Session: 06-17-05 AI Technology for Ocean Engineering - V

Submission Number: 183936

Image-Driven Regression AI for Sea State Prediction in a Confined Volume With Discrete Decomposition 

Floating structures such as surface ships are subject to wave loads. Monitoring and modeling sea states and wave conditions are essential for accurately estimating these imposed wave loads in real time. However, many existing wave models and filtering methods fail to capture nonlinear effects such as refraction, diffraction, or shoaling, which depend on the topographic and bathymetric features of the domain. Likewise, they do not account for structure- or egoshape-dependent effects such as slamming. 
Image-based methods for capturing the current free surface exist, most prominently WASS, but these techniques are, as of yet, primarily used to measure the instantaneous surface properties of the water and are not well suited for predicting future states. 
Convolutional neural networks (CNNs) with a U-Net architecture, i.e. encoding–decoding networks, have proven highly effective for analyzing tensor-like structures. In this study, we investigate the use of such an architecture, conditioned on images of the free surface, the positional encoding of the cameras capturing those images, and motion extraction, to predict a simulation-ready state. The chosen simulation tool is REEF3D, using the NHFLOW solver. The tensor-like state is estimated by the convolutional network, including elevation, velocity, and pressure for each grid point in the sigma coordinate grid. Predicting future states of the water domain is then performed using the REEF3D::NHFLOW solver. 
The resulting wave model therefore represents both the current and predicted future states of the bathymetry-following sigma coordinate grid. The strength of this representation lies in its ability to compute wave loads using direct forcing and panel theory. 
To determine which aspects of the architecture, such as motion extraction, camera position, and encoding method, have the greatest impact on the network’s performance, an ablation study will be conducted. The performance of the network will be evaluated based on the final loss of each model, which, for the time being, is defined as a weighted norm distance over the elevation, velocity, and pressure fields. 

Presenting Author: Jon Estil Krågebakk NTNU

Presenting Author Biography: Jon Estil Krågebakk holds a Master’s degree in Engineering Cybernetics, during which he completed a thesis on visual servoing. He is currently pursuing a PhD focused on applying artificial intelligence to the operation of autonomous surface vessels. His latest research investigates image-based regression models to predict a discretized decomposition of a water domain in a confined volume around a point of interest, using query images taken of that point.

Outside academia, he enjoys building fighting robots with friends, cross-country skiing, swimming, cycling, and running, perhaps a bit more enthusiastically than the average person.

Authors:

Jon Estil Krågebakk NTNU
Widar Weizhi Wang NTNU
Ekaterina Kim NTNU
Hans Bihs NTNU