Influence of sample intervals in real-sea trails on the nonparametric model of 3-DoF ship motion predictions

Abstract

The accuracy of ship manoeuvrability models based on real sea trail data heavily relies on data quality, as noise and measurement errors pose significant challenges in data processing and model training. This study employs a deep learning approach for the nonparametric modelling of 3-DoF ship manoeuvrability predictions, utilizing data from zigzag tests conducted at sea. It examines how the limitations of measurement equipment affect data analysis, emphasizing the impact of different sample intervals on model accuracy. Our study further explores the efficacy of deep neural networks in capturing low-frequency components more effectively than high-frequency ones, discussing the data sampling process and frequency analysis in the response domain for training set construction. Simulation results indicate that both excessively small and large sample intervals can significantly compromise predictions of motion, location, and heading angles. Moreover, to enhance the evaluation of the deep learning-based nonparametric model, incorporating a test case with a minimal rudder angle is crucial for assessing the prediction precision. Generally, a sample interval ranging from 0.4 s to 0.6 s is identified as optimal for data down sampling in real sea trails, balancing accuracy and computational efficiency.