基于深度卷积生成式对抗网络的船型特征认知与条件生成方法  

A ship hull offset feature cognition and generation method based on conditional deep convolutional generative adversarial networks

作　　者：杜林 李胜忠[2] 李广年 舒跃辉 刘子祥 赵峰[2] DU Lin;LI Sheng-zhong;LI Guang-nian;SHU Yue-hui;LIU Zi-xiang;ZHAO Feng(Maritime and Transportation College,Ningbo University,Ningbo 315000,China;China Ship Scientific Research Center,Wuxi 214082,China)

机构地区：[1]宁波大学海运学院,浙江宁波315000 [2]中国船舶科学研究中心,江苏无锡214082

出　　处：《船舶力学》2024年第8期1162-1174,共13页Journal of Ship Mechanics

基　　金：船舶总体性能创新研究开放基金资助项目(11322203);国家自然科学基金青年科学基金资助项目(52201368);高等学校学科创新引智计划项目(D21013)。

摘　　要：船体型值与图片一样也是序列相关型数据,所以用于生成图片的神经网络模型也能生成船型数据。由于船舶种类繁多、需求复杂,本文研究重点从船舶水线上下、船艏、舯、艉等区域位置特征,和船舶设计中普遍存在球艏、尾轴、艏部外板升高等全局几何特征的条件生成需求出发,训练条件深度卷积生成式对抗网络模型(Con⁃ditional Deep Convolutional Generative Adversarial Networks)来实现两种特征的条件认知与生成。首先,将实现船型区域位置特征与全局几何特征的条件生成作为目标,分别建立条件深度卷积生成式对抗网络模型;然后,针对两类特征设置若干从易到难的不同分割方案和特征种类,使神经网络能够循序渐进地完成条件生成任务;最后,通过对训练过程和生成结果进行对比,初步证明所研究方法用于解决船型特征条件生成问题的可行性。本研究延续了作者之前的研究成果,属于基于计算机视觉技术的船型智能设计方法领域,旨在进一步探索引入人工智能实现船型智能设计的可行性方法。The hull form modelling progress in ship design is significantly relied on the parent hull database and the professional designers well trained with CAD software,and it is usually a time and experience costly work.The conditional generation of ship hull with both geometrical and locational features by training an arti⁃ficial neural network was concerned by this paper.The geometrical feature means the overall shape variety of ship designs like bulbous bow,stern shaft,etc.,the locational feature means the shape difference between stern,front and mid-body of ships.Firstly,a conditional deep-convolutional generative adversarial network(CDC-GAN)was constructed to distinguish the geometrical and locational features individually;Secondly,the CDC-GAN was well trained to learn and generate these features with different resolutions and categories,from easy to hard;In the end,the training cost and performance of networks were compared and concluded to prove the capability of CDC-GAN in solving ship hull form generating issues.This paper is based on au⁃thors’previous investigation with regular GAN,and it provides a further exploration about the potential of CDC-GAN in ship design.

关 键 词：船型智能设计 深度卷积生成式对抗网络 计算机视觉 

分 类 号：U662.9[交通运输工程—船舶及航道工程]