20.1 Statistical Process Control Principles, Variation, and Run Rules

Statistical Process Control (SPC) is the operational methodology that applies statistical techniques to monitor, analyze, and regulate production processes. Formulated in 1924 by physicist and engineer Walter A. Shewhart of Bell Telephone Laboratories, SPC transformed modern industrial manufacturing from an era of reactive post-production inspection into a regime of proactive, real-time defect prevention. Shewhart recognized that no two manufactured parts are identical: variation is an intrinsic physical characteristic of all production environments. To maintain quality economically, an enterprise must distinguish between natural background noise and actionable operational disturbances.

---

1. Common Cause vs. Special Cause Variation

Every production system exhibits variability. Shewhart separated this total variability into two mutually exclusive, fundamentally distinct categories:

                    Total Process Variation
                               │
        ┌──────────────────────┴──────────────────────┐
        ▼                                             ▼
  Common Cause                                  Special Cause
 (Chance / Random)                             (Assignable)
────────────────────                          ─────────────────
• Inherent in equipment & design              • External shocks & disturbances
• Stable, stationary distribution             • Unstable, erratic shifts/trends
• Governed by laws of probability             • Caused by specific events
• Requires management system redesign         • Requires operator/engineer action

Common Cause (Chance / Inherent) Variation

Common cause variation represents the natural, random noise inherent in the design, construction, and daily operation of a machine or process. Examples include microscopic backlash in CNC lead screws, ambient room temperature fluctuations within HVAC limits, minor lot-to-lot chemical variations in raw polymer pellets, and normal operator pulse tremor.

• Statistical Character: A system influenced exclusively by common causes exhibits a stable distribution over time. Its mean, variance, and shape remain stationary and predictable within probabilistic bounds.
• Managerial Implication: Shop-floor operators cannot eliminate common cause variation through routine machine adjustments. Reducing common cause variation requires systemic management intervention: purchasing higher-precision tooling, upgrading environmental controls, or redesigning the product.

Special Cause (Assignable / Extraneous) Variation

Special cause variation arises from external, intermittent, or unexpected disturbances that are not part of the baseline system design. Examples include a fractured cutting insert, an untrained relief operator loading raw stock backwards, a batch of contaminated hydraulic fluid, a severe electrical surge, or a drifting temperature sensor.

• Statistical Character: Special causes disrupt the underlying probability distribution, causing unpredictable shifts in the process mean, expansions in process dispersion, or non-random cyclical oscillations.
• Operational Remedy: Special causes can be identified, isolated, and permanently removed by line technicians and industrial engineers. The immediate objective of SPC is to detect special causes rapidly so corrective action occurs before mass scrap is fabricated.

---

2. Statistical Control vs. Process Capability

A critical conceptual trap on the FE Industrial exam is conflating Statistical Control with Process Capability. These two metrics measure completely independent performance dimensions:

• Statistical Control (Voice of the Process): Answers the question: Is the process stable and predictable over time? It evaluates the internal consistency of the process using empirical control limits derived solely from process samples.
• Process Capability (Voice of the Customer): Answers the question: Does the process satisfy external customer engineering tolerances? It evaluates whether the natural 6-sigma process spread ($6\sigma$) fits inside the allowable tolerance interval ($USL - LSL$).

                 The Four Operating States Matrix

                     Statistical Control (Stability)
                        IN CONTROL        OUT OF CONTROL
                    ┌─────────────────┬─────────────────┐
                    │     STATE 1     │     STATE 3     │
           CAPABLE  │  Ideal State:   │  Threshold /    │
                    │  Predictable &  │  Precarious:    │
  Process           │   Zero Scrap    │ Passing, but    │
Capability          │                 │ Unpredictable   │
(Tolerances)        ├─────────────────┼─────────────────┤
                    │     STATE 2     │     STATE 4     │
          INCAPABLE │  Predictable    │ Chaos State:    │
                    │  Failure:       │  Unpredictable  │
                    │ Stable, but     │  & Generating   │
                    │ Steady Scrap    │  Heavy Scrap    │
                    └─────────────────┴─────────────────┘

Operational Analysis of the Four States

1. State 1: In Control and Capable (The Ideal State): The process exhibits only common-cause variation, and its $6\sigma$ natural spread is substantially narrower than the tolerance band ($C_{pk} \ge 1.33$). Defect rates are near zero and predictable.
2. State 2: In Control and Incapable (Predictable Failure): The process is statistically stable, but its natural spread exceeds customer tolerances ($C_{pk} < 1.0$) or the mean is badly off-center. Crucial Exam Concept: Adjusting machine knobs will not eliminate defects in State 2. Because the process is stable, scrap is produced at a constant, mathematically predictable rate. Only engineering redesign (better tooling, tighter feed systems) can make it capable.
3. State 3: Out of Control and Capable (The Precarious State): Individual parts currently meet engineering tolerances because customer specifications are exceptionally wide. However, special causes are actively shifting the process mean or expanding dispersion. It is only a matter of time before an unmonitored shift produces defects.
4. State 4: Out of Control and Incapable (The Chaos State): The process is unstable and generating unacceptable scrap. Immediate shutdown, containment, and root-cause troubleshooting are mandatory.

---

3. Control Limits Mechanics and the 3-Sigma Foundation

A Shewhart control chart plots a quality metric (such as subgroup average $\bar{X}$) along a vertical axis against sequential sample numbers along a horizontal time axis. It features three defining reference lines:

• Center Line ($CL$): Represents the historical mathematical average of the plotted statistic (e.g., $\mu_0$ or $\bar{\bar{X}}$).
• Upper Control Limit ($UCL$): $CL + 3\sigma_{\text{statistic}}$
• Lower Control Limit ($LCL$): $CL - 3\sigma_{\text{statistic}}$

                 Shewhart 3-Sigma Control Chart Topology

 UCL ────────────────────────────────────────────────────────── +3σ
                          Zone A (Upper)
 +2σ ---------------------------------------------------------- +2σ
                          Zone B (Upper)
 +1σ ---------------------------------------------------------- +1σ
                          Zone C (Upper)
  CL ══════════════════════════════════════════════════════════  0σ
                          Zone C (Lower)
 -1σ ---------------------------------------------------------- -1σ
                          Zone B (Lower)
 -2σ ---------------------------------------------------------- -2σ
                          Zone A (Lower)
 LCL ────────────────────────────────────────────────────────── -3σ

Why 3-Sigma Limits?

Shewhart selected 3-sigma limits based on economic optimization rather than pure mathematical convenience. In quality control, decision-makers face two competing risks:

1. Type I Error ($\alpha$, Producer's Risk / False Alarm): Concluding the process is out of control when it is actually operating stably under common causes. This triggers expensive, unnecessary production halts, tool teardowns, and engineering investigations.
2. Type II Error ($\beta$, Consumer's Risk / Missed Detection): Concluding the process is in control when an assignable cause has actually shifted the mean or expanded variance. This allows nonconforming parts to escape to downstream assembly or customers.

If control limits were set at $\pm 2\sigma$, the false alarm rate would surge to $\alpha = 0.0455$ (4.55% of samples), overwhelming production staff with phantom investigations. If limits were set at $\pm 4\sigma$, false alarms would drop, but the chart would fail to detect real process shifts (huge $\beta$ error).

Under an underlying standard normal distribution $Z \sim N(0, 1)$: $\alpha = P(|Z| > 3) = 2 \left[1 - \Phi(3)\right] = 2(0.00135) = 0.0027 \quad (0.27\% \text{ or } 2.7 \text{ per thousand})$

Average Run Length ($ARL$)

The Average Run Length ($ARL$) quantifies the expected number of subgroups plotted before an out-of-control signal is generated: $ARL = \frac{1}{P(\text{Signal})}$

• In-Control Average Run Length ($ARL_0$): When the process is operating in statistical control with pure common-cause variation: $ARL_0 = \frac{1}{\alpha} = \frac{1}{0.0027} \approx 370.4 \text{ samples}$ Interpretation: On average, a false alarm will occur once every 370 to 371 subgroups.
• Out-of-Control Average Run Length ($ARL_1$): When an assignable cause shifts the distribution such that the probability of a point exceeding control limits is $p = 1 - \beta$: $ARL_1 = \frac{1}{1 - \beta}$

The Golden Rule: Control Limits vs. Specification Limits

On the FE Industrial exam, questions frequently attempt to trick examinees into plotting Specification Limits on control charts:

• Specification Limits ($USL, LSL$): Engineering tolerances determined by design engineers, customer specifications, or regulatory codes. They apply strictly to individual items ($X_i$).
• Control Limits ($UCL, LCL$): Statistical thresholds calculated from sample statistics (such as subgroup averages $\bar{X}$). They represent the natural variability of the process.

CRITICAL EXAM TRAP: Specification limits (USL and LSL) must NEVER be drawn on an $\bar{X}$ or $R$ chart!

By the Central Limit Theorem, the standard deviation of subgroup averages is $\sigma_{\bar{X}} = \sigma / \sqrt{n}$, which is drastically smaller than the standard deviation of individual parts $\sigma$. Plotting individual-part specification limits on an average chart creates a false illusion of safety: an individual part can easily violate $USL$ even while the subgroup average $\bar{X}$ sits safely inside the control limits!

---

4. Out-of-Control Detection Rules: Western Electric and Nelson Rules

A process is not declared in statistical control simply because all points fall between $UCL$ and $LCL$. If points exhibit non-random patterns, trends, or clustering, assignable causes are active. To detect shifts and trends faster than waiting for a single point to breach the 3-sigma limits, industrial practice partitions each side of the control chart into three equal 1-sigma zones:

• Zone C: From the Center Line to $\pm 1\sigma$
• Zone B: From $\pm 1\sigma$ to $\pm 2\sigma$
• Zone A: From $\pm 2\sigma$ to $\pm 3\sigma$ (outermost zone)

The Western Electric Rules (WECO Rules)

Developed in the 1956 Western Electric Handbook, these four rules detect unnatural clustering:

                  Western Electric Rules Visualized

 UCL ──────────────────────────────────────────────────────────
               ● Rule 1: 1 point > 3σ
 +2σ ---------------------------------------------------------- Zone A
          ●       ● Rule 2: 2 of 3 points in Zone A
 +1σ ---------------------------------------------------------- Zone B
       ●     ●  ●    ● Rule 3: 4 of 5 points in Zone B or beyond
  CL ══════════════════════════════════════════════════════════ Center Line
     ● ● ● ● ● ● ● ● Rule 4: 8 consecutive points on one side of CL
 -1σ ---------------------------------------------------------- Zone C

1. Rule 1 (Extreme Outlier): Any 1 point falls beyond Zone A (outside the $3\sigma$ limits: above $UCL$ or below $LCL$).
   • Physical Meaning: Large sudden shock (tool fracture, wrong material lot loaded).
2. Rule 2 (Zone A Warning): 2 out of 3 consecutive points fall in Zone A or beyond on the same side of the center line.
   • Physical Meaning: Moderate process shift of approximately $2\sigma$ in the mean.
3. Rule 3 (Zone B Warning): 4 out of 5 consecutive points fall in Zone B or beyond on the same side of the center line.
   • Physical Meaning: Small process shift of approximately $1\sigma$ in the mean.
4. Rule 4 (Run Above/Below Center Line): 8 consecutive points fall on one side of the center line (in Zone C or beyond).
   • Physical Meaning: Sustained minor bias or slight offset in machine calibration ($P = (0.5)^8 = 0.0039$).

The Nelson Run Rules

In 1984, Lloyd S. Nelson expanded these detection patterns into eight standardized rules widely tested on engineering examinations:

• Nelson Rule 1: 1 point beyond $3\sigma$ ($> UCL$ or $< LCL$).
• Nelson Rule 2: 9 consecutive points on the same side of the center line (shift in mean).
• Nelson Rule 3: 6 consecutive points steadily increasing or steadily decreasing (continuous tool wear, temperature buildup, chemical depletion).
• Nelson Rule 4: 14 consecutive points alternating systematically up and down (sawtooth oscillation; caused by two alternating fixtures, morning vs. night shift differences, or operator over-adjustment).
• Nelson Rule 5: 2 out of 3 consecutive points $> 2\sigma$ from center line on the same side.
• Nelson Rule 6: 4 out of 5 consecutive points $> 1\sigma$ from center line on the same side.
• Nelson Rule 7 (Stratification): 15 consecutive points in Zone C (within $\pm 1\sigma$ of center line). While this looks "good" to novices, it signals an artificial variance reduction, such as an incorrect standard deviation calculation or mixing two distinct product streams into a single subgroup.
• Nelson Rule 8 (Mixture): 8 consecutive points with none in Zone C (points hug the outer zones). Caused by two distinct machines or streams feeding into the same sample without blending.

---

5. Overcontrolling, Tampering, and Deming's Funnel Experiment

One of the most profound principles in industrial quality control is the danger of tampering (overcontrol). Well-intentioned operators frequently inspect a finished part, observe that it deviates slightly from the nominal target (though well within control limits), and adjust the machine settings to "compensate."

Dr. W. Edwards Deming demonstrated the catastrophic mathematical outcome of this practice using the Funnel Experiment, where marbles are dropped through a funnel onto a tabletop target:

                   Deming's Funnel Experiment Rules

 Rule 1: No Adjustment           Rule 2: Error Feedback
 (Leave Funnel Stationary)       (Move Funnel -e_k from Target)

          Target                           Target
            │                                │
            ▼                                ▼
        ┌───────┐                        ┌───────────────┐
        │ • • • │                        │  •   •   •    │
        │•  +  •│                        │•   • + •   •  │
        │ • • • │                        │  •   •   •    │
        └───────┘                        └───────────────┘
    Stable Variance σ²                  Doubles Variance 2σ²

The Four Funnel Rules and Mathematical Consequences

1. Rule 1: Leave the Funnel Alone (Unadjusted)
   • Method: The funnel is fixed directly over the target. Marbles are dropped without adjusting the funnel.
   • Result: The resulting marble pattern forms a circular, stable normal distribution with baseline variance $\sigma^2$. This is the minimum achievable variance.
2. Rule 2: Compensate from Target Based on Previous Error
   • Method: If marble $k$ lands at vector position $z_k$ relative to the target, the operator shifts the funnel for drop $k+1$ by $-z_k$.
   • Result: The variance of the process doubles: $\sigma^2_{\text{Rule 2}} = 2\sigma^2$. The area covered by drops expands by 41% (radius $\sqrt{2}\sigma$). Adjusting the machine for common-cause error introduces the variance of the adjustment on top of the natural variance!
3. Rule 3: Adjust the Funnel Relative to the Previous Drop Position
   • Method: The operator moves the funnel from its current setting by an amount equal and opposite to the error of the last drop (setting $z_{k+1} = -e_k$).
   • Result: Exploding, unbounded variance. The marble locations oscillate with increasing amplitude until marbles roll off the table entirely.
4. Rule 4: Aim the Funnel at the Spot of the Last Marble
   • Method: The operator centers the funnel directly over wherever the previous marble landed.
   • Result: Random Walk Drift. The process wanders steadily away from the target without bound, eventually drifting across the entire factory.

Core Engineering Takeaway: A control chart tells engineers when to take action (when special causes are present) and when to leave the process alone (when only common causes are present). Compensating for common-cause deviations always increases total process variability.

---

6. Step-by-Step Worked Engineering Calculations

Worked Example 20.1.1: Multi-Rule Out-of-Control Analysis and ARL Determination

Problem Statement: A high-speed milling center produces bearing journal housings with a critical internal bore diameter. Historical baseline data establishes that when the process is in statistical control, subgroup averages of size $n = 5$ follow a normal distribution with grand mean $\mu = 50.000$ mm and standard deviation of subgroup means $\sigma_{\bar{X}} = 0.006$ mm. Engineering specifications for the bore are set at $50.000 \pm 0.030$ mm.

1. Compute the Center Line and 3-sigma control limits for the $\bar{X}$ chart.
2. Determine the boundaries for Zone A, Zone B, and Zone C on both the upper and lower sides of the center line.
3. Calculate the in-control Average Run Length ($ARL_0$) for the 3-sigma limits.
4. Evaluate the following sequence of eight consecutive subgroup averages (in mm) against the Western Electric Rules: ${50.007, 50.009, 50.011, 50.013, 50.014, 50.015, 50.017, 50.019}$. Determine whether an out-of-control condition exists, cite the specific rule violated, and state the engineering action required.

Solution:

Step 1: Calculate Center Line and Control Limits

• Center Line: $CL = \mu = 50.000$ mm
• Upper Control Limit: $UCL = CL + 3\sigma_{\bar{X}} = 50.000 + 3(0.006) = 50.018$ mm
• Lower Control Limit: $LCL = CL - 3\sigma_{\bar{X}} = 50.000 - 3(0.006) = 49.982$ mm

Step 2: Determine Zone Boundaries

• Center Line: $50.000$ mm
• Upper Zone Boundaries:
  • $+1\sigma$ boundary: $50.000 + 0.006 = 50.006$ mm
  • $+2\sigma$ boundary: $50.000 + 2(0.006) = 50.012$ mm
  • $+3\sigma$ boundary ($UCL$): $50.000 + 3(0.006) = 50.018$ mm
• Upper Zone Partitioning:
  • Zone C (Upper): $[50.000, 50.006)$ mm
  • Zone B (Upper): $[50.006, 50.012)$ mm
  • Zone A (Upper): $[50.012, 50.018]$ mm

Step 3: Calculate In-Control Average Run Length ($ARL_0$)

• False alarm probability under normality: $\alpha = 2[1 - \Phi(3)] = 2(0.00135) = 0.0027$
• In-control Average Run Length: $ARL_0 = \frac{1}{\alpha} = \frac{1}{0.0027} = 370.37 \approx 370.4 \text{ samples}$

Step 4: Evaluate the Sample Sequence Map each point to its zone:

• Subgroup 1: $50.007$ mm $\to$ Zone B ($50.006$ to $50.012$)
• Subgroup 2: $50.009$ mm $\to$ Zone B
• Subgroup 3: $50.011$ mm $\to$ Zone B
• Subgroup 4: $50.013$ mm $\to$ Zone A ($50.012$ to $50.018$)
• Subgroup 5: $50.014$ mm $\to$ Zone A
• Subgroup 6: $50.015$ mm $\to$ Zone A
• Subgroup 7: $50.017$ mm $\to$ Zone A
• Subgroup 8: $50.019$ mm $\to$ Above UCL ($> 50.018$ mm)

Rule Violations Identified:

1. Western Electric Rule 4 (8 on one side): All 8 points fall strictly on the upper side of the center line ($CL = 50.000$). At Subgroup 8, this rule is fully violated ($P = 0.5^8 = 0.0039$).
2. Western Electric Rule 3 (4 of 5 in Zone B or beyond): Looking at Subgroups 1 through 5, all 5 points fall in Zone B or beyond. The rule triggered as early as Subgroup 4 or 5!
3. Western Electric Rule 2 (2 of 3 in Zone A or beyond): Subgroups 4 and 5 are both in Zone A. This rule was violated at Subgroup 5.
4. Western Electric Rule 1 (1 point outside 3-sigma): Subgroup 8 ($50.019$ mm) exceeds $UCL = 50.018$ mm.
5. Nelson Rule 3 (Trend): All 8 points are strictly increasing ($50.007 < 50.009 < 50.011 < 50.013 < 50.014 < 50.015 < 50.017 < 50.019$). A trend of 6 consecutive increasing points was confirmed at Subgroup 6.

Engineering Action Required: The process is out of control. Despite individual parts still residing well within customer engineering specifications ($50.000 \pm 0.030 = [49.970, 50.030]$ mm), an assignable cause is actively shifting the machine (likely thermal expansion of the spindle or progressive tool insert wear). The engineer must stop production, inspect the tooling and cooling system, and eliminate the assignable cause before parts exceed $USL$.

---

7. NCEES Reference Handbook Tips & Realistic Exam Traps

• Never Plot Spec Limits on X-bar Charts: If an exam question asks where to draw $USL$ and $LSL$ on an $\bar{X}$ chart, the correct answer is: they are not drawn. Specification limits apply to individual units, not subgroup means.
• Two-Sided False Alarm Rate: Under 3-sigma limits, the total Type I error rate is $\alpha = 0.0027$ ($0.27%$, or $2.7$ per thousand). If an exam question asks for the probability of a false alarm on one side only, divide by 2: $\alpha / 2 = 0.00135$ ($0.135%$).
• Tampering Always Increases Variance: Whenever an exam problem describes an operator tweaking machine calibration after checking a sample that was inside the control limits, recognize this immediately as tampering (Deming's Funnel Rule 2). The correct answer will state that process variability will increase (variance doubles to $2\sigma^2$).
• Zones are 1-Sigma Increments: Zone C is $CL \pm 1\sigma$; Zone B is between $\pm 1\sigma$ and $\pm 2\sigma$; Zone A is between $\pm 2\sigma$ and $\pm 3\sigma$. Keep the zone letter order straight: C is closest to Center, A is furthest Away.