Seismic Data Noise Attenuation Based On Deep Residual Network

In this paper,a neural network model for noise suppression of seismic data is proposed,named Seis-Res Net(SRN).Seis-Res Net adaptively recognizes the random noise in seismic data.Seis-Res Net is a 41-layer deep residual neural network framework.After training,Seis-Res Net establishes the end-to-end network mapping between the noisy seismic signal and the separated noise.When Seis-Res Net input layer network receives noisy seismic data,Seis-Res Net can intelligently predict random noise in seismic data and output noise signal from output layer.Training data sets play an important role in deep learning and are the key factors that restrict the prediction accuracy of deep neural networks.In order to solve the problem of lacking label data in the research of artificial intelligence methods for seismic data processing,this paper introduces Online Dictionary Learning(ODL)algorithm from the field of image processing to automatically generate standardized label data sets which is adapted to the training of seismic data neural networks.After data training,Seis-Res Net can effectively suppress random noise in synthetic model and real seismic data.There is no obvious effective signal leakage in noise profile.At the end of this paper,Seis-Res Net is trained and improved by using the method of transfer learning,so that Seis-Res Net can deal with other types of seismic noise.