基于联邦泛化的非平稳船舶横摇运动预测方法  

A federated generalization prediction method for non-stationary ship roll motion

作　　者：张琴[1] 刘敦康 李嘉宾 周福娜 胡雄[1] ZHANG Qin;LIU Dun-kang;LI Jia-bing;ZHOU Fu-na;HU Xiong(School of Logistics Engineering,Shanghai Maritime University,Shanghai 201306)

机构地区：[1]上海海事大学物流工程学院,上海201306

出　　处：《船舶力学》2024年第11期1654-1665,共12页Journal of Ship Mechanics

基　　金：国家自然科学基金资助项目(62373213);国家高技术研究发展计划项目(2013AA041106)。

摘　　要：船舶易受到风浪干扰而影响海上风力发电机的安装精度和维护安全性,其中长峰波随机波浪谱下的非平稳横摇运动影响最大。为保证海上作业在复杂海况下的稳定性,需要提高预测模型的泛化性,故本文提出基于联邦泛化的非平稳船舶横摇运动预测方法。首先,利用变模态方法分解非平稳船舶横摇运动为多分量平稳序列,进而采用注意力机制的长短期记忆神经网络建立本地多维多步预测模型,并进行误差校正;其次,为了提高复杂海况下遇到新类型船舶横摇运动时的预测效果,在不共享数据的前提下联合多家船舶横摇运动数据持有方进行择优联邦建模;最后,使用最大均值差异方法选择特征相似度高的数据进行加权平均联邦训练。实验结果表明,经过联邦学习后的模型具有更高的预测精度,以及更好的泛化能力,有助于风电安装时的波浪补偿稳定控制。Ships are susceptible to wind and waves causing the declines of installation accuracy and mainte⁃nance safety of offshore wind turbines.The most seriously affected case is the non-stationary ship roll motion under long-peaked random wave spectrum.To ensure the stability of offshore operations under complex sea conditions,it is necessary to improve the generalization of the prediction model.In this paper,a preferential feature federation method was proposed.Firstly,the non-stationary ship roll motion was decomposed into multi-component stationary sequences by using the variable modal decomposition method.Then,the long and short-term memory neural network with attention mechanism was used to build a local multi-dimension⁃al multi-step prediction model with error correction.Finally,in order to improve the prediction effect of new type ship motions in complex sea conditions,a federation method was used to combine some ship motion data holders for best model parameters,which were selected with the maximum mean discrepancy method with high similarity for preferential feature federated training.The experimental results show that the federated model has higher prediction accuracy and better generalization ability,which can help the stability control of wave compensation during offshore wind turbines installation.

关 键 词：船舶横摇运动预测 变分模态分解 注意力机制 LSTM 联邦学习 

分 类 号：U661.321[交通运输工程—船舶及航道工程]