Research Of Adaptive Learning Denoising Method Based On Seismic Weak Signal Protection

With the continuous exploitation of shallow and middle-level oil reserves,traditional oil and gas exploration has been unable to meet the demand of human for the speed of resource consumption.In the ultra-deep exploration,the signal strength decreases continuously with the increase of depth,and the signal-to-noise ratio decreases continuously,so that high-resolution inversion data cannot be obtained.Due to the influence of external conditions,construction factors and instruments,the seismic signals received by the seismometer are interfered by the background noise,and the weak signals are drowned by a lot of noise.The strength of seismic data signal is relative.When the amplitude of the effective wave is weak,the energy of the background noise is too strong,the seismic signal is often submerged in the background noise,and the weak signal cannot be recognized directly from the collected data.In the process of seismic data processing,it is expected that the processed seismic data can have three properties: high signal-to-noise ratio,high resolution and high fidelity.In addition,the deep weak signal is very important for the inversion of layered information of underground structure.Therefore,various methods have to be adopted to suppress and attenuate the background noise under the weak signal so that we can improve the signal-to-noise ratio,and extract the deep weak reflected signal.Starting from the mathematical optimization algorithm,this dissertation studies the recovery of seismic deep weak signals by combining the sparse transformation and compressed sensing theory.The specific contents are as follows:Firstly,this dissertation studies the method of suppressing the blind noise of seismic signal based on coherence-based tight frame learning method.Data-driven tight frame method has been widely used in suppressing random noise of seismic data.However,this learning-based method relies on the prior knowledge of noise when removing random noise.We cannot get satisfactory results when processing real seismic data,and generally smmoth result or residual noise.In order to solve this problem,this dissertation introduces the correlation threshold operator on the basis of tight frame learning method,which avoids the need of pre-estimation of noise level,and performs adaptive sparse representation and noise removal for different seismic data blocks.Compared with the traditional tight frame learning method,the coherence-based tight frame learning methodnot only reduces the computation of artificial debugging parameters,but also achieves better noise suppression performance.Secondly,this dissertation studies the reservation of suppressing random noise and weak signal based on group sparse transform learning method.Different from the local sparsity of signals,the method proposed in this dissertation combines the sparsity and selfsimilarity of data.We take advantage of the global self-similarity of data,and represent blocks with highly similar features jointly.From the point of view of patch strategy,we can find exactly the same information in the global details and group them.Theoretically,the sparse representation coefficients of each group in the transform domain are exactly the same.However,due to the existence of noise,the coefficients of each group vary.Therefore,we use group sparse regularization to represent the signals by grouping,and then separate the signal and noise in the specific transform domain.In this way,starting from the self-similarity of signals,the global information of seismic data can be greatly used to enhance the related signals.Compared with the traditional local sparse transform learning method,our proposed methods yield better noise suppression and weak signal retention results.Finally,this dissertation studies the seismic data denoising problem based on a fast dictionary learning method.We adopt the proximal gradient method to solve the optimization problem of dictionary learning,which has low algorithm complexity and ensures the convergence of the algorithm.In terms of seismic data recovery,the sparse constraint denoising problem is divided into several sub-problems.Each sub-problem deals with different data set of non-overlapping blocks.The denoising result is obtained from the average result of the sub-problems.The experimental results show that this method reduces the redundancy.More importantly,this method avoids artifacts and produces better denoising result.