A method for detecting high-frequency continuation of a signal and related device
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CENT SOUTH UNIV
- Filing Date
- 2026-05-26
- Publication Date
- 2026-06-23
Smart Images
Figure CN122260488A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of signal processing technology, and in particular to a high-frequency extension method and related equipment for detecting signals. Background Technology
[0002] High-resolution seismic data is the core foundation for accurately identifying deep underground geological structures and characterizing reservoir properties. However, seismic waves undergo energy attenuation due to stratum absorption during underground propagation, especially high-frequency energy loss, which leads to a significant decrease in the resolution of seismic data for deep target layers with depth. This results in the broadening of the phase axis and the submergence of thin-layer information, greatly increasing the uncertainty of deep reservoir modeling. This is a core technical bottleneck restricting breakthroughs in deep oil and gas exploration.
[0003] Existing methods for seismic data resolution enhancement and high-frequency extension can be broadly categorized into two types: traditional physics-driven methods and data-driven deep learning methods. Traditional physical-driven methods are based on the physical laws of seismic wave propagation. They improve resolution by compensating for formation absorption effects and expanding the signal bandwidth. Core technologies include three main categories: deconvolution, spectral whitening, and inverse Q-filtering. Deconvolution, by compressing seismic wavelets through pulse constraints, improves longitudinal resolution and is the most widely used conventional method in industry. Spectral whitening expands the effective bandwidth of seismic signals through spectral equalization; it is simple to operate and computationally efficient, making it the mainstream solution for improving the conventional resolution of post-stack seismic data. Inverse Q-filtering, based on a formation absorption attenuation model, compensates for formation absorption through inverse filtering. The amplitude attenuation and phase distortion during seismic wave propagation are classic methods that strictly follow the physical laws of seismic wave propagation. However, these traditional methods have unavoidable inherent limitations. Their core reliance on simplified linear stationarity assumptions and manually designed a priori knowledge makes it impossible to accurately determine the nonlinear and non-stationary characteristics of seismic wave attenuation in deep and complex strata. They also have poor adaptability to complex structural areas with fault development and abrupt lithological changes. Furthermore, under low signal-to-noise ratio conditions, noise is easily amplified synchronously during high-frequency extension, resulting in the common problems of low-frequency information loss and high-frequency noise amplification, making it difficult to achieve high-fidelity resolution improvement of deep seismic data.
[0004] With the development of deep learning technology, data-driven seismic resolution enhancement methods have made groundbreaking progress. The core of these methods is to learn complex nonlinear mappings from low-resolution to high-resolution seismic data through neural networks, without relying on strict physical assumptions, and outperforming traditional methods in multiple scenarios. Among these, supervised learning methods are the mainstream research direction. Researchers have achieved vertical resolution enhancement of seismic data using the U-Net convolutional neural network, validating the effectiveness of convolutional architectures in seismic super-resolution tasks. Other researchers have proposed iterative deep neural networks to achieve simultaneous estimation of seismic wavelets and reflectivity, as well as high-resolution inversion, improving the network's interpretability. Meanwhile, generative adversarial networks (GANs) have been widely used for seismic resolution enhancement. By building a GAN-based seismic resolution enhancement network, more continuous and refined reflection phase axes have been generated than traditional methods. Some scholars have also constructed an integrated network that simultaneously achieves seismic super-resolution and denoising, further optimizing the processing effect in low signal-to-noise ratio scenarios. However, these supervised learning methods heavily rely on scarce and costly LR-HR paired label data. In actual deep seismic exploration, it is impossible to obtain real underground resolution labels. Only synthetic labels can be generated through forward modeling, which deviates significantly from the actual underground geological features, making it difficult for the methods to be industrialized.
[0005] Therefore, existing weakly supervised deep learning methods still have four core shortcomings: 1. The lack of accurate modeling of stratigraphic dip structure makes the local dip estimation of deep seismic data susceptible to high-frequency noise interference. Existing single-scale estimation methods cannot take into account both detailed structure and stratigraphic strike, have poor noise robustness, and cannot provide reliable structural guidance constraints for high-frequency extension. 2. Insufficient ability to track in-phase axes in fault-developed areas; difficulty in solving the problems of in-phase axis intersections and cross-fault connections caused by fine fractures; easy to cause misconnection or omission of in-phase axes; resulting in severe structural distortion in the high-frequency extension results of structurally complex areas. 3. Deep earthquakes lack the ability to extract relative geological age information. Existing methods have weak characterization ability and poor generalization of deep low signal-to-noise ratio data, and cannot provide large-scale stratigraphic prior constraints for high-frequency extension. They are also prone to generating false high-frequency components that do not conform to geological laws. 4. The time-frequency analysis and wavelet reconstruction accuracy is insufficient, making it impossible to accurately characterize the transient evolution characteristics of deep spatiotemporal variable wavelets. Phase distortion is prone to occur during high-frequency extension, resulting in low effective signal fidelity and making it impossible to achieve high-fidelity bandwidth extension of deep seismic data. Summary of the Invention
[0006] This invention provides a high-frequency extension method and related equipment for detecting signals, aiming to solve the technical challenges of high-fidelity, high-robust bandwidth widening and resolution improvement of deep seismic data.
[0007] To achieve the above objectives, the present invention provides a high-frequency extension method for detecting signals, comprising: Step 1: Extract seismic profile data with different depths and signal-to-noise ratios from the seismic data in the work area, and add noise to the seismic profile data to obtain high signal-to-noise ratio data; Step 2: Train the constructed multi-scale branch network using high signal-to-noise ratio data to obtain the local dip estimation network for seismic angle estimation. Input the deep seismic data of the target into the local dip estimation network for processing to obtain local dip field data. Step 3: Input the seismic profile data and the corresponding local dip field data into the constructed dynamic spatial attention network for feature extraction to obtain the phase axis edge delay features. Then, using the strong reflection peak points of the seismic data as seeds, cross-fault tracking is performed along the dip-guided direction through the phase axis edge delay features to obtain the phase axis tracking results. Step 4: Based on the phase axis tracking results and the constructed refined annotation dataset, the two-stage incremental learning model is trained to obtain the seismic facies spatial representation model. The target deep seismic data is then input into the seismic facies spatial representation model for extraction to obtain the geological age corresponding to the seismic facies spatial representation results. Step 5: Input the geological age and seismic profile data into the inversion physical model for processing to obtain elastic parameter data. Then, input the elastic parameter data, well logging data, and well-side seismic data into the forward physical model for reconstruction to obtain the time-varying wavelet spectrum characteristics. Step 6: Construct a high-frequency extension inversion objective function based on the time-varying wavelet spectrum characteristics, complete the inversion solution based on the high-frequency extension inversion objective function, obtain a high-resolution reflection coefficient sequence, and convolve the high-resolution reflection coefficient sequence with the time-varying wavelet spectrum characteristics to obtain the high-frequency extended deep seismic data.
[0008] Furthermore, multi-scale branching networks include: The backbone network consists of encoders and decoders; A large-scale subnetwork composed of large convolutional layers and pooling layers is used to capture low-frequency features of the overall strike of the strata; A medium-scale subnetwork consisting of standard convolutional layers is used for medium-sized construction features; Small-scale subnetworks consisting of dilated convolutional layers and deformable convolutional layers are used to construct detailed features; The encoder's output is connected to the input of the large-scale subnetwork, the input of the medium-scale subnetwork, and the input of the small-scale subnetwork, respectively. The outputs of the large-scale subnetwork, the medium-scale subnetwork, and the small-scale subnetwork are all connected to the input of the decoder to output local tilt field data at different scales.
[0009] Furthermore, a dynamic spatial attention network is constructed by inputting seismic profile data and corresponding local dip field data to extract features, obtaining the time delay features at the edges of the phase axes, including: A dynamic spatial attention network is constructed by inputting seismic profile data and corresponding local dip field data; The receptive field orientation is adjusted by tilt-guided convolutional layers in a dynamic spatial attention network; An adaptive spatial scale cross-correlation algorithm is used to calculate the cross-correlation coefficients between different gathers along the strike of the dip-estimated strata, thereby obtaining the edge delay characteristics of the in-phase axis.
[0010] Furthermore, the expression for calculating the cross-correlation coefficient between different Dao collections is: ; in, Indicating the first earthquake data Tao and the First Cross-correlation coefficients between channel signals Indicates the first channel signal, Indicates the first channel signal, , All represent the mean of the signal within the window. This represents the adaptive time delay offset for tilt-guided tilt. Indicates the radius of the adaptive spatial scale window. Indicates a time window.
[0011] Furthermore, using the strong reflection peaks of seismic data as seeds, cross-fault tracing is performed along the dip-guided direction using the time delay characteristics at the edge of the phase axis, yielding phase axis tracing results, including: The amplitude envelope data is calculated by performing a Hilbert transform on the seismic profile data. A sliding time window is set along the strike of the strata based on local dip field data, and local maxima in the amplitude envelope data are searched within the sliding time window as candidate peak values. The candidate peaks are filtered by a signal-to-noise ratio threshold to obtain the strong reflection peaks of the seismic data as seeds; Based on the seed, the phase axis is tracked along the dip angle. When a fault area is encountered, the phase axis matching on both sides of the fault is completed through the dynamic time warping algorithm to obtain the phase axis tracking result.
[0012] Furthermore, the two-stage incremental learning model includes: Variational autoencoder, latent space clustering layer, 3D convolutional decoder, incremental classification head; Variational autoencoders extract seismic data features through multi-scale three-dimensional convolutional layers; The latent space clustering layer is used to map the seismic data features to the latent space and then perform unsupervised clustering using K-means. A 3D convolutional decoder is used to learn general features of seismic facies by reconstructing a loss-constrained model; The incremental classification head is used to obtain the geological age corresponding to the seismic epochal spatial characterization results by adopting a batch incremental learning paradigm of strata from shallow to deep.
[0013] Furthermore, the objective function for high-frequency extension inversion based on the time-varying wavelet spectral characteristics is as follows: ; in, This indicates the core item for data fidelity. Represents the raw earthquake data. This represents the high-resolution reflectance coefficient to be determined. Represents a time-varying wavelet. , , All represent weights. Represents sparse regularization terms. This represents the combined constraint term for dip structure and seismic facies strata priori. Indicates low-frequency fidelity. This represents the low-pass filter operator.
[0014] This invention also provides a high-frequency extension device for seismic data, which applies a high-frequency extension method for detection signals. The high-frequency extension device for seismic data includes: The extraction module is used to extract seismic profile data of different depths and signal-to-noise ratios from the seismic data in the work area, and add noise to the seismic profile data to obtain high signal-to-noise ratio data. The processing module is used to train the constructed multi-scale branch network with high signal-to-noise ratio data to obtain the local dip estimation network of the earthquake, and to input the deep seismic data of the target into the local dip estimation network for processing to obtain local dip field data; The tracking module is used to extract features from the dynamic spatial attention network constructed by inputting seismic profile data and corresponding local dip field data to obtain the edge delay features of the phase axis. Using the strong reflection peak points of the seismic data as seeds, the module performs cross-fault tracking along the dip-guided direction through the edge delay features of the phase axis to obtain the phase axis tracking results. The extraction module is used to train the two-stage incremental learning model based on the phase axis tracking results and the constructed fine-labeled dataset to obtain the seismic facies spatial representation model. The target deep seismic data is then input into the seismic facies spatial representation model for extraction to obtain the geological age corresponding to the seismic facies spatial representation results. The reconstruction module is used to input geological age and seismic profile data into the inversion physical model for processing to obtain elastic parameter data. It then inputs the elastic parameter data, well logging data, and well-side seismic data into the forward physical model for reconstruction to obtain time-varying wavelet spectrum characteristics. The solution module is used to construct a high-frequency extension inversion objective function based on the time-varying wavelet spectrum characteristics, complete the inversion solution based on the high-frequency extension inversion objective function, obtain a high-resolution reflection coefficient sequence, and then convolve the high-resolution reflection coefficient sequence with the time-varying wavelet spectrum characteristics to obtain the high-frequency extended deep seismic data.
[0015] The present invention also provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a high-frequency extension method for detecting signals.
[0016] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a high-frequency extension method for detecting signals.
[0017] The above-described solution of the present invention has the following beneficial effects: Compared with existing technologies, this invention constructs a local dip estimation network to process deep seismic data to obtain local dip field data, solving the core defects of traditional single-scale methods that are susceptible to high-frequency noise interference in deep areas and cannot simultaneously consider macroscopic trends and detailed structures. By using a dynamic spatial attention network to focus on structurally complex regions and combining adaptive spatial scale cross-correlation to capture signal edge delay features, it solves the problems of in-phase axis intersections and cross-fault connections caused by fragmented faults. A two-stage incremental learning seismic facies spatial representation model is constructed to extract the geological time corresponding to the seismic facies spatial representation results. The geological time is then compared with… Seismic profile data is input into an inversion physical model for processing to obtain elastic parameter data. This elastic parameter data, well logging data, and well-side seismic data are then input into a forward physical model for reconstruction, yielding time-varying wavelet spectral characteristics. A high-frequency extension inversion objective function is constructed based on these characteristics. The inversion is then solved using this objective function, resulting in a high-resolution reflection coefficient sequence. This high-resolution reflection coefficient sequence is then convolved with the time-varying wavelet spectral characteristics to obtain the high-frequency extended deep seismic data. This approach solves the technical challenges of high-fidelity, high-robust bandwidth broadening and resolution improvement for deep seismic data.
[0018] Other beneficial effects of the present invention will be described in detail in the following detailed description section. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating an embodiment of the present invention; Figure 2This is a schematic diagram of the high-frequency extension device for seismic data in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the terminal device in an embodiment of the present invention. Detailed Implementation
[0020] To make the technical problems, solutions, and advantages of this invention clearer, a detailed description will be provided below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0021] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0022] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0023] This invention addresses existing problems by providing a high-frequency extension method and related equipment for detecting signals.
[0024] like Figure 1 As shown, an embodiment of the present invention provides a high-frequency extension method for detecting signals, comprising: Step 1: Extract seismic profile data with different depths and signal-to-noise ratios from the seismic data in the work area, and add noise to the seismic profile data to obtain high signal-to-noise ratio data; Step 2: Train the constructed multi-scale branch network using high signal-to-noise ratio data to obtain the local dip estimation network for seismic angle estimation. Input the deep seismic data of the target into the local dip estimation network for processing to obtain local dip field data. Step 3: Input the seismic profile data and the corresponding local dip field data into the constructed dynamic spatial attention network for feature extraction to obtain the phase axis edge delay features. Then, using the strong reflection peak points of the seismic data as seeds, cross-fault tracking is performed along the dip-guided direction through the phase axis edge delay features to obtain the phase axis tracking results. Step 4: Based on the phase axis tracking results and the constructed refined annotation dataset, the two-stage incremental learning model is trained to obtain the seismic facies spatial representation model. The target deep seismic data is then input into the seismic facies spatial representation model for extraction to obtain the geological age corresponding to the seismic facies spatial representation results. Step 5: Input the geological age and seismic profile data into the inversion physical model for processing to obtain elastic parameter data. Then, input the elastic parameter data, well logging data, and well-side seismic data into the forward physical model for reconstruction to obtain the time-varying wavelet spectrum characteristics. Step 6: Construct a high-frequency extension inversion objective function based on the time-varying wavelet spectrum characteristics, complete the inversion solution based on the high-frequency extension inversion objective function, obtain a high-resolution reflection coefficient sequence, and convolve the high-resolution reflection coefficient sequence with the time-varying wavelet spectrum characteristics to obtain the high-frequency extended deep seismic data.
[0025] Specifically, multi-scale branching networks include: The backbone network consists of encoders and decoders; A large-scale subnetwork composed of large convolutional layers and pooling layers is used to capture low-frequency features of the overall strike of the strata; A medium-scale subnetwork consisting of standard convolutional layers is used to characterize medium-sized structural features; A small-scale subnetwork consisting of dilated convolutional layers and deformable convolutional layers is used to capture high-frequency features of detailed structures such as fine cracks and thin layers. The encoder's output is connected to the input of the large-scale subnetwork, the input of the medium-scale subnetwork, and the input of the small-scale subnetwork, respectively. The outputs of the large-scale subnetwork, the medium-scale subnetwork, and the small-scale subnetwork are all connected to the input of the decoder to output local tilt field data at different scales.
[0026] In this embodiment of the invention, the large convolutional layer uses a 7×7 large convolutional kernel, and the standard convolutional layer uses a 3×3 convolutional kernel, which is used to characterize medium-sized structural features such as faults and folds.
[0027] This invention employs a course-learning strategy in multi-scale branch networks. First, high signal-to-noise ratio (SNR) data is used for network pre-training, then low SNR data is gradually introduced for fine-tuning, improving the network's resistance to deep high-frequency noise. Simultaneously, the Adam optimizer is used to optimize network parameters, setting... , The initial learning rate is 0.0001, decays by 50% every 500 training epochs, the batch size is 4, and the total number of training epochs is 20,000.
[0028] In this embodiment of the invention, the high signal-to-noise ratio data used for preprocessing multi-scale branch networks needs to be obtained by manual annotation and coherence calculation to obtain tilt label data. The sample size is expanded by data flipping, rotation, scaling and other enhancement strategies, thereby reducing the risk of model overfitting.
[0029] In the training process, the multi-scale branch network of this invention introduces multi-scale cross-gradient constraints to calculate the spatial gradient distribution of tilt angle estimation results at different scales. This constrains the gradient direction of the large-scale tilt angle to maintain consistency with the principal gradient directions of the medium- and small-scale tilt angles, ensuring the spatial structure consistency of the estimation results at different scales and avoiding local distortions that occur in single-scale estimation. The L1 loss is used as the main loss for tilt angle estimation, combined with multi-scale cross-gradient regularization loss and smoothing loss, to construct the expression for the joint loss function: ; in, Represents the joint loss value. This represents the L1 norm loss between the tilt angle estimation result and the label. This represents the multi-scale cross-gradient constraint loss. This indicates the loss of spatial smoothness in the tilt field. , This represents the weighting coefficient, with default values of 0.5 and 0.2.
[0030] Specifically, a dynamic spatial attention network is constructed by inputting seismic profile data and corresponding local dip field data to extract features, obtaining the time delay features at the edges of the phase axes, including: A dynamic spatial attention network is constructed by inputting seismic profile data and corresponding local dip field data; The receptive field orientation is adjusted by tilt-guided convolutional layers in a dynamic spatial attention network; An adaptive spatial scale cross-correlation algorithm is used to calculate the cross-correlation coefficients between different gathers along the strike of the dip-estimated strata, thereby obtaining the edge delay characteristics of the in-phase axis.
[0031] Specifically, the expression for calculating the cross-correlation coefficient between different Dao collections is: ; in, Indicating the first earthquake data Tao and the First Cross-correlation coefficients between channel signals Indicates the first channel signal, Indicates the first channel signal, , All represent the mean of the signal within the window. This represents the adaptive time delay offset for tilt-guided tilt. Indicates the radius of the adaptive spatial scale window. Indicates a time window.
[0032] In this embodiment of the invention, a dynamic gating mechanism is designed when extracting features of the same phase axis. Higher attention weights are assigned to complex regions with fault development and intersection of the same phase axis, focusing on feature extraction of complex regions.
[0033] For seismic data on both sides of a fault, this embodiment of the invention designs a dynamic time warping loss function for a dynamic spatial attention network. By minimizing the warping path distance, it matches the vector data features of the same phase axes on both sides of the fault, solving the matching problem caused by the misalignment of the same phase axes. At the same time, it adds a dip angle consistency constraint, requiring that the dip angle of the matched same phase axes be consistent with the dip field data, avoiding misconnection of the same phase axes.
[0034] Simultaneously, using the strong reflection peaks of seismic data as seeds, cross-fault tracing is performed along the dip-guided direction using the time delay characteristics at the edge of the seismic axis, yielding the seismic axis tracing results, including: The amplitude envelope data is calculated by performing a Hilbert transform on the seismic profile data. A sliding time window is set along the strike of the strata based on local dip field data, and local maxima in the amplitude envelope data are searched within the sliding time window as candidate peak values. The candidate peaks are filtered by a signal-to-noise ratio threshold to obtain the strong reflection peaks of the seismic data as seeds; Based on the seed, the in-phase axis is tracked along the dip angle. When a fault region is encountered, the in-phase axis matching on both sides of the fault is completed through the dynamic time warping algorithm, so as to obtain a complete and continuous in-phase axis tracking result.
[0035] Specifically, the two-stage incremental learning model includes: Variational autoencoder, latent space clustering layer, 3D convolutional decoder, incremental classification head; Variational autoencoders extract seismic data features through multi-scale three-dimensional convolutional layers; The latent space clustering layer is used to map the seismic data features to the latent space and then perform unsupervised clustering using K-means. A 3D convolutional decoder is used to learn general features of seismic facies by reconstructing a loss-constrained model; The incremental classification head is used to obtain the geological age corresponding to the seismic epochal spatial characterization results by adopting a batch incremental learning paradigm of strata from shallow to deep.
[0036] In this embodiment of the invention, seismic data is divided into blocks according to stratigraphic units based on the phase axis tracking results to obtain block results. Then, the block results are used to pre-train a two-stage incremental learning model. By learning the general latent space representation of seismic facies, seismic facies of different geological ages and different lithologies are mapped to different clusters in the latent space, thus completing the basic feature learning of the model.
[0037] This invention uses well-logged seismic data as the core, combined with regional geological priors to construct a refined labeled dataset. An incremental learning strategy is adopted, in which the labeled data is divided into multiple batches according to the stratigraphic age from shallow to deep, and gradually input into the pre-trained two-stage incremental learning model for fine-tuning. This avoids the model forgetting the general features learned from shallow layers, while gradually adapting to the feature distribution of deep seismic facies.
[0038] In the incremental learning process, this invention employs a joint loss function of contrastive learning loss and cross-entropy loss to bring the distance between samples of the same seismic facies in the latent space closer and push the distance between samples of different seismic facies further apart, iteratively improving the model's classification accuracy and representation ability for deep seismic facies. At the same time, stratigraphic chronological order constraints are added, requiring the clustering distribution of seismic facies in the latent space to conform to the geological laws of stratigraphic deposition, thereby improving the geological reliability of the representation results.
[0039] Specifically, geological age and seismic profile data are input into the inversion physical model for processing to obtain elastic parameter data, including P-wave velocity, S-wave velocity, and density. The elastic parameter data are combined with well logging data and well-side seismic data with the forward physical model. Through well-seismic calibration, time-varying wavelet spectrum characteristics at different depths and incident angles are extracted, and analytical expressions of the time-varying wavelet spectrum characteristics are fitted to clarify the evolution of wavelet dominant frequency, phase, and bandwidth with depth and stratum lithology.
[0040] In this embodiment of the invention, the spectral characteristics of the time-varying wavelet are expressed using an improved analytical expression for the Ricker wavelet with incident angle and depth correction, specifically: ; in, Represents the wavelet time variable. This indicates a two-way trip (corresponding to the depth of the stratum). Indicates the angle of incidence of seismic waves. Represents the wavelet amplitude coefficient. Indicates the clock speed. Indicates phase.
[0041] Specifically, the extraction of time-varying wavelet spectral characteristics at different depths and incident angles through well seismic calibration includes: Based on the in-phase axis tracking results and the block results, the time-frequency matrix output by the second-order transient transform is precisely aligned with each block in the block results in the time-space dimension. For each block, a sliding time-frequency window with equal time width is set along the phase axis. Within the window, the core features of the time spectrum, such as peak frequency, effective bandwidth, and phase spectrum slope, are statistically analyzed to remove outliers from abnormal noise channels. The variation law of the wavelet spectrum of this block with depth and incident angle was obtained by least squares fitting. At the same time, the calibration results of the well-side seismic data were combined for correction to obtain the time-varying wavelet spectrum characteristics.
[0042] Specifically, the objective function for high-frequency extension inversion based on the time-varying wavelet spectral characteristics is as follows: ; in, The core term representing data fidelity is the L2 norm error between the convolution results of the original seismic data and the desired high-resolution reflection coefficients and time-varying wavelet spectral characteristics, ensuring the physical consistency of the inversion. Represents the raw earthquake data. This represents the high-resolution reflectance coefficient to be determined. Represents a time-varying wavelet. , , All represent weights. This represents a sparse regularization term used to match the sparse distribution characteristics of formation reflection coefficients. This represents a combined constraint term representing the a priori relationship between dip tectonics and seismic facies strata, ensuring spatial continuity and geological rationality. This represents the low-frequency fidelity term. A low-pass filter operator is used to lock in the effective low-frequency components, preventing low-frequency distortion during the inversion process. This represents the low-pass filter operator.
[0043] In this embodiment of the invention, the high-frequency extension inversion objective function takes the spectral evolution law of the wavelet as the core constraint, and combines the seismic facies strata prior and dip angle structural guidance constraint to invert and recover the high-frequency effective components lost due to strata absorption attenuation while ensuring the fidelity of the low-frequency effective signal.
[0044] This invention employs a fast iterative shrinking threshold algorithm to complete the inversion solution and recover the high-frequency effective components. Specifically: First, the raw seismic data is low-pass filtered to extract the effective low-frequency components below 15Hz, which serve as hard constraints for the iterative process to ensure that the low-frequency signals are not distorted. During the iteration, a soft threshold operator is used to implement sparse constraints on the reflection coefficients, and spatial smoothing filtering is performed along the dip field direction to implement tectonic guidance constraints. Combined with the wavelet spectrum evolution law of seismic phase blocks, adaptive compensation is performed on the attenuated high-frequency components. After the iteration converges, a high-resolution reflection coefficient sequence is obtained, which is convolved with the time-varying wavelet spectrum characteristics to finally recover the high-frequency effective components lost due to formation absorption attenuation, thus completing the high-frequency extension of the seismic data.
[0045] The formula for the soft threshold operator is as follows: ; in, This represents the soft threshold operator. The adaptive regularization threshold representing the location in three-dimensional space is determined jointly by dip field data and seismic phase priors: ; in, Indicates the baseline threshold. This represents the local dip gradient. The smaller the gradient magnitude and the lower the threshold in the continuous region, the more thin-layer details can be preserved. This represents the lithological probability output by the seismic facies. The threshold is higher in stable sedimentary facies regions to ensure sparsity.
[0046] In the iterative process of this embodiment of the invention, a structurally guided smoothing step is added after the gradient descent step. The intermediate variables are weighted and averaged along the main dip direction to ensure the spatial continuity of the reflection coefficient. Every 5 iterations, the wavelet parameters are updated based on the seismic phase prior to ensure the consistency of the wavelet stratigraphy. After the final iteration converges, a high-resolution reflection coefficient sequence without structural distortion can be obtained.
[0047] The embodiments of this invention have completed a comprehensive feasibility verification through synthetic data simulation and actual field data testing in the work area. The core results are as follows: Synthetic data simulation validation: Synthetic seismic data was constructed based on the Marmousi complex tectonic model. Formation absorption attenuation was simulated using different Q values. Random noise of 5dB, 10dB, and 15dB was added to simulate deep low signal-to-noise ratio scenarios. Comparisons were made with traditional spectral whitening methods, inverse Q-filtering methods, and weakly supervised CycleGAN methods, as shown in Table 1 below. Table 1. Quantitative comparison of the performance of various methods under different signal-to-noise ratios.
[0048] The method provided in this invention achieves the lowest mean absolute error (MAE), the highest structural similarity (SSIM), and the highest Pearson correlation coefficient (PCC) under all signal-to-noise ratio conditions. Specifically, in a strong noise scenario with an SNR of 5 dB, the method provided in this invention achieves an MAE of only 1.28, an SSIM of 0.89, and a PCC of 0.96, significantly outperforming the comparative methods and demonstrating the method's strong noise robustness and high-resolution reconstruction accuracy.
[0049] Verification using actual field data: 3D post-stack seismic data from the deep oil and gas exploration area in western China were used, with a target layer depth of 4500~6000m. The original data had a dominant frequency of only 25Hz and an effective bandwidth of 10~45Hz, as shown in Table 2 below. Table 2. Verification results of actual field work area data
[0050] As shown in Table 2 above, after processing by the method provided in this embodiment of the invention, the dominant frequency of the target layer data is increased to 45Hz, the effective bandwidth is expanded to 8~80Hz, the high-frequency effective energy is significantly recovered, the characterization accuracy of geological structures such as thin layers and faults is greatly improved, and the continuity of the phase axis is significantly enhanced. Well-seismic calibration results show that the normalized cross-correlation coefficient between the seismic data processed by the method of this invention and the well logging composite record reaches 0.82, far higher than the 0.63 of the traditional spectral whitening method and the 0.71 of the weakly supervised CycleGAN method, demonstrating the high fidelity and geological reliability of the high-frequency extrapolation results.
[0051] Compared with existing technologies, this invention constructs a local dip estimation network to process deep seismic data to obtain local dip field data, solving the core defects of traditional single-scale methods that are susceptible to high-frequency noise interference in deep areas and cannot simultaneously consider macroscopic trends and detailed structures. By using a dynamic spatial attention network to focus on structurally complex regions and combining adaptive spatial scale cross-correlation to capture signal edge delay features, it solves the problems of in-phase axis intersections and cross-fault connections caused by fragmented faults. A two-stage incremental learning seismic facies spatial representation model is constructed to extract the geological ages corresponding to the seismic facies spatial representation results. The geological ages are then... Seismic profile data is input into an inversion physical model for processing to obtain elastic parameter data. This elastic parameter data, well logging data, and well-side seismic data are then input into a forward physical model for reconstruction, yielding time-varying wavelet spectral characteristics. A high-frequency extension inversion objective function is constructed based on these characteristics, and the inversion is solved using this function to obtain a high-resolution reflection coefficient sequence. This high-resolution reflection coefficient sequence is then convolved with the time-varying wavelet spectral characteristics to obtain the high-frequency extended deep seismic data. This approach solves the technical challenges of high-fidelity, high-robust bandwidth broadening and resolution improvement for deep seismic data.
[0052] Corresponding to the high-frequency extension method for the detection signal described in the above embodiments, such as Figure 2 As shown, this embodiment of the invention also provides a high-frequency seismic data extension device 100, which includes: The extraction module 101 is used to extract seismic profile data of different depths and signal-to-noise ratios from the seismic data in the work area, and add noise to the seismic profile data to obtain high signal-to-noise ratio data. Processing module 102 is used to train the constructed multi-scale branch network using high signal-to-noise ratio data to obtain the local dip estimation network of the earthquake, and to input the deep seismic data of the target into the local dip estimation network of the earthquake for processing to obtain local dip field data; The tracking module 103 is used to input seismic profile data and corresponding local dip field data into a dynamic spatial attention network to extract features, obtain the edge delay features of the phase axis, and use the strong reflection peak points of the seismic data as seeds to perform cross-fault tracking along the dip-guided direction through the edge delay features of the phase axis to obtain the phase axis tracking results. The extraction module 104 is used to train the two-stage incremental learning model based on the phase axis tracking results and the constructed fine-labeled dataset to obtain the seismic facies spatial representation model, and to input the target deep seismic data into the seismic facies spatial representation model for extraction to obtain the geological age corresponding to the seismic facies spatial representation results. The reconstruction module 105 is used to input geological age and seismic profile data into the inversion physical model for processing to obtain elastic parameter data, and to input the elastic parameter data, well logging data, and well-side seismic data into the forward physical model for reconstruction to obtain time-varying wavelet spectrum characteristics. The solver module 106 is used to construct a high-frequency extension inversion objective function based on the time-varying wavelet spectrum characteristics, complete the inversion solution based on the high-frequency extension inversion objective function, obtain a high-resolution reflection coefficient sequence, and then convolve the high-resolution reflection coefficient sequence with the time-varying wavelet spectrum characteristics to obtain the high-frequency extended deep seismic data.
[0053] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.
[0054] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0055] This invention also provides a terminal device, such as... Figure 3 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 3 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100. When the processor D100 executes the computer program D102, it implements the above-described high-frequency extension method for detecting signals.
[0056] The terminal device D10 can be a desktop computer, laptop, handheld computer, server, server cluster, or cloud server, etc. This terminal device may include, but is not limited to, a processor D100 and a memory D101. Those skilled in the art will understand that... Figure 3 This is merely an example of terminal device D10 and does not constitute a limitation on terminal device D10. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.
[0057] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.
[0058] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.
[0059] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.
[0060] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0061] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements a high-frequency extension method for detecting signals.
[0062] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a building device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.
[0063] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A method for detecting high-frequency extension of signals, characterized in that, include: Step 1: Extract seismic profile data with different depths and signal-to-noise ratios from the seismic data within the work area, and add noise to the seismic profile data to obtain high signal-to-noise ratio data; Step 2: Train the constructed multi-scale branch network using the high signal-to-noise ratio data to obtain the local dip estimation network, and input the target deep seismic data into the local dip estimation network for processing to obtain the local dip field data; Step 3: Input the seismic profile data and the corresponding local dip field data into the constructed dynamic spatial attention network for feature extraction to obtain the phase axis edge delay features. Then, using the strong reflection peak points of the seismic data as seeds, cross-fault tracking is performed along the dip-guided direction through the phase axis edge delay features to obtain the phase axis tracking results. Step 4: Based on the in-phase axis tracking results and the constructed refined annotation dataset, the two-stage incremental learning model is trained to obtain the seismic facies spatial representation model. The target deep seismic data is then input into the seismic facies spatial representation model for extraction to obtain the geological age corresponding to the seismic facies spatial representation results. Step 5: Input the geological age and the seismic profile data into the inversion physical model for processing to obtain elastic parameter data. Then, input the elastic parameter data, well logging data, and well-side seismic data into the forward physical model for reconstruction to obtain time-varying wavelet spectrum characteristics. Step 6: Construct a high-frequency extension inversion objective function based on the time-varying wavelet spectrum characteristics, complete the inversion solution based on the high-frequency extension inversion objective function to obtain a high-resolution reflection coefficient sequence, and convolve the high-resolution reflection coefficient sequence with the time-varying wavelet spectrum characteristics to obtain the high-frequency extended deep seismic data.
2. The high-frequency extension method for detection signals according to claim 1, characterized in that, The multi-scale branching network includes: The backbone network consists of encoders and decoders; A large-scale subnetwork composed of large convolutional layers and pooling layers is used to capture low-frequency features of the overall strike of the strata; A medium-scale subnetwork consisting of standard convolutional layers is used for medium-sized construction features; Small-scale subnetworks consisting of dilated convolutional layers and deformable convolutional layers are used to construct detailed features; The output of the encoder is connected to the input of the large-scale sub-network, the input of the medium-scale sub-network, and the input of the small-scale sub-network, respectively. The output terminals of the large-scale subnetwork, the medium-scale subnetwork, and the small-scale subnetwork are all connected to the input terminal of the decoder to output local tilt field data at different scales.
3. The high-frequency extension method for detection signals according to claim 1, characterized in that, The seismic profile data and the corresponding local dip field data are input into a constructed dynamic spatial attention network for feature extraction to obtain the time delay features at the edge of the phase axis, including: The seismic profile data and the corresponding local dip field data are input into a dynamic spatial attention network; The receptive field orientation is adjusted by tilt-guided convolutional layers in the dynamic spatial attention network. An adaptive spatial scale cross-correlation algorithm is used to calculate the cross-correlation coefficients between different gathers along the strike of the dip-estimated strata, thereby obtaining the edge delay characteristics of the in-phase axis.
4. The high-frequency extension method for detection signals according to claim 3, characterized in that, The expression for calculating the cross-correlation coefficient between different collections is: ； in, Indicating the first earthquake data Tao and the First Cross-correlation coefficients between channel signals Indicates the first channel signal, Indicates the first channel signal, , All represent the mean of the signal within the window. This represents the adaptive time delay offset for tilt-guided tilt. Indicates the radius of the adaptive spatial scale window. Indicates a time window.
5. The high-frequency extension method for detection signals according to claim 1, characterized in that, Using the strong reflection peaks of the seismic data as seeds, cross-fault tracing is performed along the dip-guided direction using the time delay characteristics at the edge of the in-phase axis to obtain the in-phase axis tracing results, including: The amplitude envelope data was calculated by performing a Hilbert transform on the seismic profile data. Based on the local dip field data, a sliding time window is set along the strike of the strata, and local maxima in the amplitude envelope data are searched within the sliding time window as candidate peak values. The candidate peaks are filtered by signal-to-noise ratio threshold to obtain the strong reflection peaks of the seismic data as seeds; Based on the seed, the phase axis is tracked along the dip angle. When a fault area is encountered, the phase axis matching on both sides of the fault is completed through the dynamic time warping algorithm to obtain the phase axis tracking result.
6. The high-frequency extension method for detection signals according to claim 1, characterized in that, The two-stage incremental learning model includes: Variational autoencoder, latent space clustering layer, 3D convolutional decoder, incremental classification head; The variational autoencoder extracts seismic data features through multi-scale three-dimensional convolutional layers; The latent space clustering layer is used to map the seismic data features to the latent space and then perform unsupervised clustering using K-means. The three-dimensional convolutional decoder is used to learn the general features of seismic facies by reconstructing a loss-constrained model; The incremental classification head is used to obtain the geological age corresponding to the seismic epochal spatial characterization results by adopting a batch incremental learning paradigm of strata from shallow to deep.
7. The high-frequency extension method for detection signals according to claim 1, characterized in that, Based on the aforementioned time-varying wavelet spectrum characteristics, the high-frequency extension inversion objective function is constructed as follows: ； in, This indicates the core item for data fidelity. Represents the raw earthquake data. This represents the high-resolution reflectance coefficient to be determined. Represents a time-varying wavelet. , , All represent weights. Represents sparse regularization terms. This represents the combined constraint term for dip structure and seismic facies strata priori. Indicates low-frequency fidelity. This represents the low-pass filter operator.
8. A high-frequency seismic data extension device, characterized in that, The high-frequency extension device for seismic data, using the detection signal high-frequency extension method as described in any one of claims 1-7, comprises: The extraction module is used to extract seismic profile data of different depths and signal-to-noise ratios from the seismic data in the work area, and add noise to the seismic profile data to obtain high signal-to-noise ratio data; The processing module is used to train the constructed multi-scale branch network using the high signal-to-noise ratio data to obtain the local dip estimation network, and to input the target deep seismic data into the local dip estimation network for processing to obtain the local dip field data. The tracking module is used to input the seismic profile data and the corresponding local dip field data into a dynamic spatial attention network to extract features, obtain the phase axis edge delay features, and use the strong reflection peak points of the seismic data as seeds to perform cross-fault tracking along the dip-guided direction through the phase axis edge delay features to obtain the phase axis tracking results. The extraction module is used to train the two-stage incremental learning model based on the phase axis tracking results and the constructed fine-labeled dataset to obtain the seismic facies spatial representation model, and to input the target deep seismic data into the seismic facies spatial representation model for extraction to obtain the geological age corresponding to the seismic facies spatial representation results. The reconstruction module is used to input the geological age and the seismic profile data into the inversion physical model for processing to obtain elastic parameter data, and to input the elastic parameter data, well logging data, and well-side seismic data into the forward physical model for reconstruction to obtain time-varying wavelet spectrum characteristics. The solution module is used to construct a high-frequency extension inversion objective function based on the time-varying wavelet spectrum characteristics, complete the inversion solution based on the high-frequency extension inversion objective function, obtain a high-resolution reflection coefficient sequence, and convolve the high-resolution reflection coefficient sequence with the time-varying wavelet spectrum characteristics to obtain the high-frequency extended deep seismic data.
9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the high-frequency extension method for the detection signal as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the high-frequency extension method for the detection signal as described in any one of claims 1 to 7.