基于可伸缩型注意力机制的神经网络地震数据去噪方法  

A method for seismic data denoising based on the neural network with a retractable attention mechanism

作　　者：张敏[1,2] 许一卓 易继东 ZHANG Min;XU Yi-Zhuo;YI Ji-Dong(School of Geosciences,China University of Petroleum(East China),Qingdao 266580,China;Key Laboratory of Deep Oil and Gas,Qingdao 266580,China)

机构地区：[1]中国石油大学(华东)地球科学与技术学院,山东青岛266580 [2]深层油气重点实验室,山东青岛266580

出　　处：《物探与化探》2024年第4期1065-1075,共11页Geophysical and Geochemical Exploration

基　　金：国家自然科学基金项目(42074133);中石油重大科技合作项目(ZD2019-183-003)。

摘　　要：地震资料中的随机噪声会影响地震数据的质量,从而影响后续处理与解释的准确性。传统去噪方法受先验条件的约束,效率低下,神经网络具有强大的特征提取能力,能够弥补这些缺点。然而,由于传统神经网络卷积核的局限性,可能会导致全局信息丢失。为了克服这个缺点,本文在卷积神经网络(CNN)的基础上,添加了可伸缩型注意力机制。该机制在网络中同时呈现密集和稀疏两种类型的自注意力模块,这两种注意力模块交替使用可以显著增强神经网络的表现能力,扩大接受场。通过卷积层和注意力模块提取地震数据浅层特征和深层特征,结合CNN的局部建模能力和Transformer的全局建模能力,有利于提升网络的全局交互作用,增强其去除噪声和处理细节的能力。最后,合成和实际地震数据实验结果均表明,该方法相较于Unet和DnCNN,具有更好的噪声压制与保留地震数据有效信息的能力,可以大幅提高信噪比,为地震数据的处理和解释提供帮助。Random noise in seismic data impairs the quality of the data,thus affecting the accuracy of subsequent processing and interpretation.Conventional denoising methods,constrained by prior conditions,exhibit low efficiency.Neural networks possess a strong feature extraction ability,which can make up for these shortcomings.However,the limitations of convolution kernels in conventional neural networks may lead to the loss of global information.Hence,this study introduced a retractable attention mechanism to the convolutional neural network(CNN).This mechanism presents both dense and sparse self-attention modules in the CNN.The alternate use of the two self-attention modules can significantly enhance the performance of the CNN and expand the receptive field.The shallow and deep features of seismic data were extracted using the convolutional layer and self-attention modules.Combined with CNN's local modeling ability and Transformer's global modeling ability,they contributed to enhancing CNN's global interaction and ability to reduce noise and deal with details.As indicated by the experimental results of synthetic and field data,the method used in this study can more effectively suppress noise and retain effective information of seismic data compared to Unet and DnCNN,significantly improving the signal-to-noise ratio and thus assisting in the processing and interpretation of seismic data.

关 键 词：随机噪声 卷积神经网络 可伸缩型注意力机制 TRANSFORMER 

分 类 号：P631.4[天文地球—地质矿产勘探]