Utility of neural networks in nondestructive waste assay measurement methods

摘要

Concepts devised to utilize nondestructive assay measurement techniques to quantify waste entrained radioactive material mass, transuranic and otherwise, must necessarily contain provisions for complexity. Such complexities are founded in the multi-variate attribute distributions associated with the typical waste form and the inherently limited nature of present-day nondestructive assay (NDA) sensors and detection techniques. For many waste forms, the attribute variables are such that commonly employed NDA techniques do not possess the capability to acquire accurate measures useful in deriving a viable solution using first principle modeling techniques. The existence of limitations in commonly employed NDA instrumentation and techniques logically leads to a search for an alternate view or paradigm for data treatment. The approach addressed in this paper shifts from model-driven algorithmic methods to data-driven empirical methods. Such empirical methods are statistical in nature, and possess desirable characteristics of adaptivity and learning. Examples of modern empirical methods include neural networks, fuzzy logic, genetic algorithms, and combinations thereof. This work provides an investigation into the utility of three neural network architectures for deriving useful information for nondestructive waste assay solutions. To illustrate the inherent capability of these data-driven techniques, a simple waste form classification exercise is performed using radial basis function, counterpropagation, and adaptive resonance theory neural networks. The classifications are derived solely from the self-organizing and adaptation capabilities of the network architectures and associated learning rules. No apriori information or models are assumed during the classification exercises.