Research On Simultaneous Reconstruction And Denoising Of Seismic Data Based On U-Ne

In the actual acquisition process of seismic data,there will be missing traces irregularly distributed in the spatial direction.In addition,random noise and high-amplitude noise are inevitably mixed in the acquired data due to the acquisition condition restriction.Reconstructing and denoising the irregularly missing seismic data containing random noise and high-amplitude noise is beneficial to improve the accuracy of subsequent seismic data processing and interpretation.The traditional compressed sensing methods for simultaneous reconstruction and denoising require to choose the processing parameters by trial and error,and may cause distorted signal and residual noise for processing complex seismic data.Existing convolutional neural network methods cannot suppress the highamplitude noise effectively.In this thesis we research the simultaneous reconstruction and denoising method of seismic data by introducing the commonly-used convolutional neural network U-Net.The main works are as follows:(1)We introduce the convolutional neural network U-Net,which is commonly used in seismic data processing,for simultaneous reconstruction of seismic data with random noise and high-amplitude noise suppression.We build an end-to-end U-Net,train it on the synthetic data,and then use transfer learning on real data to fine-tune the U-Net.The trained U-Net learns a nonlinear mapping between input data and labels by optimizing the loss function of the network.The input data contains the random noise,high-amplitude noise and irregularly-missing traces.The labels refer to the complete and clean data.The processing results on the synthetic and real data show that the U-Net method makes full use of the information of multiple training windows,has a better reconstruction result,suppresses random noise and high-amplitude noise more effectively,achieves higher signal-to-noise ratio and peak signal-to-noise ratio compared with the traditional compressed sensing and convolutional neural network methods.(2)The U-Net method does not consider the non-Gaussian distribution nature of the high-amplitude noise.In order to solve this problem,we propose the simultaneous reconstruction and denoising method based on Huber loss function according to the nonGaussian distribution nature of high-amplitude noise.The Huber loss function is used to replace the loss function of mean square error on the U-Net model.By imposing the Huber norm minimization constraint on the high-amplitude noise,the Huber-U-Net method suppresses the high-amplitude noise and protects the signal more effectively,as well as achieves a higher signal-to-noise ratio and peak signal-to-noise ratio than the U-Net method.The processing results on the synthetic and real data demonstrate the effectiveness of the proposed Huber-U-Net method.(3)In order to enhance the ability of feature expression and optimal selection of the UNet method,we propose the simultaneous reconstruction and denoising method for seismic data combining attention mechanism.The proposed method adds multiple Sequeze and Excitation Block(SE-Block)in the concatenation operation between the coding and decoding stages on the U-Net.The proposed SE-U-Net method combines the attention mechanism,assigns different weights to different features through a learnable weight matrix and enhances the feature extraction ability of seismic signals.Compared with the UNet method,the proposed SE-U-Net method suppresses more high-amplitude noise,causes less signal loss,achieves a higher signal-to-noise ratio and peak signal-to-noise ratio.The processing results on synthetic and real data demonstrates the effectiveness of the proposed SE-U-Net method.