电工技术学报2024，Vol.39Issue(19)：5914-5928,5936,16.DOI:10.19595/j.cnki.1000-6753.tces.231549
电磁轨道发射器高速下电流密度场预测
Current Density Field Prediction Method for Electromagnetic Rail Launcher at High Speed
摘要
Abstract
During the electromagnetic rail launcher's launch process,the current distribution on the rail cross-section and armature-rail contact surfaces exhibits significant inhomogeneity,with current density approaching the material's bearing limit.Calculating the current density field during the launching process is one of the necessary conditions to realize the fine design of its structure.However,due to the complexity of the problem involving multi-field coupling and nonlinearity,refining the modeling and calculations using the finite element method(FEM)is time-consuming,often spanning hours or even days.Affected by the velocity skinning effect,the electromagnetic field control equation is a convection-diffusion equation,which is difficult or even impossible to compute at high speeds.Considering that the deep neural network has a powerful feature extraction ability and can learn the time-space change law of the field from numerical simulation data,this paper proposes a current density field prediction method based on ResUnet-Trans-CGAN at high speed.The dataset is obtained using FEM,and the high-precision interpolation prediction of the current density field and the extrapolation prediction of the difficult-to-stabilize computation stage at high speed can be realized after the training is completed,and the single computation time is within seconds. First,to realize the direct generation of the current density field,this paper uses the generative adversarial network(GAN),which is currently one of the most widely used image generation models.The conditional generative adversarial network(CGAN)addresses GAN's uncontrollable image generation issue by introducing a conditional variable y to the generator and discriminator.In order to realize the one-to-one mapping from input parameters to current density field,this paper removes the random noise z in the input,and directly inputs the excitation current,armature conductivity,rail structure parameters and time t into the neural network as the conditional variables,and establishes the current density field prediction model based on CGAN.Then,to improve the feature extraction capability for complex current density field distributions,the ResUnet-Trans network is constructed as the feature generator of CGAN.The U-Net structure is used to realize the fusion of different levels of features,short Residuals are introduced to inhibit the deep neural network degradation,and the Transformer is introduced to enhance the model's ability to extract global features.Second,dropout and a cosine annealing algorithm for adaptive learning rate adjustment are introduced in the model to curb overfitting and improve prediction accuracy.Further,a transfer learning training strategy,transitioning from low to high conductivity,is applied during training to improve the model's extrapolation prediction accuracy and generalization ability.Finally,the accuracy of model interpolation and extrapolation prediction is verified with the examples of low-speed and high-speed electromagnetic rail launcher,respectively. The results show that for interpolation prediction,the ResUnet-Trans-CGAN proposed in this paper has the highest accuracy with the mean absolute percentage error(MAPE)value of current density field prediction less than 1.5％compared to CGAN,Unet-CGAN and ResUnet-Trans.After using the transfer learning training strategy from low to high conductivity,the MAPE is reduced by 9.59％and 19.0％at the rail cross-section and at the armature-rail contact surface,respectively.When extrapolating the prediction at high speeds,the MAPE of the current density field at the rail cross-section is reduced from 1.83％to 1.52％and the training speed is improved by 17.0％,and that at the armature-rail contact surface is reduced from 3.47％to 2.43％,and the training speed is improved by 22.1％,after using the Transfer Learning training strategy of moving from low to high conductivity on the test set.The model can generate the current density field directly at the stage where it is difficult to perform stable numerical simulation at high speed.关键词
电磁轨道发射器/电流密度场分布/深度学习/图预测Key words
Electromagnetic rail launcher/current density field distribution/deep learning/image prediction分类
信息技术与安全科学引用本文复制引用
金亮,宋居恒,马天赐,尹振豪,张陈源..电磁轨道发射器高速下电流密度场预测[J].电工技术学报,2024,39(19):5914-5928,5936,16.基金项目
国家自然科学基金重大研究计划(92066206)、国家自然科学基金面上项目(51977148)和中央引导地方科技发展专项自由探索项目(226Z4503G)资助. （92066206）
