Practical Statistical Sampling Inspection — From GB/T 2828.1 to Zero Defect Sampling

In the quality inspection system of manufacturing enterprises, sampling inspection is one of the two major inspection strategies alongside full inspection. Full inspection is suitable for key characteristics, small batches, or low inspection costs, while sampling inspection is based on statistical principles, using samples to infer the overall quality level, thus achieving a balance between inspection costs and quality risks. However, many companies exhibit two extremes when implementing sampling inspection: either mechanically applying the AQL values from GB/T 2828.1, ignoring changes in actual quality levels; or blindly pursuing "zero defects" and reverting to full inspection, thereby losing the efficiency advantage of sampling inspection. This article starts from the statistical foundation of sampling inspection, systematically explains the correct usage of GB/T 2828.1, and explores how to transition to zero defect sampling strategies in practice.

1. Statistical Foundation of Sampling Inspection: OC Curve and Two Types of Risk

To understand sampling inspection, one must first understand the operating characteristic (OC) curve, which is the core tool for measuring the "discrimination ability" of a sampling plan.

The OC curve describes the probability ( P_a(p) ) that a batch will be judged as acceptable (accepted) when the actual nonconforming product rate of the submitted batch is ( p ). Each sampling plan (n, Ac) — where n is the sample size and Ac is the acceptance number — corresponds to a unique OC curve. An ideal OC curve should be step-like: the acceptance probability is 100% when the nonconforming product rate is below a certain threshold, and 0% when it is above the threshold. However, in practice, due to the randomness inherent in sampling, the OC curve is a smooth S-shaped curve, meaning that there are inevitably two types of judgment risks.

The first type of risk is called the producer's risk ( \alpha ), typically set at 5% — that is, when the batch quality is actually acceptable (the nonconforming product rate reaches the acceptable quality level AQL), there is still a ( \alpha ) probability that it will be incorrectly judged as unacceptable and rejected. The second type of risk is called the consumer's risk ( \beta ), typically set at 10% — that is, when the batch quality is actually unacceptable (the nonconforming product rate reaches the lot tolerance percent defective LTPD or rejectable quality level RQL), there is still a ( \beta ) probability that it will be incorrectly judged as acceptable and accepted.

In practice, many quality managers focus only on the AQL, ignoring the fact that the acceptance probability corresponding to the AQL is only 95% (i.e., there is still a 5% rejection risk), and are even less aware of the extent to which batch quality must deteriorate before there is a sufficient probability of interception. This is where the core value of GB/T 2828.1 lies — it dynamically controls these two types of risks through transfer rules and tightened inspection, rather than having the company rigidly adhere to a fixed AQL value.

2. Correct Usage of GB/T 2828.1

GB/T 2828.1 (corresponding to the international standard ISO 2859-1) is the most widely used standard in the field of attribute sampling inspection. Its design philosophy is: based on the continuous quality performance of the submitted batches, dynamically switch between normal inspection, tightened inspection, and reduced inspection to protect the interests of the consumer in the long term while providing incentives to the producer.

The first step is to determine the AQL (acceptable quality level). AQL is not the "allowed nonconforming product rate" but a "tolerable process average" — that is, the quality level normally expected when the process is in a stable and controlled state. The determination of AQL should consider the importance of the product, customer requirements, and process capability. For key characteristics (safety, regulations), the AQL value is typically set at 0.01% to 0.1%; for important characteristics (function, performance), it is set at 0.65% to 1.0%; and for general characteristics (appearance, non-critical dimensions), it is set at 1.0% to 2.5% or higher. It is important to note that the smaller the AQL, the larger the sample size, and the higher the inspection cost. Therefore, the selection of AQL is essentially a balance between quality requirements and economics.

The second step is to determine the inspection level. GB/T 2828.1 specifies seven inspection levels, with general inspection levels I, II, and III being the most commonly used, and level II being the "normal" level. The inspection level determines the sample size code, which in turn determines the size of n. Level I has a sample size of about half that of level II, with weaker discrimination but lower cost; level III has a sample size of about 1.5 times that of level II, with stronger discrimination. For destructive inspections or extremely high-cost inspection items, special inspection levels S-1 to S-4 can be chosen (with very small sample sizes but significantly reduced discrimination).

The third step is to consult the tables to determine the sampling plan. Based on the batch size N and the inspection level, consult the sample size code table to get the code; then, based on the code and AQL, consult the main sampling table to get the (n, Ac, Re) combination. A common misconception is that the sampling plan is directly related to the batch size N. In reality, the batch size N only affects the code (i.e., the approximate range of n), while the acceptance number Ac is entirely determined by the AQL. This means that under continuous production conditions, even if the batch size fluctuates, as long as the AQL and inspection level remain unchanged, the sampling plan remains almost constant.

The fourth step, which is often the most overlooked, is to strictly enforce the transfer rules. The soul of GB/T 2828.1 is not the static (n, Ac) values but the dynamic "normal—tightened—reduced" transfer mechanism. When 2 out of 5 consecutive batches are rejected, the inspection should switch from normal to tightened; when 5 consecutive batches are accepted under tightened inspection, it can revert to normal inspection; when 5 consecutive batches under tightened inspection are still not accepted, the inspection should be suspended until the process quality is significantly improved. This mechanism ensures that when the process quality deteriorates, the sampling plan automatically tightens to protect the consumer's interests; and when the process quality is stable and better than the AQL, the inspection cost can be reduced under reduced inspection, incentivizing the producer to continuously improve.

Many companies' practical mistakes are: only checking a (n, Ac) plan and using it all the way without ever enforcing the transfer rules. This turns GB/T 2828.1 into a "static table lookup tool," losing its core value as a dynamic quality monitoring system.

3. Principles and Applicable Scenarios of Zero Defect Sampling (C=0 Plan)

"Zero defect sampling" typically refers to a sampling plan with an acceptance number Ac=0 — if even one nonconforming product is found in the sample, the entire batch is rejected. This plan is widely used in the automotive industry (IATF 16949), medical devices, and aerospace, with the core idea being: no nonconforming products are allowed to enter the next process or reach the end customer for key quality characteristics.

From the perspective of the OC curve, the C=0 plan is fundamentally different from the Ac>0 plan under the same AQL. For example, with an AQL of 0.65%, when the sample size code is G, the Ac=1 plan requires n=32, while the C=0 plan requires n=125 (about four times the former). In other words, the C=0 plan significantly increases the sample size to achieve a steeper OC curve and lower consumer risk. Under the same acceptance quality level, the C=0 plan has a significantly higher probability of detecting nonconforming products compared to the Ac>0 plan.

However, this does not mean that the C=0 plan is superior to the Ac>0 plan in all scenarios. From the perspective of inspection economics, the sample size of the C=0 plan is usually 3 to 5 times that of the Ac=1 plan, leading to a sharp increase in inspection costs. For non-key characteristics or characteristics with a fully stable process capability, using the C=0 plan results in over-inspection, which is not cost-effective.

Therefore, the reasonable application scenarios for zero defect sampling include: safety or regulatory-related characteristics, customer-mandated zero defect characteristics, initial verification stages of newly introduced products, and weak processes with a process capability ( Cpk < 1.33 ). For characteristics with a fully capable process ( ( Cpk \geq 1.67 ) ) and a stable historical quality performance, even using a conventional sampling plan, the probability of the process producing nonconforming products is extremely low. Maintaining the C=0 plan in such cases has limited practical value and wastes inspection resources.

4. Gradual Path from GB/T 2828.1 to Zero Defect Sampling

From a practical perspective, companies should not uniformly require all incoming and process inspections to use the C=0 plan, nor should they rigidly adhere to the old AQL system. A more reasonable path is "step-by-step transition and fine-grained management."

The first step is to classify and grade all inspection characteristics. Divide product characteristics into four levels: safety/regulatory characteristics, key characteristics, important characteristics, and general characteristics, each corresponding to different sampling strategies. Safety/regulatory characteristics must use the C=0 plan; key characteristics should use normal inspection with AQL ≤ 0.1% and a C=0 alternative plan; important characteristics should use normal inspection with AQL = 0.65% to 1.0%, fully implementing the transfer rules; general characteristics can use reduced inspection with AQL ≥ 1.5% to lower inspection costs.

The second step is to introduce process capability data into the dynamic adjustment of sampling plans. When a characteristic has a ( Cpk \geq 1.67 ) and no customer complaints for 12 consecutive months, the characteristic can be adjusted from the C=0 plan to normal inspection with AQL = 0.65%. Conversely, when ( Cpk < 1.33 ) or recent nonconformities occur, the characteristic should be upgraded to the C=0 plan. This data-driven dynamic adjustment ensures that inspection resources are allocated to high-risk areas while avoiding over-inspection in low-risk areas.

The third step is to promote the transition of inspection strategies from "attribute sampling" to "statistical process control." Sampling inspection is essentially a "post-event verification" method, while true quality control should be moved to the process. When the process is in a statistically controlled state (no abnormal points on the SPC control chart) and the capability is sufficient, the frequency of sampling inspection can be reduced from every batch to every n batches, or even gradually eliminated, focusing resources on real-time process monitoring. This aligns with the reduced inspection philosophy of GB/T 2828.1, but upgrades the judgment criteria from simple "consecutive batch acceptance counts" to more precise "process capability and control status."

5. Common Issues and Solutions for Implementing Sampling Inspection Systems in Enterprises

Issue 1: Inconsistent batch sizes in incoming inspection lead to frequent changes in sampling plans. The challenge in incoming inspection is the significant variation in the batch sizes of supplier shipments — sometimes thousands of items arrive at once, and sometimes only a few dozen. The solution is to establish a "fixed sample size" system: for a certain incoming material, determine a fixed sample size n based on the historical average batch size and AQL, regardless of the actual batch size of a single shipment. While this approach may cause slight fluctuations in the OC curve theoretically, it is more practical and avoids the chaos of frequent table lookups by inspectors.

Issue 2: Destructive inspection items are difficult to execute with large sample sizes. For destructive tests such as tensile testing and metallographic analysis, a large sample size can result in significant cost losses. In such cases, prioritize "replacing batch acceptance with process capability" — establish SPC control charts for destructive items, and draw very small samples (e.g., n=3 to 5) for process monitoring each shift or batch, rather than relying on large sample sizes to determine batch acceptability. This is essentially a combination strategy of "special inspection level + process monitoring."

Issue 3: Inspectors do not follow the sampling execution standards — they do not sample when they should, do not inspect when they sample, and do not record when they inspect. This is an execution-level issue within the quality management system. Solutions include: implementing LOT number management for incoming and process inspections, generating a unique inspection batch number for each batch, and having the system automatically generate the sampling plan and print sampling labels based on the preset AQL and inspection level. Inspectors should sample according to the specified locations and quantities on the labels and enter the inspection data into the system in real time, with the system automatically determining batch acceptance or rejection. This "paperless + system poka-yoke" approach can eliminate non-standard sampling issues at the process level.

Issue 4: Transfer rules are ineffective. Many companies have established transfer rules, but in practice, "tightened" or "reduced" inspections never truly occur. The reasons are usually: rejected batches are bypassed through concession acceptance (deviation release), preventing the system from recognizing the pattern of consecutive rejections; or inspectors and managers lack performance focus on the transfer rules. The solution is to have the quality information system automatically monitor the transfer conditions, and when the conditions are met, the system automatically switches the inspection level and records the reasons and dates for the switch, facilitating traceability during management reviews.

The core of sampling inspection is not to look up tables to select a plan, but to dynamically manage risks.

Knowledge code: 11.1.2

Version: v20260706

Author: Quality Think Tank Quality Think Tank is dedicated to providing systematic professional knowledge, methodologies, and practical tools for quality management practitioners, helping enterprises continuously improve their quality capabilities.