Explainable convolutional neural networks for acoustic detection of subcooled nucleate boiling; Validation via high-resolution decision visualization

植木 祥高*; 渡辺 晃也*; 相澤 康介 

Ueki, Yoshitaka*; Watanabe, Koya*; Aizawa, Kosuke

In this study, a deep-learning-based methodology was developed and validated to detect the onset of subcooled boiling and characterize its progression through acoustic signal analysis. A convolutional neural network architecture incorporating AlexNet and the synchrosqueezed wavelet transform was used to classify boiling-related acoustic signals with the highest accuracy. To improve the interpretability of the model's decision-making process, Guided Grad-CAM was used to extract and visualize salient features associated with the temporal and spectral characteristics of the boiling process. The model effectively identified intermittent bubbling patterns and their corresponding frequency components near the onset of nucleate boiling. The pressure fluctuations caused by vapor bubble generation and collapse were confirmed to be significant acoustic indicators of boiling events.