OEM / private label · Multi-factory sourcing · Export project coordination
sales@moxasourcing.comCall UsChat on WhatsApp

Process Stability and Capability: Reading the Indices

In NIST’s production-characterization discussion, stability means that the response has a constant mean, variance and distribution over time.

Process Stability and Capability: Reading the Indices: Stable process, Distribution model, Specification limits, Capability estimate
Original educational diagram by Moxa Sourcing.
Quick comparison
Evidence or termWhat it means here
StabilityWhether the process behavior is consistent over the assessed history.
CpA comparison of specification width with the selected spread estimate.
CpkA capability index that also considers proximity to the specification limits.

What does the evidence explain?

What does process stability mean before capability is discussed?

In NIST’s production-characterization discussion, stability means that the response has a constant mean, variance and distribution over time. It is a statement about how the process behaves, not about whether its output meets the buyer’s limits. The source places stability assessment before capability analysis because a changing process cannot be adequately represented by one fixed capability description.

A short collection of acceptable-looking units does not demonstrate that stability. Conversely, stable output can be consistently outside the required range. Identify the measured response and the interval represented by the data. This avoids turning the general phrase “consistent quality” into an unsupported claim that all future shipments will conform. The source provides an analytical framework, not evidence that a particular supplier’s process is stable.

Why must process observations retain their time order?

The source assesses stability by plotting observations against sample order, looking for random variation around a constant level. Control limits can be added where appropriate. The time span must be long enough to represent the behavior under investigation; the page describes collecting more than one hundred independent samples for its process-characterization approach.

That number is not a universal lot-acceptance sample size or a release obligation created for every product by this guide. It belongs to the source’s stability study. The useful record requirement is to preserve the sequence and conditions rather than sorting values into a convenient order before review. A histogram can show the overall distribution, but removing time order can hide a shift or drift. The linked numerical stationarity and chart methods require their own method selection before use.

How can a stable process still be incapable?

Capability compares the output of an in-control process with the specification limits. A process can have a stable distribution whose spread is too wide for those limits or whose center is displaced from the desired range. Stability and capability therefore answer different questions. The first concerns behavior over time; the second concerns that behavior relative to the required specification.

NIST illustrates capability using a histogram with specification limits. The limits come from the requirement, not from drawing a convenient boundary around the observed values. Do not revise the specification merely to make a process appear capable. Also distinguish this assessment from an acceptance decision on one sampled lot. A process-history analysis can inform the control plan, while the actual order retains its agreed lot definition, inspection and disposition rules.

What does the Cp index compare?

For the normal-distribution model with two specification limits, NIST defines Cp as the specification width divided by six process standard deviations: Cp = (USL − LSL) / (6σ). USL and LSL are the upper and lower specification limits. Sigma describes the process distribution, not the standard uncertainty of a measurement report. The numerator and denominator must describe the same measured characteristic in compatible units.

Cp measures width, not centering. It does not contain the process mean, so shifting the same distribution without changing its spread leaves Cp unchanged. This is why an impressive Cp alone cannot show that the output lies in the correct part of the specification interval. The index is dimensionless, but its interpretation still depends on a valid process model and meaningful specification limits.

How does Cpk account for an off-center process?

Cpk uses the smaller of two distances: the distance from the process mean to the upper limit and the distance from the mean to the lower limit, each divided by three process standard deviations. In symbols, Cpk = min[(USL − μ)/(3σ), (μ − LSL)/(3σ)]. It therefore reflects the more restrictive side of the specification interval.

If the process mean is at the midpoint, the two one-sided distances are equal and Cpk equals Cp. Moving the mean toward either limit reduces the smaller distance without necessarily changing Cp. The direct formula also makes the sign meaningful if a mean lies beyond a limit; it should not be hidden by taking an absolute value. Any acceptance threshold for the index must come from the actual technical agreement, not from assuming every source example supplies a universal pass mark.

Why distinguish a specified target from the midpoint of two limits?

NIST also defines Cpm, whose denominator includes both process variance and the squared distance of the mean from a target T. Its form is (USL − LSL) divided by six times the square root of [σ² + (μ − T)²]. It therefore includes a target-deviation term that is absent from Cp.

A nominal target and two acceptance limits are different pieces of specification information. Do not silently replace the stated target with the interval midpoint if the project defines them differently. Cpm does not remove the need to inspect the underlying limits, mean and variation; it summarizes them under the stated model. Report which capability index is used rather than describing Cp, Cpk and Cpm as interchangeable scores.

What if the characteristic has only one specification limit?

A one-sided specification uses an upper or a lower requirement rather than a bilateral interval. NIST defines Cpu = (USL − μ)/(3σ) for the upper side and Cpl = (μ − LSL)/(3σ) for the lower side. Where both limits exist, Cpk is the smaller of Cpu and Cpl, while Cp is their average.

Do not invent a second specification limit solely to calculate a two-sided index. First identify the actual requirement and choose the applicable measure. When sample estimates replace the population mean and standard deviation, label the resulting index as an estimate. The calculation’s compact form does not establish that the underlying process distribution, independence or stability assumptions have been checked.

How can identical spread give Cp of one but a smaller Cpk?

Use the source illustration’s upper limit twenty, lower limit eight and estimated standard deviation two. For the calculation, take the mean as sixteen, as used in its displayed arithmetic. The specification width is twelve and six standard deviations also equal twelve, so estimated Cp is one. The upper one-sided index is four divided by six, approximately 0.6667; the lower is eight divided by six, approximately 1.3333. Estimated Cpk is the smaller value, 0.6667.

If the same spread were centered at fourteen, the midpoint of those limits, the one-sided distances would match. The calculation isolates the effect of centering without pretending that variation improved. These are abstract source-based calculations with common units, not actual dimensions, temperatures or results for a mugwort product. The source’s prose uses an inequality around its average while its arithmetic uses sixteen; this explanation states the calculation input explicitly rather than treating that prose as a measured dataset.

Why is an estimated capability index uncertain?

A population capability index uses the process mean and standard deviation; an estimated index substitutes the sample mean and sample standard deviation. Different random samples can therefore produce different estimates even when the underlying process has not changed. NIST discusses confidence limits and cautions that variation in estimated indices is not negligible.

The page mentions roughly fifty independent values in discussing sufficiently large samples for index estimates, then discusses larger capability studies and minimum conditions for particular approximations. These are different contexts, not one universal product-inspection count. A report should state its data basis, study scope and uncertainty treatment. Do not compare two small-sample point estimates as if their last decimal places established a real process improvement. The detailed confidence-limit formulas require the applicable model and data; none is prescribed here for a current supplier.

Can a Cp value be converted directly into an observed reject rate?

NIST’s table relating capability width to tail fractions assumes a normal distribution centered in the specification interval. Under those assumptions, a specification width of six standard deviations corresponds to a modeled tail fraction of approximately 0.27 percent. Changing the centering or distribution changes the interpretation even if Cp itself is unchanged.

A theoretical tail fraction is not an observed shipment reject rate. It should not be advertised as a supplier’s achieved parts-per-million performance without the relevant process and measurement evidence. State the assumptions when using such a conversion and keep modeled expectations separate from actual inspection results. A capability index can summarize a process model; it does not count every unit or eliminate the possibility of nonconforming items.

What changes when the process data are not approximately normal?

The usual indices and their normal-tail interpretations cannot simply be carried over without checking the distribution. NIST describes two routes: a suitable transformation toward normality or indices designed for nonnormal distributions. Naming one route does not establish that it is appropriate for the data or that its assumptions have been satisfied.

If a transformation is used, the observations and specification information must be interpreted consistently in the analysis. A transformation should not conceal an unstable process or change the physical acceptance requirement. If a nonnormal index is used, identify its definition and coverage basis. Keep the original measured quantity and limits available so that a compact transformed result does not lose its connection to the actual product characteristic.

Why must percentile-based capability reports state their coverage convention?

The source notes that nonparametric capability descriptions can use different percentile ranges. Its discussion includes the 0.135th to 99.865th percentiles to mimic the normal model’s approximately 99.73 percent interval, while other conventions use the 0.5th to 99.5th percentiles for 99 percent coverage. These are different definitions of the width being compared with the specification.

Consequently, two reported nonnormal capability numbers may not be comparable merely because their symbols look similar. Identify the percentile convention, estimation method and data basis before interpreting a difference. The introductory source does not provide the unseen raw data for a current factory, and an extreme percentile requires an appropriate estimation basis. No such estimate or current supplier capability is claimed by listing these conventions.

FAQ

What does process stability mean before capability is discussed?

In NIST’s production-characterization discussion, stability means that the response has a constant mean, variance and distribution over time. It is a statement about how the process behaves, not about whether its output meets the buyer’s limits. The source places stability assessment before capability analysis because a changing process cannot be adequately represented by one fixed capability description.

Why must percentile-based capability reports state their coverage convention?

The source notes that nonparametric capability descriptions can use different percentile ranges. Its discussion includes the 0.135th to 99.865th percentiles to mimic the normal model’s approximately 99.73 percent interval, while other conventions use the 0.5th to 99.5th percentiles for 99 percent coverage. These are different definitions of the width being compared with the specification.

Sources and scope

Sources: NIST/SEMATECH process stability and capability pages, interpreted as general statistical methods. Stable, capable and accepted are different conclusions. Normality, independence, centering and sampling uncertainty remain explicit; no universal capability pass threshold, required SKU test count or actual supplier performance is asserted. NIST process stability NIST capability assessment NIST capability indices

Chat on WhatsApp+852 5725 4301