Nonintrusive identification of boiling regimes enabled by deep learning based on flow boiling acoustics

Monitoring two-phase cooling systems and identifying corresponding boiling regimes are vital to avoid device failure owing to thermal runaway. Boiling acoustic characteristics can be used to develop a nonintrusive monitoring tool because of its close relation to the bubble behavior. Compared with bubble image-based techniques, acoustic sensing is cheaper and easier to implement. However, the acoustic data available is very limited, especially for flow boiling. Also, real-time strategies are still lacking since the prediction models based on boiling acoustics are usually offline. Additionally, noise interference typically degrades model prediction performance. To promote the implementation of an automated boiling system amidst these challenges, we present a deep learning framework for real-time identification of boiling regimes using acoustic signals of flow boiling. Flow boiling experiments are first performed using HFE7100 under different conditions and then combined with deionized water data from our previous experiments to construct the acoustic dataset. We use the corresponding acoustic spectrograms to train and test a convolutional neural network which shows a prediction accuracy of 99.97 % without noise interference. A transfer learning strategy is further employed to improve the original model performance under noise interference. Finally, we add a decision logic to adapt the CNN model to acoustic data streams for real-time identification of boiling regimes. The CNN model with the decision logic can accurately identify the boiling regimes in real-time by scanning the corresponding spectrograms irrespective of working fluids and conditions. In practice, our strategy can be utilized to initiate the necessary precautions to avoid overheating-induced failure.