基于深度卷积神经网络的地震数据断层识别方法

Seismic fault interpretation based on deep convolutional neural networks

对地震数据进行断层解释一直是油气勘探开发过程中的一项重点工作。传统的断层解释主要是以人机交互方式进行的,效率低,并且人为因素可能增大断层解释结果的不确定性;而常规的断层识别方法则通常需要设置多个控制参数,导致断层识别的结果严重依赖参数设置的准确性。为此,提出一种基于卷积深度神经网络的地震数据断层识别方法,该方法利用ResNet深度残差网络可有效训练深层卷积神经网络和U-Net架构可表征多尺度、多层次特征信息的优势,将ResNet和U-Net架构联合,构建了用于地震数据断层识别的网络架构(SeisFault-Net)。其中,U-Net架构由编码和解码两个子网络组成,使SeisFault-Net以端到端的方式进行模型训练;同时,利用残差神经网络克服深层网络梯度弥散的问题,有效提高SeisFault-Net的训练效率。训练后的SeisFault-Net无需设置任何参数即可对地震数据进行断层识别,避免了常规方法中人为设置参数的经验误差和不确定性。数据实验表明,提出的SeisFault-Net方法可准确地识别断层位置,且识别的断层垂向连续性好,断层轮廓清晰。与相干算法相比,SeisFault-Net方法识别的断层细节更丰富,断层解释更准确;同时,可有效提高地震断层识别的效率。

Seismic fault interpretation has always been a key task in the process of oil and gas exploration and development.Conventional fault interpretation is mainly based on human-computer interaction,which is of low efficiency and causes the results with many uncertainties.In addition,conventional methods for fault interpretation usua-lly set multiple parameters,whose controls accuracy of the predicted faults.This paper proposes a method using seismic data based on convolutional deep neural networks.Taking the advantages of ResNet for effectively training deep convolutional neural network and U-Net architecture for characterizing multi-scale and multi-layer characteristic information,this method combines deep residual neural network and U-Net architecture to construct a network architecture(SeisFault-Net)for fault interpretation based on seismic data.The U-Net architecture consists of an encoding sub-network and a decoding sub-network.They enable the SeisFaultNet to train models in an end-to-end manner.The residual neural network can suppress the gradient dispersion of deep network,and effectively improve the training efficiency of the SeisFault-Net.After trained,the SeisFault-Net can perform fault interpretation based on seismic data without setting any parameters.This avoids the empirical error and uncertainties caused by parameters artificially set in conventional methods.Applications to raw data have proved that the SeisFault-Net method can effectively and accurately detect fault locations,and the faults have good vertical continuity and clear outlines.The detailed information of faults interpretated by the SeisFault-Net method is more abundant and accurate than the coherent algorithm.And the calculating efficiency of the SeisFault-Net method is very high in seismic fault interpretation.