A Deep Learning-Based Intelligent Method and System for Suppressing Noise in Seismic Signals
By constructing noise prior data and a deep learning suppression model, the three-dimensional seismic signal is differentially suppressed, which solves the problems of noise suppression and reflection component preservation in coal exploration, and improves the signal-to-noise ratio and interpretation reliability.
Patent Information
- Application Number
- CN202610995323.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-06
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to accurately suppress various types of noise during 3D seismic acquisition in coal exploration, while simultaneously preserving effective reflection components and the stability of phase axis continuity, resulting in a low signal-to-noise ratio and impacting the accuracy of subsequent interpretation.
By constructing noise prior data, combining noise attribution data and reflection protection data, a deep learning suppression model is used to perform differential suppression on the original 3D seismic signal data. This includes data acquisition, noise scene label construction, time-frequency characterization, dual-branch identification, and consistency feedback optimization, and outputs the target suppression result.
It effectively reduces various types of noise, improves the signal-to-noise ratio, stably preserves effective reflection components and phase axis continuity, and enhances the reliability of coal seam interface, fault interpretation, and structural interpretation.
Smart Images

Figure CN123113003A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of seismic signal suppression technology, and in particular to a method and system for intelligent suppression of seismic signal noise based on deep learning. Background Technology
[0002] In coal exploration, 3D seismic acquisition is susceptible to surface wave interference, random noise, industrial interference, coupling anomalies, and shallow scattering interference, resulting in low signal-to-noise ratios in the acquired raw 3D seismic signal data and affecting the accuracy of subsequent stacking imaging and structural interpretation. These interferences create strong vibration noise, continuous coherent noise, pulse-type anomaly noise, and pseudo-reflection responses in the raw 3D seismic signal data, significantly submerging or distorting effective reflection signals, especially weak reflection signals, continuous reflection signals, and boundary transition signals.
[0003] Existing seismic signal denoising methods primarily employ frequency filtering, threshold denoising, local transform suppression, or deep learning direct mapping to suppress noise. While traditional frequency domain or time-frequency domain methods can reduce some regular noise to a certain extent, they often struggle to effectively distinguish between noise from complex scenarios that highly overlap with the effective reflection signal in terms of frequency band distribution, amplitude variation, and local morphology. This can lead to the suppression of effective reflection signals while simultaneously weakening the noise. Although existing deep learning methods have improved the adaptability of noise suppression, most still focus on the direct reconstruction from the original input to the target output, emphasizing waveform restoration. They do not adequately consider the differences in noise sources, the noise formation mechanisms under different disturbance scenarios, or the usability of the suppressed geological interpretation. This can easily result in problems such as the inaccurate suppression of effective weak reflections, decreased reflection continuity, blurred boundary information, and insufficient stability of suppression results under different construction disturbance scenarios.
[0004] Especially in application scenarios such as exploration along urban rail transit lines, detection of underground cavities under construction disturbance conditions, and detailed exploration of strata around foundation pits, interpreters not only require the suppressed seismic signal to have a low noise level, but also require the preserved reflection events to continue to support subsequent layer tracking, fault identification, boundary interpretation, and anomaly determination. Therefore, how to intelligently identify and differentially suppress multiple types of noise in the original 3D seismic signal data under complex engineering disturbance conditions, while stably preserving continuous reflection sections and weak reflection sections, and further outputting highly reliable suppression results that can be used for subsequent interpretation, has become a problem that existing technologies urgently need to solve. Summary of the Invention
[0005] The technical problem solved by this invention is that existing technologies are unable to accurately suppress various types of noise that affect the signal-to-noise ratio of the original three-dimensional seismic signal data during the three-dimensional seismic acquisition process in coal exploration, while simultaneously ensuring the stable preservation of effective reflection components and the continuity of the phase axis.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0007] A deep learning-based intelligent noise suppression method for seismic signals includes the following steps:
[0008] Step S1: Collect the original 3D seismic signal data and noise scene label data of the target exploration area, and store the original 3D seismic signal data and noise scene label data in the database;
[0009] Step S2: Extract silent segment data, wakeband data, anomalous segment data, and repeated acquisition segment data from the original 3D seismic signal data, and construct noise prior data based on the silent segment data, wakeband data, anomalous segment data, and repeated acquisition segment data. At the same time, construct time-frequency characterization data based on the original 3D seismic signal data.
[0010] Step S3: Input the original 3D seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the deep learning suppression model, and output noise attribution data and reflection protection data.
[0011] Step S4: Suppress the original 3D seismic signal data based on noise attribution data and reflection protection data, and output the initial suppression result data;
[0012] Step S5: Construct consistency evaluation data based on the initial suppression result data and the original 3D seismic signal data, update the parameters of the deep learning suppression model based on the consistency evaluation data, and output the target suppression result data and suppression confidence data.
[0013] Step S6: Input the target suppression result data and suppression confidence data to the control terminal for display, storage and transmission.
[0014] Preferably, step S1 includes the following sub-steps:
[0015] Step S101: Acquire raw three-dimensional seismic signal data of the target exploration area using a detector array. The raw three-dimensional seismic signal data includes amplitude sequence data arranged by trace and corresponding sampling time sequence data.
[0016] Step S102: Collect noise scene label data corresponding to the original three-dimensional seismic signal data. The noise scene label data includes at least one of train disturbance label data, mechanical construction disturbance label data, vehicle disturbance label data, and coupling anomaly label data.
[0017] Step S103: Establish the correspondence between the original 3D seismic signal data and the noise scene label data, and store the corresponding original 3D seismic signal data and noise scene label data in the database.
[0018] Preferably, step S2 includes the following sub-steps:
[0019] Step S201: Extract the low-energy region before excitation based on the original three-dimensional seismic signal data and output the silent segment data;
[0020] Step S202: Extract the attenuation interval after the effective reflection ends based on the original three-dimensional seismic signal data, and output the coma band data;
[0021] Step S203: Extract amplitude mutation intervals, frequency mutation intervals, and continuity interruption intervals from the original three-dimensional seismic signal data, and output the anomaly segment data;
[0022] Step S204: Extract the repeated response intervals from the original three-dimensional seismic signal data repeatedly acquired in the same area, and output the repeated acquisition segment data;
[0023] Step S205: Calculate noise intensity data, noise continuity data, and noise repeatability data based on silent segment data, wake wave segment data, abnormal segment data, and repeated acquisition segment data, and combine the noise intensity data, noise continuity data, and noise repeatability data into noise prior data.
[0024] Step S206: Perform short-time time-frequency transformation on the original three-dimensional seismic signal data and output time-frequency characterization data.
[0025] Preferably, the logic for constructing the noise prior data is as follows:
[0026] Calculate the energy distribution data, frequency band occupancy data, and inter-channel similarity data for the silent segment data, wakeband data, abnormal segment data, and repeated acquisition segment data, respectively;
[0027] The corresponding segments are divided according to silent segment data, wake segment data, abnormal segment data and repeated acquisition segment data, and segment identification data is generated for each corresponding segment.
[0028] Noise intensity data is calculated based on the energy distribution data of each corresponding section, and the noise intensity data is then bound to the corresponding section identification data.
[0029] Noise continuity data is calculated based on the frequency band occupancy data of each corresponding segment, and the noise continuity data is bound to the corresponding segment identification data.
[0030] Noise repeatability data is calculated based on the inter-track similarity data of each corresponding section, and the noise repeatability data is bound to the corresponding section identification data.
[0031] The noise prior data is output by combining the segment identification data, noise intensity data, noise continuity data, and noise repeatability data.
[0032] Preferably, step S3 includes the following sub-steps:
[0033] Step S301: Input the original 3D seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the noise attribution branch of the deep learning suppression model, and output the noise attribution data.
[0034] Step S302: Input the original 3D seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the reflection protection branch of the deep learning suppression model, and output the reflection protection data.
[0035] Step S303: Establish the trace-level correspondence between noise attribution data and reflection protection data and the original three-dimensional seismic signal data, and output it to the suppression processing module.
[0036] Preferably, the processing logic of the deep learning suppression model is as follows:
[0037] Based on the noise scene label data, the original 3D seismic signal data is subjected to scene condition coding, and the scene coded data is output.
[0038] Based on noise prior data and time-frequency characterization data, the noise components in the original 3D seismic signal data are classified and noise attribution data is output.
[0039] Based on scene coding data, raw 3D seismic signal data, and time-frequency characterization data, continuous reflection segments and weak reflection segments in the raw 3D seismic signal data are identified, and reflection protection data is output.
[0040] Based on noise attribution data and reflection protection data, the noise component in the original 3D seismic signal data is attenuated while the reflection component is preserved, and the initial suppression result data is output.
[0041] Preferably, step S4 includes the following sub-steps:
[0042] Step S401: Obtain noise attribution data and determine the noise category and noise intensity corresponding to each segment in the original three-dimensional seismic signal data;
[0043] Step S402: Obtain reflection protection data and determine the reflection preservation weights corresponding to each segment in the original three-dimensional seismic signal data;
[0044] Step S403: Based on the noise category, noise intensity, and reflection retention weight, perform differentiated attenuation processing on each segment of the original three-dimensional seismic signal data, and output the initial suppression result data.
[0045] Preferably, step S5 includes the following sub-steps:
[0046] Step S501: Extract the phase axis continuity data, boundary stability data, and energy fidelity data based on the initial suppression result data;
[0047] Step S502: Construct consistency evaluation data based on in-phase axis continuity data, boundary stability data, and energy fidelity data;
[0048] Step S503: Feed the consistency evaluation data back to the deep learning suppression model and update the parameters of the deep learning suppression model.
[0049] Step S504: Based on the parameter-updated deep learning suppression model, the original 3D seismic signal data is suppressed again, and the target suppression result data and suppression confidence data are output.
[0050] Preferably, step S6 includes the following sub-steps:
[0051] Step S601: Obtain target suppression result data and suppression confidence data, and establish the correspondence between target suppression result data, suppression confidence data and noise scene label data;
[0052] Step S602: Input the corresponding target suppression result data and suppression confidence data to the control terminal for display, and generate suppression result record data;
[0053] Step S603: Store the suppression result record data in the database, and transmit the target suppression result data and suppression confidence data to the interpretation terminal.
[0054] A deep learning-based intelligent noise suppression system for seismic signals includes a data acquisition and storage module, a priori feature construction module, a dual-branch identification module, an intelligent suppression processing module, a consistency feedback optimization module, and a result output transmission module.
[0055] The data acquisition and storage module is used to acquire raw 3D seismic signal data and noise scene label data of the target exploration area, and store the raw 3D seismic signal data and noise scene label data into the database.
[0056] The prior feature construction module is used to extract silent segment data, wakeband data, anomalous segment data and repeated acquisition segment data from the original three-dimensional seismic signal data, and to construct noise prior data based on the silent segment data, wakeband data, anomalous segment data and repeated acquisition segment data, while constructing time-frequency characterization data based on the original three-dimensional seismic signal data.
[0057] The dual-branch identification module is used to input the original three-dimensional seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the deep learning suppression model, and output noise attribution data and reflection protection data.
[0058] The intelligent suppression processing module is used to suppress the original three-dimensional seismic signal data based on noise attribution data and reflection protection data, and output the initial suppression result data.
[0059] The consistency feedback optimization module is used to construct consistency evaluation data based on the initial suppression result data and the original three-dimensional seismic signal data, and to update the parameters of the deep learning suppression model based on the consistency evaluation data, and output the target suppression result data and suppression confidence data.
[0060] The result output transmission module is used to input the target suppression result data and suppression confidence data to the control terminal for display, storage and transmission.
[0061] The beneficial effects of this invention are as follows: By constructing noise prior data and combining it with noise attribution data and reflection protection data to differentially suppress the original three-dimensional seismic signal data, this invention can effectively reduce multiple types of noise, improve the signal-to-noise ratio, and stably retain effective reflection components and phase axis continuity during the three-dimensional seismic acquisition process in coal exploration, thereby improving the reliability of coal seam interface interpretation, fault interpretation, and structural interpretation. Attached Figure Description
[0062] Figure 1 A flowchart illustrating the steps of a deep learning-based intelligent noise suppression method for seismic signals, as provided in one embodiment of the present invention.
[0063] Figure 2 This is a system composition diagram of a deep learning-based intelligent seismic signal noise suppression system provided in one embodiment of the present invention. Detailed Implementation
[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0065] Example 1, referring to Figure 1This paper presents a deep learning-based intelligent noise suppression method for seismic signals, comprising the following steps:
[0066] Step S1: Collect the original 3D seismic signal data and noise scene label data of the target exploration area, and store the original 3D seismic signal data and noise scene label data in the database.
[0067] Step S2: Extract silent segment data, wakeband data, anomalous segment data, and repeated acquisition segment data from the original 3D seismic signal data. Construct noise prior data based on the silent segment data, wakeband data, anomalous segment data, and repeated acquisition segment data. Simultaneously, construct time-frequency characterization data based on the original 3D seismic signal data.
[0068] Step S3: Input the original 3D seismic signal data, noise scene label data, noise prior data, and time-frequency characterization data into the deep learning suppression model, and output noise attribution data and reflection protection data.
[0069] Step S4: Suppress the original 3D seismic signal data based on noise attribution data and reflection protection data, and output the initial suppression result data.
[0070] Step S5: Construct consistency evaluation data based on the initial suppression result data and the original 3D seismic signal data, update the parameters of the deep learning suppression model based on the consistency evaluation data, and output the target suppression result data and suppression confidence data.
[0071] Step S6: Input the target suppression result data and suppression confidence data to the control terminal for display, storage and transmission.
[0072] This embodiment is applied to the 3D seismic acquisition data processing scenario in coal exploration. It aims to solve the technical problem that surface wave interference, random noise, industrial interference, coupling anomaly interference, and shallow scattering interference overlap with the effective reflection components in the original 3D seismic signal data, making it difficult for existing technologies to simultaneously achieve effective noise suppression, signal-to-noise ratio improvement, and stable preservation of effective reflection components.
[0073] This invention constructs noise prior data and combines it with noise attribution data and reflection protection data to differentially suppress the original 3D seismic signal data. This can effectively reduce multiple types of noise, improve the signal-to-noise ratio, and stably preserve effective reflection components and phase axis continuity during 3D seismic acquisition in coal exploration, thereby improving the reliability of coal seam interface interpretation, fault interpretation, and structural interpretation.
[0074] Step S1 includes the following sub-steps:
[0075] Step S101: Use a detector array to acquire raw three-dimensional seismic signal data of the target exploration area. The raw three-dimensional seismic signal data includes amplitude sequence data arranged by trace and corresponding sampling time sequence data.
[0076] In step S101, let the first... The amplitude sequence data of the earthquake records are represented as follows ,in The sampling time in the sampling time sequence data, , This indicates the number of traces acquired. For each raw 3D seismic signal data trace, the sampling duration is set to... The number of sampling points is set to Then it has a discrete form. ,in .
[0077] Step S102: Collect noise scene label data corresponding to the original three-dimensional seismic signal data. The noise scene label data includes at least one of train disturbance label data, mechanical construction disturbance label data, vehicle disturbance label data, and coupling anomaly label data.
[0078] In step S102, for the first Establish scene perturbation label vectors from seismic records Its element takes the value 0 or 1, which is used to indicate whether the corresponding disturbance exists.
[0079] Step S103: Establish the correspondence between the original 3D seismic signal data and the noise scene label data, and store the corresponding original 3D seismic signal data and noise scene label data in the database.
[0080] In step S103, a trace-level correspondence is established between the original 3D seismic signal data and the noisy scene label data, denoted as a sample pair. The corresponding original 3D seismic signal data and noise scene label data are then written into the database. Through the above processing, each original 3D seismic signal data is uniquely bound to the noise scene label data of the same channel, which can be used for subsequent noise prior data construction and deep learning suppression model.
[0081] Step S2 includes the following sub-steps:
[0082] Step S201: Extract the low-energy region before excitation based on the original three-dimensional seismic signal data and output the silent segment data.
[0083] In step S201, let the sampling interval before the excitation time in each original 3D seismic signal data be denoted as . The silent segment data is then represented as ,in .
[0084] Step S202: Extract the attenuation interval after the effective reflection ends based on the original three-dimensional seismic signal data, and output the tailband data.
[0085] In step S202, let the sampling interval corresponding to the end of effective reflection be... The tailband data is then represented as ,in .
[0086] Step S203: Extract amplitude mutation intervals, frequency mutation intervals, and continuity interruption intervals from the original three-dimensional seismic signal data, and output the anomalous segment data.
[0087] In step S203, the difference between adjacent sampling points is calculated first:
[0088] ;
[0089] Then calculate the local energy:
[0090] ;
[0091] in, The width is half the width of a local window. When... Exceeding the differential threshold, or When the energy change relative to adjacent windows exceeds the energy threshold, the corresponding interval is marked as an outlier segment. .
[0092] Step S204: Extract the repeated response intervals from the original three-dimensional seismic signal data repeatedly acquired in the same area, and output the repeated acquisition segment data.
[0093] In step S204, let the two records collected repeatedly in the same area be as follows: and Then calculate the repeated correlation coefficient:
[0094] ;
[0095] when When the data exceeds the relevant threshold, the corresponding interval will be output as a repeatedly collected segment of data. .
[0096] Step S205: Calculate noise intensity data, noise continuity data, and noise repeatability data based on the silent segment data, wake wave segment data, abnormal segment data, and repeated acquisition segment data, and combine the noise intensity data, noise continuity data, and noise repeatability data into noise prior data.
[0097] The logic for constructing noisy prior data is as follows:
[0098] Calculate the energy distribution data, frequency band occupancy data, and interchannel similarity data for the silent segment data, wakeband data, abnormal segment data, and repeated acquisition segment data, respectively.
[0099] The corresponding segments are divided based on silent segment data, wake segment data, abnormal segment data, and repeated acquisition segment data, and segment identification data is generated for each corresponding segment.
[0100] Noise intensity data is calculated based on the energy distribution data of each corresponding section, and the noise intensity data is then bound to the corresponding section identification data.
[0101] Noise continuity data is calculated based on the frequency band occupancy data of each corresponding segment, and the noise continuity data is then bound to the corresponding segment identification data.
[0102] Noise repeatability data is calculated based on the inter-track similarity data of each corresponding section, and the noise repeatability data is bound to the corresponding section identification data.
[0103] The noise prior data is output by combining the segment identification data, noise intensity data, noise continuity data, and noise repeatability data.
[0104] In step S205, for the first The first of the original 3D seismic signal data There are several segments, and the amplitude sequence of each segment is denoted as... ,in , Indicates the starting sampling point of this section. This indicates the termination sampling point of the segment. First, segment identification data is generated for each segment. This data includes at least the track number, the segment's starting sampling point, and the segment's ending sampling point. The segment identification data is only used to establish the association between each segment and its corresponding calculation result during the calculation process and is not considered a final component of the noise prior data.
[0105] For energy distribution data, the mean square energy of the amplitude sequence within the segment is used for characterization. Dao Di Energy distribution data for each section The calculation is as follows:
[0106] ;
[0107] in, Used to reflect the overall level of amplitude energy within this section. When the silent section data or the wake section data corresponds to the section A higher value indicates that there may be strong background noise or continuous noise in that segment; when the abnormal segment data corresponds to the segment's... A significant sudden increase indicates the possible presence of transient noise or localized abnormal interference in that section. Based on the energy distribution data, the noise intensity data for that section can be further obtained. To enhance the comparability between different sections, in this embodiment, the energy distribution data for all sections is normalized to obtain the noise intensity data. The calculation is as follows:
[0108] ;
[0109] in, This represents the maximum value of the energy distribution data for all segments in the current processing batch. This represents the minimum energy distribution data for all segments in the current batch. The larger the normalized noise intensity data, the stronger the noise energy in the corresponding segment.
[0110] For the frequency band occupancy data, the segment amplitude sequence is first subjected to time-frequency transformation to obtain the segment spectral amplitude. ,in Indicates frequency index, This represents the time window index within the segment, with the spectral amplitude threshold set to... Then the first Dao Di Frequency band occupancy data for each segment The calculation is as follows:
[0111] ;
[0112] in, Represents a set of frequency indices. Represents the set of time window indices. This represents the characteristic function. When the spectral amplitude at a certain frequency and time window position is not lower than the spectral amplitude threshold... When the frequency band occupancy is specified, the indicator function takes a value of one; otherwise, it takes a value of zero. Used to reflect the active frequency band range of this segment in the time-frequency domain, when A larger value indicates that the energy distribution in that segment persists over a wider frequency band or a longer time period, typically corresponding to noise components with strong continuity. Based on frequency band occupancy data, this embodiment directly... This will be used as noise continuity data for this section. ,Right now .
[0113] For inter-channel similarity data, when processing silent segment data, wakeband data, and anomalous segment data, the similarity calculation is performed between the current segment and the segment at the same position in the adjacent channel; when processing repeatedly acquired segment data, the similarity calculation is performed between the corresponding segments obtained from the repeated acquisitions. Let's assume that the current segment is similar to the segment at the same position in the adjacent channel. Dao Di The comparison segment corresponding to each segment is: ,in This indicates the adjacent channel number or repeated collection number corresponding to the current segment. Dao Di Inter-lane similarity data for each segment The calculation is as follows:
[0114] ;
[0115] in, The closer the value of is to one, the closer the waveform of the current segment is to the comparison segment, indicating that the repetitive or coherent noise characteristics in this segment are more obvious; when A lower value indicates that the segment is more likely to be a segment dominated by random noise or a segment with local anomalies. Based on inter-channel similarity data, this embodiment will use the noise repeatability data for this segment. ,Right now .
[0116] In obtaining the first Dao Di After obtaining the noise intensity data, noise continuity data, and noise repeatability data for each segment, the correlation between these three types of data and the corresponding segments is established using the aforementioned segment identification data. This ensures that each calculation result can be accurately traced back to the corresponding silent segment data, wake segment data, abnormal segment data, or repeated acquisition segment data segment. After completing the correlation of all segments, the noise intensity data, noise continuity data, and noise repeatability data corresponding to all segments are combined according to channel order and segment order to form noise prior data. It is represented as:
[0117] ;
[0118] in, Indicates the first The total number of segments that have been divided in the original 3D seismic signal data.
[0119] Step S206: Perform short-time time-frequency transformation on the original three-dimensional seismic signal data and output time-frequency characterization data.
[0120] In step S206, a short-time Fourier transform is used to obtain:
[0121] ;
[0122] in, For window functions, For frequency index, This is the sampling point at the center of the window. Further modal spectrum values are then obtained:
[0123] ;
[0124] The time-frequency representation data is used as input for subsequent deep learning suppression models.
[0125] Step S3 includes the following sub-steps:
[0126] Step S301: Input the original 3D seismic signal data, noise scene label data, noise prior data, and time-frequency characterization data into the noise attribution branch of the deep learning suppression model, and output the noise attribution data.
[0127] The deep learning suppression model in this embodiment is denoted as . Its input is denoted as:
[0128] ;
[0129] Noise attribution branch is denoted as Output noise attribution data .
[0130] Noise attribution data consists of a probability sequence of noise categories output at each sampling point. Let the total number of noise categories be... Then in the first At each sampling point, the noise attribution data can be represented as:
[0131] ;
[0132] in Specifically, first, based on the scene's anti-motion tag data... Perform scene condition encoding to obtain scene encoded data:
[0133] ;
[0134] The scene coding data is then fused with the time-frequency characterization data and noise prior data, and input into the noise attribution branch. The calculation process is as follows:
[0135] ;
[0136] ;
[0137] in This represents the time series characteristics obtained after compressing time-frequency characterization data along the frequency dimension. It is a nonlinear mapping.
[0138] Step S302: Input the original 3D seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the reflection protection branch of the deep learning suppression model, and output the reflection protection data.
[0139] In step S302, the reflection protection branch is denoted as Output reflection protection data ,in A larger value indicates that the corresponding sampling point should be retained. Specifically, the reflection protection branch is calculated as follows:
[0140] ;
[0141] ;
[0142] in, This is a compression function.
[0143] Step S303: Establish the trace-level correspondence between noise attribution data and reflection protection data and the original three-dimensional seismic signal data, and output it to the suppression processing module.
[0144] The processing logic of the deep learning suppression model is as follows:
[0145] The original 3D seismic signal data is coded based on the noise scene label data, and the scene-coded data is output.
[0146] Based on prior noise data and time-frequency characterization data, the noise components in the original 3D seismic signal data are classified, and noise attribution data is output.
[0147] Based on scene coding data, raw 3D seismic signal data, and time-frequency characterization data, continuous reflection zones and weak reflection zones in the raw 3D seismic signal data are identified, and reflection protection data is output.
[0148] Based on noise attribution data and reflection protection data, the noise component in the original 3D seismic signal data is attenuated while the reflection component is preserved, and the initial suppression result data is output.
[0149] In step S303, the trace-level correspondence between noise attribution data, reflection protection data, and original 3D seismic signal data is established, that is, for each trace of original 3D seismic signal data... All output the corresponding and .
[0150] Step S4 includes the following sub-steps:
[0151] Step S401: Obtain noise attribution data and determine the noise category and noise intensity corresponding to each segment in the original three-dimensional seismic signal data.
[0152] In step S401, the noise attribution data is weighted and summarized by category to construct a noise intensity sequence:
[0153] ;
[0154] in, For the first Suppression weights corresponding to noise types.
[0155] Step S402: Obtain reflection protection data and determine the reflection preservation weights corresponding to each segment in the original three-dimensional seismic signal data.
[0156] The reflection preservation weight in step S402 is .
[0157] Step S403: Based on the noise category, noise intensity, and reflection retention weight, perform differentiated attenuation processing on each segment of the original three-dimensional seismic signal data, and output the initial suppression result data.
[0158] In step S403, the joint suppression coefficient is constructed:
[0159] ;
[0160] in, This means limiting the results to between 0 and 1, further obtaining the initial suppression result data:
[0161] ;
[0162] in, This refers to the initial suppression result data. It can be seen that when the noise intensity of a sampling point is high and the reflection protection weight is low, the joint suppression coefficient increases, and the amplitude attenuation of the corresponding sampling point is stronger; when a sampling point is identified as a continuous reflection section or a weak reflection section that should be retained, the reflection protection weight increases, and the joint suppression coefficient decreases, thereby avoiding excessive suppression of the effective reflection signal.
[0163] Step S5 includes the following sub-steps:
[0164] Step S501: Extract the phase axis continuity data, boundary stability data, and energy fidelity data based on the initial suppression result data.
[0165] In step S501, for reflection continuity data, the local coherence coefficient between adjacent channels is first calculated to obtain:
[0166] ;
[0167] in, The width of the window is half the width of the passageway. For a time window of half width, further... Statistics are for continuous reflection data For boundary stability data, the local gradient magnitude is first calculated for the original 3D seismic signal data and the initial suppression result data, denoted as... and Then define the boundary stability data:
[0168] ;
[0169] The closer the value is to one, the smaller the change in the boundary position before and after suppression. For energy fidelity data, it is defined as:
[0170] ;
[0171] in, This represents the set of high-retention-weighted segments determined by reflection protection data.
[0172] Step S502: Construct consistency evaluation data based on the continuity data of the phase axis, the boundary stability data, and the energy fidelity data.
[0173] In step S502, the consistency evaluation data is defined as:
[0174] ;
[0175] in, These are the weighting coefficients.
[0176] Step S503: Feed back the consistency evaluation data to the deep learning suppression model and update the parameters of the deep learning suppression model.
[0177] In step S503, in order to update the parameters, this embodiment constructs a total loss function:
[0178] ;
[0179] in, For noise attribution loss, For reflection protection loss, This is the consistency evaluation loss term. The noise attribution loss term can be expressed as:
[0180] ;
[0181] in, This is a reference label for noise attribution. The reflection protection loss term can be expressed as:
[0182] ;
[0183] in, For reflection protection reference labels. The consistency evaluation loss term is defined as:
[0184] ;
[0185] This shows that when the initial suppression results lead to a decrease in reflection continuity, boundary stability, or energy fidelity, the consistency evaluation data decreases, and the corresponding consistency evaluation loss term increases, thereby driving the deep learning suppression model to adjust its parameters in subsequent iterations.
[0186] Step S504: Based on the parameter-updated deep learning suppression model, the original 3D seismic signal data is suppressed again, and the target suppression result data and suppression confidence data are output.
[0187] The target suppression result data is denoted as Suppression confidence data is calculated based on the stability of noise attribution data and reflection protection data, specifically as follows:
[0188] ;
[0189] in, and This is the output of the model after parameter updates. When the conflict between the reflection protection judgment and the noise attribution judgment is small, that is, when the model has a clear tendency to believe that the point should be suppressed or retained, the suppression confidence data is high; when the conflict between the two is large, the suppression confidence data is low, which is used to prompt the interpretation terminal to focus on verification.
[0190] Step S6 includes the following sub-steps:
[0191] Step S601: Obtain target suppression result data and suppression confidence data, and establish the correspondence between target suppression result data, suppression confidence data, and noise scene label data, thus forming a record group. .
[0192] In step S602, the corresponding target suppression result data and suppression confidence data are input to the control terminal for display, and suppression result record data is generated. When displaying the data, the control terminal can display the target suppression result data in waveform profile form and overlay the suppression confidence data in confidence heat form, so that interpreters can intuitively see the structural information after suppression and the corresponding confidence level.
[0193] Step S603: The suppression result record data is stored in the database, and the target suppression result data and suppression confidence data are transmitted to the interpretation terminal for subsequent fault identification, cavity boundary interpretation or stratigraphic tracking.
[0194] In this embodiment, the deep learning suppression model can adopt an encoder and a dual-branch decoder structure. The encoder receives the original 3D seismic signal data and time-frequency characterization data, and extracts shared features; the noise attribution branch outputs noise attribution data; the reflection protection branch outputs reflection protection data; the joint suppression layer generates initial suppression result data based on the noise attribution data and reflection protection data; the consistency feedback module calculates consistency evaluation data based on the initial suppression result data and feeds back the loss, thereby updating the model parameters. During the training phase, iterative training is performed using stored sample pairs in the database. During the inference phase, only the original 3D seismic signal data, noise scene label data, noise prior data, and time-frequency characterization data are input to output the target suppression result data and suppression confidence data.
[0195] Therefore, this embodiment constructs noise prior data by using silent segment data, wake segment data, abnormal segment data, and repeated acquisition segment data. Then, it outputs noise attribution data and reflection protection data simultaneously through a deep learning suppression model. Furthermore, it uses consistency evaluation data to optimize the suppression process, enabling the target suppression result data to more stably retain continuous reflection segments and weak reflection segments while achieving noise suppression. This approach is suitable for intelligent suppression of seismic signals under complex engineering disturbance conditions.
[0196] Example 2, refer to Figure 2 This paper presents a deep learning-based intelligent noise suppression system for seismic signals, which includes a data acquisition and storage module, a priori feature construction module, a dual-branch identification module, an intelligent suppression processing module, a consistency feedback optimization module, and a result output transmission module.
[0197] The data acquisition and storage module is used to acquire raw 3D seismic signal data and noise scene label data of the target exploration area, and store the raw 3D seismic signal data and noise scene label data in the database.
[0198] The prior feature construction module is used to extract silent segment data, wakeband data, anomalous segment data, and repeated acquisition segment data from the original 3D seismic signal data, and to construct noise prior data based on the silent segment data, wakeband data, anomalous segment data, and repeated acquisition segment data. At the same time, it constructs time-frequency characterization data based on the original 3D seismic signal data.
[0199] The dual-branch identification module is used to input the original 3D seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the deep learning suppression model, and output noise attribution data and reflection protection data.
[0200] The intelligent suppression processing module is used to suppress the original 3D seismic signal data based on noise attribution data and reflection protection data, and output the initial suppression result data.
[0201] The consistency feedback optimization module is used to construct consistency evaluation data based on the initial suppression result data and the original 3D seismic signal data, and to update the parameters of the deep learning suppression model based on the consistency evaluation data, outputting the target suppression result data and suppression confidence data.
[0202] The result output transmission module is used to input the target suppression result data and suppression confidence data to the control terminal for display, storage and transmission.
[0203] This invention addresses the 3D seismic acquisition process in coal exploration, specifically the problem of multiple types of noise intertwined in raw 3D seismic signal data, including surface wave interference, random noise, industrial interference, coupling anomalies, and shallow scattering interference. It constructs prior noise data by extracting data from silent segments, wake segments, anomalous segments, and repeated acquisition segments. This allows the deep learning-based noise suppression model to not only rely on the raw 3D seismic signal data itself but also to incorporate prior constraints on noise components based on the energy distribution, continuity, and repetition characteristics of different noise segments. This improves the accuracy of identifying complex noise and enhances the targeted suppression capability of noise components in 3D seismic acquisition signals.
[0204] This invention inputs raw 3D seismic signal data, noise scene label data, noise prior data, and time-frequency characterization data into a deep learning suppression model, and outputs noise attribution data and reflection protection data respectively. This enables the suppression process to not only identify noise components that need to be weakened, but also simultaneously identify effective and weak reflection components that should be retained. Compared to existing methods that directly output suppression results through a single mapping relationship, this invention can implement differentiated suppression based on the noise category, noise intensity, and reflection retention weight of different sections, avoiding the weakening of effective reflection components along with the noise, thus improving the signal-to-noise ratio of the raw 3D seismic signal data.
[0205] This invention, after suppression processing, further extracts phase axis continuity data, boundary stability data, and energy fidelity data, and constructs consistency evaluation data to feed back into the deep learning suppression model for parameter updates. This ensures that the model optimization objective not only focuses on noise energy reduction but also on whether the phase axis is continuous after suppression, whether the effective reflection components are preserved, and whether the geological interface is clear. In this way, this invention effectively alleviates the problems of over-smoothing, phase axis breakage, and effective reflection wave distortion commonly found in existing technologies, making the suppressed target data more suitable for subsequent 3D seismic data interpretation and processing.
[0206] The target suppression result data output by this invention is accompanied by suppression confidence data, and after establishing a correspondence with the noise scene label data, it is transmitted to the 3D seismic data interpretation terminal. This allows interpreters to not only obtain the noise-suppressed seismic data when using the suppression results, but also to identify the sections with high confidence and the sections that need to be reviewed in detail, thereby improving the reliability of 3D seismic data interpretation and processing and its engineering application value.
[0207] This invention, through processing mechanisms such as noise prior data construction, noise attribution and reflection protection collaborative identification, differential suppression and consistency feedback optimization, can effectively improve the signal-to-noise ratio of the original 3D seismic signal data during the 3D seismic acquisition process in coal exploration, stably retain the effective reflection components and phase axis continuity, improve the accuracy of coal seam interface interpretation, fault interpretation accuracy and structural interpretation reliability, and has good practical application effects.
[0208] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0209] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.
Claims
1. A deep learning-based intelligent noise suppression method for seismic signals, characterized in that, Includes the following steps: Step S1: Collect the original 3D seismic signal data and noise scene label data of the target exploration area, and store the original 3D seismic signal data and noise scene label data in the database; Step S2: Extract silent segment data, wakeband data, anomalous segment data, and repeated acquisition segment data from the original 3D seismic signal data, and construct noise prior data based on the silent segment data, wakeband data, anomalous segment data, and repeated acquisition segment data. At the same time, construct time-frequency characterization data based on the original 3D seismic signal data. Step S3: Input the original 3D seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the deep learning suppression model, and output noise attribution data and reflection protection data. Step S4: Suppress the original 3D seismic signal data based on noise attribution data and reflection protection data, and output the initial suppression result data; Step S5: Construct consistency evaluation data based on the initial suppression result data and the original 3D seismic signal data, update the parameters of the deep learning suppression model based on the consistency evaluation data, and output the target suppression result data and suppression confidence data. Step S6: Input the target suppression result data and suppression confidence data to the control terminal for display, storage and transmission.
2. The intelligent noise suppression method for seismic signals based on deep learning as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Use a three-dimensional seismic acquisition device to acquire raw three-dimensional seismic signal data in the target exploration area. The raw three-dimensional seismic signal data includes amplitude sequence data arranged by trace and corresponding sampling time sequence data. Step S102: Collect noise scene label data corresponding to the original three-dimensional seismic signal data. The noise scene label data includes at least one of surface wave interference label data, random noise label data, industrial interference label data, coupling anomaly label data, and shallow scattering interference label data. Step S103: Establish the correspondence between the original 3D seismic signal data and the noise scene label data, and store the corresponding original 3D seismic signal data and noise scene label data in the database.
3. The intelligent noise suppression method for seismic signals based on deep learning as described in claim 2, characterized in that, Step S2 includes the following sub-steps: Step S201: Extract the low-energy region before excitation based on the original three-dimensional seismic signal data and output the silent segment data; Step S202: Extract the attenuation interval after the effective reflection ends based on the original three-dimensional seismic signal data, and output the coma band data; Step S203: Extract amplitude mutation intervals, frequency mutation intervals, and continuity interruption intervals from the original three-dimensional seismic signal data, and output the anomaly segment data; Step S204: Extract the repeated response intervals from the original three-dimensional seismic signal data repeatedly acquired in the same area, and output the repeated acquisition segment data; Step S205: Calculate noise intensity data, noise continuity data, and noise repeatability data based on silent segment data, wake wave segment data, abnormal segment data, and repeated acquisition segment data, and combine the noise intensity data, noise continuity data, and noise repeatability data into noise prior data. Step S206: Perform short-time time-frequency transformation on the original three-dimensional seismic signal data and output time-frequency characterization data.
4. The intelligent noise suppression method for seismic signals based on deep learning as described in claim 3, characterized in that, The logic for constructing the noise prior data is as follows: Calculate the energy distribution data, frequency band occupancy data, and inter-channel similarity data for the silent segment data, wakeband data, abnormal segment data, and repeated acquisition segment data, respectively; The corresponding segments are divided according to silent segment data, wake segment data, abnormal segment data and repeated acquisition segment data, and segment identification data is generated for each corresponding segment. Noise intensity data is calculated based on the energy distribution data of each corresponding section, and the noise intensity data is then bound to the corresponding section identification data. Noise continuity data is calculated based on the frequency band occupancy data of each corresponding segment, and the noise continuity data is bound to the corresponding segment identification data. Noise repeatability data is calculated based on the inter-track similarity data of each corresponding section, and the noise repeatability data is bound to the corresponding section identification data. The noise prior data is output by combining the segment identification data, noise intensity data, noise continuity data, and noise repeatability data.
5. The intelligent noise suppression method for seismic signals based on deep learning as described in claim 4, characterized in that, Step S3 includes the following sub-steps: Step S301: Input the original 3D seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the noise attribution branch of the deep learning suppression model, and output the noise attribution data. Step S302: Input the original 3D seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the reflection protection branch of the deep learning suppression model, and output the reflection protection data. Step S303: Establish the trace-level correspondence between noise attribution data and reflection protection data and the original three-dimensional seismic signal data, and output it to the suppression processing module.
6. The intelligent noise suppression method for seismic signals based on deep learning as described in claim 5, characterized in that, The processing logic of the deep learning suppression model is as follows: Based on the noise scene label data, the original 3D seismic signal data is subjected to scene condition coding, and the scene coded data is output. Based on noise prior data and time-frequency characterization data, the noise components in the original 3D seismic signal data are classified and noise attribution data is output. Based on scene coding data, raw 3D seismic signal data, and time-frequency characterization data, continuous reflection segments and weak reflection segments in the raw 3D seismic signal data are identified, and reflection protection data is output. Based on noise attribution data and reflection protection data, the noise component in the original 3D seismic signal data is attenuated while the reflection component is preserved, and the initial suppression result data is output.
7. The intelligent noise suppression method for seismic signals based on deep learning as described in claim 6, characterized in that, Step S4 includes the following sub-steps: Step S401: Obtain noise attribution data and determine the noise category and noise intensity corresponding to each segment in the original three-dimensional seismic signal data; Step S402: Obtain reflection protection data and determine the reflection preservation weights corresponding to each segment in the original three-dimensional seismic signal data; Step S403: Based on the noise category, noise intensity, and reflection retention weight, perform differentiated attenuation processing on each segment of the original three-dimensional seismic signal data, and output the initial suppression result data.
8. The intelligent noise suppression method for seismic signals based on deep learning as described in claim 7, characterized in that, Step S5 includes the following sub-steps: Step S501: Extract the phase axis continuity data, boundary stability data, and energy fidelity data based on the initial suppression result data; Step S502: Construct consistency evaluation data based on in-phase axis continuity data, boundary stability data, and energy fidelity data; Step S503: Feed the consistency evaluation data back to the deep learning suppression model and update the parameters of the deep learning suppression model. Step S504: Based on the parameter-updated deep learning suppression model, the original 3D seismic signal data is suppressed again, and the target suppression result data and suppression confidence data are output.
9. The intelligent noise suppression method for seismic signals based on deep learning as described in claim 8, characterized in that, Step S6 includes the following sub-steps: Step S601: Obtain target suppression result data and suppression confidence data, and establish the correspondence between target suppression result data, suppression confidence data and noise scene label data; Step S602: Input the corresponding target suppression result data and suppression confidence data to the control terminal for display, and generate suppression result record data; Step S603: Store the suppression result record data in the database, and transmit the target suppression result data and suppression confidence data to the interpretation terminal.
10. A deep learning-based intelligent noise suppression system for seismic signals, applied in any one of the deep learning-based intelligent noise suppression methods for seismic signals as described in claims 1-9, characterized in that, It includes a data acquisition and storage module, a priori feature construction module, a dual-branch identification module, an intelligent suppression processing module, a consistency feedback optimization module, and a result output and transmission module; The data acquisition and storage module is used to acquire raw 3D seismic signal data and noise scene label data of the target exploration area, and store the raw 3D seismic signal data and noise scene label data into the database. The prior feature construction module is used to extract silent segment data, wakeband data, anomalous segment data and repeated acquisition segment data from the original three-dimensional seismic signal data, and to construct noise prior data based on the silent segment data, wakeband data, anomalous segment data and repeated acquisition segment data, while constructing time-frequency characterization data based on the original three-dimensional seismic signal data. The dual-branch identification module is used to input the original three-dimensional seismic signal data, noise scene label data, noise prior data and time-frequency characterization data into the deep learning suppression model, and output noise attribution data and reflection protection data. The intelligent suppression processing module is used to suppress the original three-dimensional seismic signal data based on noise attribution data and reflection protection data, and output the initial suppression result data. The consistency feedback optimization module is used to construct consistency evaluation data based on the initial suppression result data and the original three-dimensional seismic signal data, and to update the parameters of the deep learning suppression model based on the consistency evaluation data, and output the target suppression result data and suppression confidence data. The result output transmission module is used to input the target suppression result data and suppression confidence data to the control terminal for display, storage and transmission.
