基于K-medoids聚类和正则化神经网络的电阻率反演成像
首发时间:2016-05-17
摘要:针对神经网络反演技术在电阻率成像工程勘探应用推广中存在的问题,提出了一种基于正则化神经网络的剪枝贝叶斯神经网络(PBNN)非线性反演算法和一种基于K-medoids聚类的样本构造方法。在基于K-medoids聚类的样本构造方法中,利用观测数据的聚类结果提供先验信息构造神经网络的训练样本,从而有针对性的指导神经网络的训练过程;剪枝贝叶斯神经网络在引入贝叶斯正则化的基础上,通过评估各隐节点对反演结果的影响来自适应确定神经网络的隐层结构,根据小样本条件下本文训练样本的分布特征,选择了基于广义平均的超参数矢量来引导剪枝过程。通过与地球物理领域内其它常用的自适应正则化方法相比较,验证了本文算法的有效性。模型仿真和实测数据反演的结果表明:该方法能够较好的抑制神经网络训练过程中噪声的影响,提高网络的泛化能力,其反演结果优于不含正则化参数的BPNN反演和RBFNN反演以及传统的最小二乘反演。
关键词: 电阻率成像 贝叶斯神经网络 正则化 非线性反演 K-medoids聚类
For information in English, please click here
Electrical Resistivity Imaging Inversion Based on K-medoids Clustering and Regularization Neural Networks
Abstract:To solve the application problems of Electrical Resistivity Imaging (ERI) neural network inversion technique in engineering exploration, a Pruning Bayesian Neural Network (PBNN) nonlinear inversion algorithm based on Regularization Neural Network (RNN) and a sample construction method based on K-medoids clustering algorithm are proposed in this paper. The training samples of neural network are constructed with prior information provided by the clustering results through the proposed sample construction method, thus the training process of neural network can be guided specifically. The Pruning Bayesian Neural Network, based on Bayesian regularization, selects hidden layer structure adaptively by assessing the influence of each hidden nodes to inversion results. Then hyper-parameter vectors based on generalized mean are chosen to guide pruning process according to the distribution of small training samples. The proposed algorithm is proved effectively in comparison to the other common adaptive regulation methods in geophysics. It is demonstrated by results of model simulation and filed data inversion that the proposed method is able to suppress noise effectively in neural network training stage and to increase the network generalization ability. The inversion results of the proposed method are better than those of Back Propagation Neutral Network (BPNN) and Radial Basis Function Neural Network (RBFNN) inversion without regulation parameters as well as the traditional least square inversion.
Keywords: Electrical resistivity imaging Bayesian neural network Regularization Nonlinear inversion K-medoids clustering
论文图表:
引用
No.4687360764569146****
同行评议
共计0人参与
勘误表
基于K-medoids聚类和正则化神经网络的电阻率反演成像
评论
全部评论0/1000