Process Capability

A factory may need a part thickness between an upper and lower limit, and repeated measurements show how its process actually varies, so capability analysis asks whether that spread fits comfortably inside the allowed range. Specifications come from customer or design requirements and are not the same as control limits calculated from process behaviour, so a process can be statistically stable yet make products outside specification.

First check stability with an appropriate time-ordered analysis, such as control charts, because if the process shifts unpredictably a capability estimate based on past data may say little about future output. NIST describes capability as comparing an in-control process with its specification limits, and ASQ likewise stresses stable sampling and the challenge of representative data, so a single snapshot from a convenient hour may be misleading.

Cp compares specification width with a six-standard-deviation process spread, which measures potential fit when the process is centred but does not check whether the process average sits near one boundary. Cpk considers the nearer specification limit and the process mean, so if a process is off-centre Cpk can be lower than Cp, letting a manager distinguish excessive spread from a centring problem.

For a fictional measurement with lower limit 8, upper limit 20, mean 16 and standard deviation 2, Cp equals 12 divided by 12, or 1, while Cpk is the smaller of 4 divided by 6 and 8 divided by 6, about 0.67. The example shows why Cp alone can flatter an off-centre process, and NIST gives a similar worked calculation.

It does not establish a universal pass threshold for every product or risk level. The standard deviation estimate depends on sampling design, so measurements across several shifts can reveal changes missed by a short burst, and enough independent, representative observations are needed for the stated method.

Cp and Cpk formulas commonly assume approximately normal continuous measurements, so non-normal data may need transformation or another method, and an attribute outcome such as pass or fail needs different analysis from a continuous thickness measurement. Do not assign a Cp value to a yes-or-no result without a defensible model, since ASQ notes different capability estimates for attribute data.

Measurement-system error can inflate or obscure observed variation, so check gauges, calibration and repeatability before blaming the production process; a better machine cannot correct a faulty ruler. One-sided specifications require an appropriate one-sided measure, because a lower minimum strength without an upper limit is not the same problem as a two-sided dimensional tolerance.

Confidence matters with small samples as an estimated Cpk is itself uncertain, so report the sample period and count and consider uncertainty where decisions are costly. Capability should be studied for a specific characteristic and process condition, since combining data from different machines can hide one weak machine or create artificial variation, and customers may set a capability requirement in a contract, so confirm their specification, sampling plan and acceptance method instead of using a remembered industry rule.

A capability study can guide improvement, with wide spread calling for reduced sources of variation and a mean close to a limit calling for adjusted centring while maintaining control, and actual nonconforming output should be counted and investigated too, since a model-derived index and the observed defect rate answer related but different questions. A stable process can still be incapable and a temporary good run can still be unstable, so treat stability and capability as two checks and pair an index with a distribution plot, time sequence, sample details and actual quality outcomes.