Research On Random Noise Suppression Method Of Seismic Data Based On Deep Learning

At present,the exploration and development of oil and gas fields extend to the high depth and complex blocks,where the stratigraphic response is weak and there are many influencing factors.Affected by complex geological conditions and acquisition technology,random noise is inevitably introduced into seismic data,which seriously reduces the signal-to-noise ratio（SNR）of seismic data and affects the fine description of subsequent geological structure and lithologic characteristics.Therefore,the research on the high-precision random noise suppression method is the key problem to be solved.Deep learning technology has attracted much attention in the field of seismic data processing because of its strong ability of deep feature extraction and nonlinear approximation.However,the existing random noise suppression methods based on deep learning usually only focus on the feature extraction in time-domain and design the denoising model with sufficient sample coverage as the assumption,which is easy to lead to insufficient feature extraction of seismic data,fuzzy texture after denoising,poor robustness of the model and so on.Aiming at the related problems,based on the deep learning technology,this paper expatiates the progress and existing problems of the current seismic data denoising methods,analyzes its basic idea,studies the construction method of the seismic sample set,and puts forward a denoising model combined with multi-dimensional features and a robust denoising model,to improve the accuracy and generalization ability of random noise suppression.The main research contents are as follows:（1）This paper studies the construction method of seismic data sample set,According to the characteristics of seismic data,this paper analyzes the acquisition process and format analysis method of seismic data,studies the sample organization of synthetic data and actual seismic data,and constructs the seismic data set for experimental analysis.（2）Aiming at the problems of insufficient feature extraction of seismic data and fuzzy texture after denoising,this paper analyzes the characteristics of seismic data in time and frequency domain,combines multi-dimensional feature information,designs joint loss function and expanded convolution,combines residual learning algorithm,and proposes a random noise suppression method of seismic data based on joint deep learning.Experiments show that compared with similar algorithms,this algorithm has a better effect on events feature preservation and a higher SNR.（3）Aiming at the problem of poor robustness of denoising model under the condition of incomplete coverage of seismic data samples,this paper studies the method to improve the generalization ability of the model,analyzes the relationship between noise intensity and denoising,integrates the global and local information of seismic data,designs a two-stage network,combines L₁norm,feature fusion,residual learning algorithms,and proposes a robust seismic data denoising method based on deep learning.Compared with similar algorithms,the experiments show that this algorithm has higher generalization ability.Based on the deep learning technology,this paper analyzes the problems and basic ideas of the current denoising algorithm based on deep learning,studies the construction method of the seismic sample set,and designs a denoising model combined with multi-dimensional features and a strong robust denoising model,which can effectively enhance SNR and fidelity of seismic data,and provide guarantee for the subsequent description of geological structure.