一种基于RUnet卷积神经网络的地震资料随机噪声压制方法

应用深度学习方法压制地震噪声的训练集和测试集均来自同一数据集,使得模型的泛化性受限。为解决泛化性问题,提出一种基于卷积神经网络(CNN)的残差U型网络(RUnet)压制随机噪声的方法。方法的设计思想是在基于卷积神经网络的U型网络(Unet)基础上加入残差块,以增强网络对随机噪声的捕获能力。该方法建立在端到端的编码解码网络结构上,将含噪声地震数据作为输入,由多个卷积层和残差块提取随机噪声的本质特征,构成编码;再由多个反卷积层和残差块构成解码,网络的输出即为噪声压制后的地震数据。在残差块之后加入批规范化层,采用带泄露整流函数作为非线性因子,提高网络模型对地震资料随机噪声的泛化性和敏感性。在叠后和叠前地震数据实验中将RUnet卷积神经网络方法与小波变换、离散余弦变换、三维块匹配(BM3D)算法和Unet卷积神经网络算法进行去噪效果对比,结果表明,RUnet卷积神经网络方法相比其它4种方法,对随机噪声的压制更有效,并且在一定程度上保护了有效信号。

The training and test sets used for deep learning methods to suppress seismic noise originate from the same data set,which leads to limited model generalization.This study proposes a random noise suppression method using the RUnet convolutional neural network.The design idea of the method is to add a residual block on the basis of the Unet convolutional neural network to enhance the networks ability to capture random noise.First,based on the code-decode network structure,this method used seismic data containing noise as the input,and extracted the essential features of random noise using multiple convolutional layers and residual blocks to form the coding process.Then,multiple deconvolution layers and residual blocks were used to build the decoding process; the output of the network was noise-suppressed seismic data.A batch normalization layer was added after the residual block and the leakage rectification function was used as the nonlinear factor to improve the generalization and sensitivity of the network model to seismic random noise.Testing using post-and pre-stack data showed that compared with the wavelet transform,discrete cosine transform,BM3D algorithm,and Unet convolutional neural network,the RUnet convolutional neural network was more effective in suppressing random noise and preserving signal.

国家重点研发计划深地专项项目(2016YFC0601100)和四川省科技计划项目(2019CXRC0027)共同资助。