量具年年送校准、证书一张不缺，客户审核还是开出"MSA 不充分"的不合格项。原因很简单：校准只证明量具在标准点上的偏倚受控，而 MSA 要回答的是——这套"人+量具+零件+方法+环境"组成的测量系统，输出的数据到底可不可信。计量型 MSA，看五个特性。
▌ 五性分两堆：位置管准不准，宽度管散不散
位置特性有三个——偏倚、线性、稳定性，管"准不准"；宽度特性有两个——重复性、再现性，管"散不散"。为什么两堆都要看？因为测量数据 = 零件真实变差 + 测量系统变差，测量变差吃掉公差，Cpk 算得再漂亮也是假象。
▌ 位置三性：偏倚、线性、稳定性
- 偏倚：多次测量均值与基准值之差。拿基准值已知的标准件测 10 次以上，均值减基准值即是；
- 线性：偏倚在量程内是否一致。选 5 个覆盖量程的基准件各测 10 次以上，拟合偏倚-基准值直线，斜率接近零才算好；
- 稳定性：偏倚随时间漂不漂。定期复测同一标准件，结果画 Xbar-R 控制图，持续受控即稳定。
▌ 宽度两性：重复性与再现性
重复性（EV）：同一操作者、同一量具、同一零件连续测量的散布，反映设备变差——卡尺测同一根轴，十次读数各不相同，就是它。
再现性（AV）：不同操作者用同一量具测同一零件的均值差异，反映人员变差——同一把卡尺，老师傅和新手读数不一样，就是它。两者合起来就是 GRR。
▌ GRR 怎么做、怎么判
经典方案：10 个零件 × 3 名操作者 × 3 次重复 = 90 个数据。三条铁律：
- 零件从生产过程随机抽取、覆盖过程变差——专挑同一批次会造出假合格；
- 测量顺序随机、盲测，不许互相报数；
- 记录原始读数，不做修约。
分析首选 ANOVA 法，比均值-极差法多拆出"人×零件"交互作用。判定线：GRR% 小于 10% 可接受；10%~30% 有条件接受，视重要性改进；大于 30% 不可接受，必须改进。另看 ndc ≥ 5，否则量具分不出零件好坏。
▌ 顺序不能反：先三性，再 GRR
校准不等于 MSA，实施也有先后：先做偏倚、线性、稳定性，再上 GRR。位置特性不过关，宽度特性做得再细也是白做。
记一句：测量数据不可信，SPC 和 Cpk 全是空中楼阁——数据可信，工具才有意义。
想让团队把 MSA 从"应付审核"做成"数据可信"的基本功？续创的在线实操课带你用真实数据做 GRR。欢迎联系续创信息技术咨询。
Gauges go out for calibration every year and every certificate is on file, yet the customer audit still raises a nonconformity for "insufficient MSA". The reason is simple: calibration only proves the gauge's bias is controlled at reference points, while MSA answers a bigger question — is the data from this measurement system of "people + gauge + parts + method + environment" actually trustworthy? For variable MSA, look at five properties.
▌ Five Properties in Two Piles: Location for Accuracy, Width for Spread
Three location properties — bias, linearity and stability — govern "is it accurate"; two width properties — repeatability and reproducibility — govern "is it consistent". Why watch both piles? Because measured data = true part variation + measurement system variation. When measurement variation eats into the tolerance, even the prettiest Cpk is an illusion.
▌ The Three Location Properties: Bias, Linearity, Stability
- Bias: the difference between the average of repeated measurements and the reference value. Measure a master part with a known reference value at least 10 times and subtract the reference from the mean;
- Linearity: whether bias stays consistent across the range. Pick 5 masters covering the range, measure each at least 10 times, fit the bias-versus-reference line — a slope near zero is what you want;
- Stability: whether bias drifts over time. Re-measure the same master regularly and plot the results on an Xbar-R chart; ongoing statistical control means stability.
▌ The Two Width Properties: Repeatability and Reproducibility
Repeatability (EV): the spread when the same operator, with the same gauge, measures the same part repeatedly — it reflects equipment variation. Ten caliper readings on the same shaft all coming out different? That is repeatability.
Reproducibility (AV): the difference between averages when different operators measure the same part with the same gauge — it reflects appraiser variation. Same caliper, but the veteran and the new hire read it differently? That is reproducibility. Together they make up GRR.
▌ How to Run a GRR Study and How to Judge It
The classic design: 10 parts × 3 operators × 3 repeats = 90 readings. Three iron rules:
- Parts are drawn randomly from the process and must cover the process variation — cherry-picking a single batch manufactures a fake pass;
- Measurement order is randomised and blind — no calling out readings to each other;
- Record raw readings with no rounding.
Prefer the ANOVA method, which — unlike the average-and-range method — also separates out the operator-by-part interaction. Acceptance lines: GRR% below 10% is acceptable; 10%–30% is conditionally acceptable, improved based on importance; above 30% is unacceptable and must be fixed. Also check ndc ≥ 5, otherwise the gauge cannot tell good parts from bad.
▌ The Order Is Not Negotiable: Three Properties First, Then GRR
Calibration is not MSA, and implementation has a sequence: bias, linearity and stability first, then GRR. If the location properties fail, no amount of refinement on the width properties will help.
One line to remember: if measurement data cannot be trusted, SPC and Cpk are castles in the air — trustworthy data is what makes the tools meaningful.
Want your team to build MSA into a data-integrity discipline instead of an audit-pleasing exercise? CINO's hands-on online courses guide trainees through GRR studies on real data. Contact CINO Information Technology Consulting.