Research On Key Technologies Of High-resolution Processing For Nonstationary Seismic Data

Seismic exploration methods discover the structure of subsurface media by processing the seismic wavefield obtained from the surface,so the quality of seismic data is the key to obtain high-quality images of subsurface media.The ability of using seismic wave response to distinguish subsurface structures and physical properties depends on the resolution of seismic data.Now,thin interbed,lithology,and complex reservoirs have become the key areas in seismic exploration.Conventional seismic data is difficult to meet the requirements of fine characterization for thin sand reservoir etc.,and broadband seismic data acquisition still cannot effectively solve many highresolution processing problems of seismic data.It is necessary to continuously study the high-resolution processing technologies for seismic data.Field seismic data acquisition is affected by many uncertain factors,and because of the limitation of observation conditions,the field seismic data cannot be expressed by deterministic time(or space)function,which is essentially a nonstationary random process and can be defined as nonstationary seismic data.Therefore,it is of great practical value to carry out systematic research on high-resolution processing technologies,especially for nonstationary seismic data.In seismic data processing,inverse Q filtering and deconvolution are two important methods to improve the vertical resolution and broaden the frequency band of prestack and poststack seismic data,which are of great significance to improve the interpretation accuracy of seismic data and thin reservoir identification in oil and gas exploration.In view of the above contents,the dissertation carries out systematic research on the key technologies of high-resolution processing for nonstationary seismic data.For different practical applications,the estimation method of time-varying Q value in time-frequency domain,the efficient adaptive deconvolution method based on streaming prediction-error filter,and the blind sparse spike deconvolution method based on nonstationary convolution model with timevarying Q are proposed.Time-frequency analysis method can provide joint distribution of data in time and frequency domains,which provides a necessary domain for Q estimation.The dissertation introduces the basic theory,advantages,and disadvantages of nine timefrequency analysis methods,including linear,bilinear,and nonlinear time-frequency analysis methods.Then,a new local time-frequency transform based on streaming computing framework is proposed according to the theories of short-time Fourier transform and local time-frequency decomposition,the method is named as streaming local time-frequency transform(SLTFT).In the proposed method,the target signal is matched by Fourier series with time-varying coefficients,which is adaptively updated by streaming computing framework.The new transform can effectively characterize the time-frequency characteristics of nonstationary data.Compared with the local timefrequency decomposition,the proposed method avoids the iterations and effectively saves the computational cost in the calculation process.Compared with the STFT method,the proposed method has the ability of flexible parameter selection and can autonomously select the frequency sampling interval and the frequency range,which only keeps the effective frequency information and saves the data storage.Finally,the time-frequency resolution and focusing performance of the ten time-frequency analysis methods are analyzed by using a synthetic signal.In the field data processing,different research contents and purposes always need to select different time-frequency analysis methods.The systematic analysis and summary of time-frequency analysis methods in this dissertation can provide a theoretical basis for the characterization of timefrequency attributes for nonstationary seismic data,meanwhile,the work provides technical support for the subsequent estimation of the quality factor Q in the timefrequency domain.The inverse Q filtering method can compensate and correct the amplitude attenuation and phase distortion caused by the media absorption and attenuation,thereby improves the resolution of seismic data.The compensation ability of the inverse Q filtering depends on the accuracy of Q value,so the estimation methods for Q value are studied in this dissertation.Three Q-value estimation methods in the frequencydomain or time-frequency domain are introduced.To avoid picking the instantaneous spectra of the top and bottom interface of the target layer,this dissertation proposes a new method for time-varying Q value estimation in the time-frequency domain,named as the local centroid frequency shift method(LCFS).In this method,the local centroid frequencies are defined by using the shaping regularization to constrain the inversion problem in the time-frequency domain,then the time-varying Q values are calculated by combining the local centroid frequencies with the centroid frequency shift(CFS)method.Compared with the traditional method of Q value estimation,the proposed method does not pick the top and bottom interfaces of the target layer,and can also get a relatively reasonable Q value even when the effective frequency spectra are missing.The vertical resolution of nonstationary seismic data can be effectively improved by using the time-varying Q value to implement the inverse Q filtering,which provides a feasible scheme for high-resolution automatic processing.Deconvolution mainly improves the vertical resolution of seismic data by compressing seismic wavelet,which is of great significance to solve the high-resolution processing problems of seismic data in many special scenarios.Prediction-error filter is widely used in seismic deconvolution.To adapt to the nonstationarity of seismic data,iterative or recursive methods can be used to calculate the coefficients of nonstationary prediction-error filter,however,it often leads to problems,such as slow calculation speed and high memory cost.This dissertation proposes an adaptive deconvolution method based on a streaming prediction-error filter.This method adaptively predicts time-varying seismic data by analytically calculating the underdetermined problem under the regularization condition,thereby dynamically compressing wavelets and effectively improving the seismic resolution along the time direction.To avoid the discontinuity of the high-resolution deconvolution results along the space axis,constraints are added to the objective function in both time and space directions to implement multichannel adaptive deconvolution.At the same time,the introduction of time-varying prediction step can make the deconvolution results closer to the relative amplitude relationship of the original data,and it provides high-resolution processing results with relative amplitude preservation,which provides a technical idea for trueamplitude high-resolution processing.According to the characteristics of the streaming computation,the proposed method avoids slow iterations in traditional algorithms.The new deconvolution method not only characterize the nonstationary characteristics of seismic data but also effectively reduces the computational cost,which is suitable for high-resolution processing tasks in large-scale wide-azimuth and high-density acquisition data.Conventional deconvolution method has some limitations in improving the resolution of seismic data,which can only compress seismic wavelet in a limited frequency band,and most of them need to assume that seismic wavelet is minimum phase and reflection coefficient is white noise sequence.Therefore,to avoid the two assumptions and achieve broadband seismic section,sparse constraints are used in reflection coefficient inversion.The traditional sparse spike deconvolutions assume that the seismic wavelet remains stable in wave propagation,but it is not consistent with the field situation.The amplitude,phase,and frequency of seismic wavelet always change in the process of propagation due to the influence of absorption and attenuation of strata,therefore,the traditional sparse spike deconvolutions will cause inaccurate estimation of reflection coefficient when one processes the nonstationary field data and even affects the subsequent reservoir prediction and structural interpretation.To solve this problem,the quality factor Q,which can characterize the absorption and attenuation characteristics,can be introduced into the convolution model to implement the blind sparse spike deconvolution based on the nonstationary convolution model.The method uses the iterative optimization method to estimate seismic wavelet and reflection coefficients simultaneously.This dissertation extends the nonstationary blind sparse spike deconvolution method with time-varying Q value based on the TSMF-Q framework.However,there are two main differences between the proposed method and the TSMF-Q method.One is that the LCFS method proposed in this dissertation is used to estimate the time-varying Q value.The other is to introduce the time-varying Q value attenuation function into the nonstationary convolution model.The synthetic data and field data tests show that the proposed method can reduce the errors from time variability of seismic wavelet and obtain more accurate deconvolution results compared with the traditional blind sparse spike deconvolution.Finally,an improved highresolution processing link for nonstationary seismic data is implemented.