Multivariate analysis applied to the characterization of spent nuclear fuel
MetadataShow full item record
The Multi-Isotope Process Monitor is being developed at Pacific Northwest National Laboratory as a method to verify the process conditions within a nuclear fuel reprocessing facility using the gamma spectra of various process streams. The technique uses multivariate analysis techniques such as principal component analysis and partial least squares regression applied to gamma spectra collected of a process stream in order to classify the contents as belonging to a normal versus off-normal chemistry process. This approach to process monitoring is designed to function automatically, nondestructively, and in near real-time. To extend the Multi-Isotope Process Monitor, an analysis method to char- acterize spent nuclear fuel based on the reactor of origin, either pressurized or boiling water reactor, and burnup of the fuel using nuclide concentrations as input data has been developed. While the Multi-Isotope Process Monitor uses gamma spectra as input data, nuclide activities were used in this work as an initial step before Nuclide composition information was generated using ORIGEN-ARP for different fuel assembly types, initial 235U enrichments, burnup values, and cooling times. This data was used to train, tune, and test several multivariate analysis algorithms in order to compare their performance and identify the technique most suited for the analysis. To perform the classification based on reactor type, four methods were considered: k-nearest neighbors, linear and quadratic discriminant analysis, and support vector machines. Each method was optimized, and its performance on a validation set was used to determine the best method for classifying the fuel reactor class. Partial least squares was used to make burnup predictions. Three models were generated and tested: one trained on all the data, one trained for just pressurized water reactors, and one trained for boiling water reactors. Quadratic discriminant analysis was chosen as the best classifier of reactor class because of its simplicity and its potential to be extended to classify spent nuclear fuel’s fuel assembly type, i.e, more specific classes, using nuclide concentrations as input data. In the case of predicting the burnup of spent fuel using partial least squares, it was determined that making reactor-specific partial least squares models, one trained for pressurized water reactors and one trained for boiling water reactors, performed better than a single, general model that was trained for all light water reactors. Thus, the the classifier, regression algorithm, and all the necessary intermediate data processing steps were combined into a single analysis method and implemented as a Matlab function called “burnup.” This function was used to test the analysis routine on an additional set of data generated in ORIGEN-ARP. This dataset included samples with parameters that were not represented in the development data in order to ascertain the analysis method’s ability to analyze data for which it has not been explicitly trained. The algorithm was able to achieve perfect binary classification of the reactor as being a pressurized or boiling water reactor on the dataset and made burnup predictions with an average error of 0.0297%.