Acoustic detection of boiling states by deep convolutional neural networks: Visual explanation of identification basis

With a final goal of early detection and understanding of the transition of coolant boiling events in the core of sodium-cooled fast reactors, we aim to obtain and maintain the basic knowledge necessary for developing anomaly detection technology associated with local anomalies in the core and to demonstrate basic feasibility in the present study. We constructed a deep learning method and evaluated its performance to detect the occurrence and understand the transition of subcooled boiling using acoustic identification. In this research, we acquired acoustic data during subcooled boiling of ultrapure water and learned feature quantities of the boiling in time–frequency expression. A deep learning model of a convolutional neural network for label classification was constructed. Visualization of the discrimination criteria using the gradient-weighted class activation mapping method revealed the acoustic frequency bands that the deep learning model deemed important, enabling it to detect the onset of nucleate boiling with interpretability and accountability. We also constructed a regression analysis-type deep learning model and demonstrated that boiling heat flux values can be predicted with high accuracy. • CNN trained with acoustic data in TFR accurately identify nucleate boiling states. • AlexNet trained with STFT dataset is the best among the present combinations. • Grad-CAM visualizes frequency bands CNN models focus on in label classification. • CNN regression analysis accurately predicts boiling heat flux from acoustic data.