摘要
随着深度神经网络技术的发展,卷积神经网络(CNN)被越来越多地应用于地震数据的噪声压制中。常规CNN方法一般是在时间域进行,为了提升CNN方法对地震噪声的压制效果,提出了基于连续小波变换(CWT)的CNN地震噪声压制方法。该方法首先将一维时间域信号通过CWT转换为二维时频域信号。然后,在利用CNN对时频谱进行噪声压制时,提出了两种策略:能量谱策略(策略Ⅰ)是将CWT计算的复数矩阵的振幅谱作为CNN的训练样本,保持相位谱不变;复矩阵策略(策略Ⅱ)是将复数矩阵的实部和虚部图谱作为CNN的不同通道分别进行处理。最后,对于CNN的输出结果,利用逆连续小波变换(ICWT)将二维复数矩阵还原成一维地震信号。为了定量地对比方法的效果,提出利用峰值信噪比(PSNR)、结构相似性(SSIM)和均方根误差(RMSE)3个指标对比基于两种CWT策略的CNN方法与其它常规滤波(包括低通滤波、小波滤波和中值滤波)方法的噪声压制效果。相较于常规滤波方法,数值实验表明基于CWT策略的CNN方法具有更好的随机噪声和涌浪噪声压制效果。为了提高模型处理地震数据的泛化性,引入迁移学习对预训练模型进行微调。迁移学习的成功应用表明基于CWT的CNN地震噪声压制方法可以有效且可靠地处理实际地震信号。
In seismic data processing,seismic noise suppression is closely related to improving the signal-to-noise ratio of seismic data in the context of high-precision oil and gas exploration.Effective denoising of seismic signals is a prerequisite for seismic data interpretation and processing.With the development of deep neural networks,convolutional neural networks(CNN)have been applied to the denoising of seismic data.Traditional CNN methods are usually conducted in the time domain.To improve the ability of the CNN in suppressing seismic noise,a continuous wavelet transform was proposed to transform a one-dimensional time-domain signal into a two-dimensional time-frequency domain signal.Two strategies are proposed here to suppress noise in the time-frequency images using CNN:StrategyⅠtakes the amplitude spectrum of the complex matrix calculated by CWT as the training sample of the CNN,while keeping the phase spectrum unchanged;StrategyⅡtreats the real and imaginary images of the complex matrix as different CNN channels.To output results of the CNN,the two-dimensional complex matrix was restored to one-dimensional seismic signals using inverse continuous wavelet transform(ICWT).This study quantitatively compared the noise suppression effects of two CNN strategies with other conventional filtering methods(including low-pass,wavelet,and median filtering)using three indices:peak signal-to-noise ratio(PSNR),structural similarity(SSIM),and root mean square error(RMSE).Numerical experiments showed that the CNN method used in this study performed better than conventional denoising methods in suppressing random noise and swell noise,which proved the effectiveness of the proposed CNN method.To further improve the generalization ability of the proposed model for processing seismic data,transfer learning was introduced to fine-tune the pre-trained CNN model.The results showed that the method presented in this paper can be applied effectively to actual signal processing.
作者
赵金泉
尤加春
魏俊廷
黄聪
ZHAO Jinquan;YOU Jiachun;WEI Junting;HUANG Cong(School of Geophysics,Chengdu University of Technology,Chengdu 610059,China)
出处
《石油物探》
CSCD
北大核心
2023年第3期395-405,共11页
Geophysical Prospecting For Petroleum
基金
国家自然科学基金项目(42004103,42050104,42030812)资助。
关键词
卷积神经网络
连续小波变换
涌浪噪声
随机噪声
迁移学习
时频谱
泛化性
convolutional neural network
continuous wavelet transform
swell noise
random noise
transfer learning
time-frequency images
generalization ability
