A Deep Learning-Based Method and System for Suppressing Noise in Seismic Data

Abstract

This application relates to the field of seismic data analysis, and in particular to a method and system for seismic data noise suppression based on deep learning. The method includes: acquiring raw seafloor seismic data of a target area, and marine environmental parameters that are spatiotemporally correlated with the raw seafloor seismic data; mapping the marine environmental parameters to generate a spatiotemporally continuous physical guidance noise field, the generation process of which is constrained by physical consistency, and the physical guidance noise field possessing the low-frequency, coherent, and smooth physical properties of flow-induced noise; using the physical guidance noise field as physical guidance information, adaptively separating effective seismic signals from the raw seafloor seismic data, reconstructing and outputting the noise-suppressed seafloor seismic dataset. This application, through the combination of physical reasoning and deep learning, enables the model to distinguish between signals resembling noise and noise resembling signals, improving the high fidelity and generalization ability of seismic data denoising.

Description

Technical Field

[0001] This application relates to the field of seismic data analysis, and in particular to a method and system for suppressing seismic data noise based on deep learning. Background Technology

[0002] Traditional seismic data analysis techniques, when dealing with submarine seismic data, especially data acquired from seabed nodes, are often subject to strong coherent noise interference related to seawater movement, such as persistent low-frequency noise generated by the interaction of ocean currents and seabed topography. Traditional filtering methods rely on the simple assumption of separation between signal and noise in the transform domain, making it difficult to handle complex flow-induced noise that is non-stationary, spatially variable, and overlaps with the frequency bands of weak geological signals, and easily damaging the effective signal.

[0003] Existing deep learning-based seismic denoising methods mostly employ end-to-end supervised learning, and their effectiveness heavily relies on the "clean signal" label. These methods typically treat noise as a statistical phenomenon and remove it "by memory," lacking an understanding of the physical sources of noise. When faced with novel disturbances not present in the training data, such as laminar flows under specific ocean dynamic processes like specific thermohaline structures, their generalization ability drops sharply. Summary of the Invention

[0004] This application provides a method and system for suppressing seismic data noise based on deep learning to solve the above-mentioned problems.

[0005] In a first aspect, this application provides a deep learning-based method for suppressing seismic data noise. The method includes: acquiring raw submarine seismic data of a target area and marine environmental parameters that are spatiotemporally correlated with the raw submarine seismic data; mapping the marine environmental parameters to generate a spatiotemporally continuous physical guidance noise field, wherein the generation process of the physical guidance noise field is constrained by physical consistency, and the physical guidance noise field possesses the physical properties of low frequency, coherence, and smoothness of flow-induced noise; using the physical guidance noise field as physical guidance information, adaptively separating effective seismic signals from the raw submarine seismic data, reconstructing and outputting a noise-suppressed submarine seismic dataset.

[0006] The above technical solution, through a physical guidance mechanism, achieves high-fidelity noise suppression, significantly outperforming traditional methods and general deep learning models. First, the physically guided noise field, forced by physical consistency constraints, possesses the physical properties of flow-induced noise, enabling the signal extraction network to accurately distinguish between image noise signals and image signal noise. This significantly improves the protection of subtle geological features such as faults and pinch-outs while suppressing strong coherent noise. Second, this method exhibits strong generalization and adaptability. Even under ocean current patterns not present in the training data, the noise field generator can generate reasonable noise estimates based on the input ocean environmental parameters, achieving physical inference rather than pattern memorization, thus avoiding the sharp decline in generalization ability of traditional deep learning models. Furthermore, the generated physically guided noise field, as an intermediate product, provides interpretability and can be used for noise source analysis. Finally, by employing weakly supervised labels, this method eliminates the dependence on absolutely pure signals, greatly enhancing its practicality and improving the efficiency and reliability of data processing.

[0007] Optionally, the step of mapping the marine environmental parameters to generate a spatiotemporally continuous physical guidance noise field includes: inputting the marine environmental parameters into an environmental encoder to encode and generate a high-dimensional environmental context feature vector characterizing the physical state of the current acquisition environment; constructing a spatiotemporal coordinate tensor corresponding to the dimension of the original seafloor seismic data, wherein each element in the spatiotemporal coordinate tensor is associated with the temporal and spatial trace index information of the corresponding sampling point in the marine environmental parameters; inputting the environmental context feature vector and the spatiotemporal coordinate tensor together into a noise field generator, wherein the noise field generator outputs a spatiotemporally continuous preliminary noise field through its internal differentiable neural network mapping; during the training process of the noise field generator, constraining the preliminary noise field through a physical consistency loss function to force its energy distribution and spatiotemporal variation pattern to conform to the low-frequency, coherent, and smooth physical priors of flow-induced noise, thereby generating the physical guidance noise field.

[0008] By employing the aforementioned technical solution, and using marine environmental parameters and spatiotemporal coordinates as input, combined with physical consistency constraints, the interpretability and generalization ability of the noise field generator are enhanced. First, the model no longer relies solely on statistical learning; instead, physical prior constraints ensure the physical rationality of the generated noise field. This allows the model to reason based on physical principles when faced with changes in the acquisition environment, thus exhibiting strong generalization ability. Second, the generated noise field is spatiotemporally continuous, accurately matching every sampling point of the original seismic data and providing high-precision noise estimation. Finally, the environmental encoder abstracts complex marine environmental parameters into high-dimensional feature vectors, effectively integrating physical information into the deep learning framework and providing precise physical guidance information for subsequent signal separation.

[0009] Optionally, the physical consistency loss function includes at least one of the following constraints: low-spectrum constraint, spatiotemporal smoothness constraint, and apparent velocity range constraint; the low-spectrum constraint is to penalize the high-frequency energy of the initial noise field after frequency domain transformation; the spatiotemporal smoothness constraint is to calculate the gradient of the initial noise field in the time and space dimensions and penalize its large changes; the apparent velocity range constraint is to transform the initial noise field to the τ-p domain and encourage its energy to concentrate in the low-speed range corresponding to ocean current interference.

[0010] By introducing three physical consistency constraints through the above technical solution, the generated physical guidance noise field possesses extremely high physical rationality, thereby improving the denoising performance of the subsequent signal extraction network. The low-spectrum constraint prevents valid signals from being misclassified as noise. The spatiotemporal smoothness constraint ensures the coherence of the noise field, which is beneficial for the model to accurately capture the wavefield characteristics of flow-induced noise. The apparent velocity range constraint precisely limits the noise field in terms of propagation characteristics, ensuring its correspondence with actual ocean current interference. This multi-dimensional physical constraint improves the model's generalization ability, enabling it to generate accurate noise priors even in different ocean environments.

[0011] Optionally, the step of adaptively separating effective seismic signals from the original submarine seismic data using the physical guidance noise field as physical guidance information, and reconstructing and outputting a noise-suppressed submarine seismic dataset includes: concatenating the physical guidance noise field with the original submarine seismic data along the channel dimension to form combined data; inputting the combined data into a signal extraction network, which explicitly distinguishes and processes information from the noise field channel and the data channel through its initial attention layer, and uses the physical guidance noise field as a spatial attention guidance condition to focus the signal extraction network on data regions with a morphological similarity greater than a preset similarity to the noise field, thus completing the noise prior; the signal extraction network, guided by the noise prior, learns to decouple and separate components related to the physical guidance noise field from the original submarine seismic data through hierarchical Transformer blocks, while retaining and reconstructing effective signal components related to geological reflections to obtain a preliminary denoised seismic signal; and analyzing the preliminary denoised seismic signal based on the original submarine seismic data to construct a submarine seismic dataset.

[0012] By explicitly introducing the physically guided noise field as a noise prior into the signal extraction network through the above technical solution, adaptive and high-fidelity noise separation is achieved. The design of the initial attention layer enables the model to accurately focus on the noise region, avoiding overprocessing of the effective signal region and thus reducing the risk of damage to the effective signal. The powerful long-range dependency modeling capability of the hierarchical Transformer block ensures that the continuity and integrity of the geological reflection phase axis are preserved while removing coherent noise. This physically guided attention mechanism improves the model's denoising performance and generalization ability in complex marine environments.

[0013] Optionally, the signal extraction network is a deep learning model with a visual Transformer structure as its backbone; the signal extraction network processes the input of the combined data through a self-attention mechanism and a cross-attention mechanism, wherein the physical guidance noise field is used as the key and value of the cross-attention mechanism to guide the model to pay attention to the context in the data that is related to the noise prior.

[0014] The above technical solution employs a cross-attention mechanism to enhance the noise suppression capability of the signal extraction network. Using the physically guided noise field as the key and value in the cross-attention mechanism provides an efficient and interpretable way to inject physical prior information into the deep learning model. This enables the model to adaptively decouple features based on the physical form of the noise, improving denoising accuracy and protecting weak signals. This structure avoids the confusion between noise and signal features in traditional methods, ensuring high-fidelity reconstruction of the effective signal.

[0015] Optionally, the step of analyzing the pre-denoised seismic signal based on the original submarine seismic data to construct a submarine seismic dataset includes: calculating the initial residual between the original submarine seismic data and the pre-denoised seismic signal; inputting the initial residual, the pre-denoised seismic signal, and the physical guidance noise field into a residual optimization model; the residual optimization model is a lightweight convolutional neural network used to analyze the residual noise, over-removed effective signal components, and noise components that the physical guidance noise field fails to fully model in the initial residual, and outputting a fine residual correction; adding the residual correction to the pre-denoised seismic signal to obtain the noise-suppressed submarine seismic dataset.

[0016] By employing the aforementioned technical solutions, a residual optimization model is introduced to achieve secondary refinement of the denoising results, significantly improving the signal-to-noise ratio and fidelity of the final output. The residual learning mechanism focuses on addressing subtle errors in the signal extraction network, effectively recovering over-removed effective signal components, and further suppressing residual noise, especially complex noise that the physically guided noise field has not been fully modeled. The lightweight CNN design ensures the efficiency of the correction process, avoiding the introduction of excessive computational burden.

[0017] Optionally, the training of the residual optimization model is performed jointly with the training of the signal extraction network, and the total loss function of the joint training includes a signal reconstruction loss applied to the final output and a sparsity constraint loss applied to the final residual; the signal reconstruction loss is used to measure the difference between the final output and the weakly supervised label; the sparsity constraint loss is used to encourage complete noise separation and complete preservation of effective signals by applying an L1 norm penalty to the final residual between the original seafloor seismic data and the final output, so that the final residual does not contain structural information.

[0018] Through the above technical solutions, the joint training mechanism ensures consistency in the optimization objectives of the signal extraction and residual optimization stages, achieving end-to-end optimality. The introduction of sparsity constraint loss is crucial for ensuring high-fidelity noise reduction. By forcing the final residual to contain only noise, it prevents effective signal components from being incorrectly classified as noise and removed. This constraint mechanism improves the model's ability to protect weak geological signals, particularly in the reconstruction of detailed features such as faults and pinch-outs.

[0019] Optionally, the joint training process of the signal extraction network and the residual optimization model includes: preparing a training dataset, which contains multiple sets of samples, each set of samples including: original submarine seismic data segments, spatiotemporally correlated marine environmental parameter vectors, and low-disturbance seismic signal data as the weak supervision label; using the training dataset as the objective of the total loss function, performing end-to-end joint training of the signal extraction network and the residual optimization model, so that the signal extraction network and the residual optimization model learn the differences between the original seismic data and the clean seismic signal data under different marine environmental influences, thereby enabling the signal extraction network and the residual optimization model to accurately separate and reconstruct effective seismic signals under the guidance of a noise field.

[0020] Through the aforementioned technical solutions, the end-to-end joint training strategy ensures that all components of the model can be collaboratively optimized to achieve global optimum. By utilizing spatiotemporally correlated marine environmental parameter vectors, the model learns the mapping relationship between noise and the physical environment, thus generating a reasonable physical guidance noise field even when facing environmental conditions not seen in the training set, demonstrating strong generalization ability. The weakly supervised learning strategy significantly reduces the dependence on expensive and difficult-to-obtain "clean signal" labels, greatly improving the practicality of this method.

[0021] Optionally, the low-disturbance seismic signal data is seismic data obtained by applying long-offset stacking processing to the original seafloor seismic data; the signal-to-noise ratio of the low-disturbance seismic signal data is higher than that of the original seafloor seismic data.

[0022] The above-described technical solution, employing long-offset stacking to generate weakly supervised labels, offers significant practical and technical advantages. It utilizes conventional seismic data processing workflows, is cost-effective and easy to implement, and avoids reliance on expensive "clean" labels. Long-offset stacking leverages the velocity differences between signal and flow-induced noise, ensuring high fidelity in the generated labels regarding the effective signal, which is crucial for training the signal extraction network to recover geological reflections. Finally, the quality of these weakly supervised labels is sufficient to drive the joint training process; combined with physical guidance and sparsity constraints, high-fidelity noise suppression is ultimately achieved.

[0023] Secondly, this application provides a seismic data noise suppression system based on deep learning, the system comprising: The data acquisition module is used to acquire raw submarine seismic data of the target area, as well as marine environmental parameters that are spatiotemporally correlated with the raw submarine seismic data; The data mapping module is used to map the marine environmental parameters to generate a spatiotemporally continuous physical guidance noise field. The generation process of the physical guidance noise field is subject to physical consistency constraints, and the physical guidance noise field has the physical properties of low frequency, coherence, and smoothness of flow-induced noise. The data reconstruction module is used to adaptively separate effective seismic signals from the original submarine seismic data using the physical guidance noise field as physical guidance information, and to reconstruct and output the noise-suppressed submarine seismic dataset. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of an application scenario provided in an embodiment of this application.

[0026] Figure 2 This is a flowchart of a deep learning-based seismic data noise suppression method provided in one embodiment of this application.

[0027] Figure 3 This is a schematic diagram of a seismic data noise suppression system based on deep learning, provided as an embodiment of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0029] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0030] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0031] Existing deep learning-based seismic denoising methods mostly employ end-to-end supervised learning, and their effectiveness heavily relies on the "clean signal" label. These methods typically treat noise as a statistical phenomenon and remove it "by memory," lacking an understanding of the physical sources of noise. When faced with novel disturbances not present in the training data, generated by specific ocean dynamic processes (such as laminar flow under specific thermo-salinity structures), their generalization ability drops sharply.

[0032] Based on this, this application provides a deep learning-based method and system for noise suppression of seismic data. Through a physical guidance mechanism, it achieves high-fidelity noise suppression, significantly outperforming traditional methods and general deep learning models. The physically guided noise field, generated by physical consistency constraints, possesses the physical properties of flow-induced noise, enabling the signal extraction network to accurately distinguish between "noise-like signals" and "signal-like noise." This significantly improves the protection of weak geological features such as faults and pinch-outs while suppressing strong coherent noise. Secondly, this method exhibits strong generalization and adaptability. Even under ocean current patterns not present in the training data, the noise field generator can generate reasonable noise estimates based on the input ocean environmental parameters, achieving "physical inference" rather than "pattern memory," thus avoiding the problem of a sharp decline in generalization ability in traditional deep learning models. Furthermore, the generated physically guided noise field, as an intermediate product, provides interpretability and can be used for noise source analysis. Finally, by employing weakly supervised labels, this method eliminates the dependence on elusive "absolutely pure" signals, greatly enhancing its practicality and improving the efficiency and reliability of data processing.

[0033] Figure 1 This is a schematic diagram of an application scenario provided by this application. When analyzing submarine seismic data, the method provided in this application is applied to achieve high-fidelity noise suppression.

[0034] Specifically, the method provided in this application can be applied to any server. The server interacts with sensor arrays and environmental sensors to obtain raw seafloor seismic data provided by the sensor arrays and marine environmental parameters provided by the environmental sensors. Based on real-time signal processing and physics-guided deep learning theory, a noise field generation module with physical constraints is constructed by introducing marine environmental parameters as prior information. This solves the limitations of traditional methods and general deep learning methods in dealing with complex flow-induced noise. Through the combination of "physical reasoning + deep learning", the model can distinguish between "signals that look like noise" and "noise that looks like signals", significantly improving the high fidelity and generalization ability of noise reduction. This method enables monitoring personnel to obtain seafloor seismic datasets with high reference value.

[0035] For specific implementation details, please refer to the following examples.

[0036] Figure 2 This application provides a flowchart of a deep learning-based seismic data noise suppression method according to an embodiment of the present application. The method of this embodiment can be applied to servers in the above scenarios, such as... Figure 2 As shown, the method includes: S201. Acquire the raw submarine seismic data of the target area, as well as marine environmental parameters that are spatially and temporally correlated with the raw submarine seismic data; S202. Map marine environmental parameters to generate a spatiotemporally continuous physical guidance noise field. The generation process of the physical guidance noise field is constrained by physical consistency. The physical guidance noise field has the physical properties of low frequency, coherence and smoothness of flow-induced noise. S203. Using the physical guidance noise field as the physical guidance information, the effective seismic signal is adaptively separated from the original submarine seismic data, and the submarine seismic dataset after noise suppression is reconstructed and output.

[0037] This invention, based on real-time signal processing and physically guided deep learning theory, constructs a physically constrained noise field generation module by introducing marine environmental parameters as prior information, thus overcoming the limitations of traditional methods and general deep learning methods in handling complex flow-induced noise. The core principle of this method is that the physical properties of flow-induced noise, such as low frequency, coherence, and smoothness, are strongly correlated with marine environmental parameters such as flow velocity, flow direction, temperature, and salinity. The marine environmental parameters obtained in step S201 are used to characterize the physical origin of the noise. Step S202 uses a physically consistent noise field generator to map the environmental parameters into a spatiotemporally continuous physically guided noise field. This noise field possesses the physical prior properties of flow-induced noise, and can generate reasonable noise estimates based on physical inference even in weakly labeled or unlabeled cases. In step S203, the signal extraction network uses this physically guided noise field as explicit noise prior information, guiding the model to adaptively focus on regions in the original data that are similar in morphology to the noise field, thereby achieving high-fidelity separation and reconstruction of the effective signal. This combination of physical reasoning and deep learning enables the model to distinguish between noise-like signals and noise-like signals, significantly improving the high fidelity and generalization ability of noise reduction.

[0038] This method includes three main steps, S201, S202, and S203, corresponding to... Figure 1The overall process is shown below. In step S201, raw seafloor seismic data of the target area is acquired through a sensor array deployed in the target area. This data is typically a common receiver gather or a common shot gather, with a time-space trace index. Simultaneously, marine environmental parameters closely related to the time and spatial location of this data acquisition are acquired through environmental sensors, such as surface current velocity, current direction, water depth, thermocline depth, and wave height. These parameters constitute a feature vector characterizing the physical origin of flow-induced noise. In step S202, the marine environmental parameters are input into a deep learning network to generate a physically guided noise field. The core of this generation process lies in physical consistency constraints, ensuring that the generated noise field possesses the low-frequency, coherent, and smooth characteristics of flow-induced noise. This physically guided noise field, as an interpretable intermediate product, intuitively demonstrates the noise pattern "as perceived" by the model, enhancing the model's credibility. In step S203, the signal extraction network receives the raw seafloor seismic data and the physically guided noise field. Utilizing the latter as powerful prior noise information, it decouples the effective seismic signal from the raw data through an adaptive separation mechanism and reconstructs and outputs the noise-suppressed seafloor seismic dataset. The role of the signal extraction network is to perform precise signal-noise separation using physically guided information. In the above scheme, the raw seafloor seismic data is the noisy data to be processed, the marine environmental parameters are the physical causes of flow-induced noise, and the physically guided noise field serves as a bridge connecting the physical world and the deep learning model. The signal extraction network is responsible for completing the core noise suppression and signal reconstruction tasks under physical guidance.

[0039] In the above implementation, the selection of marine environmental parameters can be adjusted according to the actual working area conditions. For example, in addition to surface flow velocity, flow direction, water depth, and thermocline depth, parameters such as seabed topographic slope, seabed sediment type, and tidal cycle can also be included to more comprehensively characterize the physical causes of flow-induced noise. These parameters can be further processed by the encoder to generate higher-dimensional environmental context feature vectors. In step S202, the physically guided noise field generation model can be replaced with other generation models, such as those based on generative adversarial networks (GANs) or variational autoencoders (VAEs), but its training process must be constrained by physical consistency constraints. For example, a loss term based on physical equations (such as a simplified form of the Navier-Stokes equations) can be used to more strictly constrain the spatiotemporal evolution of the noise field. In step S203, the signal extraction network can be replaced with other advanced deep learning structures, such as the U-Net structure based on convolutional neural networks (CNNs) or the topology based on graph neural networks (GNNs). However, the core requirement is that the network must be able to explicitly receive and utilize the physically guided noise field as prior information to guide the signal separation process. For example, an attention mechanism can be introduced into the feature fusion layer of the CNN to perform weighted fusion of the feature map of the noise field with the feature map of the original data, thereby realizing the introduction of noise prior.

[0040] In some embodiments, marine environmental parameters are input into an environmental encoder to generate a high-dimensional environmental context feature vector characterizing the physical state of the current acquisition environment; a spatiotemporal coordinate tensor corresponding to the dimensions of the original seafloor seismic data is constructed, with each element in the spatiotemporal coordinate tensor associated with the temporal and spatial trace index information of the corresponding sampling point in the marine environmental parameters; the environmental context feature vector and the spatiotemporal coordinate tensor are input together into a noise field generator, which outputs a spatiotemporally continuous preliminary noise field through its internal differentiable neural network mapping; during the training process of the noise field generator, a physical consistency loss function is used to constrain the preliminary noise field, forcing its energy distribution and spatiotemporal variation patterns to conform to the low-frequency, coherent, and smooth physical priors of flow-induced noise, thereby generating a physically guided noise field.

[0041] This implementation focuses on the physical guidance noise field generation process, involving an environmental encoder, a spatiotemporal coordinate tensor, a noise field generator, and physical consistency constraints. The environmental encoder employs a small feedforward neural network (such as a three-layer MLP) to receive input marine environmental parameters such as current velocity, current direction, water depth, and thermocline depth, encoding them into an environmental context feature vector of dimension D (e.g., D=64 or ). This feature vector carries the physical cause information of the flow-induced noise. The spatiotemporal coordinate tensor is a tensor with the same time and spatial trace dimensions as the original seafloor seismic data. Each element in the tensor is a multidimensional vector containing the time index of the corresponding sampling point, spatial trace indices such as receiver location coordinates, and acquisition timestamp information associated with the environmental parameters. The noise field generator employs a deep learning model based on MLP or Siren. It receives the concatenated environmental context feature vector and spatiotemporal coordinate tensor, performs differentiable mapping, and outputs a preliminary noise field with dimensions consistent with the original data. Physical consistency constraints: When training the noise field generator, physical consistency loss terms, such as spectral loss and gradient loss, are added to the total loss function to impose constraints on the initial noise field. For example, low-spectral constraints penalize high-frequency energy, forcing the noise field to maintain low-frequency properties; spatiotemporal smoothness constraints penalize drastic spatiotemporal changes, forcing the noise field to remain coherent and smooth. These constraints ensure that the final generated physically guided noise field conforms to the physical priors of flow-induced noise in both the spatiotemporal and frequency domains.

[0042] In the above implementation, the environmental encoder can adopt a Transformer-based structure to better capture the nonlinear interactions between different marine environmental parameters such as current velocity and thermocline depth. For example, a self-attention mechanism can be used to weight the importance of different environmental parameters, thereby generating a more representative environmental context feature vector. The noise field generator can be replaced with a model based on implicit neural representation (INR), such as a Siren network using periodic activation functions like sine functions. This helps generate a spatiotemporally continuous noise field with higher resolution and stronger smoothness, especially suitable for describing the coherence of flow-induced noise. The encoding method of the spatiotemporal coordinate tensor can be optimized. In addition to directly using time indexes and spatial trace indexes, absolute position coordinates such as latitude and longitude, water depth, and relative time such as relative to the tidal cycle can be Fourier encoded to improve the noise field generator's ability to model high-frequency details while maintaining the capture of low-frequency coherence. The implementation of physical consistency constraints can be diversified. In addition to imposing constraints on the loss function, physical priors can be used as regularization terms, or a physical model layer can be embedded in the architecture of the noise field generator. For example, a dedicated convolutional layer can be used to simulate the diffusion or advection process of the flow field in time and space, so as to more closely integrate physical laws.

[0043] In some embodiments, the physical consistency loss function includes at least one of the following constraints: a low-spectrum constraint, a spatiotemporal smoothness constraint, and an apparent velocity range constraint; the low-spectrum constraint is to penalize the high-frequency portion of the energy of the initial noise field after frequency domain transformation; the spatiotemporal smoothness constraint is to calculate the gradient of the initial noise field in the time and spatial dimensions and penalize its large changes; the apparent velocity range constraint is to transform the initial noise field to the τ-p domain and encourage its energy to concentrate in the low-velocity range corresponding to ocean current interference.

[0044] This implementation introduces three physical consistency constraints to force the generated initial noise field to conform to the physical priors of flow-induced noise across three dimensions: frequency domain, spatiotemporal domain, and apparent velocity domain. The low-spectrum constraint penalizes high-frequency energy to ensure the noise field possesses low-frequency properties, preventing the model from misclassifying high-frequency geological signals as noise. The spatiotemporal smoothness constraint penalizes spatiotemporal gradients to ensure the noise field is continuous and coherent in time and space, conforming to the characteristics of fluid disturbance fluctuations and preventing the generator from outputting high-frequency, irregular noise. The apparent velocity range constraint addresses the unique characteristics of flow-induced noise, which has an extremely low apparent velocity; by transforming the data to... The slowness-time domain clearly separates energy at different velocities. This constraint encourages the energy of the noise field to concentrate in the region of greater slowness, i.e., the low-speed range, thereby accurately simulating the characteristics of ocean current interference. These three constraints together constitute a strong physical prior, significantly improving the physical plausibility of the noise field and the guiding accuracy of the signal extraction network.

[0045] In the above implementation process, low-spectral constraints can be implemented using wavelet transform-based constraints, penalizing the coefficients corresponding to high-frequency components in the wavelet coefficients to achieve finer frequency domain control. Spatiotemporal smoothness constraints can be replaced with total variation (TV)-based constraints, i.e., penalizing the L1 norm of the gradient. This is beneficial for preserving potentially sharp boundaries in the noise field (e.g., fluid disturbance boundaries caused by abrupt changes in seabed topography) while maintaining smoothness in most areas. Apparent velocity range constraints can be enhanced by introducing a constraint based on sparse representation. For example, in... In the domain, besides penalizing high-speed energy, the sparsity of low-speed energy can be encouraged. This involves applying an L1 norm penalty in the low-speed range to make flow-induced noise appear as a few concentrated energy clusters. Furthermore, more complex wavefield decomposition methods (such as higher-order Radon transforms) can be used to replace the standard method. Transformation to improve the accuracy of capturing nonlinear coherent noise.

[0046] In some embodiments, the physically guided noise field and the original submarine seismic data are concatenated along the channel dimension to form combined data. The combined data is input into a signal extraction network. The signal extraction network explicitly distinguishes and processes information from the noise field channel and the data channel through its initial attention layer, and uses the physically guided noise field as a spatial attention guidance condition to focus the signal extraction network on data regions with a morphological similarity greater than a preset similarity with the noise field, thus completing the noise prior. Guided by the noise prior, the signal extraction network, through hierarchical Transformer blocks, learns to decouple and separate components related to the physically guided noise field from the original submarine seismic data, while retaining and reconstructing effective signal components related to geological reflections, to obtain a preliminary denoised seismic signal. Based on the original submarine seismic data, the preliminary denoised seismic signal is analyzed to construct a submarine seismic dataset.

[0047] The signal extraction network works as follows: First, by concatenating the raw data and the physically guided noise field along the channel dimension, the model obtains the physical prior information of the noise at the input stage. Second, the initial attention layer explicitly distinguishes the information of these two channels and uses the noise field as a spatial attention guidance condition. This means that the model calculates the morphological similarity between each spatiotemporal location in the raw data and the noise field, and concentrates attention resources on data regions with high similarity (i.e., high probability of noise). This completes the import of the noise prior, enabling the model to accurately know "where the noise is". Subsequently, the hierarchical Transformer block, leveraging its powerful long-range dependency modeling capabilities, learns the decoupling function between signal and noise under the guidance of the noise prior. The Transformer structure can capture the temporal and spatial coherence of seismic data, thereby accurately preserving and reconstructing effective signal components such as geological reflections while separating noise components, ultimately outputting the pre-denoised seismic signal. Finally, through residual analysis, the pre-denoised signal is optimized to construct the final submarine seismic dataset.

[0048] The core of this implementation is the structure and function of the signal extraction network, which adopts a Transformer-based architecture to achieve accurate guidance from noise priors. Combined data: consisting of raw submarine seismic data ( ) and physically guided noise field (dimension) It is pieced together along the channel dimension to form a dimension of The input tensor. Initial attention layer: This is a custom attention module that explicitly introduces a noise prior. This layer first maps the combined data to a high-dimensional feature space. Then, it uses the feature map of the noise field as a "query" or "key / value" to compute a similarity matrix between the feature maps of the original data channels and the noise field channels. This similarity matrix generates a spatial attention weight map, indicating which regions in the original data are highly similar to the morphology of the physically guided noise field. Noise prior: This is the spatial attention weight map generated by the initial attention layer. This weight map assigns high weights to regions with a high probability of noise, thus guiding subsequent hierarchical Transformer blocks to focus on these regions. Hierarchical Transformer block: Composed of multiple Transformer encoders, employing a hierarchical design (such as a pyramid structure) to capture spatiotemporal features at different scales. Within each Transformer block, a self-attention mechanism, guided by the noise prior, learns how to decompose the feature vectors in the original data into signal and noise components, achieving decoupling and separation. The seismic signal after preliminary noise reduction is the direct output of the signal extraction network and is the seismic data after preliminary noise suppression.

[0049] In the above implementation, the data concatenation method can be replaced by a feature fusion method. For example, the original data and the physically guided noise field can be extracted for features through independent convolutional layers, and then weighted summation or element-wise multiplication can be performed in the feature space as input to the signal extraction network. The initial attention layer can employ a more complex cross-attention mechanism. For example, a two-stream network can be designed, with one stream processing the original data and the other processing the noise field. Then, through the cross-attention mechanism, the features of the noise field are used as keys and values, and the features of the original data are used as queries, achieving more refined noise guidance. The hierarchical Transformer block can be replaced with other types of Transformer structures, such as SwinTransformer or ConvNeXt, to improve computational efficiency and feature extraction capabilities. Inside the Transformer block, a gating mechanism can be introduced to dynamically control the degree of influence of the noise prior on the feature decoupling process, ensuring that the influence of the noise prior is minimized in the signal region. The noise prior can be implemented using a probability density function. That is, the output of the initial attention layer is not a rigid attention weight map, but a probability distribution map representing the presence of noise at each spatiotemporal point. Subsequent Transformer blocks then perform feature weighting based on this probability distribution map.

[0050] In some embodiments, the signal extraction network is a deep learning model with a visual Transformer structure as its backbone; the signal extraction network processes the input of the combined data through a self-attention mechanism and a cross-attention mechanism, wherein the physical guidance noise field is used as the key and value of the cross-attention mechanism to guide the model to focus on the context in the data that is related to the noise prior.

[0051] In seismic data processing, coherent noise, such as flow-induced noise, exhibits long-range spatiotemporal correlations across the entire gather, which traditional CNNs struggle to capture effectively. This implementation uses ViT as the backbone of the signal extraction network to fully utilize its global modeling capabilities. The network works as follows: First, the combined data (containing the original data and the noise field) is segmented into token sequences and positionally encoded. Second, through a self-attention mechanism, the model captures the long-range spatiotemporal correlations between signals and noise across the entire dataset. The core innovation lies in the introduction of a cross-attention mechanism. In this mechanism, features from the original data (as query Q) interact with features from the physically guided noise field (as key K and value V). Since the noise field has been physically constrained to ensure its physical validity, using it as K and V allows the original data Q to query contextual information highly similar to the physical noise morphology. This mechanism explicitly injects prior noise information into the feature extraction process, guiding the model to accurately decouple noise components in the feature space, thereby achieving high-precision signal separation.

[0052] Signal Extraction Network: The backbone adopts a ViT structure, such as a hierarchical ViT (e.g., SwinTransformer), to balance local feature extraction and global context modeling. Self-Attention Mechanism: Applied within the combined data, it captures the spatiotemporal interrelationships and long-range dependencies between seismic data and noise fields. This mechanism ensures the model understands the coherence of flow-induced noise across the entire gather. Cross-Attention Mechanism: This is the core component for implementing physical guidance. In the cross-attention layer: Query (Q): Feature representation from the raw seafloor seismic data. Key (K) and Value (V): Feature representation from the physical guidance noise field. Cross-Attention Calculation An attention weight is generated by calculating the similarity between the original data feature Q and the noise field feature K. This weight indicates which features in the original data should be "influenced" or "guided" by the noise field feature V. Since K and V come from the physically constrained noise field, this guidance-driven model focuses attention on contextual information related to the physical noise, thus achieving accurate importation of noise priors. The network output is a feature representation of the pre-denoised seismic signal, which is then reconstructed into a time-domain signal through a decoder (such as an MLP or deconvolution layer).

[0053] In the above implementation, the backbone of the signal extraction network can be replaced with a hybrid structure based on convolutional neural networks and Transformers (such as ConvNeXt or CoaT) to improve the efficiency of local feature extraction while maintaining global context modeling capabilities. The implementation of the cross-attention mechanism can be adjusted. For example, multi-head cross-attention can be used, where different heads focus on different physical properties of the noise field (such as low-frequency components, coherent paths, etc.) to achieve finer guidance. The role of the physically guided noise field in the cross-attention mechanism can be extended. Besides serving as the key K and value V, its influence on the query Q can be controlled through a gating mechanism. For example, through a learnable gating parameter. Dynamic adjustment To adapt the guiding intensity under different signal-to-noise ratio conditions, the tokenization process in the ViT structure can employ overlapping patch embedding to preserve the local spatial continuity of seismic data, which is crucial for handling wavefield characteristics of coherent noise.

[0054] In some embodiments, an initial residual is calculated between the original submarine seismic data and the pre-denoised seismic signal; the initial residual, the pre-denoised seismic signal, and the physical guidance noise field are jointly input into a residual optimization model; the residual optimization model is a lightweight convolutional neural network used to analyze the residual noise, over-removed effective signal components, and noise components that the physical guidance noise field fails to fully model in the initial residual, and outputs a fine residual correction; the residual correction is added to the pre-denoised seismic signal to obtain the noise-suppressed submarine seismic dataset.

[0055] Although signal extraction networks utilize physical guidance, two types of errors may still exist in complex real-world data: residual noise (noise that has not been completely removed) and excessive removal of the effective signal (signal impairment). This implementation introduces a residual optimization model to refine the initial denoising results. Its working principle is based on the idea of ​​residual learning: the difference between the initial denoising result and the original data, i.e., the initial residual, is used as the correction target. The residual optimization model receives three inputs: the initial residual containing error information, the initial denoised signal (providing signal context), and the physically guided noise field (providing noise prior). The role of the residual optimization model (lightweight CNN) is to analyze the composition of the initial residual. For example, if there are high-frequency, incoherent components in the residual, it may correspond to residual random noise; if there are structures in the residual similar to the in-phase axis of the effective signal, it may correspond to excessively removed effective signal. By combining the contextual information of the initial denoised signal and the noise field, the model can distinguish these components and output a refined residual correction. This correction includes the recovery of the effective signal and further removal of residual noise. Finally, the correction amount is added back to the initial denoised signal to obtain the final high-fidelity submarine earthquake dataset.

[0056] Initial residual: The calculation formula is as follows ,in It is raw undersea earthquake data. This is the seismic signal after preliminary denoising. The residual represents the sum of noise and error perceived by the signal extraction network. The residual optimization model employs a lightweight convolutional neural network (such as a ResNet block or a simplified version of U-Net), designed to efficiently process the input tensor and output a correction. The model receives a multi-channel input tensor, with channels including... , and (Physically guided noise field). Output of the correction: Output of the residual optimization model. This is the precise residual correction amount. If If it is positive, it means that it needs to be obtained from Subtract residual noise from the middle; if A negative value indicates that the effective signal from the excessive removal needs to be added back. In the middle. Final output: Noise-suppressed submarine seismic dataset. The calculation formula is: .

[0057] In the above implementation process, the residual optimization model can be replaced with a sequence model based on Gated Recurrent Units (GRUs) or Long Short-Term Memory (LSTM) networks, especially when the residual noise exhibits time-series correlation, as such models can better capture its dynamic characteristics. The input to the residual optimization model can be optimized. In addition to the initial residual, the initial denoised signal, and the noise field, an environmental context feature vector (from the environmental encoder) can be used as a conditional input to help the model better understand the correlation between residual noise and environmental changes. The method of applying corrections can be replaced with a weighted average. For example, the final output... It can be ,in It is a weighted graph output by the model, dynamically determining the balance between the initial denoising result and the result of subtracting the noise field from the original data. The initial residual can be calculated using frequency domain residuals. That is, calculated in the frequency domain. This is then fed into the residual optimization model to focus more on correcting frequency domain errors, such as high-frequency random noise or low-frequency residual coherent noise.

[0058] In some embodiments, the training of the residual optimization model is performed jointly with the training of the signal extraction network, and the total loss function of the joint training includes a signal reconstruction loss applied to the final output and a sparsity constraint loss applied to the final residual; the signal reconstruction loss is used to measure the difference between the final output and the weakly supervised label; the sparsity constraint loss is used to encourage complete noise separation and complete preservation of effective signals by applying an L1 norm penalty to the final residual between the original seafloor seismic data and the final output, so that the final residual does not contain structural information.

[0059] To ensure the signal extraction network and residual optimization model work together effectively and achieve high-fidelity denoising, this implementation employs an end-to-end joint training mechanism. This joint training allows both modules to jointly optimize their parameters under the guidance of a global loss function, achieving optimal signal separation and reconstruction results. Total Loss Function It comprises two key components: signal reconstruction loss and sparsity constraint loss Signal reconstruction loss Its purpose is to ensure that the final output is as close as possible to the weakly supervised label (such as low-perturbation seismic signal data). This provides the main supervisory signal, guiding the model to learn how to recover the effective signal. (Sparseness constraint loss) This is one of the key innovations of this invention. Final residual This represents all the noise components as perceived by the model. Its physical meaning is that if the noise is completely separated and the effective signal is fully preserved, then... It should contain only noise and not any structural information (such as phase axes, reflecting interfaces, etc.). Through the analysis of... Applying an L1 norm penalty (sparseness constraint) forces the model to retain all structural information in the final output, while concentrating unstructured noise components. In the middle, the L1 norm (compared to the L2 norm) naturally encourages sparsity, which helps to completely separate noise.

[0060] Joint training: All trainable parameters of the signal extraction network and the residual optimization model are updated in the same iteration, ensuring collaborative optimization of the two modules. Total loss function. Defined as ,in and These are weighting coefficients. This is a loss of physical consistency. Signal reconstruction loss. The L1 or L2 norm is typically used to measure the final output. With weak supervision label The differences between them.

[0061] Sparse constraint loss : Acts on the final residual By imposing L1 norm penalties, we encourage... sparsity, i.e. The energy in the signal should be concentrated as much as possible in the noise component and should not contain structural signal information.

[0062] Through joint training and sparsity constraints, the model is forced to learn a decomposition function that decomposes the original data into... (High-fidelity signal) and (Sparse noise), thus avoiding excessive damage to the effective signal.

[0063] During the above implementation process, signal reconstruction loss Perceptual loss or structural similarity loss (SSIMLoss) can be used to better measure the visual and structural similarity of the final output to the weakly supervised label, especially in maintaining the continuity of the phase axis. Sparse constraint loss. Non-L1 norm sparsity measures, such as entropy-based constraints, can be used to encourage sparsity in the final residual. The probability distribution of the signal has high entropy (i.e., strong randomness and lack of structural information). During joint training, a multi-scale loss mechanism can be employed. That is, the signal reconstruction loss is calculated separately at different levels of the network. The loss function is then weighted and summed to ensure that the model can accurately separate signal and noise at different scales. To enhance the robustness of the model, an adversarial loss can be added to the total loss function, in which a discriminator is used to distinguish the final output from the weakly supervised labels, thereby further improving the realism and signal-to-noise ratio of the final output.

[0064] In some embodiments, a training dataset is prepared, which contains multiple sets of samples. Each set of samples includes: raw submarine seismic data segments, spatiotemporally correlated marine environmental parameter vectors, and low-disturbance seismic signal data as the weak supervision label. Using the total loss function as the objective, the signal extraction network and the residual optimization model are jointly trained end-to-end using the training dataset. This enables the signal extraction network and the residual optimization model to learn the differences between raw seismic data and clean seismic signal data under different marine environmental influences, thereby enabling the signal extraction network and the residual optimization model to accurately separate and reconstruct effective seismic signals under the guidance of a noise field.

[0065] The performance of deep learning models heavily depends on the quality of training data and the training strategy. A major challenge in processing submarine seismic data is the difficulty in obtaining "absolutely pure" signal labels. This implementation employs a weakly supervised learning strategy, utilizing "low-disturbance seismic signal data" as weakly supervised labels, and combines end-to-end joint training to maximize the model's practicality and generalization ability. The joint training process works as follows: each set of samples in the training dataset contains raw data (noisy), marine environmental parameters (the physical causes of the noise), and low-disturbance signal data (weakly supervised labels). By inputting the environmental parameters into a noise field generator, a physically guided noise field is generated. Then, a noise field-guided signal extraction network and a residual optimization model process the raw data. The total loss function is used... With this goal in mind, the model learns during training how to generate reasonable noise estimates based on the marine environmental parameter E. and utilize Accurately extract effective signals from raw data , making The training strategy aims to obtain labels that are as close as possible to those with weak supervision. This end-to-end training approach enables the model to explicitly learn and adapt to the effects of noise on different marine environments, thus giving it strong generalization capabilities in practical applications.

[0066] Training dataset: contains Groups of samples, each group of samples By triplet Composition, among which, This is a raw submarine seismic data segment. It is a vector of marine environmental parameters with spatiotemporal correlation. This corresponds to low-disturbance seismic signal data (weakly supervised label). The size of the data segment is typically a common-detector point gather or a subset thereof. Marine environmental parameter vector. It must include key parameters such as flow velocity, flow direction, water depth, and thermocline depth to drive the noise field generator. (Weak supervision tag) The signal-to-noise ratio (SNR) of the obtained signal data is higher than that of the original data after applying relatively reliable denoising processes such as long-offset superposition and high-precision FK filtering. However, it may still contain a small amount of residual noise. End-to-end joint training: Minimize the total loss function using optimizers such as Adam or SGD. In each training iteration, the data and The inputs to the entire network architecture include an environmental encoder, a noise field generator, a signal extraction network, and a residual optimization model. Loss function Based on the final output and The differences and sparsity of the final residuals are calculated, and all network parameters are updated by backpropagation.

[0067] In summary, the preparation of the training dataset can employ data augmentation techniques, such as random time-shifting, amplitude scaling, or adding random background noise to the original data, to improve the model's robustness. The joint training process can utilize a staged training strategy. For example, the noise field generator can be trained independently first (using physical consistency loss and a small amount of supervision information), then frozen, and finally trained end-to-end for the signal extraction network and the residual optimization model. The optimizer can be an adaptive learning rate optimizer (such as AdamW) combined with a learning rate scheduling strategy (such as Cosine Annealing) to ensure the stability and convergence speed of the training process. During training, model uncertainty estimation can be introduced. For example, Monte Carlo Dropout or Bayesian neural networks can be used to output an uncertainty map along with the final result to evaluate the reliability of the denoising results.

[0068] In some embodiments, the low-disturbance seismic signal data is seismic data obtained by applying long-offset stacking processing to the original seafloor seismic data; the signal-to-noise ratio of the low-disturbance seismic signal data is higher than that of the original seafloor seismic data.

[0069] Weakly supervised learning requires labeled data It possesses a relatively high signal-to-noise ratio (SNR), but does not need to be "absolutely pure." Flow-induced noise typically exists in a low-velocity, coherent form in near-offset data. However, in far-offset stacking, effective signals such as formation reflections have high velocities, and their in-phase axes are effectively enhanced during stacking. Meanwhile, low-velocity flow-induced noise is effectively suppressed due to its low-velocity characteristics and spatial coherence. Therefore, seismic data obtained by applying far-offset stacking to raw submarine seismic data has a significantly higher SNR than the original data, making it an economical and practical weakly supervised label. Although such labels may still contain a small amount of residual noise, their effective signal components are greatly preserved and enhanced, thus meeting the requirements for signal reconstruction loss. Requirements for label quality.

[0070] Low-disturbance seismic signal data Generation Process: Data Screening: First, the raw seafloor seismic data is screened by offset, selecting only seismic traces with offsets greater than a preset threshold (e.g., greater than 1000 meters). This is because flow-induced noise is usually strongest at near offsets. Velocity Analysis: A detailed velocity analysis is performed on the screened far-offset data to determine the superposition velocity of effective signals such as first-order reflected waves. Superposition Processing: Conventional dynamic correction and superposition processing (such as CMP superposition or common-detector superposition) are applied. Due to the extremely low apparent velocity of flow-induced noise, its energy is effectively suppressed during superposition due to insufficient coherence. Output: The superimposed data is the low-disturbance seismic signal data. Signal-to-noise ratio improvement: Through long-offset superposition processing, the amplitude of the effective signal is enhanced, while the amplitude of flow-induced noise is relatively weakened, thereby improving the signal-to-noise ratio. The signal-to-noise ratio is significantly higher than that of the original submarine seismic data. For example, the signal-to-noise ratio can be improved by 5dB to 10dB.

[0071] Low-disturbance seismic signal data Other advanced denoising methods can be used to generate weakly supervised labels. For example, methods based on High-Order Singular Value Decomposition (HOSVD) or sparse representation can be used to preprocess the original data to generate higher-quality weakly supervised labels. The offset threshold can be dynamically adjusted according to the water depth and ocean current intensity of the actual work area. In deep-water work areas, the influence of flow-induced noise is wider, and the offset threshold may need to be set higher. The overlay processing can be replaced by high-precision denoising methods, such as methods based on 3D block matching and 3D filtering (BM3D), to further suppress random noise in the far-offset data to improve the purity of the weakly supervised labels. To improve the reliability of the labels, an ensemble strategy of multiple weakly supervised labels can be adopted. For example, far-offset overlay labels and HOSVD-based denoised labels can be generated simultaneously, and the signal reconstruction loss can be minimized. The training employs a weighted average method for supervision to improve its robustness.

[0072] Figure 3 A schematic diagram of a deep learning-based seismic data noise suppression system provided in one embodiment of this application is shown below. Figure 3 As shown, a seismic data noise suppression system 300 based on deep learning in this embodiment includes: a data acquisition module 301, a data mapping module 302, and a data reconstruction module 303.

[0073] The data acquisition module 301 is used to acquire the original submarine seismic data of the target area, as well as marine environmental parameters that are spatiotemporally related to the original submarine seismic data. The data mapping module 302 is used to map the marine environmental parameters to generate a spatiotemporally continuous physical guidance noise field. The generation process of the physical guidance noise field is subject to physical consistency constraints, and the physical guidance noise field has the physical properties of low frequency, coherence and smoothness of flow-induced noise. The data reconstruction module 303 is used to adaptively separate effective seismic signals from the original submarine seismic data using the physical guidance noise field as physical guidance information, and to reconstruct and output the noise-suppressed submarine seismic dataset.

[0074] Optionally, the data mapping module 302 is specifically used for: inputting the marine environmental parameters into an environmental encoder to encode and generate a high-dimensional environmental context feature vector characterizing the physical state of the current acquisition environment; constructing a spatiotemporal coordinate tensor corresponding to the dimension of the original seafloor seismic data, wherein each element in the spatiotemporal coordinate tensor is associated with the temporal and spatial trace index information of the corresponding sampling point in the marine environmental parameters; inputting the environmental context feature vector and the spatiotemporal coordinate tensor together into a noise field generator, wherein the noise field generator outputs a spatiotemporally continuous preliminary noise field through its internal differentiable neural network mapping; during the training process of the noise field generator, constraining the preliminary noise field through a physical consistency loss function, forcing its energy distribution and spatiotemporal variation pattern to conform to the low-frequency, coherent, and smooth physical priors of flow-induced noise, thereby generating the physical guidance noise field.

[0075] Optionally, in the data mapping module 302, the physical consistency loss function includes at least one of the following constraints: low-spectrum constraint, spatiotemporal smoothness constraint, and apparent velocity range constraint; the low-spectrum constraint is to penalize the high-frequency energy of the initial noise field after frequency domain transformation; the spatiotemporal smoothness constraint is to calculate the gradient of the initial noise field in the time and space dimensions and penalize its large changes; the apparent velocity range constraint is to transform the initial noise field to the τ-p domain and encourage its energy to concentrate in the low-speed range corresponding to ocean current interference.

[0076] Optionally, the data reconstruction module 303 is specifically used for: splicing the physical guidance noise field with the original submarine seismic data in the channel dimension to form combined data; inputting the combined data into a signal extraction network, which explicitly distinguishes and processes information from the noise field channel and the data channel through its initial attention layer, and uses the physical guidance noise field as a spatial attention guidance condition to make the signal extraction network focus on the data region with a morphological similarity to the noise field greater than a preset similarity, thus completing the noise prior; the signal extraction network, guided by the noise prior, learns to decouple and separate the components related to the physical guidance noise field from the original submarine seismic data through a hierarchical Transformer block, while retaining and reconstructing effective signal components related to geological reflection, thus obtaining a preliminary denoised seismic signal; and analyzing the preliminary denoised seismic signal based on the original submarine seismic data to construct the submarine seismic dataset.

[0077] Optionally, in the data reconstruction module 303, the signal extraction network is a deep learning model with a visual Transformer structure as its backbone; the signal extraction network processes the input of the combined data through a self-attention mechanism and a cross-attention mechanism, wherein the physical guidance noise field is used as the key and value of the cross-attention mechanism to guide the model to focus on the context related to the noise prior in the data.

[0078] Optionally, when constructing the submarine earthquake dataset based on the original submarine earthquake data and analyzing the pre-denoised seismic signal, the data reconstruction module 303 is specifically used to: calculate the initial residual between the original submarine earthquake data and the pre-denoised seismic signal; input the initial residual, the pre-denoised seismic signal, and the physical guidance noise field into the residual optimization model; the residual optimization model is a lightweight convolutional neural network used to analyze the residual noise, over-removed effective signal components, and noise components that the physical guidance noise field fails to fully model in the initial residual, and output a fine residual correction amount; add the residual correction amount to the pre-denoised seismic signal to obtain the noise-suppressed submarine earthquake dataset.

[0079] Optionally, in the data reconstruction module 303, the training of the residual optimization model and the training of the signal extraction network are performed jointly, and the total loss function of the joint training includes a signal reconstruction loss applied to the final output and a sparsity constraint loss applied to the final residual; the signal reconstruction loss is used to measure the difference between the final output and the weakly supervised label; the sparsity constraint loss is used to encourage complete separation of noise and complete preservation of effective signal by applying L1 norm penalty to the final residual between the original seafloor seismic data and the final output, so that the final residual does not contain structural information.

[0080] Optionally, in the data reconstruction module 303, the joint training process of the signal extraction network and the residual optimization model is specifically used for: preparing a training dataset, the training dataset containing multiple sets of samples, each set of samples including: original seafloor seismic data segments, spatiotemporally correlated marine environmental parameter vectors, and low-disturbance seismic signal data as the weak supervision label; using the training dataset as the objective of the total loss function, performing end-to-end joint training of the signal extraction network and the residual optimization model, so that the signal extraction network and the residual optimization model learn the differences between the original seismic data and the clean seismic signal data under different marine environmental influences, thereby enabling the signal extraction network and the residual optimization model to accurately separate and reconstruct effective seismic signals under the guidance of a noise field.

[0081] Optionally, in the data reconstruction module 303, the low-disturbance seismic signal data is seismic data obtained by applying long-offset stacking processing to the original seafloor seismic data; the signal-to-noise ratio of the low-disturbance seismic signal data is higher than that of the original seafloor seismic data.

[0082] The system in this embodiment can be used to execute the methods of any of the above embodiments, and its implementation principle and technical effect are similar, so they will not be described again here.

Claims

1. A method for suppressing seismic data noise based on deep learning, characterized in that, include: Acquire raw submarine seismic data for the target area, as well as marine environmental parameters that are spatiotemporally correlated with the raw submarine seismic data; The marine environmental parameters are mapped to generate a spatiotemporally continuous physical guidance noise field, including: The marine environmental parameters are input into the environmental encoder to generate a high-dimensional environmental context feature vector that characterizes the physical state of the currently collected environment. Construct a spatiotemporal coordinate tensor corresponding to the dimensions of the original seafloor seismic data, wherein each element in the spatiotemporal coordinate tensor is associated with the temporal and spatial trace index information of the corresponding sampling point in the marine environmental parameters; The environmental context feature vector and the spatiotemporal coordinate tensor are input into the noise field generator, which outputs a spatiotemporally continuous preliminary noise field through its internal differentiable neural network mapping. During the training process of the noise field generator, the initial noise field is constrained by the physical consistency loss function, which forces its energy distribution and spatiotemporal variation pattern to conform to the low-frequency, coherent and smooth physical priors of flow-induced noise, thereby generating the physical guided noise field. Using the physical guidance noise field as physical guidance information, effective seismic signals are adaptively separated from the original submarine seismic data, and a noise-suppressed submarine seismic dataset is reconstructed and output, including: The physical guidance noise field is spliced ​​with the original submarine seismic data in the channel dimension to form combined data; The combined data is input into the signal extraction network. The signal extraction network explicitly distinguishes and processes information from the noise field channel and the data channel through its initial attention layer. It uses the physical guidance noise field as a spatial attention guidance condition to make the signal extraction network focus on the data region with a morphological similarity to the noise field greater than a preset similarity, thus completing the noise prior. The signal extraction network, guided by the noise prior, learns to decouple and separate components related to the physical guidance noise field from the original submarine seismic data through hierarchical Transformer blocks, while retaining and reconstructing effective signal components related to geological reflections, thus obtaining a preliminarily denoised seismic signal. Based on the raw submarine earthquake data, the earthquake signals after preliminary denoising are analyzed to construct the submarine earthquake dataset.

2. The method according to claim 1, characterized in that, The physical consistency loss function includes at least one of the following constraints: low-spectrum constraint, spatiotemporal smoothness constraint, and apparent velocity range constraint; The low-frequency constraint is to penalize the high-frequency portion of the energy after the initial noise field has undergone frequency domain transformation. The spatiotemporal smoothness constraint is: to calculate the gradient of the initial noise field in the time and space dimensions and penalize its large changes; The apparent velocity range constraint is to transform the initial noise field to the τ-p domain and encourage its energy to concentrate in the low-speed range corresponding to ocean current interference.

3. The method according to claim 2, characterized in that, The signal extraction network is a deep learning model with a visual Transformer structure as its backbone. The signal extraction network processes the input of the combined data through a self-attention mechanism and a cross-attention mechanism, wherein the physical guidance noise field is used as the key and value of the cross-attention mechanism to guide the model to focus on the context in the data that is related to the noise prior.

4. The method according to claim 2, characterized in that, The process of constructing the submarine earthquake dataset based on raw submarine earthquake data, analyzing the preliminarily denoised earthquake signals, includes: Calculate the initial residual between the original submarine seismic data and the pre-denoised seismic signal; The initial residual, the pre-denoised seismic signal, and the physical guidance noise field are all input into the residual optimization model; The residual optimization model is a lightweight convolutional neural network used to analyze the residual noise, over-removed effective signal components, and noise components that the physical guidance noise field failed to fully model in the initial residual, and output a fine residual correction amount. The residual correction amount is added to the pre-denoised seismic signal to obtain the noise-suppressed submarine seismic dataset.

5. The method according to claim 4, characterized in that, The training of the residual optimization model is carried out in conjunction with the training of the signal extraction network, and the total loss function of the joint training includes a signal reconstruction loss that acts on the final output and a sparsity constraint loss that acts on the final residual. The signal reconstruction loss is used to measure the difference between the final output and the weakly supervised label; The sparsity constraint loss is used to encourage complete noise separation and full preservation of effective signals by applying an L1 norm penalty to the final residual between the original seafloor seismic data and the final output, so that the final residual does not contain structural information.

6. The method according to claim 5, characterized in that, The joint training process of the signal extraction network and the residual optimization model includes: Prepare a training dataset, which contains multiple sets of samples. Each set of samples includes: raw submarine seismic data segments, spatiotemporally correlated marine environmental parameter vectors, and low-disturbance seismic signal data as the weak supervision label. Using the total loss function as the objective, the signal extraction network and the residual optimization model are jointly trained end-to-end using the training dataset. This enables the signal extraction network and the residual optimization model to learn the differences between raw seismic data and clean seismic signal data under different marine environmental influences, thereby allowing the signal extraction network and the residual optimization model to accurately separate and reconstruct effective seismic signals under the guidance of a noise field.

7. The method according to claim 6, characterized in that, The low-disturbance seismic signal data is seismic data obtained by applying long-offset stacking processing to the original seafloor seismic data. The signal-to-noise ratio of the low-disturbance seismic signal data is higher than that of the original submarine seismic data.

8. A seismic data noise suppression system based on deep learning, characterized in that, Applied to the method as described in any one of claims 1-7, comprising: The data acquisition module is used to acquire raw submarine seismic data of the target area, as well as marine environmental parameters that are spatiotemporally correlated with the raw submarine seismic data; The data mapping module is used to map the marine environmental parameters to generate a spatiotemporally continuous physical guidance noise field. The generation process of the physical guidance noise field is subject to physical consistency constraints, and the physical guidance noise field has the physical properties of low frequency, coherence, and smoothness of flow-induced noise. The data reconstruction module is used to adaptively separate effective seismic signals from the original submarine seismic data using the physical guidance noise field as physical guidance information, and to reconstruct and output the noise-suppressed submarine seismic dataset.

Images