故事 · 休哈特画的那两条线,与 Nelson 的补刀
Story · Shewhart's Two Lines and Nelson's Follow-Up
1924 年,贝尔实验室的 Walter Shewhart 在一张备忘录上画下控制图 —— 中心线加上下两条控制限。
他要回答管理者一个永恒的难题:这次波动,到底是过程本来就会有的随机抖动,还是真出了问题?
乱调一个本就受控的过程(休哈特叫它「过度调整」),只会让波动变得更大。控制限就是那条「该不该动手」的客观判据。
可仅凭「出不出界」太迟钝了 —— 过程可能早就偷偷漂移、趋势、周期,却始终卡在界内。
1984 年,Lloyd Nelson 在《质量技术杂志》上发表了 8 条判异规则,把这些「界内的异常模式」一一编码。
从此控制图不只看「有没有冲线」,更看「点子排得随不随机」。
In 1924, Walter Shewhart at Bell Labs sketched the first control chart on a one-page memo —
a center line and two control limits above and below. He was answering a question every manager faces:
is this swing just the normal random jitter of the process, or did something actually go wrong?
Tampering with a process that was already in control (Shewhart called it "over-adjustment") only widens the variation.
The control limits are the objective rule for "act, or leave it alone." But "did a point cross the line?" is a coarse signal —
a process can drift, trend, or cycle for a long time without ever leaving the band. In 1984,
Lloyd Nelson published eight rules in the Journal of Quality Technology that codify these "in-limit anomalies" one by one.
From then on a control chart was no longer just "did a point break through?" — it became "are the points arranged in a random way at all?"
1 控制图:界内的异常也逃不掉 Control Chart: In-Limit Anomalies Have Nowhere to Hide
受控In control2 两种波动,两种对策 Two Kinds of Variation, Two Responses
控制图把 ±3σ 切成 A / B / C 三个区带(每带 1σ 宽)。Nelson 规则正是基于点落在哪个区带、连续几个、朝哪个方向来判异。界内随机 = 常见原因,别动它;出界或成模式 = 特殊原因,去查它。 The chart slices the ±3σ band into three zones — A / B / C, each one σ wide. Nelson's rules trigger based on which zone a point lands in, how many points in a row, and which direction they go. In-limit and random = common cause — leave it alone; out of limit or a clear pattern = special cause — go investigate.
3 现实里的 SPC 与 Nelson 规则 SPC and the Nelson Rules in the Real World
SPC 控制图:实时把过程波动画成图,让一线就能判断「该不该动手」,是六西格玛 DMAIC「控制」阶段的主力工具。
SPC control chart: plot process variation in real time so the operator on the floor can decide "act or wait." It is the workhorse tool of the Control phase in DMAIC.
常见 vs 特殊:界内随机抖动是常见原因(系统固有),乱调它只会更糟;成模式或出界是特殊原因,必须查根因。
Common vs special: in-limit random noise is common-cause variation (built into the system) — tamper with it and you make it worse. A clear pattern or an out-of-limit point is special cause — investigate the root cause.
Nelson 8 规则:点出界、连续同侧、持续趋势、上下震荡、逼近控制限…… 它们捕捉「界内却不随机」的早期警报。
Nelson's 8 rules: out of limit, long run on one side, steady trend, alternating, hugging the control limit — they catch the early warnings that hide between UCL and LCL.
模式识别:不同模式指向不同根因 —— 趋势多是磨损 / 老化,周期多是换班 / 温控,跳变多是换料 / 换设备。
Pattern recognition: different patterns point to different root causes — trends usually mean tool wear or aging, cycles mean shift changes or thermal control, sudden jumps mean a material change or a swapped machine.
一句话In One Line
SPC 的全部智慧,浓缩成一句话:先分清波动是「天生的」还是「出事的」,再决定动不动手。
控制限解决了「该不该调」—— 别对常见原因过度反应。而 Nelson 规则解决了「没出界也可能出事」——
连续同侧暗示均值偏了,持续趋势暗示在磨损,规律震荡暗示有周期性外因。不出界 ≠ 随机,不出界 ≠ 受控。
一张会读模式的控制图,能在缺陷真正冲破规格之前,就替你提前拉响警报 —— 这正是「控制」二字的真义。
All of SPC boils down to one line: first decide whether the variation is "built in" or "something happened," then decide whether to act.
Control limits answer "should I adjust?" — do not overreact to common cause. The Nelson rules answer
"can it still be wrong even if no point breaks the limit?" — a long same-side run hints the mean has shifted;
a steady trend hints at tool wear; a rhythmic oscillation hints at a cyclical external factor.
Inside the limits ≠ random; inside the limits ≠ in control.
A control chart that reads patterns sounds the alarm long before defects break through the spec — that is the real meaning of "Control."
常见误用Common Mistakes
只看出不出界,忽略界内模式。同时套用 Nelson 规则,连续同侧 / 趋势 / 震荡都要查。
Watching only for points outside the limits and ignoring in-limit patterns. Run the Nelson rules in parallel — flag same-side runs, trends, and alternating patterns too.
对常见原因波动过度调整。受控随机的抖动别动它,乱调(追逐噪声)只会放大波动。
Over-adjusting in response to common-cause variation. Leave in-limit random jitter alone — chasing noise only widens it.
把规格限当控制限混用。控制限来自过程自身数据(±3σ),规格限来自客户需求,两者不可互换。
Confusing specification limits with control limits. Control limits come from the process itself (±3σ); specification limits come from the customer. They are never interchangeable.