基于耦合特征LSTM的船舶运动预报方法研究  

Research on ship motion prediction method based on coupled feature LSTM

作　　者：何国联 姚朝帮 孙小帅 孟凡亮 HE Guolian;YAO Chaobang;SUN Xiaoshuai;MENG Fanliang(School of Naval Architecture and Ocean Engineering,Huazhong University of Science and Technology,Wuhan 430074,China;Unit 91054,Beijing 102422,China;System Avionics Research Office,China Special Vehicle Research Institute,Jingmen 448035,Hubei China)

机构地区：[1]华中科技大学船舶与海洋工程学院,湖北武汉430074 [2]91054部队,北京102422 [3]中国特种飞行器研究所系统航电研究室,湖北荆门448035

出　　处：《华中科技大学学报（自然科学版）》2024年第4期1-7,共7页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(52071148);工信部民机专项资助项目(2022KF0031).

摘　　要：采用长短期记忆(LSTM)神经网络模型对船舶在不同海况下的运动姿态进行预报.针对LSTM模型难以优化的特点,提出了一种耦合特征LSTM神经网络模型.首先对船舶运动时间序列数据进行了归一化处理;然后基于深度学习框架TensorFlow搭建了具有输入层、隐藏层和输出层的LSTM模型;接着将原始数据按照不同特征输入形式进行划分;最后采用不同耦合特征LSTM模型分别对测试样本进行预报.结果表明:相比于其他LSTM模型,六自由度耦合特征LSTM神经网络模型的预报精度有明显优势;在四级海况下,运动预报误差降低了2.1%~12.9%;在五级海况下,运动预报误差降低了2.4%~12.3%;六特征耦合LSTM模型只须进行一次计算,就能同时输出六自由度运动,可减少51.4%~82.7%的计算时间,提升了计算效率.Long short-term memory(LSTM)neural network model was used to predict the motion of ship advancing in waves under different sea states.Aiming at the characteristic that LSTM model was difficult to optimize,a coupled feature LSTM neural network model was proposed.First,the ship motion time series data were normalized.Then,the LSTM model with input layer,hidden layer and output layer was built based on the deep learning framework TensorFlow,and the original data were divided according to different features.Finally,LSTM models with different coupling features were used to predict the tested samples.Results show that the six-feature coupled LSTM neural network model has obvious advantages compared with other LSTM models for the prediction accuracy.The motion prediction error is decreased by 2.1%~12.9%in the level-4 sea state.In the level-5 sea state,the motion prediction error is decreased by 2.4%~12.3%.The six-feature coupled LSTM model can simultaneously output six degrees of freedom(six-DOF)motion with only one calculation,which can reduce the calculation time by 51.4%~82.7%and improve the calculation efficiency.

关 键 词：六自由度运动 长短期记忆(LSTM) 时间序列 耦合特征 预报精度 

分 类 号：U661.1[交通运输工程—船舶及航道工程]