A method and system for denoising data of a passive source ocean bottom seismometer

Abstract

The present application belongs to the technical field of seismic exploration, and particularly relates to a method and system for data denoising of an active source bottom seismograph, which is used to solve the problems of low signal-to-noise ratio of measured OBS active source data and effective seismic phase being covered by complex ocean background noise. The method comprises the following steps: extracting pure noise from measured active source seismic records to construct a sample library; selecting noise blocks to generate a continuous noise field through overlapping and weighted fusion, and then superimposing the continuous noise field with pure records to construct noisy records and form a training sample pair; dividing the sample pair into blocks and transforming the sample pair into a frequency-wavenumber domain, taking the real and imaginary parts of the spectrum as double-channel input of a deep convolution network, and training to obtain a mapping from noisy spectrum to pure spectrum; dividing the data to be processed into blocks and transforming the data into a frequency-wavenumber domain, obtaining a predicted denoised spectrum through the trained network, and then inversely transforming the data back to a time-space domain, overlapping and weighting the data blocks to reconstruct the data blocks, and obtaining denoised seismic records, which can be used to improve the reliability of subsequent seismic phase analysis and velocity model inversion.

Description

Technical Field

[0001] This invention belongs to the field of earthquake detection technology, specifically relating to a method and system for denoising data from an active source seafloor seismograph. Background Technology

[0002] Active-source seismic exploration is a crucial technique for analyzing marine crustal structure and conducting deep tectonic research. During operations, seismic waves are generated using artificial sources such as air guns, and data acquisition is completed via towed cables and ocean bottom seismographs (OBS). However, compared to onshore seismic exploration, the marine observation environment is far more complex. Ocean turbulence, instrument attitude disturbances, ocean currents, and various background clutter all contribute to the overall low signal-to-noise ratio of measured seismic data. Effective reflected signals are easily masked by strong background noise, exhibiting weak amplitude and poor continuity, which reduces the imaging accuracy of subsequent stages such as velocity analysis, migration imaging, and wave impedance inversion. Therefore, seismic signal denoising is an indispensable pre-processing step in the marine seismic data processing workflow.

[0003] Current seismic denoising methods, particularly traditional model-driven methods, primarily include filtering, sparse transformation, and low-rank matrix decomposition. These methods rely on prior knowledge of noise characteristics, requiring parameter adjustments for different stations in practical applications, resulting in a large workload and relatively limited generalization ability. With the rise of deep learning, convolutional neural networks have been gradually applied to seismic denoising. However, existing methods mainly employ end-to-end time-domain processing, making it difficult to identify signals from strong noise using only time-domain waveforms. This can lead to excessive noise residue or over-smoothing that loses stratigraphic details. When transforming to the frequency-wavenumber domain (FK domain) for feature learning, conventional U-Net networks, being single-channel input / output, cannot simultaneously preserve the amplitude and phase characteristics of seismic records. Furthermore, most existing training datasets are constructed using Gaussian white noise superposition to synthesize records. However, Gaussian noise training models often perform poorly when applied to actual seabed work areas. This is because Gaussian noise models cannot reflect the true distribution of non-Gaussian composite noise at sea, making it difficult for the model to learn from real noise environments. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention provides a method and system for denoising active source seafloor seismograph data, aiming to enhance the continuity of seismic phases in OBS signals, expand the range of identifiable phases, and provide a more reliable basis for subsequent phase analysis and velocity model inversion.

[0005] A first aspect of this invention provides a method for denoising active-source seafloor seismograph data, the method comprising: Pure noise was extracted from measured active source seismic records to construct a real noise sample library; Noise blocks are selected from the real noise sample library, and the selected noise blocks are overlapped and weighted to fuse them to synthesize a noise field with continuous boundaries. The noise field is superimposed with the clean seismic record to obtain a noisy seismic record, and the noisy seismic record and its corresponding clean seismic record constitute a training sample pair. The training samples are divided into blocks and transformed to the frequency-wavenumber domain to obtain the noisy spectrum and the clean spectrum in the frequency-wavenumber domain. The real and imaginary parts of the noisy spectrum are used as two channels to input a deep convolutional network, and the real and imaginary parts of the corresponding clean spectrum are used as supervision labels for training to obtain the mapping relationship between the noisy spectrum and the clean spectrum. The seismic data to be processed is divided into blocks and transformed to the frequency-wavenumber domain. The blocks are then input into the trained deep convolutional network to obtain a denoised spectrum. The denoised spectrum is then subjected to an inverse transformation from the frequency-wavenumber domain to the spatiotemporal domain. Finally, the data blocks obtained by the inverse transformation are overlapped and weighted to reconstruct the denoised seismic record.

[0006] In some embodiments, the step of extracting pure noise from measured active-source seismic records and constructing a real noise sample library includes: From the bandpass filtered measured active source seismic records, regions that do not contain effective reflected waves, refracted waves, diffracted waves and surface waves are selected as pure noise regions, and multiple pure noise samples are obtained by cropping from the pure noise regions. A random amplitude perturbation is applied to each of the pure noise samples, wherein the random amplitude perturbation is the pure noise sample multiplied by an amplitude scaling factor randomly selected within a preset value range; A sliding window is used to slice the pure noise sample after the random amplitude perturbation to obtain multiple noise blocks, and all the noise blocks constitute the real noise sample library.

[0007] In some embodiments, the clean seismic record is obtained as follows: Based on the horizontal layered velocity model, a noise-free active source seismic record is generated by forward modeling using the wave equation, which serves as the pure seismic record. The forward modeling used a Ricker wavelet as the source, and the receivers were set up along the simulated seabed interface to simulate the location of the seabed seismograph. The shot points were set up along the survey line according to the preset shot spacing.

[0008] In some embodiments, the step of performing overlapping weighted fusion on the selected noise blocks to synthesize a noise field with continuous boundaries includes: Each selected noise block is weighted using a two-dimensional Hanning window, which is the product of a one-dimensional Hanning window in the vertical direction of the time axis and a one-dimensional Hanning window in the horizontal direction of the spatial axis. The weighted noise blocks are weighted and averaged according to their overlapping areas, and then fused to obtain a noise field with the same size as the pure seismic record. The noise field obtained by fusion is then subjected to mean removal, standard deviation normalization, and random adjustment of noise intensity.

[0009] In some embodiments, the step of dividing the training sample pairs into blocks and transforming them to the frequency-wavenumber domain to obtain noisy and clean frequencies-wavenumber spectra, using the real and imaginary parts of the noisy spectra as two channels to input a deep convolutional network, and using the real and imaginary parts corresponding to the clean spectra as supervision labels for training includes: The noisy seismic records and the clean seismic records are each divided into multiple data blocks of the same size; For each data block, a two-dimensional Fourier transform is performed without applying a window function or padding with zeros to obtain the corresponding frequency-wavenumber domain spectrum. The network input is formed by the real and imaginary parts of the data blocks corresponding to the noisy seismic records, and the supervision labels are formed by the real and imaginary parts of the data blocks corresponding to the clean seismic records, thus forming a supervised training dataset in the frequency-wavenumber domain.

[0010] In some embodiments, before transforming the training sample pairs into the frequency-wavenumber domain after partitioning them into blocks to obtain noisy and clean frequencies-wavenumber spectra, using the real and imaginary parts of the noisy spectra as two channels to input a deep convolutional network, and using the real and imaginary parts of the corresponding clean spectra as supervision labels for training, the method further includes: The spectrum of the input deep convolutional network is normalized, and the normalization is performed separately for each training sample, including: For a noisy spectrum of a training sample, construct a two-channel tensor from its real and imaginary parts. The root mean square scale of the dual-channel tensor is calculated using the following formula. : Wherein, the root mean square scale The calculation covers the real part, the imaginary part, and all frequency-wavenumber domain spectral points, and the real part and the imaginary part are normalized using the same root mean square scale; Divide the noisy spectrum by the root mean square scale. This is then used as input to the deep convolutional network, and the clean spectrum corresponding to the noisy spectrum is divided by the same root mean square scale. Later used as a supervisory label; When denoising the seismic data to be processed using the trained deep convolutional network, the root mean square scale is calculated according to the spectrum obtained by the block transformation of the seismic data to be processed. And after the output of the deep convolutional network, it is multiplied by the root mean square scale. .

[0011] In some embodiments, the deep convolutional network is a two-dimensional residual U-Net network, which includes an encoder and a decoder. The input and output of the two-dimensional residual U-Net network are both two channels, which correspond to the real part and imaginary part of the frequency-wavenumber domain spectrum, respectively. The encoder includes a first residual convolutional block, a first downsampling layer, a second residual convolutional block and a second downsampling layer connected in sequence, and a third residual convolutional block located at the bottom. The first residual convolutional block expands the number of channels from two to a first number of channels, and the second residual convolutional block expands the number of channels from the first number of channels to a second number of channels. The decoder includes, in sequence, a first upsampling, concatenating the first upsampling result with the feature map output by the second residual convolutional block in the encoder, a fourth residual convolutional block, a second upsampling, concatenating the second upsampling result with the feature map output by the first residual convolutional block in the encoder, a fifth residual convolutional block, and a convolutional layer that outputs the two channels. Each residual convolutional block includes two convolutional layers, batch normalization, and an activation function. When the number of input channels is inconsistent with the number of output channels, the number of channels in the shortcut branch is adjusted through convolution. The downsampling type is average pooling, and the upsampling type is bilinear upsampling.

[0012] In some embodiments, the loss function used to train the deep convolutional network is a weighted sum of the complex spectrum reconstruction error loss, the spectral energy error loss, and the spectral magnitude gradient error loss, expressed as: In the formula, The complex spectrum reconstruction error loss is characterized by the mean square error between the predicted spectrum and the pure spectrum output by the deep convolutional network in both the real and imaginary channels. The spectral energy error loss characterizes the error between the energy of the predicted spectrum and the energy of the noisy spectrum; The spectral amplitude gradient error loss characterizes the gradient difference between the amplitude of the predicted spectrum and the amplitude of the pure spectrum in the horizontal and vertical directions. The weighting coefficients for the spectral energy error loss are... These are the weighting coefficients for the spectral amplitude gradient error loss; The Adam optimizer is used during training, and after each training round, the current network parameters are saved when the average training loss is lower than the previously recorded minimum. The network parameters with the lowest loss during training are used as the trained deep convolutional network.

[0013] In some embodiments, the process involves dividing the seismic data to be processed into blocks and transforming it to the frequency-wavenumber domain, inputting the trained deep convolutional network to obtain a denoised spectrum, performing an inverse frequency-wavenumber domain to spatiotemporal domain transformation on the denoised spectrum, and then performing overlapping weighted reconstruction on the data blocks obtained by the inverse transformation to obtain a denoised seismic record, including: The seismic data to be processed is bandpass filtered, and the seismic data to be processed, the measured active source seismic records used to extract pure noise, and the pure seismic records use the same bandpass filtering parameters. The bandpass filtered seismic data to be processed is divided into multiple data blocks according to the preset data block size and sliding step size. The preset data block size and sliding step size are the same as those during training. A two-dimensional Fourier transform is performed on each data block to be processed to obtain the frequency-wavenumber domain spectrum to be processed. The real part and imaginary part of the spectrum to be processed are input into the trained deep convolutional network to obtain the corresponding denoised spectrum. The denoised spectrum is subjected to a two-dimensional inverse Fourier transform to obtain the corresponding denoised data block. The denoised data block is then weighted and averaged according to the overlapping area using a two-dimensional Hanning window to reconstruct the denoised seismic record.

[0014] A second aspect of this invention provides a data denoising system for active source seafloor seismometers, the system comprising: The noise library construction module is used to extract pure noise from measured active source seismic records and construct a real noise sample library. The sample synthesis module is used to select noise blocks from the real noise sample library, perform overlapping and weighted fusion on the selected noise blocks, synthesize a noise field with continuous boundaries, superimpose the noise field with the clean seismic record to obtain a noisy seismic record, and use the noisy seismic record and its corresponding clean seismic record to form a training sample pair. The network training module is used to divide the training samples into blocks and transform them to the frequency-wavenumber domain to obtain the noisy spectrum and the clean spectrum in the frequency-wavenumber domain. The real and imaginary parts of the noisy spectrum are used as two channels to input a deep convolutional network, and the real and imaginary parts of the corresponding clean spectrum are used as supervision labels for training to obtain the mapping relationship between the noisy spectrum and the clean spectrum. The denoising and reconstruction module is used to divide the seismic data to be processed into blocks and transform it to the frequency-wavenumber domain. The denoised spectrum is then input into the trained deep convolutional network to obtain the denoised spectrum. The denoised spectrum is then subjected to an inverse transformation from the frequency-wavenumber domain to the spatiotemporal domain. Finally, the data blocks obtained by the inverse transformation are overlapped and weighted to reconstruct the denoised seismic record.

[0015] The beneficial effects of this invention include: By generating noisy training samples through random splicing and superposition, compared with methods using Gaussian white noise or artificially simulated noise, the training data used in this method can more realistically reflect the characteristics of ocean noise and improve the model's adaptability to actual working conditions. Transforming OBS time-domain data to the frequency-wavenumber domain makes the effective signal and random noise have a more obvious distribution difference in the feature space, which enhances the separability of signal and noise and provides a more favorable feature representation basis for subsequent deep learning denoising. By using the real and imaginary parts of the frequency-wavenumber domain spectrum as input channels for the neural network, the spectral amplitude and phase information are preserved, avoiding information loss in traditional amplitude spectrum processing, which is beneficial to improving the seismic waveform recovery capability and the continuity of the phase axis. By learning the nonlinear mapping relationship between the noisy spectrum and the clean spectrum under real noise conditions through deep neural networks, complex non-stationary noise can be adaptively suppressed without the need to manually set filter boundaries, velocity thresholds or noise models. While effectively reducing various types of noise, it better maintains the continuity of the phase axis and wavefield structure characteristics of the strong signal from small shot distance and the weak signal from large shot distance, reduces the effective signal loss caused by traditional filtering methods, and improves the signal-to-noise ratio and interpretation reliability of seismic data. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a method for denoising data from an active-source seafloor seismograph, as shown in an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S1 in an embodiment of the present invention; Figure 3 This is a flowchart illustrating step S2 in an embodiment of the present invention; Figure 4 A flowchart illustrating step S3 in an embodiment of the present invention; Figure 5 This is another flowchart illustrating step S3 in an embodiment of the present invention; Figure 6 This is a schematic diagram of a deep convolutional network structure shown in an embodiment of the present invention; Figure 7 A flowchart illustrating step S4 in an embodiment of the present invention; Figure 8This is a comparison of the results of this method and the noise suppression method at station C10 in the OBS2016 survey line in the South Yellow Sea, as shown in an embodiment of the present invention. Figure 9 This is a partial comparison of the results of this method and the noise suppression method at station C10 in the OBS2016 survey line in the South Yellow Sea, as shown in an embodiment of the present invention. Figure 10 This is a comparison diagram of the method and the noise suppression method at station C12 in the OBS2016 survey line in the South Yellow Sea, as shown in an embodiment of the present invention. Figure 11 This is a structural block diagram of an active source seafloor seismograph data denoising system, as shown in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] The following will clearly and completely describe the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in the present invention can be combined interactively without contradicting each other.

[0020] like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for denoising active source seafloor seismograph data, the method comprising: S1. Extract pure noise from measured active source seismic records and construct a real noise sample library.

[0021] In this embodiment, the raw seismic records obtained during offshore operations using artificial seismic sources such as air guns to generate seismic waves and acquired by towed cables or seabed seismometers are used as the noise source. The measured active-source seismic records simultaneously contain effective reflected waves, refracted waves, diffracted waves, surface waves, as well as noise components such as ocean turbulence, instrument attitude disturbances, ocean current effects, and various background clutter. Pure noise refers to data segments in these records that contain no effective seismic signals and only retain the actual ambient noise from the acquisition environment.

[0022] The real noise sample library is a dataset that centrally stores these pure noise segments. The purpose of constructing the real noise sample library is to ensure that the noise added to the clean records during subsequent training comes from the real marine environment, rather than artificially simulated Gaussian noise, so that the distribution pattern learned by the model is closer to actual working conditions. The execution method involves manually selecting regions in the original records that do not contain effective seismic phases as pure noise regions, and then cropping multiple pure noise segments from these regions and incorporating them into the sample library.

[0023] S2. Select noise blocks from the real noise sample library, perform overlapping and weighted fusion on the selected noise blocks to synthesize a noise field with continuous boundaries, superimpose the noise field with the clean seismic record to obtain the noisy seismic record, and use the noisy seismic record and its corresponding clean seismic record to form a training sample pair.

[0024] First, several noise segments of fixed size are randomly selected from the sample library as noise blocks. A certain overlap area is maintained between adjacent noise blocks, and a weighted average is performed in the overlap area to ensure that the amplitude of the spliced ​​noise does not change abruptly.

[0025] The boundary-continuous noise field is obtained after overlapping and weighted fusion. Its size is consistent with the pure seismic record, and it is a whole noise field with smooth transitions between adjacent regions without obvious seams. The pure seismic record is a theoretical seismic record without any noise.

[0026] Noisy seismic records are the result of adding clean seismic records and noise fields point-by-point at the same coordinate positions. Training sample pairs refer to pairs of data consisting of a noisy seismic record and its corresponding clean seismic record; the former serves as network input, and the latter as supervision labels.

[0027] S3. After dividing the training samples into blocks, transform them to the frequency-wavenumber domain to obtain the noisy spectrum and the clean spectrum in the frequency-wavenumber domain. Use the real and imaginary parts of the noisy spectrum as two channels to input a deep convolutional network, and use the real and imaginary parts of the corresponding clean spectrum as supervision labels to train the network, thereby obtaining the mapping relationship between the noisy spectrum and the clean spectrum.

[0028] Each noisy seismic record and its corresponding clean seismic record are divided into multiple data blocks of the same size. Then, a two-dimensional Fourier transform is performed on each data block to map the time-space domain record to the frequency-wavenumber domain with frequency and wavenumber as coordinates.

[0029] The noisy spectrum and the clean spectrum are complex spectra obtained after performing a two-dimensional Fourier transform on the noisy data block and the clean data block, respectively. Since the result of the two-dimensional Fourier transform is a complex number containing real and imaginary parts, which together carry the amplitude and phase information of the spectrum, the real part is used as one channel and the imaginary part as another, forming a dual-channel input. During training, the deep convolutional network continuously adjusts its parameters to learn the correspondence between the noisy spectrum and the clean spectrum. This mapping relationship is expressed as: In the formula, This represents a deep convolutional network model; Indicates network parameters; This represents the noisy spectrum of the input; This indicates the predicted spectrum.

[0030] S4. After dividing the seismic data to be processed into blocks, transform it to the frequency-wavenumber domain, input it into the trained deep convolutional network to obtain the denoised spectrum, perform an inverse transformation from the frequency-wavenumber domain to the spatiotemporal domain on the denoised spectrum, and perform overlapping weighted reconstruction on each data block obtained by the inverse transformation to obtain the denoised seismic record.

[0031] For actual seismic data collected on-site that requires noise suppression, the data is divided into blocks in the same way as during training and subjected to a two-dimensional Fourier transform. The real and imaginary parts of each data block are then fed into a pre-trained deep convolutional network, which outputs the predicted denoised spectrum.

[0032] The denoised spectrum refers to the frequency-wavenumber domain complex spectrum obtained by network prediction after noise suppression. A two-dimensional inverse Fourier transform is performed on the denoised spectrum using an inverse transform to restore the frequency-wavenumber domain result to the time-space domain. The overlapping denoised data blocks obtained from the inverse transform are then weighted and averaged according to the overlapping regions, and stitched together to reconstruct a complete denoised seismic record.

[0033] Specifically, for active source observation data from seafloor seismometers with significant differences between station and shot point intervals, regions lacking effective phases are manually selected from the bandpass-filtered measured records, and pure noise samples are cropped and included in a real noise sample library. Subsequently, multiple noise blocks are randomly extracted from the sample library, weighted using a two-dimensional Hanning window, and then laid out according to overlapping areas and weighted averaged to synthesize a noise field with the same size and continuous boundaries as the clean seismic record. This noise field is then superimposed on the clean seismic record generated by forward modeling of the wave equation, forming pairs of noisy and clean seismic records. Next, the paired records are divided into data blocks with 256 sampling points in the time direction and 256 channels in the spatial direction. A two-dimensional Fourier transform is performed on each data block to obtain the frequency-wavenumber domain spectrum. The real and imaginary parts of the noisy spectrum are used as two input channels, and the real and imaginary parts of the clean spectrum are used as supervision labels. This data is fed into a deep convolutional network for repeated iterations to learn the mapping from the noisy spectrum to the clean spectrum. Finally, the same block division and two-dimensional Fourier transform were performed on the on-site seismic data to be processed. The spectrum was fed into the trained network to obtain the denoised spectrum. The denoised spectrum was then subjected to a two-dimensional inverse Fourier transform to restore it to the time and space domain. The overlapping denoised data blocks were then weighted and averaged using a two-dimensional Hanning window to reconstruct the complete denoised seismic record. This process suppressed complex background noise while preserving the continuity of effective seismic phases.

[0034] Among the methods for extracting pure noise regions from measured records, interpreters can manually select regions without effective phases based on the phase axis morphology and energy distribution. Alternatively, bandpass filtering can be performed first, followed by cropping in regions with weak and irregularly distributed residual energy. Frequency-wavenumber domain transformation can be achieved by directly performing a two-dimensional fast Fourier transform on each data block. Another method involves multiplying the forward and inverse transforms by the inverse of the data block area without multiplying by the normalization coefficients. Overlapping weighted reconstruction of adjacent data blocks can be achieved by using a two-dimensional Hanning window and weighted averaging based on the overlapping area. A third method involves accumulating the weighted results of each data block and then dividing by the accumulated weight to complete normalization and concatenation.

[0035] This embodiment extracts real ocean noise from measured data and synthesizes noisy training samples, enabling the training data to reflect the non-Gaussian and non-stationary characteristics of ocean noise, thereby improving the model's adaptability after being transferred to actual seabed work areas. By using a two-dimensional Fourier transform to map the time-space domain data to the frequency-wavenumber domain, the effective signal and noise show significant differences in their aggregation patterns on the frequency-wavenumber trajectory, enhancing the separability of signal and noise. The real and imaginary parts of the spectrum are treated as two separate channels, preserving both amplitude and phase information, avoiding the loss of phase information caused by using only the amplitude spectrum, and improving the continuity of the phase axis and waveform fidelity after the inverse transform. The combination of block transform and overlapping weighted reconstruction allows for the full learning of long-term sparsely distributed phases block by block, avoiding the seams caused by directly splicing together blocks after independent processing. This effectively reduces various types of noise while maintaining wavefield structure characteristics, eliminating the need to adjust parameters individually for different stations.

[0036] In some embodiments, such as Figure 2 As shown, in step S1, pure noise is extracted from the measured active source seismic records to construct a real noise sample library, including: S101. From the bandpass filtered measured active source seismic records, select the region that does not contain effective reflected waves, refracted waves, diffracted waves and surface waves as the pure noise region, and obtain multiple pure noise samples from the pure noise region.

[0037] In the formula, Indicates the first One noise sample; This represents the total number of noise samples. The extracted noise samples only contain noise information from the actual acquisition environment and do not include active source seismic signals.

[0038] This step first applies bandpass filtering to the measured seismic records, suppressing components outside the frequency band of the effective seismic phase, thus ensuring the recorded frequency band matches the subsequent clean records. Bandpass filtering is a processing method that allows only signals within a specific frequency band to pass while suppressing components outside that band. Effective reflected waves, refracted waves, diffracted waves, and surface waves are all effective seismic phases carrying information about subsurface structures. Reflected waves originate from reflections at stratigraphic interfaces, refracted waves originate from sliding at the top interface of high-velocity layers, diffracted waves originate from scattering at subsurface discontinuities, and surface waves propagate along the surface or seabed.

[0039] The purpose of selecting regions that do not contain the aforementioned effective seismic phases as pure noise regions is to ensure that the cropped samples contain only the actual acquisition noise, and to extract several data segments containing only noise from these pure noise regions.

[0040] S102. Apply a random amplitude perturbation to each pure noise sample. The random amplitude perturbation is to multiply the pure noise sample by an amplitude scaling factor randomly selected within a preset value range.

[0041] The processed noise sample can be represented as: in: This is a random amplitude scaling factor, and its value range is limited to: , For the first Noise samples after applying random amplitude perturbations to a noise sample.

[0042] Amplitude scaling factor It is a scaling factor used to amplify or reduce the amplitude of noise samples. The preset value range refers to the upper and lower limits that the coefficient can take beforehand, for example, between 0.85 and 1.15. The purpose of applying random amplitude perturbation is to make the same batch of noise samples present different intensity levels, increase the diversity of noise samples, and prevent the network from seeing only noise of a single intensity, thus reducing its generalization ability.

[0043] S103. Using a sliding window, slice the pure noise sample after random amplitude perturbation to obtain multiple noise blocks, and the real noise sample library is composed of all noise blocks.

[0044] Use a fixed size ( The window is set at a certain step size on the pure noise samples after amplitude perturbation. The sliding window is defined as follows: L is the side length of the sliding window, and P is the sliding step size. At each sliding position, data within the window range is extracted as a noise block. After slicing, a large number of noise blocks are obtained. , This represents the total number of noise blocks obtained. A certain degree of overlap is maintained between adjacent windows. All noise blocks combined constitute a real noise sample library, which preserves the spectral structure characteristics, temporal correlation, and spatial correlation of the actual acquisition environment.

[0045] Specifically, for active source records from seafloor seismometers, four bandpass filters with turn frequencies of 2Hz, 3Hz, 9Hz, and 10Hz are first used to suppress components outside the effective frequency band. Then, interpreters select regions in the filtered records that do not contain reflected waves, refracted waves, diffracted waves, or surface waves as pure noise regions, from which multiple pure noise samples are extracted. Each pure noise sample is then multiplied by an amplitude scaling factor randomly selected between 0.85 and 1.15 to introduce random fluctuations in sample intensity. Next, a window with 256 sampling points in the time direction and 256 channels in the spatial direction is used, sliding across the amplitude-perturbed samples with a step size of 128 sampling points, overlapping adjacent windows by half, to extract a large number of noise blocks. All noise blocks are compiled to form a real noise sample library. The number of samples depends on the number of input files and the selected range; the total number of noise blocks directly counted during actual library construction is the library capacity.

[0046] The selection of pure noise regions can be done manually by interpreters based on the shape and energy distribution of the in-phase shaft, or automatically after bandpass filtering based on the criterion of weak energy and irregular distribution. The amplitude scaling factor can be uniformly and randomly sampled between 0.85 and 1.15, or randomly sampled within a pre-defined wider or narrower interval to match different intensity levels. The slicing method can be a fixed window of 256 sampling points in the time direction and 256 channels in the spatial direction, with an overlapping sliding step of 128 sampling points. Alternatively, the window size and step size can be adjusted to change the number of noise blocks and the overlap ratio.

[0047] This embodiment first performs bandpass filtering on the measured records before extracting noise, ensuring that the frequency band of the noise samples is consistent with the clean records and subsequent data to be processed, thus avoiding frequency band mismatch that could affect training and inference. Random amplitude perturbations are applied to each sample, expanding the diversity of noise samples in the intensity dimension, allowing the model to see a rich distribution of noise intensity, thereby improving its adaptability to noise of varying intensities. Overlapping sliding window slicing is used to build a library, preserving the spectral structure, temporal correlation, and spatial correlation of the actual acquisition environment. This ensures that the noise blocks subsequently extracted from the library carry the characteristics of real ocean noise, laying the foundation for synthesizing noisy training samples that closely resemble real-world conditions.

[0048] In some embodiments, the clean seismic record in step S2 is obtained as follows: Based on the horizontal layered velocity model, a noise-free active source seismic record was generated by forward modeling using the wave equation as a pure seismic record. The source used in the forward modeling was the Ricker wavelet, and the receiver points were arranged along the simulated seabed interface to simulate the location of the seabed seismograph station. The shot points were arranged along the survey line according to the preset shot spacing.

[0049] In this embodiment, a horizontally layered velocity model with flattened strata interfaces is first established, and then the propagation of seismic waves in the model is calculated by forward modeling using the wave equation to obtain a theoretical seismic record without any noise.

[0050] Among them, the horizontal layered velocity model is a model that abstracts the underground medium into several horizontal layers, and assigns corresponding velocity parameters to each layer. Its depth, width and number of layers can be set with reference to existing regional models, and the layer interfaces are flattened. The velocity of each layer is set according to the corresponding layer characteristics.

[0051] Next, based on the wave equation describing wave propagation, the propagation of the wave field in the model after the source excitation and the response received by the receiver point are numerically calculated. Because this record comes from theoretical calculations and is not superimposed with ocean currents, instrument noise, or other environmental noise, it is used as a clean seismic record and employed as a target label in supervised learning. Establishing a well-structured horizontally layered model is to obtain a theoretical record that is easy to use as a label, avoiding the influence of complex structures on the label.

[0052] In the forward modeling simulation, the excitation signal is set as the Ricker wavelet, and the receiving positions are placed on the simulated seabed interface. The excitation positions are arranged at fixed intervals along the survey line. The Ricker wavelet is a commonly used theoretical source wavelet with a definite dominant frequency and symmetrical waveform. Its dominant frequency can be set, for example, 8Hz. The receiver points refer to the locations where seismic waves are received. They are arranged along the simulated seabed interface to simulate the actual locations of seabed seismometers, and the station interval can be 6km. The shot points refer to the locations where seismic waves are excited. They are arranged along the survey line at a preset shot spacing, which can be 125m. The time sampling interval can be 2ms. The wavefield propagation and the response of the seabed receiver points are calculated shot by shot according to the above settings. Then, the results are arranged according to time and shot point locations to obtain the simulated active source seismic record.

[0053] Specifically, a simplified horizontal layered velocity model was established by referencing the depth, width, and number of strata in the existing South Yellow Sea model, flattening the interface. The velocities of each layer were assigned values ​​according to the corresponding strata characteristics. A Ricker wavelet with a dominant frequency of 8 Hz was used as the forward modeling source, with a time sampling interval of 2 ms. Geophones were placed on the simulated seabed interface at 6 km intervals to simulate the location of the seabed seismograph station, and shot points were placed along the survey line at 125 m shot intervals. Wave field propagation after source excitation and the response received by the seabed geophones were calculated shot-by-shot using the wave equation. These results were then arranged according to time and shot point location to form the active source seismic record from the seabed seismograph. Since no artificial superposition of ocean currents, instrument noise, or other environmental noise occurred during the entire forward modeling process, the resulting record was a clean seismic record. This record was subsequently added to a noise field extracted and synthesized from measured data to form a noisy record, with the original forward modeling record serving as the corresponding label.

[0054] The velocity model can be established by using a simplified horizontal layered model that flattens the interface after referencing the depth / width and number of strata in an existing regional model, or by assigning velocity parameters layer by layer according to the corresponding strata characteristics. The forward modeling source can use a Ricker wavelet with a dominant frequency of 8 Hz, or a Ricker wavelet with an adjusted dominant frequency to match the target frequency band. The observation system can be deployed by placing geophones at 6 km intervals along the simulated seabed interface and shot points at 125 m intervals along the survey line, or by adjusting the geophone intervals and shot spacing according to the target station location and shot point density.

[0055] This embodiment uses wave equation forward modeling to generate noise-free records as labels, ensuring supervised learning has a strictly noise-free objective and preventing inaccurate training targets caused by noisy labels. Flattening the stratigraphic interface yields a clearly structured horizontal layered model, making theoretical records easy to use as labels without being influenced by complex structures. Geophones are deployed along the simulated seafloor interface, and shot points are arranged at preset shot intervals. This ensures that the forward modeling observation system closely resembles the station and shot point distribution of the active source of the seafloor seismograph, making the synthesized training samples after adding noise geometrically closer to the actual data, thus improving the model's adaptability when transferred to field recording.

[0056] In some embodiments, such as Figure 3 As shown, in step S2, the selected noise blocks are subjected to overlapping weighted fusion to synthesize a noise field with continuous boundaries, including: S201. A two-dimensional Hanning window is used to weight each selected noise block. The two-dimensional Hanning window is the product of a one-dimensional Hanning window in the vertical direction of the time axis and a one-dimensional Hanning window in the horizontal direction of the space axis.

[0057] Among them, for location The expression for the two-dimensional Hanning window function of the point is: Among them, for the first The weighted result of the noise blocks is: In the formula, A one-dimensional Hanning window in the vertical direction of the time axis. A one-dimensional Hanning window in the horizontal direction of the spatial axis. This represents the side length of the window.

[0058] S202. The weighted noise blocks are weighted and averaged according to their overlapping areas to obtain a noise field with the same size as the pure seismic record.

[0059] Each noise block after being weighted by a two-dimensional Hanning window The noise field is laid out with overlapping blocks at a set step size across the entire recording scale. Overlapping areas are then weighted and averaged using accumulated window weights. The overlapping area is where adjacent noise blocks overlap during the laying process. Within the overlapping area, the weighted values ​​of each noise block are summed and then divided by the accumulated window weight at the corresponding location. The result is a noise field with a size completely consistent with the clean seismic record, exhibiting a smooth transition between adjacent areas. Because Hanning window weighting and weighted averaging of overlapping areas are used, there are no abrupt amplitude changes or obvious seams between adjacent noise blocks.

[0060] S203. The noise field obtained by fusion is then subjected to mean removal, standard deviation normalization, and random adjustment of noise intensity.

[0061] Mean reduction involves subtracting the average value of the entire noise field from each value in the noise field, making the mean zero. Standard deviation normalization involves dividing the mean-reduced noise field by its standard deviation, thus standardizing the dispersion of the amplitude distribution. Random adjustment of noise intensity involves randomly setting the standard deviation of the normalized noise field to a certain range, such as between 0.05 and 0.15, thereby controlling and randomizing the noise intensity superimposed on the clean record. Performing these three steps sequentially ensures that the overall amplitude level of the synthesized noise is both stable and random.

[0062] Specifically, multiple noise blocks with 256 sampling points in the time direction and 256 channels in the spatial direction are randomly selected from a real noise sample library. Each noise block is multiplied by a two-dimensional Hanning window obtained by multiplying a one-dimensional Hanning window in the horizontal direction and a one-dimensional Hanning window in the vertical direction, so that its edge weights are smoothly attenuated. Then, the weighted noise blocks are overlapped on the whole scale with a step size of 128 sampling points, with half overlap between adjacent blocks. For each position, the weighted values ​​of all blocks are added together and divided by the cumulative window weight at that position to fuse a noise field with the same size as the clean seismic record and continuous boundaries. Next, the mean of the noise field is removed to make the mean zero, and then it is normalized by dividing by the standard deviation. Finally, the standard deviation is randomly adjusted to between 0.05 and 0.15. The continuous noise field processed in the above way is then added to the whole clean record to obtain a noisy record with smooth boundaries and random intensity.

[0063] This embodiment employs a two-dimensional Hanning window to weight noise blocks and uses a weighted average based on overlapping regions. This ensures smooth transitions between randomly selected noise blocks across the entire recording scale, avoiding seams and abrupt amplitude changes left by direct splicing at boundaries. A continuous noise field is first synthesized across the entire recording scale, allowing adjacent regions to share the same noise background context, preventing abrupt changes in noise distribution at block boundaries. The noise field is then successively mean-reduced, standard deviation-normalized, and its intensity randomly adjusted to stabilize and add randomness to the amplitude of the synthesized noise. This results in noisy samples with diverse intensities and continuous boundaries after being added to the clean recording, providing high-quality training data for the network to learn realistic and smoothly transitioning noise distributions.

[0064] In some embodiments, such as Figure 4 As shown, in step S3, the training sample pairs are divided into blocks and transformed to the frequency-wavenumber domain to obtain the noisy spectrum and the clean spectrum in the frequency-wavenumber domain. The real and imaginary parts of the noisy spectrum are used as two channels to input the deep convolutional network, and the real and imaginary parts of the corresponding clean spectrum are used as supervision labels for training, including: S301. Divide the noisy seismic records and the clean seismic records into multiple data blocks of the same size.

[0065] The entire noisy seismic record and its corresponding clean seismic record are divided into one-to-one corresponding data blocks of the same fixed size. All data blocks have the same number of sampling points in both the temporal and spatial directions; for example, 256 sampling points in the temporal direction and 256 channels in the spatial direction. Using the same segmentation method for both noisy and clean records ensures a one-to-one correspondence between their data blocks, facilitating the pairing of inputs and labels. Segmenting into fixed-size data blocks allows for the processing of long-term sparsely distributed seismic phases into local segments, improving feature learning capabilities.

[0066] S302. For each data block, perform a two-dimensional Fourier transform without applying a window function or padding with zeros to obtain the corresponding frequency-wavenumber domain spectrum.

[0067] In the formula, For the first A noisy seismic record data block For the first A block of clean seismic records. For frequency coordinates, Wavenumber coordinates and These are the noisy spectrum in the frequency-wavenumber domain and the clean spectrum in the frequency-wavenumber domain, respectively.

[0068] This step performs a two-dimensional Fourier transform directly on each segmented data block, without multiplying by a window function or padding the data block with zeros before the transform. "No window function" means that no windowing weight is applied to the data block before the two-dimensional Fourier transform, preserving its original values; "No zero padding" means that zeros are not padded around the data block to increase its size. The two-dimensional Fourier transform maps the data block in the time-space domain to the frequency-wavenumber domain, where frequency and wavenumber are the coordinates. The resulting frequency-wavenumber spectrum is a complex number, containing both real and imaginary parts. Zero padding is only used to fill the remaining size to allow for proper slicing when processing the edges of the entire data frame and the remaining size is less than a window.

[0069] S303. The network input consists of the real and imaginary parts of the data blocks corresponding to noisy seismic records, and the supervision labels consist of the real and imaginary parts of the data blocks corresponding to clean seismic records, forming a supervised training dataset in the frequency-wavenumber domain.

[0070] Since the result of the two-dimensional Fourier transform is in complex form, the frequency-wavenumber domain spectrum described above can be written as: In the formula, The real part of the spectrum, For the imaginary part of the spectrum, It is the imaginary unit.

[0071] The real and imaginary parts are used as the two input channels of the neural network, respectively: The supervision label is: Supervised training dataset for the final frequency-wavenumber domain: Specifically, each noisy seismic record and its corresponding clean seismic record are divided into one-to-one data blocks of the same size, with 256 sampling points in the time direction and 256 channels in the spatial direction. For each data block, a two-dimensional Fourier transform is performed directly without windowing or zero-padding. The forward transform is performed without multiplying by the normalization coefficient, and the inverse transform is performed by multiplying by the reciprocal of the data block area, thus allowing the forward and inverse transforms to be reconciled. Zero-padding to the sliceable size is only performed when processing the edges of the entire data frame and the remaining size is less than a window. For the resulting complex spectrum, the real part is used as one channel and the imaginary part as another. The real and imaginary parts of the noisy data blocks are used as the network input, and the real and imaginary parts of the clean data blocks at the same location are used as the supervision labels. These are compiled block by block to form a supervised training dataset in the frequency-wavenumber domain, which the network uses to learn the mapping from the noisy spectrum to the clean spectrum.

[0072] In this embodiment, noisy and clean records are segmented one-to-one with the same size, ensuring a strict positional correspondence between the input and the label, providing accurate paired data for supervised learning. A two-dimensional Fourier transform is performed directly without windowing or zero-padding, avoiding the additional changes to the spectrum caused by windowing and zero-padding. This ensures the spectrum accurately reflects the data block content, and the forward and inverse transforms can be directly connected to restore each other. Using the real and imaginary parts as channels to form the input and label respectively preserves the amplitude and phase information completely, avoiding the phase loss caused by using only the amplitude spectrum. This allows the network to learn a mapping that recovers both amplitude and phase, thereby improving the fidelity of the waveform after the inverse transform.

[0073] In some embodiments, such as Figure 5 As shown, in step S3, before training with the real and imaginary parts of the noisy spectrum as two channels input to the deep convolutional network and the corresponding real and imaginary parts of the clean spectrum as supervision labels, the spectrum of the input deep convolutional network is normalized. Normalization is performed separately for each training sample, including: S310. For a noisy spectrum of a training sample, construct a two-channel tensor from its real and imaginary parts. And calculate the root mean square scale of the two-channel tensor according to the following formula. : Among them, root mean square scale The calculation covers the real part, the imaginary part, and all frequency-wavenumber domain spectral points, and the real part and the imaginary part are normalized using the same root mean square scale.

[0074] Root mean square scale The calculation simultaneously covers the real and imaginary channels, as well as all frequency-wavenumber domain spectral points, thus obtaining a uniform scale for the entire sample. The real and imaginary parts are normalized using the same root mean square scale to maintain their original relative proportions. Normalization is performed separately for each training sample.

[0075] S320, Divide the noisy spectrum by the root mean square scale. This is then used as input to a deep convolutional network, and the clean spectrum corresponding to the noisy spectrum is divided by the same root mean square scale. It was later used as a monitoring label.

[0076] Use the root mean square scale obtained in the previous step The noisy spectrum and its corresponding clean spectrum are scaled. The real and imaginary parts of the noisy spectrum are divided by the root mean square scale to obtain a normalized two-channel tensor, which serves as the network input. The real and imaginary parts of the clean spectrum corresponding to the same noisy spectrum are also divided by the same root mean square scale to obtain the scaled result, which serves as the supervision label. Since the input and labels use the exact same scale, their relative relationship remains unchanged before and after scaling. This process ensures that the input magnitudes of different samples are consistent, facilitating stable network training.

[0077] S330. When using a trained deep convolutional network to denoise the seismic data to be processed, calculate the root mean square scale based on the unprocessed spectrum of the seismic data. And multiply by the root mean square scale after the output of the deep convolutional network. .

[0078] The inference phase uses the same scaling method as the training phase. For each spectrum of the seismic data to be processed, the real and imaginary parts are constructed into a two-channel tensor, and its root mean square scale is calculated. The noisy spectrum is then divided by this scale and fed into the network. The network output is a denoised spectrum with normalized magnitude, which needs to be multiplied by the same root mean square scale to restore the magnitude to the same level as the original noisy data. Then, the restored real and imaginary parts are reconstructed into a complex spectrum, and a two-dimensional inverse Fourier transform is performed to restore it to the time-space domain. Because the real and imaginary parts are not scaled separately, their original relative proportions and phase relationships are preserved.

[0079] Specifically, for each noisy data block spectrum, the real and imaginary parts are superimposed into a two-channel tensor. Regarding this tensor Square all elements, take the mean, then take the square root to get the root mean square, and add the value to it. The root mean square (RMS) scale is obtained by taking the smallest positive number. The real and imaginary parts of the noisy spectrum are divided by this scale and used as network input. The real and imaginary parts of the clean spectrum at the same location are also divided by the same scale and used as supervision labels. Normalization is performed individually for each sample before being fed into the network for training. When denoising the data to be processed in the field, the RMS scale is calculated for each spectrum in the same way, divided by this scale, and then fed into the trained network. The denoised spectrum output by the network is multiplied back by the same scale to restore the magnitude. Then, the real and imaginary parts are reassembled into a complex spectrum, and a two-dimensional inverse Fourier transform is performed to recover the time-space domain signal. Throughout the entire process, the real and imaginary parts always share the same scale, and the phase relationship remains unchanged.

[0080] This embodiment calculates the root mean square scale separately for each sample and normalizes the input and label using the same scale, making the magnitudes of different samples more consistent and mitigating training instability caused by excessive amplitude differences. The real and imaginary parts share the same scale, maintaining their relative proportion and phase relationship, and avoiding the destruction of phase information by separate scaling. During the inference stage, the scale is calculated based on the spectrum to be processed and multiplied back after output, ensuring that the magnitude of the denoised result is consistent with the original noisy data. The waveform amplitude remains undistorted after the inverse transform, thus improving the fidelity of the denoising result while maintaining consistency between training and inference.

[0081] In some embodiments, such as Figure 6 As shown, the deep convolutional network in step S3 is a two-dimensional residual U-Net network with an encoder-decoder structure. The input and output of the two-dimensional residual U-Net network are both two channels, which correspond to the real part and imaginary part of the frequency-wavenumber domain spectrum, respectively.

[0082] The network employed adopts an encoder-decoder structure that first compresses and then restores the signal at each stage, incorporating residual connections. The encoder-decoder structure refers to a symmetrical structure where the first half progressively downsamples to extract features, and the second half progressively upsamples to restore resolution. The 2D Residual U-Net introduces shortcut branches in this symmetrical structure and uses 2D convolution to process the 2D spectrum. The two input channels represent the real and imaginary parts of the noisy frequency-wavenumber domain spectrum, respectively, while the two output channels represent the real and imaginary parts of the predicted denoised spectrum. Using two channels corresponding to the real and imaginary parts allows the network to simultaneously process and restore amplitude and phase information.

[0083] The encoder includes a first residual convolutional block, a first downsampling layer, a second residual convolutional block and a second downsampling layer connected in sequence, and a third residual convolutional block located at the bottom. The first residual convolutional block expands the number of channels from two to the first number of channels, and the second residual convolutional block expands the number of channels from the first number of channels to the second number of channels. The decoder includes a first upsampling, concatenating the first upsampling result with the feature map output by the second residual convolutional block in the encoder, a fourth residual convolutional block, a second upsampling, concatenating the second upsampling result with the feature map output by the first residual convolutional block in the encoder, a fifth residual convolutional block, and a convolutional layer that outputs two channels. Each residual convolutional block includes two convolutional layers, batch normalization, and an activation function. When the number of input channels is inconsistent with the number of output channels, the number of channels in the shortcut branch is adjusted through convolution. The downsampling type is average pooling, and the upsampling type is bilinear upsampling.

[0084] Specifically, the real and imaginary parts of the noisy frequency-wavenumber domain spectrum are fed into a two-dimensional residual U-Net as two channels. At the encoding end, the number of channels is first expanded from 2 to 32 via a first residual convolutional block, then passed through a first downsampling layer with 2x2 average pooling, followed by a second residual convolutional block expanding the number of channels from 32 to 64, then a second downsampling layer with 2x2 average pooling, reaching the bottom third residual convolutional block and maintaining 64 channels. At the decoding end, bilinear upsampling is performed to amplify the signal by a factor of 2, concatenating it with the 64 channels from the encoding end to obtain 128 channels, then compressed to 32 channels via a fourth residual convolutional block, then bilinear upsampling is performed again to a factor of 2, concatenating it with the previously retained 32 shallow features, then processed to 32 channels via a fifth residual convolutional block, and finally output as two channels (real and imaginary parts) using a 1x1 convolution. Each residual convolutional block contains two layers of 3x3 convolutions, batch normalization, and a linear unit with leakage correction and a negative slope of 0.05. A 1x1 convolution is used to adjust the number of shortcut branch channels when the number of input and output channels is inconsistent. The network learns the complete frequency-wavenumber plane, with the input directly taking the real and imaginary parts of the two-dimensional Fourier transform result.

[0085] In this embodiment, both the input and output have two channels, corresponding to the real and imaginary parts respectively, enabling the network to simultaneously recover amplitude and phase information. The encoder downsamples step by step and expands the number of channels from 2 to 32 and then to 64, extracting more abstract features while reducing the size. The decoder upsamples step by step and concatenates the features with the corresponding number of channels at the encoder, restoring resolution while fusing shallow details, which helps maintain the continuity of the in-phase axis. Each residual convolutional block alleviates the training difficulties of deep networks and promotes feature transfer through shortcut branches and channel adjustments, making it easier for the network to converge when learning the mapping from noisy spectrum to clean spectrum, thus better preserving the waveform structure after denoising.

[0086] In some embodiments, the loss function used to train the deep convolutional network is a weighted sum of the complex spectrum reconstruction error loss, the spectral energy error loss, and the spectral magnitude gradient error loss, expressed as: In the formula, The complex spectrum reconstruction error loss is characterized by the mean square error between the predicted spectrum and the pure spectrum output by the deep convolutional network in both the real and imaginary channels. The spectral energy error loss characterizes the error between the energy of the predicted spectrum and the energy of the noisy spectrum; The spectral amplitude gradient error loss characterizes the gradient difference between the amplitude of the predicted spectrum and the amplitude of the pure spectrum in the horizontal and vertical directions. The weighting coefficients for the spectral energy error loss are... is the weighting coefficient for the gradient error loss of the spectral amplitude.

[0087] Among them, the complex spectrum reconstruction error loss Represented as: In the formula, This represents the number of training samples in the batch. and The height and width of the frequency-wavenumber domain spectrum of the training samples; Indicates the real part of the spectrum channel; Indicates the imaginary part of the spectrum channel; Predict the spectrum for the network; The target is a pure spectrum.

[0088] Among them, spectral energy error loss Represented as: In the formula, For the network output spectrum at location The energy at the location, The energy of the noisy input spectrum; For the network output spectrum at location The real part of the place, For the network output spectrum at location The virtual part of the location; For the noisy input spectrum at position The real part of the place, For the noisy input spectrum at position The virtual part of the location.

[0089] Since the spectral energy error loss constrains the energy level of the network output based on the energy of the noisy spectrum, while the denoising is driven by the complex spectrum reconstruction error loss, the energy of weak signal spectral points will not be suppressed to zero, thus avoiding over-smoothing.

[0090] Among them, the spectral amplitude gradient error loss Represented as: The horizontal difference is defined as: The vertical difference is defined as: In the formula, The amplitude of the network output spectrum. The amplitude of the target spectrum.

[0091] The Adam optimizer is used during training. After each training round, the current network parameters are saved when the average training loss is lower than the previously recorded minimum. The network parameters with the lowest loss during training are used as the trained deep convolutional network.

[0092] The Adam optimizer is an optimization method that adaptively adjusts the update step size based on the first and second moments of the gradient. An initial learning rate can be set, for example, to 0.0001. After each training epoch, the average training loss for that epoch is calculated. If the loss is lower than the previously recorded minimum, the current network parameters are saved, and the recorded minimum value is updated. After all training epochs are completed, the set of parameters saved when the average training loss was lowest throughout the entire training process is used as the new network parameters.

[0093] During training, the neural network is represented as a mapping function. ,in This represents the network parameters. The goal of network training is to find the optimal parameters. This minimizes the weighted sum of the complex spectrum reconstruction error, spectral energy error, and spectral amplitude gradient error between the network output spectrum and the target pure spectrum, i.e.: Specifically, in each training batch, the network outputs a denoised spectrum from the input noisy spectrum. First, it calculates the mean square error between the predicted spectrum and the clean spectrum in both the real and imaginary channels as the complex spectrum reconstruction error loss. Then, it calculates the error between the energy of the sum of the squares of the real and imaginary parts at each position of the predicted spectrum and the corresponding energy of the noisy spectrum as the spectral energy error loss. Next, it calculates the difference between the amplitudes of the predicted spectrum and the clean spectrum at adjacent positions in the horizontal and vertical directions as the spectral amplitude gradient error loss. The spectral energy error loss is multiplied by a weighting factor of 0.1, and the spectral amplitude gradient error loss is multiplied by a weighting factor of 0.05, and then added to the complex spectrum reconstruction error loss to obtain the total loss. The Adam optimizer is used with an initial learning rate of 0.0001, iteratively updating the parameters with 16 samples per batch for a total of 30 training rounds. After each round, the average training loss is calculated, and if it falls below the previous minimum, the current parameters are saved. Finally, the parameters saved when the loss is lowest are used as the trained network.

[0094] This embodiment uses a weighted sum of complex spectrum reconstruction error loss, spectral energy error loss, and spectral amplitude gradient error loss as the loss function. This allows the network to approximate both the real and imaginary parts while considering energy distribution and amplitude spatial variations, thus more comprehensively constraining the predicted spectrum. The complex spectrum reconstruction error loss constrains the overall approximation of the real and imaginary parts, the spectral energy error loss constrains the energy level, and the spectral amplitude gradient error loss constrains the amplitude variation trend in the horizontal and vertical directions. These three factors complement each other, which is beneficial for the continuity of the phase axis and waveform fidelity. The Adam optimizer is used to adaptively adjust the step size and save the parameters with the lowest loss, making the training more stable and avoiding selecting intermediate states with high losses. As a result, the network has both amplitude recovery and structure preservation capabilities during denoising.

[0095] In some embodiments, such as Figure 7 As shown, in step S4, the seismic data to be processed is divided into blocks and transformed to the frequency-wavenumber domain. The blocks are then input into a trained deep convolutional network to obtain a denoised spectrum. An inverse frequency-wavenumber domain to spatiotemporal domain transformation is performed on the denoised spectrum, and the data blocks obtained from the inverse transformation are reconstructed by overlapping and weighting to obtain the denoised seismic record, including: S401. Bandpass filtering is applied to the seismic data to be processed, and the same bandpass filtering parameters are used for the seismic data to be processed, the measured active source seismic records used to extract pure noise, and the clean seismic records.

[0096] For seismic data requiring noise suppression in the field, bandpass filtering is first applied, ensuring that this data uses the same bandpass filtering parameters as the measured records used for database construction and the clean records obtained from forward modeling. All three datasets use the same set of corner frequencies, for example, 2Hz, 3Hz, 9Hz, and 10Hz. Using the same bandpass filtering parameters for the clean records used in training, the measured noise used in database construction, and the data to be processed in inference is to avoid inconsistencies between the frequency bands observed during model training and the frequency bands of the field inference data.

[0097] S402. Divide the bandpass-filtered seismic data to be processed into multiple data blocks according to the preset data block size and sliding step size. The preset data block size and sliding step size are consistent with those during training. Perform a two-dimensional Fourier transform on each data block to be processed to obtain the frequency-wavenumber domain spectrum to be processed. Input the real and imaginary parts of the spectrum to be processed into the trained deep convolutional network to obtain the corresponding denoised spectrum.

[0098] The frequency-wavenumber domain spectrum to be processed is represented as follows: Extract the real and imaginary parts of the spectrum to be processed: Inputting it into the trained deep convolutional network yields the corresponding denoised spectrum: S403. Perform a two-dimensional inverse Fourier transform on the denoised spectrum to obtain the corresponding denoised data block. Use a two-dimensional Hanning window to perform a weighted average of each denoised data block according to the overlapping area, and reconstruct the denoised seismic record.

[0099] Perform a two-dimensional inverse Fourier transform on the predicted denoised spectrum: The denoised seismic record is reconstructed by fusing the data using the Hanning window weighted average method, and the result is as follows: in: Indicates the first One denoised data block; Σ represents the weight of the corresponding Hanning window; Σ represents the summation of all overlapping data blocks.

[0100] Specifically, the seismic data to be processed in the field is first bandpass filtered using the same parameters as the actual measured records and the clean forward modeling records, with the four corner frequencies being 2Hz, 3Hz, 9Hz, and 10Hz respectively. Then, using the same block size as during training (256 sampling points in the time direction, 256 channels in the spatial direction, and 128 sampling points), the filtered data is divided into multiple data blocks to be processed. A two-dimensional Fourier transform is performed on each block to obtain the spectrum to be processed. The real and imaginary parts are used as two channels and fed into the trained network to obtain the corresponding denoised spectrum. Next, a two-dimensional inverse Fourier transform is performed on each denoised spectrum to recover the denoised data block. A two-dimensional Hanning window (obtained by multiplying one-dimensional Hanning windows in the horizontal and vertical directions) is used to weight each denoised data block. The weighted results are summed according to overlapping areas and divided by the cumulative weight to reconstruct the complete denoised seismic record, thus suppressing the complex background noise of the actual acquisition environment while preserving the effective seismic phase structure.

[0101] like Figure 8 The figure shown is a comparison of the results of this method and the noise suppression method at station C10 in the OBS2016 survey line in the South Yellow Sea, as illustrated in an embodiment of the present invention. Figure 8 The upper middle section shows the processing result of the method of this invention, and the lower section shows the processing result of noise combination suppression. This method showed an effective seismic phase at 150km to the right of the C10 station and an effective seismic phase at 100km to the left. Compared with the noise combination suppression processing method, the distance of the identifiable seismic phase has increased by 50%.

[0102] like Figure 9The figure shown is a partial comparison of the results of this method and the noise combined suppression method at station C10 in the OBS2016 survey line in the South Yellow Sea, as illustrated in the embodiment of the present invention. Compared with the noise combined suppression method, the strata reflection phase axis in the circle is horizontally continuous and the bending shape is complete and traceable. After processing by the noise combined suppression method, the continuity of the small shot distance detection signal is not good.

[0103] like Figure 10 The figure shown is a comparison of the results of this method and the noise suppression method at station C12 in the OBS2016 survey line in the South Yellow Sea, as illustrated in this embodiment of the invention. This method revealed an effective seismic phase 151 km to the right of station C12, and compared to the noise suppression method, the distance of the identifiable seismic phase increased by 200%.

[0104] In this embodiment, the measured records and clean records used for database construction employ the same bandpass filtering parameters to ensure consistent data frequency bands between training and inference, avoiding a decrease in denoising performance due to frequency band mismatch. The data is sliced ​​and transformed using the same block size and step size as the training data, unifying the data organization between inference and training, ensuring the network faces the same data structure during inference. Each denoised data block is reconstructed using a two-dimensional Hanning window with weighted averages based on overlapping areas. This avoids abrupt amplitude changes, seams, and discontinuities in phase axes that would result from directly splicing together blocks after independent denoising at boundaries. This reconstructs denoised records with smooth boundaries and complete structure, achieving high-fidelity recovery of seismic data under complex noise conditions.

[0105] The following are system embodiments of the present invention, which can be used to execute the active-source seafloor seismometer data denoising method involved in the present invention. For details not disclosed in the system embodiments of the present invention, please refer to the method embodiments of the active-source seafloor seismometer data denoising method involved in the present invention.

[0106] like Figure 11 As shown in the figure, an active source seafloor seismograph data denoising system 50 provided in this embodiment of the invention includes: The noise library construction module 501 is used to extract pure noise from measured active source seismic records and construct a real noise sample library. The sample synthesis module 502 is used to select noise blocks from the real noise sample library, perform overlapping and weighted fusion on the selected noise blocks, synthesize a noise field with continuous boundaries, superimpose the noise field with the clean seismic record to obtain the noisy seismic record, and use the noisy seismic record and its corresponding clean seismic record to form a training sample pair. The network training module 503 is used to divide the training sample pairs into blocks and transform them to the frequency-wavenumber domain to obtain the noisy spectrum and the clean spectrum in the frequency-wavenumber domain. The real and imaginary parts of the noisy spectrum are used as two channels to input the deep convolutional network, and the real and imaginary parts of the corresponding clean spectrum are used as supervision labels for training to obtain the mapping relationship between the noisy spectrum and the clean spectrum. The denoising and reconstruction module 504 is used to divide the seismic data to be processed into blocks and transform it to the frequency-wavenumber domain. The denoised spectrum is then input into the trained deep convolutional network to obtain the denoised spectrum. The denoised spectrum is then subjected to an inverse transformation from the frequency-wavenumber domain to the spatiotemporal domain. Finally, the data blocks obtained by the inverse transformation are overlapped and weighted to reconstruct the denoised seismic record.

[0107] It should be noted that the active source seafloor seismograph data denoising system provided in the above embodiments is only illustrated by the division of the above functional modules when performing data denoising. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the active source seafloor seismograph data denoising system will be divided into different functional modules to complete all or part of the functions described above.

[0108] Furthermore, the active source seafloor seismometer data denoising system and the active source seafloor seismometer data denoising method provided in the above embodiments belong to the same concept. The specific way each module performs its operation has been described in detail in the method embodiments, and will not be repeated here.

[0109] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for denoising active-source seafloor seismograph data, characterized in that, The method includes: Pure noise was extracted from measured active source seismic records to construct a real noise sample library; Noise blocks are selected from the real noise sample library, and the selected noise blocks are overlapped and weighted to fuse them to synthesize a noise field with continuous boundaries. The noise field is superimposed with the clean seismic record to obtain a noisy seismic record, and the noisy seismic record and its corresponding clean seismic record constitute a training sample pair. The training samples are divided into blocks and transformed to the frequency-wavenumber domain to obtain the noisy spectrum and the clean spectrum in the frequency-wavenumber domain. The real and imaginary parts of the noisy spectrum are used as two channels to input a deep convolutional network, and the real and imaginary parts of the corresponding clean spectrum are used as supervision labels for training to obtain the mapping relationship between the noisy spectrum and the clean spectrum. The seismic data to be processed is divided into blocks and transformed to the frequency-wavenumber domain. The blocks are then input into the trained deep convolutional network to obtain a denoised spectrum. The denoised spectrum is then subjected to an inverse transformation from the frequency-wavenumber domain to the spatiotemporal domain. Finally, the data blocks obtained by the inverse transformation are overlapped and weighted to reconstruct the denoised seismic record.

2. The method for denoising active-source seafloor seismograph data according to claim 1, characterized in that, The extraction of pure noise from measured active-source seismic records and the construction of a real noise sample library include: From the bandpass filtered measured active source seismic records, regions that do not contain effective reflected waves, refracted waves, diffracted waves and surface waves are selected as pure noise regions, and multiple pure noise samples are obtained by cropping from the pure noise regions. A random amplitude perturbation is applied to each of the pure noise samples, wherein the random amplitude perturbation is the pure noise sample multiplied by an amplitude scaling factor randomly selected within a preset value range; A sliding window is used to slice the pure noise sample after the random amplitude perturbation to obtain multiple noise blocks, and all the noise blocks constitute the real noise sample library.

3. The method for denoising active-source seafloor seismograph data according to claim 1, characterized in that, The pristine seismic records were obtained in the following manner: Based on the horizontal layered velocity model, a noise-free active source seismic record is generated by forward modeling using the wave equation, which serves as the pure seismic record. The forward modeling used a Ricker wavelet as the source, and the receivers were set up along the simulated seabed interface to simulate the location of the seabed seismograph. The shot points were set up along the survey line according to the preset shot spacing.

4. The method for denoising active-source seafloor seismograph data according to claim 1, characterized in that, The step of performing overlapping weighted fusion on the selected noise blocks to synthesize a noise field with continuous boundaries includes: Each selected noise block is weighted using a two-dimensional Hanning window, which is the product of a one-dimensional Hanning window in the vertical direction of the time axis and a one-dimensional Hanning window in the horizontal direction of the spatial axis. The weighted noise blocks are weighted and averaged according to their overlapping areas, and then fused to obtain a noise field with the same size as the pure seismic record. The noise field obtained by fusion is then subjected to mean removal, standard deviation normalization, and random adjustment of noise intensity.

5. The method for denoising active-source seafloor seismograph data according to claim 1, characterized in that, The process involves dividing the training samples into blocks and transforming them to the frequency-wavenumber domain to obtain noisy and clean frequencies-wavenumber spectra. The real and imaginary parts of the noisy spectra are then used as two channels to input a deep convolutional network. Training is performed using the corresponding real and imaginary parts of the clean frequencies as supervision labels. This includes: The noisy seismic records and the clean seismic records are each divided into multiple data blocks of the same size; For each data block, a two-dimensional Fourier transform is performed without applying a window function or padding with zeros to obtain the corresponding frequency-wavenumber domain spectrum. The network input is formed by the real and imaginary parts of the data blocks corresponding to the noisy seismic records, and the supervision labels are formed by the real and imaginary parts of the data blocks corresponding to the clean seismic records, thus forming a supervised training dataset in the frequency-wavenumber domain.

6. The method for denoising active-source seafloor seismograph data according to claim 1, characterized in that, Before transforming the training sample pairs into the frequency-wavenumber domain to obtain noisy and clean frequencies-wavenumber spectra, and using the real and imaginary parts of the noisy spectra as two channels to input a deep convolutional network, and using the real and imaginary parts of the clean spectra as supervision labels for training, the process further includes: The spectrum of the input deep convolutional network is normalized, and the normalization is performed separately for each training sample, including: For a noisy spectrum of a training sample, construct a two-channel tensor from its real and imaginary parts. The root mean square scale of the dual-channel tensor is calculated using the following formula. : Wherein, the root mean square scale The calculation covers the real part, the imaginary part, and all frequency-wavenumber domain spectral points, and the real part and the imaginary part are normalized using the same root mean square scale; Divide the noisy spectrum by the root mean square scale. This is then used as input to the deep convolutional network, and the clean spectrum corresponding to the noisy spectrum is divided by the same root mean square scale. Later used as a supervisory label; When denoising the seismic data to be processed using the trained deep convolutional network, the root mean square scale is calculated according to the spectrum obtained by the block transformation of the seismic data to be processed. And after the output of the deep convolutional network, it is multiplied by the root mean square scale. .

7. The method for denoising active-source seafloor seismograph data according to claim 1, characterized in that, The deep convolutional network is a two-dimensional residual U-Net network, which includes an encoder and a decoder. The input and output of the two-dimensional residual U-Net network are both two channels, which correspond to the real part and imaginary part of the frequency-wavenumber domain spectrum, respectively. The encoder includes a first residual convolutional block, a first downsampling layer, a second residual convolutional block and a second downsampling layer connected in sequence, and a third residual convolutional block located at the bottom. The first residual convolutional block expands the number of channels from two to a first number of channels, and the second residual convolutional block expands the number of channels from the first number of channels to a second number of channels. The decoder includes, in sequence, a first upsampling, concatenating the first upsampling result with the feature map output by the second residual convolutional block in the encoder, a fourth residual convolutional block, a second upsampling, concatenating the second upsampling result with the feature map output by the first residual convolutional block in the encoder, a fifth residual convolutional block, and a convolutional layer that outputs the two channels. Each residual convolutional block includes two convolutional layers, batch normalization, and an activation function. When the number of input channels is inconsistent with the number of output channels, the number of channels in the shortcut branch is adjusted through convolution. The downsampling type is average pooling, and the upsampling type is bilinear upsampling.

8. The method for denoising active-source seafloor seismograph data according to claim 1, characterized in that, The loss function used to train the deep convolutional network is a weighted sum of the complex spectrum reconstruction error loss, the spectral energy error loss, and the spectral magnitude gradient error loss, expressed as: In the formula, The complex spectrum reconstruction error loss is characterized by the mean square error between the predicted spectrum and the pure spectrum output by the deep convolutional network in both the real and imaginary channels. The spectral energy error loss characterizes the error between the energy of the predicted spectrum and the energy of the noisy spectrum; The spectral amplitude gradient error loss characterizes the gradient difference between the amplitude of the predicted spectrum and the amplitude of the pure spectrum in the horizontal and vertical directions. The weighting coefficients for the spectral energy error loss are... These are the weighting coefficients for the spectral amplitude gradient error loss; The Adam optimizer is used during training, and after each training round, the current network parameters are saved when the average training loss is lower than the previously recorded minimum. The network parameters with the lowest loss during training are used as the trained deep convolutional network.

9. The method for denoising active-source seafloor seismograph data according to claim 1, characterized in that, The process involves dividing the seismic data to be processed into blocks and transforming it to the frequency-wavenumber domain. The resulting blocks are then input into the trained deep convolutional network to obtain a denoised spectrum. An inverse frequency-wavenumber domain to spatiotemporal domain transformation is performed on the denoised spectrum, and the resulting data blocks are then overlapped and weighted to reconstruct the denoised seismic record, including: The seismic data to be processed is bandpass filtered, and the seismic data to be processed, the measured active source seismic records used to extract pure noise, and the pure seismic records use the same bandpass filtering parameters. The bandpass filtered seismic data to be processed is divided into multiple data blocks according to the preset data block size and sliding step size. The preset data block size and sliding step size are the same as those during training. A two-dimensional Fourier transform is performed on each data block to be processed to obtain the frequency-wavenumber domain spectrum to be processed. The real part and imaginary part of the spectrum to be processed are input into the trained deep convolutional network to obtain the corresponding denoised spectrum. The denoised spectrum is subjected to a two-dimensional inverse Fourier transform to obtain the corresponding denoised data block. The denoised data block is then weighted and averaged according to the overlapping area using a two-dimensional Hanning window to reconstruct the denoised seismic record.

10. A data denoising system for an active-source seafloor seismograph, characterized in that, The system includes: The noise library construction module is used to extract pure noise from measured active source seismic records and construct a real noise sample library. The sample synthesis module is used to select noise blocks from the real noise sample library, perform overlapping and weighted fusion on the selected noise blocks, synthesize a noise field with continuous boundaries, superimpose the noise field with the clean seismic record to obtain a noisy seismic record, and use the noisy seismic record and its corresponding clean seismic record to form a training sample pair. The network training module is used to divide the training samples into blocks and transform them to the frequency-wavenumber domain to obtain the noisy spectrum and the clean spectrum in the frequency-wavenumber domain. The real and imaginary parts of the noisy spectrum are used as two channels to input a deep convolutional network, and the real and imaginary parts of the corresponding clean spectrum are used as supervision labels for training to obtain the mapping relationship between the noisy spectrum and the clean spectrum. The denoising and reconstruction module is used to divide the seismic data to be processed into blocks and transform it to the frequency-wavenumber domain. The denoised spectrum is then input into the trained deep convolutional network to obtain the denoised spectrum. The denoised spectrum is then subjected to an inverse transformation from the frequency-wavenumber domain to the spatiotemporal domain. Finally, the data blocks obtained by the inverse transformation are overlapped and weighted to reconstruct the denoised seismic record.

Images