起源故事 · 十几张图,怎么不迷路
Origin Story · A Dozen Charts, How Not to Get Lost
从 1924 年 Shewhart 的 X̄-R 图起,半个世纪里控制图越长越多:连续型有 X̄-R、X̄-S、I-MR,计数型有 p、np、c、u……新手最常犯的错不是不会画,
而是一开始就选错了图。质量界的应对办法很朴素 —— 把选择过程画成一棵决策树,写进每本 SPC 教材的扉页。
只要诚实回答两三个关于数据本质的问题,路径会自动收敛到唯一正确的那张图。它把"凭经验拍脑袋"变成了"照树走一遍"。
From Shewhart's 1924 X̄-R chart onward, the family kept growing for half a century: X̄-R, X̄-S, I-MR for variables; p, np, c, u for attributes. The rookie mistake isn't drawing the chart badly —
it's picking the wrong chart on day one. The quality community's fix is refreshingly plain: draw the selection process as a decision tree and paste it in the front matter of every SPC textbook.
Answer two or three honest questions about what your data actually is, and the path converges to the single correct chart. Gut-feel becomes "just walk the tree."
1 点击节点,走出一条路径 Click a Node, Trace a Path
2 七张图速查:一句话记住每张 Seven Charts Cheat Sheet: One Line Each
3 现实里的图型选择 Chart Selection in the Real World
SPC 部署第一步:上线前先用决策树确定每个监控点该挂哪张图,避免后期推倒重来。
First step of SPC rollout: walk every monitoring point through the tree before go-live, so you don't rip and replace later.
培训新工程师:决策树是最快讲清七张图区别的工具,一张图胜过十页文字。
Training new engineers: a decision tree explains the seven charts faster than ten pages of text — one picture says it all.
化工/制药:批量小、取样难,常落到 I-MR;切忌硬套需要子组的 X̄-R。
Chemicals / pharma: small batches and slow sampling almost always land on I-MR; don't force X̄-R when you can't form a real subgroup.
检验数据:出货抽检批量不一,自动收敛到 p 图或 u 图,控制限随样本量自适应。
Inspection data: outgoing lots are sampled at different sizes, so the tree naturally lands on p or u — limits flex with n automatically.
一句话In One Line
决策树之所以可靠,是因为每个分叉问的都是数据本身的客观属性,而非主观偏好:连续还是计数?子组多大?数不良品还是缺陷?
这些答案唯一确定了数据背后的概率分布(正态 / 二项 / 泊松),而控制限公式正是从分布推出来的。
所以选图本质是"认清你的数据服从什么分布"。把这一步走对,后面的控制限、判异规则才有意义 —— 选图错,全盘错。
The decision tree works because every fork asks an objective property of the data, not a personal preference: variable or attribute? Subgroup size? Defectives or defects?
Those answers pin down the underlying probability distribution (normal / binomial / Poisson) — and the control-limit formula is derived straight from that distribution.
So choosing a chart is really "naming the distribution your data follows." Get that right and everything downstream — limits, run rules, alarms — is meaningful. Get it wrong and nothing else can save you.
常见误用Common Mistakes
不管数据类型,一律套 X̄-R。先分连续/计数,计数数据用 p/np/c/u。
Defaulting to X̄-R regardless of data type. Split variable vs attribute first — attribute data belongs on p/np/c/u.
单值数据硬凑子组算 X̄-R。一次只取一个值就用 I-MR(单值-移动极差)。
Fabricating fake subgroups from individual readings to plot X̄-R. One reading at a time means I-MR (individuals and moving range).
子组大却还用极差 R。子组 ≥10 用标准差 S(X̄-S),R 在大子组下估计效率低。
Sticking with range R on large subgroups. At n ≥ 10 use the sample standard deviation S (X̄-S); R becomes inefficient on big subgroups.