Machine learning based classification of boiling and burnout heat flux using acoustic signals in nuclear thermal systems

Abstract

To prevent fuel damage and reactor instability, precise detection of boiling and burnout heat flux conditions is essential for nuclear power plant thermal safety. Using high-dimensional acoustic spectrum data acquired from controlled tests at high pressure thermo-physical bench, this paper investigates the use of supervised ML algorithms for the classification of thermal states, including normal boiling and burnout. Each of the 173 samples in the dataset is defined by 200 frequency-domain characteristics. A stratified 5-fold cross-validation pipeline was used to train seven ML models: Multilayer Perceptron, Logistic Regression, Support Vector Machine (RBF kernel), k-Nearest Neighbors, Random Forest, LightGBM, and CatBoost. Hyperparameters were adjusted using RandomizedSearchCV. Model interpretability was assessed with the use of SHAP values, permutation importance, and Gini scores, while feature selection was carried out using ANOVA F-statistics and Recursive Feature Elimination. Random Forest outperformed the other models in terms of test accuracy (88.57 %), recall consistency, and overall performance. Although they were not quite as stable in terms of interpretability, SVM and CatBoost also showed strong classification capabilities with high AUC values (≥ 0.82). The results show that ensemble-based classifiers work well in reactor settings with limited data and running in real-time. In order to provide insights into the performance of the models and their interpretability for safety-critical applications, this study builds a methodology for acoustic-based thermal diagnostics in nuclear systems.