Particle Filtering Based On Empirical Mode Decomposition And Its Application On Random Noise Attenuation For Seismic Exploration

Seismic exploration is an important method to detect underground oil and gasresources. The quality of seismic data is severely reduced by the complexity of theexploration area of geological structure and the diversity of the causes of noise. Itbrings a lot of inconvenience for the extraction of useful information. Especially forstrong random noise, some conventional denoising methods are difficult to achievesatisfactory results. So it is very significant to develop the algorithm in the low SNRwhich can both suppress random noise effectively, and fully retain effective signals soas to improve signal to noise ratio (SNR).Particle Filtering completing the Monte Carlo estimation under the recursiveBayesian framework is a statistical filtering method based signal generation model.Acording to the signal generation model of its own information as a reference, thefiltered signals more easily obtain time-varying signal characteristic information.Furthermore, Bayesian filtering based on the combination of observations andexperiences of information which can make use of a priori and causal knowledge tosolve the conditional probability of the signal under the optimal filtering criteria. So itis firstly introduced particle filtering technology to suppress random noise of seismicdata. However, in practice, the particles represented posterior distribution issusceptible to noise. Especially for exploration data in low SNR, strong random noiseseverely affect the distribution of particles to deviate from true posterior distributionresulting in reducing filtering accuracy. To solve the problem, the particle filteringbased on empirical mode decomposition (EPF) which behaves better resolution oftime and frequency domain is proposed to eliminate random noise of seismic data.In application of random noise attenuation for seismic exploration, firstly, we weconstruct a state-space model of seismic records in order to complete the predictionand update of Bayesian estimation and use varying autoregressive(AR) model todescribe the characteristics of the seismic signal. Secondly, according to the truesituation, we select the importance density function and resampling method to solvethe degradation problem caused by long iteration time. Finally, according to the lowSNR characteristics of seismic data, we combine empirical mode decomposition.Themain idea is: by time iteration of systerm, each particle will develop the time series.We decomposite the sequence by EMD.Then, we obtain different IMF components ofdifferent frequencies. By removing the high frequency component, retaining the lowfrequency close to signal so as to reduce deviations interference from random noiseand reconstructing the particles of left frequency components which are closer to trueposterior distribution, we can improve filtering accuracy. In order to verify the effectiveness the proposed algorithm, PF and EPF areapplied to artificial seismic records. By comparison of time-domain waveform andfrequency-domain spectral before and after denoising,and analyzing the results ofSNR and RMSE evaluated filtered performance, the results show that EPF preservesthe advantages of amplitude preservation and denoising, and improves the signal tonoise ratio to11dB after filtering. At last, the proposed algorithm is applied to realseismic data, the results show that shot records restorated by EPF not only effectivelysuppress random noise but also recover characteristics of valid events of seismic data.