How do you improve Cpk when a supplier reports a low Cpk on a machined feature?

What does a low Cpk on a machined feature actually tell you?

A Cpk below 1.33 says the nearer specification limit sits less than four standard deviations from the process mean, so the tail of the distribution reaches closer to that limit than the target allows and the out-of-tolerance rate the model predicts is higher than agreed. It does not say why, and the same number can come from a process aimed at the wrong size, a process that is too variable, a gauge that is too noisy, or data that do not fit the normal model behind the index.

The NIST/SEMATECH e-Handbook writes Cpk as Cp(1 − k), where k is how far the mean sits from the middle of the tolerance as a share of half the tolerance, and states that Cpk can never exceed Cp. Cp compares the tolerance width with the process spread alone; Cpk also charges for aim. That pair is the first diagnostic tool you have, because Cp is the ceiling: centering a process perfectly raises Cpk to Cp and no further. If Cp itself is below the target, no offset change will get the characteristic there.

The AIAG PPAP manual grades an initial-study index below 1.33 as not meeting acceptance and asks the supplier to take the result to the customer; where the target cannot be reached in time, the manual expects a corrective action plan and a control plan that normally adds 100 percent inspection. Before treating the number as a verdict on the supplier, check that it is a real number. A study of a few dozen parts has a wide confidence interval: using the normal approximation the NIST handbook gives, an estimated Cpk of 1.33 from 30 parts has a one-sided 95 percent lower bound of about 1.03, and an estimate of 1.67 from 30 parts only proves about 1.30. A study with points outside the control limits describes no stable process at all; and a study computed against the wrong limits, for example a position tolerance that should have earned bonus at MMC, is judging the part against a tolerance the drawing does not impose.

How do you diagnose why Cpk is low before agreeing to a fix?

Ask for the raw readings in production order and work through four checks: is the process stable, is it centered, is the gauge good enough, and do the data fit the model the index assumes? Each check takes minutes with a spreadsheet and rules out a whole class of wrong fixes.

1. Stability. Plot the readings on an X-bar and R chart or an individuals chart (NIST e-Handbook sections 6.3.2.1 and 6.3.2.2). Points outside the limits, or patterns caught by the run and trend rules in the handbook’s variables control chart section, mean the process was changing during the study. Fix that first; an index from an unstable process predicts nothing.
2. Centering. Compare Cpk with Cp. When Cpk is close to Cp, the process is centered and the problem is spread. When Cpk is well below Cp, the problem is aim, and a re-centered process will reach roughly the current Cp. Worked example using the NIST formulas: a bore with limits of 19.970 and 20.030 mm, a mean of 20.014 mm and a standard deviation of 0.0065 mm gives Cp = 0.060 / 0.039 = 1.54 but Cpk = 0.016 / 0.0195 = 0.82; moving the mean to 20.000 mm lifts Cpk to 1.54 with no other change. Check the drawing for unequal bilateral or unilateral tolerances, where the nominal is not the middle of the tolerance zone.
3. Measurement. Ask for a gauge R&R on the gauge that produced the readings. In the AIAG MSA manual, a result below 10 percent is generally acceptable, 10 to 30 percent may be acceptable depending on the application, and above 30 percent the gauge needs improving. Measurement variance adds to process variance, so a noisy gauge lowers the observed index even when the parts are fine; the table below shows how much.
4. Distribution and limits. Look at the histogram. A skewed shape on a GD&T callout, or two peaks, points to a model or stream problem, not a machining problem. Confirm that the limits in the software match the drawing, including any bonus tolerance at MMC, which ASME Y14.5 allows as the feature departs from its maximum material size.

Only when all four checks pass is the low number a true statement that the process cannot hold the tolerance. That outcome is common on features whose tolerance is close to what the operation can hold in production, and it is the case where a tolerance review becomes a fair question rather than a request to lower the bar.

What should the drawing, RFQ and PO say so a low Cpk does not recur?

Say which characteristics need capability, which index and target apply, how the data are collected and measured, and what happens when the result is short. A low Cpk discovered after the order is placed is often a requirement that nobody priced.

On the drawing, flag the critical characteristics with your key-characteristic symbol and add one note that states the requirement, for example: “Flagged characteristics: initial study index 1.67 minimum, ongoing index 1.33 minimum, distribution model chosen to fit the data per ISO 22514-2; zero-bounded geometric tolerances evaluated one-sided against the upper limit.” Put the rest in the RFQ or quality agreement:

• Each flagged balloon number with its index and its target.
• Which gauge measures each one, with its R&R result supplied before the study starts.
• The run, the subgroup plan, and that readings are reported in production order by machine, spindle and fixture.
• The reaction plan below target: who is told, when 100 percent inspection starts, and when it stops.
• That any deviation is in writing, limited in quantity or time, and tied to a corrective action.

Keep the list short. A capability requirement on every dimension multiplies cost without improving the part; flag the few that decide whether the part fits, seals or fails safe, and leave the rest to first article and lot inspection, as described in the gage R&R guide for the measurement side.