故事 · 当样本只有一个的时候
Story · When the Sample Is a Sample of One
Walter Shewhart 在贝尔实验室发明控制图时,用的是子组:每次抓 4 - 5 个样本,靠组内极差估波动。可工程师很快遇到难题 ——
有些过程天生没有子组:一整炉钢只测一次成分、一天只取一个 pH 读数、一批昂贵试剂只够做一次化验。
于是 I-MR 图应运而生:既然没法在同一时刻取多个样,就退而求其次,用相邻两个时刻之差(移动极差 MR)来估计过程的短期波动。
那个常数 2.66 = 3 / d₂(n = 2 时 d₂ ≈ 1.128),把 MR̄ 还原成标准差再乘 3,于是单值数据也能画出标准的 ±3σ 控制限。
这让 SPC 走出了批量制造,进入化工、制药、环境监测这些"一次一个值"的领域。
When Walter Shewhart invented the control chart at Bell Labs, he relied on subgroups: grab 4 to 5 samples at a time and estimate variation from within-group range. But engineers ran into a wall —
some processes simply cannot be subgrouped: a whole heat of steel yields one composition assay, a day produces one pH reading, an expensive reagent batch supports only one test.
Enter the I-MR chart. If you can't pull multiple samples at the same moment, fall back on the next best thing: use the difference between consecutive time points (the moving range MR) to estimate short-term variation.
The constant 2.66 = 3 / d₂ (with d₂ ≈ 1.128 at n = 2) converts MR̄ back into a standard deviation and multiplies by three — giving you standard ±3σ limits even for single measurements.
This is what took SPC out of mass manufacturing and into chemicals, pharma, and environmental monitoring — the world of one-value-at-a-time data.
1 I 图:每个单值对比 ±2.66·MR̄ 控制限 I Chart: Each Individual Tested Against ±2.66·MR̄ Limits
受控in control2 MR 图:相邻两点之差,监控波动本身 MR Chart: Consecutive Differences, Monitoring Variation Itself
MRi = |xi − xi−1|。MR 图的 中心线 = MR̄、UCL = 3.267·MR̄(无下限)。 它专门盯波动有没有突然变大 —— 这是 I 图看不出来的另一层信号。 MRi = |xi − xi−1|. The MR chart's centerline = MR̄ and UCL = 3.267·MR̄ (no lower limit). Its sole job is to catch sudden inflation of variation — a second layer of signal that the I chart cannot see.
3 现实里的 I-MR 图 The I-MR Chart in the Real World
单值数据:每批一个化验值、每天一个读数,没法分子组的过程,I-MR 是首选 SPC 工具。
Individual-value data: one assay per batch, one reading per day — when subgrouping isn't possible, I-MR is the go-to SPC tool.
化工 / 批量过程:一炉钢成分、一锅料的黏度、一次反应的产率,天然是"一批一个值"。
Chemicals & batch processes: a steel heat's composition, a kettle's viscosity, a reaction's yield — by nature one value per batch.
移动极差:用相邻两点之差估短期 σ,是没有子组时唯一可行的波动估计法。
Moving range: using consecutive differences to estimate short-term σ is the only viable variation estimator when subgroups are unavailable.
低频测量:环境监测、设备日检、月度指标这类采样稀疏的场景,I-MR 让 SPC 仍然可用。
Low-frequency sampling: environmental monitoring, daily equipment checks, monthly KPIs — wherever sampling is sparse, I-MR keeps SPC alive.
一句话In One Line
控制图的核心永远是同一句话:中心线 ± 3σ。难点只在于怎么估这个 σ。
X̄-R 图靠组内极差,可一次只能测一个值时根本没有"组内"。I-MR 图的巧思是:把相邻两个时刻之差当作一个虚拟的 n = 2 子组,
用 MR̄ / d₂ 还原出短期标准差,于是 I 图控制限就变成 x̄ ± (3/d₂)·MR̄ = x̄ ± 2.66·MR̄。
这就是为什么单值数据也能做 SPC —— 你没有横向的样本,但时间序列本身提供了纵向的波动信息。
Every control chart says the same thing: centerline ± 3σ. The only real question is how to estimate that σ.
The X̄-R chart leans on within-group range — but when each time point gives only one value, there's no "within group" to lean on. The I-MR chart's elegant move: treat the difference between consecutive time points as a virtual n = 2 subgroup,
then use MR̄ / d₂ to recover short-term standard deviation, so the I-chart limits become x̄ ± (3/d₂)·MR̄ = x̄ ± 2.66·MR̄.
That's why SPC works on individual-value data — you don't have cross-sectional samples, but the time series itself supplies the longitudinal variation you need.
常见误用Common Mistakes
有子组也用 I-MR。能分子组就用 X̄-R / X̄-s,它们对均值漂移更灵敏;I-MR 是"没办法分组"时的退路。
Using I-MR when subgroups are available. If you can subgroup, use X̄-R / X̄-S — they detect mean shifts much faster. I-MR is the fallback when subgrouping is impossible.
数据强烈自相关还硬画 I-MR。相邻点高度相关(如温度连续记录)会让 MR̄ 偏小、控制限假性收窄,要先建时间序列模型。
Applying I-MR to strongly autocorrelated data. When consecutive points are highly correlated (e.g. continuous temperature logs), MR̄ collapses and the limits shrink artificially — fit a time-series model first.
只看 I 图忽略 MR 图。MR 图专抓波动突变;只盯 I 图会漏掉"均值没变但波动变大"的失控。
Reading the I chart and ignoring the MR chart. MR is built to catch sudden inflation of variation. Watch I alone and you'll miss the "mean steady, spread blowing up" failure mode.