基于深度学习的中心回线瞬变电磁全区视电阻率计算  

The calculation of full-region apparent resistivity of central loop TEM based on deep learning

作　　者：吴国培 张莹莹 张博文 赵华亮 WU Guo-Pei;ZHANG Ying-Ying;ZHANG Bo-Wen;ZHAO Hua-Liang(School of Geology and Mining Engineering,Xinjiang University,Urumqi 830047,China)

机构地区：[1]新疆大学地质与矿业工程学院,新疆乌鲁木齐830047

出　　处：《物探与化探》2021年第3期750-757,共8页Geophysical and Geochemical Exploration

基　　金：新疆维吾尔自治区自然科学基金项目(2017D01C064);新疆维吾尔自治区研究生科研创新项目(XJ2020G044)。

摘　　要：深度学习是人工神经网络算法的扩展,对复杂函数有很好的逼近能力,本文将其引入用于瞬变电磁视电阻率计算。首先,建立归一化感应电动势与瞬变场参数单一映射关系的5层深度神经网络,通过对单一隐含层不同神经元个数所训练的误差情况进行分析,确定5层深度神经网络各隐含层神经元个数为13,8,5,8,13。训练算法选择了改进的具有自适应学习率的Nadam算法,该算法可加速训练过程。对训练好的深度神经网络模型进行仿真实验,采用典型地电模型加以验证,发现其对不同的地电模型均具有较好的反映,证明本文采用的基于深度学习计算视电阻率的可行性。应用结果表明训练好的深度神经网络模型可快速准确计算视电阻率。Deep learning is an extension of the artificial neural network algorithm, which has a good approximation ability for complex functions. This paper introduces this means for the calculation of transient electromagnetic apparent resistivity. First, a 5-layer deep neural network is established with a single mapping relationship between the normalized induced electromotive force and the transient field parameters. By analyzing the error conditions trained by different numbers of neurons in a single hidden layer, the hidden layers of the 5-layer deep neural network are determined. The number of layered neurons is 13,8,5,8,13. The training algorithm chooses the improved Nadam algorithm with adaptive learning rate, which can speed up the training process. The trained deep neural network model is simulated and verified by a typical electrical model, and it is found that it has a good response to different geoelectric models, which proves the feasibility of calculating apparent resistivity based on deep learning put forward in this paper. The actual application results show that the trained deep neural network model can quickly and accurately calculate the apparent resistivity, and its effectiveness is verified by drilling.

关 键 词：瞬变电磁法 深度学习 人工神经网络 视电阻率 

分 类 号：P631[天文地球—地质矿产勘探]