Research On Seismic Data Denoising Denoising And Reconstruction Using Deep Learning

With the continuous improvement of oil and gas exploration,the object of seismic exploration is gradually shifting from structural reservoirs to lithological reservoirs.Thus,seismic data with high signal-to-noise ratio,high resolution and high fidelity is needed to describe the fine structure of underground oil and gas reservoir.The increasingly complex exploration environment usually makes the seismic data incomplete or irregular and polluted by various random noise.But the noise and the incompleteness will affect the fidelity,resolution,and signal-to-noise ratio of the subsequent data processing,which in turn reduce the reliability of the final data interpretation.Therefore,it is necessary to suppress the noise and reconstruct the missing traces in the seismic data.However,the conventional reconstruction methods,which are limited by the Nyquist sampling theorem,usually requires a higher sampling rate of seismic data and results in a rather high cost for data acquisition.Furthermore,conventional denoising and reconstruction methods usually rely more on the prior information of seismic data.In recent years,the emerging deep learning has achieved excellent performance in many fields.This method aims to let the model learn the features of the data at different depths first,and then use the trained model to predict the unknown data.To solve the mathematical problem of seismic data reconstruction and denoising,this paper explored the mathematical model based on deep learning,which can be uniformly applied to seismic data denoising and reconstruction.This paper first established a simple and effective data preprocessing method,which made the neural network achieve rather good results of training and test.In the denoising of seismic data,a deep convolutional neural network was first constructed,then the training data was established as the patch pair of the noisy data and its corresponding noise,and finally the network learned to predict the noise in the data through supervised learning.On this basis,a model,or called a mathematical framework,for uniformly predicting different levels of noise in 2D and 3D seismic data was established,which realizes that different levels of noise can be separated from the data without any prior information and the whole process will not produce artifacts.In the reconstruction of 2D seismic data,a deep convolutional neural network was first constructed,then the patch pairs composed of randomly sampled data by 50% and their corresponding complete data were used as the training data for the neural network.The trained network realized high-precision reconstruction for 2D seismic data after regular and irregular sampling by 50% or even far below 50%.On this basis,a model was established for unified reconstruction for 2D and 3D seismic data and simultaneous denoising and reconstruction,realizing not only the reconstruction of 2D regular or irregular seismic data whose sampling does not satisfy Nyquist theorem without any prior information and pre-interpolation but also simultaneous denoising and reconstruction of 3D seismic data.In the denoising and reconstruction of 3D seismic data,based on the non-interference working mechanism of the detectors,a processing strategy was proposed that 3D seismic data can be disassembled into a number of 2D seismic profiles to process.And the above trained neural network that can predict different levels of noise was used to suppresse the noise in 3D seismic data effectively.Then the randomly sampled data by 50% with noise and their corresponding complete data were regarded as new patch pairs and were used as new training data to retrain the above network for reconstructing 2D seismic data,realizing the simultaneous denoising and reconstruction of 3D seismic data with regular and irregular sampling.This processing strategy greatly improves the processing efficiency of 3D seismic data and the utilization rate of trained neural networks while ensuring the quality of the results.The results of numerical experiments verify the effectiveness of the seismic data denoising and reconstruction method based on deep learning.Deep learning will provide a new idea for seismic data denoising and reconstruction with high fidelity,high resolution and high precision.At the same time,it will also provide a new solution for seismic data to achieve high-efficiency and low-cost acquisition.