摘要
人工神经网络是通过从大量训练数据中学习来拟合复杂非线性函数的有效方法,属于一种数据驱动的机器学习方法.人工神经网络应用于地震反演时可以得到更高分辨率和精度的结果,有着优于传统反演方法的泛化能力和非线性拟合能力.本文对人工神经网络的发展脉络进行了回顾,梳理了基于梯度的学习过程中代价函数的作用,反向传播学习算法的思路,激活函数的不同类型,以及万能近似定理等.特别是对热门的深度神经网络,按照时间先后顺序总结了带卷积核的LeNet-5、AlexNet、VGGNet、GoogLeNet、ResNet、UNet、自编码器和GANs等经典模型.在此基础上,本文分析了深度神经网络在反射系数和子波反演、速度反演、波阻抗反演和地震结构反演中不同网络的拓扑结构、学习算法、激活函数和训练样本等.最后,本文归纳和讨论了用于地震反演的有监督端到端学习网络的流程和关键影响因素等,展望了融入物理规律、基于反演目标函数展开的专用地震反演网络.
Artificial Neural Network(ANN)is an effective method to fit complex nonlinear functions by learning from a large amount of training data,which is a data-driven machine learning method.When it is applied to seismic inversion problems,the resolution and accuracy obtained are higher.Artificial neural network has generalization ability and nonlinear fitting ability superior to conventional seismic inversion methods.In this paper,the development of artificial neural networks are reviewed.The role of cost function in the gradient-based learning process,the idea of back propagation learning algorithm,the different types of activation functions,and the universal approximation theorem are combed.Especially for popular deep neural networks,the classic models with convolution kernels such as LeNet-5,AlexNet,VGGNet,GoogLeNet,ResNet,UNet,autoencoder and GANs are summarized in chronological order.On this basis,this paper analyzes the topology,learning algorithm,activation function and training samples of different networks in seismic wavelet and reflectivity inversion,velocity inversion,wave impedance inversion and seismic structure inversion of deep neural network.Finally,this paper summarizes and discusses the process and key influencing factors of supervised end-to-end learning networks for seismic inversion,and looks forward to a dedicated seismic inversion network that incorporates physical laws and unfolds based on the inversion objective function.
作者
王竟仪
王治国
陈宇民
高静怀
WANG JingYi;WANG ZhiGuo;CHEN YuMin;GAO JingHuai(School of Mathematics and Statistics,Xi'an Jiaotong University,Xi'an 710049,China;School of Information and Communication Engineering,Xi'an Jiaotong University,Xi'an 710049,China)
出处
《地球物理学进展》
CSCD
北大核心
2023年第1期298-320,共23页
Progress in Geophysics
基金
国家自然科学基金面上项目(41974137)资助。
关键词
人工神经网络
地震反演
代价函数
反向传播算法
激活函数
Artificial Neural Network(ANN)
Seismic inversion
Cost function
Back propagation algorithm
Activation function
