None.
The present invention relates generally to a learning-based approach for fast denoising with applications to fast-track processing. The method is designed to directly work on raw data without separating each noise type and character. The automatic attenuation of noise is attained by performing migration/demigration guided sparse inversion.
Seismic provides unique insights for exploration and development. Improving signal-to-noise ratio is vital in seismic processing to meet the needs of geophysical applications. In general, noise attenuation is performed in many steps of processing, aiming various data domains and methods for iterative reduction. Conventional denoising methods often targeted a single type of noise such as harmonics or ice-break noise (e.g., Li et al., 1995; Gulunay, 2014), or noise with similar characters such as random noise or guided waves (e.g., Soubaras, 1995; Spitzer et al., 2001). A typical workflow in processing starts with identifying dominant noise types and then based on the experience, cascades a series of filtering and inversion tools for attenuation. Depending on the strength and complexity of noise, such a process could require extensive testing in multiple data domains and interaction with other processing sequences, possibly adding weeks or even months of time. A practical fast denoising solution is indeed imperative for fast-track processing with compressed timeline.
In recent times, machine learning has gained increasing popularity in geophysics. It offers a new perspective on seismic data processing. Taking denoising problems for example, one can train a filter or an optimizer to adapt to the local wavefield morphology and obtain clean signal estimates (e.g., Beckouche and Ma, 2014; Mandelli et al., 2019). These methods rely on the learning ability to suppress noise. Li et al. (2018, 2019) introduced a weak signal recovery technique which integrated the conventional denoising methods with dictionary learning and inversion. The versatility was illustrated through examples of attenuating both incoherent and coherent noises on land. We here further expand the technique to develop a fast denoising approach using migration and demigration. The premise is that migration could focus the seismic signal into coherent events, and meanwhile, defocus the noise. Given sufficient fold, the stacking power shall further increase the signal-to-noise ratio. Demigrated gathers from the stacked image then serve as a training set to guide noise removal from raw data. The approach is intended to substitute manual and iterative process with algorithmic computation, reducing time and effort associated with designing complex noise attenuation workflow. We discuss examples from Permian Basin to show the impacts on fast-track and production processing.
Noise is an unavoidable issue in seismic acquisition and processing. In particular with modern land acquisition design trending towards point source, point sensor and high trace density, the raw seismic quality would be further compromised due to the lack of stacking power for individual traces. For decades, many authors have put forward their solutions to cope with noise. Nevertheless, these methods often targeted a single type of noise, such as harmonics or ice-break noise, or noise with similar characters, such as random noise or guided waves. In processing, seismic denoising is often performed in data domains and incorporated into multiple steps to account for different noise characters. Even though most denoising methods are less computationally intensive comparing to velocity model building and imaging, the procedure could be still time consuming with iterations of manual testing, quality control and parameterization.
Leveraging migration and demigration, here we propose a learning-based approach for fast denoising with applications to fast-track processing. The method is designed to directly work on raw data without separating each noise type and character. The automatic attenuation of noise is attained by performing migration/demigration guided sparse inversion. By discussing examples from a Permian Basin dataset with very challenging noise issues, we attest the feasibility of this learning-based approach as a fast turnaround alternative to conventional denoising methodology.
The invention more particularly includes a process for automatically suppressing noise in raw seismic data (d) by estimating the reference image ({tilde over (m)}) using a seismic image processing technique selected from automatic gain control, migration and stacking, or other imaging techniques; demigrating the data from the reference image ({tilde over (m)}) and building a dictionary from the demigrated data by unsupervised learning; solving a dual-domain sparse inversion problem to simultaneously invert signal and noise and obtain a denoised copy of the raw input data (d).
A process for imaging a subterranean reservoir under a survey area; designing a seismic survey to image the subterranean formations; acquiring raw data using seismic sources and receivers; obtaining raw data (d) from a seismic survey; estimating the stacked image ({tilde over (m)}); demigrating the data from the stacked image ({tilde over (m)}) to attain a clean reference; building a reference from the demigrated data; solving an unsupervised dictionary learning problem; resolving a dual-domain sparse inversion to leverage sparsity and guide attenuating noise of the raw input data (d); inverting simultaneously the dual-domain sparse inversion to achieve a successful reduction or more and stronger noise; and comparing a direct demigration to the learned inversion to obtain denoised data with a proper amplitude versus offset.
A process for producing hydrocarbons from a subterranean reservoir selecting a subterranean formation for hydrocarbon production; designing a seismic survey to image the subterranean formation; acquiring raw data using seismic sources and receivers; obtaining raw data (d) from a seismic survey; estimating the stacked image ({tilde over (m)}) with mitigated artifacts; demigrating the data from the stacked image ({tilde over (m)}) to attain a clean reference; building a reference from the demigrated data; solving an unsupervised dictionary learning problem; resolving a dual-domain sparse inversion to leverage sparsity and guide attenuating noise of the raw input data (d); inverting simultaneously the dual-domain sparse inversion to reduce noise; comparing a direct demigration to the learned inversion to obtain denoised data with a proper amplitude versus offset; and constructing a well in said subterranean reservoir to produce hydrocarbons.
In some embodiments the subterranean reservoir is an oil reservoir, an unconventional reservoir, heavy oil reservoir, gas reservoir, coal-bed-methane reservoir, or a source rock.
Seismic data noise may include many different kinds coherent and incoherent noise from harmonics, random noise, guided waves, and combinations thereof.
The stacked image may be estimated by
{tilde over (m)}=LTWd,
where LT denotes the migration operator and W applies a filter.
Mitigated artifacts may be filtered with an automatic gain control (AGC) and other noise control techniques.
The denoised data may be demigrated from the image estimate ({tilde over (m)}) with the velocity model
{tilde over (d)}=L{tilde over (m)}=LLTWd,
where {tilde over (d)} share the same source/receiver locations, kinematics and spectrums as input data d, but with less noise.
The unsupervised dictionary learning problem may be
where the dictionary D and sparse coefficients x are resulted by a sparse linear decomposition of the demigrated data , and T is the sparsity threshold.
The trained dictionary (D) problem is used to guide attenuating noise on the raw input data (d):
where x1 and x2 are the sparse representations of signal and noise, respectively, S is the sparsity basis for noise, and a and E are the tradeoff parameter and error tolerance.
As used herein, equations include the equation as written, rearranged, or substituted with equivalent variables such that {tilde over (m)}=LTWd is equivalent to LT={tilde over (m)}(Wd)−1 is equivalent to Wd={tilde over (m)}/LT. Other equations may be similarly manipulated to solve for different variables or substitute equivalent terms.
The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee. A more complete understanding of the present invention and benefits thereof may be acquired by referring to the follow description taken in conjunction with the accompanying drawings.
Turning now to the detailed description of the preferred arrangement or arrangements of the present invention, it should be understood that the inventive features and concepts may be manifested in other arrangements and that the scope of the invention is not limited to the embodiments described or illustrated. The scope of the invention is intended only to be limited by the scope of the claims that follow.
Pre-stack migration and stacking is a natural means to suppress noise in data. Given noisy data i, one can quickly estimate the stacked image {tilde over (m)}by
{tilde over (m)}=LTWd,
where LT denotes the migration operator and w applies automatic gain control (AGC) to the input data. The goal of AGC is to mitigate artifacts from large-amplitude noise during migration. Though it could be substituted by other filtering methods, AGC is deemed to be appropriate due to its simple parameterization and robustness.
Demigration can be regarded as the adjoint operator of migration, which has been widely used to implement least-squares migration (Schuster, 1993). To attain a clean reference, we demigrate the data from the image estimate with the same velocity model, i.e.,
{tilde over (d)}=L{tilde over (m)}=LLTWd,
With careful implementation, we ensure the demigrated data {tilde over (d)} share the same source/receiver locations, kinematics and spectrums as input data d, but with little noise. Since demigration is performed on the stacked image with AGC applied, {tilde over (d)} cannot maintain the true amplitude and correct amplitude versus offset (AVO) response, both of which are critical for subsequent quantitative analysis. In order to preserve the amplitude of primaries in denoising, we propose to build a reference from the demigrated data and then count on learning and inversion to recover the signal.
We exploit the spatial and temporal redundancies in the demigrated data by solving an unsupervised dictionary learning problem, i.e.,
Here the dictionary D and sparse coefficients x are resulted by a sparse linear decomposition of the demigrated data {tilde over (d)}, and T is the sparsity threshold. The premise of sparsity holds in general since seismic data are, by nature, five-dimensional measurements of three-dimensional subsurface.
The trained dictionary D is next used to guide attenuating noise on the raw input d. Specifically, we aim to resolve a dual-domain sparse inversion problem which leverages the sparsity again and simultaneously inverts for target signal and noise. That is to solve
where x1 and x2 are the sparse representations of signal and noise, respectively, S is the sparsity basis for noise, and a and E are the tradeoff parameter and error tolerance. Note that S can be defined using either non-adaptive transforms or a similar learning model to train from noise. By inverting simultaneously, we lift the problem into a higher dimension which permits a successful reduction with much stronger noise. Solving both optimization problems are non-trivial but omitted here due to limited space (Aharon et al., 2006; Li et al., 2013). The denoised data from d is given by
dout=Dx⊥,
Comparing to a direct demigration, the result of learning and inversion adapts to the true amplitude of signal, giving rise to proper AVO responses. The workflow of the proposed approach is summarized in
The following examples of certain embodiments of the invention are given. Each example is provided by way of explanation of the invention, one of many embodiments of the invention, and the following examples should not be read to limit, or define, the scope of the invention.
In 2018, we conducted a high-density acquisition over a producing field in the Permian Basin, in an effort to increase image quality for unconventional reservoir development. The survey was acquired using point source, point receiver in conjunction with CSI technology (Mosher et al., 2017) to maximize the operation efficiency. The field is known as a difficult seismic data area—salt dissolution in near surfaces causes strong back-scattering noise. In addition, increased amount of surface activities further deteriorates the noise issue.
In parallel with production processing using conventional denoising means and cascaded workflow, we ran a test using the above approach. The raw data after applying AGC were directly migrated using Kirchhoff prestack depth migration with a reasonably good velocity model. Then Kirchhoff demigration (Zhang et al., 2002) was employed to build the reference model in the time domain.
We compared the result to the one obtained from production fast-track processing at a similar stage. The production denoising workflow is composed of several cascaded steps, mainly targeting narrow-band noise bursts, spikes, ground roll and scattering noise. Because of the high complexity of the noise in the area, establishing such a workflow using conventional methods involved iterations of manual testing. In the end, we spent about 5 weeks to clean up the data to a quality that met the requirements of fast-track processing. On the other hand, thanks to the data-driven nature of the proposed approach, the fast denoising was completed in 5 days with minimal manual efforts.
We further examined the data by imaging. We ran Kirchhoff prestack depth migration on all three datasets shown in
Besides the usage of fast denoising, the proposed approach can also be adapted to clean up gathers at various stages of production processing. For instance, by means of omitting AGC and inserting a remigration step after demigration in the workflow shown in
Traditionally, it has been a time consuming and manually intensive task to come up with an effective denoising workflow for the areas affected by strong complex noise. Leveraging migration and demigration, we put forward a fast denoising approach which turns noise attenuation problem into learning and inversion. It is applicable to a direct attenuation of various noises in the raw data with a unified approach and little tuning. Examples from Permian Basin proved that it could achieve similar or higher quality comparing to conventional methodologies commonly used in seismic processing, with a reduced timeline from weeks to days. This approach can also be adapted to clean up gathers before or after migration to facilitate production processing.
In closing, it should be noted that the discussion of any reference is not an admission that it is prior art to the present invention, especially any reference that may have a publication date after the priority date of this application. At the same time, each and every claim below is hereby incorporated into this detailed description or specification as a additional embodiments of the present invention.
Although the systems and processes described herein have been described in detail, it should be understood that various changes, substitutions, and alterations can be made without departing from the spirit and scope of the invention as defined by the following claims. Those skilled in the art may be able to study the preferred embodiments and identify other ways to practice the invention that are not exactly as described herein. It is the intent of the inventors that variations and equivalents of the invention are within the scope of the claims while the description, abstract and drawings are not to be used to limit the scope of the invention. The invention is specifically intended to be as broad as the claims below and their equivalents.
All of the references cited herein are expressly incorporated by reference. The discussion of any reference is not an admission that it is prior art to the present invention, especially any reference that may have a publication data after the priority date of this application. Incorporated references are listed again here for convenience:
This application is a non-provisional application which claims benefit under 35 USC § 119(e) to U.S. Provisional Application Ser. No. 63/142,626 filed Jan. 28, 2021, entitled “SEISMIC DENOISING,” which is incorporated herein in its entirety.
Number | Date | Country | |
---|---|---|---|
63142626 | Jan 2021 | US |