Seismic Background Noise Suppression Based On Trainable Nonlinear Reaction Diffusion Network

In the field of mineral resources exploration,seismic exploration technology is the most important technical method.Using geophones to receive artificially seismic signals,and based on the received seismic exploration record,the geological structure can be verified and used to search mineral resources.In recent years,because of the rapid consumption of oil-gas resources,the reserves of conventional oil-gas reservoirs have decreased significantly.Therefore,it is particularly important to develop unconventional oil-gas reservoirs with abundant reserves.The natural environment and geological conditions of the area where the unconventional oil-gas reservoirs are located are extremely complicated.The effective signal energy in the seismic exploration data received by the geophones is extremely weak,and the background noise interference energy is extremely strong,so that the signal-to-noise ratio(SNR)of the seismic data is extremely low.Especially for seismic prospecting data in the desert region,the effective signals and the background noise are both low-frequency,therefore,the spectrum aliasing causes a large amount of effective signals to be submerged by low-frequency background noise.Therefore,suppressing the low-frequency background noise,extracting the effective signals,restoring the continuity of the effective signals and improving the SNR of seismic prospecting data are the key and difficult issues in the field of seismic signal processing.Because desert background noise has complex characteristics such as non-stationary,non-Gaussian and low-frequency,it leads to the conventional filtering method and the time-frequency domain transform combined with threshold filtering method,while suppressing the low-frequency desert background noise,the energy and amplitude of the effective signals are attenuated.Unable to achieve high SNR and high amplitude preservation requirements.In recent years,deep learning methods and matrix low-rank decomposition algorithms have become hot issues in various fields.Therefore,in order to eliminate the low-frequency desert noise without attenuation of the effective signals,a new method called R-TNRD for suppressing the low-frequency desert background noise is proposed in this paper,which using trainable nonlinear reaction diffusion(TNRD)network assisted by robust principle component analysis(RPCA)algorithm.RPCA algorithm is the generalization of sparse representation theory on matrix,so RPCA has very good sparse representation characteristics.Different from the random noise contained in general land microseismic signals and downhole microseismic signals,which is mainly Gaussian white noise,the low-frequency desert background noise shows non-Gaussian.Therefore,the classical PCA matrix decomposition algorithm is not applicable,and RPCA does not assume that the noise satisfies the Gaussian distribution.It can represent the input noisy desert seismic data as a superposition of two matrices under the optimization criterion.One matrix is called a low-rank matrix and the other is called a sparse matrix.It is a very effective matrix decomposition algorithm.After decomposing noisy desert seismic data into two matrices in the time domain using RPCA algorithm,because the background noise energy in the desert seismic data is extremely strong and the spectrum aliasing of effective signal and background noise is serious,it is impossible to apply conventional filters and threshold to filter the two matrices obtained by decomposition.Therefore,the TNRD network model is introduced into desert seismic data denoising for extracting the effective signals accurately.The TNRD network is a deep learning method based on partial differential equations(PDEs).This network convolves the input data with a series of filters and influence functions,and the loss function and the desert seismic data training set are used to control the network training,and then the optimal network parameters can be automatically obtained,so as to construct TNRD network model suitable for desert seismic data denoising.By using the trained network model to denoise the two matrices obtained by decomposition,we can suppress the low-frequency desert background noise.In the last experimental part of the paper,we applied R-TNRD to the denoising of synthetic desert seismic data and real desert seismic data.The experimental results show that the denoising effect of this algorithm is very good,and the algorithm is compared with conventional filtering methods.The comparative experimental results demonstrate that the proposed method can suppress low-frequency desert background noise more effectively than other conventional methods,and the energy loss of the effective signals is minimal.