| Many geological information can be obtained by seismic exploration,which is very significant to the exploitation of oil and gas reserves.The demand for such resources is gradually increasing with the development of the society,especially nonrenewable resources.Easily accessible conventional resources are increasingly consumed and facing depletion.Therefore,conventional oil and gas resources exploitation cannot have satisfied modern humans and society's need for its growing,and unconventional oil and gas reservoir exploitation has been one of the research focus.Unconventional reservoir of oil and gas is much harder to obtain than that of conventional reservoir,and the environment around them is more harsh.So,a large amount of seismic data collected by us will be accompanied by strong noise,and the signal-to-noise ratio(SNR)of the data will become very low,which brings great difficulties to our subsequent seismic record interpretation.Therefore,one of the problems we need to solve is how to remove the strong noise in these seismic records,restore the effective signal,and improve the accuracy and signal-to-noise ratio of seismic data.This paper mainly focuses on desert areas,because there are a large amount of unused oil and gas resources in desert areas,and successful exploration will be of great significance.As is known to all,the environments are extremely harsh in desert areas.Apart from the difficulty of mining,seismic records collected can also be seriously polluted by noise,which poses a great challenge to seismic exploration in desert areas.Desert noise is characterized by low frequency,non-gaussian,non-stationary,high energy and seriously aliasing with the effective signal in the frequency domain.Therefore,eliminating noise from desert seismic record is a huge challenge.Since the development of seismic exploration noise suppression methods,many scholars at home and abroad have put forward many excellent algorithms,such as bandpass filter,time-frequency peak filtering,adaptive filtering,Radon transform,wavelet transform and Shearlet transform.These typical algorithms have been used in the field of seismic signal processing.Although they have a certain effect in reducing random noise,the processing effect of more complex desert noise is not ideal,and the classical algorithms still need to be improved.Based on the above problem,we designed a novel network DnResNeXt to process desert seismic data.In terms of network structure improvement,DnResNeXt utilized skip connections,residual learning and grouped convolution,which can extract and transmit more features than Dn CNN or Res Net.Moreover,our network doesn't need to apply Batch Normalization(BN)that is an indispensable layer for preventing gradient explosion and disappearance when CNN is processing the image.In this way,the DnResNeXt can extract the original characteristics from desert noise data directly.The overall algorithm flow is as follows.First of all,we build the training sets of high quality,including the effective signal set and the noise set.And then we use the training sets to train the DnResNeXt network.At last,the trained model is utilized to process the desert seismic data.And then the denoised result could be calculated by subtracting the network's output(the predicted result)from the noisy record.In synthetic and field desert seismic record processing,the method proposed in this paper are compared with the classical method.And the aspects of time domain,frequency domain,SNR enhancement and mean square error(MSE)are investigated.It can be seen that the performance of the proposed method is much better than that of the classical noise suppression methods in both synthetic and field record.In the denoising results of our method,the effective signal is clearly and continuously restored.The noise is suppressed to the maximum extent,and the SNR is improved greatly,which amply demonstrates that the effectiveness of our method,and it has a big advantage over the classical method in processing desert seismic records. |
PDF Full Text Request