故事 · 容差不是写给图纸看的
Story · Tolerance isn't written for the drawing — it's written for the machine
年轻工程师最爱在图纸上写紧到发抖的容差,仿佛越紧越显本事。可老师傅会反问一句:「这个尺寸,你的机器 Cpk 多少?」
设计阶段定的容差只是一张支票,验证阶段才是去兑现 —— 用实测数据算 Cpk,看产线到底能不能稳定地把零件做进这个范围。
Cpk 同时盯两件事:分布有没有居中(μ 离规格中心远不远)、波动有没有变胖(σ 大不大)。
只要其中一头失守,Cpk 就掉下 1.33,意味着量产时会有看得见的废品流出。
这就是为什么 DFSS 不允许「画完图就放行」 —— 容差和能力必须当面对账。
Young engineers love tightening tolerances on the drawing — as if narrower numbers prove they're sharper.
A veteran will fire back one question: "What's your machine's Cpk on that dimension?"
The tolerance set during design is just a cheque. Verification is when you go to cash it — by computing Cpk from real measurements
and seeing whether the line can actually park parts inside that band.
Cpk watches two things at once: is the distribution centered (how far μ sits from the spec mid-point) and is the spread too fat (how large σ is).
Lose either one and Cpk slips below 1.33 — meaning visible scrap will flow off the line in production.
That's why DFSS refuses to ship "drawing released, we're done" — tolerance and capability have to settle the bill face to face.
1 分布 vs 规格限:Cpk 实时验证 Distribution vs Spec Limits: Cpk Verified Live
达标PASS2 Cpk 怎么算,门槛怎么读 How Cpk Is Computed and How to Read the Threshold
公式:
Cpk = min(USL−μ, μ−LSL) / 3σ。取离中心较近那侧规格限的距离,除以 3σ —— 木桶由最短的板决定。
Formula: Cpk = min(USL−μ, μ−LSL) / 3σ. Take the distance to the closer spec limit and divide by 3σ — a barrel is sized by its shortest stave.
居中性:μ 偏离规格中心,min 取到的就是更近那侧,Cpk 被拉低 —— 偏心比单纯变胖更隐蔽地伤能力。
Centering: shift μ off the spec mid-point and "min" grabs the closer side — Cpk drops. Off-centering hurts capability more insidiously than just fatter spread.
1.33 门槛:Cpk = 1.33 约对应 ±4σ 余量、63 PPM;1.0 只有 ±3σ、2700 PPM;1.67 接近六西格玛水平。
The 1.33 threshold: Cpk = 1.33 buys about ±4σ of margin and ~63 PPM; Cpk = 1.0 leaves only ±3σ and ~2,700 PPM; Cpk = 1.67 is approaching Six Sigma territory.
容差 vs 能力:容差是分子的天花板,能力(σ)是分母。两者必须匹配 —— 紧容差配胖过程,Cpk 必崩。
Tolerance vs capability: tolerance is the ceiling on the numerator, capability (σ) is the denominator. They must match — pair a tight tolerance with a fat process and Cpk will crash.
3 现实里的 Cpk 验证 Cpk Verification in the Real World
汽车零部件:PPAP 提交几乎都要求关键尺寸 Cpk ≥ 1.33,安全件甚至要 1.67,否则不放量产。
Automotive parts: nearly every PPAP submission demands Cpk ≥ 1.33 on key characteristics — and ≥ 1.67 on safety items. Miss it and production is not released.
容差-能力匹配:设计端发现 Cpk 不够,先问「容差能否合理放宽」,再谈「要不要升级设备」 —— 别只逼工艺。
Tolerance capability balance: when Cpk falls short, design should first ask "can the tolerance be loosened without hurting function?" before "do we need a better machine?" — don't only squeeze process engineering.
1.33 门槛:客户验收、首件鉴定、过程审核普遍以 1.33 为及格线,把它当成设计放行的硬性闸门。
The 1.33 gate: customer acceptance, first-article qualification and process audits all treat 1.33 as the pass mark — a hard gate before the design is released.
设计验证:用试产真实数据而非仿真去算 Cpk,是 DFSS 验证阶段「图纸兑现」的关键一步。
Design verification: compute Cpk from real pilot-production data — not simulation — that's how DFSS turns the drawing into a delivered process.
一句话In one line
Cpk 的厉害之处,是它把两件原本各说各话的事 ——「设计想要多严」和「产线能做多稳」——
压进了同一个数字里。容差是分子,过程波动是分母,谁也别想单方面赢:
容差画得再漂亮,σ 一胖,分子那点余量瞬间被吃光;过程再稳,μ 一偏心,min 就咬住离得近的那侧,照样把 Cpk 拖下水。
所以 DFSS 的验证阶段从不接受「设计说没问题」这种空头支票,它要的是实测数据兑现的 Cpk。
但也别把 1.33 当成神谕:它只是个工程惯例,背后假设了分布近似正态、过程已经受控。
分布若严重偏态、或样本根本不稳,Cpk 这个数字本身就不可信 —— 先确认过程稳了,Cpk 才有意义。
What makes Cpk powerful is that it forces two arguments that used to talk past each other —
"how tight the designer wants it" and "how steady the line can hold it" — into a single number.
Tolerance is the numerator, process spread is the denominator, and neither side can win alone:
draw a beautiful tolerance and let σ grow, the margin in the numerator vanishes; hold the process steady but let μ drift off-center, "min" snaps to the nearer limit and drags Cpk down anyway.
That's why DFSS verification refuses the "design says it's fine" rubber-check — it wants a Cpk cashed by real measurement.
But don't worship 1.33 either: it's an engineering convention that assumes a roughly normal distribution and a process already in statistical control.
If the distribution is strongly skewed, or the sample isn't stable, the Cpk number itself isn't trustworthy — stabilise the process first; only then does Cpk mean anything.
常见误用Common mistakes
只看 Cp,忽略 Cpk。Cp 假装分布居中,Cpk 才反映偏心;偏心严重时两者天差地别。
Looking only at Cp and ignoring Cpk. Cp pretends the distribution is centered; only Cpk catches off-centering — and when the mean has drifted, the two numbers diverge sharply.
过程没受控就急着算 Cpk。Cpk 假设过程稳定,先用控制图确认受控,再谈能力。
Computing Cpk before the process is in control. Cpk assumes statistical control — confirm stability with a control chart first, then talk capability.
Cpk 不够就一味放宽容差。放宽容差可能牺牲功能;先分清是容差不合理还是过程该改进。
Loosening the tolerance whenever Cpk is short. A wider tolerance may quietly sacrifice function. First decide whether the tolerance is genuinely unreasonable, or the process is the one that needs to improve.