| Due to the influence of sensor sensitivity,acquisition method,field noise and other factors,the effective information often contains various kinds of noise.If we want to get a clear and effective reflection signal,we must suppress or eliminate these noises.Traditional denoising methods usually use a mathematical transformation method,which needs to go through a tedious process of manual selection of data and parameters.Improper threshold will lead to poor denoising effect,and even damage of effective information.The convolution neural network has been applied to the denoising of seismic data by predecessors,and good results have been achieved.However,it is difficult to get real clean data and noise from actual seismic data,which leads to the unsatisfactory effect of this method in practical application.In order to solve this problem,this paper develops a denoising method of seismic data based on the geological structure guidance and convolution neural network,which uses the geological structure guidance filtering method to generate labeled data sets to train the network,overcomes the problem of the lack of actual data samples in the conventional convolution neural network,and avoids the process of the traditional denoising method to select parameters manually.It realizes the efficient de-noising of seismic data,preserves the details of in-phase axis edge,fault and so on,and improves the signal-tonoise ratio of data.The main research contents are as follows:1.Research on guided filtering method of geological structure.The traditional guided filtering method of geological structure is not enough to protect the information of geological structure such as the same phase axis and fault,and has no amplitude preserving property,and has poor anti noise property in the calculation of dip angle.A Gaussian beam filtering method based on coherent attribute is proposed,which only has filtering weight in the direction of construction extension,making the filtering result more fidelity.The accuracy of angle calculation determines the quality of de-noising.The structural tensor method used in this paper can obtain the direction information of signal efficiently and accurately,avoid large-scale direction measurement,and has a certain degree of anti noise.At the same time,the coherent attribute is introduced to control the filtering scale,so that the geological structure information such as faults can be effectively protected.2.Research on seismic data denoising method based on convolution neural network.Many existing denoising methods need to face the process of parameter or threshold optimization,which has a direct impact on the denoising effect.When the signal-to-noise ratio is low,it is difficult to accurately select the appropriate threshold,resulting in poor denoising effect.The method of seismic data denoising based on deep learning convolution neural network can directly let the machine learn the features of the data,so as to identify the noise,avoid the tedious process of parameter selection,and solve the problem that the noise is difficult to remove under the low SNR.Combined with batch standardization and adaptive moment estimation,the method of seismic data denoising based on convolution neural network is realized.Through the test,the application effect of this method in model seismic data and actual seismic data is compared and analyzed.3.Research on seismic data denoising method based on geological structure guidance and convolution neural network.Because the actual seismic data samples are few,the convolution neural network will damage the effective information in the actual seismic data denoising.To solve this problem,a new strategy of sample set construction is proposed.The geological structure guided filtering method is introduced into the convolutional neural network,and the actual seismic data and noise are obtained by Gaussian beam filtering based on coherent attributes,which makes up for the shortage of training samples of the actual data of the convolutional neural network.The diversity of training data set is increased by increasing training samples.The method of residual learning is used to improve the convergence speed and training accuracy of the network.Through the test of model data and actual data,it shows that the method of seismic data denoising based on geological structure guidance and convolution neural network proposed in this paper can effectively remove the random noise in the seismic data and protect the effective information.The results of 3D seismic data processing and attribute extraction show that this method can provide a good data base for subsequent seismic data processing and interpretation. |
