Data-driven predictive analysis and visualisation of air–water dynamics in an air vessel

ABSTRACT Transient flow issues, particularly pressure surges in air vessels, significantly challenge the safe and efficient operation of water systems. This paper explores a hybrid approach, integrating machine learning, including deep learning, to address these challenges through predictive analysis. Focusing on a prevalent but often overlooked transient flow issue, this method combines visual data, from a high-speed camera, with numerical data from pressure transducers and a velocity profiler. A U-Net deep learning model performs image segmentation to quantify air–water mixture patterns, providing crucial input for subsequent pressure predictions. Three neural network models are developed, incorporating visual information derived from the segmentation. These models predict pressure variations within an air vessel, crucial for managing pressure surges and ensuring system safety. Experimental data from transient flow tests are used for training and validation. Results demonstrate that incorporating visual data significantly improves pressure prediction accuracy, generalising to both interpolation and extrapolation scenarios. The models, despite being trained with limited data, yield satisfactory predictions. Key challenges include dataset limitations for image segmentation and the practical acquisition of high-resolution visual data in real-world settings. These findings lay the groundwork for more effective real-time monitoring and control of water systems, contributing to improved safety and efficiency.