重力梯度张量的拟BP神经网络反演.pdfVIP

第 42 卷第 12 期 中南大学学报( 自然科学版) Vol.42 No.12 2011 年 12 月 Journal of Central South University (Science and Technology) Dec. 2011 重力梯度张量的拟BP 神经网络反演 郭文斌，朱自强，鲁光银 ( 中南大学 地球科学与信息物理学院，有色金属成矿预测教育部重点实验室，湖南 长沙，410083) 摘要：基于重力梯度张量是反映重力场空间变化率的参数，比传统的重力异常具有更高的分辨率和更丰富的信息， 将改进的 BP 神经网络算法应用于重力梯度张量的反演中并分析其反演效果。该算法是一种基于 RPROP 算法的 拟 BP 神经网络反演算法，采用三层神经网络结构，用隐层神经元表示物性单元的密度值，根据 RPROP 算法自 动修改各单元密度值，从而得出场源空间的密度分布。研究结果表明：采用这种算法对重力梯度张量进行反演计 算，收敛速度快，对初始模型依赖性小，可准确反映出异常体形态特征和密度特征。 关键词：重力梯度张量；拟 BP 神经网络；RPROP 算法；反演  中图分类号：P631 文献标志码：A 文章编号：1672−7207(2011)12−3797−07   Quasi-BP neural network inversion of gravity gradient tensor GUO Wen-bin, ZHU Zi-qiang, LU Guang-yin (Key Laboratory of Metallogenic Prediction of Nonferrous Metals, Ministry of Education, School of Info-Physics and Geomatics Engineering, Central South University, Changsha 410083, China) Abstract: Based on the fact that gravity gradient tensor is a parameter which can reflect the spatial variation of gravity field, and that it has a higher resolution compared to the traditional gravity anomaly, a method for interpretation of gravity gradient tensor was proposed. The method is a kind of quasi-BP neural network algorithm which is based on RPROP algorithm. A three-layer network and the hidden layer neurons denote physics value were used. The physics value was automatically modified according to RPROP algorithm, and the physical distribution of field source was gotten. The results show that the method has a fast convergence speed and little dependence on initial model u