Core Discharge Monitor Machine Learning Algorithm For Candus With Fuel Separator Method

Angelo Antonio Alessandrello - International Atomic Energy Agency
Federica Mingrone - International Atomic Energy Agency
Stephan Bertl - International Atomic Energy Agency
A large fraction of verification activities performed by the IAEA nuclear safeguards inspectors is on safeguards data coming from CANDU reactor facilities. Traditionally the number of fuel bundles discharged from a CANDU reactor are counted using the VXI Integrated Fuel Monitor (VIFM) detector suite, in particular the Core Discharge Monitor (CDM), which counts the number of CANDU bundles discharged from the core, and the Bundle Counter (BC), which counts the number of bundles transferred to the fuel pond. Previously, the IAEA used an internally developed software that was counting the bundles by analyzing the specific signature of the discharge signal in the CDM detectors. This review software has been updated with the Integrated Review and Analysis Package (IRAP) that allows correlated analysis, merging the CDM results with the BC results. However, this correlated analysis could not be performed efficiently to some CANDUs that use the fuel separator method. In such cases, the CDM signal is not easily interpreted and deconvolved in bundle counting using traditional methods. Recently, the IAEA has evaluated different machine-learning methodologies for signal classification and bundle detection, including Recurrent Neural Networks (RNN), Decision Tree Classifiers (DTC) and Random Forest Classifiers (RFC). Data preparation was performed by extracting specific features from the raw neutron and gamma counts. The RNN based approached provided good results but required CPU-intensive calculations and were more complex to integrate in IRAP. DTC and RFC provided equal or better performance, with the advantage of being able to map non-linear functions. Furthermore, the results of such classifiers are not complex to interpret, and can be transformed in any target programming language. Based on these results, DTC was implemented as the primary algorithm for signal classification and bundle detection in IRAP analysis for CANDUs with separator method. However, further analysis of other algorithms might be performed in future for possible improvements. Starting from the description on how to build core discharging events using IRAP, the paper will provide an overview of the techniques used to develop novel algorithms that are capable of counting CDM bundles with an efficiency greater than 20% if compared to traditional algorithms.