论文详情
随机噪声会干扰地震数据中的有效信号并降低数据的信噪比, 进而影响地震数据的后续处理。常规基于监督学习的深度学习噪声衰减方法需要大量的标签来训练网络, 但是, 在真实地震数据中制作无噪声的标签用于训练深度神经网络是非常具有挑战性的工作。因此, 提出端到端的无监督学习框架来衰减随机噪声, 并提取多维地震资料中的有效信号信息。首先, 建立由全连接模块、编码器模块和解码器模块组成的深度神经网络框架, 并在编码器和解码器之间添加类似残差结构的跳跃链接以提高去噪表现。为了提高网络的去噪表现, 使用适用于地震资料的数据增强方法, 将输入的多维大尺度含噪地震数据分割为大量的小尺度一维数据进行迭代。对地震数据进行数据增强时, 选择合适的切分和滑动尺寸将提高网络的计算效率和去噪效果。合成数据和渤海油田实际数据的应用结果表明, 相较于传统地震噪声衰减方法, 本文提出的方法具有更好的随机噪声衰减能力和有效信号提取能力。
Random noises contaminate seismic signals and reduce the signal-to-noise ratio of seismic data, which will affect subsequent seismic data processing. A denoising method based on supervised deep learning usually requires a large number of labels to train the network, but it is very challenging to make noise-free labels using observed seismic data. To attenuate random noises and extract useful signals from multi-dimensional seismic data, we propose an end-to-end neural network based on unsupervised deep learning, which consists of a fully connected module, an encoder module and a decoder module. A skip connection similar to a residual structure is added between the encoder and decoder to improve the performance of denoising. To strengthen the network further, a data enhancement method is used to segment large-scale multi-dimensional noisy seismic data into a large number of small-scale one-dimensional data for iteration. Appropriate slicing and sliding sizes for data enhancement could improve the calculation efficiency and denoising effect of the network. The application to synthetic data and actual data acquired in Bohai oilfield shows that the proposed method is better than traditional denoising techniques in random noise attenuation and signal extraction.