Advancing QE Skills (7) | Introduction to Measurement Uncertainty: From GR&R to Uncertainty Reports

1. A "Qualified" GR&R Report Can't Stop a Client's Question

A manufacturing company's hole diameter characteristic is 10.00±0.05 mm. The GR&R report for the measuring instrument shows: repeatability EV=0.0012, reproducibility AV=0.0008, %GR&R=8.7%, which is deemed "acceptable" according to the criteria. Based on this, the inspection specification was approved for over a year. Mid-year, during a process audit by the client, a batch of parts with actual measurement values of 10.028 mm was found, and the company rejected them as out of tolerance. The client's engineer simply asked, "What is your measurement uncertainty? 10.028 mm is only 0.022 mm away from the upper limit of 10.05 mm. How do you prove it is truly out of tolerance?" The inspection department head had no answer—there was only a percentage in the GR&R report, with no absolute value that could be directly used for judgment. In the end, both parties had to handle the situation based on the worst-case scenario, and the cost of dealing with this batch exceeded the entire budget for measuring instruments that year.

The root cause lies in the confusion of two issues: GR&R answers whether the "measurement system is adequate," while measurement uncertainty answers how "reliable the reading is in millimeters."

2. Key Principles: Two Ways to Express the Same Data

The coordinate systems for the two are different. %GR&R is a relative proportion of the tolerance (6σ_ms/T), dimensionless; uncertainty U is an absolute value (mm, g, °C), with units and can be used for judgment. Therefore, using GR&R to answer whether a part is "qualified" is like using a proportion to answer a dimensional problem.

The underlying logic of uncertainty is the additivity of variances: when sources are independent, the combined standard uncertainty u_c = √(Σu_i²), and the expanded uncertainty U = k × u_c. In engineering, k is typically taken as 2 (approximately 95% confidence).

The conversion rules from GR&R to u_i are as follows:

• Repeatability EV, Reproducibility AV: directly take the standard deviation σ, do not multiply by 6;
• Calibration certificate provides U and k: u = U/k (if k=2, divide by 2); if only the maximum permissible error ±MPE is given: u = MPE/√3;
• Instrument resolution d: u = (d/2)/√3 = d × 0.289;
• Environmental temperature: u = ΔT × α × L/√3 (α is the linear expansion coefficient).

Using the data from the previous example, the individual components and their proportions are:

After combining, u_c = 0.0039 mm, U = 0.0079 mm, T/U = 12.7:1 (T is the tolerance band width 0.10 mm). Note an counterintuitive result: GR&R only contributes 13.5% of u_c², while resolution and calibration account for 80.0%. This is the mechanism behind why "GR&R qualified" does not equal "U very small"—GR&R does not cover the resolution and calibration components.

Another quantitative relationship worth remembering: if only the measurement system itself is considered, the total width of the guard band 2U/T = (2/3) × %GR&R. In the previous example, this calculation suggests only 5.8% of the tolerance would be consumed, but the actual measurement consumes 15.7%, a difference of 2.7 times, all from components not covered by GR&R.

3. Practical Five Steps: From GR&R Report to Usable Uncertainty

Step 1: List the Components. Write out the five major categories—repeatability, reproducibility, calibration, resolution, and environment—and specify the data source for each (which report, which certificate). Criterion: the list must have at least 5 items; any list missing resolution or calibration sources is directly deemed incomplete.

Step 2: Convert to Standard Uncertainty. Convert each component according to the rules mentioned above, ensuring all are in the same unit (e.g., mm). Criterion: each component must be labeled with the distribution assumption (normal/uniform) and data source, otherwise, it cannot be verified later.

Step 3: Combine and Choose k. u_c = √(Σu_i²), U = k × u_c. Typically, k is taken as 2; when the effective degrees of freedom in the components are small (e.g., repeatability is based on fewer than 6 trials, or multiple components come from small samples), a more conservative k of 2.1 to 2.5 is advisable. Criterion: U should be reported with two significant digits, and if k is not specified, the report is deemed不合格 (unqualified).

Step 4: Calculate T/U and Determine the Judgment Limits. T/U = tolerance band width / U; the guard band width w can be chosen in three ways (strict w=U, moderate w=U/2, no guard band for simple acceptance). The acceptance limit AL = lower limit + w, upper limit - w. In the previous example, if w=U, the judgment interval is 9.9579 to 10.0421. Any reading within this interval cannot be simply accepted or rejected; it must follow a predetermined shared risk handling process. Criterion: simple acceptance is only applicable if T/U ≥ 4:1 and process Cpk ≥ 1.33; if T/U < 3:1, a guard band or a higher precision instrument must be used.

Step 5: Issue the Report and Incorporate into Specifications. The report should include six elements: the measured quantity and specification, the measurement value, u_c, U and k, the judgment rule, and the judgment conclusion. The critical action is to incorporate the judgment rule (whether to use a guard band, its width, and how to handle shared risk) into the inspection specification or contract. Criterion: the rule must be pre-determined, and selecting a rule post-measurement is equivalent to an invalid judgment.

4. How to Choose the Judgment Rule: Balancing Three Approaches

Simple acceptance implies that both parties accept the shared risk—any measurement value within the specification line is considered qualified, without a guard band, provided that T/U is sufficiently large or the risk is acceptable. The most conservative guard band w=U almost never misjudges out-of-tolerance parts as qualified, but the cost is high: in the previous example, 15.7% of the tolerance band would fall into the "unclear judgment" area, leading to increased scrap and retesting on the production line. A moderate approach is to take w=U/2 (in the previous example, AL = 9.9539 to 10.0461), which is the balance point for most manufacturing sites, but this line must be confirmed by both quality and the client, not decided by the inspector on the spot.

Once the approach is agreed upon, it must be consistent in the inspection specification, control plan, and client agreement. Using different rules for the same characteristic in three documents is the easiest discovery item during audits.

When a part falls into the shared risk area, the most common practice is "retest several times and take the average." The effectiveness of this approach depends on the composition of the components: retesting m times and taking the average only reduces the random component to 1/√m. In the previous example, when m=3, U decreases from 0.0079 to 0.0045, and the judgment interval widens to 9.9545 to 10.0455; when m=9, U decreases to 0.0026, and the interval narrows to 9.9526 to 10.0474—10.028 mm always falls within the interval. The reason is that 80.0% of u_c in this example comes from resolution and calibration, which are components that do not shrink with retesting. Retesting is only effective when repeatability is a high proportion (e.g., manual reading, unstable clamping). Therefore, there should be a simple judgment rule on-site: if the proportion of repeatability in u_c² is less than 50%, retesting plans are not adopted, and only the measuring instrument or concession negotiation can be changed.

Additionally, uncertainty reports are not one-time tasks: when the measuring instrument is changed, the calibration level is adjusted, the specification is tightened, or the measurement environment conditions change (relocation, air conditioning failure), or when the client makes a request, any of these four triggering conditions should prompt a re-evaluation and update of the judgment rule with the new value.

5. Common Misconceptions

1. Adding U to the Measurement Value as an Error. 10.028 + 0.0079 = 10.036 is not the "true value," but one endpoint of the confidence interval; uncertainty describes dispersion, not bias correction.
2. Not Dividing the U from the Certificate by k. Using it directly as the standard uncertainty would overestimate the component by 2 times, leading to a significant overestimation when combined.
3. Taking 6σ as u. The 6σ in GR&R is the numerator of the "tolerance proportion," not the standard uncertainty; using it directly would amplify repeatability by 6 times.
4. Believing %GR&R Qualification Means No Need to Calculate Uncertainty. In the previous example, %GR&R = 8.7% is fully qualified, but the largest component is resolution (54.1%), leading to a completely different conclusion.
5. Reporting Only a ± Value. Not providing k, the component table, or the judgment rule makes it impossible for the client to verify, rendering the report ineffective.
6. Using Single Repeatability to Represent u_c. Missing calibration and resolution, it typically underestimates by more than one fold.

6. Self-Check List

• Does the component list cover repeatability, reproducibility, calibration, resolution, and environment, with each item noting the data source?
• Is the U from the calibration certificate divided by k? Is the resolution calculated as d × 0.289?
• Does the report provide u_c, U, k, and the T/U ratio (not just a ± value)?
• Are the guard band width and judgment rules pre-specified in the inspection specification or contract, and consistent with the control plan and client agreement?
• For parts falling into the shared risk area, is there a clear handling process (retest, tighten, concession, or client negotiation)?

First, calibrate your own measuring instrument, then judge the parts of others.

Knowledge code: 11.2.2

Version: v20260917

Author: QTank QTank is dedicated to providing systematic knowledge, methodologies, and practical tools for quality management professionals, helping companies continuously improve their quality capabilities.