基于改进U-Net++网络的地震数据随机噪声压制方法  

Random noise suppression of seismic data based on improved U-Net++network

作　　者：邓聪 李明桐 张文祥 杨文博 彭海龙 DENG Cong;LI Mingtong;ZHANG Wenxiang;YANG Wenbo;PENG Hailong(Zhanjiang Branch,CNOOC China Limited,Zhanjiang,Guangdong 524057,China)

机构地区：[1]中海石油(中国)有限公司湛江分公司,广东湛江524057

出　　处：《世界石油工业》2025年第2期69-80,共12页World Petroleum Industry

基　　金：中国海洋石油集团有限公司科技项目(KJZH-2024-2104)。

摘　　要：为了解决常规的U-Net++网络模型在开展地震资料随机噪声压制时容易出现地质体边缘信息模糊和构造细节信息丢失的问题,采用基于改进U-Net++网络的地震数据随机噪声压制方法,即在U-Net++网络结构中加入构造导向滤波模块和多尺度密集型残差模块用于提取不同尺度下地震数据体的构造特征信息,并实现不同层次空间的构造特征融合;对模型中的损失函数进行改进,在原有L2损失函数的基础上增加多尺度结构相似性损失函数和边缘保持损失函数,通过差异性权重因子控制不同损失函数的影响,有效提升网络模型的压制噪声能力和边缘细节信息保护能力。在网络模型训练中,通过消融实验确定不同损失函数的权重因子,获取具有地区差异性的损失函数构成因子。模型测试及实际资料应用结果表明,改进的U-Net++网络模型具有显著的噪声压制效果,各项视觉指标均优于常规去噪方法,地震数据中的地质体信息清晰度最优。结论认为,该方法能够在保护地质体边缘有效信息和细节信息的基础上提高数据的信噪比,有利于后续开展的地震资料解释工作,在地震数据处理工作中具有推广意义。The conventional U-Net++network model is prone to problems such as blurred geological body edge information and loss of structural details when dealing with seismic random noise.Therefore,a seismic data random noise suppression method based on improved N-Net++network is proposed.This method incorporates a construction oriented filtering module and a multi-scale dense residual module into the U-Net++network structure to extract the construction feature information of seismic data volumes at different scales,and achieve the fusion of construction feature in different hierachical spaces;At the same time,the loss function in the model was improved by adding multi-scale structural similarity loss function and edge preservation loss function on the basis of the original L2 loss function,The influence of different loss function was controlled by different weight factors,effectively enhancing the network model’s ability to suppress noise and protect edge detail information.In the training of network models,the weight factors of different loss functions are determined through ablation experiments to obtain loss function composition factors with regional differences.The model test and practical data application show that the improved U-Net++network model has significant noise supression effect,and all visual indicators are superior to conventional denoising methods.The geological information clarity in seismic data is the best.This method can improve the signal-to-noise ratio of data while protecting the effective and detailed information of geological body edges,which is beneficial for seismic interpretation and has good promotion significance in seismic data processing.

关 键 词：随机噪声 地震边缘信息 U-Net++网络模型 信噪比 损失函数 

分 类 号：TE12[石油与天然气工程—油气勘探] P631.4[天文地球—地质矿产勘探]