Impact of noisy data on the characterisation of diversions in spent nuclear fuel assemblies using artificial neural networks

Abstract

Spent Nuclear Fuel (SNF) contains fissile materials that can be used to build nuclear weapons, therefore it’s important to safeguard it. Non-Destructive Analysis (NDA) techniques are used to detect potential defects in SNF. A method currently under research relies on interpreting the results of these NDA measurements via Artificial Neural Networks (ANN:s) in order to detect missing fuel pins. The quality of the predictions made by this ANN are affected by noise in data used to train and test the model. The aim of this study was to evaluate the ANN:s performance using data with simulated noise and to see how different network architectures affected the quality of the models predictions. It was concluded that the two alternative architectures that were tested had a negative impact on the performance of the ANN. It was also concluded that noisy data negatively impacts the performance of the ANN.