Research On Seismic Data Denoising Method Based On Deep Convolutional Neural Network

Data acquired by seismic exploration contain a lot of noise,which reduces the quality of the data and is not conducive to the subsequent processing of velocity analysis and migration imaging.Therefore,denoising is a necessary part of seismic data processing.Traditional seismic data denoising methods are based on model driven,and some parameters need to be set manually.However,the noise level of field-acquired seismic data is unknown,which makes the denoising performance largely depend on the experience of researchers.On the other hand,its calculation process is complicated,which seriously affects the real-time performance of denoising.Due to the powerful feature extraction ability and completely data-driven,deep learning technology has achieved better results than traditional methods in the field of image denoising,and has faster processing speed.In this paper,deep learning technology is applied to seismic data processing,and the deep convolutional neural network models and algorithms are improved for better solving the problem of seismic noise suppression.The main research work is as follows:1.A seismic data denoising algorithm is proposed by improving the classical denoising model DnCNN.The output of the model DnCNN is changed to learn the effective seismic signal by analyzing the difference between seismic data and image data.Experiments show that the deep convolutional neural network can accurately extract the characteristics of seismic useful signal,so as to achieve better denoising performance.2.A seismic data denoising algorithm is proposed based on convolutional neural network combining Inception structure and residual learning unit.The above seismic denoising algorithm based on improved DnCNN loses some in-phase structure information while denoising.In this paper,the Inception structure is added to convolutional neural network to obtain the multi-scale features of seismic data,and the residual learning unit is also embedded prevent the phenomenon of gradient disappearance.Experiments show that the Inception structure and residual learning unit can improve the denoising performance of deep convolution neural network for seismic data.3.A seismic data denoising algorithm is proposed based on U-shaped convolutional neural network guided by attention mechanism.The two denoising models mentioned in this article have complex structures and a large amount of calculation.Therefore,the UNet with a narrow network architecture is used for seismic denoising,and two innovations have been realized.(1)Guided by attention mechanism,UNet can extract important features of seismic data and effectively retain the detailed information of seismic profile.(2)Using multi-scale loss function to train the network,the network can learn more detailed features of seismic useful signals.Experiments show that the algorithm is superior to the previous two algorithms in detail protection ability,SNR and computation efficiency.