NO343015B1 - Method for denoising seismic data from a seafloor array - Google Patents

Abstract

A method for denoising seismic data from a seafloor array with several hydrophones and several seismic receivers, each seismic receiver comprising three mutually orthogonal motion sensors. The method comprises the steps of: transforming (110, 115) recorded motion data (111) from the seismic receivers and recorded hydrophone data (116) from the hydrophones into a transform domain where acoustic sources are separated; and matching (120) transformed hydrophone data (117) to transformed motion data (112), thereby obtaining matched hydrophone data (121). The separated sources enable more reliable automatic filtering and matching than similar processing in the TX domain. The method further comprises the steps of: performing an inverse transform (130) of the matched hydrophone data (121), thereby obtaining matched P-wave data (131) in the TX domain; and subtracting the matched P-wave data (131) from the recorded motion data (111), thereby obtaining enhanced motion data (141) in the TX domain. Coherent noise transferred through the seafloor can be removed in a later step.

Description

BACKGROUND

Field of the invention

[0001] The present invention concerns a method for denoising seismic data from a seafloor array.

Prior and related art

[0002] On different scales, earthquakes and microseismic events release acoustic energy that travel through the Earth's crust. An array of seismic sensors records the resulting waves. Later processing reveal, for example, the epicentre of an earthquake, location of a microseismic event as well as the size and type of an event. The seismic waves also carry information about the structure of the underground. An earthquake generates low frequency, long wavelength seismic waves that reveal structures on a global scale and/or down to the molten core of the Earth, whereas a microseismic event generates waves with shorter wavelengths that reveal finer structures closer to the solid surface. The seismic sensors may be different on different scales, but the techniques used for processing are similar.

[0003] In contrast to the passive monitoring above, active seismic acquisition involves a manmade acoustic source. In marine seismic acquisition, the acoustic source is usually an array of airguns. The acoustic energy penetrates into the underground, where boundaries between layers with different densities reflect and refract the acoustic waves. In some instances, a survey vessel towing an array of hydrophone cables at the sea surface samples the resulting wavefield. However, towing an array of streamers is sometimes impossible or impractical, for example near rigs and other installations on a production field or in shallow waters in a transition zone between dry land and the ocean. In these and similar cases, active seismic acquisition involve deploying ocean bottom cables or seismic sensors permanently or temporarily over the field to be surveyed.

[0004] In somewhat greater detail, body waves are seismic waves able to propagate through solids, and comprise pressure waves (P-waves) and shear waves (S-waves). In P-waves, particle motion is parallel to the direction of propagation. S-waves have particle motion perpendicular to the direction of propagation. At boundaries between different materials, mode conversion means that a fraction of the body waves shift mode from P-waves to S-waves and vice versa. Some of the acoustic energy is also converted to boundary waves. Scholte waves are boundary waves traveling along the seafloor, i.e. the boundary between a body of water and the sediments or rock beneath. Boundary waves are typically large amplitude, low frequency waves that may mask events, especially from deeper layers.

[0005] As used herein, a hydrophone is any seismic sensor detecting pressure and a motion sensor is any seismic sensor detecting particle motion, e.g. a MEMS-accelerometer or a geophone. We reserve the term 'seismic receiver' for any device that comprises three mutually orthogonal motion sensors. The seismic receiver, i.e. orthogonal motion sensors, detects both P-waves and S-waves, but a hydrophone does not detect S-waves. A 4C (four component) receiver contains one hydrophone and three orthogonal motion sensors. Design of seismic sensors as such is not part of the invention and need no further description herein.

[0006] Also as used herein, a 'seafloor array' is any array with several hydrophones and several seismic receivers temporarily or permanently deployed on a seafloor, regardless of whether the seismic sensors are designed for passive or active acquisition or both.

[0007] S-waves do not travel through fluids, so hydrophones are needed to detect P-waves in towed streamers. For the same reason, motion sensors in close contact with the seafloor are needed to detect S-waves. Some towed streamers contain motion sensors to distinguish upgoing wavefields of interest from downgoing wavefields due to reflections from the sea surface. In active surveying, removal of multiple reflections is known as 'deghosting'.

Amundsen and Reitan (1995) [1] describes a similar use of multicomponent seafloor data.

[0008] The desired P and S-wave signals are usually weak compared to ambient noise, i.e. the signal to noise ratio (SNR) is usually low. In microseismic monitoring, adding power to the signal is impossible so noise reduction is important to improve the SNR. Some noise may be attenuated by deploying the receiver array in trenches, e.g. in sediments on the seafloor. Incoherent noise can be removed by summation or averaging techniques in which coherent signals add and incoherent (random) signals cancel. Rigs and subsea installations may contribute with low frequency waves, typically less than 15-20 Hz, traveling along the seafloor. Like the Scholte waves, these waves may mask low frequency events from deeper layers. A significant fraction of the remaining noise is P-waves traveling through the water from various sources such as vessels and nearby installations. Thresholding or logarithmic averaging may exclude a few such coherent sources because they are orders of magnitude greater than the signals of interest. For instance, propeller noise in shallow waters may fall into this category. Band-stop filters may exclude a few more, e.g. water borne rig noise below 20Hz. However, band-stop filters also remove signals of potential interest. Thus, there is a need to remove coherent noise from various sources, in particular water borne noise from fixed sources such as rigs, subsea installations, pumps etc.

[0009] A few decades ago, computing resources were more expensive than they are today. Then, a human such as a geophysicist removed outliers corresponding to defect devices manually from common source or common receiver gathers. Filtering and muting in a transform domain were frequently used to save processor time and computer memory.

Degraded data quality and artefacts introduced by the forward and inverse transforms were considered acceptable, and a geophysicist was required to interpret the results during data processing.

[0010] The widely used Fourier transform may serve as an example. A digitised time series may be transformed efficiently into a frequency domain by the DFT (Discrete Fourier Transform) and the FFT (Fast Fourier Transform) algorithm. A spatial equivalent maps X-data to wavenumbers k, and a two-dimensional Fourier transform maps TX-signals into an FK domain. Convolutions in the TX-domain involves integration of a signal multiplied by a kernel, here a sinusoid. In the FK-domain, the convolution reduces to relatively simple multiplications, thereby facilitating deconvolution. However, the transform introduces challenges. For example, three different events in TX may map to a wedge shaped region flanked by two slanted lines in the FK domain. Removing one of the slanted lines without removing part of the wedge shaped region is difficult without human supervision.

[0011] A second example concerns Radon transforms, which were first described in 1917 by Johann Radon as a purely mathematical concept. An early practical application of these transforms was in computer tomography, the development of which was awarded the 1979 Nobel prize in medicine. Diebold and Stoffa (1981) [2] introduced Radon transforms in seismic applications. Their seminal article describes seismic travel time data in terms of an intercept time τ and instantaneous slope or slowness p = dt/dx = 1/v, where v is the phase velocity or speed of sound. A Radon transform using a linear kernel t = τ px, the τ-p transform, maps 'straight' events such as direct arrivals and head waves to points. Hyperbolic seismic events maps to ellipses in the τ-p domain. For reflections, sums of such ellipses form τ-p trajectories, and all turning rays (whether refracted or postcritically reflected) combine to form an easily identified trajectory. A velocity‐depth function can be derived from such a trajectory from reflected or refracted arrivals.

[0012] A related transform, τ-q, uses a parabolic kernel t = τ qx<2>. The τ-q transform may remove residual moveout and perform other tasks in seismic acquisition and processing. Radon transforms with hyperbolic and other kernels are also well known and studied.

[0013] Current computer technology enables matching and filtering by statistical methods, e.g. a Wiener filter computing cross-correlation and auto-correlation in large datasets. Still, computers lack the human brain's capability of recognising patterns. Hence, geophysicists are still involved in standard seismic processing, e.g. to identify τ-p trajectories in a Radon panel or decide if a residual moveout correction or τ-q transform is required. However, general advances in electronics, signal processing etc. also enable increasingly large datasets. Human picking in large datasets is often more prone to errors than comparable automatic algorithms, and human supervision may even be impossible in real-time applications.

[0014] US 7433265 concerns converted wave energy removal from seismic data. Coherent wave noise energy is removed data by modelling both the P-wave primary energy and the coherent wave noise energy. The P-wave primary energy is modelled first and then subtracted from the input data. The data with the P-wave primary energy removed is used as the input for coherent wave energy removal. The coherent wave energy is modelled and subtracted from the original input data, i.e. the data input into P-wave primary removal. This leaves a dataset with P-wave primary energy and noise energy not related to coherent waves. This method can be utilised to remove all types of coherent noise with a velocity difference to the desired P-wave primary energy or with a different type of moveout (change of time arrival with sourcereceiver distance).

[0015] The method in US 7433265 depends on extensive modelling and multiple transforms, all of which potentially degrade the result due to approximations. Moreover, extensive modelling, multiple transforms and numerical filtering require considerable computer resources. A more accurate and less demanding algorithm is desirable for an automatic real-time application.

[0016] Poole et al. (2012) [3] regards ocean-bottom seismic data processing. According to the authors, the pressure component P is in general of good quality, while high levels of noise are often observed on the vertical component Z. Nonetheless, Z is needed to achieve complete pre-stack wavefield separation and also to drive processes such as mirror imaging and updown deconvolution. The authors propose a new method to address this problem, based on upgoing and downgoing wavefield properties and pressure-particle velocity relationships. The sparse τ-p domain plays a key role in this 3D method, allowing for better signal-noise discrimination than other domains. Application of the proposed algorithm to a deep-water ocean-bottom survey attenuates noise while preserving amplitudes.

[0017] A main objective of the present invention is to provide a method for real-time denoising of seismic data while retaining the benefits of prior art.

[0018] US 2013182533 A1 and US 2011213556 A1 disclose a method for denoising seismic data from several hydrophones and several seismic receivers, each seismic receiver comprising three mutually orthogonal motion sensors.

[0019] Marine seismic mapping with multiple seismic sensors, and a noise cancellation method is known from US 2014278118 (A1).

[0020] The following three patents relate to seismic mapping with hydrophones and geophones on the seabed. GB 2539097 (A) describes data processing including separation of wave fields, transformation domains; images are generated by fusion of data. US 2013107664 (A1) discloses a noise cancellation process. US 2013021873 (A1) discloses a method for data processing includes separation of wave fields, transformation of data sets, and combination and calibration of data sets, inverse transformation and eventually subtraction of wavefield data to generate estimates of the separated wave fields separately.

SUMMARY OF THE INVENTION

[0021] This and other objectives are achieved by the method according to claim 1. Further features and benefits appear in the dependent claims.

[0022] More particularly, the invention concerns a method for denoising seismic data from several hydrophones and several seismic receivers comprising transforming recorded motion data from the seismic receivers and recorded hydrophone data from the hydrophones, the method is performed in real time and in a TX domain, the seismic data comprising S-wave signals, P-wave signals and waterborne noise, each seismic receiver comprising three mutually orthogonal motion sensors, the method further comprising the steps of:

- the recorded motion data and the recorded hydrophone data from the hydrophones are transformed into a transform domain involving a τ-p transform

- where acoustic sources are separated; and

- matching transformed hydrophone data by involving a filter such as a Wiener filter to transformed motion data, thereby obtaining matched hydrophone data

[0023] Recorded data are one trace per hydrophone and seismic receiver in the TX domain. The traces may represent responses to a shot or an arbitrary time interval in a microseismic monitoring application. Integrating acceleration or differentiating displacement with respect to time is trivial. Contrary to traditional noise filtering and muting, the purpose of the present transform is to separate acoustic sources, not saving CPU-time or computer memory. An unsupervised computer algorithm is likely to detect the separated sources in the transform domain, whereas filtering in the TX domain may require human supervision for reliable results. Hence, the separate sources facilitate reliable and automatic matching and filtering. Current processors or clusters are likely to handle the transform and filtering in near real-time.

[0024] A currently preferred method further comprises the steps of:

- performing an inverse transform of the matched hydrophone data, thereby obtaining matched P-wave data in the TX domain; and

- subtracting the matched P-wave data from the recorded motion data, thereby obtaining enhanced motion data in the TX domain.

[0025] The enhanced motion data comprises the S-wave and P-wave signals detected by the motion sensors. The P-wave signals detected by the hydrophones contain water-borne noise and are removed. Further processing of the enhanced motion data in the TX-domain is of course possible.

[0026] An alternative method comprises the step of subtracting the matched hydrophone data from the transformed motion data, thereby obtaining enhanced motion data in the transform domain. This alternative permits further processing, e.g. filtering and/or muting, in the transform domain. This step is usually followed by an inverse transform to obtain enhanced motion data in the TX domain as in the currently preferred embodiment.

[0027] In both alternatives, transforming may involve a τ-p transform. This transform separates sources based on relative arrival time τ as well as slope p = dt/dx.

[0028] Matching may involve a Wiener-filter. This does not require any assumptions about the shape or size of source representation in the transform domain, e.g. τ-p trajectories.

[0029] The method above removes P-wave noise detected by the hydrophones, but not coherent waves traveling along the seafloor such as Scholte waves or rig-noise not carried through the water. This coherent noise may be identified and removed from the enhanced motion data by conventional means, e.g. by autocorrelation techniques.

[0030] As the method above does not require human supervision, implementing a near realtime system is mainly a matter of scaling. In method terms, this is expressed as steps to obtain the enhanced motion data in the TX domain within, for example, one second from end of recording the motion data and hydrophone data. Such steps may include parallelisation, increasing processing power etc. as known in the art.

BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The invention will be explained with reference to exemplary embodiments and the accompanying drawings, in which

Fig. 1 illustrates a method according to the invention

DETAILED DESCRIPTION OF A PREFERRED EMBODIMENT

[0032] The drawings are schematic and not necessarily to scale. Numerous details known to the skilled person are omitted.

[0033] Figure 1 illustrates a method 100 according to the invention. In the present example, a seafloor array comprises 32 hydrophones and 32 seismic receivers as defined in the introduction. In the present example, the mutually orthogonal motion sensors are MEMS-accelerometers. Input data 111, 116 in the time-space (TX) domain are traces lasting 0.8 seconds, and may represent an arbitrary interval in a passive monitoring application or responses to a shot during active acquisition. The traces are aligned such that the significant amplitudes of an event appear along straight lines. The alignment of traces is fixed because the receiver positions are fixed on the seafloor.

[0034] MEMS data 111 in the TX domain comprise S-wave signals, P-wave signals and water borne noise from various sources. The significant amplitudes appear from about 0.2 to about 0.5 s along the T-axis. This time interval grows linearly with the number of traces.

[0035] In step 110, a forward τ-p transform maps MEMS-data 111 to a τ-p domain. As seen in panel 112, the sources are separated along the τ and p axes. The separation facilitates automatic processing at the cost of a τ-p transform and its inverse.

[0036] Specifically, the human brain recognises patterns in a way computers cannot mimic. In the MEMS-data 111, a human easily identifies 'lines' of significant amplitudes, each corresponding to an acoustic source. The human analyst would also identify and remove any outlier breaking the pattern. In contrast, a computer algorithm must break down the problem and, for example, assign amplitudes to a dataset before fitting a line through the dataset using linear regression. In the TX-data 111, the difficulties of assigning amplitudes to separate datasets grow as the number of closely spaced and crossing 'lines' increase. In the τ-p data 112, the straight 'lines' from input data 111 appear as short trajectories. A computer algorithm can identify these short curves more rapidly and reliably than the 'lines' in a TX-domain.

[0037] Assuming, for example, that each significant trajectory extends over at least three traces, only 1/3 of the traces need a scan over the entire time interval. Moreover, each trajectory extends for a short time compared to the length of a trace. Thus, once a computer algorithm has identified one significant amplitude, a search for nearby significant amplitudes can be limited to a small patch in the transform domain, at least in theory.

[0038] Step 115 maps TX pressure data 116 from hydrophones to τ-p data 117. The differences between step 115 and step 110 are due to differences in the sensors, e.g. scaling, sampling intervals etc.

[0039] Step 120 involves matching the hydrophone data to MEMS-data in the τ-p domain. In a currently preferred embodiment, the matching filter is a Wiener filter. The resulting dataset 121 comprises two matched trajectories that represent separate acoustic sources. The two encircled trajectories in panel 121 correspond to two similar curves in hydrophone data 117 and two of three curves in τ-p MEMS-data 112.

[0040] Step 130 performs an inverse τ-p transform on the dataset 121 to obtain matched hydrophone data in the TX domain. The resulting dataset 131 shows a cross of two intersecting 'legs' of significant amplitudes. Similar 'crosses' appear in the input data 111 from the MEMS-accelerometers and in the input data 116 from the hydrophones.

[0041] Step 140 subtracts the matched hydrophone data 131 from the MEMS-data 111. The resulting dataset 141 comprises a strong 'line' of amplitudes corresponding to S-wave data in the input 111 and a weaker 'line' corresponding to one 'leg' of the 'cross' appearing in TX panels 111, 116 and 131. The weaker 'line' corresponds to P-wave data recorded by the MEMS-accelerometers that are not recorded by the hydrophones.

[0042] In an alternative embodiment, the matched hydrophone data in panel 121 could be subtracted from the transformed MEMS-data in panel 112 in the τ-p domain. This would leave enhanced MEMS-data in the τ-p domain, possibly for further processing, before an inverse transform yields enhanced MEMS-data 141 in the TX-domain.

[0043] The matched hydrophone data 131 look similar to the original hydrophone data 116 with some noise added due to approximations in the forward and inverse τ-p transforms. This is similar to processing a few decades ago, when filtering and muting in a transform domain was required due to limited computer resources, and slightly degraded data were unavoidable. As noted in the introduction, current processors may filter and match the hydrophone data in the TX-domain during standard processing. In contrast, we propose using the advances in computer technology to perform relatively demanding operations such as a τ-p transform and Wiener filtering to improve reliability in an automatic real-time application. In particular, the present invention

- use a transform for separating acoustic sources, not noise,

- matches sources in the transform domain to improve reliability in a process without human supervision and

- runs in time intervals far shorter than those associated with standard processing involving human geophysicists or seismic analysts.

[0044] In short, the transforms 110, 115, 130 have a different purpose and the matching in step 120 is different from traditional filtering and muting of noise in a transform domain. Of course, step 120 may include such traditional filtering and muting. Furthermore, the proposed method merely removes water borne noise detected by hydrophones. Coherent noise such as Scholte waves and rig noise transferred through the seafloor can be removed in a later step.

[0045] The proposed method works with any seismic motion sensors, and is not limited to MEMS accelerometers. Furthermore, any transform capable of separating acoustic sources may replace the τ-p transform. The matching filter in step 120 matches sources, but does not apply any noise filtering or muting in the transform domain. Any known matching filter may replace or enhance the Wiener filter in step 120.

[0046] Finally, we remark that an S-wave source on the seafloor may further improve the desired S-wave signal in the enhanced MEMS-data 141, or in general enhanced data 141 from any motion sensor as defined above.

[0047] While the invention has been described with reference to specific examples, the full scope of the invention is defined in the appended claims.

References

[1] Amundsen L. and Reitan A.: "Decomposition of multicomponent sea-floor data into upgoing and downgoing P- and S-waves", Geophysics 60(2), March 1995, pp. 563-572

[2] Diebold, John B and Stoffa, Paul L: "The traveltime equation, tau-p mapping, and inversion of common midpoint data", Geophysics 46(3), March 1981, pp. 238-254

[3] Poole et al., "Sparse τ-p Z-Noise Attenuation for Ocean-Bottom Data", 2012 SEG Annual Meeting, 4-9 November 2012, Las Vegas, Nevada

Claims (9)

Claims

1. A method for denoising seismic data from several hydrophones and several seismic receivers comprising transforming (110, 115) recorded motion data (111) from the seismic receivers and recorded hydrophone data (116) from the hydrophones, characterized in that the method is performed in real time and in a TX domain, the seismic data comprising S-wave signals, P-wave signals and waterborne noise, each seismic receiver comprising three mutually orthogonal motion sensors,

the method further comprising the steps of:

- the recorded motion data (111) and the recorded hydrophone data (116) from the hydrophones are transformed into a transform domain involving a τ-p transform - where acoustic sources are separated; and

- matching (120) transformed hydrophone data (117) by involving a filter such as a Wiener filter to transformed motion data (112), thereby obtaining matched hydrophone data (121) comprising two separate acoustic sources.

2. The method according to claim 1, further comprising the steps of:

- performing an inverse transform (130) of the matched hydrophone data (121), thereby obtaining matched P-wave data (131) in the TX domain; and

- subtracting the matched P-wave data (131) from the recorded motion data (111), thereby obtaining enhanced motion data (141) in the TX domain.

3. The method according to claim 1, further comprising the step of:

- subtracting the matched hydrophone data (121) from the transformed motion data (112) to obtain enhanced motion data in the transform domain.

4. The method according to claim 3, further comprising the step of:

- performing an inverse transform (130) on the enhanced motion data in the transform domain to obtain enhanced motion data (141) in the TX domain.

5. The method according to any preceding claim, wherein matching involves a Wienerfilter.

6. The method according to any preceding claim, further comprising filtering and/or muting in the TX domain.

7. The method according to any preceding claim, further comprising filtering and/or muting in the transform domain.

8. The method according to any claim 2-7, further comprising the steps of identifying and removing coherent noise in the enhanced motion data (141) by way of autocorrelation.

9. The method according to any claim 2-8, further comprising steps to obtain the enhanced motion data (141) in the TX domain within one second from end of recording the motion data (111) and hydrophone data (116).