利用高斯径向基函数的拟神经网络重力反演方法

Gravity inversion method based on quasi-neural network featuring Gaussian radial basis function

为了提高重力反演分辨率,提出一种利用高斯径向基函数的拟神经网络反演方法。该方法利用高斯径向基函数压缩模型空间,在保证复杂模型表征能力的前提下,实现了反演参数的降维;以高斯径向基函数为激活函数,构建一种拟神经网络结构,不需要训练,可以克服建立训练数据集的困难。该方法较好地解决了重力反演不适定性所导致的趋肤、垂向分辨率低、多解性强和严重依赖先验约束等问题,并从重力数据中最大程度地提取有效信息以提高反演结果的分辨率,增强可靠性。模型实验证明了该方法具有较高的精度和分辨率,能较准确地反演模型的位置、边界和密度。应用该方法反演车镇凹陷重力数据,得到了垂向分辨率较高的剩余密度模型,从中提取密度界面和剖面开展构造解释,揭示了下古生界构造格局和潜山发育规律,证明了该方法的实用价值和应用潜力。

An inversion method based on a quasi-neural network featuring Gaussian radial basis function(RBF)is presented in this paper to improve the resolution of gravity inversion.The model space is compressed through the Gaussian RBF,and the dimension of the inversion parameters is reduced without influencing the representation ability of complex models.A quasi-neural network structure is proposed which takes the Gaussian RBF as the activation function and saves the difficulty of establishing a training set in that it does not require training.The proposed method solves the problems of skinning,low vertical resolution,strong multi-solution,and severe dependence on priori constraints caused by the ill-posedness of gravity inversion.In addition,it can extract effective information from gravitational data to enhance the resolution and reliability of the inversion results.Model experiments show that the method,with high accuracy and resolution,can accurately obtain the position,boundary,and density of the model through inversion.A residual density model with a high vertical resolution is obtained by inverting the gravitational data of the Chezhen Depression.Density interfaces and profiles are extracted from the density model for structural interpretation,which in turn reveals the structural pattern of Lower Paleozoic and the development law of buried hills and proves the practical value and application potential of this method.