A multi-domain seismic diffraction wave field intelligent separation method and system

Abstract

This invention belongs to the field of intelligent processing technology for geophysical exploration data, and discloses a multi-domain intelligent separation method and system for seismic diffraction wavefields. The method acquires full-field seismic data, including common offset domain data and / or dip domain data; constructs a dual-channel joint diffusion model with the full-field seismic data as input, which includes a dual-output noise prediction network for predicting the noise components of diffracted and reflected waves respectively; trains the model using a joint loss function that simultaneously constrains the prediction errors of diffracted and reflected wave noise; independently samples two initial noises from a standard normal distribution and inputs them into the trained model; using the full-field seismic data as conditions, iteratively generates separated diffracted and reflected wave data through a reverse denoising process. Experimental results show that the technical solution of this invention can be applied to different stages such as pre-stack processing and post-imaging processing, and has good application flexibility and engineering practical value.

Description

Technical Field

[0001] This invention belongs to the field of intelligent processing technology for geophysical exploration data, and particularly relates to an intelligent separation method and system for multi-domain seismic diffraction wavefields. Background Technology

[0002] In the field of oil and gas exploration, seismic diffraction waves contain high-resolution information about local discontinuities in the subsurface medium. These discontinuities include key geological elements such as fault fracture zones, fractured areas, lithological pinch-out lines, and small-scale geological anomalies. Accurate identification and characterization of these geological bodies are crucial for reservoir evaluation and oil and gas reservoir development. However, diffraction waves are typically 1-2 orders of magnitude weaker than reflected waves and are severely aliased with strong reflected waves in conventional seismic records. Traditional seismic imaging processes often treat this as noise suppression, resulting in irreversible loss of important geological information. Therefore, developing efficient and reliable diffraction wave separation and imaging techniques is essential for accurately characterizing complex geological structures.

[0003] Traditional diffraction wave separation techniques are mainly based on transform domain processing strategies, including Radon transform, plane wave decomposition (PWD), and common diffraction surface (CDS) superposition. These techniques typically require accurate velocity field models and meticulous parameter adjustments. They are easily affected by strong reflected wave interference in complex wavefield environments or under low signal-to-noise ratio conditions, limiting separation accuracy and algorithm robustness. Furthermore, filtering-based methods struggle to completely eliminate reflected wave residues, leading to noticeable artifacts in subsequent diffraction imaging results.

[0004] In recent years, with the rapid development of artificial intelligence technology, supervised learning methods such as Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) have been introduced into the field of diffraction wave recognition and have made some progress. However, these methods rely heavily on large-scale labeled samples for network training, and their generalization ability on real complex wave field data is still insufficient, and they are difficult to effectively handle diffraction signals with extremely weak energy.

[0005] Deep generative models, especially denoising diffusion probabilistic models, have demonstrated powerful data distribution learning and high-quality sample generation capabilities in recent years in fields such as image generation and computer vision. Compared to generative adversarial networks (GANs) and variational autoencoders (VAEs), diffusion models offer advantages such as stable training and less susceptibility to pattern collapse. Diffusion models generate data through carefully designed forward denoising and backward denoising processes, and can achieve precise and controllable generation by introducing conditional inputs. This powerful generative capability provides a new technical approach for seismic data processing.

[0006] Based on the above analysis, the problems and shortcomings of existing diffraction wave separation technology in practical applications are as follows: (1) Existing transform domain-based methods typically rely on a large number of manually adjusted parameters, such as the range of p-values ​​in the Radon transform, the dip search range for plane wave decomposition, the cutoff wavenumber for fk-domain filtering, the dip threshold in the dip domain, and the number of modes in VMD decomposition. The selection of these parameters heavily depends on the experience of the processors and often requires repeated trial and error. Different parameter combinations are often required in different work areas, or even in different regions of the same data volume, making it difficult to standardize and automate the processing flow. In addition, traditional methods are mostly based on simplified geological model assumptions. When the actual geological conditions are complex—such as severe tectonic fluctuations, variable dip angles of reflected waves, intertwined wave fields, or highly heterogeneous velocity fields—the original assumptions are broken, often leading to problems such as multiple or singular dip angles in plane wave decomposition and severe aliasing in the Radon transform. As a result, the effectiveness of traditional methods is significantly reduced.

[0007] (2) Existing supervised learning methods rely on a large number of labeled samples. However, in diffraction wave separation tasks, the labeling cost is high and there is a lack of absolute truth. At the same time, supervised models are prone to overfitting to non-essential features in the training data and are sensitive to different geological backgrounds, acquisition parameters and processing procedures, resulting in limited generalization ability.

[0008] (3) The lack of a universal diffraction wave separation framework that is applicable to both the common offset domain and the tilt domain limits the practicality and flexibility of the method. Summary of the Invention

[0009] To overcome key technical bottlenecks in related technologies, such as strong dependence on velocity models, complex parameter adjustment, insufficient ability to extract weak diffraction waves, poor adaptability to complex structures, difficulties in supervised learning annotation, weak model generalization ability, and the single data domain and lack of targeted network design in existing diffusion models, this invention discloses an intelligent multi-domain seismic diffraction wavefield separation method and system. The technical solution is as follows: This invention is implemented as follows: a smart separation method for multi-domain seismic diffraction wavefields, comprising the following steps: Acquire seismic full-wave field data, which includes common offset domain data and / or dip domain data, and the seismic full-wave field data is represented as a linear superposition of diffracted wave components and reflected wave components; A dual-channel joint conditional denoising and diffusion probability model is constructed, in which the seismic full-wavefield data is continuously used as the conditional guiding input at each time step in the reverse denoising process. The dual-channel joint conditional denoising and diffusion probability model is then extended into a dual-channel joint diffusion model, which includes a dual-output noise prediction network. The dual-output noise prediction network predicts the diffraction noise component and the reflection noise component, respectively, thereby achieving mutual constraint collaborative modeling of the two types of wavefields. The dual-channel joint diffusion model is trained using a joint loss function, which simultaneously constrains the prediction error of diffraction noise and the prediction error of reflection noise. The first initial noise and the second initial noise are independently sampled from the standard normal distribution and used as the initial states of the diffracted wave and the reflected wave, respectively. The first initial noise and the second initial noise are input into the trained dual-channel joint diffusion model, and the separated diffracted wave data and reflected wave data are generated iteratively through the reverse denoising process, using the earthquake full wave field data as a condition.

[0010] Furthermore, the conditional denoising diffusion probability model is a conditional diffusion model, and the reverse denoising process uses full-field seismic conditional data at each step. Conditional guidance is represented as follows: ; In the formula, For a given first Step-by-step noisy diffraction wave field samples and seismic full-wavefield condition data Under the condition, the first denoising diffusion probability model predicts the th The conditional probability distribution of the step sample. It follows a Gaussian distribution. For the parameter is The conditional Gaussian distribution mean predicted by the neural network. For the parameter is The conditional Gaussian distribution covariance predicted by the neural network. For the first Step-by-step noisy diffraction data, For the first Step-by-step noisy diffraction data, This is seismic full-wavefield condition data. For time steps, These are the learnable parameters for the noise prediction network; Input with noise Step-noisy diffraction wave data Seismic full-wavelength condition data and time step Seismic full-wave field condition data It remains unchanged throughout the reverse process and is consistent with noisy diffraction wave data. Having the same spatial dimensions; the training loss function is expanded to: ; In the formula, For the conditional denoising diffusion probability model in the th The training loss function corresponding to the reverse denoising process. For the first Step-noisy diffraction wave data diffusion time step Seismic full-wavelength condition data and follows a standard normal distribution random noise The desired mathematical expectation, To expand the noise prediction network, To obtain from the standard normal distribution The actual noise in the mid-sample, These are the network parameterization estimates of the mean and variance of the inverse process, respectively.

[0011] Furthermore, the dual-channel joint diffusion model achieves multi-domain coordinated separation of diffracted and reflected waves in the following manner: Seismic full-wave field condition data As a conditional input, the noise prediction function is expanded to In the reverse denoising process, full-wave field information is introduced as a constraint to achieve conditional generation of diffracted wave components. A dual-channel modeling framework is constructed to represent seismic full-wavefield data as a superposition of diffracted and reflected waves. Simultaneously predict the output diffraction wave noise prediction component. and reflected wave noise prediction components ;in, For diffraction wave data, This is the data for reflected waves; During training, common offset domain data and / or tilt domain data are used as training samples to learn the response differences between reflected and diffracted waves in different domains. For common offset domain samples, the model learns the difference between the hyperbolic morphological characteristics of diffracted waves and the straight in-phase axis of reflected waves. For tilt domain samples, the model learns the difference between the arc characteristics of diffracted waves and the straight in-phase axis of reflected waves.

[0012] Furthermore, the joint loss function is: ; In the formula, To add the diffracted wave component during the forward diffusion process True Gaussian noise, For a dual-channel conditional denoising network, at the input of the first... Step-noisy diffraction wave data , No. Step-noise reflected wave data Seismic full-wavelength condition data and time step Then, the predicted diffraction wave noise component, To add the reflected wave component during the forward diffusion process True Gaussian noise, The noise component of the reflected wave predicted by the dual-channel conditional denoising network; In the reverse denoising process, from the standard normal distribution Independent sampling of initial noisy state and random disturbance noise The trained dual-channel joint diffusion model is used to predict the noise component of diffracted waves. and reflected wave noise components Then, following the iterative formula, the separated diffracted waves are gradually generated from the full seismic wavefield data. and reflected waves The expression is: ; ; In the formula, and The first The generated diffraction wave data and reflected wave data, For the first Step noise scheduling coefficient, and The first Step-by-step noisy diffraction wave data and noisy reflection wave data, and These are the diffraction noise component and the reflected noise component predicted by the dual-channel conditional denoising network, respectively. This is seismic full-wavefield condition data. and These are random Gaussian noises injected into the diffracted wave channel and the reflected wave channel, respectively, both of which follow a standard Gaussian distribution. , For cumulative parameters, The standard deviation of the random perturbation in the reverse denoising step; in, , For the forward diffusion process Noise variance scheduling parameters for each step; cumulative parameters .

[0013] Furthermore, the training process of the dual-channel joint diffusion model includes: Randomly select seismic full-wavefield data in each iteration cycle Pure diffraction wave data and pure reflected wave data and from the time interval Uniform sampling within one time step Random noise is sampled from a standard Gaussian distribution, and noisy data is constructed through a forward diffusion process. The noisy data and the seismic full-wave field data as conditions are fed into the dual-output noise prediction network for noise prediction. The joint loss function is calculated to evaluate the error between the predicted noise and the actual added noise. The gradient is calculated based on the loss and the network parameters are updated. The sampling process of the dual-channel joint diffusion model includes: from time step Begin with a pair of random noises and As the initial state, the noise is denoised step by step according to the iterative formula, and intermediate states with decreasing time steps are generated in sequence until the separated diffraction wave data and reflection wave data are finally generated.

[0014] Furthermore, the dual-channel joint diffusion model is based on the U-Net architecture and includes: The number of residual modules and the size of the convolution kernel in the encoder can be flexibly adjusted according to the size of the input data. In this embodiment, a preferred setting of 5 residual modules is adopted. Each residual module has two 3×3 convolutional layers, followed by a normalization layer and a Swish activation function to extract multi-scale features. The decoder connects to the corresponding layer of the encoder via skip connections to fuse shallow detail information; The self-attention mechanism module is used to capture global dependencies in the feature map; The time embedding module is used to incorporate the time step information of the current backpropagation process into the network; The noise prediction network takes three input channels: noisy diffraction wave data, noisy reflected wave data, and full-field condition data, and outputs the predicted diffraction wave noise and reflected wave noise.

[0015] The offset domain training dataset for the dual-channel joint diffusion model is constructed in the following manner: A reflectivity sequence is constructed, the shape of which is controlled by two sinusoidal trends, and the number and position of reflectivity are randomly selected. The reflectivity model is convolved with the Ricker wavelet, and the same wavelet is used for each reflection phase axis. The amplitude and dominant frequency of the wavelet of different phase axes vary randomly. The diffraction wave gather is generated by generating hyperbolic reflectivity with random number, position and opening angle, and the apex of the hyperbola is located above the reflection interface. Each reflectivity is convolved with the Ricker wavelet of random amplitude and dominant frequency, and the same wavelet parameters are used for diffracted waves along the same reflection phase axis; the diffraction energy is set to attenuate as the vertex distance increases; a segmented combination strategy is adopted to augment the generated pure reflection and pure diffraction gathers. The segmented combination strategy is as follows: different numbers of pure reflection wave segments and pure diffraction wave segments are randomly truncated and spliced ​​sequentially to form a complete co-offset gather, thereby expanding the diversity and coverage of the training data based on a limited sample. The tilt domain training dataset for the dual-channel joint diffusion model is constructed in the following manner: Finite difference forward modeling and Gaussian beam migration imaging were performed on the velocity model containing only continuous reflection interfaces and the same model with added random diffraction points, respectively, to obtain pure reflection wave gathers and full wavefield gathers containing reflection and diffraction. The diffraction wave gathers were separated from the full wavefield through a convolutional neural network to form training sample pairs containing seismic full wavefield data, corresponding pure diffraction data, and pure reflection data.

[0016] Furthermore, the training process of the dual-channel joint diffusion model includes: Using the full-wave field gather as the conditional input and the corresponding pure reflected wave or pure diffraction wave gather as the target output, the conditional distribution for gradually recovering a clean signal from noisy target data is learned. The following training parameters are all optimized configurations obtained through experiments. They are specific settings for the data scale and hardware conditions of this embodiment and are not essential features of the technical solution of this invention. Those skilled in the art can make corresponding adjustments according to the actual data size, computing resources, and task requirements. For offset domain data, the network input / output size is fixed at 128×128, the model is trained for a total of 277 epochs, the batch size is 4, the Adam optimizer is used, and the initial learning rate is set to 0.0001. The total number of steps in the diffusion process is 1000, and the noise scheduling adopts a cosine decay strategy. For tilt domain data, the network input / output size is 512×64, the training epochs are 1500, the batch size is 16, and the learning rate is fixed at 1×10. -4 All input data is linearly normalized before being fed into the network. Interval.

[0017] Another object of the present invention is to provide a multi-domain seismic diffraction wavefield intelligent separation system, which is used to implement the multi-domain seismic diffraction wavefield intelligent separation method. The system includes: The data acquisition module is used to acquire seismic full-wave field data, which includes common offset domain data and / or dip domain data. The model building module is used to build a conditional denoising diffusion probability model. The seismic full wavefield data is used as the conditional input, and the conditional denoising diffusion probability model is extended into a dual-channel joint diffusion model. The dual-channel joint diffusion model includes a dual-output noise prediction network, which is used to predict the diffraction noise component and the reflection noise component, respectively. The training module is used to train the dual-channel joint diffusion model using a joint loss function, which simultaneously constrains the prediction error of diffraction noise and the prediction error of reflection noise. The generation module is used to independently sample the first initial noise and the second initial noise from the standard normal distribution, which are used as the initial states of the diffracted wave and the reflected wave, respectively. The first initial noise and the second initial noise are input into the trained dual-channel joint diffusion model, and the separated diffracted wave data and reflected wave data are generated iteratively through the reverse denoising process, using the earthquake full wave field data as a condition.

[0018] Another object of the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent separation method for multi-domain seismic diffraction wavefields.

[0019] Combining all the above technical solutions, the beneficial effects of this invention are as follows: First, this invention develops an intelligent diffraction wave separation technology system, establishes a complex feature training dataset covering multiple geological scenarios, multiple acquisition parameters, and multiple noise types, constructs a deep learning separation network with weak signal recognition capabilities, and forms a high-precision dual-domain collaborative processing method applicable to both the common offset domain and the dip domain. This invention significantly improves the model's ability to capture extremely weak diffraction signals through network architecture optimization design tailored to the temporal characteristics and multi-scale nature of seismic data, including an enhanced multi-scale feature extraction module, an improved conditional information fusion mechanism, and an adaptive noise scheduling strategy. This invention utilizes physical forward modeling to construct unsupervised training samples, eliminating reliance on manual annotation, and learns the intrinsic feature differences between diffracted and reflected waves through a data-driven approach, achieving weak dependence on the velocity model and robust separation in complex tectonic regions.

[0020] This invention addresses the long-standing core problem of diffraction wave energy loss under complex geological conditions. It effectively suppresses strong reflected wave interference while maintaining the integrity of weak diffraction waves, improving diffraction wave separation quality in complex structural zones, areas with difficult velocity modeling, and under extremely low signal-to-noise ratio conditions. The invention enhances the model's cross-regional generalization ability through domain adaptation technology and reduces computational costs using fast sampling algorithms and network optimization, making the method practical for industrial applications. Ultimately, this invention can provide high-quality input data for high-resolution imaging of key reservoir features such as small-scale geological anomalies, fault ends, fracture systems, cave boundaries, and channel pinch-outs, promoting the widespread application of diffraction wave separation technology in actual seismic exploration and production, including fine reservoir characterization, unconventional oil and gas exploration, and deep-ultra-deep resource evaluation.

[0021] Secondly, compared with existing technologies, the intelligent seismic diffraction wave separation method based on a dual-domain cooperative conditional diffusion model proposed in this invention has the following significant advantages: (1) The high fidelity of extremely weak diffraction signals effectively solves the contradiction between weak signal protection and strong reflection suppression. Compared with traditional plane wave decomposition (PWD) and conventional U-Net methods, this invention utilizes the powerful distribution learning capability of the denoising diffusion probability model (DDPM) to accurately identify and separate extremely weak diffraction waves with 1-2 orders of magnitude lower energy (such as diffraction below salt domes or at fault endpoints) from a strong reflection background. Experimental results show that this method completely eliminates strong reflection wave interference while avoiding the phase reversal, waveform destruction, and high-frequency information loss problems common in traditional methods, significantly improving the resolution of subsequent imaging.

[0022] (2) Strong generalization ability, overcoming excessive dependence on velocity models and artificial parameters. Unlike traditional Radon transform or plane wave filters, this method does not require a precise velocity model in advance, nor does it require repeated adjustment of parameters (such as dip angle range, cutoff wave number, etc.) for different work areas, realizing a highly automated processing flow. Test results (such as Pluto model, Sigsbee 2A model and actual data) prove that even with extremely complex geological structures (such as subsalt structures and drastic lateral velocity variations), the model can still maintain stable separation effect, showing excellent cross-work area generalization ability.

[0023] (3) It has a wide range of applications and realizes dual-domain collaborative processing of offset and dip domains. This invention constructs adapted training and inference processes for the common offset domain and dip domain respectively. This allows the method to be used for pre-stack processing to optimize velocity analysis and post-imaging processing to enhance structural interpretation, flexibly meeting the needs of different exploration stages.

[0024] (4) High fidelity of weak diffraction signals and strong suppression capability of reflected waves. This invention utilizes the stepwise denoising mechanism and distribution learning capability of the conditional denoising diffusion probability model (c-DDPM) to effectively identify and separate weak diffraction signals with significantly lower energy than reflected waves (such as diffraction waves below salt domes or at fault endpoints) in strong reflection backgrounds. Experimental results combining the three-layer model, Pluto model, and Sigsbee 2A model show that this method can effectively suppress strong reflected wave interference during separation while maintaining the phase consistency and waveform continuity of diffraction waves, avoiding the phase reversal and waveform distortion problems common in traditional plane wave decomposition (PWD) methods, and also improving the high-frequency information loss phenomenon in conventional U-Net methods. In addition, in weak signal regions (such as diffraction points below salt domes), the scheme of this invention can still maintain good diffraction wave energy integrity, enabling the separation results to exhibit higher resolution and clearer structural response in subsequent imaging processes. The results show that the scheme of this invention can effectively protect weak diffraction signals while achieving strong reflection suppression, significantly improving the diffraction wave separation quality.

[0025] (5) Strong generalization ability, reducing dependence on velocity models and artificial parameters. Compared with traditional Radon transform and plane wave filtering methods, this invention does not rely on accurate velocity models, nor does it require repeated adjustment of processing parameters (such as dip range, cutoff wave number, etc.) for different work areas, thus reducing the dependence on human experience. Test results on Pluto model, Sigsbee 2A model and actual seismic data show that under complex geological conditions (including salt dome structures, fault-developed areas and areas with significant lateral velocity variations), this method can still maintain stable diffraction wave separation effect without obvious energy leakage or misjudgment. At the same time, for diffraction types not included in the training data (such as deep internal diffraction points), the model can still effectively identify and separate them, demonstrating good generalization ability. The results show that the scheme of this invention has strong adaptability and stability, and can achieve reliable diffraction wave separation under different geological conditions.

[0026] (6) Wide applicability, enabling collaborative processing of offset and dip domains. This invention constructs a dual-domain collaborative processing framework, establishing corresponding training and inference processes in the common offset and dip domains respectively, to achieve joint utilization of multi-domain information. In the common offset domain processing, this method can effectively extract diffraction wave information, providing high-quality input data for velocity analysis; in the dip domain processing, it can enhance the response of diffraction waves in imaging, improving the ability to identify faults, salt dome boundaries, and small-scale discontinuities. Experimental results show that the diffraction wave imaging results obtained based on the scheme of this invention are superior to traditional methods in terms of structural clarity, detail characterization, and weak anomaly identification, and can more accurately reflect the characteristics of complex underground structures. The results show that the scheme of this invention can be applied to different stages such as pre-stack processing and post-imaging processing, and has good application flexibility and engineering practical value.

[0027] Third, this invention can be directly applied to commercial software and services for seismic data processing in oil and gas exploration. By providing high-quality diffraction wave separation results, it can significantly improve the identification accuracy of reservoir features such as faults, fractures, and caverns, reduce exploration risks, increase drilling success rates, and thus reduce ineffective exploration investment. This method has automated processing capabilities, which can significantly reduce the workload and technical threshold of professional interpreters, and has significant cost-saving and efficiency-enhancing value. In high-value exploration fields such as unconventional oil and gas (shale gas, tight oil) and deep-ultra-deep resource evaluation, this invention is expected to become a core processing module for fine reservoir characterization, with broad commercialization prospects and high economic conversion value.

[0028] Currently, there is no universal intelligent diffraction wave separation framework applicable to both the common offset domain and the dip domain. Existing methods either target only a single data domain, rely on accurate velocity models, or require a large number of manually labeled samples, all of which fail to achieve automated, high-precision separation across domains. This invention, for the first time, extends the conditional denoising diffusion probability model into a dual-channel joint modeling framework, simultaneously achieving the coordinated separation of diffraction and reflection waves within a single model. This fills the technical gap in the field of multi-domain seismic wavefield separation using deep generative models and provides a new technical path for subsequent research in this direction.

[0029] High-fidelity extraction of extremely weak diffraction waves under complex geological conditions has long been a core challenge in seismic exploration. Diffraction waves beneath salt domes and at fault endpoints often have energy levels one to two orders of magnitude lower than surrounding reflected waves. Traditional methods, while suppressing strong reflections, inevitably damage weak diffraction signals, leading to artifacts or missing information in subsequent imaging. This invention, through a stepwise denoising mechanism of a diffusion model and dual-channel collaborative constraints, achieves for the first time complete and faithful extraction of extremely weak diffraction signals under strong reflection backgrounds, effectively solving this long-standing technical problem in seismic exploration.

[0030] There has long been a technical bias in the industry that diffusion models are only applicable to single-objective generation tasks and struggle to handle multi-component wavefields with clear physical decomposition relationships. This invention, by constructing a dual-channel joint diffusion model, demonstrates that diffusion models can simultaneously model two types of wavefield components with physical coupling relationships within a unified framework, and that the synergistic constraint effect of joint modeling is significantly better than that of two independent models handling these components separately. Furthermore, while traditional views hold that deep learning methods must rely on a large amount of manually labeled real seismic data, this invention breaks this cognitive limitation by constructing training samples through physical forward modeling, verifying the effective generalization ability of unsupervised sample construction strategies on real, complex geological data. Attached Figure Description

[0031] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the disclosure of this invention and, together with the description, serve to explain the principles of the disclosure of this invention. Figure 1 This is a flowchart of the intelligent separation method for multi-domain seismic diffraction wavefields provided in an embodiment of the present invention; Figure 2 These are in-phase axis morphological feature diagrams of reflected waves and diffracted waves in the common offset domain provided by embodiments of the present invention; wherein, (a) is the in-phase axis diagram of reflection, and (b) is the in-phase axis diagram of diffraction; Figure 3 These are in-phase axis morphological feature diagrams of reflected and diffracted waves in the tilt domain provided by embodiments of the present invention; wherein, (a) is a schematic diagram of a diffraction point with lateral offset, (b) is a schematic diagram of a diffracted wave as a horizontal straight line, and (c) is a schematic diagram of a reflected wave when the velocity is inaccurate. Figure 1 (d) is a schematic diagram of the reflected wave when the velocity is inaccurate. Figure 2 The blue curve represents the reflection in-phase axis, and the orange curve represents the diffraction in-phase axis. Figure 4 This is a schematic diagram of the forward and backward processes of the denoising diffusion probability model provided in the embodiments of the present invention; Figure 5 This is a schematic diagram of the forward and backward processes of the conditional denoising diffusion probability model provided in the embodiments of the present invention; Figure 6 This is a diagram illustrating the training and sampling process of the conditional diffusion model provided in this embodiment of the invention; wherein, (a) is a schematic diagram of the diffusion process of the conditional diffusion model, and (b) is a schematic diagram of the sampling process of the conditional diffusion model. Figure 7 This is a diagram of the U-Net network architecture provided in an embodiment of the present invention. The inputs are noisy diffraction data, noisy reflection data, and seismic full wavefield data. The network outputs are the predicted diffraction noise and reflection noise. Figure 8These are schematic diagrams of four typical offset domain training data samples provided in this embodiment of the invention; wherein, (a)-(d) are full-wave field gather diagrams, and (e)-(h) are corresponding pure diffraction gather diagrams; each set of data adopts a segmented combination construction method: (a) and (e) are single-segment constructions, (b) and (f) are four-segment constructions, (c) and (g) are three-segment constructions, and (d) and (h) are two-segment constructions; Figure 9 This is a schematic diagram of tilt domain training data samples provided in an embodiment of the present invention; wherein, (a) is a full-wave field data diagram of earthquake, (b) is a diffraction wave data diagram, and (c) is a reflection wave data diagram; Figure 10 This is a comparison diagram of zero offset data of a three-layer model provided in an embodiment of the present invention; wherein, (a) is a full-wave field data diagram of the earthquake, (b) is a diffraction wave diagram separated based on the plane wave suppression method, and (c) is a diffraction wave diagram separated based on the c-DDPM method; Figure 11 This is a schematic diagram of the diffraction wave imaging results of the three-layer model provided in the embodiment of the present invention; wherein, (a) is the reverse time migration imaging result of the seismic full wavefield data, (b) is the imaging result of the diffraction wave separated based on the plane wave suppression method, and (c) is the imaging result of the diffraction wave separated based on the c-DDPM method. Figure 12 This is a schematic diagram of the zero-offset data separation effect of the Pluto model provided in the embodiment of the present invention; wherein, (a) is the full wavefield data diagram of the earthquake, (b) is the result diagram of the diffraction wave separation by the plane wave suppression method, (c) is the result diagram of the diffraction wave separation by U-Net, and (d) is the result diagram of the diffraction wave separation by c-DDPM. Figure 13 These are comparison images of Pluto model data imaging provided in the embodiments of the present invention; wherein, (a) is a reverse time migration imaging image of seismic full wavefield data, (b) is an imaging image of diffracted waves separated by plane wave suppression method, (c) is an imaging image of diffracted waves separated by U-Net, and (d) is an imaging image of diffracted waves separated by c-DDPM. Figure 14 These are test result diagrams provided in the embodiments of the present invention; wherein, (a) is a schematic diagram of the full-wavefield tilt common imaging gather of the Sigsbee 2A model, and (b) is a schematic diagram of the diffraction wave separation result of the Sigsbee 2A model; Figure 15 These are imaging schematic diagrams provided in the embodiments of the present invention; wherein, (a) is a schematic diagram of the Sigsbee 2A velocity model provided in the embodiments of the present invention, (b) is a schematic diagram of the full-wave field imaging result, (c) is a schematic diagram of the diffraction wave imaging obtained by separating using the c-DDPM method, and (d) is a schematic diagram of the diffraction wave imaging obtained by separating using the U-Net method; Figure 16These are schematic diagrams of full-wavelength imaging provided in the embodiments of the present invention; wherein, (a) is a full-wavelength imaging diagram of actual data provided in the embodiments of the present invention, (b) is a schematic diagram of diffraction separation results based on the c-DDPM method, (c) is a comparison diagram of single-channel seismic full-wavelength data and separation results, (d) is a schematic diagram of diffraction wave imaging separated by the c-DDPM method, and (e) is a schematic diagram of diffraction wave imaging separated by the U-Net method. Detailed Implementation

[0032] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0033] The innovation of this invention lies in its combination of a conditional diffusion generation model and a dual-channel joint modeling mechanism, which for the first time constructs a multi-domain intelligent separation framework for seismic diffraction waves that is applicable to both the common offset domain and the dip domain. Specific innovations are as follows: (1) A seismic wavefield separation architecture based on conditional denoising diffusion probability model (c-DDPM) is proposed: the diffusion model in generative models is introduced into the field of geophysical exploration, and a seismic full wavefield data separation architecture is designed. As a conditional input, guide the inverse denoising process. A novel network architecture. It breaks through the bottleneck of performance degradation of traditional discriminative models (such as CNN and SVM) at low signal-to-noise ratios, and achieves high-probability reconstruction of non-steady-state and extremely weak diffraction signals by utilizing the randomness and generative ability of the diffusion process.

[0034] (2) A dual-channel joint prediction and noise scheduling strategy is proposed: Unlike conventional single-output networks, this invention designs a dual-channel U-Net structure, which simultaneously accepts noisy diffraction data, noisy reflection data and full-wave field condition data as input, and simultaneously outputs diffraction noise. and reflected wave noise The predicted value. Through joint loss function constraints, the network implicitly learns the full wave field. The physical conservation relationship significantly reduces the signal leakage and background residue problems commonly encountered when extracting a single component.

[0035] (3) A U-Net network structure optimized for seismic wavefield characteristics was constructed: Based on the standard U-Net, a self-attention mechanism and a time embedding module were integrated. The time embedding module enables the network to perceive the noise level of the diffusion process, and the self-attention mechanism enhances the network's ability to capture global geometric features of seismic phase axes (such as the long-range correlation of hyperbolas), thereby enabling the differentiation of wavefield components with similar local morphology but different global features.

[0036] (4) Physics-Driven Data Synthesis and Domain Adaptive Training Method: A data construction process without real labels was invented. In the offset domain, a reflectivity convolution synthesis strategy with random parameters was adopted; in the dip domain, a strategy of finite difference forward modeling combined with Gaussian beam migration was adopted to simulate complex wavefield interferometry modes. The training set constructed by this method covers the characteristics of various geological anomalies such as faults, fractures, and pinch-outs, enabling the model to directly process complex real-world data without actual data training, thus solving the data hunger problem in the application of deep learning in the geophysical field.

[0037] Example 1, such as Figure 1 As shown, the intelligent separation method for multi-domain seismic diffraction wavefields provided in this embodiment of the invention includes the following steps: Acquire seismic full-wave field data, which includes common offset domain data and / or dip domain data, and the seismic full-wave field data is represented as a linear superposition of diffracted wave components and reflected wave components; A dual-channel joint conditional denoising and diffusion probability model is constructed, in which the seismic full-wavefield data is continuously used as the conditional guiding input at each time step in the reverse denoising process. The dual-channel joint conditional denoising and diffusion probability model is then extended into a dual-channel joint diffusion model, which includes a dual-output noise prediction network. The dual-output noise prediction network predicts the diffraction noise component and the reflection noise component, respectively, thereby achieving mutual constraint collaborative modeling of the two types of wavefields. The dual-channel joint diffusion model is trained using a joint loss function, which simultaneously constrains the prediction error of diffraction noise and the prediction error of reflection noise. The first initial noise and the second initial noise are independently sampled from the standard normal distribution and used as the initial states of the diffracted wave and the reflected wave, respectively. The first initial noise and the second initial noise are input into the trained dual-channel joint diffusion model, and the separated diffracted wave data and reflected wave data are generated iteratively through the reverse denoising process, using the earthquake full wave field data as a condition.

[0038] This invention has been validated in the following scenarios: In pre-stack common offset domain processing, the invention was applied to common offset gathers of a three-layer velocity model and a Pluto subsalt tectonic model, achieving accurate extraction of diffracted waves without residual reflected energy. This avoids the phase reversal and waveform destruction problems that occur in traditional plane wave suppression methods. The separated diffracted waves can be clearly focused on each diffraction point after reverse time migration imaging, effectively solving the problem of diffracted wave energy loss in complex tectonic areas. In post-imaging dip domain processing, the invention was applied to dip common imaging point gathers of the Sigsbee 2A model and actual seismic data. It can accurately separate diffracted waves, including those with extremely weak energy (the energy of diffracted points below the salt dome is two orders of magnitude lower than the surrounding area). The separation results, when superimposed along the dip direction, can clearly image the salt dome boundary, fault, and weak subsalt diffracted points. Compared with the U-Net method, it has higher imaging resolution and less noise, and is suitable for different stages such as pre-stack processing and post-imaging processing.

[0039] The steps described above are explained in detail below.

[0040] I. Specific implementation of step S1: data acquisition and feature analysis; 1. The characteristic differences between diffracted and reflected waves in different data domains; 1.1 Analysis of the characteristics of reflected and diffracted waves in the common offset domain; In co-offset seismic data, reflected waves and diffracted waves exhibit distinctly different morphologies and distribution characteristics. Reflected waves originate from continuous geological interfaces and appear as continuous, smooth, in-phase axes on seismic profiles. In co-offset trace collections, reflected waves exhibit a gently inclined, approximately straight-line shape, with the dip angle gradually increasing with the offset distance. Reflected waves have strong energy and good continuity, making them easy to track and identify throughout the entire profile. Diffracted waves, on the other hand, are generated from isolated underground scattering points, such as fault endpoints, fracture edges, and karst cave boundaries—locally discontinuous geological bodies. In co-offset trace collections, diffracted waves appear as incomplete hyperbolic segments, with key characteristics including: (1) poor spatial continuity, appearing only in local areas and difficult to track laterally; (2) weak energy, typically only a few percent of the energy of adjacent reflected waves, easily submerged by strong reflection backgrounds; (3) relatively high frequency components, but rapid attenuation; and (4) a hyperbolic shape with significantly increased curvature at far offsets, showing dramatic changes in apparent dip angle.

[0041] Figure 2 It demonstrates the typical in-phase axis morphological characteristics of reflected and diffracted waves in the common offset domain. Figure 2 Figure (a) shows the reflection phase axis, which represents a series of events in various forms, reflecting the underground geological interfaces with different occurrences; Figure 2Figure (b) shows the diffraction phase axis, which exhibits an approximately downward-curving hyperbola shape, with the vertex corresponding to the location of the underground diffraction point. Due to the spatial discontinuity, weak energy, and frequent overlap of diffracted waves with reflected waves, directly separating the two in the co-offset domain presents a significant challenge.

[0042] 1.2 Analysis of Reflected and Diffracted Wave Characteristics in the Dip-Angle Domain In tilt-coordinated common-image point gathers (DACIGs), reflected waves exhibit a hyperbolic shape, while diffracted waves show a weaker linear in-phase axis. This morphological difference can be used to achieve wavefield separation. The shapes of diffracted and reflected waves in the tilt domain can be described analytically by the following expressions: ; ; In the formula, The dip angle of the strata. For the imaging angle, For offset velocity, This refers to the actual underground speed. , ( The horizontal position after imaging. (The lateral position of the diffraction point). For the common imaging point collection at the tilt angle, the corresponding imaging angle is the imaging angle. Imaging depth below The reference imaging depth at zero tilt angle. offset speed Compared with actual underground speed The ratio, Lateral position after imaging Lateral position relative to the diffraction point The lateral offset between the two points is used to characterize the horizontal deviation of the imaging position relative to the diffraction point position.

[0043] Equations (1) and (2) originate from the kinematic relationship between reflected and diffracted waves in angle-domain imaging theory. These formulas theoretically characterize the imaging depth of reflected and diffracted waves in diagonal domain common imaging point gathers (DACIGs). With imaging angle The relationship between these two types of waves reveals the differences in their geometric response characteristics in the dip domain: reflected waves typically exhibit a nonlinear curve shape with a stationary point, while diffracted waves present as linear or approximately linear phase axes without a stationary point. This invention utilizes the geometric characteristic differences revealed by the above formula as one of the theoretical bases for intelligent separation of multi-domain seismic wavefields, using it to guide the construction of dip domain features and the design of discrimination rules, thereby achieving effective differentiation between reflected and diffracted waves.

[0044] In DACIGs, diffraction and reflection events exhibit different geometric forms, such as Figure 3 As shown. When the offset velocity is accurate and the diffraction point is directly above the gather, the diffracted wave is a horizontal straight line, as shown. Figure 3 As shown in Figure (b), the reflected wave is a "smile" shaped curve with a stationary point. If there is a lateral shift in the diffraction point, the diffracted wave becomes a straight line without a stationary point, as shown in Figure (b). Figure 3 As shown in Figure (a). Figure 3 Figure (c) in the middle and Figure 3 Figure (d) illustrates the situation when velocity is inaccurate: the reflected wave retains its upward-convex "smile" shape, while the diffracted wave is distorted and loses its linear characteristics. The slope of the curve is controlled by ρ, and the depth position is determined by z0. Therefore, diffracted and reflected waves can be distinguished in DACIGs by identifying the differences in their shapes.

[0045] II. Specific implementation of step S2: Model construction; 2. Basic framework of the Denoising Diffusion Probability Model (DDPM); Denoising Diffusion Probabilistic Model (DDPM) is a type of generative model, such as... Figure 4 As shown, its core idea is to destroy the data structure through a forward diffusion process that gradually adds noise, and then recover the data through a learned reverse denoising process. This model has achieved significant success in fields such as image generation and speech synthesis, and this invention innovatively applies it to the task of seismic diffraction wave separation.

[0046] The forward diffusion process is defined as a Markov chain, starting from the original data. Begin, at each time step Gradually add Gaussian noise until... Post-step data It becomes pure noise with a standard normal distribution. The single-step diffusion process is represented as: ; in, Indicates a normal distribution. These are noise scheduling parameters that control the amount of noise added at each step. It is the identity matrix. (Definition) and cumulative parameters Then you can start from Direct sampling obtains the value at any time. No need for step-by-step iteration: ; The goal of the reverse denoising process is to learn the inverse trajectory of the forward process, from pure noise... Gradually recover the original data The reverse process is represented as a Markov chain: ; By predicting the noise added during the forward process To indirectly predict the mean. Define a neural network. For predicting noise, the inverse mean is expressed as: ; The loss function for model training is defined as: ; The sampling generation process starts from standard normal noise. Begin by iteratively applying the reverse denoising steps: ; Equations (3) to (8) are the standard mathematical expressions of the diffusion probability model. Equations (3) and (4) describe the forward diffusion process, that is, by gradually adding Gaussian noise to the original data, the seismic wavefield data is mapped to random variables that follow a standard normal distribution. Equations (5) and (6) describe the reverse denoising process, which gradually recovers the original data by constructing a parameterized model. Equation (7) is the loss function for model training, which is used to constrain the network's ability to predict noise. Equation (8) is the sampling process, which is used to generate target data. This invention utilizes the stepwise denoising capability of the above diffusion model and introduces it into the seismic wavefield separation task to achieve the gradual recovery of diffraction wave signals from complex full-wavefield seismic data. Furthermore, conditional constraints are introduced on the basis of this model, using full-wavefield seismic data as conditional input to guide the model to generate the corresponding diffraction wave components. Unlike traditional diffusion models, this invention combines the multi-domain characteristics of seismic data (including offset domain and dip domain) and the differences in geometric characteristics between reflected waves and diffracted waves to improve the model input and constraint mechanism, thereby improving the accuracy and stability of diffraction wave separation.

[0047] The innovation of this invention lies not in the diffusion model itself, but in its introduction into the multi-domain seismic wave field separation problem, and its improvement by combining multi-domain characteristics such as the dip domain and physical constraints, thereby achieving high-precision separation of diffracted waves.

[0048] Specifically, in the conditional diffusion model, the full-wavefield seismic data y is first used as a conditional input, thus expanding the noise prediction function from its original form to: ; This allows for the introduction of full-wavefield information as a constraint during the reverse denoising process, enabling conditional generation of the diffracted wave components. Furthermore, within the dual-channel modeling framework, the seismic full-wavefield data is represented as a superposition of diffracted and reflected waves: ; And construct a dual-output noise prediction model: ; The noise components of diffracted and reflected waves are predicted separately, thus enabling co-modeling of the two types of wave fields. Furthermore, by combining dip and offset domain data as model inputs, the model can simultaneously utilize multi-domain feature information during training to learn the response differences between reflected and diffracted waves in different domains, thereby improving separation accuracy.

[0049] Through the above improvements, the diffusion model is extended from a traditional univariate generation model to a conditional generation model oriented towards multi-component wave field decomposition, thereby achieving effective separation of diffracted and reflected waves.

[0050] 3. Conditional Diffusion Model (DDPM); To achieve the diffraction wave separation task, it is necessary to generate corresponding diffraction wave data given the full seismic wavefield data y. This necessitates extending DDPM to a conditional generation model. Conditional Diffusion Model (c-DDPM) achieves this by introducing conditional information into the denoising network, such as... Figure 5 As shown.

[0051] In the conditional diffusion model, the inverse denoising process is conditional at each step. For guidance: ; The noise prediction network was modified accordingly. Simultaneously input noisy diffraction wave data Full-field conditions and time step The training loss function becomes: ; 4. Dual-channel joint diffusion model; Traditional diffusion models typically model a single data distribution, outputting a single variable. They are primarily used for image generation or single-target reconstruction tasks, but do not address the problem of co-generating multiple components with clearly defined physical decomposition relationships. In seismic wavefield separation tasks, diffracted and reflected waves are highly coupled in the data, and they differ significantly in amplitude, geometry, and energy distribution. How to simultaneously model both types of wavefields within a unified model framework while ensuring the physical consistency of the separation results is not straightforward. To further improve separation accuracy, this invention extends the conditional diffusion model into a dual-channel joint model, enabling the model to simultaneously predict both diffracted and reflected wave components, thereby explicitly characterizing the coupling relationship between the two types of wavefields. Compared to traditional single-channel models, this method effectively reduces energy leakage and aliasing during wavefield separation.

[0052] Specifically, seismic full-wavefield data can be decomposed into two components: diffracted waves and reflected waves. The dual-channel model predicts two outputs simultaneously: diffracted wave. and reflected waves The joint loss function is defined as: ; After training, c-DDPM learns a conditional distribution. By sampling from the learned distribution, this invention can generate data that is consistent with the conditional input (i.e., seismic full-wavefield data). The corresponding new separated diffracted and reflected waves. First from the standard normal distribution Medium sampling And use the trained network to predict Subsequently, according to equations (12) and (13), the separated diffracted waves are generated from the full seismic wavefield data. and reflected waves .

[0053] ; ; Equations (9) and (10) are the standard expressions of the conditional diffusion probability model, belonging to the basic methods in existing generative models. Among them, conditional information is introduced during the reverse denoising process. (Seismic full-wavefield data) enables conditional modeling of the target data distribution. Based on this, this invention improves upon the traditional conditional diffusion model by proposing a dual-channel joint diffusion model to simultaneously generate diffracted and reflected wave components. Specifically, the seismic full-wavefield data is represented as a superposition of diffracted and reflected waves. A dual-output noise prediction network was constructed to predict the diffraction wave noise components separately. and reflected wave noise components .

[0054] For this dual-channel structure, a joint loss function (Equation (11)) is constructed to simultaneously constrain the noise prediction errors of the two types of wave fields, thereby achieving collaborative modeling of diffracted and reflected waves. Compared with the traditional single-channel diffusion model, this method can effectively characterize the coupling relationship between the two types of wave fields and reduce energy leakage and aliasing during the wave field separation process.

[0055] During the sampling phase, based on the proposed dual-channel diffusion model, reverse denoising processes (Equations (12) and (13)) are performed on the diffracted and reflected waves respectively to achieve joint generation of the two types of wave fields. This process improves the accuracy and stability of the separation results by introducing cross-correlation condition information to ensure that the generated diffracted and reflected waves are physically consistent.

[0056] The innovation of this invention lies in introducing a dual-channel joint modeling mechanism on the basis of the model. By simultaneously predicting the diffracted wave and reflected wave components and constructing a joint loss function and the corresponding dual-channel sampling process, the co-generation and effective separation of the two types of wave fields are realized, thereby improving the separation accuracy and reducing wave field aliasing.

[0057] The training and sampling process of the conditional diffusion model is as follows: Figure 6 As shown, during the training phase, a pair of training samples is randomly selected in each iteration cycle. and and from the time interval Uniform sampling within one time step Subsequently, random noise is sampled from the standard Gaussian distribution, and the input of noisy data is constructed by calculation using formula (4). This noisy data, along with the seismic full-wave field data used as a condition, is fed into a denoising network (the network used here is the U-Net network) for noise prediction, and the prediction error is evaluated by calculating the loss function formula (11) between the predicted noise and the actual added noise. and The gradient is calculated based on this loss, and the network parameters are updated accordingly to complete one training iteration.

[0058] During the data generation phase, the sampling process begins at time step T, using a pair of random noise. and As the initial state, noise is denoised iteratively through formulas (12) and (13), generating intermediate states with decreasing time steps in turn until a clear target image is finally generated.

[0059] 5. Network Architecture; This invention designs a conditional denoising network based on the U-Net architecture as the core of the diffusion model, and its structure is as follows: Figure 7 As shown, the network consists of three parts: an encoder, intermediate modules, and a decoder. The encoder contains five residual modules, each with two 3×3 convolutional layers, followed by normalization and the application of the Swish activation function to extract multi-scale features. The network incorporates a self-attention mechanism to effectively capture global dependencies within the feature maps. Each upsampling layer in the decoder is connected to the corresponding layer in the encoder via skip connections to preserve the details of the original data. Furthermore, the network embeds a temporal embedding module, which integrates the current backdivergence time step information into the network, helping to improve the accuracy of noise prediction. The network input has three channels: diffraction wave data... Reflected wave data and full-field condition data The output is the predicted noise. and The input data is a tilt-angle common imaging gather of size 512×64. The model's diffusion and sampling processes are as follows: Figure 6 Figure (a) and Figure 6 As shown in Figure (b), during the training process, the network learns the mapping relationship between the input and noise by combining seismic full-wave field data, reflection data and diffraction data, and compares the predicted noise with the actual added noise to update the network parameters accordingly.

[0060] The innovation of this invention lies in improving the network input and output form by combining the characteristics of seismic wavefield separation tasks, constructing a multi-channel input and dual-channel output structure, and combining it with a dual-channel diffusion model to achieve collaborative modeling and separation of diffracted and reflected waves, thereby improving the separation effect.

[0061] III. Specific implementation of step S3: Model training; 6. Training Dataset Preparation; After determining the network architecture, a training dataset needs to be prepared. The methods for constructing training samples in the offset domain and tilt domain are described below.

[0062] 6.1 Preparation of Training Dataset in Offset Domain; In the offset domain, the phase axes of reflected waves exhibit different morphologies of potential geological interfaces, while diffracted waves exhibit an approximately downward-curving hyperbolic shape. Based on the differences in waveform characteristics between diffraction and reflection, this invention adopts a data construction strategy based on convolution synthesis: First, a reflectivity sequence is constructed, the shape of which is controlled by two sinusoidal trends, and the number and position of reflectivity are randomly selected within a certain range. Then, the reflectivity model is convolved with the Ricker wavelet, where the same wavelet is used for each reflection phase axis, and the wavelet amplitude and dominant frequency vary randomly for different phase axes. The generation of diffracted wave gathers adopts a similar two-step process: First, hyperbolic reflectivity with random numbers, positions, and opening angles is generated, and the hyperbola vertex is located above a certain reflection interface. Then, each reflectivity is convolved with the Ricker wavelet with random amplitude and dominant frequency, and diffracted waves along the same reflection phase axis use the same wavelet parameters. To simulate the dynamic characteristics of diffracted waves, the diffraction energy is set to attenuate as the distance to its vertex increases. Furthermore, a segmented combination strategy is employed to augment the generated pure reflection and pure diffraction trajectory sets. The training dataset is as follows: Figure 8 As shown.

[0063] 6.2 Preparation of Pitch Domain Training Dataset; For the dip co-imaging point gathers, a training dataset was constructed using a method based on finite difference forward modeling and Gaussian beam migration. The specific process was as follows: forward modeling and migration imaging were performed on the velocity model containing only continuous reflection interfaces and the same model with added random diffraction points, respectively, resulting in two sets of dip gathers: one set of pure reflection wave gathers and the other set of full-wavefield gathers containing both reflection and diffraction. The diffraction wave gathers were separated from the full-wavefield using a convolutional neural network, thus forming a complete training sample pair containing three components: seismic full-wavefield data, corresponding pure diffraction data, and pure reflection data. The training dataset is as follows: Figure 9 As shown.

[0064] 7. Network Training; After setting the network architecture and preparing the training dataset, the network can be trained. For the two different input data formats—offset domain and tilt domain—this invention constructs corresponding c-DDPMs for training. The overall architecture of the two networks is consistent, with their core consisting of a U-Net for noise prediction. This network adopts a typical encoder-decoder design. The encoding part extracts multi-scale features step-by-step through five residual modules, while the decoding part fuses shallow detail information through skip connections. The network also introduces a temporal embedding module and a self-attention mechanism. The former incorporates the current time step information into the computation, while the latter helps capture global dependencies in the feature map, thereby improving the accuracy of noise prediction.

[0065] During training, the model uses the full-wave field gathers generated by the aforementioned method as conditional input and the corresponding pure reflected wave or pure diffracted wave gathers as target output, learning the conditional distribution for gradually recovering clean signals from noisy target data. For the offset domain, the network input / output size is fixed at 128×128, the model is trained for 277 epochs, the batch size is 4, the Adam optimizer is used, and the initial learning rate is set to 0.0001. The total number of diffusion steps is 1000, and a cosine decay strategy is used for noise scheduling. The input / output size of the tilt domain model is 512×64, the number of training epochs is increased to 1500, the batch size is 16, and the learning rate is fixed at 1×10. -4 All input data is linearly normalized to the [-1,1] interval before being fed into the network to ensure the stability of the training process.

[0066] To systematically evaluate the effectiveness of the proposed c-DDPM method, a traditional U-Net model was simultaneously trained on the same training dataset and with identical hyperparameter settings for comparison. The U-Net network structure is identical to the noise prediction network used in c-DDPM, but it directly takes full-field seismic data as input and outputs separated diffracted wave data. This comparative experiment aims to directly compare the learning ability and separation performance of the conditional diffusion model and the classical supervised learning method under the same data distribution.

[0067] Example 2: The multi-domain seismic diffraction wavefield intelligent separation system provided in this embodiment of the invention includes: The data acquisition module is used to acquire seismic full-wave field data, which includes common offset domain data and / or dip domain data. The model building module is used to build a conditional denoising diffusion probability model. The seismic full wavefield data is used as the conditional input, and the conditional denoising diffusion probability model is extended into a dual-channel joint diffusion model. The dual-channel joint diffusion model includes a dual-output noise prediction network, which is used to predict the diffraction noise component and the reflection noise component, respectively. The training module is used to train the dual-channel joint diffusion model using a joint loss function, which simultaneously constrains the prediction error of diffraction noise and the prediction error of reflection noise. The generation module is used to independently sample the first initial noise and the second initial noise from the standard normal distribution, which are used as the initial states of the diffracted wave and the reflected wave, respectively. The first initial noise and the second initial noise are input into the trained dual-channel joint diffusion model, and the separated diffracted wave data and reflected wave data are generated iteratively through the reverse denoising process, using the earthquake full wave field data as a condition.

[0068] To further demonstrate the positive effects of the above embodiments, the present invention conducted the following experiments based on the above technical solutions. The experiments covered four sets of tests: synthetic data (three-layer velocity model), complex geological models (Pluto subsalt tectonic model, Sigsbee 2A model), and actual seismic data. The effectiveness, generalization ability, and engineering applicability of the method of the present invention were comprehensively verified from two dimensions: the common offset domain and the dip domain.

[0069] 1. Common-offset domain diffraction wave separation; First, tests were conducted on a three-layer velocity model, which includes three media layers, a fault, and eight diffraction points. Three diffraction points are located at the reflecting interface, and five are distributed in the third layer. An observation system with 149 shots and 201 receivers per shot was used, with a Ricker wavelet at a dominant frequency of 25 Hz. Common offset gathers were obtained using the finite difference method. Figure 10 Figure (a) shows the zero-offset seismic full-wavefield data, which was cut into 128×128 data blocks with 75% overlap and input into the c-DDPM model. The separation results are as follows. Figure 10 As shown in Figure (b), the diffracted wave is accurately extracted with no residual reflected energy. Figure 10As shown in Figure (c), compared to the traditional plane wave suppression method, the latter exhibits phase reversal and waveform disruption at the diffraction vertices (red and yellow arrows). It is worth noting that the training data only contains interface diffraction points, but the model successfully separated the internal diffraction points in the third layer that were not included in the training, demonstrating good generalization ability.

[0070] The separated diffracted waves were subjected to reverse time migration imaging. The diffracted wave imaging results of the three-layer model are as follows: Figure 11 As shown. The results of the reverse-time migration imaging of the full-wavefield seismic data are as follows. Figure 11 As shown in Figure (a), in full-wave imaging, diffraction points and faults are obscured by the reflecting interface, while diffraction wave imaging based on c-DDPM separation can clearly focus on each diffraction point and has high resolution, such as... Figure 11 As shown in Figure (b), imaging based on the plane wave suppression method exhibits obvious phase reversal and artifacts caused by tomography (red arrows), such as... Figure 11 As shown in Figure (c).

[0071] The applicability of the method under actual complex geological conditions was further verified using the Pluto subsalt tectonic model. Its zero-offset full-wavefield seismic data is as follows: Figure 12 Figure (a) shows various types of diffracted waves generated by faults (white arrows), salt domes (yellow arrows), and interlayer discontinuities (red arrows). Figure 12 The c-DDPM separation results shown in Figure (d) demonstrate that low-amplitude diffracted waves can be effectively extracted from the strongly reflective background with good waveform fidelity. For comparison, as shown in Figure (d), Figure 12 As shown in (b), the separation results of the traditional plane wave suppression method exhibit obvious phase reversal and spurious frequency noise, and some steeply tilted reflections are misidentified as diffracted waves (blue arrows). Figure 12 As shown in Figure (c), although the U-Net separation results are generally comparable to c-DDPM, there is a lack of diffraction wave energy in local regions (red boxes). The imaging comparison results of the Pluto model data are as follows... Figure 13 As shown. Figure 13 As shown in Figure (a), when the separated data are subjected to reverse time migration imaging, both c-DDPM and U-Net can clearly highlight the fault structure (white arrows), while... Figure 13 The plane wave suppression rule shown in Figure (b) suffers from severe image quality degradation due to vertex suppression and phase issues. (Similar to...) Figure 13 Compared to the traditional U-Net method shown in Figure (c), the model used in this invention introduces a dual-channel joint modeling mechanism for diffracted and reflected waves in its structure. The c-DDPM separated diffracted wave imaging results are shown below. Figure 13As shown in Figure (d), experimental results show that U-Net suffers from energy loss in diffraction waves in local regions, while the solution of this invention can completely recover weak diffraction signals. This indicates that simultaneously modeling both diffraction and reflection wave components helps reduce energy aliasing between the two wave fields, thereby improving separation accuracy.

[0072] 2. Diffraction wave separation in the tilt region; like Figure 15 As shown in Figure (a), the Sigsbee 2A model was used for testing. This model includes high-speed salt domes and surrounding small-scale faults and scattering points, and can excite multiple types of diffracted waves. Tilt gathers were generated through finite-difference forward modeling and Gaussian beam migration. All gathers were then stacked along the tilt direction to obtain a full-field image, as shown in Figure (a). Figure 15 As shown in Figure (b), the diffraction point indicated by the yellow arrow below the salt dome has extremely weak energy and is almost completely obscured by the reflective layer. Before testing, the data was preprocessed with bandpass filtering and automatic gain control to match the spectrum of the training data and enhance the weak diffraction signal.

[0073] Input all full-wave field channels into the trained network sequentially for further testing, and extract the test results of one channel, such as... Figure 14 As shown. Figure 14 Figure (a) shows the full-wavefield imaging of this channel. The red arrows indicate the diffracted waves with higher energy and more distinct characteristics, while the yellow arrows indicate the diffracted waves with weaker energy. Figure 14 Figure (b) shows the diffraction wave gather separated by the c-DDPM method. It can be seen that the method used can accurately separate all diffracted waves, including those with low energy, while preserving kinematic characteristics and minimizing reflection residue. The diffraction separation results of all channels are superimposed along the tilt direction to obtain a diffraction stacking image. Figure 15 Figure (c) shows the imaging results based on offset c-DDPM, clearly revealing the salt dome boundary indicated by the red arrow, the left fault indicated by the blue arrow, and the weak diffraction point under the salt indicated by the yellow arrow. It is worth mentioning that due to the strong shielding effect of the salt dome, the energy of the three diffraction points under the salt dome is two orders of magnitude weaker than the surrounding diffraction energy. However, based on the method used in this study, these extremely weak diffraction waves can still be separated at high resolution. Figure 15 Figure (d) shows the separation results based on the U-Net network. It can be seen that the imaging resolution is low, the noise is strong, the separation performance is significantly poor, and it cannot even separate slightly weak diffraction waves.

[0074] The same trained network was then applied to real-world data, which was then used for full-wavelength imaging, as shown in the example. Figure 16 As shown in Figure (a), diffraction is only visible to the naked eye, as indicated by the red arrow. Actual data is input into the network channel by channel for separation, and one channel is extracted, such as... Figure 16Figure (b) shows full-wavelength imaging, as shown in the image. Figure 16 Figure (c) shows the diffraction separation results based on the c-DDPM method. It can be seen that even with the reflected wave suppressed, the diffracted wave can be clearly and accurately separated. Furthermore, all separation results are superimposed along the tilt angle direction to obtain a diffraction superposition image. Figure 16 Figure (d) shows the imaging results based on offset c-DDPM, all of which can clearly image hidden weak discontinuities. Figure 16 Figure (e) shows diffraction stacking imaging based on the U-Net method, which has lower imaging resolution and more noise. The results show that the c-DDPM-based method of this invention can better characterize subsurface complexity, enabling the network to effectively learn diffraction features. Its ability to suppress noise and extract weak signals is significantly better than the U-Net method, providing a reliable tool for high-resolution imaging of complex areas such as salt domes and faults.

[0075] Based on the combined results of the above-mentioned multiple models and actual data experiments, it can be seen that the scheme of this invention outperforms traditional methods and single-channel models in terms of diffraction wave separation accuracy, weak signal fidelity, and adaptability to complex structures. The dual-channel joint modeling mechanism introduced in this invention can simultaneously characterize the coupling relationship between diffracted and reflected waves, effectively reducing wave field aliasing, thereby verifying the effectiveness of the structural design.

[0076] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention and within the spirit and principles of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A multi-domain seismic diffracted wave field intelligent separation method, characterized in that, The method includes the following steps: Acquire seismic full-wave field data, which includes common offset domain data and / or dip domain data, and the seismic full-wave field data is represented as a linear superposition of diffracted wave components and reflected wave components; A dual-channel joint conditional denoising and diffusion probability model is constructed, in which the seismic full-wavefield data is continuously used as the conditional guiding input at each time step in the reverse denoising process. The dual-channel joint conditional denoising and diffusion probability model is then extended into a dual-channel joint diffusion model, which includes a dual-output noise prediction network. The dual-output noise prediction network predicts the diffraction noise component and the reflection noise component, respectively, thereby achieving mutual constraint collaborative modeling of the two types of wavefields. The dual-channel joint diffusion model is trained using a joint loss function, which simultaneously constrains the prediction error of diffraction noise and the prediction error of reflection noise. The first initial noise and the second initial noise are independently sampled from the standard normal distribution and used as the initial states of the diffracted wave and the reflected wave, respectively. The first initial noise and the second initial noise are input into the trained dual-channel joint diffusion model, and the separated diffracted wave data and reflected wave data are generated iteratively through the reverse denoising process, using the earthquake full wave field data as a condition.

2. The intelligent separation method for multi-domain seismic diffraction wavefields according to claim 1, characterized in that, The conditional denoising diffusion probability model is a conditional diffusion model, and the reverse denoising process is performed on the seismic full wave field conditional data at each step is guided by the condition, which is represented as: ； In the formula, For a given first Step-by-step noisy diffraction wave field samples and seismic full-wavefield condition data Under the given conditions, the first denoising diffusion probability model predicts the th... The conditional probability distribution of the step sample. It follows a Gaussian distribution. For the parameter is The conditional Gaussian distribution mean predicted by the neural network. For the parameter is The conditional Gaussian distribution covariance predicted by the neural network. For the first Step-by-step noisy diffraction data, For the first Step-by-step noisy diffraction data, This is seismic full-wavefield condition data. For time steps, These are the learnable parameters for the noise prediction network; Input with noise Step-noisy diffraction wave data Seismic full-wavelength condition data and time step Seismic full-wave field condition data It remains unchanged throughout the reverse process and is consistent with noisy diffraction wave data. Having the same spatial dimensions; the training loss function is expanded to: ； In the formula, For the conditional denoising diffusion probability model in the th The training loss function corresponding to the reverse denoising process. For the first Step-noisy diffraction wave data diffusion time step Seismic full-wavelength condition data and follows a standard normal distribution random noise The desired mathematical expectation, To expand the noise prediction network, To obtain from the standard normal distribution The actual noise in the sampled data.

3. The intelligent separation method for multi-domain seismic diffraction wavefields according to claim 2, characterized in that, The dual-channel joint diffusion model achieves multi-domain coordinated separation of diffracted and reflected waves in the following manner: Seismic full-wave field condition data As a conditional input, the noise prediction function is expanded to In the reverse denoising process, full-wave field information is introduced as a constraint to achieve conditional generation of diffracted wave components. A dual-channel modeling framework is constructed to represent seismic full-wavefield data as a superposition of diffracted and reflected waves. Simultaneously predict the output diffraction wave noise prediction component. and reflected wave noise prediction components ;in, For diffraction wave data, This is the data for reflected waves; During training, common offset domain data and / or tilt domain data are used as training samples to learn the response differences between reflected and diffracted waves in different domains. For common offset domain samples, the model learns the difference between the hyperbolic morphological characteristics of diffracted waves and the straight in-phase axis of reflected waves. For tilt domain samples, the model learns the difference between the arc characteristics of diffracted waves and the straight in-phase axis of reflected waves.

4. The intelligent separation method for multi-domain seismic diffraction wavefields according to claim 2, characterized in that, The joint loss function is: ； In the formula, To add the diffracted wave component during the forward diffusion process True Gaussian noise, For a dual-channel conditional denoising network, at the input of the first... Step-noisy diffraction wave data , No. Step-noise reflected wave data Seismic full-wavelength condition data and time step Then, the predicted diffraction wave noise component, To add the reflected wave component during the forward diffusion process True Gaussian noise, The noise component of the reflected wave predicted by the dual-channel conditional denoising network; In the reverse denoising process, from the standard normal distribution Independent sampling of initial noisy state and random disturbance noise Predicting diffraction wave noise components using a trained dual-channel joint diffusion model. and reflected wave noise components Then, following the iterative formula, the separated diffracted waves are gradually generated from the full seismic wavefield data. and reflected waves The expression is: ； ； In the formula, and The first The generated diffraction wave data and reflected wave data, For the first Step noise scheduling coefficient, and The first Step-by-step noisy diffraction wave data and noisy reflection wave data, and These are the diffraction noise component and the reflected noise component predicted by the dual-channel conditional denoising network, respectively. This is seismic full-wavefield condition data. and These are random Gaussian noises injected into the diffracted wave channel and the reflected wave channel, respectively, both of which follow a standard Gaussian distribution. , It is the identity matrix. represents the standard deviation of the random perturbation in the reverse denoising step; where, , For the forward diffusion process Noise variance scheduling parameters for each step; cumulative parameters .

5. The intelligent separation method for multi-domain seismic diffraction wavefields according to claim 4, characterized in that, The training process of the dual-channel joint diffusion model includes: Randomly select seismic full-wavefield data in each iteration cycle Pure diffraction wave data and pure reflected wave data and from the time interval Uniform sampling within one time step Random noise is sampled from a standard Gaussian distribution, and noisy data is constructed through a forward diffusion process. The noisy data and the seismic full-wave field data as conditions are fed into the dual-output noise prediction network for noise prediction. The joint loss function is calculated to evaluate the error between the predicted noise and the actual added noise. The gradient is calculated based on the loss and the network parameters are updated. The sampling process of the dual-channel joint diffusion model includes: starting from time step T, using a pair of random noise... and As the initial state, the noise is denoised step by step according to the iterative formula, and intermediate states with decreasing time steps are generated in sequence until the separated diffraction wave data and reflection wave data are finally generated.

6. The intelligent separation method for multi-domain seismic diffraction wavefields according to claim 1, characterized in that, The dual-channel joint diffusion model is based on the U-Net architecture and includes: The hyperparameters of the number of residual modules and the size of the convolutional kernels in the encoder are flexibly adjusted according to the size of the input data. An optimal setting of 5 residual modules is adopted. Each residual module has two 3×3 convolutional layers, followed by a normalization layer and a Swish activation function to extract multi-scale features. The decoder connects to the corresponding layer of the encoder via skip connections to fuse shallow detail information; The self-attention mechanism module is used to capture global dependencies in the feature map; The time embedding module is used to incorporate the time step information of the current backpropagation process into the network; The noise prediction network takes three input channels: noisy diffraction wave data, noisy reflected wave data, and full-field condition data, and outputs the predicted diffraction wave noise and reflected wave noise.

7. The intelligent separation method for multi-domain seismic diffraction wavefields according to claim 1, characterized in that, The offset domain training dataset for the dual-channel joint diffusion model is constructed in the following manner: A reflectivity sequence is constructed, the shape of which is controlled by two sinusoidal trends, and the number and position of reflectivity are randomly selected. The reflectivity model is convolved with the Ricker wavelet, and the same wavelet is used for each reflection phase axis. The amplitude and dominant frequency of the wavelet of different phase axes vary randomly. The diffraction wave gather is generated by generating hyperbolic reflectivity with random number, position and opening angle, and the apex of the hyperbola is located above the reflection interface. Each reflectivity is convolved with the Ricker wavelet of random amplitude and dominant frequency, and the same wavelet parameters are used for diffracted waves along the same reflection phase axis; the diffraction energy is set to attenuate as the vertex distance increases; a segmented combination strategy is adopted to augment the generated pure reflection and pure diffraction gathers. The segmented combination strategy is as follows: different numbers of pure reflection wave segments and pure diffraction wave segments are randomly truncated and spliced ​​sequentially to form a complete co-offset gather, thereby expanding the diversity and coverage of the training data based on a limited sample. The tilt domain training dataset for the dual-channel joint diffusion model is constructed in the following manner: Finite difference forward modeling and Gaussian beam migration imaging were performed on the velocity model containing only continuous reflection interfaces and the same model with added random diffraction points, respectively, to obtain pure reflection wave gathers and full wavefield gathers containing reflection and diffraction. The diffraction wave gathers were separated from the full wavefield through a convolutional neural network to form training sample pairs containing seismic full wavefield data, corresponding pure diffraction data, and pure reflection data.

8. The intelligent separation method for multi-domain seismic diffraction wavefields according to claim 7, characterized in that, The training process of the dual-channel joint diffusion model includes: Using the full-wave field gather as the conditional input and the corresponding pure reflected wave or pure diffraction wave gather as the target output, the conditional distribution for gradually recovering a clean signal from noisy target data is learned. For offset domain data, the network input / output size was fixed at 128×128, the model was trained for 277 epochs, the batch size was 4, the Adam optimizer was used, and the initial learning rate was set to 0.0001; the total number of diffusion steps was 1000, and a cosine decay strategy was used for noise scheduling; for tilt domain data, the network input / output size was 512×64, the training epochs were 1500, the batch size was 16, and the learning rate was fixed at 1×10. -4 All input data is linearly normalized before being fed into the network. Interval.

9. A multi-domain intelligent separation system for seismic diffraction wavefields, characterized in that, This system is used to implement the intelligent separation method for multi-domain seismic diffraction wavefields as described in any one of claims 1 to 8, and the system comprises: The data acquisition module is used to acquire seismic full-wave field data, which includes common offset domain data and / or dip domain data. The model building module is used to build a conditional denoising diffusion probability model. The seismic full wavefield data is used as the conditional input, and the conditional denoising diffusion probability model is extended into a dual-channel joint diffusion model. The dual-channel joint diffusion model includes a dual-output noise prediction network, which is used to predict the diffraction noise component and the reflection noise component, respectively. The training module is used to train the dual-channel joint diffusion model using a joint loss function, which simultaneously constrains the prediction error of diffraction noise and the prediction error of reflection noise. The generation module is used to independently sample the first initial noise and the second initial noise from the standard normal distribution, which are used as the initial states of the diffracted wave and the reflected wave, respectively. The first initial noise and the second initial noise are input into the trained dual-channel joint diffusion model, and the separated diffracted wave data and reflected wave data are generated iteratively through the reverse denoising process, using the earthquake full wave field data as a condition.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the intelligent separation method for multi-domain seismic diffraction wavefields as described in any one of claims 1 to 8.

Images