Research On Random Noise Suppression Of Seismic Data Using Dual-channel Neural Network

Suppression of random noise has always been a focus in seismic data processing.The existence of random noise leads to great interference and low Signal-to-Noise Ratio of seismic data,which is not conducive to distinguish effective signals.Therefore,enhancing Signal-to-Noise Ratio is an indispensable research step in the field of seismic data processing and analysis.Traditional methods such as F-X deconvolution and Curvelet have some limitations in solving seismic denoising,which are mainly reflected in two aspects: For one thing,they are difficult to balance the removal of noise and the retention of signal.For another,they are not effective.In recent years,the data-driven depth learning methods have obvious advantages in image noise attenuation.Therefore,it is of great value for the research and application of depth learning method in the field of seismic noise attenuation.First,aiming at the shortcomings of traditional methods,we adopt Denoising Convolution Neural Network to attenuate random noise in seismic data.The experimental results show that the denoising convolutional neural network is feasible in the task of seismic noise attenuation and realizes fast intelligent denoising.Then,aiming at the insufficient learning of effective information in denoising convolutional neural network,we propose to send the seismic data into two convolutional neural networks with different structures for learning to extract the complementary information from seismic data,and integrate the outputs of the two networks through concatenate;In view of the high similarity of the effective information between each trace set in seismic data,the dilated convolution is introduced to increase the receptive field of the network and fully capture the neighborhood information in seismic data,so as to retain more useful texture detail information;Aiming at the influence of activation function on the convergence of network and the possible phenomenon of “dying Re LU”,using Swish activation function instead of Re LU activation function can improve the convergence speed and noise attenuation performance of network.Based on these three improvements,a Dual-channel Convolutional Neural Network(SDC-CNN)is proposed.Finally,based on Signal-toNoise Ratio(SNR),Peak Signal-to-Noise Ratio(PSNR)and Root Mean Squared Error(RMSE)denoising evaluation indexes,we compare the performance of the SDC-CNN with those of the common traditional methods and neural network model on a synthetic Common Middle Point(CMP)gather,preprocessed field seismic data,pre-stack seismic data and post stack seismic data.The results of these experiments show that,compared with F-X deconvolution,Curvelet,U-Net and Denoising Convolutional Neural Network(Dn CNN),SDC-CNN is most effective in attenuating random noise as well as in reducing the loss of texture detail information,and has faster data processing speed.