基于卷积神经网络的瞬变电磁数据快速成像算法研究

Fast imaging of transient electromagnetic data based on convolutional neural network

瞬变电磁一维反演技术已经相对成熟,但对野外实测数据进行实时成像仍然有一定难度.本文将卷积神经网络引入到时间域瞬变电磁数据成像中,通过训练网络近似瞬变电磁响应与地电模型之间的函数关系,将复杂的反演过程转化为矩阵映射过程,进而实现瞬变电磁数据的实时快速成像.考虑到传统算法大多针对单点进行成像,难以实现面积性数据快速处理,本文尝试将接收点相对发射源的坐标信息作为网络输入参数,这不仅使得该算法在野外成像过程中更加灵活,同时也大大减少了训练过程中样本集数量.为验证算法的有效性,我们首先在大量理论数据上进行测试,检验了卷积网络的优越性以及加入坐标信息可在不影响成像精度的前提下极大提升该方法对不规则测点电磁数据的成像灵活性.最后,我们通过对实测数据分别进行神经网络成像和Occam反演,进一步验证本文神经网络成像方法的有效性.

The one-dimensional(1D)inversion of Transient Electromagnetic(TEM)data has been well developed,however,it is still difficult to conduct real-time imaging of field survey data.In this paper,we introduce a Convolutional Neural Network(CNN)into TEM data imaging.By training the network to approximate the relationship between the time-domain EM responses and geoelectrical model parameters,the complex inversion process is transformed into a matrix mapping process,realizing the real-time and fast imaging of TEM data.Considering that the traditional imaging algorithms are mostly executed on a point-by-point basis,it is difficult to achieve rapid processing of large survey dataset,we try in this paper to take the coordinate information of the receivers relative to the transmitting source as the network input parameter,which not only makes the algorithm more flexible in the imaging of survey data,but also greatly reduces the number of samplings in the training process.To verify the effectiveness,we first test our algorithm on synthetic data and demonstrate that by adding coordinate information into the network our CNN method has the flexibility of imaging irregular survey points without affecting the imaging accuracy.Finally,we compare our imaging results with the Occam's inversion for a survey dataset to further verify the effectiveness of our imaging method.