基于CNN-GAN组合模型的可控源电磁反演方法  

A Controllable Source Electromagnetic Inversion Method Based on CNN-GAN Combinatorial Model

作　　者：马琰祁 李伟勤[1] 吴育涵 钟连诚 MA Yanqi;LI Weiqin;WU Yuhan;ZHONG Liancheng(School of Electrical Engineering and Information,Southwest Petroleum University,8 Xindu Road,Chengdu 610500,China)

机构地区：[1]西南石油大学电气信息学院,成都市610500

出　　处：《大地测量与地球动力学》2023年第10期1095-1100,共6页Journal of Geodesy and Geodynamics

基　　金：四川省科技计划(2020YFG0182);四川省安全生产监督管理局项目(四川-0004-2016AQ);中国石油勘探开发研究院科技项目(RIPED.CN-2019-CL-53)。

摘　　要：提出一种基于卷积神经网络(CNN)和生成对抗网络(GAN)组合模型的可控源电磁反演方法。为突出异常体信息,将总场进行差分预处理,同时修改损失函数以增强GAN的稳定性。将差分总场送入CNN,得到地面接收电场与地下异常体电导率数据之间的结构化因果关系;输出异常体的粗略电导率模型,再将其作为GAN的输入,在GAN中提取特征进行训练,得到高精准和高对比度的电导率反演结果,以达到工程应用的要求。通过将CNN-GAN组合模型与传统神经网络模型的预测结果进行对比发现,CNN-GAN组合模型优于传统神经网络模型,能够更好地预测地下异常体的电导率模型,相似度高达94.38%,比传统CNN和GAN模型分别提高约48%和78%。We propose a controllable source electromagnetic inversion method based on convolution neural network(CNN)and generated countermeasure network(GAN).In order to highlight the information of the abnormal body,we preprocess the total field by difference,and modify the loss function to enhance the stability of the GAN.By sending the difference total field into CNN,we obtain the structural causal relationship between the ground receiving electric field and the conductivity data of the underground abnormal body;the rough conductivity model of the abnormal body is output.Then as the input of GAN,the features are extracted in GAN for training,and we obtain the conductivity inversion results with high precision and high contrast,which meets the requirements of engineering application.In comparison,the CNN-GAN combined model is better than the traditional neural network model,and can better predict the electrical conductivity model of underground abnormal bodies;the similarity is as high as 94.38%,which is about 48%and 78%higher than the traditional CNN and GAN model,respectively.

关 键 词：卷积神经网络 生成对抗网络 差分总场 可控源电磁法 网络结构 参数评价 

分 类 号：P318[天文地球—固体地球物理学]