起源故事 · 田口的瓷砖窑
Origin Story · Taguchi's Tile Kiln
战后,田口玄一(Genichi Taguchi)投身重建日本制造业。一家瓷砖厂受困于尺寸波动:窑内温度不均,靠窑壁的砖和中心的砖尺寸差很多。
按常理该改造窑炉(昂贵)。田口却建议只调配方 —— 增加石灰石含量,让瓷砖对温度波动不那么敏感。
结果几乎零成本就把废品率砍了下来。这就是稳健设计的精髓:与其消灭噪声,不如让产品无视噪声。质量损失函数和信噪比,是把这个理念量化的工具。
After the war, Genichi Taguchi threw himself into rebuilding Japanese manufacturing. A tile factory was bleeding money over dimensional drift:
the kiln baked unevenly, so tiles near the wall came out a different size from tiles in the middle. The textbook fix was to rebuild the kiln — expensive.
Taguchi proposed something else: just tweak the clay recipe — add more limestone so the tile body was less sensitive to temperature swings.
Scrap rate collapsed at near-zero cost. That's the heart of robust design: don't kill the noise — design a product that doesn't care. The loss function and S/N ratio are the math that turns that mindset into numbers.
1 同样的噪声,陡区翻江倒海,平区纹丝不动 Same Noise — Chaos in the Steep Zone, Calm in the Flat
陡区Steep zone2 质量损失函数:偏一点点,损失是平方级 Quadratic Loss: Drift a Little, Pay a Lot
田口认为:只要偏离目标就有损失,且损失 L = k·(偏差)² 是抛物线 —— 不是"在规格内就零损失"的台阶。输出波动越大,落在抛物线两翼的概率越高,平均损失就越大。稳健点同时压低波动和损失。 Taguchi's claim: any deviation from target carries a cost, and the cost is a parabola L = k·(deviation)² — not the step-function "zero loss inside spec, full loss outside". Wider output spread → more mass on the parabola's flanks → higher average loss. The robust point shrinks spread and loss at the same time.
3 现实里的稳健设计 Robust Design in the Wild
对环境不敏感:电路在 -20℃ 到 60℃ 都稳定工作,靠选对工作点而非恒温箱。
Environment-insensitive: a circuit that holds spec from −20 °C to 60 °C — by picking the right operating point, not by buying a climate box.
对批次不敏感:田口瓷砖窑 —— 调配方让产品扛住原料和温度的批次波动。
Batch-insensitive: Taguchi's tile kiln — recipe tweaks let the product shrug off raw-material and temperature variation between batches.
参数设计:先用控制因子把波动压到最小,再单独调"调整因子"把均值挪到目标。
Parameter design: first use control factors to crush the spread, then nudge the mean onto target with a separate adjustment factor.
容差设计:稳健做到位后,对真正敏感的少数零件再收紧容差 —— 把钱花在刀刃上。
Tolerance design: once the design is robust, tighten tolerances only on the few parts that still matter — spend the money where it actually buys quality.
一句话
In One Line
西方质量观长期是"在规格内就合格",田口把它推翻:偏离目标就有损失,损失按偏差的平方增长。
于是优化目标从"别超界"变成"把波动压到最小"。而压波动最省钱的办法,不是花钱削噪声,是利用响应曲线的非线性,
把工作点挪到平坦区 —— 用信噪比 S/N 量化"稳不稳",谁的 S/N 高谁就赢。这是质量从"检验"走向"设计"的分水岭。
For decades the Western view was "inside spec = good enough." Taguchi tore it up: any deviation from target costs you, and the cost grows as the square of the deviation.
The optimization target flips from "don't cross the line" to "crush the spread." And the cheapest way to crush spread is not to pay to suppress noise — it's to exploit the nonlinearity of the response curve,
sliding the operating point into the flat region. The S/N ratio turns "how steady" into a number — highest S/N wins. This is the watershed where quality stops being inspection and starts being design.
常见误用
Common Mistakes
认为"在规格内就零损失"。损失是抛物线,越偏离目标损失越大,并非到边界才突然出现。
Believing "inside spec means zero loss". Loss is a parabola — the further from target, the higher the cost. It doesn't suddenly appear at the spec line.
先调均值再想稳健。田口两步法:先用控制因子最大化 S/N 压波动,再用调整因子对准目标。
Centering the mean before thinking about robustness. Taguchi's two-step: first use control factors to maximize S/N and shrink the spread, then use an adjustment factor to slide the mean onto target.
一上来就花钱削噪声。先免费地把工作点挪到平坦区,稳健做尽了再谈收容差、降噪。
Throwing money at noise reduction first. Slide the operating point into the flat zone for free. Only after the design is as robust as it can be do you talk about tighter tolerances or quieter inputs.