Research On High-precision Reconstruction Method Of Multi-source Seismic Wavefield Based On Deep Learning Framework

Seismic exploration technology plays an extremely important role in engineering inspection,tectonic illumination,mineral deposit detection and so on.In recent years,with a variety of new methods and technologies constantly proposed,seismic exploration methods still have a good performance in the face of more complex problems.Multi-source seismic exploration includes passive-source seismic exploration and active-source simultaneous-source seismic exploration.Passive source seismic exploration utilizes the seismic waves generated by the vibration of passive sources existing underground.These waves carry a large amount of real low-frequency information propagating upward from the depth,which is of great significance for detecting the deep structure.Active source simultaneous source technology improves acquisition efficiency and data quality by having multiple sources excited simultaneously in a short period of time,resulting in faster acquisition and more coverage.The reconstruction of noise sources using seismic interferometry will inevitably result in the appearance of coherent signals that interfere with the reconstructed active signal.The passive source signals need to be acquired continuously for a long time to ensure that the data have a high signal-to-noise ratio,so the time required for acquisition is costly.In the basic assumptions of seismic interferometry,it is required that the distribution of underground sources is uniform and the medium is not damaged.When the number of underground sources is small or the distribution is not uniform,the application of seismic interferometry to reconstruct the effective signals of the virtual shot records from the passive source seismic records will be weak in energy,poor in continuity,or waveform anomalies.For the illumination of large-scale underground structures,the arrangement of geophones tends to be sparse and the interval between sampling points is large,resulting in poor continuity of reconstructed records.In active-source simultaneous-source acquisition seismic wavefields,neighboring source records in the common shot point domain have similar morphology and comparable frequencies and energies,and therefore interfere with each other and are more coherent.Conventional deblending methods require precise source excitation times,pseudo-separation and channel-set conversion operations,and the final deblending effect is related to the means of suppressing noise and does not allow for true simultaneous excitation.Whether the wavefield is a noise source passive source virtual source or an active source simultaneous source,there is a significant amount of coherent noise as multisource seismic data.In addition,since each receiver receives signals from multiple sources,it is sensitive to the denseness of the receiver channel,and the absence of each channel results in a large amount of signal loss.It can clearly be seen,while having the advantages of multi-source seismic data wavefields,some disadvantages of multisource wavefields are inevitable.Compared with the traditional means of seismic exploration data processing,the invocation of deep learning technology within the field of seismic exploration can significantly simplify the processing process,avoid repetitive operations,and reduce the impact of the subjectivity of the processor,thus allowing seismic exploration to solve more complex and more workload processing problems.Aiming at the common problems of multi-source wavefields and the characteristic problems of different wavefields,this paper examines several methods based on deep learning to solve the problems existing in multi-source wavefields.The main methods proposed and the research results obtained in this paper can be summarized as follows:(1)Aiming at the problem that passive source seismic data requires a long acquisition period to improve the signal-to-noise ratio,a deep neural network-based signal enhancement technique for the virtual source wavefield of short-period-acquired passive source seismic data is proposed.The method utilizes a deep convolutional neural network to mine and enhance the effective signals and suppress the coherent signals in the virtual source wavefield of the passive source seismic data acquired in a short period of time.After experiments,it is proved that the deep neural network-based can enhance the effective signals in the wavefield of the virtual source of the passive source seismic data acquired in a short period,improve the signal-to-noise ratio of the wavefield.By utilizing 5 seconds of passive source seismic records for interferometric reconstruction,the resulting virtual shot record can achieve the effect of the virtual shot record obtained by interfering with 50 seconds of passive source seismic records,which significantly reduces the time required for passive source seismic data acquisition.(2)In order to solve the problem that passive source seismic exploration is sensitive to the number and distribution of underground seismic sources,a wavefield enhancement technique based on UNet with adaptive attention mechanism is proposed for passive source virtual seismic source recordings with inactive seismic sources.It is experimentally demonstrated that,for the problem of weak effective signal energy appearing in the reconstructed virtual source wavefield from passive source seismic records with a small number of sources,the network can suppress the coherent noise more thoroughly to enhance the effective signal.For the problem of waveform anomalies of the effective signal in the virtual source wavefield reconstructed from passive source seismic records with uneven source distribution,the network can correct the waveform to the correct shape.By using the seismic records acquired from 30 passive sources for interferometric reconstruction,the resulting virtual shot record can achieve the effect of the virtual shot record obtained by interfering with the seismic records of 200 passive sources,which greatly reduces the sensitivity of passive source seismic data acquisition to the number of sources.For some passive sources with extreme source distribution,the virtual shot records can also achieve high-fidelity waveform correction,which reduces the sensitivity of passive source seismic data acquisition to the location of the source distribution.(3)Responding to problems of sparse arrangement of geophones or large sampling intervals in passive-source seismic exploration,an interpolation technique based on a global multiscale fusion residual shrinkage network of the wavefield of passive-source virtual seismic sources is proposed.It is experimentally demonstrated that the network is more effective for channel interpolation and sample point interpolation of virtual source records while suppressing coherent signals.Under the condition of supplemented active source seismic data,the network is more effective in processing.For wavefields with missing non-uniform channels,the network encrypts the missing channels while supplementing them.It can expand the number of channels of irregularly missing sparsely acquired seismic data by three times,and realize the highfidelity interpolation reconstruction of seismic records.This method can improve the continuity of the wavefield of virtual sources for passive source seismic surveys and improve the quality of seismic data in large-scale illumination of subsurface structures.(4)In confronting problems of active-source simultaneous source blending acquisition of seismic wavefields,the coherence of neighboring source recordings within the common shot channel set is strong,and the conventional method requires accurate source excitation time and complicated operation.The separation technique of active source simultaneous source seismic wavefield with common cannon point channel set based on improved multi-output Unetr network is proposed.It is experimentally demonstrated that the method does not require the excitation time of the source,skips the redundant processing flow,and separates the more coherent activesource simultaneous-source seismic wavefields into single-source wavefields directly within the set of common-shot channels.Meanwhile,the network has good processing capability for active source simultaneous multi-source seismic wavefields with multiple sub-waves,active source simultaneous seismic wavefields with sparse channels,and noise-containing active source simultaneous seismic wavefields.It is possible to separate multiple single-source seismic wavefields while expanding the seismic traces of sparsely acquired simultaneous source seismic records by three times.This method can reduce the processing technical requirements of active source simultaneous source seismic wavefields,has no dependence on the excitation time,improves the application capability of active source simultaneous source seismic wavefields,and shortens the time required for the separation of active source simultaneous source seismic wavefields.In this paper,we use deep learning methods to target the aspects of acquisition time,number and distribution of passive sources,geophone spacing and sampling point interval to break through the limitations of non-anthropogenic and anthropogenic factors on the passive source virtual shot wavefields and to improve the quality of the passive source virtual shot wavefields.And to simplify the separation problem of active-source simultaneous-source seismic wavefield from the perspective of getting rid of the excitation time and complicated operation steps required for shot separation.