无相位远场数据反演散射障碍的神经网络方法

针对无相位信息反演障碍物位置及形状的问题, 提出一种两层门控循环单元(GRU)神经网络对门控循环单元神经网络的方法(MGNN), 并给出该方法的收敛性分析. 首先, 以无相位远场数据与障碍物边界曲线方程参数作为输入和输出, 通过GRU神经网络控制门思想与长期记忆功能, 有选择性地更新网络状态, 保存数据特征; 其次, 应用梯度下降算法更新模型权重和偏置, 解决了无相位信息的远场数据反演障碍物位置及形状的难题; 最后, 利用数值实验说明该方法的有效性.

Neural Network Method for Inverse Scattering Obstacles with Phaseless Far-Field Data

Aiming at the problem of  the position and shape of inverse obstacles with  phaseless far-field data, we proposed  a two-layer gated recurrent unit (GRU) neural network to gated recurrent unit neural network (MGNN) method, and gave the convergence analysis of the proposed  method. Firstly, using the phaseless far-field data and the obstacle boundary curve equation parameters as input and output, through GRU neural network control gate idea and long-term memory function, the network states were selectively updated and the  data characteristics were saved. Secondly, we applied the gradient descent algorithm to update the weights and bias of the model, and  solved the difficulty of  the position and shape of inverse obstacles with phaseless far-field data. Finally,  the effectiveness of the proposed  method was demonstrated by numerical experiments.