Untitled

Abstract

To evaluate the time-domain positioning performance of arctic marine structures, it is necessary to generate an ice load appropriate for the current position and heading of the structure. The position and orientation angle of a floating body continuously change with time. Therefore, an ice load is required for any attitude in the time-domain simulation. In this study, we present a fundamental technique for analyzing ice loads in the frequency domain based on data measured at various angles in the ice-water tank experiment. We perform spectral analysis instead of general FFT to analyze the ice load, which has the characteristics of a random signal. To generate the necessary ice load in the time domain, we must first interpolate the measured data in the frequency domain. Using the Blackman-Tukey method, we estimate the spectrum for the measured data, then process the data to generate the training set required for machine learning. Based on the results, we perform regression analysis by applying four representative techniques, including linear regression, random forest, or neural network, and compare the results with MSE. The deep neural network method performed best, but we provide further discussion for each model.