1.3 Applied Engineering Statistics, Population Distributions & Quality Metrics

1.3 Applied Engineering Statistics, Population Distributions & Quality Metrics

Quick Answer: Statistical quality control in welding engineering prevents catastrophic service failures by quantifying variability. For small test coupon lots ($n < 30$), sample variance requires Bessel's correction ($n - 1$). Process capability is evaluated via $C_p = (\text{USL} - \text{LSL})/(6s)$ and centered capability $C_{pk} = \min[(\text{USL} - \bar{x})/(3s), (\bar{x} - \text{LSL})/(3s)]$. Automated lines target $C_{pk} \ge 1.33$ ($< 63\text{ PPM}$ defects). Linear regression ($y = mx + b$) models dilution as a function of heat input, quantified by $R^2$.

---

Descriptive Statistics & Degrees of Freedom in Mechanical Testing

Welding qualification standards—such as AWS B2.1, AWS D1.1, and ASME Section IX—require mechanical testing of weldments (transverse tensile tests, all-weld-metal tensile tests, side/root/face bend tests, and Charpy V-Notch impact tests). Because welding involves metallurgical solidification, variable thermal cycles, and microstructural heterogeneity, test data inevitably exhibit scatter.

Measures of Central Tendency

• Sample Mean ($\bar{x}$): The arithmetic average of $n$ observations: $$\bar{x} = \frac{1}{n} \sum_{i=1}^n x_i$$
• Median: The middle value of an ordered dataset. For an odd sample size, it is the exact center; for an even sample size, it is the average of the two central numbers.

Measures of Dispersion and Bessel's Correction

When assessing a full population of size $N$, variance is divided by $N$. However, welding engineers almost exclusively test small samples ($n = 3\text{ to }5$ qualification coupons). Calculating variance with $n$ creates a systematically biased underestimation of population variance because sample deviations are measured from the sample mean $\bar{x}$ rather than the true population mean $\mu$.

To correct for this bias, Bessel's correction uses $(n - 1)$ degrees of freedom in the denominator:

$$\text{Sample Variance } s^2 = \frac{1}{n - 1} \sum_{i=1}^n (x_i - \bar{x})^2$$ $$\text{Sample Standard Deviation } s = \sqrt{\frac{\sum_{i=1}^n (x_i - \bar{x})^2}{n - 1}}$$

   Degrees of Freedom: ν = n - 1
   Why? One degree of freedom is lost in estimating the sample mean x̄.
   Dividing by n on small samples (n = 3) underestimates true variance by 33%!

Skewness in Charpy V-Notch (CVN) Toughness Data

Tensile and yield strength test results generally follow a symmetric Gaussian normal distribution. In contrast, Charpy V-Notch (CVN) absorbed energy data frequently violate normality assumptions, especially when tested within the Ductile-to-Brittle Transition Temperature (DBTT) regime.

   Frequency
       |
       |    Cleavage / Brittle Mode
       |        *           * Microvoid Coalescence / Ductile Mode
       |       ***         ***
       |      *****       *****
       |     *******     *******  <-- Bimodal Skewed Distribution
     0 +----+-------+---+-------+---- Energy (Joules)

Within the transition range, microscopic variations in grain size or the presence of a single local brittle zone (LBZ, such as grain-coarsened HAZ) cause some specimens to fail by transgranular cleavage (absorbing only $15\text{ J}$) while adjacent specimens fail by ductile shear (absorbing $120\text{ J}$). This creates a bimodal or heavily skewed distribution where the arithmetic mean is misleading. Codes recognize this phenomenon:

• AWS D1.1 and ASME Section IX require testing sets of three specimens, stipulating both an allowable average energy (e.g., $27\text{ J}$ at $-20^\circ\text{C}$) and an absolute individual minimum (e.g., no single specimen below $20\text{ J}$).

---

Gaussian Normal Distribution, Z-Scores & Confidence Intervals

For large populations of automated welds, dimensional attributes (bead width, penetration, throat size) and ultimate tensile strengths conform to the Gaussian probability density function:

$$f(x) = \frac{1}{\sigma \sqrt{2\pi}} \exp\left( -\frac{(x - \mu)^2}{2\sigma^2} \right)$$

Standard Normal Variate (Z-Score)

To evaluate probabilities across arbitrary means $\mu$ and standard deviations $\sigma$, observations are normalized to standard normal variable $Z \sim N(0, 1)$:

$$Z = \frac{x - \mu}{\sigma} \quad \text{or for sample data:} \quad Z = \frac{x - \bar{x}}{s}$$

The $Z$-score represents the exact number of standard deviations an individual test value lies from the mean.

Confidence Intervals for the Population Mean

When certifying that a welding procedure consistently meets code tensile minimums, the engineer constructs a two-sided confidence interval for the population mean $\mu$:

$$\bar{x} - t_{\alpha/2, \nu} \left(\frac{s}{\sqrt{n}}\right) \le \mu \le \bar{x} + t_{\alpha/2, \nu} \left(\frac{s}{\sqrt{n}}\right)$$

where:

• $s / \sqrt{n}$ = Standard Error of the Mean ($SE$)
• $t_{\alpha/2, \nu}$ = Critical value from Student's $t$-distribution with $\nu = n - 1$ degrees of freedom at significance level $\alpha$ (for $95\%$ confidence, $\alpha = 0.05$).

As sample size grows ($n \ge 30$), the $t$-distribution converges to the standard normal distribution ($Z_{0.025} = 1.960$).

---

Process Capability Indices (Cp and Cpk) in Automated & Robotic Welding

In high-volume manufacturing (automotive body-in-white GMAW, structural beam submerged arc lines, pipe mills), quality cannot be inspected into the product through post-weld NDE; it must be designed into the welding process. Process Capability Indices quantify whether an automated welding cell is statistically capable of manufacturing within specification limits.

          LSL                       Target                      USL
           |                          |                          |
           |                     |--- μ ---|                     |
           |                   /             \                   |
           |                 /                 \                 |
           |               /                     \               |
           |             /                         \             |
      -----+-----------+-----------------------------+-----------+-----
           |           |<---------- 6σ ---------->|  |           |
           |<-- D_l -->|                             |<-- D_u -->|

Potential Process Capability ($C_p$)

$C_p$ evaluates the spread of the process against the total engineering tolerance width, ignoring whether the process is centered:

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

where $\text{USL}$ is the Upper Specification Limit and $\text{LSL}$ is the Lower Specification Limit.

Actual Process Capability Index ($C_{pk}$)

$C_{pk}$ penalizes the index for process mean shift away from the nominal midpoint:

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

One-Sided Specifications ($C_{pk} = C_{pl}$ or $C_{pu}$)

Many welding engineering requirements have only a single specification limit. For example:

• Minimum Specified Tensile Strength (SMTS): Only an $\text{LSL}$ exists.
• Minimum weld throat thickness: Only an $\text{LSL}$ exists.
• Maximum HAZ hardness (e.g., NACE MR0175 limit of $250\text{ HV} / 22\text{ HRC}$ for sour service): Only a $\text{USL}$ exists.

For a lower specification limit only:

$$C_{pk} = C_{pl} = \frac{\bar{x} - \text{LSL}}{3s}$$

Quality Benchmark Thresholds

---

Ordinary Least Squares Linear Regression for Process Modeling

Welding engineers frequently establish empirical predictive relationships between independent operating variables ($x$, such as heat input or wire feed speed) and dependent quality outputs ($y$, such as weld metal dilution percentage, penetration depth, or ferrite number).

Linear Model Formulation

$$\hat{y} = m x + b$$

where $m$ is the slope and $b$ is the $y$-intercept.

Using the method of Ordinary Least Squares (OLS) to minimize the sum of squared vertical residuals $\sum (y_i - \hat{y}_i)^2$:

$$m = \frac{n \sum (x_i y_i) - (\sum x_i)(\sum y_i)}{n \sum x_i^2 - (\sum x_i)^2}$$ $$b = \bar{y} - m \bar{x} = \frac{\sum y_i - m \sum x_i}{n}$$

Coefficient of Determination ($R^2$)

The coefficient of determination quantifies the proportion of total variation in dependent variable $y$ explained by linear regression on $x$:

$$R^2 = \frac{\left[ n \sum (x_i y_i) - (\sum x_i)(\sum y_i) \right]^2}{\left[ n \sum x_i^2 - (\sum x_i)^2 \right] \left[ n \sum y_i^2 - (\sum y_i)^2 \right]}$$

• $R^2 = 1.0$: Perfect linear correlation; $100\%$ of dilution variation is dictated by heat input.
• $R^2 = 0.0$: No linear relationship; parameter $x$ provides zero predictive utility.

---

Comprehensive Worked Numerical Example: Automated Fillet Weld Capability Analysis

Problem Statement

An automotive chassis facility uses robotic GMAW to produce thousands of lap fillet welds per shift. The structural design drawings establish the effective fillet weld leg size specification as:

• Nominal leg size: $8.0\text{ mm}$
• Lower Specification Limit ($\text{LSL}$): $7.2\text{ mm}$
• Upper Specification Limit ($\text{USL}$): $9.2\text{ mm}$

A quality audit randomly extracts $n = 35$ macro-etched cross-sections over one production week. Statistical analysis of the leg size data yields:

• Sample Mean $\bar{x} = 7.65\text{ mm}$
• Sample Standard Deviation $s = 0.20\text{ mm}$

Calculate: (a) Potential process capability $C_p$, (b) Upper capability $C_{pu}$ and Lower capability $C_{pl}$, (c) Actual process capability index $C_{pk}$, (d) The $Z$-score for the critical specification limit, and (e) The expected defect non-conformance rate in Parts Per Million (PPM) assuming a normal distribution.

Step-by-Step Solution

Step 1: Calculate potential process capability ($C_p$)

$$C_p = \frac{\text{USL} - \text{LSL}}{6s} = \frac{9.2\text{ mm} - 7.2\text{ mm}}{6(0.20\text{ mm})} = \frac{2.00}{1.20} = 1.667$$

Interpretation: With $C_p = 1.67$, the total intrinsic process spread ($6s = 1.20\text{ mm}$) is well within the $2.00\text{ mm}$ tolerance band. The cell possesses excellent inherent precision.

Step 2: Calculate upper ($C_{pu}$) and lower ($C_{pl}$) capability components

$$C_{pu} = \frac{\text{USL} - \bar{x}}{3s} = \frac{9.2\text{ mm} - 7.65\text{ mm}}{3(0.20\text{ mm})} = \frac{1.55}{0.60} = 2.583$$ $$C_{pl} = \frac{\bar{x} - \text{LSL}}{3s} = \frac{7.65\text{ mm} - 7.20\text{ mm}}{3(0.20\text{ mm})} = \frac{0.45}{0.60} = 0.750$$

Step 3: Determine actual process capability index ($C_{pk}$)

$$C_{pk} = \min(C_{pu}, \, C_{pl}) = \min(2.583, \, 0.750) = 0.750$$

Engineering Evaluation: Because $C_{pk} = 0.750 < 1.0$, the process is statistically incapable! Even though process spread is small ($C_p = 1.67$), the process mean has shifted severely downward toward the $\text{LSL}$ ($\bar{x} = 7.65\text{ mm}$ vs. target $8.20\text{ mm}$), resulting in non-conforming undersized welds.

Step 4: Compute the $Z$-score for the Lower Specification Limit

$$Z_{\text{LSL}} = \frac{\text{LSL} - \bar{x}}{s} = \frac{7.20 - 7.65}{0.20} = \frac{-0.45}{0.20} = -2.25$$

Step 5: Determine expected non-conformance rate (PPM) Referring to standard cumulative normal distribution tables for $Z = -2.25$:

$$P(Z < -2.25) = 0.01222 = 1.222\%$$

Converting to Parts Per Million (PPM):

$$\text{Defect Rate} = 0.01222 \times 1,000,000 = 12,220\text{ PPM}$$

Corrective Action: The welding engineer adjusts the robotic program by increasing wire feed speed by $6\%$ or decreasing travel speed from $650\text{ mm/min}$ to $610\text{ mm/min}$ to shift the mean to $\bar{x} = 8.20\text{ mm}$ (dead center). Once centered, $C_{pk} = C_p = 1.67$, reducing defects to $< 1\text{ PPM}$.

---

Real-World Engineering Scenarios & Exam Pitfalls

Practical Industrial Scenario

A pipeline contractor welding high-strength X70 line pipe experienced unexpected bend test failures during procedure qualification. Five Charpy specimens tested at $-30^\circ\text{C}$ produced absorbed energy values of: $18\text{ J}$, $22\text{ J}$, $85\text{ J}$, $92\text{ J}$, and $88\text{ J}$. The junior engineer calculated the simple average as $(18 + 22 + 85 + 92 + 88)/5 = 61\text{ J}$, concluding the procedure passed the specification average of $45\text{ J}$. The CWEng auditor rejected the PQR: the two low values indicated cleavage fracture initiated by coarse-grained heat-affected zone local brittle zones (LBZs). In a small sample, extreme skewness masked an unacceptably high probability of low-energy brittle fracture under subsea operating stresses.

Common Exam Traps

Exam Trap 1: Dividing by n Instead of (n - 1) for Sample Variance When calculating standard deviation on a coupon test set of 3 to 5 specimens, candidates often use $n$ (population formula) rather than $n - 1$ (sample formula). On $n = 3$, this introduces a massive error of $\sqrt{3/2} = 1.22$ ($22\%$ underestimation of standard deviation), falsely inflating $C_{pk}$ and failing the exam question.

Exam Trap 2: Confusing Cp with Cpk Questions frequently ask: "A process has $C_p = 1.80$. Can you state that the process produces zero non-conforming welds?" The answer is NO. $C_p$ measures only the tolerance width relative to $6\sigma$. If the mean drifts outside the tolerance band, a process with $C_p = 2.0$ could produce $100\%$ defective welds! $C_{pk}$ is always required to establish true compliance.

Exam Trap 3: Applying Symmetric Two-Sided Capability to One-Sided Code Limits In tension testing, codes specify a minimum tensile strength (e.g., AWS D1.1 Table 5.4 mandates $485\text{ MPa}$ for E70 filler metal). There is no upper limit on tensile strength unless notch toughness or seismic overstrength criteria apply. Do not attempt to compute $C_p = (\text{USL} - \text{LSL})/(6s)$ when $\text{USL}$ does not exist. Use the one-sided index $C_{pl} = (\bar{x} - \text{LSL})/(3s)$.