故事 · Shewhart 的一页备忘录
Origin Story · Shewhart's One-Page Memo
1924 年,贝尔实验室的 Walter Shewhart 给上司写了一页备忘录,画了一张带中心线和上下限的图 —— 这就是控制图的起点,也被视为现代质量管理的开端。
他洞察到一个致命陷阱:工人看到产品偏大就往小调、偏小就往大调,结果越调波动越大。因为他们把常见原因(过程固有的随机波动)误当成了需要纠正的问题。
Shewhart 用 ±3σ 控制限划出一条线:线内的起伏是过程在"正常喘气",别动它;只有越线才是真出了特殊原因,才该停机追查。
X̄ 图盯中心位置、R 图盯波动大小,两张图配合,过程的脉搏就一目了然。
In 1924 Walter Shewhart at Bell Labs handed his boss a one-page memo with a chart — a center line and a pair of limits. That sketch is the birth of the control chart, and arguably of modern quality management itself.
He had spotted a deadly trap: operators who saw a part run high would dial the machine down, saw it run low would dial it back up — and ended up amplifying the very variation they were trying to suppress. They were mistaking common-cause noise for a problem that needed correction.
Shewhart drew ±3σ control limits as a hard line: wiggles inside the line are the process "breathing normally" — leave it alone; only a point across the line is a real special cause worth stopping the line for.
The X̄ chart watches the location, the R chart watches the spread — together you take the pulse of the process at a glance.
1 X̄ 图盯中心,R 图盯波动,出界即报警 X̄ Watches the Center, R Watches the Spread, Out-of-Limits = Alarm
受控in control2 两种波动:一种别碰,一种快查 Two Kinds of Variation: One You Leave Alone, One You Hunt Down
常见原因(线内随机)
过程固有的微小波动,来自无数个无法单独识别的小因素。它稳定、可预测 —— 想减小它要改进系统本身,逐点调反而添乱。 Common cause (random, inside the limits)
Tiny noise baked into the process, the sum of countless small factors that can't be singled out. It's stable and predictable — to shrink it you have to redesign the system itself, and tweaking point-by-point only adds to it.
过程固有的微小波动,来自无数个无法单独识别的小因素。它稳定、可预测 —— 想减小它要改进系统本身,逐点调反而添乱。 Common cause (random, inside the limits)
Tiny noise baked into the process, the sum of countless small factors that can't be singled out. It's stable and predictable — to shrink it you have to redesign the system itself, and tweaking point-by-point only adds to it.
特殊原因(出界 / 趋势)
外来的、可识别的异常:换错料、刀具磨损、设备故障。它不可预测,控制图一亮红灯就该停机追根因,把它从过程里清除。 Special cause (out of limits / trends)
An identifiable intrusion from outside: wrong material loaded, tool wear, a failing motor. It's unpredictable — the moment the chart lights up, stop the line, find the root cause, and drive it out of the process.
外来的、可识别的异常:换错料、刀具磨损、设备故障。它不可预测,控制图一亮红灯就该停机追根因,把它从过程里清除。 Special cause (out of limits / trends)
An identifiable intrusion from outside: wrong material loaded, tool wear, a failing motor. It's unpredictable — the moment the chart lights up, stop the line, find the root cause, and drive it out of the process.
3 现实里的 X̄-R 图 X̄-R Charts in the Real World
SPC 监控:车间最常用的连续型质量监控,n=2~9 的小子组定时取样,实时看过程脉搏。
SPC monitoring: the workhorse chart for continuous variables on the floor — small subgroups of n = 2 to 9 sampled on a fixed cadence, taking the process pulse in real time.
A2 / D3 / D4:控制限常数都内含 d2,把短期 σ 估计直接折进 ±3σ 公式,车间不用算 σ。
A2 / D3 / D4: every control-limit constant carries d2 inside it, folding the short-term σ estimate directly into the ±3σ formula — the operator never has to compute σ.
先看 R 图:R 图不稳则波动失控,X̄ 图的控制限也不可信,所以判读顺序是先 R 后 X̄。
Read R first: if the R chart isn't stable then spread is out of control, which means the X̄ limits (built on R̄) aren't trustworthy. Always interpret R before X̄.
Westgard / Nelson 判异:除了出界,还看连续上升、连续单侧等趋势规则,更早捕捉漂移。
Westgard / Nelson rules: beyond a single point out of limits, watch for runs, trends, and one-sided streaks — they catch drift earlier than a hard out-of-limits hit.
一句话In One Line
控制图的全部智慧浓缩成一句话:把噪声和信号分开。
常见原因是噪声 —— 过程天生就在这条带子里抖,越界以内的每一次起伏都是它的正常呼吸,逐点反应只会引入新波动。
特殊原因是信号 —— 它把点甩出 ±3σ 之外,概率小到几乎不可能是偶然,所以一旦出现就值得停机追查。
X̄ 图管"中心有没有漂",R 图管"波动有没有变大",两张图一上一下,过程的位置与散布同时被盯死。
这就是为什么 SPC 既能止住"过度调整",又能在异常刚冒头时就拉响警报。
The entire idea of the control chart compresses to one line: separate the noise from the signal.
Common cause is noise — the process was born jittering inside that band, every wiggle within the limits is just it breathing, and reacting point-by-point only injects fresh variation.
Special cause is signal — it throws a point past ±3σ, an event so unlikely under noise alone that it justifies stopping the line and hunting down the root cause.
The X̄ chart asks "has the center drifted?", the R chart asks "has the spread grown?" — stacked one above the other, they pin down location and spread at the same time.
That's why SPC simultaneously stops over-adjustment and raises the alarm the moment a real anomaly pokes its head up.
常见误用Common Mistakes
把控制限当规格限。控制限来自过程自身波动(±3σ),规格限来自客户要求,二者无关,不能混用。
Treating control limits as spec limits. Control limits come from the process's own ±3σ variation; spec limits come from the customer. They are independent — don't conflate them.
线内点也逐个去调过程。受控状态下的起伏是常见原因,瞎调(过度调整)只会放大波动。
Tweaking the process on every in-control wiggle. Wiggles inside the limits are common cause; chasing them (over-adjustment) only amplifies variation.
R 图失控也照算 X̄ 控制限。R 图不稳 → R̄ 不可靠 → X̄ 限失真,必须先让 R 图受控再判 X̄。
Computing X̄ limits while the R chart is still out of control. Unstable R → unreliable R̄ → distorted X̄ limits. Get R in control first, then interpret X̄.