Acceptance Sampling and Conformity Decisions

Introduction

Acceptance sampling is a statistical procedure used to decide whether to accept or reject a lot of products, materials, components, or records on the basis of a sample. It is particularly valuable when 100% inspection is destructive, expensive, slow, or less reliable than a properly designed sample inspection.

Acceptance sampling is not the same as process control. Process control asks whether a process is stable and predictable; acceptance sampling asks whether a particular lot should be accepted. A sampling plan therefore provides evidence for a decision, not absolute proof that every item in the lot is conforming.

In machine technology, acceptance sampling may be applied to castings, heat-treated components, dimensional features, surface defects, purchased bearings, fasteners, and supplier batches. Decisions must account not only for sample results but also for the risks of accepting poor quality, rejecting acceptable quality, and making measurements with uncertainty.

Learning Objectives

By the end of this chapter, the reader should be able to:

1. Explain the purpose and limitations of acceptance sampling.
2. Distinguish single, double, and sequential sampling plans.
3. Interpret AQL, LTPD, producer's risk, and consumer's risk.
4. Calculate and interpret an operating characteristic curve.
5. Use attribute and variable sampling concepts appropriately.
6. Incorporate measurement uncertainty into conformity decisions.
7. Make risk-based lot-acceptance and supplier-control decisions.

1. Foundations of Lot Acceptance

A lot is a defined quantity of product produced under sufficiently similar conditions. A sample is a subset selected from the lot for inspection. The sampling plan specifies the sample size and the rule for accepting or rejecting the lot.

For attribute inspection, each sampled item is classified as conforming or nonconforming. The number of nonconforming items is counted. For variable inspection, a measured characteristic such as diameter, hardness, or length is used, usually under assumptions concerning the distribution and process variability.

Important distinctions are:

• A conforming item meets the specified requirement.
• A nonconforming item fails at least one requirement.
• A defect is a specific departure from a requirement.
• A defective unit is a unit containing one or more defects.
• A sampling plan is the technical rule for selecting and judging a sample.
• A sampling scheme is a system containing several related plans and switching rules.

Acceptance sampling should normally be used after the supplier, process, specification, inspection method, and lot definition have been clearly established. If the process is demonstrably capable and stable, process control may provide better protection than repeated lot-by-lot acceptance sampling.

2. Single-Sampling Plans

A single-sampling plan is defined by two numbers: sample size n and acceptance number c. Select n items at random, inspect them, and let X be the number of nonconforming items.

The sample must be representative. Random selection, stratification across production time, and prevention of substitution are essential. A plan cannot compensate for biased sampling.

2.1 Probability of acceptance

If sampling is approximately independent and the lot is sufficiently large, the number of nonconforming items can be modelled by a binomial distribution. If the fraction nonconforming in the lot is p, the probability of acceptance is:

Here, P_a(p) is the probability of accepting a lot whose nonconforming fraction is p, n is sample size, c is the acceptance number, and C(n,x) is the binomial coefficient. This model assumes random sampling, a stable lot quality level, and a binary classification with negligible classification error.

For sampling without replacement from a finite lot of size N, the hypergeometric model is more exact:

where D is the number of nonconforming units in the lot. The binomial model is commonly adequate when the sample is small relative to the lot.

Worked example 1: Single plan

A plan uses n=50 and c=2. Estimate the probability of acceptance when the lot is 4% nonconforming. With p=0.04,

The terms are approximately 0.1299, 0.2706, and 0.2763. Therefore:

Thus, a lot containing 4% nonconforming units has about a 67.7% chance of acceptance under this plan. This result illustrates why a single sample does not guarantee that a lot is nearly perfect.

3. Operating Characteristic Curves

The operating characteristic (OC) curve plots the probability of acceptance against the actual or assumed lot fraction nonconforming. It describes the discriminatory power of a sampling plan.

A steep OC curve provides strong separation between good and poor quality. A shallow curve provides weak discrimination and produces substantial uncertainty in decisions. The OC curve should be evaluated at quality levels that matter operationally, not merely at convenient percentages.

The curve is affected by:

• sample size: larger samples generally improve discrimination;
• acceptance number: larger values make acceptance easier;
• lot size and sampling fraction;
• inspection accuracy and classification errors;
• whether the plan is single, double, or sequential.

The OC curve is a long-run probability model. It does not state the actual quality of one particular lot after a decision.

4. AQL, LTPD, and Decision Risks

The acceptable quality limit (AQL) is a quality level that is considered satisfactory as a process average or reference level for a sampling scheme. AQL does not mean that every lot at that quality level must be accepted, nor does it mean that production at the AQL is a target for poor processes.

The limiting lot tolerance percent defective (LTPD), also called the rejectable quality level in some frameworks, is a poor quality level that the consumer wishes to reject with high probability. Terminology varies among standards, so the governing standard and contract must be identified.

Producer's risk, denoted α, is the probability of rejecting a lot at a designated good quality level, often associated with the AQL:

Consumer's risk, denoted β, is the probability of accepting a lot at a designated poor quality level, often associated with the LTPD:

These definitions apply only after the designated quality points and sampling plan have been specified. AQL and LTPD are not universal physical constants; they are contractual or design parameters.

Example 2: Interpreting risks

Suppose a plan is designed so that P_a(AQL)=0.95 and P_a(LTPD)=0.10. Then:

• producer's risk is α=0.05, or 5%;
• consumer's risk is β=0.10, or 10%.

The plan does not guarantee acceptance of every AQL lot or rejection of every LTPD lot. It establishes long-run probabilities under repeated comparable sampling.

5. Double-Sampling Plans

A double-sampling plan can reduce average inspection when the first sample often gives a clear decision. It consists of a first sample size, first acceptance number, first rejection number, and—if necessary—a second sample and combined decision rule.

A typical plan is written as:

• first sample: n₁;
• accept immediately if X₁≤c₁;
• reject immediately if X₁≥r₁;
• otherwise inspect a second sample of size n₂;
• accept if X₁+X₂≤c₂, otherwise reject.

The exact rules must be taken from the selected standard or designed statistically. The first and second samples should be selected independently and randomly unless the governing procedure specifies otherwise.

Double sampling is attractive when inspection is costly but delay and administrative complexity are acceptable. Its average sample number depends on the quality distribution: good and very poor lots often receive early decisions, whereas borderline lots require both samples.

6. Sequential Sampling

In sequential sampling, items are inspected one at a time, or in small groups, and the cumulative evidence is compared with acceptance and rejection boundaries. Sampling stops as soon as the evidence is sufficiently strong; otherwise inspection continues until a maximum sample size or a specified termination rule is reached.

Sequential probability ratio testing (SPRT) compares competing hypotheses, such as:

• H₀: the lot has acceptable quality;
• H₁: the lot has unacceptable quality.

For a binomial model, the likelihood ratio after observing x nonconforming units in m observations is:

where p₀ and p₁ are the quality levels represented by the two hypotheses, with p₁>p₀. In a simplified two-risk design, accept the good-quality hypothesis when Λ≤A, reject it when Λ≥B, and continue sampling between the boundaries. The constants A and B are selected from the desired risks and depend on the exact test formulation.

Sequential plans can reduce average inspection, but they require trained inspectors, controlled recording, clear stopping rules, and software or tables validated for the chosen design. They are less convenient when inspection results are slow or destructive.

7. Attribute and Variable Sampling

Attribute plans count conforming or nonconforming units. They are simple and robust, but they may require larger samples. Variable plans use measured values and can be more efficient when the measurement distribution and variability assumptions are credible.

For a two-sided dimensional specification, the measured characteristic must be compared with lower and upper specification limits, LSL and USL. A common process capability index is:

where μ is the process mean and σ is the within-process standard deviation. This is a process-capability measure, not itself an acceptance-sampling rule. Its use assumes a stable process and an appropriate distributional model, commonly approximate normality for the measured characteristic.

Variable sampling can be inappropriate when the distribution is strongly non-normal, when measurement error is large, or when the cost of estimating variability exceeds the benefit. The chosen method must match the specification and the governing standard.

8. Measurement Uncertainty and Conformity Decisions

A sampling decision is only as reliable as the inspection system. Measurement uncertainty arises from resolution, calibration, repeatability, reproducibility, temperature, fixturing, operator influence, and method variation.

Let a measured result be y with expanded uncertainty U, commonly reported for a stated coverage factor and confidence convention. A simple guard-band rule for an upper specification limit is:

For a lower specification limit:

For two-sided specifications, both conditions must be satisfied. These rules reduce the probability of accepting a nonconforming item, but they increase the probability of rejecting a conforming item. They are therefore risk policies, not universal laws.

If no guard band is used, a result just inside a limit may still be compatible with the item being outside the specification. The decision rule must be stated before inspection, including the treatment of borderline results, rounding, calibration status, and uncertainty estimation. ISO/IEC 17025 principles emphasize that a conformity statement should identify the decision rule and its relationship to measurement uncertainty.

Measurement-system analysis should precede acceptance sampling. Useful studies include bias, linearity, stability, repeatability, reproducibility, and gauge capability. A gauge with excessive variation can cause both false acceptance and false rejection.

9. Designing a Risk-Based Sampling Plan

Plan design begins with the consequences of failure rather than with a table lookup. The following sequence is practical:

1. Define the lot, characteristic, specification, and classification of defects.
2. Identify critical, major, and minor failure consequences.
3. Establish acceptable and unacceptable quality points.
4. Select producer's and consumer's risks.
5. Choose attribute or variable inspection.
6. Determine sample size and decision numbers using a recognized standard or validated calculation.
7. Verify the OC curve at relevant quality levels.
8. Confirm sampling randomness and measurement-system capability.
9. Define escalation, reinspection, and supplier-corrective-action rules.
10. Review performance using inspection and field-failure data.

High-severity characteristics may require zero-acceptance criteria, 100% inspection, automated inspection, process controls, or special approval rather than ordinary sampling. A zero-acceptance plan does not imply zero risk: a sample can contain no observed failures even when the lot contains some nonconforming units.

10. Supplier Control and Quality Governance

Acceptance sampling should be integrated with supplier management. A supplier with stable capability, effective traceability, and strong corrective action may qualify for reduced inspection. A supplier with repeated escapes, weak measurement systems, or process instability may require tightened inspection, source inspection, audits, or suspension.

Supplier decisions should be based on evidence such as:

• lot acceptance and rejection history;
• defect severity and escape consequences;
• process capability and control-chart performance;
• audit findings and corrective-action effectiveness;
• measurement-system results;
• delivery and traceability performance;
• field returns and warranty information.

Changing a sampling plan after observing an undesirable result can create biased performance records. Switching rules should be predetermined. Reinspection of rejected lots must also be controlled; repeated sampling until acceptance can substantially increase the chance of eventually accepting poor material and invalidates the original risk statements.

11. Illustrative Case: Machined Hydraulic Sleeves

Background

A manufacturer purchases lots of 2,000 precision hydraulic sleeves. The critical internal diameter must lie between 25.000 mm and 25.040 mm. An out-of-limit sleeve can cause leakage or seizure. The supplier currently provides a certificate based on its own inspection.

Problem

Three lots have passed the existing sample plan, but a customer later finds oversized sleeves. Investigation shows that the supplier's gauge has poor reproducibility and that samples were taken only from the beginning of each production run.

Analysis

The customer first defines the lot as one production run and requires sampling throughout the run. A gauge repeatability and reproducibility study is performed. The measurement uncertainty near the upper limit is found to be ±0.008 mm under the stated reporting convention. Consequently, a result of 25.038 mm cannot automatically be treated as safely conforming.

Because leakage is a serious failure mode, the customer adopts a documented guard-band rule: an individual measurement must satisfy y+U≤25.040 and y−U≥25.000 for unconditional acceptance. Results in the guard band are escalated for a higher-resolution method or engineering review.

The customer replaces the previous informal plan with a recognized sampling plan whose OC curve provides 5% producer's risk at the selected acceptable quality point and 10% consumer's risk at the selected limiting quality point. The supplier must also demonstrate process stability and capability. For the first three lots, tightened inspection is used; after sustained satisfactory performance, normal inspection may resume under the predefined switching rules.

Outcome

The supplier discovers tool wear late in the production run and introduces tool-life controls, hourly verification, and stratified sampling. Subsequent lots meet the acceptance requirements, and customer escapes decline. The organization also learns that the original problem was not solved merely by increasing sample size: sampling bias and measurement uncertainty were major contributors.

Learning points

1. A certificate is not a substitute for an effective measurement system.
2. Sampling must cover the entire lot.
3. Risk levels must be connected to failure consequences.
4. A conformity decision requires a stated decision rule.
5. Supplier development and process control often provide greater protection than inspection alone.

12. Limitations and Contemporary Practice

Acceptance sampling cannot detect every defect, prove process capability, or replace prevention. It may be inefficient for very small lots, highly heterogeneous lots, or products with catastrophic failure consequences. Poorly designed plans can encourage suppliers to aim at the acceptance boundary rather than improve the process.

Modern manufacturing increasingly combines sampling with automated vision, in-line sensors, digital traceability, and statistical process monitoring. Machine-learning systems may identify patterns in inspection data, but they must be validated for drift, false alarms, data bias, and explainability. AI should support—not silently replace—the approved conformity rule. Any automated classifier must have known performance, controlled software changes, and clear responsibility for final decisions.

Regulatory and ethical responsibilities include truthful certificates, prevention of deliberate sample selection, protection of customer and public safety, data integrity, and transparent communication of uncertainty. Contractual requirements and applicable standards always take precedence over informal statistical practice.

Best-Practice Checklist

• Define the lot and sampling frame precisely.
• Use random or scientifically justified stratified sampling.
• Select plans from recognized standards or validated statistical designs.
• Record AQL, LTPD, producer's risk, consumer's risk, and the decision rule.
• Verify the measurement system before relying on results.
• Include uncertainty and guard-band policy where appropriate.
• Use process control and supplier development alongside acceptance sampling.
• Predetermine switching, escalation, and reinspection rules.
• Review OC performance and actual field outcomes periodically.
• Preserve traceable, tamper-resistant inspection records.

Chapter Summary

Acceptance sampling makes a lot-level decision from sample evidence. A single plan uses sample size and acceptance number; double plans permit an early decision or a second sample; sequential plans continue inspection until evidence crosses an acceptance or rejection boundary. The probability of acceptance is described by the OC curve and can be calculated using binomial or, for finite lots, hypergeometric models.

AQL represents a designated satisfactory quality level, while LTPD represents a designated poor quality level. Producer's risk is the probability of rejecting quality regarded as good; consumer's risk is the probability of accepting quality regarded as poor. These are properties of a specified plan and decision framework, not guarantees for individual lots.

Sound decisions require representative sampling, capable measurement systems, explicit uncertainty treatment, and risk-based supplier controls. Sampling should complement process capability, traceability, corrective action, and prevention. For critical characteristics, inspection alone may be insufficient, and automated or 100% controls may be justified.

Key Terms

Acceptance number: Maximum number of nonconforming units permitted for acceptance.

Acceptance sampling: Statistical inspection used to accept or reject a lot.

AQL: Designated satisfactory quality level used in sampling-plan design.

Consumer's risk: Probability of accepting a lot at a designated poor quality level.

Conformity decision: Formal decision that a result meets or does not meet specified requirements.

Double-sampling plan: Plan in which a second sample may be required after an inconclusive first sample.

Defective unit: Unit containing one or more defects.

Guard band: Conservative region used to account for measurement uncertainty.

LTPD: Limiting lot tolerance percent defective; designated poor quality level.

Lot: Defined quantity produced or supplied under stated conditions.

Nonconforming unit: Unit that fails at least one specified requirement.

OC curve: Curve showing probability of acceptance versus lot quality.

Producer's risk: Probability of rejecting a lot at a designated good quality level.

Sequential sampling: Sampling in stages until evidence supports acceptance or rejection.

Single-sampling plan: Plan based on one sample and one acceptance rule.

Stratified sampling: Sampling from defined portions of a lot to improve representation.

Type I risk: Often used for producer's risk in hypothesis-testing language.

Type II risk: Often used for consumer's risk in hypothesis-testing language.

Measurement uncertainty: Parameter describing dispersion associated with a measured result.

Review Questions

Short Answer Questions

1. What is the difference between acceptance sampling and process control?
2. Define AQL and LTPD.
3. What do the sample size and acceptance number specify in a single-sampling plan?
4. Why is random sampling important?
5. What is the purpose of a guard band?

Long Answer Questions

1. Explain the construction and interpretation of an OC curve.
2. Compare single, double, and sequential sampling plans, including their advantages and limitations.
3. Discuss producer's risk and consumer's risk using a manufacturing example.
4. Explain why measurement-system capability is essential to conformity decisions.
5. Develop a risk-based framework for controlling a supplier of critical machined components.

Analytical / Application Questions

1. A plan has n=50 and c=2. Explain what the decision means when three nonconforming units are found, and discuss why rejection does not prove that every other unit is nonconforming.
2. A measured diameter is 25.038 mm, the upper specification limit is 25.040 mm, and the expanded uncertainty is 0.008 mm. Apply the upper-limit guard-band rule and explain the resulting decision.
3. A supplier has repeated lot acceptance but several field failures. Identify at least four possible weaknesses in the acceptance system and propose a corrective-control strategy.

Multiple Choice Questions with Answers

Question 1

In a single-sampling plan, what does the acceptance number c represent?

A. The number of lots inspected

B. The maximum number of nonconforming units allowed for acceptance

C. The consumer's risk

D. The lot size

Correct Answer: B. The maximum number of nonconforming units allowed for acceptance

Question 2

For a binomial attribute plan, what does p represent in P_a(p)?

A. Sample size

B. Acceptance number

C. Fraction nonconforming in the lot

D. Measurement uncertainty

Correct Answer: C. Fraction nonconforming in the lot

Question 3

If P_a(AQL)=0.95, the producer's risk is:

A. 0.05

B. 0.10

C. 0.95

D. 1.95

Correct Answer: A. 0.05

Question 4

Consumer's risk at the LTPD is the probability of:

A. Rejecting a good lot

B. Accepting a poor-quality lot

C. Measuring the lot twice

D. Finding zero defects in every lot

Correct Answer: B. Accepting a poor-quality lot

Question 5

Which plan can require a second sample after an inconclusive first sample?

A. Single-sampling plan

B. Double-sampling plan

C. Capability index

D. Calibration plan

Correct Answer: B. Double-sampling plan

Question 6

For an upper specification limit, which guard-band rule is conservative?

A. y−U≤USL

B. y+U≤USL

C. y×U≤USL

D. y/U≤USL

Correct Answer: B. y+U≤USL

Question 7

What does an OC curve primarily show?

A. Machine speed versus power

B. Probability of acceptance versus lot quality

C. Uncertainty versus temperature only

D. Supplier cost versus delivery time

Correct Answer: B. Probability of acceptance versus lot quality

Question 8

A lot contains 4% nonconforming units. Under a plan with n=50 and c=2, the approximate probability of acceptance calculated in the chapter is:

A. 0.05

B. 0.32

C. 0.68

D. 0.96

Correct Answer: C. 0.68

Question 9

Which condition most directly threatens the validity of an acceptance-sampling result?

A. Random selection across the lot

B. A capable measurement system

C. Sampling only from the beginning of production

D. A documented decision rule

Correct Answer: C. Sampling only from the beginning of production

Question 10

For a critical characteristic with severe failure consequences, which response may be more appropriate than ordinary sampling alone?

A. Ignore measurement uncertainty

B. Use only supplier certificates

C. Consider 100% inspection, process control, or automated inspection

D. Increase the AQL without review

Correct Answer: C. Consider 100% inspection, process control, or automated inspection