基于深度神经网络的重力异常反演

Inversion of gravity anomalies based on a deep neural network

为解决传统线性反演方法容易陷入局部极小,计算效率低等问题,本文提出了一种基于深度学习的重力异常反演方法。该方法首先构造不同形状的二维密度模型,正演得到重力异常,组成数据集;然后用该数据集训练深度神经网络;最后将重力异常数据输入到训练好的深度神经网络,直接得到反演结果。实验结果表明,该方法能快速、准确地反演出地下异常体的位置和形态,且具有较好的泛化能力和抗噪声能力,可用于重力异常反演。

Traditional linear inversion of gravity anomalies is liable to encounter local minima and suffer low computational efficiency.Given this,this paper proposed a deep learning-based inversion of gravity anomalies.Specifically,two-dimensional density models of various shapes were firstly established,and gravity anomalies were obtained through forward simulation using these models to form a dataset.Then,a deep neural network was trained using the dataset.Finally,gravity anomaly data were input into the deep neural network to directly yield inversion results.Experimental results show that the inversion method proposed in this study can determine the locations and shapes of underground anomalies quickly and accurately,with high generalization ability and anti-noise ability.Therefore,this method can be used for the inversion of gravity anomalies.