Detection Of Partial Defects In Spent Fuel Assemblies With The Help Of Machine Learning

The detection of partial defects, i.e. missing fuel pins, in spent fuel assemblies is a challenging tasks during a safeguards inspection. Spent fuel assemblies are composed of a large number of fuel pins (e.g. 264 pins in a PWR 17x17 assembly geometry) and a proliferator can develop a practically infinite number of diversion scenarios to replace fuel pins with dummy pins. Safeguards inspections on spent fuel rely on non-destructive assays (NDA) to verify the presence of spent fuel and to detect fuel pins diversions. One of the NDA that has been developed specifically for the detection of fuel pins diversions is the Partial Defect Tester (PDET). The PDET instrument combines several detectors to measure the neutron and gamma-ray fluxes in several positions across the fuel assembly cross-section. Given the almost infinite number of possible diversion scenarios in a fuel assembly and the multivariate information resulting from a PDET measurement of a fuel assembly, machine learning (ML) has been chosen to tackle the data analysis. First, a large dataset has been developed with the results of Monte Carlo simulations representing measurement of complete fuel assemblies as well as assemblies with dummy pins. Then, a class label has been assigned to each dataset observation according to the percentage of replaced fuel pins. The problem of partial defect in a fuel assembly has been treated as a classification problem, using as input features the detector responses from the PDET instrument and as output values the class label of the observation. Both decision tree and k-nearest neighbor (kNN) approaches have been applied to develop the ML models. The results show that the use of multiple detectors in different locations within and outside the fuel assembly is helpful in the classification problem. By tuning the hyper-parameters for each ML approach the decision tree and kNN models were able to reach a 95% and >99% classification accuracy, respectively. A detailed discussion highlights the observations that are misclassified, with particular attention to the missed detections where assemblies with missing fuel pins are classified as complete assemblies by the ML models.