控制图上没有一个点子出界，客户审核却一句话把你打回来："这个过程不能报 Cpk，只能报 Ppk。"很多人当场懵了：这不都受控了吗？2026 版 AIAG-VDA SPC 手册 8.4 节把这个误区讲透了：受控不等于统计受控，关键看过程分布的均值和方差是否随时间变化。
▌ 四种模型：位置和变差随不随时间变
手册按 ISO 22514-2，用两个维度给过程分类：位置（均值）和变差（方差）是否随时间恒定。
- 模型 A：位置和变差都恒定——真正的统计受控；
- 模型 B：位置恒定，但变差随时间变化；
- 模型 C：变差恒定，但位置随时间变化；
- 模型 D：位置和变差都在变。
其中 A 和 C 还有细类：A1 表示瞬时分布为正态，A2 为非正态；C1~C3 对应位置变化的不同方式，比如刀具系统性磨损、换刀、批次更换。
▌ 谁能报 Cpk：统计受控与受控稳定是两回事
只有模型 A1、A2 代表统计受控（统计稳定），这时才适合用 Cp/Cpk。模型 B 和 D 因位置或变差发生变化，不处于统计受控，只能用 Pp/Ppk 报性能。
模型 C1~C3 属于"受控稳定"：通过回顾性控制图和定期监控，把特性值维持在控制限内，可以在受控条件下使用 Cp/Cpk——但它不是统计稳定。手册的态度很明确：从质量角度，统计稳定优于仅受控稳定；不过受控稳定运行成本更低，是困难情况下可以接受的折中方案。
▌ 模型怎么验证：先看图，再算数
- 目视检查：用概率图或直方图，看理论分布与实际数据是否匹配；
- 数值评价：用净相关系数做量化判断，避免只凭眼睛下结论；
- 抓关键区域：先评估整个数据集，再重点关注距离规格限最近的 25% 数据——这正是计算 Pmk、Ppk、Cpk 的关键区域。
注意：模型分配只描述过程随时间的分布形态，与能力或性能无关——它还没有涉及公差。
▌ 现场启示：刀具磨损正在把你推进模型 C
加工现场的刀具磨损、启用新刀、批次更换，这些难以避免的影响，常常让过程落入模型 C。有两件事必须记住：
- 改进之后如果位置或变差发生了变化，必须重新评价控制限——老控制限配新过程，等于没有限；
- 别再把"控制图没出界"当成全部标准：先判断过程属于哪种时间相关分布模型，再决定配什么指数、敢不敢写 Cpk。
记一句：控制图告诉你过程"稳不稳"，分布模型告诉你"稳的是哪一种"——指数报错了，再漂亮的 Cpk 也只是纸上数字。
想让团队按 2026 版 SPC 手册把能力指数用对、用出说服力？续创的 SPC 统计方法实操课用真实过程数据边学边练。欢迎联系续创信息技术咨询。
Not a single point beyond the control limits, yet the customer auditor sends you back with one sentence: "You cannot report Cpk for this process — only Ppk." Many engineers are baffled: isn't the process in control? Clause 8.4 of the 2026 AIAG-VDA SPC manual clears up the misconception: "in control" is not the same as "statistically controlled" — what matters is whether the process mean and variance change over time.
▌ Four Models: Do Location and Variation Change Over Time?
Following ISO 22514-2, the manual classifies processes along two dimensions: whether location (the mean) and variation (the variance) remain constant over time.
- Model A: both location and variation are constant — true statistical control;
- Model B: location is constant, but variation changes over time;
- Model C: variation is constant, but location changes over time;
- Model D: both location and variation change.
Models A and C have sub-types: A1 means the instantaneous distribution is normal, A2 non-normal; C1–C3 describe different ways the location can shift, such as systematic tool wear, tool changes or batch changes.
▌ Who May Report Cpk: Statistically Stable vs. Merely Controlled-Stable
Only models A1 and A2 represent statistical control (statistical stability), and only then are Cp/Cpk appropriate. Models B and D, with location or variation changing, are not statistically controlled — only Pp/Ppk may be reported as performance indices.
Models C1–C3 are "controlled-stable": through retrospective control charts and regular monitoring, the characteristic is kept within the control limits, so Cp/Cpk may be used under controlled conditions — but this is not statistical stability. The manual is explicit: from a quality standpoint, statistical stability is preferable to mere controlled stability; however, controlled-stable operation is cheaper to run and an acceptable compromise in difficult situations.
▌ How to Verify the Model: Look First, Then Calculate
- Visual check: use a probability plot or histogram to see whether the theoretical distribution matches the actual data;
- Numerical assessment: use the net correlation coefficient for a quantitative judgement instead of relying on the eye alone;
- Focus on the critical region: assess the whole data set first, then pay special attention to the 25% of data closest to the specification limits — exactly the region that drives Pmk, Ppk and Cpk.
Note: model assignment only describes how the distribution behaves over time; it says nothing about capability or performance, because tolerances are not yet involved.
▌ Shop-Floor Reality: Tool Wear Is Pushing You into Model C
Tool wear, new tools and batch changes — influences that are hard to avoid in machining — often land the process in Model C. Two things must be remembered:
- If an improvement changes the location or variation, the control limits must be re-evaluated — old limits on a new process are no limits at all;
- Stop treating "no points out of control" as the complete criterion: first identify which time-dependent distribution model your process follows, then decide which index fits and whether you dare write Cpk.
One line to remember: a control chart tells you whether the process is stable; the distribution model tells you what kind of stability it is — report the wrong index, and even a beautiful Cpk is just a number on paper.
Want your team to apply the 2026 SPC manual and use capability indices correctly and convincingly? CINO's hands-on SPC courses practise on real process data. Contact CINO Information Technology Consulting.