The Application Of LS-SVM Model On Land Seismic Random Noise Modeling And Its Attenuation

Land seismic exploration is a normal method to explore oil and gas resourcesnowadays. During the real seismic records acquisition, the signal must be influencedby various kinds of noise, which heavily reduced the signal-to-noise ratio of realseismic data. And it is not conducive to analyze the real seismic data.The characteristics of seismic random noise are influenced by the surfacevegetation and the landform characteristics of seismic regions. In explorationseismology, the decreasing of gas resources which are easily explored pushes theexploration targets to deep exploration, shallow exploration and lithologicalexploration. It leads the random noise in real seismic data is more complex. As amain noise of real seismic data, attenuating the ambient noise effectively becomes akey technology in the filed of seismic data processing, in order to meet therequirement of a high signal to noise ratio (SNR).The LS-SVM(Least square support vector machine) theory has been always ahotspot since it’s put forward in the field of machine learning. It’s widely used in thestudy of non-linear modeling because it has less parameters, stronger generalizationability to learn and higher modeling precision. What’s more, the LS-SVM modelalso has a broad application prospect in the field of soft measurement, medicaldiagnosis, pattern recognition and so on.In this paper, the LS-SVM regression model of real seismic random noise in theregions of mountain, desert and loess tableland of China are mainly studied. First ofall, the Duffing chaotic time series are adopted to prove that the LS-SVM regressionmodel can model chaotic series effectively. And we analyze the influence of the errorbetween the real data and the predict data by different values of two parameters, thepunishment factor and the width of the RBF kernel function. Then the LS-SVMmodel is applied to model the real seismic random noise which is proved to bechaotic. In the experiment, different lengths of random noise of different traces aremodeled and predicted in different regions. The results prove that the LS-SVM has agood ability to nonlinear filter. What’s more, the values of RMSE and NRMSEbetween the real seismic noise data and the predict data are all low, which prove that the LS-SVM model has a high predict precision. Based on these results, in this paperwe adopt the dividing directly measure to attenuate the random noise which isproved to be chaotic in the seismic records.To prove the feasibility of the approach proposed in this paper, an artificialseismic record of40traces is simulated and the real seismic random noise from thedesert common record is added into it. The experiment divides into two steps. First,the noise data of each trace are modeled by the LS-SVM model and predicted therest data. Then, in the noise seismic record the dividing directly measure is adoptedto attenuate the random noise. By analyzing the time domain waveform and thefrequency spectrum of a single seismic before and after the random noise attenuation,the approach proposed in this paper is effective. It offers a new way to attenuateseismic random noise.