点源反演问题的神经网络方法研究

Research on Neural Network Methods for Inverse Point Source Problems

针对通过相关的远场测量重构点源位置的声波反源问题,构建基于神经网络和门控思想的点源位置参数反演模型。首先,以单频远场数据与点源位置参数分别作为输入和输出序列,通过门控思想和长期记忆函数,有选择性地更新网络状态并保存数据特征。其次,采用自学习方法更新模型的权重和偏置,进而反演点源的位置。最后,数值实验说明该方法可有效解决点源位置反演问题。

Aiming at the inverse point source problem of recovering unknown point source by the associated far-field measurements in sound field,a location parameter inversion model is constructed based on neural network with gating thought.Firstly,the single frequency far-field measurements and the location parameters of the point source are taken as the input and output sequences,respectively. The network state is selectively updated to preserve the specific structure of the underlying data through a gated idea and the use of the long-term memory function. Secondly,the weights and the offsets of the network are updated by self-learning algorithm so as to reconstruct the location of the point source. Finally,numerical experiments show that this method can effectively solve the inverse point source problem.