In the last years, a database of simulated spent fuel observables was developed at SCK•CEN by combining the results of depletion-evolution codes and the responses of several detectors obtained by Monte Carlo models. We analysed the large amount of generated data with Artificial Neural Networks, by using the MATLAB toolbox. In this paper we focus on the application of Artificial Neural Networks to simulated Self-Interrogation Neutron Resonance Densitometry observables with the aim to quantify the 239Pu content in spent fuel. In view of a realistic application of the method, the number of data in the training and validation sets was limited to 20 spent fuel assemblies; the obtained performance when using randomly selected spent fuel assembly was compared with the one obtained when the spent fuel assemblies were selected by expert judgement. The average deviation between the nominal 239Pu content and the calculated 239Pu content in the testing data set was 0.2% with a standard deviation of 3.5% and a maximum deviation of 10%. It was found that the selection of spent fuel assemblies based on expert judgement results in better performances and therefore speeds up the data analysis when compared to a pure random selection of the data; hence the term natural, as opposite to artificial, is present in the title of the paper.
|Number of pages||7|
|State||Published - 1 Jun 2019|