5.1 SPC Fundamentals & Process Variation

5.1 SPC Fundamentals & Process Variation

In modern precision manufacturing, quality cannot be inspected into a component after it has already been machined, stamped, molded, or assembled. Post-production sorting is inherently late, labor-intensive, and unreliable—routine manual inspection rarely catches more than 80% to 85% of nonconformances. Statistical Process Control (SPC) represents a fundamental operational paradigm shift: shifting the technician's focus from inspecting finished parts to monitoring, controlling, and stabilizing the manufacturing process itself.

As an ASQ Certified Quality Technician (CQT), you are the front-line guardian of statistical process control. Mastering the theoretical foundations established by Dr. Walter A. Shewhart and Dr. W. Edwards Deming enables you to separate true process disruptions from background noise, prevent costly over-adjustment tampering on the shop floor, establish scientifically sound rational subgrouping sampling plans, and maintain processes in a predictable state of statistical control.

---

1. Historical Foundations: Walter A. Shewhart and the Genesis of SPC

During the early 1920s, telephone exchange equipment manufactured at Western Electric's Hawthorne Works in Chicago suffered from inconsistent reliability. Despite exhaustive post-production inspection, switching components failed frequently in the field. In 1924, a young physicist and statistician at Bell Telephone Laboratories named Dr. Walter A. Shewhart recognized the core flaw: inspectors were evaluating parts against static blueprint limits without understanding the dynamic behavior of the machines that produced them.

On May 16, 1924, Shewhart authored a historic one-page internal memorandum introducing the world's first control chart. He formalized his methodology in his landmark 1931 treatise, Economic Control of Quality of Manufactured Product. Shewhart realized that manufacturing processes are inherently variable, but that this variation exhibits distinct mathematical patterns that can be categorized and controlled economically.

Shewhart established that the objective of statistical control is not perfection, but economic balance:

1. Minimizing the Alpha Risk ($\alpha$): The economic loss incurred by looking for trouble (adjusting machines, stopping assembly lines, replacing tooling) when the process is actually operating normally.
2. Minimizing the Beta Risk ($\beta$): The economic loss incurred by failing to detect a true process shift or assignable disruption, allowing defective parts to be produced and shipped.

By placing control limits at exactly $\pm 3$ standard errors ($3\sigma$) from the process average, Shewhart discovered the empirical 'sweet spot' that minimizes the total economic penalty of both error types across virtually all industrial processes.

---

2. The Nature of Variation in Manufacturing

A fundamental physical truth of manufacturing is that no two parts are ever identical. If a digital micrometer measures two turned shafts as exactly $25.000\text{ mm}$, measuring them with an optical interferometer or laser coordinate measuring machine (CMM) to six decimal places will immediately reveal differences at $25.000142\text{ mm}$ versus $24.999887\text{ mm}$.

Variation arises from the complex physical interaction of the classic 6Ms:

• Machine: Spindle runout, thermal growth of castings, bearing friction, backlash in ball screws.
• Material: Slight alloy composition differences, tensile strength variations, surface scale, supplier batch-to-batch hardness shifts.
• Method: Feed rates, depth of cut, clamping sequences, coolant flow rates.
• Measurement: Gage repeatability and reproducibility (GR&R), operator parallax, micrometer ratchet pressure, temperature calibration drift.
• Man (Operator): Fatigue, subtle differences in part loading technique, hand-tightening torque.
• Mother Nature (Environment): Shop-floor ambient temperature swings, humidity changes, building vibrations from adjacent stamping presses.

Because variation cannot be completely eliminated, quality technicians must distinguish between variation that is an inherent, unavoidable part of the system and variation that stems from an identifiable, external disturbance.

---

3. Common Cause vs. Special Cause Variation

Shewhart and Deming divided all process variation into two fundamentally distinct categories: Common Cause Variation and Special Cause Variation. Confusing these two types of variation is the single most expensive mistake made on the shop floor.

A. Common Cause Variation (Chance, Inherent, Chronic, Random Noise)

Common cause variation is the natural, random variation inherent in a process that is operating under standard conditions. It is the cumulative effect of hundreds of tiny, unavoidable variables: minor electrical line fluctuations, normal lubricant viscosity shifts, microscopic grain structure variations in raw steel, and standard mechanical clearance in machine slides.

Key characteristics of common cause variation:

• Ubiquitous and Inherent: Present in every operational step, every shift, every piece.
• Mathematically Stable: When a process experiences only common causes, the individual measurements fluctuate randomly within predictable, stationary statistical boundaries. The distribution of data remains constant in location (mean), spread (standard deviation), and shape.
• Accountability: Deming demonstrated through his famous Red Bead Experiment that 85% to 94% of all quality problems and process variation stem from common causes. Because common causes are built into the design of the equipment, tooling, plant environment, and purchasing policies, front-line operators cannot fix them. Only management and engineering possess the authority to change the system (e.g., purchasing precision ground ball screws, installing temperature-controlled clean rooms, or upgrading raw material specs).
• Proper Technician Response: Do not adjust the process! Leave the process running. Document baseline performance. If variation is too wide to meet customer specifications, form an engineering project team to systematically redesign the manufacturing system.

B. Special Cause Variation (Assignable, Sporadic, External, Signal)

Special cause variation (also called assignable cause variation) arises from specific, identifiable, external events that are not part of the standard process design. It represents an abnormal intrusion into an otherwise stable system.

Key characteristics of special cause variation:

• Intermittent and Sporadic: Appears suddenly or drifts over time, disrupting the baseline distribution.
• Mathematically Unstable: Shifts the process mean, expands the process spread, or distorts the distributional shape, rendering future process output statistically unpredictable.
• Accountability: Deming identified that only 6% to 15% of process problems stem from special causes. These causes are typically local to the workstation: a chipped carbide cutting insert, an unseated collet, a ruptured hydraulic seal, a batch of contaminated solvent, or an untrained substitute operator.
• Proper Technician Response: Immediate local intervention! The quality technician and machine operator must quarantine suspect parts produced since the last good check, identify the specific root cause, eliminate the assignable cause, and restore the process to its original baseline state.

Comparison Reference: Common Cause vs. Special Cause Variation

| Attribute | Common Cause Variation | Special Cause Variation | |:---|:---|:---|| | Historical Terminology | Chance causes, Inherent variation, Chronic noise | Assignable causes, Sporadic variation, Signal | | Physical Origin | Design of the system, machine base capability, standard raw materials | Specific mechanical failure, tool fracture, human error, bad lot of stock | | Percentage of Quality Losses | 85% – 94% (Deming 94/6 Rule) | 6% – 15% (Deming 94/6 Rule) | | Statistical Behavior | Stable, stationary, predictable over time | Unstable, shifting, mathematically unpredictable | | Control Chart Appearance | Points scatter randomly between limits; no runs or trends | Points fall beyond $3\sigma$ control limits, or exhibit non-random patterns | | Responsibility for Correction | Management & Engineering (System overhaul) | Quality Technicians & Operators (Local containment & fix) | | Proper Operational Action | Leave machine settings alone; do not adjust | Stop, quarantine suspect lots, identify root cause, eliminate disturbance | | Danger of Misdiagnosis | Tampering: Adjusting the machine increases total variance by 41% to $\infty$ | Inaction: Scrap proliferates and escapes to the customer unchecked |

---

4. Deming's Funnel Experiment and the Four Rules of Tampering

One of the most dangerous behaviors on the shop floor is tampering—treating common cause variation as if it were a special cause. When an operator measures a part that is within control limits but slightly above nominal and immediately dials back the machine offset, they are tampering.

To prove mathematically that adjusting a stable process increases overall variation, Dr. Deming devised the Funnel Experiment. In this physical demonstration, a funnel is suspended on a stand above a target on a tabletop. A marble is dropped through the funnel, rolls across the table, and comes to rest at a specific point. Deming evaluated four distinct operational rules for adjusting the funnel position.

                    [ Funnel ]
                        |
                        v
                  ( Marble Drop )
                        :
                      . : .
                    .   :   .
                  .     :     .
           [Target]     :     [Resting Point (x_k)]
              X --------+------------ O
                     Error (e_k)

Rule 1: No Adjustment (The Baseline of Statistical Control)

• Action: The funnel is aimed directly over the target and left completely untouched for all successive drops.
• Mathematical Outcome: The marble resting positions scatter in a circular pattern around the target due to inherent, random physical factors (air currents, table grain, slight surface friction). This represents pure common cause variation.
• Process Variance: $\sigma^2$ (The minimum possible variance for the system).

Rule 2: Adjust from Previous Position ($z_{k+1} = z_k - e_k$)

• Action: After drop $k$ comes to rest at position $x_k$, the error vector from the target is calculated: $e_k = x_k - \text{Target}$. The funnel is moved from its current setting $z_k$ by an amount equal and opposite to the error: $z_{k+1} = z_k - e_k$.
• Shop-Floor Analogy: A CNC lathe operator turns a shaft pin. Blueprint nominal is $20.000\text{ mm}$. The pin measures $20.004\text{ mm}$ (an error of $+0.004\text{ mm}$). The operator immediately adjusts the tool wear offset $-0.004\text{ mm}$. The next part measures $19.995\text{ mm}$ ($-0.005\text{ mm}$ error), so the operator offsets $+0.005\text{ mm}$.
• Mathematical Outcome: The variance of the process doubles: $\sigma_{\text{Rule 2}}^2 = 2\sigma^2$ The standard deviation increases by a factor of $\sqrt{2} \approx 1.414$, representing a 41.4% increase in process variation! By attempting to correct for normal common cause noise, the operator introduces an artificial second random variable, dramatically widening the dimensional spread.

Rule 3: Adjust from Target ($z_{k+1} = -e_k$)

• Action: After drop $k$ lands with error $e_k$, the funnel is reset not from its current position, but from the original target by $-e_k$.
• Shop-Floor Analogy: An operator overcompensates relative to blueprint nominal on every piece, resetting machine gages back and forth.
• Mathematical Outcome: The process enters a state of severe hunting and oscillation. The variance explodes to twice the original variance ($2\sigma^2$), accompanied by severe negative autocorrelation ($r = -0.50$). The output swings wildly back and forth across nominal, increasing the probability of producing extreme outliers.

Rule 4: Adjust over Last Resting Point ($z_{k+1} = x_k$)

• Action: The funnel is moved and positioned directly over the spot where the last marble came to rest.
• Shop-Floor Analogy:
  1. Training a new worker by having them shadow a seasoned worker, who was trained by shadowing another worker, without ever referencing standard operating procedures (worker-to-worker training drift).
  2. In sheet metal forming, setting the brake press stop against the edge of the last formed bracket rather than using the original master template or digital backstop.
  3. Adjusting a cutting tool setup to match the diameter of the last part cut rather than measuring against a calibrated master gage block.
• Mathematical Outcome: The process executes a classic random walk. The funnel wanders steadily and inexorably away from the target without bound. The variance increases linearly with each drop toward infinity: $\sigma_{\text{Rule 4}}^2 \longrightarrow \infty$

[!IMPORTANT] The Cardinal Rule of Tampering Never adjust a process that is in statistical control! Any adjustment made to an in-control process experiencing only common cause variation degrades quality by artificially inflating process variance. Adjustments should only be made when a validated out-of-control signal indicates the presence of a special cause.

---

5. The Concept of Statistical Control (Process Stability)

A manufacturing process is defined as being in a state of statistical control (or simply stable) when all assignable special causes of variation have been identified and removed, leaving only inherent common cause variation.

                 STABLE PROCESS                     UNSTABLE PROCESS
           (In Statistical Control)              (Out of Control)

  UCL  --------------------------------    --------------------------------
             *       *       *       *       *                           *
           *   *   *   *   *   *   *   *           *       *       *   *
  CL   --*-------*-------*-------*-------    ----*-------*---*---*-------*-
           *   *   *   *   *   *   *               *   *   *       *
             *       *       *                       *                   *
  LCL  --------------------------------    --------------------------------
         Location, spread, and shape        Location shifts, spread widens,
          remain constant over time.         future performance unpredictable.

Why is statistical control the absolute foundation of quality engineering?

1. Predictability: A stable process behaves like a repeatable statistical random variable. Quality technicians can predict with mathematical confidence the percentage of output that will fall within any given dimensional band tomorrow, next week, or next month.
2. Prerequisite for Process Capability: Calculating capability indices such as $C_p$ and $C_{pk}$ requires computing the sample standard deviation $\sigma$. If a process is out of control, the process mean and dispersion are drifting, making the calculated standard deviation mathematically invalid. Capability has no meaning for an unstable process.
3. Basis for Continuous Improvement: You cannot improve a process that is not in control. Attempting to optimize an unstable process is like trying to tune an engine while the cylinders are misfiring randomly.

---

6. Rational Subgrouping Principles

A control chart is only as reliable as the sampling strategy used to populate it. The method used to organize samples is called rational subgrouping, a concept formulated by Dr. Shewhart.

The Core Objective of Rational Subgrouping

Rational subgroups must be selected such that:

1. Within-Subgroup Variation is Minimized: The variation among individual parts within a single subgroup should represent only short-term, inherent common cause variation ($\sigma_{\text{within}}$).
2. Between-Subgroup Variation Exposes Special Causes: The variation between different subgroups should capture long-term process shifts, drifts, and assignable disruptions ($\sigma_{\text{between}}$).

$\sigma_{\text{total}}^2 = \sigma_{\text{within}}^2 + \sigma_{\text{between}}^2$

Because control limits on an $\bar{X}$ chart are computed directly from the average within-subgroup variation (via $\bar{R}$ or $\bar{s}$), keeping within-subgroup variation clean and small ensures narrow, highly sensitive control limits that immediately flag any between-subgroup shift.

The Golden Rules of Rational Subgrouping

To achieve this separation, quality technicians must adhere to three strict rules:

1. Sample Consecutive Pieces: Take $n$ parts produced in immediate succession (e.g., 5 consecutive parts right off the spindle over a 2-minute window). This guarantees that ambient temperature, tool condition, raw material lot, and operator setup remain virtually identical within the subgroup.
2. Never Mix Sources Within a Subgroup:
   • Do NOT take 1 part from Shift 1, 1 part from Shift 2, and 1 part from Shift 3 to make a subgroup of $n=3$.
   • Do NOT take 1 part from Machine A, 1 from Machine B, 1 from Machine C, and 1 from Machine D to make a subgroup of $n=4$.
   • Do NOT take 1 part from each cavity of a 4-cavity plastic injection mold to make a subgroup of $n=4$.
3. The Danger of Mixing (Stratification): If you mix parts from different machines, shifts, or mold cavities within a single subgroup, the differences between those sources become part of the within-subgroup range ($R$). This artificially inflates $\bar{R}$, blowing the control limits on the $ar{X}$ chart wide open. The chart becomes desensitized and completely blind to true process shifts!

[!TIP] Multi-Spindle and Multi-Cavity Best Practice When monitoring an 8-cavity mold or a 4-spindle screw machine, the technician should either maintain separate control charts for each individual cavity/spindle, or sample consecutive parts from one designated cavity over time. Never average all cavities into a single subgroup.

---

7. Subgroup Sample Size ($n$) and Sampling Frequency

Selecting the subgroup sample size ($n$) and sampling interval represents an engineering trade-off between statistical detection power and inspection economics.

Subgroup Sample Size ($n$) Selection Guidelines

For Shewhart $\bar{X}$ and $R$ charts, subgroup sizes almost universally range from $n = 4$ to $n = 5$:

• The Central Limit Theorem (CLT): The CLT states that regardless of the shape of the underlying population of individual parts (even if skewed or slightly non-normal), the distribution of subgroup averages $\bar{X}$ rapidly approaches a normal distribution as $n$ increases. A sample size of $n=4$ or $5$ provides sufficient normalization for nearly all industrial processes.
• Statistical Efficiency of Range ($R$): The sample range is an extraordinarily simple and efficient estimator of dispersion for small samples (95.5% relative efficiency at $n=4$; 91.0% at $n=5$). When $n \ge 10$, range loses efficiency and standard deviation ($s$) must be used.
• Sensitivity to Shifts: The standard error of the mean is $\sigma_{\bar{X}} = \sigma / \sqrt{n}$. As $n$ increases, $\sigma_{\bar{X}}$ shrinks, tightening control limits and accelerating the detection of small process shifts. A subgroup size of $n=5$ reliably detects a $1.5\sigma$ process shift within 1 to 2 subgroups.

Sampling Frequency Guidelines

How often should a technician pull a subgroup? Sampling frequency is governed by:

1. Process Drift Rate: Rapidly changing processes (e.g., abrasive grinding wheel breakdown or high-wear stamping dies) require frequent sampling (every 30 to 60 minutes). Highly stable CNC milling operations with diamond tooling may require sampling only once every 4 hours or once per shift.
2. Production Rate and Containment Risk: If a machine produces 1,000 parts per hour, pulling a subgroup once every 4 hours puts 4,000 parts at risk of quarantine if an out-of-control condition is detected. The cost of sorting or scrapping those 4,000 parts dictates tighter sampling intervals (e.g., every 30 minutes).
3. Setup and Material Changes: Subgroups should always be drawn immediately following any setup adjustment, tooling replacement, new batch of raw stock, or operator shift changeover.