Heuristics for managing trainable binary inspection systems

Binary inspection systems such as those based on ideal templates, neural networks, fuzzy logic, and genetic algorithms are trained by presenting them with exemplars of acceptable work. The system inspects new work by comparing it to the exemplars. The operator may not always agree with the judgment of the system and may decide to retrain it during production. This article explores the quality risks of using and of modifying trainable systems. Risk reduction heuristics used in software development are explored and adapted for use with trainable inspection systems. The use of these heuristics is illustrated in a series of scenarios.