A two-stage neural network approach for heat flux quantification from boiling images using vision transformers and transfer learning

Abstract

Pool boiling, a fundamental heat transfer process, has been a subject of extensive research due to its significance in various industrial applications. Accurate heat flux quantification is essential for assessing heat transfer performance, but traditional methods face limitations such as complex modeling and intrusive measurement techniques. Recent advances in deep learning have enabled the use of visual data for heat flux quantification, yet challenges such as high dataset labeling costs, small sample sizes leading to overfitting, and the demand for high accuracy in fine-grained tasks persist. This paper proposes a two-stage neural network approach to address these challenges. In the first stage, a self-supervised learning model is pre-trained on public boiling image datasets to extract useful features without requiring labeled data. The second stage involves fine-tuning this model on a small, labeled in-house dataset for precise heat flux quantification. This approach significantly reduces the reliance on large labeled datasets while maintaining good predictive accuracy and effectiveness, even with limited data availability. The proposed method achieved an accuracy of 0.953 (ACC1) and 0.929 (ACC2) on the test set. Even when trained on smaller samples where traditional one-stage models experience a significant drop in accuracy, the two-stage training strategy ensures more effectively maintained prediction accuracy.