Research On Seismic Data Reconstruction Method Based On Compressed Sensing

Seismic exploration is an important part of the oil exploration process.The accurate representation and analysis of seismic data is an important basis for geological rock stratum classification and oil gas reserves prediction.In actual exploration,due to environmental,equipment or human factors,there are a large amount problems of data missing and incomplete,which seriously affects the subsequent data interpretation.In addition,with the development of the petroleum exploration field,the accuracy requirements for seismic reconstruction data are getting higher.According to the limitations of the Nyqusit sampling theorem,high-precision reconstruction data often means an increase in the sampling frequency,which also means an increase in exploration costs.Reconstructing irregularly incomplete seismic data more accurately,or reconstructing seismic data that meets certain accuracy requirements using lower sampling frequencies is significant and also a huge challenge.In order to better solve the problem of seismic reconstruction,this paper is based on the theory of compressed sensing.The new seismic data reconstruction method is proposed,starting from the sparse representation of seismic data,the compressed sensing algorithm with space-time continuity constraints and the compressed sensing algorithm based on generative adversarial networks(GAN).The specific research work is as follows:(1)Seismic data has the characteristics of multi-scale,multi-directional and local variation.It is impossible to use the traditional single sparse transformation method to perform sparse representation of seismic data,which will lead to the unsatisfactory reconstruction effect.This paper proposes to use the K-SVD dictionary learning algorithm to train an over-complete dictionary and use it as a sparse transform base in the process of seismic data reconstruction,which can better perform sparse representation of seismic data.Experiments show that this method can reconstruct seismic data better than other sparse transform methods.(2)The traditional theory of compressed sensing is discussed for single-frame data.However,seismic data has spatio-temporal continuity information.Using traditional compressed sensing methods to reconstruct seismic data only satisfies the rationality of single-frame data reconstruction,without considering continuity of adjacent frames.In this paper,the traditional compressed sensing framework is modified,and the space-time continuity constraint is added to improve the accuracy of reconstruction.In addition,the SAMP is improved,and the initial sparsity estimation and variable step size strategies are added.The calculation speed is greatly improved on the basis of ensuring the reconstruction accuracy.Finally,the superiority of the algorithm is verified on real seismic data and electrical imaging data.(3)This paper proposes to combine the generative adversarial networks(GAN)with the theory of compressed sensing to avoid the sparse representation of seismic data.In order to improve the reconstruction effect and training stability,this paper uses the DCGAN framework and uses Wasserstein Distance as the evaluation index in the training process.The experimental results show that the proposed method can reconstruct seismic data that meets certain accuracy requirements at a lower sampling rate.