The Fast Modeling Method Of Composite Scattering Of Target And Rough Surface Based On Neural Network

When analyzing the composite scattering of a target and a rough surface,numerical methods are usually used if the problem is small.When the problem becomes lager,a high-frequency approximate method or hybrid high-frequency and numerical method is more favorable.In these method,the target and the rough surface should be discretized and treated as a whole problem.If the position,shape or number of the targets changes,the entire composite scattering needs to be recalculated,which greatly increases the amount of repetitive calculations.In this case,we can consider using the numerical Green’s function method,which treats the rough surface as part of the background.Then we can calculates the numerical Green’s function of the background,and uses the numerical Green’s function to establish an integral equation.In this manner,if the target changes,only the target needs to be recalculated.However,if there are too many unknowns on the rough surface,it is difficult to obtain the numerical Green’s function.Neural networks have the characteristics of high prediction accuracy,and fewer training sets.In this method,only a small number of sampling points on the rough surface should be calculated through numerical methods as the data sets for training in the neural network for training.Then,the trained neural network can be used to rapidly predict the required numerical Green’s functions,and hence greatly reduce the computational workload.First of all,this thesis studies the fully connected neural network to solve the composite scattering of a dielectric target and a dielectric rough surface.First,part of the numerical Green’s function values are calculated by the method of moments in the offline process and sent to the neural network as the sampling points for training.Consequently,numerical Green’s function of any the field-source coordinates can be predicted online by the trained fully connected neural network.Then the induced current generated by the target is calculated by the electric field integral equation,and the scattering field generated by the target is obtained by convoluting the numerical Green function with the induced current.Then,this thesis studies the solution of composite scattering by a dielectric target and a conducting rough surfaces using fully connected neural networks.Unlike the case the of dielectric targets and metal rough surfaces,when establishing integral equations,it is no longer possible to use volume integral equations for both target and rough surface.Instead,surface integral equations should be applied to the conducting rough surface to generate the impedance matrix and volume integral equations is used to yield the impedance matrix for the dielectric targets.Numerical results have shown the effectiveness of this the method.Finally,this thesis studies the generalized regression neural network to solve the composite scattering of a metallic target and a conducting rough surface.First,the method of moments is invoked repeatedly to obtain the statistical mean value of the half-space Green’s function of the region above the rough surface,and the half-space Green’s function value is predicted by the trained generalized regression neural network.In addition,when solving the composite scattering of rough surface and metallic target with arbitrary shape,the two-dimensional bilinear interpolation technique is introduced to overcome the difficulty that the location of the discretized target surface is different from the sampling Green’s function points.Consequently,the composite scattering of rough surface and target can be solved with high accuracy.It is shown from numerical results that compared with the fully connected neural network,the generalized regression neural network is superior in prediction speed and nonlinear fitting ability.