基于深度学习的地震随机噪声衰减方法研究进展  

Research progress of random noise attenuation methods for seismic data based on deep learning

作　　者：崔扬 王燕楠 陈万利 张虹 朱丹丹 白敏 CUI Yang;WANG Yannan;CHEN Wanli;ZHANG Hong;ZHU Dandan;BAI Min(Key Laboratory of Exploration Technologies for Oil and Gas Resources(Yangtze University),Ministry of Education,Wuhan,Hubei 430100,China;School of Geophysics and Petroleum Resources,Yangtze University,Wuhan,Hubei 430100,China)

机构地区：[1]长江大学油气资源与勘探技术教育部重点实验室,湖北武汉430100 [2]长江大学地球物理与石油资源学院,湖北武汉430100

出　　处：《石油地球物理勘探》2023年第5期1269-1283,共15页Oil Geophysical Prospecting

基　　金：国家自然科学基金项目“基于时间域高斯束变换的多震源数据高精度分离与高效偏移方法研究”(42174159)和“基于字典学习的多震源数据高效高精度最小二乘偏移方法研究”(41904110);湖北省自然科学基金项目“基于时间域高斯束变换的多震源数据分离、去噪和插值重建方法研究”(2021CFB498);油气资源与勘探技术教育部重点实验室青年创新团队项目“智能驱动的地震资料高分辨率处理方法”(KPI2021-01);长江大学大学生创新创业训练计划项目“基于深度学习的地震资料去噪方法(Yz2020043)”联合资助

摘　　要：分析了卷积神经网络(CNN)、去噪卷积神经网络(DnCNN)、U⁃Net深度神经网络、前反馈(BP)神经网络、空洞卷积神经网络(DCNN)、残差网络(ResNet)、迁移学习等为代表的深度学习方法的概念、发展现状、方法原理、去噪效果以及优缺点等;对比了传统去噪方法、字典学习及深度学习方法的去噪效果;展望了深度学习技术在地震去噪领域的发展前景。获得以下认识:①深度学习方法的实际去噪效果优于传统方法和字典学习方法,不需要设定结构模型,泛化性更强,且计算时间短、精度更高。②深度学习方法存在诸多不足:实际数据的去噪效果往往差于合成数据;普适性不强;神经网络的“黑匣子”特性使其物理可解释性大大降低;网络性能与训练数据的泛化性密切相关;用于训练网络的数据集因人而异,难以评价网络性能。③期待深度学习在以下方面取得进展和突破:搭建适用于不同噪声的去噪神经网络结构,并将更优的网络结构引入地震随机噪声压制;将地震信号转换到变换域构造网络的损失函数;改进学习策略的同时制作更具代表性的数据集,尽可能地使训练数据覆盖所有类型的解,提高网络泛化性;自动化的参数调优;结合模型驱动与数据驱动的方法。This paper analyzes the concept,development status,method principle,denoising performance,and advantages and disadvantages of deep learning methods represented by a convolutional neural network(CNN),denoised convolutional neural network(DnCNN),U⁃net deep neural network,forward feedback(BP)neural network,Dilated convolutional neural network(DCNN),residual network(ResNet),and transfer learning.The denoising effects of traditional denoising methods,dictionary learning,and deep learning methods are compared,and the development prospect of deep learning technology in the field of seismic denoising is forecasted.The following conclusions are obtained:①The actual denoising effect of the deep learning method is better than that of traditional methods and dictionary learning methods.It does not need to set the structural model and has stronger generalization,shorter computation time,and higher precision.②There are many shortcomings in deep learning methods:The denoising effect of actual data is often worse than that of synthetic data;the universality is not strong;the“black box”characteristic of the neural network makes its physical interpretability greatly reduced.Network performance is closely related to the generalization of training data.The data sets used to train the network vary from person to person,making it difficult to evaluate network performance.③It is expected that deep learning will make progress and breakthroughs in the following aspects:building a denoised neural network structure suitable for different noises and introducing a bet⁃ter network structure to suppress seismic random noise,constructing the loss function of the network by converting the seismic signal to the transform domain,improving the learning strategy and making a more representative data set,making training data cover all solutions as much as possible,enhancing network generalization,and achieving au⁃tomatic parameter tuning and methods combining model⁃driven and data⁃driven features.

关 键 词：模型驱动 数据驱动 深度学习 字典学习 随机噪声衰减 卷积神经网络 研究进展 

分 类 号：P631[天文地球—地质矿产勘探]