故事 · 田口玄一与「损失」的新定义
Origin Story · Taguchi Rewrites What "Loss" Means
传统质检是「合格 / 不合格」的二元世界:在公差内就 OK,出界才算坏。
田口玄一对此不以为然 —— 他说,只要偏离目标值,社会就开始承受损失,
而且这损失随偏差的平方增长,不是到了公差线才突然出现。
一个勉强卡在上限的零件,和一个正中靶心的零件,质量天差地别。
更妙的是他的对策:噪声(温湿度、原料批次、磨损)很难也很贵去消除,
那就通过参数设计,让产品的关键响应对这些噪声天生不敏感 ——
这就是稳健设计。一台无论冬夏都好打火的发动机,胜过一台需要恒温车间伺候的「精密」发动机。
Classical QC lives in a binary world: inside the spec = good, outside = scrap.
Genichi Taguchi refused to buy it — he argued that society starts bearing loss the instant you drift off target,
and that loss grows with the square of the deviation, not as a step function at the spec line.
A part barely scraping the upper limit and a part dead-on target are not the same quality at all.
His remedy was even sharper: noise (temperature, humidity, raw-material batch, wear) is hard and expensive to kill —
so use parameter design to make the product's critical response naturally numb to that noise.
That's robust design. An engine that fires happily in any season beats a "precision" engine that demands a climate-controlled room.
1 损失函数抛物线 + 过程分布 Loss Function Parabola + Process Distribution
拖中心 / 方差drag mean / variance2 「勉强在公差内」 vs 「稳健对准目标」 "Barely Inside Spec" vs "Robust, On-Target"
红色:偏置 + 宽分布,虽然大体在公差内,但期望损失很高。 绿色:对准目标 + 窄分布,期望损失大幅下降。 两者都「合格」,质量却差一个数量级 —— 这正是田口要点破的盲区。 Red: biased mean + wide spread — mostly inside the spec but expected loss is high. Green: on target + narrow spread — expected loss drops sharply. Both are "in-spec", yet the quality is an order of magnitude apart — exactly the blind spot Taguchi calls out.
3 现实里的稳健设计 Robust Design in the Real World
稳健设计:让关键响应对温湿度、批次、磨损等噪声不敏感,免去昂贵的环境控制。
Robust design: make the critical response numb to temperature, humidity, batch drift, and wear — no need for an expensive climate-controlled environment.
损失函数:质量损失 ∝ 偏差²,勉强合格也有可观损失,正中靶心才最优。
Loss function: quality loss scales with deviation². Even barely-in-spec parts carry real loss — dead-on target is the only optimum.
信噪比:用 S/N 比同时衡量「准不准」和「稳不稳」,参数设计就是把它最大化。
S/N ratio: a single number that tracks "on target" and "low variation" together — parameter design simply maximizes it.
参数设计:先调可控因子让产品抗噪(不花钱),再用容差设计精修关键容差(花钱)。
Parameter design first: tune controllable factors for noise immunity (free), then use tolerance design to tighten the few tolerances that truly matter (costs money).
一句话In One Line
田口最深刻的两点:第一,质量是离目标的平方损失,不是公差内外的开关 ——
这把人们从「卡线合格就行」的惰性里拽了出来,逼着去对准靶心。
第二,便宜的稳健 > 昂贵的精密 ——
与其花大价钱把噪声压到极小(恒温车间、超纯原料),不如先通过参数设计让产品天生对噪声免疫。
先用不花钱的可控因子(配方、几何、工艺窗口)把信噪比顶上去,最后才在真正卡脖子的地方花钱收紧容差(下一页)。
顺序很重要 —— 先免疫,再花钱。
注意:S/N 比的选型(望目 / 望大 / 望小)要对得上响应类型,用错公式,优化方向就反了。
Taguchi's two deepest punches: First, quality is a squared loss around the target, not a pass/fail switch at the spec line —
that single move drags engineers out of "barely inside is fine" and forces them to aim for the bullseye.
Second, cheap robustness beats expensive precision —
instead of burning money to squeeze noise (climate rooms, ultra-pure materials), use parameter design to make the product naturally immune to noise first.
Push S/N up using free controllable factors (formulation, geometry, process window) and only then spend money tightening the tolerances that genuinely matter (next page).
The order matters — immunize first, then spend.
Caveat: the S/N variant (nominal-the-best / larger-the-better / smaller-the-better) must match the response type — wrong formula = optimization in reverse.
常见误用Common Mistakes
只要在公差内就算好。偏离目标即有平方损失,要对准靶心、收窄分布,不是卡线了事。
Treating "in-spec" as "good enough". Any drift off target carries squared loss — aim for the bullseye and narrow the spread, don't just hug the spec line.
先砸钱压噪声,再谈设计。先用不花钱的参数设计求稳健,最后才用容差设计花钱精修。
Throwing money at noise before redesigning. Use free parameter design for robustness first; spend on tolerance design only at the very end.
信噪比公式张冠李戴。望目 / 望大 / 望小要对应响应类型,用错 S/N 优化方向就反了。
Mixing up S/N formulas. Match nominal-the-best / larger-the-better / smaller-the-better to the response type — the wrong S/N flips the optimization direction.