2026 版 SPC 手册把设备验收指标统一为 Pm/Pmk，但很多工厂还在用老习惯报 Cpk，报告一摊开就露馅。设备验收的样品怎么抽、公式怎么算、门槛守多少？这篇把抽样策略、样品数量、计算公式一次讲清。
▌ Pm/Pmk 评的是设备，不是过程
设备验收时，要把人、料、法、环尽量固定，让波动主要来自设备本身。这就是两组指数的分工：
- Pm/Pmk → 设备固有性能，只看设备；
- Cpk/Ppk → 完整过程能力，人机料法环全参与。
设备性能决定过程能力的上限——设备不行，后面怎么调都白搭。所以顺序必须是先验收设备，再谈过程能力；顺序反了，报告再漂亮也没用。
▌ 50 件怎么抽：连续、固定、一次取完
典型做法是连续生产 50~100 件，一次性取完，同时守住四个条件：
- 同一名熟练操作工；
- 同一批次原材料；
- 中途不换刀、不调参数；
- 设备预热充分、环境稳定。
多型腔设备要特别注意：每个型腔单独取样、单独计算，不能混在一起平均——平均会掩盖单个型腔的异常。抽样条件写进验收文件，比事后解释重要得多。
▌ 50 件是底线，少了要抬门槛
样本越少，估计的不确定性越大，门槛必须跟着抬高：
- n ≥ 50 件：Pm ≥ 2.0，Pmk ≥ 1.67；
- n = 30 件：Pm ≥ 2.35，Pmk ≥ 1.96。
关键特性守住 Pm≥2.0、Pmk≥1.67，相当于自带 5σ 余量才敢放行。抽不满 50 件、门槛还不升，验收就是自欺欺人。
▌ 公式怎么读：Pm 看散布，Pmk 看偏移
标准差用样本标准差 s = √[Σ(xi − 均值)² ÷ (n−1)]，注意不是控制图上的估算值。
- Pm = (USL − LSL) ÷ 6s：看散布宽窄。6s 就是 ±3 个标准差，覆盖约 99.7% 的产品；
- Pmk = min(USL − 均值, 均值 − LSL) ÷ 3s：看中心偏移，取上下两侧中较小的那侧。散布窄但中心偏了，Pmk 照样难看。
审核供应商验收报告时，见过太多"只抽 20 件、Pmk 照样写 1.67"的案例——设备性能不足，量产后再补救，成本翻十倍。记住三步：抽足（50 件、连续、固定条件）、算对（Pm 看散布、Pmk 看偏移）、守住（门槛随样本量动态调整）。设备验收不是走过场。
想帮团队建立规范的设备验收流程，把供应商验收报告审出真问题？欢迎联系续创信息技术咨询。
The 2026 SPC manual standardises machine acceptance on Pm/Pmk, yet many factories still report Cpk out of habit — and the report gives them away at first glance. How should the sample be taken, how are the indices calculated, and what thresholds must hold? This article covers the sampling strategy, sample size and formulas in one pass.
▌ Pm/Pmk Judges the Machine, Not the Process
During machine acceptance, people, material, method and environment are held as fixed as possible, so the variation comes mainly from the machine itself. That is the division of labour between the two index families:
- Pm/Pmk → inherent machine performance, machine only;
- Cpk/Ppk → full process capability, with man, machine, material, method and environment all in play.
Machine performance sets the ceiling for process capability — if the machine is not up to it, no amount of tuning will save you. So the order must be accept the machine first, then talk about process capability; reverse it and even the prettiest report is useless.
▌ How to Take the 50 Pieces: Consecutive, Fixed Conditions, One Run
The typical approach is a continuous run of 50–100 pieces, taken in one go, under four conditions:
- The same skilled operator;
- The same batch of raw material;
- No tool changes and no parameter adjustments mid-run;
- Machine fully warmed up, environment stable.
Multi-cavity equipment needs extra care: sample and calculate each cavity separately — never pool them together, because averaging hides a single misbehaving cavity. Write the sampling conditions into the acceptance document; it matters far more than explaining afterwards.
▌ 50 Pieces Is the Floor — Fewer Means a Higher Bar
The smaller the sample, the greater the estimation uncertainty, so the thresholds must rise accordingly:
- n ≥ 50 pieces: Pm ≥ 2.0, Pmk ≥ 1.67;
- n = 30 pieces: Pm ≥ 2.35, Pmk ≥ 1.96.
Holding critical characteristics to Pm ≥ 2.0 and Pmk ≥ 1.67 amounts to building in a 5-sigma margin before release. Take fewer than 50 pieces without raising the bar, and the acceptance is self-deception.
▌ How to Read the Formulas: Pm for Spread, Pmk for Offset
Use the sample standard deviation s = √[Σ(xi − mean)² ÷ (n−1)] — not an estimate read off a control chart.
- Pm = (USL − LSL) ÷ 6s: judges the spread. 6s is ±3 standard deviations, covering roughly 99.7% of the output;
- Pmk = min(USL − mean, mean − LSL) ÷ 3s: judges the centring, taking the weaker of the two sides. A narrow spread with an off-centre mean still produces an ugly Pmk.
Auditing supplier acceptance reports, we have seen too many cases of "only 20 pieces sampled, Pmk still written as 1.67" — insufficient machine performance, remedied after mass production, costs ten times more. Remember the three steps: sample enough (50 pieces, consecutive, fixed conditions), calculate right (Pm for spread, Pmk for offset), and hold the line (thresholds adjust dynamically with sample size). Machine acceptance is not a formality.
Want to build a disciplined machine acceptance process for your team and audit supplier acceptance reports for real problems? Contact CINO Information Technology Consulting.