论文部分内容阅读
针对传统神经网络在电阻率成像反演中存在的过拟合和易陷入局部极值等问题,提出了一种基于剪枝贝叶斯神经网络(PBNN)的非线性反演算法和一种基于K-medoids聚类的样本设计方法。在基于K-medoids聚类的样本设计方法中,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,从而有针对性地指导神经网络的训练过程;剪枝贝叶斯神经网络是在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,根据小样本条件下训练样本的先验分布特征,选择了基于广义平均的超参数αk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。理论数据和实测数据反演的结果表明:该方法能够较好地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BPNN反演、RBFNN反演和RRBFNN反演以及传统的最小二乘反演。
Aiming at the problems of over fitting and easily falling into local extremum of traditional neural network in resistivity imaging inversion, a non-linear inversion algorithm based on pruned Bayesian neural network (PBNN) K-medoids cluster sample design method. In the sample design method based on K-medoids clustering, the training samples of neural network are constructed by using the a priori information from the clustering results of observation data so as to guide the training process of the neural network in a targeted manner. The pruned Bayesian neural network Based on Bayesian regularization, the hidden layer structure of neural network is adaptively identified by evaluating the influence of hidden nodes on the inversion results. Based on the prior distribution of training samples under small sample conditions, Hyper-parameter αk to guide the pruning process. Compared with other commonly used adaptive regularization methods in the field of geophysics, the effectiveness of the proposed algorithm is verified. The results of theoretical data and measured data show that this method can restrain the influence of noise in neural network training and improve the generalization ability of network, and its inversion result is better than BPNN inversion, RBFNN inversion and RRBFNN inverse Acting as well as the traditional least square inversion.