A method and system for near-trace residual multiple suppression based on deep learning

Abstract

The application provides a near-trace residual multiple wave suppression method and system based on deep learning, and belongs to the field of oil and gas exploration and development. The near-trace residual multiple wave suppression method based on deep learning comprises steps 1) to 19). The application obtains a large amount of label data which is consistent with the near-trace data characteristics of the actual seismic trace set and does not contain multiple wave signals by performing AVO background trend correction, spectrum equalization processing, time-varying wavelet inversion and time-varying wavelet convolution processing on far-trace data in actual seismic data. Deep learning training is only performed in a water-containing background range, the influence of oil and gas on the AVO characteristics of the seismic trace set is avoided, and then a deep neural network for suppressing multiple waves can be more accurately constructed, the near-trace residual multiple wave suppression effect can be improved, and reliable trace set data can be provided for subsequent AVO analysis and pre-stack reservoir prediction.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas exploration and development, particularly the field of seismic gather data processing, and relates to a method for suppressing near-path residual multiples, specifically a method and system for suppressing near-path residual multiples based on deep learning. Background Technology

[0002] Multiple suppression is a crucial step in the seismic data processing workflow. Based on the kinematic and dynamic differences between multiples and primary waves, various multiple suppression techniques are commonly used in industry to perform multiple suppression in combination.

[0003] Chinese patent application CN115267911A discloses a seismic multiple suppression method based on model- and data-driven deep learning. It constructs a model for multiple suppression and a data-driven deep learning model, extracts seismic data labels, and uses a full-wavefield shot gather containing both primary and multiple waves as input to the model. The model's output is the primary wave shot gather after multiple suppression. This invention combines Fourier operators and residual networks in the network architecture using residual Fourier modules to extract multiple wave features in the channel, time, and frequency domains. This enables the network to accurately and efficiently suppress multiple waves, allowing for deep simulation of complex multiple suppression processes.

[0004] After conventional multiple suppression processing, seismic data generally shows good multiple suppression results in imaging profiles. However, in the near-path data of gathers, a significant amount of residual multiple signals are often not effectively suppressed. Existing multiple suppression methods for near-path residual multiples mainly include near-path data reconstruction methods based on the AVO equation and intelligent signal processing methods.

[0005] A journal article (Zhou Donghong, A Method for Suppressing Residual Multiples Based on Amplitude Fitting and Signal Matching Reconstruction, 2013) addresses the problem that f-k and τ-p filtering methods cannot effectively attenuate near-offset multiples in marine data. It proposes a method for suppressing residual multiples based on amplitude fitting and signal matching reconstruction, combined with filtering methods. First, an amplitude curve is fitted using mid-to-long-range data with good multiple removal performance. Based on the amplitude curve, a theoretical seismic dataset without multiples is approximately calculated. The matching degree between the theoretical and actual seismic data is calculated, and signal reconstruction is performed based on this matching degree to obtain the seismic data with suppressed residual multiples. This method can effectively attenuate multiples with different generation mechanisms, has high computational efficiency, and can adapt to massive amounts of 3D data.

[0006] A journal article (Wang Kunxi, Suppressing Seismic Multiples Based on Data Augmentation Training, 2021) addresses the issue of multiples in seismic data affecting migration imaging and misleading the interpretation of seismic data. These multiples are often treated as coherent noise and removed. To intelligently attenuate multiples, a method using a deep neural network based on data augmentation training is proposed. The deep neural network is used to identify and reconstruct primary wave data and suppress multiple wave data. The deep neural network includes convolutional encoding and convolutional decoding processes. The convolutional encoding process learns the primary wave features in the full wavefield data, and the convolutional decoding process uses these features to reconstruct the primary wave and suppress multiples and random noise. During the training phase, to improve the noise resistance and generalization ability of the deep neural network, the amount of training data is increased to reduce the risk of overfitting. Through transfer learning, the deep neural network is given the ability to suppress multiples across different work areas, making it applicable to suppressing seismic multiples under complex conditions.

[0007] Applying existing technologies to near-path residual multiple suppression will result in the following problems: (1) The near-path data reconstruction method obtains the AVO equation by fitting the mid-to-far-path seismic signal at each seismic sampling point, and realizes the near-path data reconstruction of the sampling point according to the AVO equation. This method has high requirements for the signal-to-noise ratio of seismic data. If there is a certain degree of noise interference in the mid-to-far-path data, the reconstructed near-path data will deviate significantly from the real data. In actual data application, this method is difficult to achieve good results; (2) Existing intelligent signal processing methods use data without multiple signals as label data, use deep learning algorithms, and use training datasets to establish a complex mapping relationship between the original data and the label data, and then extend it to actual seismic data to realize multiple suppression. For actual seismic data, existing methods are difficult to obtain label data without multiple data. They often use simple theoretical models to create a small amount of training dataset for deep learning training, resulting in poor generalization ability of the constructed deep learning network, which is difficult to use for multiple suppression of actual seismic data.

[0008] Due to the aforementioned problems, and the fact that near-path residual multiples can significantly alter the AVO response characteristics of earthquakes, thereby affecting the pre-stack reservoir prediction effect, this invention provides a deep learning-based near-path residual multiple suppression method and application to solve the problem of poor near-path residual multiple suppression effect. Summary of the Invention

[0009] This invention addresses the problems of existing technologies by providing a deep learning-based method and system for suppressing near-channel residual multiples. The invention performs AVO background trend correction, spectral equalization, time-varying wavelet inversion, and time-varying wavelet convolution on far-channel data from actual seismic gathers. This yields a large amount of labeled data that is consistent with the characteristics of near-channel data from actual seismic gathers and does not contain multiple signals. By training deep learning only within a water-bearing background range, the influence of hydrocarbons on the AVO characteristics of seismic gathers is avoided. This allows for the more accurate construction of a deep neural network for suppressing multiples, improving the suppression effect of near-channel residual multiples and providing reliable gather data for subsequent AVO analysis and pre-stack reservoir prediction.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] In a first aspect, the present invention provides a near-channel residual multiple suppression method based on deep learning, comprising the following steps:

[0012] 1) Collect well logging data and seismic gather data that have undergone conventional multiple wave processing within the work area;

[0013] 2): Based on the geological understanding of the work area and the characteristics of seismic data, identify the range of water-bearing background time windows in the work area where multiple waves are developed but oil and gas are relatively underdeveloped;

[0014] 3): Within the water-bearing background time window defined in step 2), calculate the root mean square amplitude of the seismic gather data collected in step 1) for each offset or incident angle; with offset or incident angle as the independent variable and root mean square amplitude as the dependent variable, construct a linear regression equation for the change of root mean square amplitude of seismic gather data with offset or incident angle using the linear regression method.

[0015] 4): Using the logging data collected in step 1), AVO forward modeling is performed to obtain the forward model. The root mean square amplitude of the forward model is calculated for each offset or incident angle within the water-bearing background time window of the forward model. The linear regression method is used to construct a linear regression equation for the root mean square amplitude of the forward model as a function of offset or incident angle.

[0016] 5): Use the linear regression equations from steps 3) and 4) to calculate the AVO background trend correction factor for each offset or incident angle;

[0017] The AVO background trend correction factor for each offset or incident angle is multiplied with the seismic data collected in step 1) to obtain the corrected AVO water-bearing background trend seismic gather data.

[0018] 6): Based on the actual characteristics of the short-path data, set the range of offsets or incident angles where residual multiples exist in the short-path. Stack the offsets or incident angles containing residual multiples to obtain stacked seismic data of the short-path containing multiples; stack the seismic data within the range of mid-to-far offsets or incident angles that do not contain multiples to obtain stacked seismic data without multiples; record the median of each offset or incident angle participating in the mid-to-far offset or incident angle stacking calculation as the offset or incident angle corresponding to the stacking profile.

[0019] 7): The near-path stacked seismic data containing multiples obtained in step 6) are divided into time windows from shallow to deep layers to obtain a series of time-varying seismic wavelets.

[0020] 8): Perform spectral equalization on the mid-to-long-range stacked seismic data without multiples obtained in step 6) to obtain spectral equalization seismic data; extract seismic wavelets from shallow to deep layers in time-division window from the spectral equalization seismic data to obtain time-varying seismic wavelets; perform time-varying relative wave impedance inversion on the obtained time-varying seismic wavelets and the spectrally flattened seismic data to obtain relative wave impedance data.

[0021] 9): Using the time-varying seismic wavelet obtained in step 7) and the relative wave impedance data obtained in step 8), a convolution operation is performed to achieve spectral correction of the mid-to-far channel stacked data, resulting in seismic data with the same spectral characteristics as the near channel data but without multiples.

[0022] 10): Substitute the offset distance or incident angle value of the multiple waves in step 6) into the linear regression equation obtained in step 4) to calculate the root mean square amplitude corresponding to each offset distance or incident angle.

[0023] 11): Substitute the offset distance or incident angle of the remaining multiple waves in step 6) into the linear regression equation obtained in step 4) to calculate the root mean square amplitude corresponding to each offset distance or incident angle.

[0024] 12): Divide the root mean square amplitude value obtained in step 10) and the root mean square amplitude obtained in step 11) to obtain the amplitude correction factor that varies with offset distance or incident angle.

[0025] 13): Within the offset or incident angle range where residual multiples exist in the near path set in step 6), multiply the amplitude correction factor obtained in step 12) and the seismic data obtained in step 9) to obtain seismic gather data that has the same energy and spectral characteristics as the seismic near path in step 5, but does not contain multiples.

[0026] 14): Within the water-bearing background area defined in step 2), the seismic gather data obtained in step 13) is used as the label data for deep learning;

[0027] 15): Within the water-bearing background range defined in step 2), according to step 6), the offset distance or incident angle range of residual multiples in the near path is set, and the corresponding seismic gathers are extracted from the seismic gather data obtained in step 5) to be used as input data for training the deep learning network.

[0028] 16): Use the label data obtained in step 14) and the input data in step 15) as the training set of the deep learning network, and use deep learning algorithms to construct the mapping relationship between the input data and the label data, and train the deep neural network corresponding to all shortcuts.

[0029] 17): Based on the offset or incident angle range set in step 6), extract the corresponding seismic data from the seismic gather data obtained in step 5) and use it as the input data for the deep neural network constructed in step 16).

[0030] 18): Substitute the input data from step 17) into the deep neural network from step 16) to calculate the seismic short-path data after multiple wave suppression based on the offset or incident angle.

[0031] 19): The seismic near-path data output in step 18) and the mid-to-far-path data in the seismic gather obtained in step 5) are spliced ​​together to obtain the seismic gather data after suppressing the residual multiples of the near-path, and then output.

[0032] Preferably, the conventional multiple processing in step 1) includes SRME processing or Radon transform processing, etc.

[0033] Preferably, the linear regression method described in steps 3) and 4) employs the least squares method.

[0034] Preferably, the calculation method for the AVO background trend correction factor in step 5) is as follows:

[0035] For each offset or incident angle, substitute it into the linear regression equations of steps 3) and 4) respectively to calculate the root mean square amplitude values ​​of the seismic data and the forward model.

[0036] The AVO background trend correction factor for each offset or incident angle is obtained by dividing the root mean square amplitude value of the forward model and the root mean square amplitude value of the seismic data.

[0037] Preferably, in step 6), the incident angle is less than 10 degrees in the range of offset distance or incident angle of the remaining multiple waves in the near path; and the incident angle range is 10-30 degrees in the range of offset distance or incident angle of the medium and far paths.

[0038] Preferably, the spectrum equalization process in step 8) includes:

[0039] The spectral equalization process was performed on the mid-to-long-range stacked seismic data without multiples obtained in step 6) using the spectral flattening technique to obtain the spectral equalized seismic data.

[0040] Preferably, the deep learning algorithm in step 16) includes a BP neural network algorithm, a convolutional neural network algorithm, a recurrent neural network algorithm, or a long short-term memory network algorithm. In one embodiment, the deep learning algorithm uses a BP neural network algorithm; the BP neural network algorithm has 20 hidden layers, 50 neurons in each hidden layer, and 5000 iterations, and uses the conjugate gradient algorithm to iteratively update the weights in the deep neural network.

[0041] Preferably, the deep learning-based near-path residual multiple suppression method further includes: performing a quality assessment on the seismic gather data after near-path residual multiple suppression output in step 19) to verify the suppression effect.

[0042] Secondly, the present invention provides a system for the deep learning-based near-channel residual multiple suppression method described in the above technical solution, comprising the following modules:

[0043] 1) Data acquisition module, used to collect well logging data and seismic gather data processed by conventional multiple wave processing within the work area;

[0044] 2) The water-bearing background segmentation module is used to identify the time window range of water-bearing backgrounds in the work area where multiple waves are developed but oil and gas are relatively underdeveloped, based on the geological understanding and seismic data characteristics of the work area.

[0045] 3) Root mean square amplitude calculation module, used to calculate the root mean square amplitude value of seismic data in the water-bearing background time window range by offset or incident angle.

[0046] 4) Linear regression module, used to construct a linear regression equation with offset or incident angle as independent variable and root mean square amplitude as dependent variable;

[0047] 5) AVO forward modeling module, used to perform AVO forward modeling simulation using well logging data and calculate the root mean square amplitude value of the forward model;

[0048] 6) AVO background trend correction module, used to calculate the AVO background trend correction factor and perform correction;

[0049] 7) The overlay module is used to overlay near-track and mid-to-far-track data to obtain near-track overlay seismic data containing multiples and overlay seismic data without multiples.

[0050] 8) Wavelet acquisition module, used to acquire seismic wavelets from shallow to deep layers in a time-division window;

[0051] 9) Spectrum equalization module, used to perform spectrum equalization processing on mid-to-long-range stacked seismic data that does not contain multiples;

[0052] 10) Wave impedance inversion module, used to perform time-varying relative wave impedance inversion to obtain relative wave impedance data;

[0053] 11) Convolution module, used to perform convolution operations to achieve spectral correction of mid-to-long-range superimposed data;

[0054] 12) Amplitude correction module, used to calculate the amplitude correction factor and perform amplitude correction;

[0055] 13) Deep learning module, used to build deep learning networks, train deep neural networks, and perform multiple wave suppressions;

[0056] 14) Data stitching module, used to stitch together the near-channel data and mid-to-far-channel data after multiple wave suppression to obtain the final seismic gather data.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] This invention obtains a large amount of labeled data that is consistent with the characteristics of the near-channel data of the actual seismic gather and does not contain multiple signals by performing AVO background trend correction, spectral equalization processing, time-varying wavelet inversion and time-varying wavelet convolution processing on the far-channel data of the actual seismic data. Then, by performing deep learning training only within the water-bearing background range, the influence of oil and gas on the AVO characteristics of the seismic gather is avoided, and the deep neural network constructed for suppressing multiples is more accurate.

[0059] This invention can effectively suppress near-path residual multiples. After processing with this invention, the consistency of seismic phase axes at near-intermediate-far angles is significantly improved, thus providing reliable gather data for subsequent AVO analysis and pre-stack reservoir prediction. The deep learning-based near-path residual multiple suppression method can be used for clastic reservoirs as well as carbonate or other lithological reservoirs, and has broad application prospects. Attached Figure Description

[0060] Figure 1 The flowchart shows a deep learning-based short-path residual multiple suppression method.

[0061] Figure 2 The image shows the collected seismic gather data, where the horizontal axis represents the incident angle (in degrees) within each CDP, the vertical axis represents the seismic sampling time (in milliseconds), and the legend represents the normalized seismic amplitude (in micrometers).

[0062] Figure 3The image shows the root mean square amplitude characteristics of seismic gathers and AVO forward models within a water-bearing background area, varying with the incident angle. The horizontal axis represents the incident angle (in degrees), the vertical axis represents the amplitude attenuation (in decibels), the blue coordinates and lines represent seismic gathers, and the red coordinates and lines represent AVO forward models.

[0063] Figure 4 This is a graph showing the root mean square amplitude offset of the actual gather before and after AVO background trend correction, varying with the incident angle. The horizontal axis represents the incident angle (in degrees), the vertical axis represents the amplitude attenuation (in decibels), the black coordinates and lines represent the AVO background trend before correction, and the purple coordinates and lines represent the AVO background trend after correction.

[0064] Figure 5 This is a near-angle stacked profile containing residual multiples; where the horizontal axis represents the trace number, the vertical axis represents the seismic sampling time (in milliseconds), and the legend represents the normalized seismic amplitude (in micrometers);

[0065] Figure 6 This is a mid-to-long-angle stacked profile without multiples; where the horizontal axis represents the trace number, the vertical axis represents the seismic sampling time (unit: milliseconds), and the legend represents the normalized seismic amplitude (unit: micrometers);

[0066] Figure 7 A comparison of seismic spectral characteristics of mid-to-long-range angle overlay data before and after spectral equalization processing; where the horizontal axis represents frequency and the vertical axis represents frequency amplitude (unit: decibel);

[0067] Figure 8 This is a mid-to-long-range angle superimposed profile after spectral equalization processing; where the horizontal axis represents the trace number, the vertical axis represents the sampling time (unit: milliseconds), and the legend represents the normalized seismic amplitude (unit: micrometers);

[0068] Figure 9 This is a seismic wavelet plot extracted from mid-to-long-range angle superimposed data after spectral equalization processing, using a time-division window. The horizontal axis represents the wavelet sampling time, and the vertical axis represents the amplitude (unit: micrometers).

[0069] Figure 10 The graph shows the results of the relative impedance inversion; where the horizontal axis represents the trace number, the vertical axis represents the seismic sampling time, and the legend represents the amplitude of the relative impedance (unit: m / s*g / cm³).

[0070] Figure 11 This is a seismic wavelet plot extracted from near-angle overlay data using time-division windows; where the horizontal axis represents the wavelet sampling time and the vertical axis represents the amplitude (unit: micrometers);

[0071] Figure 12 This is a seismic profile obtained by convolution of near-angle time-varying wavelet and relative wave impedance; where the horizontal axis represents the trace number, the vertical axis represents the seismic sampling time (in milliseconds), and the legend represents the normalized seismic amplitude (in micrometers);

[0072] Figure 13 This is a comparison chart of labeled data and deep neural network output data; where the horizontal axis represents the amplitude value (unit: micrometers), the vertical axis represents the earthquake sampling time (unit: milliseconds), red represents the neural network output result, and black represents the labeled data;

[0073] Figure 14 This is a plot of gather data after suppression by residual multiples in the near-path; where the horizontal axis represents the offset within each CDP (in degrees), the vertical axis represents the seismic sampling time (in milliseconds), and the legend represents the normalized seismic amplitude (in micrometers). Detailed Implementation

[0074] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The specific implementation methods of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0075] Example

[0076] Deep learning-based near-path residual multiple suppression methods, such as... Figure 1 As shown, the specific steps are as follows:

[0077] 1) Collect seismic gather data within the work area that has undergone conventional multiple wave processing such as SRME and Radon transform. Figure 2 ) and well logging data.

[0078] 2) Based on geological understanding, the residual multiples of the near-channel in the tunnel are caused by the Permian volcanic rock strong reflection strata above the target layer (the bottom is above 3500ms), while the target layer is below 4500ms. Therefore, the water-bearing background time window can be set to 3500ms to 4500ms.

[0079] 3) Using the well logging data collected in step 1), perform AVO forward modeling to obtain the forward model. Within the water-bearing background time window defined in step 2), calculate the root mean square amplitude values ​​of the seismic gather data collected in step 1) and the forward model as a function of the incident angle (incident angle range: 1 degree - 30 degrees). With the incident angle as the independent variable and the root mean square amplitude value as the dependent variable, obtain the linear regression equation through regression analysis, such as... Figure 3As shown, the linear regression equation corresponding to the gather data is Y = 0.26 + 0.083 * X, and the linear regression equation corresponding to the forward model is Y = 0.44 - 0.074 * X, where the horizontal axis X represents the incident angle and the vertical axis Y represents the root mean square amplitude, with the unit being decibels.

[0080] 4) For each incident angle, substitute the gather data from step 3) and the corresponding linear regression equation of the forward model to calculate the root mean square amplitude values ​​of the seismic data and the forward model.

[0081] The AVO background trend correction factor for each incident angle is obtained by dividing the root mean square amplitude value of the forward model and the root mean square amplitude value of the seismic data.

[0082] Multiply the AVO background trend correction factor for each incident angle by the actual gather data of the seismic data collected in step 1) to obtain the corrected AVO background trend seismic gather data. Figure 4 ).

[0083] 5) Based on the characteristics of seismic gather data, the range of near-path residual multiples concentrated within an incident angle of 10 degrees is analyzed. Gather data within 10 degrees and gather data above 10 degrees are then stacked separately to obtain near-path stacked seismic data containing residual multiples (profile as shown). Figure 5 (as shown) and mid-to-long-range channels without residual multiples (profile as shown) Figure 6 As shown in the figure, the incident angle of the mid-to-far path ranges from 10 degrees to 30 degrees, with a median of 20 degrees. The median of each incident angle participating in the mid-to-far path stacking calculation is recorded as the incident angle corresponding to the stacking profile; the obtained near-path stacking seismic data containing multiples are divided into time windows from shallow to deep to obtain a series of time-varying seismic wavelets;

[0084] 6) The mid-to-far channel stacked seismic data profile without multiples obtained in step 5) is subjected to spectral equalization using spectral flattening technology to obtain a relatively flat spectrum. Figure 7 earthquake data () Figure 8 ),in Figure 7 A comparison of seismic spectral characteristics of mid-to-long-range angle overlay data before and after spectral equalization processing.

[0085] 7) Seismic wavelets were extracted from the superimposed data after spectral equalization processing by time window division. In this embodiment, five wavelets were extracted with midpoints at 3250ms, 3750ms, 4250ms, 4750ms, and 5250ms respectively. Figure 9 ).

[0086] 8) Using the time-varying wavelet extracted in step 7 and the superimposed data obtained in step 6, perform relative wave impedance inversion to obtain the relative wave impedance data volume ( Figure 10 ).

[0087] 9) Extract seismic wavelets from the near-channel stacked data in step 5 by time window. In this embodiment, five wavelets were extracted with the midpoint of the time window at 3250ms, 3750ms, 4250ms, 4750ms, and 5250ms respectively. Figure 11 ).

[0088] 10) Using the time-varying wavelet extracted in step 9 and the relative wave impedance data obtained in step 8, perform convolution operations to obtain seismic data with spectral characteristics consistent with the near-channel stacked data but without the development of multiple waves. Figure 12 ).

[0089] 11) Substitute each incident angle within 10 degrees into Y = 0.44 - 0.074 * X to obtain the Y corresponding to each incident angle X. x Substituting the 20-degree incident angle into Y = 0.44 - 0.074 * X, we obtain the corresponding Y. 20 For each incident angle X, Y x Divide by Y 20 The amplitude correction factor that varies with the incident angle is obtained.

[0090] 12) In this embodiment, the incident angle within 10 degrees includes 4 degrees, 5 degrees, 7 degrees, and 9 degrees. Within the water-bearing background range, 4-degree data from the channel set is first extracted as input data for the neural network training set, and then Y4 is divided by Y... 20 As a correction factor, this correction factor is multiplied by the data obtained in step 10, and used as a label in the deep neural network training set. The deep neural network is trained using the BP neural network algorithm to obtain a deep neural network suitable for suppressing near-channel multiples at an incident angle of 4 degrees. Subsequently, data at 5 degrees, 7 degrees, and 9 degrees are extracted from the channel set as input data for the neural network training set, and the above process is repeated to obtain deep neural networks suitable for suppressing near-channel multiples at incident angles of 5 degrees, 7 degrees, and 9 degrees, respectively. In this embodiment, the deep neural network has 20 hidden layers, each with 50 neurons, and 5000 iterations. The conjugate gradient algorithm is used to iteratively update the weights within the deep neural network. A comparison of the label data and the deep neural network output data is provided. Figure 13 As shown.

[0091] 13) Extract the data at 4 degrees, 5 degrees, 7 degrees and 9 degrees from the fourth track set, and substitute them into their respective neural networks to obtain the data after suppression of the remaining multiple waves at the above angles.

[0092] 14) The data obtained in step 13 and the data from the gather after 10 degrees in step 4 are spliced ​​together by angle to obtain the gather data after suppressing the residual multiples near the path. Figure 14 ), and output.

[0093] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Simple modifications or equivalent substitutions made by those skilled in the art to the technical solution of the present invention do not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A deep learning based near-trace residual multiple suppression method, characterized in that, Includes the following steps: 1) Collect well logging data and seismic gather data that have undergone conventional multiple wave processing within the work area; 2): Based on the geological understanding of the work area and the characteristics of seismic data, identify the range of water-bearing background time windows in the work area where multiple waves are developed but oil and gas are relatively underdeveloped; 3): Within the water-bearing background time window defined in step 2), calculate the root mean square amplitude of the seismic gather data collected in step 1) for each offset or incident angle; with offset or incident angle as the independent variable and root mean square amplitude as the dependent variable, construct a linear regression equation for the change of root mean square amplitude of seismic gather data with offset or incident angle using the linear regression method. 4): Using the logging data collected in step 1), AVO forward modeling is performed to obtain the forward model. The root mean square amplitude of the forward model is calculated for each offset or incident angle within the water-bearing background time window of the forward model. The linear regression method is used to construct a linear regression equation for the root mean square amplitude of the forward model as a function of offset or incident angle. 5): Use the linear regression equations from steps 3) and 4) to calculate the AVO background trend correction factor for each offset or incident angle; The AVO background trend correction factor for each offset or incident angle is multiplied with the seismic data collected in step 1) to obtain the corrected AVO water-bearing background trend seismic gather data. 6): Based on the actual characteristics of the short-path data, set the offset or incident angle range where residual multiples exist in the short-path. Stack the offset or incident angle data containing residual multiples to obtain short-path stacked seismic data containing multiples; stack the seismic data within the mid-to-far offset or incident angle range without multiples to obtain stacked seismic data without multiples. Record the median of each offset or incident angle involved in the superposition calculation of mid-to-long-range offset or incident angle as the offset or incident angle corresponding to the superposition profile; 7): The near-path stacked seismic data containing multiples obtained in step 6) are divided into time windows from shallow to deep layers to obtain a series of time-varying seismic wavelets. 8): Perform spectral equalization on the mid-to-long-range stacked seismic data without multiples obtained in step 6) to obtain spectral equalization seismic data; extract seismic wavelets from shallow to deep layers in time-division window from the spectral equalization seismic data to obtain time-varying seismic wavelets; perform time-varying relative wave impedance inversion on the obtained time-varying seismic wavelets and the spectrally flattened seismic data to obtain relative wave impedance data. 9): Using the time-varying seismic wavelet obtained in step 7) and the relative wave impedance data obtained in step 8), a convolution operation is performed to achieve spectral correction of the mid-to-far channel stacked data, resulting in seismic data with the same spectral characteristics as the near channel data but without multiples. 10): Substitute the offset distance or incident angle value of the multiple waves in step 6) into the linear regression equation obtained in step 4) to calculate the root mean square amplitude corresponding to each offset distance or incident angle. 11): Substitute the offset distance or incident angle of the remaining multiple waves in step 6) into the linear regression equation obtained in step 4) to calculate the root mean square amplitude corresponding to each offset distance or incident angle. 12): Divide the root mean square amplitude value obtained in step 10) and the root mean square amplitude obtained in step 11) to obtain the amplitude correction factor that varies with offset distance or incident angle. 13): Within the offset or incident angle range where residual multiples exist in the near path set in step 6), multiply the amplitude correction factor obtained in step 12) and the seismic data obtained in step 9) to obtain seismic gather data that has the same energy and spectral characteristics as the seismic near path in step 5, but does not contain multiples. 14): Within the water-bearing background area defined in step 2), the seismic gather data obtained in step 13) is used as the label data for deep learning; 15): Within the water-bearing background range defined in step 2), according to step 6), the offset distance or incident angle range of residual multiples in the near path is set, and the corresponding seismic gathers are extracted from the seismic gather data obtained in step 5) to be used as input data for training the deep learning network. 16): Use the label data obtained in step 14) and the input data in step 15) as the training set of the deep learning network, and use deep learning algorithms to construct the mapping relationship between the input data and the label data, and train the deep neural network corresponding to all shortcuts. 17): Based on the offset or incident angle range set in step 6), extract the corresponding seismic data from the seismic gather data obtained in step 5) and use it as the input data for the deep neural network constructed in step 16). 18): Substitute the input data from step 17) into the deep neural network from step 16) to calculate the seismic short-path data after multiple wave suppression based on the offset or incident angle. 19): The seismic near-path data output in step 18) and the mid-to-far-path data in the seismic gather obtained in step 5) are spliced ​​together to obtain the seismic gather data after suppressing the residual multiples of the near-path, and then output.

2. The deep learning based near-trace residual multiple attenuation method of claim 1, wherein, The calculation method for the AVO background trend correction factor in step 5) is as follows: For each offset or incident angle, substitute it into the linear regression equations of steps 3) and 4) respectively to calculate the root mean square amplitude values ​​of the seismic data and the forward model. The AVO background trend correction factor for each offset or incident angle is obtained by dividing the root mean square amplitude value of the forward model and the root mean square amplitude value of the seismic data.

3. The deep learning based near-trace residual multiple attenuation method of claim 1, wherein, The spectrum equalization process in step 8) includes: The spectral equalization process was performed on the mid-to-long-range stacked seismic data without multiples obtained in step 6) using the spectral flattening technique to obtain the spectral equalized seismic data.

4. The deep learning based near-trace residual multiple attenuation method of claim 1, wherein, The deep learning algorithms in step 16) include BP neural network algorithm, convolutional neural network algorithm, recurrent neural network algorithm or long short-term memory network algorithm.

5. The method of claim 4, wherein, The deep learning algorithm described uses the BP neural network algorithm. In the BP neural network algorithm, the deep neural network has 20 hidden layers, each hidden layer has 50 neurons, and the number of iterations is 5000. The conjugate gradient algorithm is used to iteratively update the weights in the deep neural network.

6. The deep learning based near-trace residual multiple attenuation method of claim 1, wherein, In step 6), the short path has a range of offset distances or incident angles of residual multiple waves, where the incident angle is less than 10 degrees.

7. The deep learning-based short-channel residual multiple suppression method according to claim 1, characterized in that, In step 6), the range of the mid-to-long offset distance or incident angle is 10-30 degrees.

8. The deep learning-based short-channel residual multiple suppression method according to claim 1, characterized in that, Also includes: The quality of the seismic gather data after suppression of the near-path residual multiples output in step 19) is evaluated to verify the suppression effect.

9. The deep learning-based short-channel residual multiple suppression method according to claim 1, characterized in that, The linear regression method described in steps 3) and 4) uses the least squares method.

10. The system of the deep learning-based near-channel residual multiple suppression method according to any one of claims 1-9, characterized in that, Includes the following modules: 1) Data acquisition module, used to collect well logging data and seismic gather data processed by conventional multiple wave processing within the work area; 2) The water-bearing background segmentation module is used to identify the time window range of water-bearing backgrounds in the work area where multiple waves are developed but oil and gas are relatively underdeveloped, based on the geological understanding and seismic data characteristics of the work area. 3) Root mean square amplitude calculation module, used to calculate the root mean square amplitude value of seismic data in the water-bearing background time window range by offset or incident angle. 4) Linear regression module, used to construct a linear regression equation with offset or incident angle as independent variable and root mean square amplitude as dependent variable; 5) AVO forward modeling module, used to perform AVO forward modeling simulation using well logging data and calculate the root mean square amplitude value of the forward model; 6) AVO background trend correction module, used to calculate the AVO background trend correction factor and perform correction; 7) The overlay module is used to overlay near-track and mid-to-far-track data to obtain near-track overlay seismic data containing multiples and overlay seismic data without multiples. 8) Wavelet acquisition module, used to acquire seismic wavelets from shallow to deep layers in a time-division window; 9) Spectrum equalization module, used to perform spectrum equalization processing on mid-to-long-range stacked seismic data that does not contain multiples; 10) Wave impedance inversion module, used to perform time-varying relative wave impedance inversion to obtain relative wave impedance data; 11) Convolution module, used to perform convolution operations to achieve spectral correction of mid-to-long-range superimposed data; 12) Amplitude correction module, used to calculate the amplitude correction factor and perform amplitude correction; 13) Deep learning module, used to build deep learning networks, train deep neural networks, and perform multiple wave suppressions; 14) Data stitching module, used to stitch together the near-channel data and mid-to-far-channel data after multiple wave suppression to obtain the final seismic gather data.

Images