Deep K-SVD Algorithm And Its Application For Desert Seismic Noise Attenuation

Energy sources such as oil and natural gas are of great significance to the development of China's society.Seismic exploration is widely used as a means of proving underground information.The scene of seismic exploration has gradually moved from plain areas to complex geological conditions such as woodlands and deserts.The seismic data in desert area is often accompanied by a large amount of random noise,which is obtained by superimposing natural noise and human noise.Its characteristics are complex,with non-stationary,non-Gaussian,and non-uniform characteristics.It is a low-frequency colored noise which frequency is close to signal frequency.Therefore,it is difficult for traditional seismic noise suppression methods to meet the denoising requirements.In order to achieve the three high requirements for desert seismic data,it is necessary to conduct in-depth research on desert seismic noise suppression schemes to improve the quality of seismic data and subsequent interpretation of underground structures.Deep learning has achieved remarkable results in image processing field.In this paper,the deep K-SVD denoising algorithm is introduced into seismic signal processing to suppress random noise in desert seismic data.The deep K-SVD algorithm combines deep learning to implement the K-SVD framework.By using the multi-layer perceptrons to predict the priori information and sparse coefficients of unknown noise,the dictionary is obtaind to sparse represent the seismic signal.The algorithm uses an iterative threshold shrinkage algorithm for sparseness.It shows that the denoising is completed by taking into account the local and global information in observation data,and thus the problem of signal distortion is solved.The algorithm uses supervised learning to train the weights and deviations of each node in the multi-layer perceptron,so that it can adaptively obtain shrinkage threshold parameters according to different input noises.In view of the problem that deep K-SVD takes a long time in sparse representation stage,in this paper,the fast iteration threshold shrinkage algorithm is adopted to improve the deep K-SVD algorithm(FDK-SVD),reducing the number of iterations.In the proposed method,the fast iteration threshold shrinkage algorithm is used to perform sparse coding,making it possible to obtain the best denoising effect with as few iterations as possible.Therefore,it is more suitable for a large number of seismic data processing.Simulation analysis and field seismic data processing results verify the denoising performance of the improved algorithm.Under the same number of iterations,the FDK-SVD improves the denoising effect and reduces the energy loss of the signal during the denoising process.It also improves the calculation efficiency.In addition,comparing with wavelet algorithm and non-local mean filtering algorithm,FDK-SVD proposed in this paper has superiority indesert seismic data attenuation.