基于代理模型的水下结构物基座阻抗特性快速预报

Fast prediction of mechanical impedance of an underwater foundation based on surrogate models

[目的]基座结构的阻抗值直接影响水下结构物的隐蔽性能,通常借助有限元仿真获得基座阻抗特性。直接将阻抗有限元分析嵌套在基座优化设计中过于耗时,不利于优化设计,探究采用代理模型的方式直接拟合参数输入-阻抗输出的映射关系,取代有限元仿真,提高优化设计效率。[方法]选取水下结构物典型平台基座结构作为研究对象,首先计算分析其阻抗特性,然后分别通过4种常用代理模型(SVR,BPNN,RBF和Kriging)构造阻抗预报代理模型,对比分析其拟合精度。在此基础上,结合阻抗频响函数特性和工程实际,提出针对性的数据前处理方法,定量评估数据前处理方法对代理模型预报精度的提升效果。[结果]不同的代理模型在拟合平台基座阻抗特性精度上存在差异,其中Kriging代理模型在决定系数、标准化均方根误差和标准化最大绝对误差3项评价指标上均取得了最好的结果。基于前处理数据构建的代理模型与初始代理模型相比,标准化均方根误差和标准化最大绝对误差分别减小了9.80%和15.43%。[结论]Kriging代理模型对于拟合基座参数输入-阻抗输出的映射关系具备较好的适用性,提出的阻抗数据前处理方法可以进一步提高代理模型的预测精度。

[Objectives]The mechanical impedance of foundations can directly affect the stealth performance of underwater structures,which is commonly obtained by finite element simulation.However,this method is usually very time-consuming and inefficient when applied directly to optimal design.In order to enhance the efficiency,surrogate model technology is used to take the place of computer simulation.[Methods]The mechanical impedance of a typical underwater foundation is studied.The comparison among support vector regression(SVR)method,back propagation neural network(BPNN)method,radial basis function(RBF)method and Kriging model is conducted for their ftting accuracy.Besides,a data preprocessing method is put forward by combining the practical engineering requirement and characteristics of mechanical impedance.The effect of the data preprocessing method on the fitting accuracy of the surrogate model is studied.[Results]According to the calculated results,different surrogate models have different accuracy in fitting the mechanical impedance of the foundation.And the kriging model achieves the best results among four surrogate models.Compared with the initial model,the data preprocessing method has increased the normalized root mean square error and the normalized maximum absolute error by 9.80%and 15.43%respectively to the modified model.[Conclusions]The kriging model has the best applicability on fting the parameter input-impedance output mapping.And the proposed data preprocessing method can be used to improve the accuracy of the surrogate model.