真身是： 「以短期能力命名(6σ)，按长期表现计费(4.5σ)」。 理解了这一点，你就不会再把 3.4 当成精确的物理常数 —— 它是一个让全公司朝同一方向努力的、刻意保守的工程约定。 Sigma level answers one question: how many standard deviations sit between the mean and the spec? When centered, the tail beyond 6σ is two-in-a-billion — near perfection. But Motorola never believed in "always centered" — real processes drift. So they slid the whole distribution by 1.5σ and then counted the spilled-over tail: 6 − 1.5 = 4.5σ tail ≈ 3.4 DPMO. The true identity of this totem number is: "named by short-term capability (6σ), billed by long-term performance (4.5σ)". Once you see that, you stop treating 3.4 as a precise physical constant — it's a deliberately conservative engineering convention that gets the whole company pulling in the same direction.
常见误用Common Mistakes
把 1.5σ 当成普适常数硬套。它是经验通约，实际漂移须用控制图测量，不同过程差别很大。 Treating 1.5σ as a universal physical constant. It's an empirical convention — measure your actual drift with control charts; different processes vary widely.
混用短期 / 长期 σ 水平。报告时注明是含偏移（长期）还是居中（短期），否则相差 1.5σ 量级。 Mixing short-term and long-term sigma levels. Always state whether the figure includes the shift (long-term) or is centered (short-term) — they differ by a full 1.5σ.
机会数随意定，灌水 DPMO。机会须是有意义、可独立出错的环节，乱加机会会人为压低 DPMO。 Padding the opportunity count to deflate DPMO. Opportunities must be meaningful, independently failable steps — inflating them artificially drives DPMO down.

DPMO 与西格玛水平