21.1 Process Capability Indices and Defect Levels

Process capability analysis bridges statistical quality control and design engineering. While statistical process control (SPC) charts monitor an active process to maintain stability over time, capability analysis evaluates whether that stable process can consistently produce parts that conform to customer engineering specifications. On the FE Industrial and Systems exam, mastery of capability formulas ($C_p$, $C_{pk}$, $P_p$, $P_{pk}$), the centering parameter $k$, standard normal $Z$-score defect conversions, and the distinction between short-term within-subgroup variation and long-term overall variation is essential.

---

1. Preconditions for Process Capability Analysis

Before computing any process capability index, an industrial engineer must satisfy two non-negotiable statistical preconditions:

               Prerequisites for Valid Capability Analysis
               
    ┌───────────────────────────┐     ┌───────────────────────────┐
    │   Statistical Stability   │     │    Normality of Data      │
    │   (In Statistical Control)│     │   X ~ N(mu, sigma^2)      │
    └─────────────┬─────────────┘     └─────────────┬─────────────┘
                  │                                 │
                  ▼                                 ▼
          No Assignable Causes              Goodness-of-Fit / Normality
          (Common Causes Only)              (Anderson-Darling / W-Test)
                  │                                 │
                  └────────────────┬────────────────┘
                                   ▼
                   Proceed to Calculate Cp, Cpk, PPM

1. The Process Must Be in Statistical Control

A process is in a state of statistical control when it exhibits only natural, random, common-cause variation and is free from assignable (special) causes. Stability must be verified using Shewhart control charts (such as $\bar{X}-R$ or $\bar{X}-s$ charts) over an extended baseline period (typically at least 20 to 25 rational subgroups):

• All subgroup sample points must fall strictly within the 3-sigma control limits ($UCL$ and $LCL$).
• The plotted points must exhibit random scatter without non-random patterns, runs, trends, cycles, or stratification.
• Exam Rule: If a process is not in statistical control, calculating $C_p$ or $C_{pk}$ is completely invalid. An unstable process has a shifting mean $\mu(t)$ and/or time-varying variance $\sigma^2(t)$; predicting future fraction defective or conformance percentage from an unstable distribution is mathematically meaningless.

2. Normality of the Quality Characteristic

Standard capability indices assume that the underlying continuous quality characteristic $X$ is normally distributed: $X \sim N(\mu, \sigma^2)$.

• Normality should be evaluated using normal probability plots and statistical goodness-of-fit tests (such as the Anderson-Darling, Shapiro-Wilk, or Kolmogorov-Smirnov tests).
• If the distribution is significantly skewed or non-normal (such as flatness, roundness, surface roughness, or cycle times), standard formulas produce severe errors in estimated defect levels. In non-normal cases, the industrial engineer must either:
  1. Transform the raw data into an approximately normal distribution using mathematical transformations (e.g., Box-Cox power transformation or Johnson system of distributions).
  2. Apply non-parametric percentile-based capability models, where natural process width is defined by the 0.135th and 99.865th percentiles ($P_{0.135}$ and $P_{99.865}$): $C_{pk} = \min\left[ \frac{\text{USL} - P_{50}}{P_{99.865} - P_{50}}, \, \frac{P_{50} - \text{LSL}}{P_{50} - P_{0.135}} \right]$

---

2. Potential Process Capability Index ($C_p$)

The potential capability index ($C_p$) evaluates the inherent precision of a manufacturing process. It compares the allowable engineering tolerance band against the natural dispersion of the process, completely disregarding where the process distribution is centered:

$C_p = \frac{\text{USL} - \text{LSL}}{6\sigma}$

Where:

• $\text{USL}$ = Upper Specification Limit (customer-defined maximum allowable dimension)
• $\text{LSL}$ = Lower Specification Limit (customer-defined minimum allowable dimension)
• $\text{USL} - \text{LSL}$ = Total specification tolerance band ("Voice of the Customer")
• $6\sigma$ = Natural process spread ("Voice of the Process"), containing $99.73%$ of output under a normal distribution
• $\sigma$ = Inherent process standard deviation

                   Potential Capability (Cp) Anatomy

              LSL                  Target                  USL
               │                     │                      │
               ├─────────────────────┼──────────────────────┤
               │             Tolerance = USL - LSL          │
               │                                            │
                        ┌───┐
                      ┌─┘   └─┐
                    ┌─┘       └─┐       Process Spread = 6*sigma
                   ┌┘           └┐      (Voice of the Process)
               ────┴─────────────┴────────────────────────────
               -3sigma    mu    +3sigma

Properties and Limitations of $C_p$

1. Theoretical Upper Bound: $C_p$ represents the maximum capability that the process could achieve if its operating mean $\mu$ were adjusted to sit perfectly at the nominal specification midpoint $M = \frac{\text{USL} + \text{LSL}}{2}$.
2. Insensitivity to Location: $C_p$ does not depend on the process mean $\mu$. A machining center could have an outstanding $C_p = 2.50$, but if its tool offset is improperly calibrated such that $\mu > \text{USL}$, every single manufactured part will be oversized scrap. Thus, $C_p$ measures potential, never actual conformance.
3. Bilateral Specifications Only: $C_p$ can only be calculated when both an upper and lower specification limit are defined. For unilateral (one-sided) specifications, $C_p$ is undefined.

---

3. Actual Process Capability Index ($C_{pk}$) and Centering Factor ($k$)

To account for process centering relative to specification limits, industrial engineers compute the actual capability index ($C_{pk}$). $C_{pk}$ measures the clearance between the operating mean $\mu$ and the nearest specification limit, scaled against half the natural process spread ($3\sigma$).

Unilateral Capability Indices

First, define capability relative to each individual specification boundary:

• Upper Capability Index ($C_{pu}$): $C_{pu} = \frac{\text{USL} - \mu}{3\sigma}$
• Lower Capability Index ($C_{pl}$): $C_{pl} = \frac{\mu - \text{LSL}}{3\sigma}$

Defining Actual Capability ($C_{pk}$)

The overall actual capability index is the minimum of the unilateral indices:

$C_{pk} = \min(C_{pu}, \, C_{pl}) = \min\left( \frac{\text{USL} - \mu}{3\sigma}, \, \frac{\mu - \text{LSL}}{3\sigma} \right)$

Because $C_{pk}$ is bounded by the specification limit closest to the process mean, it directly reflects the worst-case tail probability of producing defective parts.

The Centering Factor ($k$)

The degree of process off-centering is quantified by the non-dimensional parameter $k$ (also called the bias factor). Let $M$ be the nominal specification midpoint:

$M = \frac{\text{USL} + \text{LSL}}{2}$

The centering factor $k$ expresses the distance between the process mean $\mu$ and midpoint $M$ as a fraction of the allowable half-tolerance band:

$k = \frac{|M - \mu|}{\frac{\text{USL} - \text{LSL}}{2}} = \frac{|\text{USL} + \text{LSL} - 2\mu|}{\text{USL} - \text{LSL}}$

The Fundamental Identity Linking $C_p$ and $C_{pk}$

Using $k$, the relationship between potential and actual capability is formulated as:

$C_{pk} = C_p (1 - k)$

                 Process Centering Regimes and k Values

     k = 0              0 < k < 1             k = 1             k > 1
  (Centered)          (Off-Center)        (On Spec Limit)    (Outside Spec)
       │                   │                     │                 │
   Cpk = Cp           0 < Cpk < Cp            Cpk = 0           Cpk < 0
  Zero Bias           Partial Shift       50% Defective      >50% Defective

Diagnostic Interpretation of $k$:

• $k = 0$ ($C_{pk} = C_p$): The process mean is perfectly centered at $M$. The process operates at its maximum theoretical capability.
• $0 < k < 1$ ($0 < C_{pk} < C_p$): The process mean is shifted away from target, but remains within the specification boundaries. As off-centering $|\mu - M|$ increases, $C_{pk}$ drops linearly.
• $k = 1$ ($C_{pk} = 0$): The process mean falls exactly on one of the specification limits ($\mu = \text{USL}$ or $\mu = \text{LSL}$). In this condition, exactly $50%$ of production violates that specification limit under a normal distribution.
• $k > 1$ ($C_{pk} < 0$): The process mean lies entirely outside the specification limits. More than half of total production is defective.

Exam Rule: $C_{pk}$ can never exceed $C_p$ ($C_{pk} \le C_p$ always). If your calculation yields $C_{pk} > C_p$, an arithmetic error has occurred.

---

4. Interpretation of Capability Values and Defect Benchmarks

Industrial manufacturing systems classify process capability according to established thresholds. Conformance is traditionally quantified in parts per million (PPM) or defects per million opportunities (DPMO).

                 Process Tolerance vs. Specification Limits

  Cpk = 1.00 (3-sigma): Spec width = 6*sigma
  LSL                           Nominal                           USL
   ├───────────────┬───────────────┼───────────────┬───────────────┤
   │◄───────────── 3*sigma ───────►│◄───────────── 3*sigma ───────►│
  Defects: 1350 PPM                                 Defects: 1350 PPM

  Cpk = 1.33 (4-sigma): Spec width = 8*sigma (Safety Margin = 1*sigma per side)
  LSL                           Nominal                           USL
   ├─────────┬─────┬───────────────┼───────────────┬─────┬─────────┤
   │ Margin  │◄─── 3*sigma ───────►│◄───────────── 3*sigma ───────►│ Margin  │
  Defects: 31.7 PPM                                 Defects: 31.7 PPM

  Cpk = 2.00 (6-sigma): Spec width = 12*sigma (Safety Margin = 3*sigma per side)
  LSL                           Nominal                           USL
   ├─── Margin ────┬───────────────┼───────────────┬──── Margin ───┤
   │◄── 3*sigma ──►│◄─── 3*sigma ─►│◄─── 3*sigma ─►│◄── 3*sigma ──►│

The Motorola 1.5-Sigma Shift and 3.4 DPMO

In standard Six Sigma methodology (developed at Motorola), real-world manufacturing processes are observed to experience non-random long-term drifts due to temperature cycles, tool wear, bearing play, and raw material batch variance. This drift is empirically modeled as a $1.5\sigma$ shift in the process mean over the long run:

• If a process has short-term capability of $C_p = 2.00$, the specification limits lie at $\pm 6\sigma$ from nominal.
• When the process mean drifts by $1.5\sigma$ toward the upper limit, the clearance to the nearest limit decreases from $6\sigma$ to $6\sigma - 1.5\sigma = 4.5\sigma$.
• The resulting capability drops to: $C_{pk} = \frac{4.5\sigma}{3\sigma} = 1.50$
• The probability of exceeding the near specification limit is: $P(Z > 4.5) = 1 - \Phi(4.5) = 3.39767 \times 10^{-6} \approx 3.4 \times 10^{-6}$
• This corresponds to exactly $3.4\text{ DPMO}$ (defects per million opportunities).

Exam Trap: A centered 6-sigma process produces 0.002 PPM (2 parts per billion). The famous 3.4 DPMO figure applies only when assuming the standard $1.5\sigma$ off-center drift!

---

5. Short-Term Capability ($C_p, C_{pk}$) vs. Long-Term Performance ($P_p, P_{pk}$)

A frequent source of confusion on the FE exam is the operational distinction between capability ($C_p, C_{pk}$) and performance ($P_p, P_{pk}$). The difference lies entirely in how process variability ($\sigma$) is estimated.

Engineering Diagnostics Using $C_{pk}$ and $P_{pk}$

Because total variation incorporates all sources of disturbance, the sample standard deviation is virtually always larger than within-subgroup variation: $s \ge \hat{\sigma}_{\text{within}}$. Consequently: $P_p \le C_p \quad \text{and} \quad P_{pk} \le C_{pk}$

• Case 1 ($C_{pk} \approx P_{pk}$): The process is operating with negligible between-subgroup drift. Total variation is dominated by pure white-noise within-subgroup variation. The process is stable and predictable.
• Case 2 ($C_{pk} \gg P_{pk}$): Inherent machine capability is excellent, but significant assignable causes, day-to-day temperature fluctuations, raw material batch shifts, or setup differences exist between subgroups. Improvement efforts must focus on process standardization, setup controls, and environmental regulation rather than purchasing a higher-precision machine.

---

6. Calculating Expected Defect Rates (PPM) Using Z-Scores

To compute the expected fraction defective ($p$) or defect rate in parts per million (PPM) for a normally distributed process, calculate the standard normal $Z$-scores relative to each specification limit:

Upper and Lower Z-Scores

$Z_{\text{USL}} = \frac{\text{USL} - \mu}{\sigma} = 3 C_{pu}$ $Z_{\text{LSL}} = \frac{\mu - \text{LSL}}{\sigma} = 3 C_{pl}$

Notice the direct relationship between $Z$-scores and unilateral capability indices: the $Z$-score distance to a specification limit is identically three times the corresponding unilateral capability index ($Z = 3C$).

Computing Tail Probabilities

Using the cumulative standard normal distribution function $\Phi(z) = P(Z \le z)$:

• Fraction Defective Above USL ($p_{\text{upper}}$): $p_{\text{upper}} = P(X > \text{USL}) = 1 - \Phi(Z_{\text{USL}})$
• Fraction Defective Below LSL ($p_{\text{lower}}$): $p_{\text{lower}} = P(X < \text{LSL}) = P(Z < -Z_{\text{LSL}}) = 1 - \Phi(Z_{\text{LSL}})$
• Total Fraction Defective ($p_{\text{total}}$): $p_{\text{total}} = p_{\text{upper}} + p_{\text{lower}} = [1 - \Phi(Z_{\text{USL}})] + [1 - \Phi(Z_{\text{LSL}})]$
• Expected Defect Rate (PPM): $\text{PPM} = p_{\text{total}} \times 10^6$

---

7. Step-by-Step Worked Engineering Calculations

Worked Example 21.1.1: Precision Shaft Grinding Capability Analysis

Problem: A precision automotive grinding cell manufactures cylindrical transmission shafts with an outside diameter specification of $25.000 \pm 0.080\text{ mm}$ (hence $\text{LSL} = 24.920\text{ mm}$ and $\text{USL} = 25.080\text{ mm}$). An industrial engineer gathers quality data from 25 rational subgroups of size $n = 5$ while the process is verified to be in statistical control. The grand mean and average range are:

• Grand average: $\bar{\bar{X}} = 25.025\text{ mm}$
• Average range: $\bar{R} = 0.04652\text{ mm}$
• Control chart factor for $n = 5$: $d_2 = 2.326$
• An overall pooled calculation of all 125 parts yields a sample standard deviation of $s = 0.0260\text{ mm}$.

Required:

1. Estimate the within-subgroup process standard deviation $\hat{\sigma}$.
2. Compute the potential capability index $C_p$ and actual capability index $C_{pk}$.
3. Compute the centering factor $k$ and verify the identity $C_{pk} = C_p(1 - k)$.
4. Compute the expected defect rate in parts per million (PPM) for both upper and lower tails.
5. Compute the long-term performance indices $P_p$ and $P_{pk}$, and interpret the process behavior.

Solution:

Step 1: Estimate Within-Subgroup Standard Deviation $\hat{\sigma} = \frac{\bar{R}}{d_2} = \frac{0.04652\text{ mm}}{2.326} = 0.0200\text{ mm}$

Step 2: Calculate Potential Capability ($C_p$)

• Specification tolerance: $\text{USL} - \text{LSL} = 25.080 - 24.920 = 0.160\text{ mm}$
• Process spread: $6\hat{\sigma} = 6 \times 0.0200 = 0.120\text{ mm}$ $C_p = \frac{\text{USL} - \text{LSL}}{6\hat{\sigma}} = \frac{0.160}{0.120} = 1.333$ Interpretation: The inherent variability of the grinding machine is capable of achieving 4-sigma performance ($C_p \ge 1.33$).

Step 3: Calculate Actual Capability ($C_{pk}$)

• Unilateral Upper Capability: $C_{pu} = \frac{\text{USL} - \bar{\bar{X}}}{3\hat{\sigma}} = \frac{25.080 - 25.025}{3(0.0200)} = \frac{0.055}{0.060} = 0.9167$
• Unilateral Lower Capability: $C_{pl} = \frac{\bar{\bar{X}} - \text{LSL}}{3\hat{\sigma}} = \frac{25.025 - 24.920}{3(0.0200)} = \frac{0.105}{0.060} = 1.7500$
• Actual Capability Index: $C_{pk} = \min(C_{pu}, \, C_{pl}) = \min(0.9167, \, 1.7500) = 0.9167 \approx 0.92$ Conclusion: Although potential capability is satisfactory ($C_p = 1.33$), the actual process is incapable ($C_{pk} = 0.92 < 1.00$) because the process mean is operating off-center toward the upper specification limit!

Step 4: Compute Centering Factor ($k$) and Verify Identity

• Specification midpoint: $M = \frac{25.080 + 24.920}{2} = 25.000\text{ mm}$
• Half-tolerance: $\frac{\text{USL} - \text{LSL}}{2} = \frac{0.160}{2} = 0.080\text{ mm}$
• Off-centering distance: $|\bar{\bar{X}} - M| = |25.025 - 25.000| = 0.025\text{ mm}$ $k = \frac{0.025}{0.080} = 0.3125$
• Verify identity: $C_{pk} = C_p (1 - k) = 1.3333 \times (1 - 0.3125) = 1.3333 \times 0.6875 = 0.9167$ The mathematical identity holds exactly.

Step 5: Compute Expected Defect Rate (PPM)

• Upper specification $Z$-score: $Z_{\text{USL}} = \frac{\text{USL} - \bar{\bar{X}}}{\hat{\sigma}} = \frac{25.080 - 25.025}{0.0200} = 2.75$ Using standard normal cumulative probability: $\Phi(2.75) = 0.99702$ $p_{\text{upper}} = 1 - 0.99702 = 0.00298 = 2,980\text{ PPM}$
• Lower specification $Z$-score: $Z_{\text{LSL}} = \frac{\bar{\bar{X}} - \text{LSL}}{\hat{\sigma}} = \frac{25.025 - 24.920}{0.0200} = 5.25$ Because $Z_{\text{LSL}} = 5.25 > 4.0$, lower tail defects are negligible: $p_{\text{lower}} = 1 - \Phi(5.25) \approx 7.9 \times 10^{-8} \approx 0.08\text{ PPM}$.
• Total expected defects: $\text{PPM}_{\text{total}} \approx 2,980\text{ PPM}$ Takeaway: Because the grinding wheel offset is off by only $+0.025\text{ mm}$, the line generates nearly $3,000\text{ PPM}$ oversized scrap. Adjusting the CNC wheel offset by $-0.025\text{ mm}$ to center the process at $25.000\text{ mm}$ immediately restores $C_{pk}$ to $1.333$, cutting total scrap to just $63.3\text{ PPM}$ without spending any capital on tooling.

Step 6: Compute Long-Term Performance Indices ($P_p, P_{pk}$) Using overall sample standard deviation $s = 0.0260\text{ mm}$: $P_p = \frac{\text{USL} - \text{LSL}}{6s} = \frac{0.160}{6 \times 0.0260} = \frac{0.160}{0.156} = 1.026$ $P_{pu} = \frac{\text{USL} - \bar{\bar{X}}}{3s} = \frac{25.080 - 25.025}{3 \times 0.0260} = \frac{0.055}{0.078} = 0.705$ $P_{pl} = \frac{\bar{\bar{X}} - \text{LSL}}{3s} = \frac{25.025 - 24.920}{3 \times 0.0260} = \frac{0.105}{0.078} = 1.346$ $P_{pk} = \min(0.705, 1.346) = 0.705$ Diagnostic: Note that $s = 0.0260 > \hat{\sigma} = 0.0200$, leading to $P_{pk} = 0.71 < C_{pk} = 0.92$. This confirms that additional long-term between-batch variation (such as grinding wheel thermal expansion and dress cycles) further degrades process performance over time.

---

8. NCEES Reference Handbook Tips & Realistic Exam Traps

• Trap: Confusing $C_p$ with $C_{pk}$: Questions frequently ask: "Is the process capable of meeting customer specifications?" If the process mean is off-center, evaluate $C_{pk}$, not $C_p$. $C_p$ measures only potential spread under ideal centering. Conformance requires $C_{pk} \ge 1.33$.
• Trap: Inverting Z-score Conversions: Remember that $Z_{\text{USL}} = 3 C_{pu}$ and $Z_{\text{LSL}} = 3 C_{pl}$. A common mistake is multiplying by 6 instead of 3 because $C_p$ has $6\sigma$ in its denominator. Unilateral capability indices evaluate half the distribution ($3\sigma$), so the multiplier is always 3.
• Trap: Estimating $\sigma$ from Subgroup Ranges: The NCEES handbook provides the divisor $d_2$ for subgroup ranges ($\hat{\sigma} = \bar{R}/d_2$) and $c_4$ for standard deviations ($\hat{\sigma} = \bar{s}/c_4$). Make sure you look up the factor corresponding to the subgroup sample size $n$ (typically $n = 4$ or $5$), not the number of subgroups $m$ ($m = 20$ or $25$).
• Trap: Negative $C_{pk}$ Values: Do not discard negative capability values as computational errors. If the process mean $\mu$ drifts outside the tolerance band ($\mu > \text{USL}$ or $\mu < \text{LSL}$), the numerator of either $C_{pu}$ or $C_{pl}$ is negative, resulting in $C_{pk} < 0$. This indicates that greater than $50%$ of production is defective.