基于RBF神经网络的船舶冲击谱速度数据挖掘与预报

Data mining and prediction of ship shock spectral velocity based on RBF neural network

船舶水下非接触爆炸载荷作用下的冲击环境计算分析是船舶设备抗冲击的一项关键工作,而如何快速有效的对船舶冲击环境进行准确预报是人们关心的问题。建立了10余艘尺寸分布合理、形式多样的船舶有限元计算模型,每艘船舶沿船长在不同甲板层上选取了分布均匀的400组样本测点,载入水下非接触爆炸载荷进行计算,形成120多万条船舶冲击环境数据,以此建立了大规模船舶冲击环境数据库,并以径向基神经网络为框架搭建了船舶冲击环境预报模型,分别将船舶的主尺度参数、船舶水下爆炸数值仿真的工况设置参数以及考察点的位置坐标作为神经网络的输入参数,以船舶考察点的谱速度作为唯一输出对搭建的RBF网络进行训练,并通过聚类算法对网络参数进行优化处理,模型训练完成后对未知船舶在给定工况下的冲击环境进行了预报及分析。预报结果表明,经过优化算法优化后径向基神经网络预报模型不仅具有较高的预报精度,且具有较好的泛化和鲁棒性能。该方法可为设计阶段船舶冲击环境的快速预报提供一种新型方法。

Calculation and analysis of ship shock environment under action of underwater non-contact explosion load are a key work of ship equipment anti-shock design,and how to quickly and effectively predict ship shock environment is a problem of concern.Here,the finite element(FE)calculation models of more than 10 ships with reasonable size distribution and various forms were established.400 groups of sample measuring points for each ship were chosen and they were evenly distributed on different deck layers along ship length.After loading underwater non-contact explosion load,FE calculations were performed to form more than 1.2 million ship shock environment data,and establish a large-scale ship shock environment database.The radial basis function(RBF)neural network was used as the framework to establish the ship shock environment prediction model.Ship main-scale parameters,working condition setting parameters for numerical simulation of ship underwater explosion and position coordinates of investigation points were taken as input parameters of the neural network,and spectral velocity of ship investigation point was taken as the only output to train the established RBF network model.The network parameters were optimized using the clustering algorithm.After model training,shock environments of unknown ships under given working conditions were predicted and analyzed.The prediction results showed that the RBF neural network prediction model optimized using the optimization algorithm can not only have higher prediction accuracy,but also have better performances of generalization and robustness;this method can provide a new method for rapid prediction of ship shock environment in design stage.