故事 · 田口玄一与“质量损失函数”
Origin Story · Genichi Taguchi and the "Quality Loss Function"
日本工程师 田口玄一(Genichi Taguchi)提出了一个颠覆性观点：质量不是“合格 / 不合格”的二分，而是连续的损失。
他的质量损失函数说，产品特性每偏离目标值 T 一点，社会就承担一份损失，且损失正比于偏差的平方 L = k(y − T)²。
一台略偏标准的电视、一根略偏尺寸的轴，虽然“没超规格”，却已悄悄累积了损失。1988 年 Chan、Cheng、Spiring 把这一思想写进能力指数，
提出 Cpm —— 在分母用 √(σ² + (μ − T)²) 取代单纯的 σ，于是“偏离目标”第一次被正式罚进了过程能力。
Japanese engineer Genichi Taguchi overturned the classic view of quality: it isn't a binary "pass / fail" — it's a continuous loss.
His Quality Loss Function says that every bit of drift from the target T inflicts a cost on society, and that cost rises with the square of the deviation: L = k(y − T)².
A TV that's slightly off-spec, a shaft a hair off nominal — both quietly accrue loss even though they "pass". In 1988, Chan, Cheng and Spiring baked this idea into a capability index,
introducing Cpm: replace σ in the denominator with √(σ² + (μ − T)²), and "off-target distance" finally gets formally docked against process capability.
1 在规格里，却离靶心越来越远 Still Inside Spec, Yet Drifting Off Target
Cpk 0.00 / Cpm 0.002 Cpk 说“还行”，Cpm 说“你跑偏了” Cpk Says "Good Enough", Cpm Says "You're Off Target"
Cpk 只在分布快撞规格线时才下降，Cpm 一旦偏离目标就立刻走低。居中时两者接近，偏心时 Cpm 明显更低 —— 这就是“勉强达标”与“对准靶心”的差距。 Cpk only drops once the distribution starts brushing a spec line; Cpm falls the instant you drift from target. They agree when centered and diverge when off-center — that gap is the price of "just passing" vs "hitting the bullseye".
3 现实里的田口思想 Taguchi Thinking in the Real World
目标值优先：精密装配、光学、半导体里，名义尺寸比规格上下限更重要，离靶就是隐性损失。
Target-first thinking: in precision assembly, optics and semiconductors, the nominal dimension matters more than the spec limits — every micron off target is a hidden loss.
损失函数：损失 ∝ (y − T)²，越偏越贵，这是把质量从“合格率”推向“持续逼近目标”的根基。
Loss function: cost ∝ (y − T)² — the further off, the costlier. This is what pushes quality from "yield" thinking to "always close in on the target".
Cpm vs Cpk：Cpk 容忍规格内的偏心，Cpm 不容忍；客户要求“瞄准目标”时该报 Cpm。
Cpm vs Cpk: Cpk tolerates off-center as long as it's in spec — Cpm does not. When the customer demands "aim at the target", report Cpm.
稳健设计：田口方法用 Cpm 思想指导参数设计，先把均值拉回 T，再压缩波动。
Robust design: Taguchi's parameter design uses Cpm-style thinking — first pull the mean back onto T, then squeeze the variation.
一句话In One Line
Cpk 的分母是 σ —— 它只关心“散布多宽、离最近规格线多远”。
Cpm 的分母换成 √(σ² + (μ − T)²) —— 田口把“偏离目标的距离”当成另一种散布，和真正的波动平方相加。
于是当 μ = T 时两项相同（Cpm 退回到类似 Cp 的形式），可一旦 μ 偏离 T，那个 (μ − T)² 立刻撑大分母、压低 Cpm。
一句话：Cpk 问“你出界了吗”，Cpm 问“你对准目标了吗”。 质量的最高境界不是不出界，而是稳稳钉在靶心。
Cpk's denominator is just σ — it only cares about spread width and the distance to the nearest spec line.
Cpm swaps that for √(σ² + (μ − T)²) — Taguchi treats off-target distance as another flavor of spread and adds it in quadrature with the real jitter.
When μ = T the two reduce to the same thing (Cpm collapses back to a Cp-like form), but the moment μ slips away from T, that (μ − T)² term inflates the denominator and drags Cpm down.
In one line: Cpk asks "did you cross the line?"; Cpm asks "did you hit the target?" The highest form of quality isn't staying inside the box — it's nailed to the bullseye.
常见误用Common Mistakes
规格内就不管偏心。田口认为偏离目标即有损失，对中心要求高的特性该用 Cpm 监控。
Ignoring off-center drift just because it's inside spec. Taguchi treats every off-target unit as a loss — use Cpm for any characteristic where centering matters.
把目标值默认成规格中点。T 由设计意图决定，可能不在 (LSL+USL)/2，须按图纸名义值取 T。
Defaulting T to the spec midpoint. T comes from design intent and is often not at (LSL + USL)/2 — pull it from the drawing's nominal value.
只看 Cpm 不看尾部废品。Cpm 衡量“离靶损失”，仍须配合 Cpk / PPM 评估实际超规格比例。
Watching Cpm alone and ignoring the tails. Cpm captures off-target loss — you still need Cpk / PPM to see how much actually escapes the spec.