摘要
悬链式单点系泊需要建立基于基础条件、工作条件、自存条件、运动和力要求等输入的模拟环境,并进行多点测试来寻找最佳设计.通过仿真计算构建2个主要数据集,即oper-ation 数据集和 self 数据集.对self数据集进行预测,并将数据分为局部、全局和全局加局部3类进行训练和验证,使用4层全连接神经网络来预测回归问题,准确率可达90％以上.将该模型应用于更复杂的operation数据集时的效果并不理想.采用DNN+BN+ReLU作为最小分量自建模型DBRNet12复杂网络处理operation的数据得到86％的平均准确率.依据残差思想在DBRNet12基础上自建RNet40网络取得了 90％的平均准确率.在网络架构方面,搭建了深度神经网络,通过全连接层进行预测,并对网络结构进行了持续的优化.最后,通过相对误差的评估来衡量预测效果的优劣,并利用残差网络进行优化.
Abstract
A catenary single-point mooring requires a simulation environment based on inputs such as basic conditions,operating conditions,self-storage conditions,motion and force requirements,and multi-point testing to find the optimal design.The paper uses deep learning method to solve the mooring system prediction problem.Firstly,two datasets,i.e.,self dataset and operation dataset are acquired by simulation calculation.Then the self dataset is predicted and the data is divided into three categories:local,global,and global plus local for training and validation,and a two-layer full-connected neural network is used to predict the regression problem with an accuracy of over 90％.As the results are not satisfactory when the model is applied to more complex operation datasets,a self-built model DBRNet12 complex network using DNN+BN+ReLU as the minimum component is added to handle more operation data,thus obtaining an average accuracy of 86％.The self-built RNet40 network based on the idea of residuals on DBRNet12 achieves a 90％average accuracy.In terms of network architecture,a deep neural network is built to predict parameters through fully connected layers,and the network structure is continuously optimized.Finally,the evaluation of relative error is used to evaluate the effectiveness of the prediction and the residual network is used for optimization.Through this procedure,the application effect of deep learning methods in mooring system prediction problems is achieved,and the ideas provide references for further research and practice in this field.
基金项目
国家重点研发计划深海关键技术与装备重点专项(2017YFC0307800-02)
国家海洋工程装备研发及产业化专项(船重科[2014]735号)
NETL
NSTL
万方数据