A seismic frequency domain data prediction method, device, equipment, medium and product

Abstract

The application provides a seismic frequency domain data prediction method, device, equipment, medium and product, and relates to the technical field of seismic frequency domain data prediction. The method comprises the following steps: performing Fourier transform on normalized seismic data to obtain seismic frequency domain data; obtaining first type and second type frequency domain data from a preset frequency threshold and the seismic frequency domain data, performing mask processing on the first type frequency domain data to obtain masked first type frequency domain data; obtaining masked frequency domain data from the masked first type frequency domain data and the second type frequency domain data; constructing a sinusoidal wave basis function in a Fourier domain reparameterization encoder of a pre-trained seismic frequency domain data prediction model, and obtaining frequency domain depth features from the masked frequency domain data and the sinusoidal wave basis function; and obtaining predicted seismic frequency domain data from the frequency domain depth features and a decoder. The frequency domain mask forces the encoder to mine high-frequency information in the missing frequency band and breaks the inertia of the model that preferentially learns low-frequency features.

Description

Technical Field

[0001] This application relates to the field of earthquake frequency domain data prediction technology, and in particular to an earthquake frequency domain data prediction method, apparatus, equipment, medium and product. Background Technology

[0002] With the increasing exploration and development of conventional shallow oil and gas resources, deep, ultra-deep, deep-sea, unconventional oil and gas, complex structural, and complex lithological reservoirs have become the core areas of current and future oil and gas exploration and development. Compared with conventional shallow oil and gas reservoirs, the above-mentioned target areas generally have characteristics such as greater burial depth, complex stratigraphic structure, strong reservoir heterogeneity, severe seismic wave propagation attenuation, and greater difficulty in oil and gas reservoir identification, which place unprecedentedly stringent requirements on the signal-to-noise ratio, resolution, and fidelity of seismic data.

[0003] Currently, existing exploration seismic basic models focus only on spatial structure information, ignoring the inherent frequency dependence of seismic signals. Furthermore, these models tend to prioritize learning low-frequency features, making it difficult to effectively reconstruct high-frequency information. In addition, the limited resolution of seismic data itself makes it insufficient for characterizing complex structures, heterogeneous reservoirs, and thin interbedded lithology, ultimately leading to an imbalance in frequency band reconstruction. Summary of the Invention

[0004] This application provides a method, apparatus, device, medium, and product for earthquake frequency domain data prediction, which solves the technical problem of frequency band reconstruction imbalance caused by the prior art focusing only on spatial structure information and prioritizing the learning of low-frequency features.

[0005] In a first aspect, embodiments of this application provide a method for predicting earthquake frequency domain data, including:

[0006] Seismic data is acquired and flattened to obtain flattened seismic data.

[0007] The flattened seismic data is then normalized to obtain normalized seismic data.

[0008] The normalized seismic data is subjected to Fourier transform to obtain seismic frequency domain data;

[0009] Based on a preset frequency threshold and the earthquake frequency domain data, a first type of frequency domain data and a second type of frequency domain data are obtained, and the first type of frequency domain data is masked to obtain the masked first type of frequency domain data.

[0010] Based on the first type of frequency domain data and the second type of frequency domain data after masking, the masked frequency domain data is obtained;

[0011] A sinusoidal basis function is constructed in the Fourier domain reparameterized encoder of the pre-trained seismic frequency domain data prediction model, and the frequency domain depth feature is obtained based on the masked frequency domain data and the sinusoidal basis function.

[0012] Based on the frequency domain depth features and the decoder, the predicted earthquake frequency domain data is obtained.

[0013] In one possible implementation, the preset frequency threshold includes a first preset frequency threshold and a second preset frequency threshold, and the first type of frequency domain data includes a third type of frequency domain data and a fourth type of frequency domain data.

[0014] Accordingly, obtaining the first type of frequency domain data and the second type of frequency domain data based on the preset frequency threshold and the seismic frequency domain data includes:

[0015] When the frequency corresponding to the earthquake frequency domain data is less than the first preset frequency threshold, the earthquake frequency domain data is determined as the third type of frequency domain data;

[0016] When the frequency corresponding to the earthquake frequency domain data is greater than the second preset frequency threshold, the earthquake frequency domain data is determined as the fourth type of frequency domain data;

[0017] When the frequency corresponding to the seismic frequency domain data is greater than or equal to the first preset frequency threshold and less than or equal to the second preset frequency threshold, the seismic frequency domain data is determined as the second type of frequency domain data.

[0018] In one possible implementation, the step of masking the first type of frequency domain data to obtain masked first type of frequency domain data includes:

[0019] The amplitude of the first type of frequency domain data is set to zero to obtain the masked first type of frequency domain data.

[0020] In one possible implementation, the Fourier domain reparameterization encoder of the pre-trained seismic frequency domain data prediction model includes a preset number of Fourier linear blocks, each of which includes a preset number of Fourier domain reparameterization linear layers.

[0021] Accordingly, the construction of sinusoidal basis functions in the Fourier domain reparameterized encoder of the pre-trained seismic frequency domain data prediction model, and the obtaining of frequency domain depth features based on the masked frequency domain data and the sinusoidal basis functions, includes:

[0022] Sine wave basis functions are constructed in each of the Fourier domain reparameterized linear layers;

[0023] The frequency domain depth features are obtained based on the masked frequency domain data and the sinusoidal basis function.

[0024] In one possible implementation, after obtaining the predicted seismic frequency domain data based on the frequency domain depth features and the decoder, the method further includes:

[0025] In the first fine-tuning stage, the parameters of the Fourier domain reparameterized encoder and the decoder are kept unchanged. The frequency domain data after the first mask is spliced ​​with the first predicted earthquake frequency domain data in the channel dimension to obtain the first spliced ​​data. The downstream task adaptation network is trained based on the first spliced ​​data.

[0026] In the second fine-tuning stage, the parameters of the Fourier domain reparameterized encoder are kept unchanged. The frequency domain data after the second masking is concatenated with the frequency domain data of the second predicted earthquake in the channel dimension to obtain the second concatenated data. The downstream task adaptation network is trained based on the second concatenated data.

[0027] In one possible implementation, the downstream task adaptation network includes a temporal prediction network, the output of which is cascaded with a linear regression head for handling regression-type continuous prediction tasks; or, the output of the temporal prediction network is cascaded with a normalized exponential function activation classification head for handling discrete classification tasks.

[0028] Secondly, embodiments of this application provide a seismic frequency domain data prediction apparatus, comprising:

[0029] The flattening module is used to acquire seismic data and flatten the seismic data to obtain flattened seismic data.

[0030] The preprocessing module is used to normalize the flattened seismic data to obtain normalized seismic data.

[0031] The Fourier transform module is used to perform Fourier transform on the normalized seismic data to obtain seismic frequency domain data.

[0032] The first masking processing module is used to obtain a first type of frequency domain data and a second type of frequency domain data according to a preset frequency threshold and the earthquake frequency domain data, and to perform masking processing on the first type of frequency domain data to obtain the masked first type of frequency domain data.

[0033] The second masking processing module is used to obtain the masked frequency domain data based on the masked first type frequency domain data and the second type frequency domain data.

[0034] The encoding module is used to construct a sinusoidal basis function in the Fourier domain reparameterized encoder of the pre-trained earthquake frequency domain data prediction model, and to obtain frequency domain depth features based on the masked frequency domain data and the sinusoidal basis function.

[0035] The decoding module is used to obtain the predicted earthquake frequency domain data based on the frequency domain depth features and the decoder.

[0036] Thirdly, embodiments of this application provide an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0037] The memory stores computer-executed instructions;

[0038] The processor executes computer execution instructions stored in the memory to implement the earthquake frequency domain data prediction method provided in the first aspect of this application.

[0039] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the seismic frequency domain data prediction method provided in the first aspect of this application.

[0040] Fifthly, embodiments of this application provide a computer program product, including a computer program, which, when executed by a processor, is used to implement the seismic frequency domain data prediction method provided in the first aspect of this application.

[0041] This application provides a method, apparatus, device, medium, and product for earthquake frequency domain data prediction, comprising: acquiring earthquake data and flattening the earthquake data to obtain flattened earthquake data; normalizing the flattened earthquake data to obtain normalized earthquake data; performing a Fourier transform on the normalized earthquake data to obtain earthquake frequency domain data; obtaining a first type of frequency domain data and a second type of frequency domain data based on a preset frequency threshold and the earthquake frequency domain data, and masking the first type of frequency domain data to obtain masked first type of frequency domain data; obtaining masked frequency domain data based on the masked first type of frequency domain data and the second type of frequency domain data; constructing a sine wave basis function in the Fourier domain reparameterized encoder of a pre-trained earthquake frequency domain data prediction model, and obtaining frequency domain depth features based on the masked frequency domain data and the sine wave basis function; and obtaining predicted earthquake frequency domain data based on the frequency domain depth features and the decoder. The above methods achieve the following technical effects: By flattening and normalizing the seismic data, standardized input is provided for the model; Fourier transform of the normalized seismic data yields seismic frequency domain data, shifting the model's learning focus from spatial structure to frequency components, thus solving the problem of traditional models only focusing on spatial structure; the construction and information preservation of the frequency domain mask force the encoder to actively mine high-frequency information in missing frequency bands, breaking the model's inertia of prioritizing low-frequency features and enhancing the ability to reconstruct high-frequency details; and the Fourier domain reparameterized encoder, which introduces sinusoidal basis functions, extracts cross-frequency band depth representations through preset multi-frequency basis functions, explicitly models the frequency dependence characteristics of seismic signals, and combines this with the decoder to output complete predicted frequency domain data, achieving balanced reconstruction and high-fidelity output of frequency band information. Ultimately, this significantly enhances the model's generalization performance and ability to characterize complex structures and thin interbedded lithology on actual seismic data in the work area. Attached Figure Description

[0042] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0043] Figure 1 This is a schematic diagram illustrating an application scenario of a seismic frequency domain data prediction method provided in an embodiment of this application.

[0044] Figure 2 A flowchart illustrating a seismic frequency domain data prediction method provided in an embodiment of this application;

[0045] Figure 3 A schematic diagram illustrating the process of constructing a frequency mask-based Fourier domain reparameterized seismic basic model for embodiments of this application;

[0046] Figure 4A schematic diagram illustrating the training method of the Fourier domain reparameterized seismic basic model based on frequency masking provided in this application embodiment;

[0047] Figure 5 A schematic diagram of the structure of a seismic frequency domain data prediction device provided in an embodiment of this application;

[0048] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0049] Explanation of reference numerals in the attached figures:

[0050] 110 - Terminal; 120 - Server; 801 - Processor; 802 - Memory; 803 - Communication Components; 804 - Bus.

[0051] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0052] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0053] In the embodiments of this application, the terms "first" and "second" are used to distinguish identical or similar items with substantially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply difference. It should be noted that in the embodiments of this application, words such as "exemplary" or "for example" are used to indicate examples, illustrations, or explanations. Any embodiment or design scheme described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design schemes. Specifically, the use of words such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner. In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more.

[0054] It should be noted that the earthquake frequency domain data prediction method provided in this application embodiment is only an example, and the earthquake frequency domain data prediction method may include more or less content.

[0055] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0056] In order to clearly understand the technical solution of this application, the solutions of related technologies are described in detail.

[0057] Currently, existing exploration seismic models focus only on spatial structure information, ignoring the inherent frequency dependence of seismic signals. Furthermore, these models tend to prioritize learning low-frequency features, making it difficult to effectively reconstruct high-frequency information. In addition, the limited resolution of seismic data itself hinders its ability to characterize complex structures, heterogeneous reservoirs, and thin interbedded lithology, ultimately leading to an imbalance in frequency band reconstruction.

[0058] In summary, how to design a technical solution to the frequency band reconstruction imbalance caused by existing technologies focusing only on spatial structure information and prioritizing the learning of low-frequency features is a problem that this application urgently needs to solve.

[0059] Therefore, in view of the above-mentioned technical problems existing in related technologies, this application provides a method, apparatus, equipment, medium and product for earthquake frequency domain data prediction, which aims to effectively solve the imbalance of frequency band reconstruction.

[0060] The following describes the application scenarios of a seismic frequency domain data prediction method, apparatus, device, medium, and product provided in the embodiments of this application. The following application scenarios are merely examples, intended to help those skilled in the art understand the technical content of this application, but do not imply that the embodiments of this application cannot be used in other devices, systems, environments, or scenarios.

[0061] Band-limited seismic data analysis in deep, ultra-deep, deep-sea, and complex tectonic zones: This type of scenario typically relies on the coordinated operation of seismic data acquisition, preprocessing, frequency domain analysis, and prediction models to support reservoir characterization, structural identification, and subsequent development deployment. The seismic frequency domain data prediction method, apparatus, equipment, medium, and product provided in this application's embodiments obtain seismic frequency domain data by performing a Fourier transform on normalized seismic data, shifting the model's learning focus from spatial structure to frequency components. The construction and information preservation of the frequency domain mask force the encoder to actively mine high-frequency information in missing frequency bands, breaking the model's inertia of prioritizing low-frequency features and enhancing the ability to reconstruct high-frequency details.

[0062] Figure 1This is a schematic diagram illustrating an application scenario of a seismic frequency domain data prediction method provided in an embodiment of this application, such as... Figure 1 As shown, it includes: terminal 110 and server 120.

[0063] Terminal 110 can be a user device that initiates the earthquake frequency domain data prediction process. Server 120 is used to control the acquisition of predicted earthquake frequency domain data.

[0064] Figure 2 This is a flowchart illustrating a seismic frequency domain data prediction method provided in an embodiment of this application. The seismic frequency domain data prediction method provided in this embodiment includes the following steps:

[0065] S101. Acquire seismic data and flatten the seismic data to obtain flattened seismic data.

[0066] In this embodiment, the seismic data can be various types of seismic data, such as one-dimensional single-channel seismic data, two-dimensional or three-dimensional post-stack seismic data. The acquired seismic data is uniformly flattened into a single-channel sequence to obtain the flattened seismic data.

[0067] S102. Normalize the flattened seismic data to obtain normalized seismic data.

[0068] In this embodiment, the flattened seismic data undergoes global normalization preprocessing to map the seismic amplitude to a standard interval, thereby eliminating dimensional differences between different work areas and different magnitudes.

[0069] S103. Perform Fourier transform on the normalized seismic data to obtain seismic frequency domain data.

[0070] In this embodiment, the normalized seismic data is converted to the frequency domain by Fast Fourier Transform to obtain seismic frequency domain data, which includes the amplitude corresponding to different seismic frequency points.

[0071] S104. Obtain first type frequency domain data and second type frequency domain data based on preset frequency threshold and seismic frequency domain data, and perform masking processing on the first type frequency domain data to obtain masked first type frequency domain data.

[0072] In this embodiment, the first type of frequency domain data is seismic frequency domain data that requires masking, and the second type of frequency domain data is seismic frequency domain data that does not require masking. After obtaining the first type of frequency domain data, the amplitude of the first type of frequency domain data is masked to obtain the masked first type of frequency domain data.

[0073] S105. Obtain the masked frequency domain data based on the first type of frequency domain data and the second type of frequency domain data after masking.

[0074] In this embodiment, after obtaining the masked first type of frequency domain data, the masked frequency domain data is obtained by combining the masked first type of frequency domain data and the second type of frequency domain data. For example, the seismic frequency domain data before masking includes frequency point A [frequency: 10, amplitude 0.800], frequency point B [frequency: 15, amplitude 0.400], frequency point C [frequency: 25, amplitude 0.200], frequency point D [frequency: 35, amplitude 0.100], and frequency point E [frequency: 45, amplitude 0.300]. The first preset frequency threshold is 20Hz, and the second preset frequency threshold is 40Hz. Then, the masked first type of frequency domain data is frequency point A [frequency: 10, amplitude 0], frequency point B [frequency: 15, amplitude 0.400], frequency point C [frequency: 25, amplitude 0.200], frequency point D [frequency: 35, amplitude 0.100], frequency point E [frequency: 45, amplitude 0.300]. The first preset frequency threshold is 20Hz, and the second preset frequency threshold is 40Hz. B [Frequency: 15, Amplitude 0], frequency point E [Frequency: 45, Amplitude 0], the second type of frequency domain data without masking is frequency point C [Frequency: 25, Amplitude 0.200], frequency point D [Frequency: 35, Amplitude 0.100], the frequency domain data after masking is frequency point A [Frequency: 10, Amplitude 0], frequency point B [Frequency: 15, Amplitude 0], frequency point C [Frequency: 25, Amplitude 0.200], frequency point D [Frequency: 35, Amplitude 0.100], frequency point E [Frequency: 45, Amplitude 0].

[0075] This application proposes a frequency-masking-based Fourier domain reparameterization self-supervised pre-training framework. It implements a directional masking strategy in the frequency domain that preserves high signal-to-noise ratio intermediate frequencies while masking missing low and high frequencies. Furthermore, it innovatively introduces a Fourier domain reparameterization coding layer composed of a sine wave basis and energy balance initialization. Through joint constraints of time-frequency multi-objective self-supervised loss, the model is forced to deeply mine the physical mapping between components across the entire frequency band during the unlabeled pre-training stage, overcoming the inherent frequency bias defect of deep learning from both the weight layer and the input end.

[0076] S106. Construct a sine wave basis function in the Fourier domain reparameterized encoder of the pre-trained seismic frequency domain data prediction model, and obtain the frequency domain depth features based on the masked frequency domain data and the sine wave basis function.

[0077] In this embodiment, the masked frequency domain data is represented by high-dimensional vectorization and feature extraction through the model's built-in linear projection and Fourier coding modules, and the decoder predicts the complete full-band spectrum. This masking strategy forces the model to break away from its dependence on low-frequency high-energy signals and reconstructs broadband seismic signals by deeply mining the intrinsic physical relationship between the mid-frequency band and the low and high frequency bands. This mechanism effectively overcomes the frequency bias defect of traditional deep learning, enabling the model to naturally possess bandwidth broadening and noise reduction capabilities during the label-free pre-training stage.

[0078] The core technology design of this application's embodiments is divided into three main modules. First, it constructs an encoder-decoder architecture based on Fourier domain reparameterization. For processing 3D big data, the scheme adopts a channel-to-channel spatial flattening strategy to efficiently reshape the 3D seismic body into a set of multiple independent 1D time-series waveforms. The core of the pre-trained seismic frequency domain data prediction model uses Fourier linear blocks as the basic feature extraction unit, parameterizing the weights into a linear combination of sine and cosine, explicitly coupling frequency, phase, and amplitude parameters. By constructing a learnable feature mapping based on multi-band sine waves, a physical frequency control mechanism is introduced at the weight level, fundamentally alleviating the frequency bias problem in model training. This overcomes the limitation of existing models that focus primarily on spatial features and consider the essence of the frequency domain less, achieving high-fidelity reconstruction of seismic signals. Second, a time-frequency dual-domain linkage frequency domain mask self-supervised pre-training paradigm is designed to replace the traditional random space masking strategy. In this application embodiment, the input seismic signal is converted to the frequency domain through fast Fourier transform, and masks are set for the low and high frequency bands, retaining only the mid-frequency. In the design of the training loss function, a joint closed-loop constraint system composed of frequency domain mask reconstruction residual, full spectrum residual and time domain shape residual is constructed to force the model to deeply explore the inherent physical coupling law between the known mid-frequency band and the missing frequency bands at both ends, so that the pre-trained basic model naturally has a strong ability to broaden the frequency band and capture macro trends. Third, the industrial migration path of implementing dual-channel feature fusion and progressive unfreezing and layered fine-tuning is adopted. The embodiments of this application adopt a strategy that combines general pre-training and two-step progressive fine-tuning. First, general self-supervised pre-training is completed using massive unlabeled seismic data in the work area to obtain a general interpretation representation of the seismic signal with a wide frequency band. When adapting to different core downstream tasks such as high-resolution seismic processing, wave impedance inversion, and physical property prediction, the feature fusion layer splices the original time-domain seismic traces with the virtual extended frequency seismic traces generated by the extrapolation of the pre-trained model and assembles them into a dual-channel time-series feature map, which is input into a vertical time-series mapping network composed of bidirectional gated recurrent units to fully capture the sedimentary sequence evolution law of deep strata. In the first step of progressive fine-tuning, all weights of the pre-trained basic model are completely frozen, and a differential learning rate is used for joint collaborative fine-tuning of the two networks. That is, the basic model part adopts a minimum learning rate, and the time-series mapping network adopts a standard learning rate, so that the frequency band extrapolation capability and the sparse well constraint task achieve optimal physical synergy.

[0079] The embodiments of this application significantly reduce the model's dependence on well logging tag samples. In work area scenarios with very few well logging tags, its processing accuracy and spatial topology generalization performance are significantly better than traditional small models. At the same time, the solution has strong noise resistance and robustness, can flexibly adapt to multiple types of seismic data and seamlessly integrate into existing industrial exploration processes such as seismic data processing, inversion and interpretation, providing a new intelligent exploration paradigm with high fidelity and strong robustness for the fine characterization of complex three-dimensional reservoirs.

[0080] S107. Based on the frequency domain depth features and the decoder, the predicted earthquake frequency domain data is obtained.

[0081] In this embodiment, the decoder consists of three sets of cascaded linear blocks. Each set of linear blocks uses a basic structure of a linear layer + GELU activation function + linear layer. The first linear block increases the number of feature channels from 1 to 256, the second linear block increases the number of feature channels from 256 to 1024, and the last linear block reduces the number of feature channels from 1024 to 1, ensuring consistent input and output shapes. This maps the frequency domain depth features extracted by the encoder back to the original dimension of the seismic signal, completing the reconstruction of the seismic signal. Higher dimensions capture richer high-dimensional general features, but also increase the number of parameters.

[0082] The input to the first linear block is the frequency domain depth feature output from the Fourier domain reparameterized encoder. For the first linear block:

[0083]

[0084] Among them, W 1,1 b is the weight matrix of the first linear layer of the first linear block. 1,1 W is the bias vector of the first linear layer of the first linear block; 1,2 b is the weight matrix of the second linear layer of the first linear block. 1,2 is the bias vector of the second linear layer of the first linear block; u1 is the output of the first linear block; v is the frequency domain depth feature.

[0085] For the second linear block:

[0086]

[0087] Among them, W 2,1 b is the weight matrix of the first linear layer of the second linear block. 2,1 W is the bias vector of the first linear layer of the second linear block. 2,2 b is the weight matrix of the second linear layer of the second linear block. 2,2 u1 is the bias vector of the second linear layer of the second linear block; u2 is the output of the first linear block, which is the input of the second linear block; u2 is the output of the second linear block.

[0088] For the third linear block:

[0089]

[0090] Among them, W 3,1 b is the weight matrix of the first linear layer of the third linear block. 3,1W is the bias vector of the first linear layer of the third linear block. 3,2 b is the weight matrix of the second linear layer of the third linear block. 3,2 u1 is the bias vector of the second linear layer of the third linear block; u2 is the output of the second linear block, which is the input of the third linear block; u3 is the output of the third linear block, which is the predicted seismic frequency domain data, and u3 is the amplitude of each frequency point in the reconstructed masked frequency domain data.

[0091] All the aforementioned weight matrices and bias vectors are used as trainable parameters. They are assigned initial values ​​through random initialization. Then, during model training, the backpropagation algorithm and gradient descent optimizer are used to minimize the total loss function. These parameters are iteratively updated using a large amount of training data, gradually converging to the optimal value. The final values ​​of these parameters are entirely driven by the training data and the loss function, and are independent of the preset fixed hyperparameters such as the frequency and phase of the basis functions. They are the core learnable part that the network learns from the data.

[0092] The masked frequency domain data is input into the encoder-decoder architecture, and the model outputs reconstructed seismic frequency domain data. The pre-trained optimization objectives are frequency masking loss and full-spectrum reconstruction loss. The original unmasked seismic data and the masked low-frequency and high-frequency information are used as self-supervised labels. The multi-objective joint optimization loss function expression is as follows:

[0093]

[0094]

[0095]

[0096]

[0097] in, For predicted earthquake frequency domain data; y is the actual earthquake frequency domain data, i.e., the earthquake frequency domain data without masking; M is the mask matrix, with the same dimension as the number of frequency points, and elements of 0 or 1, used to indicate which frequency points are valid; MSE() is the mean square error calculation formula, used to calculate the mean square of the difference between the predicted value and the true value; L mask L represents the spectral mask loss values ​​for low and high frequencies predicted by the network compared to the original data. full F represents the full-spectrum reconstruction loss value between the network prediction results and the original data; -1 () represents the inverse Fourier transform; The combined complex spectrum; x true For the actual time-domain signal; N is the total number of time-domain sampling points; L time The earthquake reconstruction loss value is the result of the network prediction compared to the original data; w mFor frequency domain mask loss weighting coefficients; w f The weighting coefficients for the full-spectrum reconstruction loss; w t For time-domain earthquake reconstruction loss weighting coefficients; w m w f and w t Initialize to 1, 0.5, and 0.5 respectively; L total This represents the total loss value.

[0098] The pre-training optimizer uses the AdamW optimizer, with a weight decay coefficient set to 1e-4 and an initial learning rate set to 1e-3. A cosine annealing learning rate decay strategy is adopted, and the total number of pre-training rounds is set to 1000. This completes the general pre-training of the Fourier domain masked autoencoder, resulting in a basic earthquake model with broadband general representation capabilities for seismic signals.

[0099] By flattening and normalizing the seismic data, standardized inputs were provided for the model. Fourier transform of the normalized seismic data yielded frequency domain data, shifting the model's learning focus from spatial structure to frequency components, thus addressing the problem of traditional models focusing solely on spatial structure. The construction and information preservation of the frequency domain mask forced the encoder to actively mine high-frequency information in missing frequency bands, breaking the model's inertia of prioritizing low-frequency features and enhancing its ability to reconstruct high-frequency details. Furthermore, the Fourier domain reparameterized encoder, incorporating sinusoidal basis functions, extracted cross-frequency band depth representations through preset multi-frequency basis functions, explicitly modeling the frequency dependence characteristics of seismic signals. Combined with the decoder's output of complete predicted frequency domain data, this achieved balanced reconstruction and high-fidelity output of frequency band information, ultimately significantly enhancing the model's generalization performance on actual seismic data from real-world work areas and its ability to characterize complex structures and thin interbedded lithology.

[0100] This application provides a method for predicting earthquake frequency domain data, comprising: acquiring earthquake data and flattening the earthquake data to obtain flattened earthquake data; normalizing the flattened earthquake data to obtain normalized earthquake data; performing a Fourier transform on the normalized earthquake data to obtain earthquake frequency domain data; obtaining first-type frequency domain data and second-type frequency domain data based on a preset frequency threshold and the earthquake frequency domain data, and masking the first-type frequency domain data to obtain masked first-type frequency domain data; obtaining masked frequency domain data based on the masked first-type frequency domain data and the second-type frequency domain data; constructing a sine wave basis function in the Fourier domain reparameterized encoder of a pre-trained earthquake frequency domain data prediction model, and obtaining frequency domain depth features based on the masked frequency domain data and the sine wave basis function; and obtaining predicted earthquake frequency domain data based on the frequency domain depth features and the decoder. The above methods achieve the following technical effects: By flattening and normalizing the seismic data, standardized input is provided for the model; Fourier transform of the normalized seismic data yields seismic frequency domain data, shifting the model's learning focus from spatial structure to frequency components, thus solving the problem of traditional models only focusing on spatial structure; the construction and information preservation of the frequency domain mask force the encoder to actively mine high-frequency information in missing frequency bands, breaking the model's inertia of prioritizing low-frequency features and enhancing the ability to reconstruct high-frequency details; and the Fourier domain reparameterized encoder, which introduces sinusoidal basis functions, extracts cross-frequency band depth representations through preset multi-frequency basis functions, explicitly models the frequency dependence characteristics of seismic signals, and combines this with the decoder to output complete predicted frequency domain data, achieving balanced reconstruction and high-fidelity output of frequency band information. Ultimately, this significantly enhances the model's generalization performance and ability to characterize complex structures and thin interbedded lithology on actual seismic data in the work area.

[0101] Based on the above embodiments, this application provides a method for predicting earthquake frequency domain data. In the earthquake frequency domain data prediction method provided in this embodiment, the preset frequency threshold includes a first preset frequency threshold and a second preset frequency threshold; the first type of frequency domain data includes a third type of frequency domain data and a fourth type of frequency domain data; and step S104, obtaining the first type of frequency domain data and the second type of frequency domain data based on the preset frequency threshold and the earthquake frequency domain data, includes the following steps:

[0102] S201. When the frequency corresponding to the seismic frequency domain data is less than the first preset frequency threshold, the seismic frequency domain data is determined as the third type of frequency domain data.

[0103] In this embodiment, the seismic frequency domain data is first divided into three intervals based on the frequency band distribution of the seismic signal: low-frequency band seismic frequency domain data, mid-frequency band seismic frequency domain data, and high-frequency band seismic frequency domain data. A simultaneous low-frequency and high-frequency masking strategy is then adopted for the three frequency bands.

[0104] Optionally, the first preset frequency threshold is 20Hz, which is the low-frequency band mask frequency threshold. No specific restriction is placed on the first preset frequency threshold here. The third type of frequency domain data is low-frequency band frequency domain data. If the frequency corresponding to a certain frequency point in the seismic frequency domain data is less than the first preset frequency threshold, then that frequency point and its corresponding amplitude are determined as the third type of frequency domain data.

[0105] S202. When the frequency corresponding to the seismic frequency domain data is greater than the second preset frequency threshold, the seismic frequency domain data is determined as the fourth type of frequency domain data.

[0106] In this embodiment, optionally, the second preset frequency threshold is 40Hz, which is a high-frequency band mask frequency threshold. No specific limitation is imposed on the second preset frequency threshold here. The fourth type of frequency domain data is high-frequency band frequency domain data. If the frequency corresponding to a certain frequency point in the seismic frequency domain data is greater than the second preset frequency threshold, then that frequency point and its corresponding amplitude are determined as the fourth type of frequency domain data.

[0107] The preset frequency threshold division is based on the typical frequency band distribution characteristics of seismic signals. Low-frequency bands typically contain low-energy noise and macroscopic geological trend information, while high-frequency bands are susceptible to absorption and attenuation by geological formations. Mid-frequency bands, on the other hand, possess a high signal-to-noise ratio and complete structural information. By masking the low and high-frequency bands and retaining only the mid-frequency band as input, the model is forced to actively learn the physical mapping patterns of the missing frequency bands during the decoding process. This breaks the model's dependence on high-energy low-frequency signals and achieves balanced reconstruction of frequency band information.

[0108] S203. When the frequency corresponding to the seismic frequency domain data is greater than or equal to the first preset frequency threshold and less than or equal to the second preset frequency threshold, the seismic frequency domain data is determined as the second type of frequency domain data.

[0109] In this embodiment, if the frequency corresponding to a certain frequency point in the seismic frequency domain data is greater than or equal to the first preset frequency threshold and less than or equal to the second preset frequency threshold, then the frequency point and its corresponding amplitude are determined as the second type of frequency domain data.

[0110] Based on the first and second preset frequency thresholds, the low-frequency, mid-frequency, and high-frequency bands of seismic frequency domain data are clearly distinguished, and the masking processing objects are clearly defined, which can improve the consistency of frequency band correspondence and the stability of data organization in subsequent frequency domain prediction modeling.

[0111] Based on the above embodiments, this application provides a method for predicting seismic frequency domain data. In the seismic frequency domain data prediction method provided in this embodiment, step S104, which involves masking the first type of frequency domain data to obtain masked first type frequency domain data, includes the following steps:

[0112] S301. The amplitude of the first type of frequency domain data is set to zero to obtain the masked first type of frequency domain data.

[0113] In this embodiment, the first type of frequency domain data includes frequency points with frequencies less than a first preset frequency threshold and their corresponding amplitudes, and frequency points with frequencies greater than a second preset frequency threshold and their corresponding amplitudes. The amplitudes corresponding to the frequency points with frequencies less than the first preset frequency threshold and the amplitudes corresponding to the frequency points with frequencies greater than the second preset frequency threshold are both set to zero to obtain the masked first type of frequency domain data.

[0114] By setting the amplitude of the first type of frequency domain data to zero, the corresponding amplitude information can be removed while preserving the frequency distribution structure. This allows the model to learn the mapping relationship of the missing frequency bands mainly based on the remaining frequency domain context during subsequent reconstruction, thereby enhancing the constraint and consistency in the frequency domain completion process and providing a clear mask input basis for subsequent frequency domain deep feature extraction.

[0115] Based on the above embodiments, this application provides a method for predicting earthquake frequency domain data. In the earthquake frequency domain data prediction method provided in this embodiment, the Fourier domain reparameterization encoder of the pre-trained earthquake frequency domain data prediction model includes a preset number of Fourier linear blocks, and each Fourier linear block includes a preset number of Fourier domain reparameterization linear layers. S106 includes the following steps:

[0116] S401. Construct sinusoidal basis functions in each Fourier domain reparameterized linear layer.

[0117] In this embodiment, the pre-trained seismic frequency domain data prediction model consists of two parts: a Fourier domain reparameterization encoder and a decoder. The Fourier domain reparameterization encoder is the core feature extraction unit of the pre-trained seismic frequency domain data prediction model, and it consists of four cascaded Fourier linear blocks. Each Fourier linear block adopts a stacked structure of Fourier domain reparameterization linear layer, GELU activation function, and LayerNorm normalization, that is, each Fourier domain reparameterization linear layer is followed by a GELU activation function and a LayerNorm normalization layer. Each Fourier linear block has four Fourier domain reparameterization linear layers. The first linear layer increases the number of feature channels from 1 to 256, the second linear layer increases the number of feature channels from 256 to 1024, the third linear layer decreases the number of feature channels from 1024 to 256, and the last linear layer decreases the number of feature channels from 256 to 1. This ensures consistent input and output shapes. Higher dimensions capture richer high-dimensional general features, but also increase the number of parameters. The Fourier domain reparameterization linear layer abandons traditional linear mapping and instead constructs learnable feature maps based on sine waves of different frequencies, phases, and amplitudes. Its core expression is:

[0118]

[0119] Among them, f k The preset frequency of the k-th basis function; W is the preset phase of the k-th basis function; k Let A be the input weight matrix of the k-th basis function; kj The amplitude weight of the k-th basis function with respect to the j-th output; x is the amplitude corresponding to the effective frequency domain data; y j For the j-th output value, j is the j-th element in the final output vector of this layer. For example, if the output dimension is 256, then j ranges from 1 to 256.

[0120] S402. Based on the masked frequency domain data and the sinusoidal basis function, obtain the frequency domain depth features.

[0121] In this embodiment, for the first Fourier domain reparameterized linear layer of the first Fourier linear block, the input is the masked frequency domain data. For example, the masked frequency domain data includes frequency point A [frequency: 10, amplitude 0], frequency point B [frequency: 15, amplitude 0], frequency point C [frequency: 25, amplitude 0.200], frequency point D [frequency: 35, amplitude 0.100], and frequency point E [frequency: 45, amplitude 0]. Only the mid-frequency domain data: frequency points C [frequency: 25, amplitude 0.200] and D [frequency: 35, amplitude 0.100] are valid frequency domain data. Assume three basis functions are used, i.e., k=1, 2, 3, the preset frequency f1 of the first basis function is 0.5Hz, and the preset phase... =0.1rad, input weight matrix W1=0.2; preset frequency f2=1.5Hz for the second basis function, preset phase =0.5rad, input weight matrix W2=0.1; preset frequency f3=2.5Hz for the third basis function, preset phase =0.8rad, input weight matrix W3=-0.4.

[0122] Calculate the 256-dimensional output row vector corresponding to the frequency point C [frequency: 25, amplitude: 0.200]. 1) First, calculate the independent variable t of each basis function. k =W k×x. For the first basis function: t1 = W1 × x = 0.2 × 0.200 = 0.04; for the second basis function: t2 = W2 × x = 0.1 × 0.200 = 0.02; for the third basis function: t3 = W3 × x = -0.4 × 0.200 = -0.08. 2) Substitute the values ​​into the sine function and calculate the basis function values. For the activation value of the first basis function: sin(2π×0.5×0.04+0.1)=sin(0.04π+0.1)≈0.2237; For the activation value of the second basis function: sin(2π×1.5×0.02+0.5)=sin(0.06π+0.5)≈0.6346; For the activation value of the third basis function: sin[2π×2.5×(-0.08)+0.8]=sin(-0.4π+0.8)≈-0.4406. 3) Calculate the value of the first output neuron. That is, the first component of the 256-dimensional output row vector: = ×Activation value of the first basis function + ×Activation value of the second basis function+ ×Activation value of the 3rd basis function; Calculate the value of the 2nd output neuron. That is, the second component of the 256-dimensional output row vector: = ×Activation value of the first basis function + ×Activation value of the second basis function+ × The activation value of the 3rd basis function; and so on, calculate the value y of the j-th output neuron. j That is, the 256th component of the 256-dimensional output row vector: y 256 =A 1,256 ×Activation value of the first basis function + A 2,256 ×Activation value of the second basis function + A 3,256 × The activation value of the third basis function. After the above calculation, the frequency point C is mapped to a 256-dimensional row vector: [y1, y2, y3, y4, ..., y 256 Following the same method, the frequency point D is also mapped to a 256-dimensional row vector. Therefore, the tensor shape of the output of the first Fourier domain reparameterized linear layer of the first Fourier linear block is [2, 256], the effective frequency domain data length in the sequence remains unchanged at 2, and the number of features at each position is expanded from 1 to 256.

[0123] For the second Fourier domain reparameterized linear layer of the first Fourier linear block, the input is the output of the first Fourier domain reparameterized linear layer of the first Fourier linear block. Since each frequency point corresponds to a 256-dimensional row vector, W... k Given a 256-dimensional row vector, in calculating W...k When ×x, it is W. k Take the dot product with x. Similarly, if each frequency point corresponds to a 1024-dimensional row vector, then W k This is a 1024-dimensional row vector. Following the same method, frequency points C and D can be mapped to a 1024-dimensional row vector. This process continues until the output of the fourth Fourier domain reparameterized linear layer of the first Fourier linear block is obtained. This output is then used as the input of the first Fourier domain reparameterized linear layer of the second Fourier linear block, and so on. The output of the fourth Fourier domain reparameterized linear layer of the fourth Fourier linear block is then used as the frequency domain depth feature.

[0124] As an alternative implementation, shallower layers (e.g., the first Fourier linear block) or layers with lower feature dimensions are configured with fewer basis functions (e.g., k=3) to control computational load, while deeper layers (e.g., the second Fourier linear block) or layers with increased feature dimensions are configured with more basis functions (e.g., k=8 or k=16) to capture more complex patterns. If different Fourier linear blocks handle significantly different data domains or frequency ranges—for example, the first Fourier linear block handles the original frequency domain signal, while subsequent Fourier linear blocks handle higher-order features—the k values ​​can be adjusted independently for each to adapt to their respective input distributions. Essentially, the k value for each layer is an independent hyperparameter. A candidate range of k values ​​can be set, starting with smaller k values ​​and gradually increasing them, observing the model's performance on the validation set until performance no longer significantly improves. The k value that optimizes model performance is then selected. Alternatively, a more systematic approach can be used, such as cross-validation or generalized cross-validation, to obtain an optimal k value by comparing the generalization error of the model under different k values. In practice, to reduce the difficulty of parameter tuning, an initial k value is often set uniformly first, for example, k=8 for all layers. Then, based on the gradient statistics or output distribution of each layer, specific layers are adjusted by adding or subtracting values.

[0125] The preset frequency is determined by the hyperparameter F. The basis functions are divided into a low-frequency basis function set {-F, -(F-1), ..., -1} and a high-frequency basis function set {1, 2, ..., F}, which are combined into a total of 2F discrete frequency values. These frequency values ​​are fixed before training and do not participate in learning. They are used directly as the preset frequency. The specific values ​​depend on the setting of F. For example, the low-frequency basis function set decreases from -1 to -F, and the high-frequency basis function set increases from 1 to F.

[0126] P phase values ​​are uniformly sampled within the interval [0, 2π] to obtain a preset phase. The phase values ​​are fixed before training and are not updated. Each frequency and each phase combination constitutes a basis function, so the total number of basis functions is M = 2F × P.

[0127] As an optional implementation, F and P values ​​are typically used as hyperparameters. The F value can be estimated based on the frequency range of the input signal to ensure that the low-frequency and high-frequency groups cover the target frequency band, while the P value is set according to the phase diversity requirements. Alternatively, values ​​can be combined within a candidate range, the model trained, and the validation set performance compared to select the optimal combination. For example, the candidate range could be F∈{4,8,16,32}, and P∈{2,4,8,16}. The total number of basis functions M=2F×P directly affects the number of parameters and computational cost. Constraints on the F and P values ​​can also be imposed based on hardware limitations and efficiency requirements. Bayesian optimization or Hyperband algorithms can also be used to efficiently search for the optimal F and P within a specified range. The number of sampling points in the sampling sequence is consistent with the number of input features, and the sampling points start from... Uniform sampling within the interval.

[0128] The introduction of sinusoidal basis functions allows for explicit modeling of frequency, phase, and amplitude parameters, enabling the model to dynamically adjust the weight allocation across different frequency bands. For example, low-frequency basis functions focus on modeling macroscopic trends in the formation, while high-frequency basis functions focus on extracting detailed features. Through linear combinations of multiple frequency basis functions, the model can extract depth representations across frequency bands, thus breaking the traditional model's preference for low-frequency features and achieving accurate reconstruction of high-frequency information, thereby solving the problem of frequency band imbalance.

[0129] For the learnable amplitude weight matrix A i Design an initialization strategy adapted to Fourier basis energies, with weights following a uniform distribution:

[0130]

[0131]

[0132]

[0133] Among them, A i k,j The amplitude weight of the k-th basis function in the i-th Fourier domain reparameterized linear layer with respect to the j-th output; d is the scaling factor; M is the total number of basis functions for the i-th Fourier domain reparameterized linear layer; d n-1 Let B be the input feature dimension of the i-th Fourier domain reparameterized linear layer, i.e., the number of sampling points in the i-th Fourier domain reparameterized linear layer. The input feature dimension of the first Fourier domain reparameterized linear layer is 1, the input feature dimension of the second Fourier domain reparameterized linear layer is 256, the input feature dimension of the third Fourier domain reparameterized linear layer is 1024, and the input feature dimension of the fourth Fourier domain reparameterized linear layer is 256. i k,tLet B be the basis function response value of the k-th basis function in the i-th Fourier domain reparameterized linear layer at the t-th sampling point. i k,t The elements form a Fourier basis matrix; w k Let w be the angular frequency of the k-th basis function in the i-th Fourier domain reparameterized linear layer. k =2πf k ; z is the preset phase of the k-th basis function in the i-th Fourier domain reparameterized linear layer; t Let t be the coordinates of the t-th sampling point.

[0134] For example, for the first Fourier domain reparameterized linear layer, d n-1 =1, that is, from Uniformly sample one point. Let F=2, then the interval is... Uniformly sample one point z1=0. Assume the first Fourier domain reparameterized linear layer has 3 basis functions, the preset frequency of the first basis function is f1=0.5Hz, and the preset phase is... =0.1rad; the preset frequency of the second basis function is f2=1.5Hz, and the preset phase is... =0.5rad; the preset frequency of the third basis function is f3=2.5Hz, and the preset phase is... =0.8 rad. Therefore, the angular frequency of the first basis function = 2π × 0.5 = π ≈ 3.1416, the angular frequency of the second basis function = 2π × 1.5 = 3π ≈ 9.4248, and the angular frequency of the third basis function = 2π × 2.5 = 5π ≈ 15.7080. Regarding the first basis function: For the second basis function: For the third basis function: ,but scaling factor , Therefore, the amplitude weights of the k-th basis function with respect to the j-th output in the first Fourier domain reparameterized linear layer are changed from a uniform distribution. Independent sampling in the middle.

[0135] For the second Fourier domain reparameterized linear layer, d n-1 =256, that is, from 256 points are sampled uniformly. Let F=2, then the interval is... 256 points were uniformly sampled and denoted as z1, z2, ..., z 256Following the above method, the amplitude weights of the k-th basis function with respect to the j-th output in the second Fourier domain reparameterized linear layer can be obtained. Similarly, the amplitude weights of the k-th basis function with respect to the j-th output in the third and fourth Fourier domain reparameterized linear layers can be obtained. This initialization strategy achieves a perfect match between the initial weights and the Fourier basis energies, ensuring that the network can learn frequency features in a balanced manner from the initial stage, thus alleviating the frequency bias problem in model training from the root cause of the weight parameters.

[0136] By performing layer-by-layer reparameterization on the masked frequency domain data using Fourier linear modules, the sinusoidal basis functions enhance the encoder's response to frequency changes. This results in frequency domain depth features that more comprehensively reflect the distribution patterns and hierarchical relationships of the seismic frequency domain data, providing stable input for subsequent frequency domain prediction. By constructing sinusoidal basis functions in each linear layer, the encoder's representation of missing frequency bands becomes more targeted, thereby improving the representational ability and prediction consistency of frequency domain reconstruction in complex seismic scenarios.

[0137] Based on the above embodiments, this application provides a method for predicting seismic frequency domain data. In the seismic frequency domain data prediction method provided in this embodiment, after S107, the following steps are also included:

[0138] S501. In the first fine-tuning stage, the parameters of the Fourier domain reparameterized encoder and decoder are kept unchanged. The frequency domain data after the first mask is spliced ​​with the first predicted earthquake frequency domain data in the channel dimension to obtain the first spliced ​​data. The downstream task adaptation network is trained based on the first spliced ​​data.

[0139] In this embodiment, addressing the pain point of extremely scarce well logging label samples in oil and gas exploration sites, this application breaks through the limitations of traditional end-to-end direct fine-tuning and designs an innovative adaptation process of general pre-training + dual-channel feature fusion + two-stage progressive fine-tuning. Only a very small number of well-seismic matching label samples are needed to retain the broadband general representation capability of the pre-trained model, achieving high-precision adaptation and noise-resistant output for multiple downstream tasks. Specifically, when adapting to downstream tasks, this application embodiment does not directly send the original standardized seismic traces into the task head. Instead, it inputs the original standardized seismic signal into the pre-trained basic model to obtain reconstructed broadband seismic traces, which are then spliced ​​and combined with the original standardized seismic traces in the feature dimension to construct a dual-channel sequence input matrix. Optionally, if low-frequency data is available, a three-channel sequence input downstream task can be constructed through splicing and fusion. This mechanism forces the downstream network to simultaneously capture the difference features between the original band-limited signal and the physical extension signal, providing richer frequency information for downstream tasks such as high-resolution, impedance inversion, and complex reservoir prediction.

[0140] In the first fine-tuning stage, while keeping the parameters of the Fourier domain reparameterization encoder and decoder unchanged, the frequency domain data after the first mask is input to the Fourier domain reparameterization encoder to obtain the first predicted earthquake frequency domain data output by the decoder. The frequency domain data after masking is the original normalized seismic trace, the predicted earthquake frequency domain data is the reconstructed broadband seismic trace, and the first stitched data is the dual-channel sequence input matrix obtained by stitching and combining the reconstructed broadband seismic trace and the original normalized seismic trace in the feature dimension.

[0141] To avoid catastrophic forgetting and gradient oscillation overfitting in small sample scenarios, this application's embodiments design a strict two-step unfreezing training path. The first stage is broadband general feature extraction, which completely freezes the weights of the encoder and decoder of the pre-trained earthquake frequency domain data prediction model, and independently trains the back-end temporal mapping network (e.g., Bi-GRU) with a large initial learning rate (e.g., 1e-3). This stage aims to enable the randomly initialized downstream network to quickly adapt to the temporal waveform distribution and prevent the messy gradients from backpropagation from destroying the general frequency features extracted by the pre-training.

[0142] S502. In the second fine-tuning stage, the parameters of the Fourier domain reparameterized encoder are kept unchanged. The frequency domain data after the second mask is spliced ​​with the frequency domain data of the second predicted earthquake in the channel dimension to obtain the second spliced ​​data. The downstream task adaptation network is trained based on the second spliced ​​data.

[0143] In this embodiment, the second stage of partial unfreezing and physical coordination keeps the encoder of the pre-trained seismic frequency domain data prediction model strictly frozen, and only partially unfreezes the decoder of the pre-trained seismic frequency domain data prediction model. At the same time, a differentiated learning rate mechanism is introduced, which assigns a small learning rate (e.g., 1e-5) to the unfrozen decoder for fine-tuning, and keeps the initial learning rate (e.g., 1e-3) on the task adapter head. This strategy locks in the high-dimensional frequency generalization representation while enabling the generation module to achieve optimal mathematical and physical coordination.

[0144] In the first fine-tuning stage, when the parameters of the Fourier domain reparameterization encoder remain unchanged, the frequency domain data after the second mask is input to the Fourier domain reparameterization encoder to obtain the second predicted earthquake frequency domain data output by the decoder. The frequency domain data after masking is the original normalized seismic trace, the predicted earthquake frequency domain data is the reconstructed broadband seismic trace, and the second stitched data is the dual-channel sequence input matrix obtained by stitching and combining the reconstructed broadband seismic trace and the original normalized seismic trace in the feature dimension.

[0145] Through the above methods, the pre-trained expressive power of the Fourier domain reparameterized encoder and decoder is maintained, and the training process mainly focuses on the downstream task adaptation network, thereby enabling it to more stably adapt to frequency domain inputs at different stages. By training the two sets of spliced ​​data separately, the downstream task adaptation network can obtain more sufficient frequency domain correlation information, thereby improving the adaptation effect to complex seismic frequency domain scenarios.

[0146] The results show that traditional small models, under direct supervised learning, suffer from severe overfitting and lateral discontinuities due to extremely limited training samples and noise interference. In contrast, the fine-tuned results of the frequency mask-based model demonstrate significant advantages. This indicates that the new method and apparatus benefit from the rich seismic prior features learned by the base model during pre-training, effectively resisting strong noise interference and overcoming the bottleneck of small samples. In characterizing geological targets, the profiles processed by the new method not only show significantly improved resolution and sharper, more accurate boundary characterization of complex geological bodies such as channels, but also exhibit better lateral continuity of phase axes. Combined with the broadband general representation capabilities obtained by the base model during pre-training, the advantages of the new method and apparatus are further extended in subsequent quantitative seismic interpretation tasks. Similarly, when applied to three-dimensional impedance inversion, porosity prediction, and lithology identification, the prediction results show higher spatial resolution and consistency with geological patterns.

[0147] This application breaks with conventional feature extraction paradigms. The pre-trained model can provide broadband general characterization capabilities for earthquakes, transforming the pre-trained model into a new paradigm for seismic data bandwidth extension preprocessing. In downstream tasks, the full-band extended seismic traces output by the pre-trained model are deeply stitched together with the original normalized seismic traces at the channel dimension to construct a dual-channel feature sequence, which is then input into a small model. This architecture incorporates both seismic structure and broadband prior constraints for downstream inversion, enabling efficient fine-tuning of downstream tasks such as high-resolution data processing, impedance inversion, and reservoir parameter prediction. It significantly reduces label dependence and improves the processing, inversion accuracy, and characterization capabilities of complex tectonic regions such as thin interlayered structures.

[0148] This application proposes a dual-drive progressive fine-tuning strategy for work areas with few labels. In the first stage, the backbone weights of the pre-trained model are completely locked, focusing on training the downstream lightweight temporal network with a high learning rate, thus protecting the broad-spectrum general physical prior knowledge learned by the large model (those with hundreds of millions of records) from being corrupted. In the second stage, the encoder is kept frozen while the decoder of the pre-trained model is unfrozen, allowing for joint fine-tuning of multiple networks with a low learning rate. This strategy effectively solves the catastrophic forgetting problem during model transfer, significantly reduces reliance on logging labels, and significantly improves the customization adaptability for small and medium-sized work areas.

[0149] To further verify the reliability of this method in actual three-dimensional space, a cross-section comparison including the verification well is presented. The predicted results and the actual logging curves maintain a high degree of consistency in both trend and detail. Table 1 compares the quantitative results of different models in downstream tasks. As shown in Table 1, the results are obtained by using the Pearson Correlation Coefficient (PCC) and the coefficient of determination (R-squared). 2 The results, including the Structural Similarity Index (SSIM), demonstrate that the methods in this application not only overcome the heavy reliance of traditional deep learning on massive amounts of high-precision labels, but also fundamentally improve the inversion accuracy and fine characterization capability for complex structural regions such as thin interlayers. The rows in the table represent metrics, and the columns represent downstream tasks.

[0150] Table 1

[0151]

[0152] Based on the above embodiments, this application provides a method for earthquake frequency domain data prediction. In the earthquake frequency domain data prediction method provided in this embodiment, the downstream task adaptation network includes a time-series prediction network, the output of which is cascaded with a linear regression head for handling regression-type continuous prediction tasks; or, the output of the time-series prediction network is cascaded with a normalized exponential function activation classification head for handling discrete classification tasks.

[0153] In this embodiment, replaceable task adapters are set after the feature fusion module for different downstream tasks. For continuous regression tasks such as high-resolution reconstruction and acoustic impedance inversion, time-series networks such as multi-layer bidirectional gated recurrent units are used to fully explore the sedimentary coupling relationship of the vertical stratigraphic sequence, and the results are output by cascading linear regression heads. For discrete classification tasks such as lithology identification, different time-series networks are switched with classification heads activated by normalized exponential functions to achieve flexible adaptation of the basic model to empower multiple downstream tasks.

[0154] By configuring different output heads at the end of the time-series prediction network, the same network entity can complete matching outputs for both continuous prediction and discrete classification tasks, thereby giving the downstream task adaptation network strong task coverage capabilities. Since regression output and classification output are directly completed by their respective heads, the network's output format can remain consistent with the task labels, thus improving the consistency of downstream task training and analysis and enhancing its applicability in different seismic interpretation tasks.

[0155] Figure 3 This is a flowchart illustrating the process of constructing a frequency-masked Fourier domain reparameterized seismic basic model according to an embodiment of this application. Figure 4 A schematic diagram illustrating the training method of the frequency mask-based Fourier domain reparameterized seismic foundation model provided in this application embodiment is shown below. Figure 3 and Figure 4 As shown, this embodiment first collects synthetic seismic data, actual seismic data, and well logging data to establish a rich database for self-supervised pre-training strategies. Secondly, a pre-training self-supervised learning framework is constructed. By performing Fourier transform on a 60,000-channel pre-training seismic dataset, low-frequency information below the lower absolute bandwidth limit and high-frequency information above the upper absolute bandwidth limit are masked. The unmasked dominant frequency seismic information is used as input data, and the entire frequency band information is reconstructed and recovered through Fourier domain reparameterization. Pre-training is achieved through a self-supervised method of frequency masking reconstruction using the seismic data itself. The high signal-to-noise ratio dominant frequency band data enables the billion-parameter network to understand the frequency band expansion pattern and preserves the pre-trained basic model weight parameters. Finally, a two-step progressive fine-tuning strategy based on frequency masking pre-training is constructed. A downstream task adaptation network is designed. Actual seismic data is input into the pre-training network to obtain initial broadband seismic data. The two are then spliced ​​and fused as new seismic data input to a small model for fine-tuning. Existing well logging data from the actual work area is processed and used as labels for learning. Specifically, the first step involves freezing all weights during the encoding and decoding phases of the base model's pre-training, and then initially training the downstream task adaptation network to form the initial weights of the small model. The second step involves continuing to train the scenario-specific small model through supervised training of downstream tasks, simultaneously updating the overall parameters of both the initial weights of the small model and the weights of the pre-trained base model. This achieves a two-step progressive fine-tuning strategy for downstream tasks and extends it to the entire seismic work area. The downstream tasks in this embodiment include seismic data processing, inversion, and interpretation, specifically high-resolution seismic data processing, elastic parameter inversion, lithology prediction, and physical property parameter prediction. This embodiment is significant in supporting future research on basic seismic models, guiding well location deployment and scale optimization in actual 3D seismic work areas, and improving its exploration and development application effectiveness.

[0156] The embodiments of this application can effectively solve the problems of insufficient frequency modeling capability, serious frequency bias, limited downstream scenario task adaptation, and poor generalization and adaptation capability in scenarios with few labels in existing large seismic models. It can achieve efficient and high-precision adaptation to different tasks in multiple scenarios such as high-resolution seismic processing, wave impedance inversion, porosity prediction, and lithology identification under the condition of a small number of labeled samples. It provides a highly generalizable and practical intelligent seismic detection technology solution for complex oil and gas exploration and development scenarios, and ultimately provides solid technical support for oil and gas well location optimization, reservoir fine characterization, and oil and gas field reserve and production increase.

[0157] This application addresses the limitations of existing large-scale seismic models, which are constrained by the inherent bandwidth of conventional seismic data, insufficient frequency domain modeling capabilities, and prominent frequency bias issues in model training. These limitations result in low vertical geological reservoir identification rates, blurred boundary characterization, limited adaptability to downstream scenarios, and poor spatial generalization of deep learning models under sparse control well conditions, making them prone to overfitting. Therefore, this application aims to provide a new device and equipment for frequency domain masking and Fourier domain reparameterization of seismic fundamental models for band-limited seismic data. This fundamentally optimizes the core industry pain point of weak adaptability of seismic fundamental models to band-limited seismic data, significantly improving the processing, interpretation, and inversion accuracy, efficiency, generalization performance, and industrial applicability of seismic models in complex oil and gas exploration scenarios. This application can significantly broaden the effective bandwidth of seismic data, improve the identification resolution of thin and complex reservoirs, and exhibit strong spatial topological generalization ability and noise resistance in work areas with minimal well constraints. It provides a high-fidelity, robust, intelligent integrated prediction method for processing, inversion, and interpretation of complex 3D reservoirs.

[0158] This application's embodiments overcome the limitations of existing large-scale seismic models that emphasize spatial domains while neglecting frequency domains, deeply integrating frequency awareness into the model architecture design and the entire self-supervised pre-training process. Through a full-band balanced modeling mechanism, it achieves accurate and coordinated representation of low-, mid-, and high-frequency seismic signals, fundamentally enhancing the model's understanding and characterization of the frequency domain of seismic signals. It introduces a Fourier domain reparameterization training method, constructing a full-band control mechanism at the weight level, and employs a frequency domain mask pre-training strategy to systematically eliminate the inherent frequency bias problem of existing models, thoroughly eliminating the inherent frequency bias defect of deep learning from the source of the weights. It constructs a frequency domain mask self-supervised pre-training paradigm, driving the model to learn the inherent coupling laws between different frequency components of seismic signals without any labeled samples, giving the pre-trained model natural bandwidth completion and noise suppression capabilities. Through the pre-trained model learning broadband general representation capabilities, downstream tasks adopt a physical enhancement architecture that splices the pre-trained network with the dual-channel features of the original signal, combined with a two-step progressive fine-tuning strategy, greatly reducing the dependence on well logging label data. With limited sample size, the model demonstrates significantly higher accuracy than existing mainstream models in key tasks such as high-resolution processing, acoustic impedance inversion, and porosity prediction, making it particularly suitable for customized applications in small- to medium-sized work areas and complex geological conditions. The model architecture boasts high compatibility and scalability, flexibly supporting various input formats including one-dimensional seismic traces and two-dimensional / three-dimensional post-stack seismic data. It can be seamlessly integrated into existing commercial seismic data processing, inversion, and interpretation industrial processes. As a universal intelligent foundation for the entire integrated oil and gas exploration and development process, it effectively supports the refined exploration and efficient development of complex oil and gas reservoirs in deep-to-ultra-deep, deep-sea, and unconventional environments.

[0159] Figure 5This is a schematic diagram of the structure of a seismic frequency domain data prediction device provided in an embodiment of this application. Figure 5 As shown, in this embodiment, the seismic frequency domain data prediction device includes:

[0160] The flattening module 701 is used to acquire seismic data and flatten the seismic data to obtain flattened seismic data.

[0161] The preprocessing module 702 is used to normalize the flattened seismic data to obtain normalized seismic data.

[0162] The Fourier transform module 703 is used to perform Fourier transform on the normalized seismic data to obtain seismic frequency domain data.

[0163] The first masking processing module 704 is used to obtain first type frequency domain data and second type frequency domain data according to a preset frequency threshold and seismic frequency domain data, and to perform masking processing on the first type frequency domain data to obtain masked first type frequency domain data.

[0164] The second mask processing module 705 is used to obtain the masked frequency domain data based on the masked first type frequency domain data and the masked second type frequency domain data.

[0165] The encoding module 706 is used to construct a sinusoidal basis function in the Fourier domain reparameterization encoder of the pre-trained seismic frequency domain data prediction model, and obtain frequency domain depth features based on the masked frequency domain data and the sinusoidal basis function.

[0166] The decoding module 707 is used to obtain the predicted earthquake frequency domain data based on the frequency domain depth features and the decoder.

[0167] The seismic frequency domain data prediction device provided in this embodiment can perform... Figure 2 The technical solution of the earthquake frequency domain data prediction method embodiment shown herein, its implementation principle and technical effect are similar to Figure 2 The earthquake frequency domain data prediction method shown in the example is similar and will not be described in detail here.

[0168] Meanwhile, the earthquake frequency domain data prediction device provided by the present invention further refines the earthquake frequency domain data prediction device based on the earthquake frequency domain data prediction device provided in the previous embodiment.

[0169] Optionally, in this embodiment, the preset frequency threshold includes a first preset frequency threshold and a second preset frequency threshold, the first type of frequency domain data includes a third type of frequency domain data and a fourth type of frequency domain data, and the first mask processing module 704 is further used for:

[0170] When the frequency corresponding to the seismic frequency domain data is less than the first preset frequency threshold, the seismic frequency domain data is determined as the third type of frequency domain data; when the frequency corresponding to the seismic frequency domain data is greater than the second preset frequency threshold, the seismic frequency domain data is determined as the fourth type of frequency domain data; when the frequency corresponding to the seismic frequency domain data is greater than or equal to the first preset frequency threshold and less than or equal to the second preset frequency threshold, the seismic frequency domain data is determined as the second type of frequency domain data.

[0171] Optionally, in this embodiment, the first mask processing module 704 is further configured to:

[0172] The amplitude of the first type of frequency domain data is set to zero to obtain the masked first type of frequency domain data.

[0173] Optionally, in this embodiment, the Fourier domain reparameterization encoder of the pre-trained seismic frequency domain data prediction model includes a preset number of Fourier linear blocks, each Fourier linear block including a preset number of Fourier domain reparameterization linear layers, and the encoding module 706 is further configured to:

[0174] Sine wave basis functions are constructed in each Fourier domain reparameterized linear layer; based on the masked frequency domain data and the sine wave basis functions, the frequency domain depth features are obtained.

[0175] Optionally, this embodiment also includes a fine-tuning module, which is used for:

[0176] In the first fine-tuning stage, the parameters of the Fourier domain reparameterization encoder and decoder are kept unchanged. The frequency domain data after the first mask is concatenated with the frequency domain data of the first predicted earthquake in the channel dimension to obtain the first concatenated data. The downstream task adaptation network is trained based on the first concatenated data. In the second fine-tuning stage, the parameters of the Fourier domain reparameterization encoder are kept unchanged. The frequency domain data after the second mask is concatenated with the frequency domain data of the second predicted earthquake in the channel dimension to obtain the second concatenated data. The downstream task adaptation network is trained based on the second concatenated data.

[0177] Optionally, in this embodiment, the downstream task adaptation network in the fine-tuning module includes a temporal prediction network, the output of which is cascaded with a linear regression head for processing regression-type continuous prediction tasks; or, the output of the temporal prediction network is cascaded with a normalized exponential function activation classification head for processing discrete classification tasks.

[0178] Figure 6This is a schematic diagram of an electronic device provided as an embodiment of this application. The electronic device is intended for use with various electronic devices capable of performing seismic frequency domain data prediction methods, such as microcomputers, microcontrollers, and other suitable computers. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0179] like Figure 6 As shown, the electronic device includes at least one processor 801 and a memory 802. The electronic device also includes a communication component 803. The processor 801, memory 802, and communication component 803 are connected via a bus 804.

[0180] In the specific implementation process, at least one processor 801 executes computer execution instructions stored in memory 802, causing at least one processor 801 to execute the seismic frequency domain data prediction method executed on the electronic device side as described above.

[0181] The specific implementation process of processor 801 can be found in the above-described embodiment of the earthquake frequency domain data prediction method. Its implementation principle and technical effect are similar, and will not be repeated here.

[0182] In the above embodiments, it should be understood that the processor 801 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor 801 can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0183] The memory 802 may include high-speed random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0184] Bus 804 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Bus 804 can be divided into address bus, data bus, control bus, etc. For ease of illustration, the bus 804 in the accompanying drawings of this application is not limited to only one bus or one type of bus.

[0185] The above description of the functions implemented by electronic devices and main control devices has introduced the solutions provided in the embodiments of this application. It is understood that, in order to implement the above functions, the electronic device or main control device includes hardware structures and / or software modules corresponding to the execution of each function. Based on the units and algorithm steps of the various examples described in the embodiments disclosed in this application, the embodiments of this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for various specific applications, but such implementation should not be considered beyond the scope of the technical solutions of the embodiments of this application.

[0186] Any of the components, modules, units, parts, methods, and operations described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Alternatively or additionally, any functionality described herein can be executed at least in part by one or more hardware logic components, such as, but not limited to, a central processing unit (CPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), an application-specific standard product (ASSP), a system-on-chip (SoC), a complex programmable logic device (CPLD), a microcontroller unit (MCU), etc. The terms "system," "computing device," or "apparatus" as used herein encompass various means, devices, and machines for processing data, including, for example, one or more programmable processors, computers, SoCs, or combinations thereof. The apparatus may also include code that creates an execution environment for the computer program in question, such as code constituting processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or a combination thereof. The aforementioned computer program (also known as a program, software, software application, app, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for a computing environment.

[0187] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-mentioned seismic frequency domain data prediction method.

[0188] The aforementioned computer-readable storage medium can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0189] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. The readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in application-specific integrated circuits (ASICs). Alternatively, the processor and the readable storage medium can exist as discrete components in an electronic device or a host device.

[0190] The memory 802 is the non-transitory computer-readable storage medium provided by this invention. The non-transitory computer-readable storage medium of this invention stores computer instructions that cause a computer to execute the seismic frequency domain data prediction method provided by this invention.

[0191] The memory 802, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules. The processor 801 executes various functional applications and data processing by running the non-transitory software programs, instructions, and modules stored in the memory 802, thereby implementing the seismic frequency domain data prediction method in the above method embodiments.

[0192] In addition, this embodiment also provides a computer program product, including a computer program, which, when executed by a processor, is used to implement the seismic frequency domain data prediction method of the above embodiment.

[0193] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0194] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0195] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0196] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0197] When integrated units / modules are implemented in hardware, the hardware can be digital or analog circuits. The physical implementation of the hardware structure includes, but is not limited to, transistors and memristors. Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, Graphics Processing Unit (GPU), FPGA, DSP, and ASIC. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random-Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), and Hybrid Memory Cube (HMC).

[0198] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0199] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0200] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0201] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for predicting earthquake frequency domain data, characterized in that, include: Seismic data is acquired and flattened to obtain flattened seismic data. The flattened seismic data is then normalized to obtain normalized seismic data. The normalized seismic data is subjected to Fourier transform to obtain seismic frequency domain data; Based on a preset frequency threshold and the earthquake frequency domain data, a first type of frequency domain data and a second type of frequency domain data are obtained, and the first type of frequency domain data is masked to obtain the masked first type of frequency domain data. Based on the first type of frequency domain data and the second type of frequency domain data after masking, the masked frequency domain data is obtained; A sinusoidal basis function is constructed in the Fourier domain reparameterized encoder of the pre-trained seismic frequency domain data prediction model, and the frequency domain depth feature is obtained based on the masked frequency domain data and the sinusoidal basis function. Based on the frequency domain depth features and the decoder, the predicted earthquake frequency domain data is obtained.

2. The method according to claim 1, characterized in that, The preset frequency thresholds include a first preset frequency threshold and a second preset frequency threshold; the first type of frequency domain data includes a third type of frequency domain data and a fourth type of frequency domain data. Accordingly, obtaining the first type of frequency domain data and the second type of frequency domain data based on the preset frequency threshold and the seismic frequency domain data includes: When the frequency corresponding to the earthquake frequency domain data is less than the first preset frequency threshold, the earthquake frequency domain data is determined as the third type of frequency domain data; When the frequency corresponding to the earthquake frequency domain data is greater than the second preset frequency threshold, the earthquake frequency domain data is determined as the fourth type of frequency domain data; When the frequency corresponding to the seismic frequency domain data is greater than or equal to the first preset frequency threshold and less than or equal to the second preset frequency threshold, the seismic frequency domain data is determined as the second type of frequency domain data.

3. The method according to claim 1, characterized in that, The step of masking the first type of frequency domain data to obtain the masked first type of frequency domain data includes: The amplitude of the first type of frequency domain data is set to zero to obtain the masked first type of frequency domain data.

4. The method according to claim 1, characterized in that, The Fourier domain reparameterization encoder of the pre-trained earthquake frequency domain data prediction model includes a preset number of Fourier linear blocks, and each Fourier linear block includes a preset number of Fourier domain reparameterization linear layers. Accordingly, the construction of sinusoidal basis functions in the Fourier domain reparameterized encoder of the pre-trained seismic frequency domain data prediction model, and the obtaining of frequency domain depth features based on the masked frequency domain data and the sinusoidal basis functions, includes: Sine wave basis functions are constructed in each of the Fourier domain reparameterized linear layers; The frequency domain depth features are obtained based on the masked frequency domain data and the sinusoidal basis function.

5. The method according to claim 1, characterized in that, After obtaining the predicted earthquake frequency domain data based on the frequency domain depth features and the decoder, the method further includes: In the first fine-tuning stage, the parameters of the Fourier domain reparameterized encoder and the decoder are kept unchanged. The frequency domain data after the first mask is spliced ​​with the first predicted earthquake frequency domain data in the channel dimension to obtain the first spliced ​​data. The downstream task adaptation network is trained based on the first spliced ​​data. In the second fine-tuning stage, the parameters of the Fourier domain reparameterized encoder are kept unchanged. The frequency domain data after the second masking is concatenated with the frequency domain data of the second predicted earthquake in the channel dimension to obtain the second concatenated data. The downstream task adaptation network is trained based on the second concatenated data.

6. The method according to claim 5, characterized in that, The downstream task adaptation network includes a temporal prediction network, the output of which is cascaded with a linear regression head for handling regression-type continuous prediction tasks; or, the output of which is cascaded with a normalized exponential function activation classification head for handling discrete classification tasks.

7. A seismic frequency domain data prediction device, characterized in that, include: The flattening module is used to acquire seismic data and flatten the seismic data to obtain flattened seismic data. The preprocessing module is used to normalize the flattened seismic data to obtain normalized seismic data. The Fourier transform module is used to perform Fourier transform on the normalized seismic data to obtain seismic frequency domain data. The first masking processing module is used to obtain a first type of frequency domain data and a second type of frequency domain data according to a preset frequency threshold and the earthquake frequency domain data, and to perform masking processing on the first type of frequency domain data to obtain the masked first type of frequency domain data. The second masking processing module is used to obtain the masked frequency domain data based on the masked first type frequency domain data and the second type frequency domain data. The encoding module is used to construct a sinusoidal basis function in the Fourier domain reparameterized encoder of the pre-trained earthquake frequency domain data prediction model, and to obtain frequency domain depth features based on the masked frequency domain data and the sinusoidal basis function. The decoding module is used to obtain the predicted earthquake frequency domain data based on the frequency domain depth features and the decoder.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the seismic frequency domain data prediction method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the seismic frequency domain data prediction method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, is used to implement the seismic frequency domain data prediction method as described in any one of claims 1 to 6.

Images