基于迁移学习的多源波浪载荷融合方法研究  

Multi-source wave load fusion method based on transfer learning

作　　者：陈帅 蒋彩霞[1,2] 王子渊 张凡 王艺陶[1,2] CHEN Shuai;JIANG Cai-xia;WANG Zi-yuan;ZHANG Fan;WANG Yi-tao(China Ship Scientific Research Center,Wuxi 214082,China;Taihu Laboratory of Deepsea Technological Science,Wuxi 214082,China)

机构地区：[1]中国船舶科学研究中心,江苏无锡214082 [2]深海技术科学太湖实验室,江苏无锡214082

出　　处：《船舶力学》2023年第10期1431-1444,共14页Journal of Ship Mechanics

基　　金：国家自然科学基金项目(52001284);江苏省自然科学基金项目(BK20200166)。

摘　　要：为了泛化模型试验小样本数据,减少试验成本,本文以某船波浪载荷短期预报为研究对象,基于迁移学习建立多源波浪载荷融合(multi-source wave load fusion,MSWLF)方法。首先,将拉丁超立方抽样(Latin hypercube sampling,LHS)、数据处理与分析、切片理论和深度神经网络(deep neural networks,DNN)结合建立低频波浪载荷智能预报(low frequency wave load intelligent prediction,LFWLIP)方法,实现线性波浪载荷的智能预报。其次,将贝叶斯优化算法和DNN结合,构建优化数学模型,通过计算得到精度较高的LFWLIP模型。最后,以理论计算波频数据作为源域一、规则波模型试验波频数据作为源域二、不规则波模型试验合成数据(波频与砰击叠加)为目标域,基于迁移学习对不规则波工况下的波浪载荷预报进行二次修正。结果表明:在任意工况下,MSWLF方法能以较高的精度预测不规则工况下的波浪载荷短期预报值,载荷预测值与试验值误差小于20%。该方法对于不同海况尤其是高海况下船舶波浪载荷快速评估具有重要意义。In order to generalize the small sample data of model test and reduce the test cost,a multi-source wave load fusion method was established based on transfer learning for short-term wave load prediction of a ship.Firstly,the Latin hypercube sampling,data processing and analysis,strip theory and deep neural networks were combined to establish a low-frequency wave load intelligent prediction method,and the intelligent prediction of linear wave loads was realized.Secondly,by combining the Bayesian optimization algorithm with DNN,the optimization mathematical model was constructed,and the LFWLIP model with a high accuracy was obtained by calculation.Finally,taking the theoretical calculation wave frequency data as Source Domain 1,the regular wave model test wave frequency data as Source Domain 2,and the irregular wave model test synthetic data of wave frequency and slamming superposition as the target domain,the wave load prediction under wave conditions was revised twice based on transfer learning.The results show that the MSWLF method can predict the short-term prediction value of wave load under irregular conditions with a high accuracy under any condition,and the error between the predicted value and the test value is less than 20%.This method is of great significance for rapid assessment of wave load under different sea conditions,especially high sea conditions.

关 键 词：MSWLF方法 LFWLIP方法 迁移学习 贝叶斯优化算法 模型试验 小样本数据 

分 类 号：U661.1[交通运输工程—船舶及航道工程]