基于神经网络电性源半航空视电阻率反演研究
Apparent resistivity inversion of electrical source semi-airborne electromagnetic data based on neural network
-
摘要:
为解决目前电性源时域半航空视电阻率反演效率低、适用性差的问题,在快速计算电性源时域半航空感应电动势基础上,将理论计算的感应电动势进行等效变换,建立感应电动势与电阻率单一映射关系的样本集,选用三层误差反向算法（Back Propagation, BP）神经网络和Levenberg-Marquardt算法对样本集进行训练,优选出收敛速度快、误差小的网络参数．基于层状模型,对比了神经网络反演方法与传统方法反演效果,对于均匀半空间模型,神经网络反演结果相对误差在5％以下,传统反演结果相对误差大于20%;对于层状大地模型,结果较传统方法更接近真实模型．基于神经网络的反演方法明显优于传统方法,这为半航空电磁数据提供了新的处理与解释方法．
Abstract:The efficiency and applicability of time-domain electrical source semi airborne apparent resistivity inversion is poor. Due to that, we conduct an equivalent conversion of the induced voltage based on fast forward modeling and build single sample set.Then, we choose three-layer BP neural network and Levenberg-Marquardt algorithm for training sample set, optimizing the network parameter with convergence speed and small error. Comparing the inversion results of neural network inversion method with the traditional method,for homogeneous half space model, the relative error of neural network inversion results are less than 5％, while the traditional inversion results is more than 20％; for layered model, the result is closer to the real model than the traditional method. It proves that inversion method based on neural network is superior to traditional method, provides a new method for processing and interpretation of semi airborne electromagnetic data.
下载: