Research On Low-frequency Noise Suppression For Desert Seismic Data Based On Wide Inference Network

As an extremely important non renewable resource,oil and gas have been closely linked with the development of the world economy,which is an important thrust of the development of the world economy.At present,the conventional oil and gas resources distributed in the shal ow and easy to exploit areas are gradually facing depletion.The unconventional oil and gas resources which are difficult to exploit in the area with complex deep underground structure have gradually become the focus of exploratio n.There are many desert areas in Northwest China,which often contain a large number of unexplored and undeveloped oil and gas resources.The geological conditions and survey conditions in desert area are relatively bad,and the noise is complex.Lowfrequency noise is one of the most common noise types in desert seismic exploration records.Low-frequency noise seems to be simple,but its statistical characteristics and mathematical representation are very complex,with large energy and strong randomness.Low-frequency noise and effective signal will overlap in frequency domain,showing weak similarity.This makes the difference between low-frequenc y noise and effective signal in space-time domain reduce.High quality seismic exploration data is the basis of formation structure imaging and interpretation,so we need to find an effective data processing strategy for desert seismic exploration.At present,most of the traditional methods are in the transform domain,through the energy,frequency or other physical quantities to distinguish the effective signal and noise,there are good examples in the traditional seismic exploration data processing.However,in the face of the complex low-frequency noise in desert seismic data,these traditional algorithms show some problems,such as incomplete noise suppression,difficult to identify effective signals and so on.Based on the above problems,we propose a method of noise suppression for desert seismic exploration data based on Wide Inference Network(WIN).First,we set up a five layer network model,each layer contains convolution processing,batch normalization processing and activatio n function.Then we build a high-quality training set,and add it to our designed network model for training.In the process of network training,we introduce residual learning to minimize the error between the predicted value and the target value,and finally get a set of optimal parameters,such as weight and bias.WIN is based on big data,using the optimal network model parameters obtained by training to get the feature informat io n of low-frequency noise and effective signal from large seismic exploration data,so as to separate low-frequency noise and effective signal.In the data processing of desert seismic exploration,we verify the feasibility and effectiveness of the proposed method through a large number of comparative tests.The method proposed in this paper establishes an optimal nonlinear mapping model between noisy records and effective signals.The trained network model is a data-driven automatic de-noising device without manual adjustment of parameters.Both the overall denoising effect and the single channel contrast image in time domain are better than the traditional methods.Based on the quantitative analysis of data,especially in the case of low SNR,win can improve the SNR from-13.7dB to 3.71 dB in the simulat io n experiment.At the same time,we compare WIN with 1)- deconvolut io n,Variational Mode Decomposition(VMD)and Shearlet transform denoising results.WIN has advantages in building nonlinear mapping model between low-frequenc y noise and effective signal,which is obviously more suitable for desert seismic exploration data.