Automatic detection of the onset of film boiling using convolutional neural networks and Bayesian statistics

Abstract

The present article shows that a combination of Bayesian statistics and convolutional neural networks can be used to successfully detect the transition from nucleate to film boiling from visualization, even if the heater is not visible in the visualization window of an on-wire boiling process. Using a trained convolutional neural network to classify boiling heat transfer regimes, this paper builds upon previous studies that show that machine learning algorithms can accurately infer boiling heat transfer regimes from visualization, and proposes the utilization of Bayesian statistics to be able to detect the transition from nucleate to film boiling with arbitrarily large confidence within seconds. Results suggest that the precise detection can be potentially done sooner than conventional temperature measurement sensors such as commercial thermocouples and RTDs. Finally, this paper presents several lower bounds to the time to detection of film boiling deflagration, which indicate that sub-second non-intrusive automatic detection of film boiling may be reached, especially with state-of-the-art machine learning algorithms.