Session: 12-03-01 Deterministic Wave and Motion Prediction

Paper Number: 78601

78601 - Application of Machine Learning for the Generation of Tailored Wave Sequences 

Model tests on waves and their effects on ships and offshore structures are still indispensable for validation purposes, both in the field of research and development and industrial application. Hereby, the generation of waves in seakeeping basins is a critical prerequisite for wave-structure investigations. In most cases, the standard model of ocean waves, which is based on the superposition of linear wave components, is sufficient, e.g., for the generation of stochastic irregular sea states. Thus, the linear dispersion relation of the independent wave components enables the description of the sea states in space and time along the wave tank. However, the generation of deterministic sea states, i.e., the reproduction of a predefined wave group at a desired target location in the wave tank, requires more sophisticated approaches, particularly for steep wave groups which do not obey linear wave theory. So far, only complex, empirical or time-consuming methods are available hindering a straightforward plug-and-play solution for test facilities.

This paper explores the applicability of machine learning for the generation of tailored wave sequences. The objective is to evaluate if machine learning methods can be utilized for the task-related generation of predefined wave sequences. For this purpose, a fully convolutional neural network model is implemented relating the target wave sequence at the target location in time domain to the respective wave board motion control signal. In a first step, a neural network architecture search is performed by means of a cross-validation and hyper parameter study, to identify the model architecture which fulfils the desired quality criteria with minimum complexity. Training and validation data are acquired by combining the standard model of ocean waves with linear transfer function of the wave board, which allows the generation of large synthetic data sets for model training. Next, physical data sets are obtained from model tests by varying the parameters of wave steepness and carrier frequency as well as distance to the wave board focussing on short wave sequences. Finally, the machine learning model is validated against the experimental data featuring complex non-linear wave groups.

Presenting Author: Svenja Ehlers Hamburg University of Technology

Authors:

Marco Klein Hamburg University of Technology
Merten Stender Hamburg University of Technology
Mathies Wedler Hamburg University of Technology
Svenja Ehlers Hamburg University of Technology
Moritz Hartmann Hamburg University of Technology
Nicolas Desmars Hamburg University of Technology
Marc-André Pick Hamburg University of Technology
Robert Seifried Hamburg University of Technology
Norbert Hoffmann Hamburg University of Technology