7.1 Process Stability Prerequisite & Capability Basics

7.1 Process Stability Prerequisite & Capability Basics

Introduction: The Voice of the Process vs. The Voice of the Customer

In modern manufacturing and quality engineering, two fundamental forces govern every production operation:

1. The Voice of the Process (VOP): The statistical reality of what the process is actually delivering in terms of location (mean, $\mu$) and dispersion (spread, $\sigma$). The process communicates through statistical process control (SPC) data and control charts.
2. The Voice of the Customer (VOC): The technical requirements and functional tolerances defined on engineering drawings, CAD models, and purchase specifications. The customer communicates through engineering specification limits—the Upper Specification Limit (USL) and Lower Specification Limit (LSL).

Process capability analysis is the scientific study that compares the Voice of the Process against the Voice of the Customer. It answers a vital manufacturing question: Can this process reliably produce parts that satisfy engineering specifications over time?

As an ASQ Certified Quality Technician (CQT), you are responsible for collecting baseline capability data, verifying statistical prerequisites, calculating capability metrics, and interpreting the results for engineering and production teams. Understanding the boundary between process behavior and customer demand is essential for process qualification, machine runoff approval, and defect prevention.

---

The Cardinal Rule: Statistical Stability as an Absolute Prerequisite

The single most critical concept in capability analysis—and one of the most frequently tested principles on the ASQ CQT exam—is the process stability prerequisite:

[!IMPORTANT] The Cardinal Rule of Process Capability: A process MUST be brought into a state of statistical control before capability can be calculated or have any predictive validity. Calculating capability on an out-of-control process is mathematically invalid and practically dangerous.

Why Statistical Stability Must Precede Capability

Statistical capability indices such as $C_p$ and $C_{pk}$ are not merely descriptive summaries of past parts; they are predictive models used to forecast future yield and expected nonconformance rates.

Under Dr. Walter Shewhart's foundational principles of statistical process control:

• A process operating in statistical control is influenced solely by a stable, constant system of common cause (chance) variation. Because the underlying distribution of outcomes remains constant over time, historical data can legitimately predict future performance.
• A process affected by special cause (assignable) variation is inherently unstable and unpredictable. The mean may shift unexpectedly, the spread may expand or contract erratically, and the underlying probability distribution changes from hour to hour.

If you calculate capability indices on an unstable process, the resulting numbers represent a fleeting snapshot of an unpredictable system. A $C_{pk}$ calculated at 9:00 AM on an unstable machine provides zero guarantee of quality at 2:00 PM. Confounding common cause variation with assignable cause variation creates a false sense of security, leading to unexpected customer rejections and massive scrap costs.

+-------------------------------------------------------------------------+
|                   THE CAPABILITY SEQUENCE OF EXECUTION                  |
|                                                                         |
|   [Step 1: Eliminate Special Causes]                                    |
|          │                                                              |
|          ▼                                                              |
|   [Step 2: Establish Statistical Control (Control Charts)]              |
|          │                                                              |
|          ▼                                                              |
|   [Step 3: Verify Data Normality (Q-Q Plot / Anderson-Darling)]         |
|          │                                                              |
|          ▼                                                              |
|   [Step 4: Calculate Process Capability (Cp, Cpk)]                      |
|          │                                                              |
|          ▼                                                              |
|   [Step 5: Compare Against Benchmarks & Implement Improvements]         |
+-------------------------------------------------------------------------+

---

Natural Process Limits vs. Engineering Specification Limits

A quality technician must strictly differentiate between the natural limits established by the machine and the specification limits dictated by design engineers.

Defining Natural Process Limits

When a process is operating in statistical control and follows a normal distribution, its natural boundaries are defined by its Natural Process Limits (NPL)—also referred to as Natural Tolerance Limits (NTL):

• Upper Natural Process Limit (UNPL / UNTL): $\text{UNPL} = \mu + 3\sigma$
• Lower Natural Process Limit (LNPL / LNTL): $\text{LNPL} = \mu - 3\sigma$
• Natural Process Spread: $\text{Spread} = \text{UNPL} - \text{LNPL} = 6\sigma$

Under the normal curve, the interval $\mu \pm 3\sigma$ contains 99.73% of all individual parts produced by the process. The remaining 0.27% (representing 2,700 parts per million, or 1,350 PPM in each tail) falls beyond these natural limits.

Defining Specification Limits

Engineering specification limits are established independently of the production equipment:

• Upper Specification Limit (USL): The highest physical dimension, property, or characteristic value acceptable to the customer or mating assembly.
• Lower Specification Limit (LSL): The lowest physical dimension, property, or characteristic value acceptable to the customer or mating assembly.
• Specification Tolerance Width: $\text{Tolerance} = USL - LSL$

The Three Process-to-Tolerance Relationships

Comparing the $6\sigma$ natural process spread to the specification tolerance band reveals three distinct capability states:

1. $6\sigma < (USL - LSL)$ (Potentially Capable): The natural process spread is narrower than the customer's tolerance band. If the process is properly centered, 100% of production will comfortably conform to specifications with virtually zero defects.
2. $6\sigma = (USL - LSL)$ (Marginally Capable): The natural process spread exactly equals the specification tolerance band. Even if perfectly centered, the process will produce approximately 0.27% nonconforming product (2,700 PPM). Any slight drift in the mean will immediately generate nonconforming parts.
3. $6\sigma > (USL - LSL)$ (Inherently Incapable): The natural process spread exceeds the engineering tolerance band. Regardless of how perfectly the process mean is centered, nonconforming parts are mathematically guaranteed to occur in both tails. Centering adjustments cannot resolve this issue; the process variation must be reduced through engineering redesign, tooling replacement, or equipment overhaul.

Comprehensive Boundary Comparison

[!CAUTION] Critical CQT Exam Rule: Specification limits must NEVER be drawn on an $\bar{X}$ control chart! An $\bar{X}$ chart monitors subgroup averages, which vary by $\sigma/\sqrt{n}$, whereas specifications apply strictly to individual pieces. Plotting specifications on an $\bar{X}$ chart confuses operators and obscures out-of-control conditions.

---

The Normality Assumption and Verification Methods

Standard capability indices ($C_p, C_{pk}, P_p, P_{pk}$) are mathematically derived under the assumption that the quality characteristic follows a Normal (Gaussian) distribution.

Why Normality Matters

The normal distribution has well-defined mathematical properties where the mean $\pm 1\sigma$ covers 68.27%, $\pm 2\sigma$ covers 95.45%, and $\pm 3\sigma$ covers 99.73% of the total area under the curve. When an engineer calculates $C_p = 1.0$, they assume that 0.27% of parts fall outside the limits.

However, if the underlying process distribution is skewed (such as runout or flatness bounded by zero) or heavy-tailed (kurtotic), the area beyond $\pm 3\sigma$ can be dramatically higher than 0.27% (sometimes 2% to 5%), resulting in massive unexpected defect rates despite an apparently acceptable capability score. Quality technicians must test and confirm data normality before reporting standard capability indices.

Method 1: Normal Probability Plot (Quantile-Quantile / Q-Q Plot)

A Normal Probability Plot arranges ordered sample data values against the theoretical percentiles of a standard normal distribution:

• If the plotted data points form an approximately straight diagonal line, the normality assumption is satisfied.
• An S-shaped curve indicates heavy tails (leptokurtic) or light tails (platykurtic).
• A concave or convex (banana-shaped) curve indicates severe skewness (asymmetry in the process).

      NORMAL DISTRIBUTION               SKEWED (NON-NORMAL) DISTRIBUTION
   +-------------------------+            +-------------------------+
   |                   *     |            |                   * * * |  <-- Tail diverges
   |                *        |            |              * *        |
   |             *           |            |         * *             |
   |          *              |            |     * *                 |
   |       *                 |            |  *                      |
   |    *                    |            | *                       |  <-- Sharp curvature
   +-------------------------+            +-------------------------+
      Linear = Normal Data                   Curved = Skewed Data

Method 2: The Anderson-Darling ($A^2$) Test

The Anderson-Darling test is the statistical industry standard for testing normality in quality engineering. It compares the empirical cumulative distribution function of your sample data to the theoretical normal cumulative distribution, placing substantial mathematical weight on the tails of the distribution where capability failures occur.

• Hypotheses:
  • $H_0$ (Null Hypothesis): The data follow a normal distribution.
  • $H_a$ (Alternative Hypothesis): The data do not follow a normal distribution.
• Decision Rule (using significance level $\alpha = 0.05$):
  • If $p\text{-value} \ge 0.05$: Fail to reject $H_0$. The data can be treated as normally distributed, and standard capability analysis may proceed.
  • If $p\text{-value} < 0.05$: Reject $H_0$. The data depart significantly from normality. Standard $C_p$ and $C_{pk}$ formulas cannot be used directly without data transformation or non-normal capability methods.

---

Sampling Requirements and Rational Subgrouping for Capability Studies

Conducting a statistically defensible capability study requires strict adherence to sample size protocols and rational subgrouping logic.

Sample Size Guidelines

A capability study based on a handful of parts is statistically meaningless due to sampling error. According to standard ASQ and AIAG guidelines:

• Minimum Number of Subgroups: A baseline capability study must collect at least 20 to 25 rational subgroups ($k \ge 20\text{ to }25$).
• Subgroup Size: Typically $n = 4$ or $n = 5$ consecutive units per subgroup.
• Total Sample Size: The overall dataset must contain a minimum of $N \ge 100$ to $125$ individual parts ($25 \text{ subgroups} \times 5 \text{ parts} = 125 \text{ parts}$).

Why Large Sample Sizes are Required

The standard deviation $\sigma$ is an estimate subject to sampling error. In small samples ($N < 50$), the confidence interval surrounding the calculated capability index is extremely wide. For example, a calculated $C_{pk}$ of 1.33 based on only 30 parts might have a 95% confidence interval ranging from 1.02 to 1.64. By collecting at least 100 to 125 parts across 25 subgroups, the confidence interval tightens significantly, providing high statistical confidence in production capability.

Rational Subgrouping Logic

Data must be collected using rational subgrouping:

• Within each subgroup: Samples must be produced consecutively over a very short time interval under virtually identical conditions (same operator, same raw material coil, same machine setting). This ensures that variation within the subgroup represents pure random common cause variation.
• Between subgroups: Subgroups should be collected periodically across shifts, days, and material batches. This allows control charts to detect whether between-subgroup shifts or tool wear (assignable causes) are occurring over time.

---

Step-by-Step Worked Numerical Examples

Worked Example 1: Comparing Natural Process Limits to Engineering Specifications

Scenario: A precision machine shop manufactures hardened steel bushings on a CNC horizontal lathe. The engineering drawing specifies an outer diameter of $35.000 \pm 0.045\text{ mm}$ ($USL = 35.045\text{ mm}, LSL = 34.955\text{ mm}$).

A capability study is conducted after confirming the process is in statistical control. The study of 25 subgroups of $n = 5$ parts yields a grand average of $\bar{\bar{X}} = 35.005\text{ mm}$ and an estimated process standard deviation of $\hat{\sigma} = 0.009\text{ mm}$.

Step 1: Calculate the Engineering Specification Width

$\text{Tolerance Width} = USL - LSL = 35.045 - 34.955 = 0.090\text{ mm}$

Step 2: Calculate the Natural Process Limits

$\text{UNPL} = \bar{\bar{X}} + 3\hat{\sigma} = 35.005 + 3(0.009) = 35.005 + 0.027 = 35.032\text{ mm}$ $\text{LNPL} = \bar{\bar{X}} - 3\hat{\sigma} = 35.005 - 3(0.009) = 35.005 - 0.027 = 34.978\text{ mm}$

Step 3: Calculate the Natural Process Spread

$\text{Natural Process Spread} = 6\hat{\sigma} = 6 \times 0.009 = 0.054\text{ mm}$

Step 4: Compare Natural Spread to Specification Width

$\text{Comparison: } 6\hat{\sigma} = 0.054\text{ mm} < 0.090\text{ mm} = (USL - LSL)$

• The natural process spread ($0.054\text{ mm}$) consumes only $\frac{0.054}{0.090} \times 100% = 60.0%$ of the total available tolerance band.
• Conclusion: The process is potentially capable. Furthermore, because both natural limits ($\text{LNPL} = 34.978\text{ mm}$, $\text{UNPL} = 35.032\text{ mm}$) fall comfortably inside the specification band ($34.955\text{ mm}$ to $35.045\text{ mm}$), the process is currently producing virtually 100% conforming parts.

---

Worked Example 2: Interpreting Normality Testing Output

Scenario: A quality technician evaluates two separate production lines making hydraulic piston rings. The technician runs an Anderson-Darling normality test on 100 parts from each line at $\alpha = 0.05$:

• Line A: Anderson-Darling statistic $A^2 = 0.312$, $p\text{-value} = 0.564$.
• Line B: Anderson-Darling statistic $A^2 = 1.945$, $p\text{-value} = 0.002$.

Technical Assessment:

• Line A Evaluation: Since the $p\text{-value} (0.564) \ge 0.05$, the technician fails to reject the null hypothesis of normality. The data from Line A can be safely treated as normally distributed. Standard capability formulas ($C_p, C_{pk}$) may be calculated with confidence.
• Line B Evaluation: Since the $p\text{-value} (0.002) < 0.05$, the technician rejects the null hypothesis. The data from Line B show statistically significant departure from normality. Standard capability formulas must not be reported; the technician must investigate physical root causes (e.g., tool wear, mixed material batches, or one-sided mechanical boundaries) or apply non-normal capability methods (such as Box-Cox transformations or percentile techniques).

---

Technician Inspection Scenarios & Common Exam Traps

Real-World Shop Scenario: Automotive Transmission Valve Body Bore

A quality technician at a transmission plant is tasked with qualifying a multi-spindle drilling machine. Production is eager to begin shipping parts and demands an immediate $C_{pk}$ calculation after machining only 12 valve bodies. The technician reviews the data and observes that the range chart is showing an upward trend, with two points exceeding the upper control limit ($UCL_R$).

If the technician yields to pressure and calculates $C_{pk}$, they violate basic quality engineering standards. The out-of-control range chart indicates that tool chatter or thermal expansion is actively destabilizing within-subgroup variation. The technician must halt the capability study, inform engineering of the assignable cause, verify that corrective actions stabilize the process, and collect a fresh sequence of 25 subgroups before releasing a formal capability report.

Common Exam Traps for CQT Candidates

• Exam Trap 1: Calculating Capability on Unstable Processes: Any exam question describing a control chart with out-of-control points, runs, or trends that asks for $C_p$ or $C_{pk}$ is testing this prerequisite. The correct technical response is always: The process must be brought into statistical control before capability can be determined.
• Exam Trap 2: Confusing Control Limits with Natural Process Limits: Control limits apply to subgroup averages ($\pm 3\sigma/\sqrt{n}$) and are plotted on control charts. Natural process limits apply to individual parts ($\pm 3\sigma$) and represent the actual spread of individual product output.
• Exam Trap 3: Confusing Specification Limits with Control Limits: Specification limits represent customer blueprints (VOC) and are never calculated from process data. Control limits represent the inherent variation of the machine (VOP) and are never taken from engineering drawings.
• Exam Trap 4: Assuming High Sample Size Fixes Non-Normality: The Central Limit Theorem states that subgroup averages ($\bar{X}$) tend toward normality as sample size increases, but it does not normalize the distribution of individual part measurements ($X$). Capability is concerned with individual parts meeting specifications, so individual non-normality remains a critical concern regardless of sample size.