Machine learning classification of boiling regimes with low speed, direct and indirect visualization. (October 2018)

• Record Type: Journal Article
• Authors: Hobold, Gustavo M.; da Silva, Alexandre K.
• Is Part Of: International journal of heat and mass transfer. Volume 125(2018)
• Page Start: 1296
• Page End: 1309
• Publication Date: 2018-10
• Subjects: Pool boiling -- Film boiling -- Machine learning -- Boiling regimes -- Visualization; Heat -- Transmission -- Periodicals; Mass transfer -- Periodicals; Chaleur -- Transmission -- Périodiques; Transfert de masse -- Périodiques; Electronic journals; 621.4022
• DOI: 10.1016/j.ijheatmasstransfer.2018.04.156
• Languages: English
• ISSNs: 0017-9310

Abstract

Highlights: Low speed/resolution image acquisition used for boiling regime classification. Machine learning algorithms can identify pool boiling regimes with over 93% accuracy. Non-intrusive, fast and accurate boiling regime classification methodology suggested. Abstract: Multiphase flow pattern identification is of utmost importance to the energy industry, given that thermohydraulic operating conditions are drastically affected by flow and heat transfer regimes. In industrial boilers and nuclear reactors, for instance, the heat transfer coefficient – and hence the heater temperature – is significantly affected by the boiling regime, where the onset of film boiling can be catastrophic to the equipment and cause irreparable damage. In this paper, it is shown that a machine can learn from visualization and successfully classify and separate natural convection, nucleate boiling and film boiling regimes using low speed and low resolution image frames acquired from visualization of an on-wire pool boiling experimental setup (direct visualization) even when only the departed, ascending bubbles are considered – i.e., the heater is suppressed from the image (indirect visualization). While not the main objective of this paper, principal component analysis of the frames is shown to provide information regarding bubble dynamics and hence is used for dimensionality reduction. Two types of classifiers, namely support vector machines and neural networks, are shown to be able to classify pool boiling frames with over 93% accuracy sufficiently fast, possibly enabling real-time execution and classification, even during indirect visualization and, hence, providing a non-intrusive and low-cost pool boiling regime identification.