21.2 Acceptance Sampling Concepts and Operating Characteristic Curves

Acceptance sampling is one of the oldest statistical methods in quality engineering, originating in military procurement standards during World War II. Unlike Statistical Process Control (SPC), which monitors ongoing production processes to prevent defects, acceptance sampling is a post-production sentencing procedure. It evaluates a representative sample drawn from a lot of incoming raw materials, subassemblies, or finished goods to decide whether to accept or reject the entire lot. On the FE Industrial and Systems exam, examinees must understand the operational purpose of sampling plans, calculate acceptance probabilities using Binomial and Poisson distributions, and analyze the mathematical properties of Operating Characteristic (OC) curves.

---

1. The Role of Acceptance Sampling vs. Statistical Process Control

A fundamental conceptual distinction tested on the FE exam is that acceptance sampling is not quality control:

                    Quality Strategy Comparison

     Attribute        Statistical Process Control (SPC)     Acceptance Sampling
  ─────────────────────────────────────────────────────────────────────────────
  Primary Purpose     Process monitoring and prevention    Lot sentencing (Accept / Reject)
  Timing              During active production (in-line)   Post-production / Receiving dock
  Quality Impact      Reduces variability, prevents defects Does NOT change product quality
  Feedback Loop       Direct feedback to machine/operator   Administrative lot disposition
  Primary Tool        Shewhart Control Charts (X-bar, R)    OC Curves, Sampling Tables

Why Acceptance Sampling Does Not Control Quality

Acceptance sampling simply screens lots into conforming or non-conforming categories. It does not alter the physical characteristics of the parts inside an accepted lot. An accepted lot containing $2%$ defective units still enters production with $2%$ defective units. The only way acceptance sampling improves average outgoing quality over time is through rectifying inspection (where rejected lots are $100%$ screened and defective units are replaced with conforming parts).

When Acceptance Sampling Is Justified

Acceptance sampling is economically and technically appropriate under specific conditions:

1. Destructive Testing: When testing physically destroys the component (e.g., tensile strength tests, ammunition firing tests, vehicle crash safety tests, electrical fuse burst ratings). In destructive environments, $100%$ inspection would destroy $100%$ of the product, making sampling the only viable option.
2. High Cost of Inspection Relative to Defect Cost: When $100%$ inspection is prohibitively expensive, labor-intensive, or complex compared to the financial consequence of passing an occasional defective part.
3. Inspector Fatigue and Error: When lot sizes are massive, human inspectors experience physical fatigue, boredom, and sensory adaptation. Studies show that manual $100%$ inspection rarely achieves $100%$ accuracy, typically missing $10%$ to $20%$ of defects. A well-designed sampling plan with rigorous, unhurried inspection of a smaller sample often yields superior quality auditing.
4. Supplier Accountability: Rejecting an entire shipment of 10,000 units due to a failed sample imposes strong economic pressure on a supplier to improve their manufacturing processes, far more than simply returning individual defective parts.

---

2. Lot-by-Lot Single Sampling Plans

The most common attribute sampling plan is the single sampling plan, defined completely by three parameters:

• $N$ = Lot size (total number of units in the submitted shipment)
• $n$ = Sample size (number of units randomly selected from the lot for inspection)
• $c$ = Acceptance number (maximum allowable number of defective units in the sample)
• $r$ = Rejection number (typically $r = c + 1$ for single sampling)

                  Single Sampling Plan Decision Flow

                       Submitted Lot of Size N
                                  │
                                  ▼
                      Draw Random Sample of Size n
                                  │
                                  ▼
                     Inspect All n Sample Units;
                     Count Number of Defectives (d)
                                  │
                 ┌────────────────┴────────────────┐
                 ▼                                 ▼
            If d <= c                          If d > c
        ┌──────────────────┐             ┌──────────────────┐
        │    ACCEPT LOT    │             │    REJECT LOT    │
        └──────────────────┘             └──────────────────┘
        Pass to Production / Stock       Return to Vendor or 100% Screen

Execution Protocol

1. Ensure the lot is homogeneous (produced by the same machines, operators, and raw material batches under uniform conditions).
2. Draw a representative, random sample of $n$ units. Every unit in the lot must have an equal probability of selection.
3. Inspect each sample unit against established quality criteria, classifying each as conforming or defective (attribute inspection).
4. Let $d$ be the observed count of defective units:
   • If $d \le c$: Accept the entire lot of $N$ units.
   • If $d > c$: Reject the entire lot of $N$ units for disposition (quarantine, return, or screening).

---

3. Operating Characteristic (OC) Curve Foundations

The Operating Characteristic (OC) curve is the definitive performance profile of any sampling plan. It displays the probability of accepting a lot ($P_a$) on the vertical axis as a function of the incoming lot fraction defective ($p$) on the horizontal axis.

                   Typical Operating Characteristic (OC) Curve

       Pa
      1.0 ┼───────┐ Shoulder
          │        \ 
      0.95│.........\....... (AQL, 1 - alpha)  <- Producer's Risk Point
          │          \
      0.8 │           \
          │            \
      0.5 │             \  Indifference Quality
          │              \
          │               \
      0.10│................\... (LTPD, beta)   <- Consumer's Risk Point
          │                 \___
      0.0 ┴───────┬──────────┬──────────┬────────►
          0      AQL        LTPD       1.0        Incoming Fraction Defective (p)

Type A vs. Type B OC Curves

The mathematical distribution used to calculate $P_a$ depends on whether the sampling plan evaluates isolated lots or a continuous process stream:

The Hypergeometric Model (Type A)

For a finite lot of size $N$ containing exactly $D = N \cdot p$ defective units:

$P_a = P(d \le c) = \sum_{d=0}^c \frac{\binom{D}{d} \binom{N - D}{n - d}}{\binom{N}{n}}$

The Binomial Model (Type B)

When lots are large relative to the sample ($N \ge 10n$), the removal of sample units does not appreciably alter the defect fraction of the remaining lot. Sampling is treated as independent Bernoulli trials:

$P_a = P(d \le c) = \sum_{d=0}^c \binom{n}{d} p^d (1 - p)^{n - d}$

---

4. Mathematical Calculation of $P_a$ via Poisson Approximation

When sample size $n$ is moderately large ($n \ge 20$) and the incoming fraction defective $p$ is small ($p \le 0.05$ or $np \le 5$), the Poisson distribution provides an outstanding, computationally efficient approximation to the Binomial distribution. On the FE exam, the Poisson approximation is almost universally preferred because it eliminates complex combinatorial evaluations.

The Poisson Approximation Formula

Let parameter $\lambda = n \cdot p$ represent the expected number of defectives in the sample:

$P_a = P(d \le c) \approx \sum_{d=0}^c \frac{e^{-np} (np)^d}{d!} = \sum_{d=0}^c \frac{e^{-\lambda} \lambda^d}{d!}$

Explicit Formulations for Common Acceptance Numbers

• For $c = 0$ Plans: $P_a = P(0) = e^{-np}$
• For $c = 1$ Plans: $P_a = P(0) + P(1) = e^{-np} + np \cdot e^{-np} = e^{-np} (1 + np)$
• For $c = 2$ Plans: $P_a = P(0) + P(1) + P(2) = e^{-np} \left( 1 + np + \frac{(np)^2}{2} \right)$
• For $c = 3$ Plans: $P_a = e^{-np} \left( 1 + np + \frac{(np)^2}{2} + \frac{(np)^3}{6} \right)$

---

5. Effect of Sample Size ($n$) and Acceptance Number ($c$) on OC Curves

The shape and position of an OC curve govern the discriminating power of the sampling plan (its ability to distinguish between high-quality lots and low-quality lots).

      Effect of Sample Size (n)               Effect of Acceptance Number (c)
       (Fixed Acceptance Number c)               (Fixed Sample Size n)

  Pa                                       Pa
 1.0 ┼───┐                                1.0 ┼───────┬───────┐
     │   │\  n=20                             │       │\      │\   c=3
     │   │ \                                  │       │ \     │ \ 
     │   │  \                                 │       │  \    │  \ c=1
 0.5 │   │   \  n=50                      0.5 │       │   \   │   \
     │   │    \                               │ c=0   │    \  │    \
     │   │     \   n=100                      │  \    │     \ │     \
 0.0 ┴───┴──────┴────────►                0.0 ┴───┴───┴──────┴┴──────┴──►
     0          Fraction Defective (p)        0          Fraction Defective (p)

1. Effect of Increasing Sample Size ($n$)

• Holding the acceptance number $c$ constant (or holding the ratio $c/n$ constant), increasing $n$ makes the OC curve steeper.
• A steeper OC curve narrows the transition zone between high acceptance and high rejection, dramatically increasing discriminating power.
• Larger sample sizes reduce both Producer's Risk ($\alpha$) and Consumer's Risk ($\beta$), but incur higher inspection labor and gauging costs.

2. Effect of Acceptance Number ($c$)

• Holding sample size $n$ constant, increasing $c$ shifts the entire OC curve to the right.
• A higher acceptance threshold increases the acceptance probability $P_a$ across all values of incoming fraction defective $p$. This protects the producer from unwarranted lot rejections, but exposes the consumer to higher risk of accepting sub-standard lots.

3. The Severe Anomaly of Zero-Acceptance ($c = 0$) Plans

In many modern manufacturing organizations, quality managers adopt a "zero defects" slogan and intuitively mandate $c = 0$ sampling plans (i.e., accept only if 0 defects are found; reject if $d \ge 1$).

While well-intentioned, $c = 0$ plans possess terrible mathematical properties:

1. Lack of an Inflection Shoulder: For $c = 0$, the acceptance probability is strictly exponential: $P_a = (1 - p)^n \approx e^{-np}$ The derivative at $p = 0$ is $\left.\frac{dP_a}{dp}\right|_{p=0} = -n$. The curve drops precipitously the instant $p$ rises above zero, forming a strictly concave upward curve with no horizontal shoulder.
2. Extreme Producer's Penalty: Because there is no shoulder, even exceptionally high-quality lots face severe rejection rates. For example, if a supplier produces a world-class lot with only $1%$ defective ($p = 0.01$) and the customer uses a plan with $n = 50, c = 0$: $\lambda = 50 \times 0.01 = 0.50$ $P_a = e^{-0.50} = 0.6065$ The customer will reject nearly $40%$ of these excellent lots (Producer's Risk $\alpha = 39.35%$). This triggers supplier disputes, needless shipping costs, and plant shutdowns.
3. Superiority of $c \ge 1$ Plans: Plans with $c = 1, 2,$ or $3$ (with proportionally larger $n$ to maintain consumer protection) exhibit a classic sigmoidal (S-shaped) curve. They possess a broad, flat horizontal shoulder at low defect levels (ensuring $P_a \ge 0.95$ for high-quality lots), followed by a steep plunge to protect the consumer.

---

6. The Ideal Operating Characteristic Curve

To understand what real sampling plans strive to achieve, consider the theoretical concept of the Ideal OC Curve:

                      The Ideal Operating Characteristic Curve

       Pa
      1.0 ┼───────────────────────┐
          │                       │
          │     Always Accept     │
          │      (Pa = 1.0)       │
          │                       │
          │                       │  Vertical Cliff at p = p*
          │                       │  (Perfect Discrimination)
          │                       │
          │                       │
          │                       │      Always Reject
          │                       │       (Pa = 0.0)
      0.0 ┴───────────────────────┼───────────────────────────►
          0                      p*                          1.0
                                 Quality Threshold

Mathematical Formulation of the Ideal Plan

Let $p^*$ represent the dividing threshold between acceptable and unacceptable incoming quality:

• If $p \le p^*$, then $P_a = 1.00$ (every conforming lot is accepted with certainty; Producer's Risk $\alpha = 0$).
• If $p > p^*$, then $P_a = 0.00$ (every non-conforming lot is rejected with certainty; Consumer's Risk $\beta = 0$).

Why the Ideal OC Curve Is Physically Unattainable in Sampling

1. Sample Size Requirements: For an OC curve to achieve an infinitely steep vertical slope at $p^*$, the sample size $n$ must approach infinity ($n \to \infty$). In finite lot sampling, perfect discrimination requires inspecting the entire lot ($n = N$).
2. Human and Mechanical Inspection Limits: Even under $100%$ inspection ($n = N$), real inspection systems never achieve the ideal step function due to gauge repeatability and reproducibility (GR&R) errors, sensor false-calls, visual fatigue, and transcription errors.
3. Practical Engineering Trade-Off: Because the ideal curve is impossible, all industrial acceptance sampling represents a statistical compromise between sample size cost ($n$) and the risks of misclassifying lots ($\alpha$ and $\beta$).

---

7. Step-by-Step Worked Engineering Calculations

Worked Example 21.2.1: Single Sampling Plan Evaluation and $c = 0$ vs. $c = 1$ Comparison

Problem: A medical device assembly plant receives electronic power converters in large lots of size $N = 2,500$. The receiving quality engineer must evaluate two candidate single sampling plans:

• Plan A ($c = 0$ Plan): $n = 40, c = 0$
• Plan B ($c = 1$ Plan): $n = 90, c = 1$

The supplier's manufacturing process historically produces lots with an average incoming fraction defective of $p = 0.015$ ($1.5%$ defective). A batch with $p = 0.060$ ($6.0%$ defective) is considered unacceptable.

Required:

1. Using the Poisson approximation, compute the probability of accepting a good lot ($p = 0.015$) under both Plan A and Plan B. Determine the Producer's Risk ($\alpha$) for each plan.
2. Using the Poisson approximation, compute the probability of accepting a poor lot ($p = 0.060$) under both Plan A and Plan B. Determine the Consumer's Risk ($\beta$) for each plan.
3. Compare the discriminating power of the two plans and recommend which plan to implement.

Solution:

Step 1: Evaluate Conformance at Good Quality ($p = 0.015$)

• Plan A ($n = 40, c = 0$): $\lambda_A = n \cdot p = 40 \times 0.015 = 0.60$ $P_a = e^{-\lambda_A} = e^{-0.60} = 0.5488$ $\alpha = 1 - P_a = 1 - 0.5488 = 0.4512 \approx 45.1\%$ Under Plan A, more than $45%$ of good lots are rejected!

• Plan B ($n = 90, c = 1$): $\lambda_B = n \cdot p = 90 \times 0.015 = 1.35$ $P_a = e^{-\lambda_B} (1 + \lambda_B) = e^{-1.35} (1 + 1.35) = 0.25924 \times 2.35 = 0.6092$ $\alpha = 1 - P_a = 1 - 0.6092 = 0.3908 \approx 39.1\%$

Now, let's look at what happens if we adjust Plan B to a standard industrial plan of $n = 50, c = 1$: $\lambda_{B2} = 50 \times 0.015 = 0.75$ $P_a = e^{-0.75}(1 + 0.75) = 0.47237 \times 1.75 = 0.8266 \implies \alpha = 17.3\%$

Step 2: Evaluate Conformance at Poor Quality ($p = 0.060$)

• Plan A ($n = 40, c = 0$): $\lambda_A = 40 \times 0.060 = 2.40$ $P_a = e^{-2.40} = 0.0907$ $\beta = P_a = 0.0907 \approx 9.1\%$
• Plan B ($n = 90, c = 1$): $\lambda_B = 90 \times 0.060 = 5.40$ $P_a = e^{-5.40} (1 + 5.40) = 0.004517 \times 6.40 = 0.0289$ $\beta = P_a = 0.0289 \approx 2.9\%$

Step 3: Engineering Comparison and Synthesis

                 Plan Performance Comparison Summary
┌─────────┬──────────────┬──────────────┬────────────────┬────────────────┐
│ Plan    │ Parameters   │ Pa (p=0.015) │ Pa (p=0.060)   │ Discrimination │
├─────────┼──────────────┼──────────────┼────────────────┼────────────────┤
│ Plan A  │ n = 40, c = 0│   0.5488     │    0.0907      │ Sluggish       │
│ Plan B  │ n = 90, c = 1│   0.6092     │    0.0289      │ Sharp          │
└─────────┴──────────────┴──────────────┴────────────────┴────────────────┘

Notice that Plan B ($n = 90, c = 1$) reduces Consumer's Risk from $9.1%$ down to $2.9%$ (providing $3\times$ better consumer protection against dangerous $6%$ defective shipments) while simultaneously increasing acceptance of good lots from $54.9%$ to $60.9%$. Plan B exhibits much steeper slope and superior discriminating power.

---

8. NCEES Reference Handbook Tips & Realistic Exam Traps

• Trap: Conflating Acceptance Sampling with Defect Prevention: If an exam question asks: "Which quality tool should an engineer implement to reduce process variability and prevent defects during assembly?", the answer is Statistical Process Control / Control Charts, NOT acceptance sampling. Acceptance sampling only sorts lots after the fact.
• Trap: Poisson Parameter Miscalculation: The Poisson parameter is $\lambda = np$. Ensure that $p$ is expressed as a decimal fraction, not a percentage. If incoming defect rate is $2%$, $p = 0.02$, so for $n = 50$, $\lambda = 50 \times 0.02 = 1.0$. Using $p = 2$ yields $\lambda = 100$, producing an immediate arithmetic error.
• Trap: Type A vs. Type B Distribution Selection: If the problem explicitly states "a single finite lot of 50 parts is evaluated without replacement", the exact calculation requires the Hypergeometric distribution. If the lot is large or from a continuous process stream, use Binomial or Poisson.