ISO/CD TR 24962

Sampling-based conformity assessment

Scope

This TR provides an overview of the use of prior information in acceptance sampling. The methods described in the present TR can be applied for the inspection of both processes and lots.
Not only do manufacturing or production processes lie within the scope of the present TR; but the scope also covers any process whose outcome are discrete physical or digital units whose conformity can be assessed. In particular, the method described in the present TR can also be applied to AI-based classification systems. The production unit would then consist of the pair (object to be assigned to a class, assigned class) and a conforming unit would be defined as a correct classification.
As far as lots are concerned, the scope of the present TR includes both the inspection of isolated lots and serial lot inspection. The term “isolated lot inspection” does not mean that the consumer has no access to information regarding lot quality prior to the inspection of the current lot. Rather “isolated lot inspection” means that there are no switching rules and that the acceptance sampling plan is calculated separately for each new lot. In particular, the consumer having past experience with or knowledge regarding the producer of the lot currently under inspection is perfectly compatible with the concept of “isolated lot inspection.”
This TR consists of three main parts.
First, a risk-based approach is described (Section 7). This approach is based on concepts (such as specific consumer’s risk and conformance probability) defined in JCGM 106. It is shown how the sample size and the acceptance number can be calculated once a region for lot conformance and a tolerance for the specific consumer risk have been specified. In addition, an overview of information-based risks is provided.
Second, a utility-based approach is described (Section 8). This approach replaces the underlying principle of risk aversion with a rational cost-benefit calculus. Indeed, risk-based approaches consider neither the testing & sampling costs, nor hidden costs such as administrative overhead, nor the potential benefits associated with lot acceptance. By contrast, in the utility-based approach, all potential benefits, costs, losses and damages—including testing & sampling costs, potential costs associated with recalling a lot or healthcare, reputational costs, costs caused by the ingestion of contaminated food, and the bureaucratic and administrative costs associated with the implementation of regulations—are internalized in one utility function. In this sense, the utility approach bridges the gap between the “old world” of risks and the “new world” of utility. Tables with standard plans for various cost-structures and lot size values are provided.
Third, an approach for serial lot inspection is described (Section 9). In this approach, a Bayesian updating framework is provided in which data-ageing is reflected in a downweighting mechanism for older data.
Prior to these three parts, there are five preliminary sections:

·        Introduction (Section 2)
·        Background information regarding the plans in the ISO 2859 and ISO 3951 standards (Section 3)
·        Background regarding prior and posterior distributions (Section 4)
·        Overall framework in which the classical and information-based risks can be seen as complementary (Section 5)
·        Guidance for deriving a prior distribution and selecting an appropriate approach for the design of an acceptance sampling plan (Section 6)

Throughout this TR, the aim is to propose relatively straightforward and pragmatic methods for the design of acceptance sampling plans. Nonetheless, in parallel to the approach, the background theory is provided and illustrated with examples.
The present TR focuses on lots consisting of discrete items and lot inspection by attributes. The methods presented here can be extended for lots consisting of bulk material and for inspection by variables. This will be the subject of subsequent work.
Measurement and inspection error are not considered in this TR. The methods described here can be adjusted to take various error sources into account. This will be the subject of subsequent work.
Finally, in this TR, the testing outcome is usually considered to follow a binomial distribution. Strictly speaking, this is more appropriate for the underlying process than for the lot. The methods described in this TR can be adjusted in such a way that the testing outcome is modelled via the hypergeometric distribution, see Section 5.4.2.
A glossary of the most important terms and notation is provided at the end for convenient reference (Section 10).