Deep Weak Seismic Signal Detection Technology Based On Compressive Sensing

High signal-to-noise ratio,high resolution and high fidelity are three important tasks in seismic data processing.High signal-to-noise ratio is foundation of the following further seismic data processing and interpretation.Because of the restriction of economic and terrain problems,the field acquisition seismic data is always irregular.As a consequence,the following techniques in seismic data processing,like multiple attenuation,AVO analysis,AVAZ analysis and wave-equation migration,are badly affected.Seismic data denoising and reconstruction are crucial to seismic data processing.In recent years,dictionary learning and sparse representation are widely applied to seismic data processing and obtain desired results.This method make initial dictionary change into certain dictionary which is adaptive to dataset to be processed.On basis of the dictionary,we can sparsely represent noisy or decimated seismic data in order to obtain denoised or reconstruction results.Commonly used algorithms,K-SVD,remains some problems like large amount of computation time and a few of useless atoms.Aimed at settling these problems,we make an improvement in the process of dictionary learning and sparse representation and we apply this algorithm to 3D seismic data processing.Details are as follows:(1)Soft threshold is taken part in sparse representation avoiding the influences from noisy data.So that compared with K-SVD,the learned dictionary is more representative and universal.Signal can be represented with less coefficients which verifies the superiority.(2)In the process of updating atoms,singular value decomposition is replaced with alternating least square combined with the restriction of 2-norm.And orthogonal matching pursuit(OMP)is replaced with batch-OMP.So that we can further reduce the fitting of noise and accelerate the speed of sparse representation.(3)The calculated amount is reduced by half by changing time domain to frequency domain.Furthermore,parallel computing technology is utilized to save time.The denoising or reconstruction results acquired by dictionary learning in frequency domain are smoother and more continuous than traditional dictionary learning algorithms.The proposed method also solves issues that seismic sections exist much spininess caused by the inconsistency of data along time axis.