Seismic Random Noise Suppression Based On Adaptive Expected Patch Log-likelihood Algorithm

Seismic exploration is a major and effective method in exploration of oil and gas resources.In the seismic exploration process,it will be affected by the factors such as acquisition environment,humanistic activities and record equipment,which leads to a large amount of random noise in the collected seismic data.Seismic random noise hinders the identification and picking of the reflection signals,which contains the underground structural information,negatively affecting the interpretation and imaging oil and gas resource distribution.Therefore,suppressing seismic noise and improving the signal-to-noise ratio of seismic data is the primary and fundamental part in seismic signal processing.As the increasing economic development and the demand of oil and gas resources,the seismic exploration gradually turns to the regions with increasingly complex landforms,and the underground media.Affected by complex landforms and underground media,the effective signal and noise collected are complex in nature Therefore,it is urgent to design a more efficient seismic noise suppression algorithm for seismic signals recovery and noise attenuation.In order to avoid patch effects,Expected Patch Likehood(EPLL)method learns the Gaussian mixture models from external data as seismic signal priority toconstrain image denoising problem.the EPLL algorithm introduces auxiliary variables by semi-secondary splitting method,to optimize the image denoising problems under the priori of image patch,which is solved by Alternate optimization two sub-problems:image patch denoising and image reconstruction.The EPLL algorithm first divides the noisy data into patches,uses the Gaussian component that best matches the image patches to denoise,and weights the denoised patches and the noisy data to reconstruct the denoised seismic image.In the EPLL algorithm,regularization parameters associated with noise variances plays a crucial role in adjusting denoising intensity and retention signal details.However,when the non-stationary desert seismic exploration data is denoched,if only the noise variance level is set to handle the nonsteady seismic signal,it is necessary to ensure that the strong signal data patch is not distorted,which will cause weak signal patch noise suppression.It is difficult to simultaneously balance the strong signal fidelity and weak signal recovery.This requires EPLL regularization parameters to adjust according to each data patch feature.In this thesis,the EPLL algorithm is not adapted to a non-steady seismic signal problem,and the space adaptive EPLL(SAEPLL)seismic low frequency random noise noise reduction method is proposed.The SAEPLL method controls the regularization parameters according to the patch signal-to-noise ratio,so that it is adaptively adjusted with the time and space of the non-steady seismic exploration signal strength,thereby balancing the resumption of local signal details and the recovery of global feature.Further,during the signal reconstruction process,the signal loss is reduced by the patch signal-to-noise comparison signal patch,and the signal loss is reduced.The experimental results of synthetic and actual seismic data prove that the SA-EPLL algorithm can effectively restore the non-stationary signals and suppress the random noise in the seismic exploration data.In the EPLL method,the regularization parameters are often set by manual tuning or the noise variance.However,in the actual denoising process,the above-mentioned methods are not only inefficient,but also need to be adjusted every time they are faced with different seismic data,and the robustness is too poor.Therefore,this paper embeds the EPLL algorithm into the deep learning framework and proposes an Unbalanced Deep Expected Patch Log Likehood(UNDEPLL)to suppress seismic noise.The denoising network consists of the expected patch log-likelihood denoising main network and an unbalanced multilayer perceptron parameter estimation network.The patch signal-to-noise ratio is estimated through an unbalanced multilayer perceptron network,and the patch signal-to-noise ratio is input as a regularization parameter into the main network to suppress seismic noise.Since the result of patch denoising is robust to high-level SNR errors,but more sensitive to low-level SNRs,in order to reduce this unbalanced sensitivity to signal-to-noise ratio estimation errors,nonequalized sensitivity is used.The equalization loss function imposes a greater penalty on the estimated low-level signal-to-noise ratio error,and adjusts the learning of the multi-layer perceptron network parameters.The unbalanced multilayer perceptron parameter estimation network can accurately estimate the regular term parameters for each patch,avoiding manual parameter adjustment,better controlling the denoising intensity of each patch,and improving the patch denoising effect.The experimental results of synthetic and actual seismic data prove that UNDEPLL is superior to traditional seismic denoising algorithms in suppressing strong desert noise and maintaining signal details.The spatial adaptive EPLL algorithm and the unbalanced deep EPLL network proposed in this paper on the basis of the EPLL algorithm solve the problem of regularization parameter selection under non-stationary seismic signals,and realize the effective suppression of noise under seismic non-stationary signals.Interpretation of seismic exploration data provides a basis.