一种基于卷积网络的地震探测数据随机噪声去除方法

为有效提高地震数据信噪比,通过卷积神经网络(convolutional neural network,CNN)的方法研究了地震勘探数据去除随机噪声问题.该方法包含17个卷积层,使用线性整流(rectified linear unit,ReLU)激活函数避免梯度消失,使用批量标准化(batch normalization,BN)提高网络的泛化能力.所构建的网络应用残差学习策略,即输入为含噪地震正演叠前数据,输出为CNN网络学习获得的随机噪声.然后从地震记录中减去网络预测的噪声数据,从而达到去除随机噪声的目的.同时,根据地震勘探数据振幅随探测时间衰减的规律,在网络训练过程中进行深度加权,使得CNN对于深部噪声的学习效果更好.网络在PyTorch框架下训练,应用图形处理器并行计算可以有效提高网络训练速度.利用训练好的网络进行去噪实验,结果表明与传统的时空域预测滤波法相比,该网络能更好地压制随机噪声.可见针对地震勘探数据,CNN能够有效提取含噪数据中的噪声信息,证明了该方法在去除随机噪声方面的合理性与有效性.

Random noise removal method for seismic detection data based on convolutional network

In order to effectively improve the signal-to-noise ratio of seismic data, convolutional neural network (CNN) was used to investigate the problem of removing random noise from seismic exploration data. The method contains 17 convolutional layers, and ReLU activation function is used to avoid the disappearance of the gradient, and Batch Normalization (BN) is used to improve the generalization ability of the network. The constructed network adopts residual learning strategies, that is, the input is pre-stack data including seismic random noise, and the output is random noise learned from CNN. And then the noise predicted by the network is subtracted from the seismic recoreds, so as to achieve the purpose of removing random noise. At the same time, according to the characteristics that amplitude of seismic exploration data attenuates with detection time, we performed depth weighting during the network training to make CNN learn better for deep noise. The network is trained under the PyTorch framework. Using GPU parallel computing can increase the speed of network training. The results show that using the trained network for denoising experiments can remove random noise better than the traditional f-x domain prediction filtering method. It is concluded that for seismic exploration data, CNN can effectively extract noise information in noisy data, which proves the rationality and effectiveness of this method in removing random noise.