8.1 Sampling Fundamentals & Operating Characteristic (OC) Curves

8.1 Sampling Fundamentals & Operating Characteristic (OC) Curves

Acceptance Sampling vs. 100% Inspection

Acceptance sampling is an operational quality procedure in which a predetermined number of units ($n$) is randomly drawn from a production lot ($N$) and inspected against engineering specifications to determine whether the entire lot should be accepted or rejected. It is critical for quality inspectors to recognize what acceptance sampling is and what it is not:

• Acceptance sampling is an audit and sentencing tool: Its sole purpose is to make a lot disposition decision (accept, reject, screen, or return to vendor).
• Acceptance sampling does not control or improve quality: Inspecting parts does not build quality into them; quality is manufactured into the product through process control (SPC), machine capability, and robust tooling. Sampling merely sorts lots based on statistical risk.

The Limitations of 100% Inspection

A widespread misconception among novice manufacturing personnel is that 100% inspection guarantees zero defects reaching the customer. Decades of human factors engineering and empirical metrology studies have proven that 100% manual visual or dimensional inspection is rarely more than 80% to 85% effective in practice. The remaining 15% to 20% of defects escape undetected due to:

1. Vigilance Decrement & Fatigue: Monotonous, repetitive inspection of thousands of identical components induces rapid cognitive fatigue, eye strain, and sensory habituation within 20 to 30 minutes of continuous sorting.
2. Gage Handling & Environmental Inconsistencies: Over the course of screening an entire production run, variations in gage clamping pressure, operator posture, lighting glare, and ambient part thermal expansion introduce measurement errors.
3. False Sense of Security: When downstream assemblers or customers believe a batch was "100% inspected," they reduce their own verification vigilance, amplifying the impact of defect escapes.
4. Destructive Testing Impossibility: Whenever verification testing damages or destroys the product—such as tensile pull testing of fasteners, Rockwell hardness indentation on precision sealing surfaces, metallurgical weld cross-sectioning, or burst pressure testing of hydraulic hoses—100% inspection is physically impossible, as it would destroy 100% of the deliverable product.
5. Economic Inefficiency: The direct labor, machine downtime, and gage wear associated with 100% screening often exceed the scrap or rework value of nonconforming parts by orders of magnitude.

---

Lot Formation, Homogeneity, and Traceability

Acceptance sampling plans are valid only if the samples drawn accurately reflect the statistical properties of the entire population. This requires strict adherence to standardized lot formation rules:

The Rule of Homogeneity

A lot must be homogeneous—meaning all units within the lot were produced under essentially identical manufacturing conditions:

• Fabricated from the same heat, melt, batch, or lot of certified raw material.
• Machined on the same work center, spindle, or tooling setup.
• Produced by the same machine operators across consecutive, uninterrupted shifts.
• Processed through the same heat treatment quench or chemical surface finishing tank.

Consequence of Non-Homogeneous Lots: If an inspector mixes parts from two different CNC lathes (one holding tight tolerances and one experiencing tool drift) into a single inspection lot, the resulting distribution becomes bimodal or skewed. The sampling plan's operating characteristic curve becomes completely invalid, leading to either severe consumer risk (accepting clusters of bad parts) or unwarranted producer risk (rejecting good parts).

Lot Size ($N$) and Traceability

While sample size ($n$) governs statistical protection far more than lot size ($N$), larger lots generally offer economic advantages because the required sample size does not increase proportionally with lot size under standard sampling tables (e.g., ANSI/ASQ Z1.4). However, inspectors must balance economic sample sizing against lot containment risk: if a lot of 50,000 units is rejected, the financial impact of quarantine, sorting, or scrap is catastrophic compared to rejecting a lot of 1,000 units. Every lot must maintain unambiguous traveler, work order, and barcode traceability from raw stock receiving through final packaging.

---

Representative Random Sampling Methods

A sample is representative only if every individual piece in the lot has an equal and independent probability of being selected. Quality inspectors must strictly reject shop floor "convenience sampling" (such as pulling five easy-to-reach parts from the very top of a deep wire basket) or "judgment sampling" (inspectors picking parts that look particularly smooth or rough).

Standard Random Sampling Techniques

1. Simple Random Sampling: Every part is assigned a virtual index ($1$ to $N$). A random number table or computer algorithm generates $n$ random integers, and those exact corresponding parts are retrieved from the container.
2. Stratified Random Sampling: Used when a lot consists of identifiable physical sub-groups (strata), such as parts packed across 10 separate tote boxes, or parts from a 4-cavity injection mold. The inspector divides the total required sample size $n$ proportionally across all strata (e.g., drawing $n/10$ random pieces from each tote), ensuring that variations between layers or cavities are faithfully captured.
3. Systematic Sampling: The inspector selects every $k$-th unit emerging from a conveyor or production line, where $k = N / n$. For example, if a lot of $N = 2,000$ parts requires a sample of $n = 50$, the inspector pulls every $40^{\text{th}}$ part. Caution: Systematic sampling introduces extreme bias if the sampling frequency matches a cyclical process harmonic (such as a periodic thermal expansion cycle or tool-indexer indexing cadence).

---

The Two Risks in Acceptance Sampling

Because acceptance sampling evaluates a fraction of the lot rather than the whole, sampling error is inevitable. In statistical hypothesis testing terminology:

• Null Hypothesis ($H_0$): The production lot conforms to specified quality standards.
• Alternative Hypothesis ($H_1$): The production lot does not conform to specified quality standards.

This leads to two distinct error states that every quality inspector must master:

                                 True Lot Quality State
                          Lot Conforms           Lot is Nonconforming
                     +-----------------------+-----------------------+
        Accept Lot   |   Correct Decision    |     Type II Error     |
                     |     (1 - alpha)       |    Consumer's Risk    |
Sampling             |    Confidence = 95%   |       (beta = 10%)    |
Decision             +-----------------------+-----------------------+
        Reject Lot   |     Type I Error      |   Correct Decision    |
                     |    Producer's Risk    |      (1 - beta)       |
                     |     (alpha = 5%)      |     Power = 90%       |
                     +-----------------------+-----------------------+

Producer's Risk ($\alpha$, Type I Error)

• Definition: The probability that an acceptable, high-quality lot having a fraction nonconforming at or better than the Acceptable Quality Level (AQL) will be erroneously rejected by the sampling plan.
• Symbol: Denoted by Greek letter $\alpha$ (alpha).
• Standard Value: Conventionally set at $\alpha = 0.05$ (or $5%$). This means the producer has a $95%$ probability of lot acceptance ($1 - \alpha = 0.95$) when submitting product at the agreed AQL.
• Shop Floor Consequence: Good parts are returned to manufacturing for unnecessary 100% sorting, tying up inspection benches, causing production delays, and increasing internal scrap costs.

Consumer's Risk ($\beta$, Type II Error)

• Definition: The probability that an unsatisfactory, poor-quality lot having a fraction nonconforming at or worse than the Lot Tolerance Percent Defective (LTPD / Limiting Quality LQ) will be mistakenly accepted by the sampling plan.
• Symbol: Denoted by Greek letter $\beta$ (beta).
• Standard Value: Conventionally set at $\beta = 0.10$ (or $10%$). This means the customer faces a $10%$ probability of accepting a batch that is at the undesirable LTPD threshold.
• Shop Floor Consequence: Defective product enters downstream assembly lines or reaches end-users, causing assembly jams, warranty claims, safety recalls, and customer dissatisfaction.

Quality Benchmarks: AQL vs. LTPD (LQ)

• Acceptable Quality Level (AQL): The maximum percent nonconforming (or maximum number of nonconformities per 100 units) that, for the purpose of acceptance sampling, can be considered satisfactory as a process average. AQL is a producer-oriented index associated with low rejection risk ($1 - \alpha \approx 95%$).
• Lot Tolerance Percent Defective (LTPD) or Limiting Quality (LQ): The designated unacceptable level of quality that the consumer wishes to restrict to a low probability of acceptance. LTPD is a consumer-oriented index associated with Consumer's Risk ($P_a = \beta = 10%$).

---

Operating Characteristic (OC) Curves

The Operating Characteristic (OC) curve is the mathematical fingerprint of an acceptance sampling plan. It displays the probability of lot acceptance ($P_a$) along the vertical Y-axis as a function of the incoming lot fraction nonconforming ($p$) along the horizontal X-axis.

Anatomy of the OC Curve

• When incoming quality is perfect ($p = 0$), acceptance probability is certain: $P_a = 1.0$ ($100%$).
• As the fraction nonconforming $p$ increases, $P_a$ drops monotonically toward zero.
• At $p = \text{AQL}$, $P_a = 1 - \alpha$ (typically $0.95$).
• At $p = \text{LTPD}$, $P_a = \beta$ (typically $0.10$).
• The zone between AQL and LTPD is the region of indifference, where the plan does not strongly discriminate between good and bad quality.

  Probability of
  Acceptance (Pa)
   1.0 +-------------.
       |              `.
       |                \
  0.95 +                 \  <-- (AQL, 1 - alpha)
       |                  \
       |                   \
       |                    \
       |                     \
       |                      \
  0.10 +                       `--.\  <-- (LTPD, beta)
       |                           `----------------
   0.0 +------+--------------------+----------------+--> Incoming Fraction
       0     AQL                  LTPD              1.0   Nonconforming (p)

Ideal vs. Actual OC Curves

• The Ideal OC Curve: An ideal OC curve would be a vertical step function: $P_a = 1.0$ for all quality levels $p \le \text{AQL}$, dropping instantly to $P_a = 0.0$ for any $p > \text{AQL}$. To achieve this vertical wall of discrimination would require 100% perfect inspection of an infinite population without human error—an operational impossibility.
• Effect of Sample Size ($n$) on Discrimination Power: As sample size $n$ increases (maintaining the same ratio of acceptance number $c$ to $n$), the OC curve becomes noticeably steeper. A steeper curve compresses the region of indifference between AQL and LTPD, providing much sharper discrimination between acceptable and unacceptable lots.

Type A vs. Type B OC Curves

• Type A OC Curve: Calculated for individual, finite isolated lots of size $N$. Because sampling without replacement changes the probabilities of successive draws from a finite population, Type A curves are mathematically modeled using the Hypergeometric distribution: $P(d) = \frac{\binom{D}{d} \binom{N - D}{n - d}}{\binom{N}{n}}$ where $N$ is lot size, $D$ is total nonconforming units in the lot, $n$ is sample size, and $d$ is nonconforming units in the sample.
• Type B OC Curve: Calculated for an infinite sequence of continuous lots or a continuous manufacturing process stream. Sampling is treated as sampling with replacement or sampling from an infinite population. Type B curves are mathematically modeled using the Binomial distribution: $P(d) = \binom{n}{d} p^d (1 - p)^{n - d}$ When sample size is moderately large ($n \ge 16$) and incoming fraction nonconforming is small ($p \le 0.10$), the Binomial is conveniently approximated by the Poisson distribution with parameter $\lambda = n p$: $P(d) = \frac{e^{-np} (np)^d}{d!}$

---

Rectifying Inspection: AOQ, AOQL, and ASN

In many receiving inspection and in-house manufacturing operations, rejected lots are not scrapped; rather, they undergo rectifying inspection (also known as screening). Under rectifying inspection, any lot rejected by the sampling plan is subjected to 100% inspection, during which all nonconforming units are removed and replaced with certified conforming units.

Average Outgoing Quality (AOQ)

Average Outgoing Quality represents the expected average fraction nonconforming in product leaving the inspection department over many lots, taking into account both accepted lots (which contain uninspected nonconforming units) and rejected lots (which have been cleaned to $0%$ nonconforming): $AOQ = \frac{P_a \cdot p \cdot (N - n)}{N}$ When the sample size $n$ is small relative to total lot size $N$ (i.e., $n / N < 0.10$), the formula simplifies to: $AOQ \approx P_a \cdot p$

Average Outgoing Quality Limit (AOQL)

As incoming quality $p$ varies, the AOQ curve exhibits a unique mathematical shape:

• If incoming quality is near-perfect ($p \to 0$), outgoing quality is naturally near-perfect ($AOQ \to 0$).
• If incoming quality is terribly poor ($p \to 1$), virtually all lots are rejected and 100% screened, so outgoing quality is once again near-perfect ($AOQ \to 0$).
• Between these extremes, the AOQ curve rises to a maximum peak and then declines. This peak represents the Average Outgoing Quality Limit (AOQL). The AOQL is the worst possible average outgoing percent nonconforming that can emerge from rectifying inspection in the long run, regardless of how bad the incoming process quality becomes.

Average Sample Number (ASN)

The Average Sample Number is the average number of parts inspected per lot to reach a disposition decision. For a single sampling plan, the sample size is fixed, so $ASN = n$ regardless of incoming quality. For double or multiple sampling plans, lots with exceptionally high or exceptionally low defect rates are accepted or rejected on the first sample ($n_1$). Only borderline lots require drawing the second sample ($n_2$). Consequently, double and multiple sampling plans yield a significantly lower ASN under good or very bad process conditions, saving inspection labor hours.

---

Real Shop Inspection Scenarios & Common Exam Traps

• Exam Trap: Confusing AQL with a Guarantee of Zero Defects: A candidate often mistakenly assumes that an AQL of $1.0%$ guarantees that no accepted lot will ever contain more than $1.0%$ nonconforming parts. This is mathematically false! AQL is a producer-oriented process average benchmark where lots have a high probability of acceptance (around $95%$). Individual accepted lots can and will contain higher defect percentages due to random sampling variability.
• Exam Trap: Assuming Lot Size Dictates Sample Size Proportionality: Exam questions often test whether an inspector should take a fixed percentage sample (e.g., "always inspect $10%$ of the lot"). A $10%$ sampling rule is statistically flawed: for a lot of 100 parts, $n=10$ gives virtually no discrimination power, while for a lot of 50,000 parts, $n=5,000$ represents massive, wasteful over-inspection. Sampling plan sample sizes depend on statistical confidence and OC curve slope, not rigid percentage ratios.
• Exam Trap: Confusing $\alpha$ and $\beta$ Association: Remember the mnemonic pairing:
  • Producer wants high quality acknowledged $\rightarrow$ associates with AQL $\rightarrow$ Producer's Risk ($\alpha$) = Type I Error.
  • Consumer fears poor quality acceptance $\rightarrow$ associates with LTPD/LQ $\rightarrow$ Consumer's Risk ($\beta$) = Type II Error.