论文部分内容阅读
在介绍阵列侧向电极系工作原理及快速正演计算的基础上,通过计算分析了Marquardt法线性反演过于依赖初值及迭代反演速度慢的不足。为避免迭代反演的缺陷,研究了采用BP神经网络进行测井快速反演的方法。通过对测井响应进行井眼校正等预处理,可有效减少网络训练所需样本数,提高反演效率。通过对比不同网络结构的计算精度与计算效率,选取了22个隐含层节点的3层网络,反演结果与真值吻合较好。该反演方法不需给定初值,不需迭代计算,反演100组参数用一般计算机仅耗时3s。
Based on the working principle and fast forward calculation of the array lateral electrode system, the Marquardt method’s linear dependence on initial value and slow iterative inversion are analyzed. In order to avoid the shortcoming of iterative inversion, the method of rapid inversion of logging using BP neural network was studied. By pre-processing logging response such as borehole correction, the number of samples required for network training can be effectively reduced and the inversion efficiency can be improved. By comparing the calculation accuracy and computational efficiency of different network structures, we select 22 layers of hidden layer nodes, and the inversion results agree well with the true values. The inversion method does not need to be given initial value, without iterative calculation, inversion of 100 parameters using the general computer only takes 3s.