Reconsitution Of Irregular Seismic Data Based On Recurrent Neural Network

Seismic data is the innate basis of inversion and final data interpretation,and its quality directly affects the subsequent processing of seismic data.The use of deep learning technology to interpolate seismic data is a research hotspot in recent years.However,currently widely used convolutional autoencoders,generative adversarial networks,etc,all treat the interpolation problem of seismic data as images for processing,and the methods lack a certain degree of flexibility.This article starts from the seismic data itself,by treating the seismic data in the spatial domain as related sequences with the characteristics of "time-series",a new interpolation method of "traces and traces" is proposed.And realized the reconstruction of irregular missing seismic data by using Recurrent Neural Networks.We use the Gated Recurrent Unit(GRU)in the Recurrent Neural Network to predict the seismic data in space,reconstruct the seismic data under irregular missing by inputting the known seismic trace and outputting the missing seismic trace.First of all,we normalize the data and select Curvelet Transform as the sparse representation base to generate the prior trace data of missing seismic traces,and use the prior trace to fit and learn the corresponding label trace;Secondly,in the aspect of model design,we improve the single-layer Gated Recurrent Unit.In this paper,we design a depth(three layers)Bidirectional Gated Recurrent Unit(Bi GRU)is to learn the advanced features of seismic data,and use a skip connection mechanism to alleviate the common gradient vanishing problem,and complete the reconstruction of seismic data by letting the network continuously learn the optimal residual between the prior trace and the label track;Finally,we verify the effectiveness of the algorithm through a large number of experiments on 2D synthetic seismic records and real seismic data.The experimental results show that the Bi GRU model proposed in this paper has a good interpolation effect on the reconstruction of seismic data.The SNR has been significantly improved,which can reduce the error in reconstructed data to a certain extent and suppress the generation of spatial aliasing.