A comparative study of machine learning approaches for identification of perturbed fuel assemblies in WWER-type nuclear reactors

Enhancing the safety of nuclear power plants relies on the prompt and accurate identification of potential anomalies within the reactor. This paper explores the application of machine learning techniques for the identification and localization of perturbed fuel assemblies in WWER-type reactors. Various machine learning classifiers, spanning the decision tree, random forest, k-nearest neighbors, multilayer perceptron, support vector machine, and 1D-convolutional neural network, are scrutinized for their performance under diverse conditions.

The methodology encompasses data collection, data preprocessing, hyperparameter tuning, and model evaluation. The necessary dataset is generated using DYNOSIM to simulate all conceivable scenarios related to fuel assembly vibration in a WWER-type reactor. In addition to assessing the models under clear and complete input conditions, a sensitivity analysis is performed to gauge the models’ resilience to detector failures and the introduction of white noise. A comparative analysis of the six machine learning classification models reveals that multilayer perceptron, support vector machine, and 1D-convolutional neural network display the most sturdy classification performance, achieving accuracies of 76.38 %, 70.85 %, and 74.64 %, respectively.