基于卷积自编码器的地震数据处理  

Seismic data processing based on convolutional autoencoder

作　　者：江金生 任浩然 李瀚野 JIANG Jin-sheng;REN Hao-ran;LI Han-ye(Key Laboratory of Geoscience Big Data and Deep Resource of Zhejiang Province,School of Earth Sciences,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]浙江大学地球科学学院,浙江省地学大数据与地球深部资源重点实验室,浙江杭州310027

出　　处：《浙江大学学报（工学版）》2020年第5期978-984,共7页Journal of Zhejiang University：Engineering Science

基　　金：国家自然科学基金资助项目(41674123);浙江省自然科学基金资助项目(LY19D040002);中石化地球物理重点实验室开放基金资助项目;多方法大数据智能反演技术及软件研发基金资助项目(2018YFC0603604).

摘　　要：引入以深度学习为代表的数据驱动方法,加速地震数据的处理流程,获得更精确的地下介质信息.卷积自编码器方法在地震数据压缩降维的同时,利用数据的空间局部相关性自动提取信号特征,避免数学物理模型的假设依赖.通过设计不同地质模型的地下速度结构,利用波动方程正演模拟构建大量不同特征的地震数据训练集和测试集.与模型驱动的地震随机噪声压制和地震道插值方法不同,数据驱动下的卷积自编码器方法能够从含随机噪声地震数据和地震道缺失数据中,直接识别和提取出其中的有效地震信号,从而压制随机噪声以及重建原始地震数据,实验结果验证了该方法的有效性.卷积自编码器方法不需要人工阈值控制,具有更高的处理效率.Data-driven methods represented by deep learning were introduced into seismic data processing,which can speed up the processing process of seismic data and obtain more accurate information about underground media.The convolutional autoencoder(CAE)method can extract data features by using the spatial correlation of seismic data while compressing it,which avoids the assumption dependence of mathematical and physical model.The underground velocity structures of different geological models were designed,and the training and test sets with different features were obtained through wave equation forward modeling.CAE method is different from the traditional model-driven method of seismic random noise attenuation and seismic data interpolation.CAE method can identify and extract the valid seismic signals directly from seismic data with random noise and missing traces,to attenuate random noise and reconstruct the original seismic data.Experimental results verified the effectiveness of this method.CAE method gets rid of the artificial threshold,therefore has higher processing efficiency.

关 键 词：卷积自编码器 地震道插值 地震随机噪声压制 深度神经网络 稀疏表达 

分 类 号：P3[天文地球—地球物理学]