吉林大学学报(信息科学版) ›› 2022, Vol. 40 ›› Issue (2): 231-239.

• • 上一篇    下一篇

基于卷积神经网络的地震数据去噪方法

刘玉敏, 魏海军, 袁 硕, 安志伟   

  1. 东北石油大学 电气信息工程学院, 黑龙江 大庆 163318
  • 收稿日期:2021-09-15 出版日期:2022-06-11 发布日期:2022-06-12
  • 通讯作者: 魏海军(1995— ), 男, 黑龙江肇东人,东北石油大学硕士研究生, 主要从事深度学习应用于地震数据去噪研究, ( Tel) 86-18746434357 ( E-mail) 1660255916@qq. com.
  • 作者简介:刘玉敏(1978— ), 女, 辽宁昌图人, 东北石油大学副教授, 硕士生导师, 主要从事智能算法及其在地震数据处理与分析中的应用研究, (Tel)86-13936827553(E-mail)liuyumin330@ 163. com.
  • 基金资助:
    黑龙江省自然科学基金资助项目(TD2019D001)

Random Noise Removal of Seismic Data Based on Convolutioncycle Generation Adversarial Network

LIU Yumin, WEI Haijun, YUAN Shuo, AN Zhiwei   

  1. College of Electrical Information Engineering, Northeast Petroleum University, Daqing 163318, China
  • Received:2021-09-15 Online:2022-06-11 Published:2022-06-12

摘要: 由于地球物理勘探环境的复杂性, 采集的地震数据常被现场的随机噪声覆盖, 如何对地震数据去噪成为 关键问题, 为此将基于卷积循环生成对抗网络(CycleGAN: Cycle Generation Adversarial Network)引入地震数据 去噪中, 该方法的关键是构建良好的生成器与判别器的结构并且优化合适的网络参数。 该网络输入数据与 标签数据是双向生成, 形成环状网络结构, 故而能引入循环一致性损失使网络参数训练的精度更高; 将非局部 神经网络模块作为CycleGAN生成器的残差链接以提高特征提取能力。以实际地震数据进行实验, 通过对数据 可视化、峰值信噪比和均方误差的对比, 验证该方法的可实行性。结果表明, 与普通生成对抗网络(GAN: Generation Adversarial Network)、残差神经网络(ResNet: Residual Network) 去噪效果相比, 该方法在不同噪声 水平下去噪性能更好, 峰值信噪比较高, 数据成像更清晰, 对实际生产工作有一定的指导意义。

关键词: 非局部神经网络;  , 循环生成对抗网络,  , 地震数据;  , 随机噪声

Abstract: Due to the complexity of geophysical exploration environment, the seismic data collected are often covered by random noise, therefore a convolution based cyclic generation adversarial network (CycleGAN) is introduced into seismic data denoising. The key of the network is to build a good structure of generator and discriminator and optimize the appropriate network parameters. The input data and label data of the network are generated each other to form a ring network structure, so the loss of cyclic consistency can be introduced to improve the accuracy of network parameter training. The non local neural network module is used as the residual link of CycleGAN generator to improve the feature extraction ability. Through the comparison of data visualization, peak signal-to-noise ratio and mean square error, the feasibility of this method is verified. The results show that compared with the general gan network and the residual network denoising effect, this method has better noise performance, higher peak signal-to-noise ratio and can clearer data imaging under different noise levels, which has a certain guiding significance for the actual production work.

Key words: non local neural network;  , cycle generation adversarial network;  , seismic data;  , random noise

中图分类号: 

  • TP273