How to use the process capability calculator
- Enter the data. Paste from Excel, upload a .csv / .xlsx file or download the template. Put one rational subgroup per row (for example 5 consecutive parts every hour), or paste one long column and give the subgroup size. A subgroup size of 1 treats the data as individual readings and uses the moving range.
- Enter the specification. LSL, USL and, if the drawing has one, a target. Leave one limit blank for a one-sided characteristic such as roughness or tensile strength.
- Read the report. The verdict compares Cpk with your requirement; the diagnosis tells you whether the problem is centering, spread or instability. Fill in “Report details” and print or save the PDF.
The data must be in time order. The within-subgroup estimate and the stability chart are meaningless if the rows have been sorted.
The formulas
All capability indices compare the tolerance with the process spread. They differ only in which standard deviation is used and whether the location of the mean is taken into account.
| Index | Formula | What it measures |
|---|---|---|
| Cp | (USL − LSL) / (6·σwithin) | Potential capability if the process were centred |
| Cpu, Cpl | (USL − x̄) / (3·σwithin), (x̄ − LSL) / (3·σwithin) | Capability against each limit |
| Cpk | min(Cpu, Cpl) | Short-term capability including centering |
| Pp, Ppk | Same formulas with σoverall = s | Actual long-term performance |
| Cpm | (USL − LSL) / (6·√[s² + (x̄ − T)²]) | Performance relative to the target T |
| k | |M − x̄| / ((USL − LSL)/2), M = midpoint | Off-centre ratio; Cpk = Cp·(1 − k) |
Within versus overall standard deviation
Reconcile different capability reports
The comparison table holds measurements, subgroup membership and specifications fixed. The existing pooled option uses sp without correction; the added option uses sp/c₄(d+1), where d=Σ(nᵢ−1). Match the estimator and correction used in your agreed analysis, rather than choosing whichever produces the largest Cpk. Overall sample SD and Pp/Ppk remain unchanged. This is not a complete reproduction of another software package.
Hand-check example: paste two rows 8,10,12 and 8,10,12, then set LSL=4 and USL=16. Each subgroup contributes squared deviations totaling 8; d=4 and sp=2. Since c₄(5)=3√(2π)/8≈0.939985603, corrected sigma≈2.127692162. Uncorrected Cp=Cpk=1, corrected Cp=Cpk≈0.939985603; overall s=√(16/5), so Pp=Ppk≈1.118033989. This tiny synthetic example checks arithmetic, not process capability.
Reference: Minitab normal capability sigma methods. Normality, rational subgrouping and stability assumptions still apply.
σ within is estimated only from the variation inside each subgroup: R̄/d₂ (the default for subgroup sizes up to 10), S̄/c₄ or the pooled standard deviation, and MR̄/1.128 for individual readings. With a suitable sampling plan and rational subgroups, it estimates short-term variation; the estimator alone does not establish common-cause control. σ overall is the ordinary sample standard deviation of all readings, so it also contains shifts and drifts between subgroups. This follows the AIAG SPC manual and Montgomery: Cp/Cpk use σ within, Pp/Ppk use σ overall. When the process is in statistical control the two are nearly equal and Cpk ≈ Ppk. A difference between Ppk and Cpk calls for checking subgrouping, sampling and estimator settings as well as process behavior; it does not alone establish instability.
PPM, Z.bench and sigma level
The expected fraction outside each limit is the normal tail area: P(x > USL) = Φ(−Z.usl) with Z.usl = (USL − x̄)/σ, and similarly for LSL. Multiply by 10⁶ for PPM. Z.bench = Φ⁻¹(1 − Ptotal) converts the total into a single Z value. The calculator reports Z.bench from σ within (short-term) and from σ overall (long-term), plus the conventional Six Sigma “sigma level”, which adds a 1.5σ shift to the long-term value — see the sigma level calculator for that convention. The observed PPM is simply the count of readings outside the limits and assumes no distribution.
How to interpret Cpk
| Cpk | Nearest limit at | PPM beyond the nearest limit | PPM if centred (both tails) | Typical use |
|---|---|---|---|---|
| 0.67 | 2.0σ | 22,216 | 44,431 | Not capable — sort 100% or fix the process |
| 1.00 | 3.0σ | 1,350 | 2,700 | Barely capable — the 3σ spread just fills the tolerance |
| 1.33 | 4.0σ | 33.0 | 66.1 | Common minimum for existing processes |
| 1.67 | 5.0σ | 0.272 | 0.544 | Common minimum for new processes and safety / critical characteristics |
| 2.00 | 6.0σ | 0.00099 | 0.0020 | Six Sigma design goal (short-term) |
The PPM columns are computed from the normal distribution with no mean shift. Diagnose with pairs of indices:
- Cp high, Cpk low — a centering problem. Adjust the mean; the spread is already good enough.
- Cp low — a spread problem. Cpk can never exceed Cp, so re-centring cannot rescue it; reduce variation or widen the tolerance.
- Cpk high, Ppk low — check sampling, subgrouping and estimator choices, and investigate possible process changes with an X̄-R chart (or an I-MR chart for individual readings) before using capability indices to predict future output.
Worked example
The sample data are 25 subgroups of 5 shaft diameters with a specification of 25.000 ± 0.018 mm (LSL 24.982, USL 25.018).
x̄ = 25.00131 mm and R̄ = 0.01028 mm, so σwithin = R̄/d₂ = 0.01028 / 2.326 = 0.00442 mm. The sample standard deviation of all 125 readings is s = 0.00454 mm.
Cp = 0.036 / (6 × 0.00442) = 1.36. Cpu = (25.018 − 25.00131) / (3 × 0.00442) = 1.26 and Cpl = (25.00131 − 24.982) / (3 × 0.00442) = 1.46, so Cpk = 1.26. With s: Pp = 1.32 and Ppk = 1.23.
Verdict against 1.33: marginal. Because Cp passes and Cpk does not, the tool reports a centering problem — the mean is 0.0013 mm high. Cpk and Ppk are close and this X̄ screen shows no signals; these checks alone do not establish stability. Review time order, sampling and control-chart evidence before predicting performance.
Assumptions you must check
- Stability. Indices describe the future only for a process in statistical control. The tool screens for points beyond the 3σ limits and runs of 8, and warns you.
- Normality. Every expected PPM comes from the normal model. The Anderson–Darling test and probability plot are included; with p < 0.05 use the normality test and histogram maker to find out why (outliers, mixed streams, a naturally skewed characteristic).
- Sample size. An index is an estimate. The 95% interval uses Bissell’s approximation, index ± 1.96·√[1/(9N) + index²/(2ν)]. With 30 readings a Cpk of 1.33 has an interval of roughly ±0.36; aim for 100 readings or more.
- Measurement system. Gauge variation is part of σ. If the gauge consumes much of the tolerance, run a Gage R&R study first.
Frequently asked questions
What is the difference between Cpk and Ppk?
Both use the same formula, min(USL − mean, mean − LSL) / (3 × sigma). Cpk uses the within-subgroup standard deviation (R-bar/d2, S-bar/c4, pooled, or moving range), which estimates short-term variation when the sampling plan and subgrouping are appropriate. Ppk uses the overall sample standard deviation of all readings, which also includes shifts and drift between subgroups. They may be close for a stable process, but sampling and estimator choices also affect their difference. A difference does not alone establish instability; review the time-ordered control chart and process evidence.
What is a good Cpk value?
Cpk 1.00 means the nearest specification limit is three standard deviations from the mean (about 1,350 PPM beyond that limit). Most customers require at least 1.33 for existing processes and 1.67 for new processes or critical characteristics; Six Sigma programmes aim for a short-term Cpk of 2.00. Always use the value in your customer or internal requirement.
How many data points do I need for a capability study?
The usual guideline is at least 100 individual readings, typically 25 subgroups of 4 or 5 parts, collected over a period long enough to include normal variation in material, operators and set-ups. With 30 readings the 95% confidence interval of a Cpk near 1.33 is roughly plus or minus 0.36, so small studies cannot distinguish a capable process from a marginal one.
Can I calculate Cpk with only one specification limit?
Yes. With only an upper limit Cpk equals Cpu = (USL − mean) / (3 sigma); with only a lower limit it equals Cpl = (mean − LSL) / (3 sigma). Cp and Pp are not defined because they need the full tolerance width. Leave the missing limit blank in the calculator.
What if my data are not normally distributed?
Cp, Cpk, Pp, Ppk and the expected PPM assume a normal distribution, and the PPM estimate is the most sensitive to that assumption. First look for the cause: outliers, two machines or cavities mixed together, or an unstable process. Characteristics that are naturally skewed, such as flatness or runout, need a transformation or a non-normal capability method. The observed PPM count does not depend on the distribution.
Why does this Cpk differ slightly from my other software?
Usually because of the within-subgroup sigma estimator. This calculator defaults to R-bar/d2 for subgroup sizes up to 10; many statistical packages default to the pooled standard deviation, sometimes with an unbiasing constant. Match the data, subgroup membership, specification limits, estimator and correction settings before comparing results. The comparison table helps isolate estimator differences; agreement with every software package is not guaranteed. The method section of the report states exactly which estimator and constants were used.