Research And Application Of Seismic Data Denoising Based On Online Dictionary Learning Algorithm

In the field of seismic data processing,the quality of seismic data denoising directly affects the effectiveness and reliability of subsequent processing.With the development of seismic exploration,more and more complex reservoirs are preferred,and clean seismic data are difficult to obtain.Therefore,the application of seismic data denoising is an important and continuous research content in the field of seismic data processing.Based on the similarity between seismic data and image and speech signals,this paper applies the fast developing dictionary learning method in the field of image and speech signals to seismic data denoising.Dictionary learning is a sparse representation algorithm based on feature learning.Firstly,the signal is useful information with structural features in seismic data.Generally,noise is random and unstructured,and can not be sparsely represented.The dictionary learning method trains and learns the seismic data to get the data structure matched with the signal in the seismic data,and then gets the excellent sparse representation of the seismic signal according to the reorganization of the data structure.The sparse representation process separates the signal from the noise.Because the computational complexity of conventional dictionary learning method is too large to process large seismic data,this paper introduces online dictionary learning method which can easily process large data sets into seismic data processing.The online dictionary learning method uses the idea of slicing the data to be processed and updating the dictionary column by column,which greatly reduces the computational complexity.In practical application,it is found that if seismic data contain strong and regular noise,the structural characteristics of noise will be learned into dictionaries,and the dictionaries obtained are poor.In order to solve this problem,EMD method is used to optimize the training set,and a few structural features are sacrificed,but the training set does not contain strong coherent noise and random noise.Then the training set is used to learn a high-quality dictionary.The improved online dictionary learning method can also effectively denoise seismic data with strong coherent noise.This paper is divided into the following aspects: Firstly,the content of the traditional sparse representation method based on fixed basis function is expanded,which is widely used in wavelet transform and Curvelet transform.The denoising experiments of synthetic seismogram and actual seismic data are carried out by using these two methods,and the denoising effect is compared;secondly,the dictionary learning method based on sparse representation is studied.In this paper,two dictionary updating algorithms,MOD and K-SVD,are summarized,and three sparse reconstruction algorithms,MP,OMP and LARS,are combined with more advanced K-SVD dictionary and OMP algorithm with higher computational efficiency.The actual seismic data denoising experiments are carried out,and the shortcomings of dictionary learning method in denoising under strong coherent noise are found.Thirdly,the dictionary learning method can not deal with a large number of problems.The problem of seismic data is that online dictionary learning method is applied to seismic data denoising;fourthly,in practical seismic data denoising applications,online dictionary learning method can not remove the shortcomings of strong energy coherent noise.EMD method is used to optimize training set,and the improved online dictionary learning method can effectively remove different types of noise;fifthly,using earthquakes in different work areas.The data is used as training set to get dictionary,and the denoising experiment is carried out.It is concluded that the training set and the data to be processed may not be the same data.As long as the training set contains the structural characteristics of the data to be processed and is clean,the denoising process can be effectively carried out.According to this conclusion,the intelligent processing model of seismic data denoising is prospected.Finally,through the denoising processing results of synthetic seismograms and actual seismic data with different types and intensity of noise,the effect evaluation of various denoising methods in this paper is obtained,and the conclusion is drawn that the improved online dictionary learning method in this paper can process actual seismic data.Compared with the method based on fixed basis function,it can remove random noise more effectively;compared with the single dictionary learning method,it can remove strong energy coherent noise effectively;and break the limitation of training set size of dictionary learning method,it can process seismic data in actual production,learn more seismic signal characteristics,it can make seismic signals have a better sparse representation.