Seismic data resolution enhancement method and apparatus based on a physically constrained diffusion model

Abstract

The application discloses a physical constraint-based diffusion model seismic data resolution enhancement method and device, and comprises the following steps: performing low-pass filtering on original seismic data to obtain low-resolution seismic data; performing a forward diffusion process on the original seismic data to obtain noise-added data; constructing a training set by using the low-resolution data, the noise-added data and the original seismic data; constructing a neural network model and defining a total loss function; preliminarily training the neural network model by using the training set; inputting the original seismic data into the latest neural network model, and obtaining preliminary predicted high-resolution seismic data through a physical-guided diffusion model reverse diffusion process; repeatedly iteratively training by reconstructing the training set through the predicted high-resolution seismic data; and inputting the seismic data to be predicted into the neural network model to obtain final predicted high-resolution seismic data. The application has strong generalization ability, effectively accelerates the convergence speed of the model, and realizes double optimization of efficiency and precision.

Description

Technical Field

[0001] This invention belongs to the field of oil and gas resource exploration data processing technology, specifically relating to a method and device for enhancing the resolution of seismic data based on a diffusion model with physical constraints. Background Technology

[0002] As exploration targets become increasingly complex (such as thin interbedded layers and concealed reservoirs), improving seismic data resolution has become an urgent need for oil and gas exploration. Limited by wave propagation attenuation, environmental noise, and acquisition costs, post-processing technologies have become crucial. Existing methods are mainly divided into four categories: (1) Deconvolution, which aims to compress the wavelet to broaden the frequency band, but its effect depends heavily on accurate wavelet estimation and stationarity assumptions, making it difficult to handle non-stationary signals in practice, and unable to recover the attenuated high-frequency components; (2) Inverse Q filtering improves resolution by performing amplitude compensation and phase correction on the absorption attenuation of the formation, but its performance is limited by the Q-value model that is difficult to obtain accurately, and it amplifies noise while compensating for effective signals; (3) Spectral whitening improves high-frequency visibility by balancing the energy of different frequency components, but it may destroy the amplitude and phase relationship, affecting the reliability of subsequent quantitative interpretation, and the parameter settings depend on experience; (4) Deep learning provides a new idea for resolution enhancement. Existing methods can be divided into two categories: one is based on reflectivity inversion, but usually requires the premise of knowing the wavelet; the other is learning the mapping from low frequency to high frequency, but it depends heavily on a large amount of "low-high" resolution paired data as labels, and such paired data is extremely difficult to obtain in practice.

[0003] In recent years, diffusion models have attracted much attention due to their outstanding performance in unsupervised / self-supervised learning. These models typically do not require a large amount of labeled data and the training process is more robust. Therefore, they have great potential for application in the field of seismic data resolution enhancement, but they still face two major challenges: first, relying purely on data-driven methods may generate false information that does not conform to geophysical laws; second, most existing methods still rely on supervised learning frameworks. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the prior art by proposing a method and device for enhancing the resolution of seismic data in a diffusion model based on physical constraints. The core innovations of this method are: First, it employs a novel self-supervised learning strategy, requiring only the low-resolution raw seismic data itself for training, cleverly linking the training objective with the input data itself; Second, it uses seismic wavelets as prior knowledge and integrates the convolutional physics mechanism into the backsampling process of the diffusion model, ensuring that the model's output satisfies the physical constraints of Robinson's convolution law, thus improving training efficiency and prediction accuracy; Third, it uses a physically guided composite loss function, introducing physical constraint loss, sparse constraint loss, and total variational regularization loss in addition to the conventional mean squared error loss, ensuring that the prediction results conform to the statistical laws of the data, satisfy the physical constraints of the convolution law, and possess good signal-to-noise ratio and lateral continuity.

[0005] The above-mentioned objectives of the present invention are achieved through the following technical means:

[0006] A physical constraint-based diffusion model-based method for enhancing seismic data resolution includes the following steps:

[0007] Step 1: Transfer the raw seismic data Low-resolution seismic data is obtained after low-pass filtering. .

[0008] Step 2: Transfer the raw seismic data Noisy data is obtained through a forward diffusion process. .

[0009] Step 3: Construct the training set .

[0010] Step 4: Construct a neural network model Define the total loss function .

[0011] Step 5: Use the training set generated in Step 3 to train the neural network model constructed in Step 4. Training is performed to obtain a pre-trained neural network model. .

[0012] Step 6: Transfer the raw seismic data Input the latest neural network model By using a physical-guided diffusion model to reverse the diffusion process, preliminary high-resolution seismic data predictions were obtained. .

[0013] Step 7: Process the preliminary high-resolution seismic data obtained in Step 6. As new raw earthquake data Construct a new training set For the latest neural network models The network was trained again, resulting in a higher-performing augmentation network. .

[0014] Step 8: Analyze the preliminary high-resolution seismic data output from Step 6. As new raw earthquake data The augmentation network trained in step 7 As the latest neural network model Repeat step 6 to obtain new preliminary high-resolution seismic data. .

[0015] Step 9: Repeat steps 7 and 8 until the initial prediction of high-resolution seismic data meets the requirements, or until the number of iterations is reached, at which point the training will terminate.

[0016] Step 10: Input the earthquake data to be predicted into the neural network model that has terminated training. This allows us to obtain the final high-resolution earthquake data for prediction.

[0017] As mentioned above, noisy data Obtained based on the following formula:

[0018] ,

[0019] Among them, parameters , These are noise variance scheduling parameters. ,noise , For time step number, For time steps, This is the initial step data; during the forward diffusion process, the initial step data... Raw earthquake data .

[0020] As described above, neural network models It includes an input layer, an encoder, a decoder, and an output layer. The encoder includes multiple downsampling layers, and the decoder includes upsampling layers corresponding to the downsampling layer levels. There are skip connections between the downsampling layers and the corresponding upsampling layers. Finally, a set number of consecutive downsampling and upsampling layers are embedded with spatial attention modules.

[0021] As described above, the downsampling layer with embedded spatial attention module sequentially includes a downsampling module, a first convolutional layer A, a second convolutional layer A, a spatial attention module A, and an activation function layer A. The upsampling layer with embedded spatial attention module sequentially includes an upsampling module, a first convolutional layer B, a second convolutional layer B, a spatial attention module B, and an activation function layer B.

[0022] The total loss function as described above Based on the following formula:

[0023] ,

[0024] in, These are the mean squared error loss, physical constraint loss, sparse constraint loss, and total variation regularization loss, respectively. These are the mean squared error loss weight, physical constraint loss weight, sparse constraint loss weight, and total variation regularization loss weight, respectively.

[0025] As described above, mean square error loss Based on the following formula:

[0026] ,

[0027] in, For the labeled data, during the first training in step 5, the labeled data... The raw seismic data will be used as the basis for iterative training in step 7, along with the label data. For the previous iteration of the neural network model Output preliminary high-resolution seismic data , For noisy data and low-resolution seismic data Input to neural network model The obtained output, It is an L2 norm.

[0028] The physical constraint loss Based on the following formula:

[0029] ,

[0030] in, It is a seismic wavelet. This represents the convolution operation.

[0031] The sparse constraint loss Based on the following formula:

[0032] .

[0033] The total variation regularization loss Based on the following formula:

[0034] ,

[0035] in, and Neural network models The output predicts the height and width of high-resolution seismic data. It is a very small positive number. and for The first in Line number The data in the column and the first Line number The data in the column.

[0036] As described above, the reverse diffusion process using a physically guided diffusion model is based on the following steps:

[0037] Calculate new intermediate high-resolution seismic data after adding physical constraints (convolutional physics mechanism). :

[0038] ,

[0039] in, For time steps Corresponding intermediate high-resolution seismic data The previous time step predicted Intermediate high-resolution seismic data, This represents the mean squared error operator. This indicates calculating the gradient. This represents the convolution operation. To control the scaling factor of the correction magnitude, It is a seismic wavelet.

[0040] Based on new intermediate-high resolution seismic data The following formula can be used to estimate the time step at each time step. Corresponding initial step data , Initial step data at time Preliminary predictions of high-resolution seismic data are expected. ,

[0041] ,

[0042] in, For time steps Corresponding new intermediate high-resolution seismic data The corresponding noise.

[0043] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the above-described method for enhancing the resolution of diffusion model seismic data based on physical constraints.

[0044] A computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described method for enhancing the resolution of diffusion model seismic data based on physical constraints.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] 1. Strong generalization ability. This invention employs a novel self-supervised learning strategy, which can complete training using only low-resolution raw seismic data without relying on external labeled data. By establishing a correlation between the training objective and the intrinsic features of the input data, this method effectively improves the model's adaptability to the geological features of different exploration areas, demonstrating significant regional generalization ability.

[0047] 2. High Efficiency. Compared to unsupervised neural network methods without physical constraints, this invention innovatively uses seismic wavelets as prior knowledge and embeds the convolution physical mechanism into the backsampling process of the diffusion model, ensuring that the model's output strictly follows Robinson's convolution law. This design improves prediction accuracy while effectively accelerating the model's convergence speed, achieving a dual optimization of efficiency and accuracy.

[0048] 3. High precision. The seismic data resolution enhancement results of this invention are superior to the traditional sparse pulse deconvolution (SSI) and U-net methods in terms of mean square error and structural similarity. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the process of the present invention.

[0050] Figure 2 The diagram shows the reflection coefficient model of the synthetic data and the synthetic low-resolution seismic data. (a) is a schematic diagram of the reflection coefficient of the Overthrust model, and (b) is a schematic diagram of the synthetic low-resolution seismic data.

[0051] Figure 3 The diagram shows the prediction results of noiseless synthetic data. (a) is a schematic diagram of noiseless low-resolution seismic data, and (b), (c), and (d) are schematic diagrams of the final predicted high-resolution seismic data results of SSI, U-net, and the present invention, respectively.

[0052] Figure 4 The diagram shows the prediction results of noisy synthetic data, where (a) is a schematic diagram of noisy low-resolution seismic data, and (b), (c), and (d) are schematic diagrams of the final predicted high-resolution seismic data results of SSI, U-net, and the present invention, respectively.

[0053] Figure 5The diagram shows the prediction results of actual data, where (a) is a schematic diagram of actual low-resolution seismic data, and (b), (c), and (d) are schematic diagrams of the final predicted high-resolution seismic data results of SSI, U-net, and the present invention, respectively. Detailed Implementation

[0054] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to embodiments. The embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0055] Example 1:

[0056] A physical constraint-based diffusion model seismic data resolution enhancement method, with input data including raw seismic data. And known seismic wavelets. The detailed flowchart is as follows: Figure 1 As shown:

[0057] Step 1: First, process the raw seismic data. Low-resolution seismic data is obtained after low-pass filtering. Low-resolution seismic data This will serve as a physical constraint for the diffusion model.

[0058] Step 2: Transfer the raw seismic data Noisy data is obtained through a forward diffusion process. Forward diffusion process as follows:

[0059] (1)

[0060] in, For time steps, Greater than or equal to 1 This is the initial step data; during the forward diffusion process, the initial step data... Raw earthquake data , These are predefined noise variance scheduling parameters. It is the identity matrix. Representing a normal distribution, it defines the distribution from noisy data. To noisy data The random probability density function, which represents the noisy data. Satisfying the mean is The variance is It follows a normal distribution.

[0061] By using reparameterization techniques, we can derive data from the initial step data. (i.e., raw earthquake data) ), directly calculate any time step Noisy data :

[0062] (2)

[0063] Among them, parameters ,parameter ,noise noise It follows a random distribution with a mean of 0 and a variance of 1. This is the time step number.

[0064] Step 3: Transfer the low-resolution seismic data from Step 1 (As a physical constraint of the diffusion model), the noisy data in step 2 and relatively high-resolution raw seismic data Together, we construct a training set of low-to-high resolution data. Used for the initial training of neural network models, where The data is input into the neural network model in a dual-channel format, i.e., low-resolution seismic data. and noisy data Assembled in the passageway.

[0065] Step 4: Construct the neural network model from Step 3 The basic architecture uses the classic U-net, but the loss function has been improved and innovated. The specific steps are as follows:

[0066] Step 4.1: Neural Network Model The core architecture is a 5-layer U-Net neural network model. It follows the classic encoder-decoder design.

[0067] Neural network model The system comprises an input layer, an encoder, a decoder, and an output layer. The encoder includes multiple downsampling layers, and the decoder includes upsampling layers corresponding to the downsampling layer levels. There are skip connections between the downsampling layers and the corresponding upsampling layers. The encoder is responsible for extracting multi-scale features of seismic data by progressively downsampling through each downsampling layer, while the decoder recovers and generates high-resolution seismic detail features by progressively upsampling through each corresponding upsampling layer.

[0068] Neural network model The specific configuration is as follows:

[0069] In this embodiment, both the downsampling and upsampling layers have 5 levels. The initial number of channels in the first downsampling layer is set to 64. The channel multiplication factors of the feature maps output by the downsampling layers at other levels compared to the feature maps input to the first level are 1, 2, 4, 8, and 8, respectively. This means that as the network depth increases, the number of channels in the feature maps output by the downsampling layers at each level increases to 64, 128, 256, 512, and 512, respectively, enabling the neural network model to capture and process increasingly complex feature representations at different scales. To effectively integrate seismic physics priors into the learning process, this invention designs a physical perception attention mechanism. The core of this mechanism is a spatial attention module, which can dynamically calculate the importance weights of each spatial location in the feature map, thereby guiding the model to focus on areas crucial for resolution enhancement, such as geological boundaries like faults and pinch-out points. This invention specifically embeds this spatial attention module into the downsampling and upsampling layers at levels 3, 4, and 5, ensuring that the neural network model can prioritize processing information most relevant to seismic physics features at the key scale with the highest degree of feature abstraction. Specifically, these downsampling layers sequentially include a downsampling module, a first convolutional layer A, a second convolutional layer A, a spatial attention module A, and an activation function layer A. The spatial attention module A is integrated in a post-processing manner, meaning that after the first and second convolutional layers A at this level complete feature extraction, the output features of the second convolutional layer A are further refined by the spatial attention module A. The dimensions of the output features of the second convolutional layer A are defined as (C1, H1, W1), where C1 is the number of channels in the output features of the second convolutional layer A, and H1 and W1 are the height and width of the output features of the second convolutional layer A, respectively. The spatial attention module operates by calculating a weight map (1, H1, W1) with the same dimensions as the output features of the second convolutional layer A. The values ​​of the weight map are between 0 and 1, representing the importance of each spatial location. Then, the weight map is multiplied element-wise with the output features of the second convolutional layer A. The dimensions of the output features of the spatial attention module A are the same as those of the output features of the second convolutional layer A, both being (C1, H1, W1). The output features of spatial attention module A are then fed into the downsampling layer of this level and the subsequent activation function layer A for further processing. The output features of activation function layer A are then passed to the upsampling layer of the corresponding level of the decoder via skip connections. At the same time, the output features of activation function layer A are transmitted to the next level of downsampling layer.The spatial attention module is embedded in the same position as in the upsampling layer. Specifically, in the upsampling layers at levels 3, 4, and 5, these upsampling layers sequentially contain an upsampling module, a first convolutional layer B, a second convolutional layer B, a spatial attention module B, and an activation function layer B. The features passed from the skip connections and the features passed from the previous level are fused through channels and then sequentially processed by the upsampling module, the first convolutional layer B, and the second convolutional layer B for feature extraction. The output features of the second convolutional layer are then refined by the spatial attention module B. The output features of the spatial attention module B are then fed into the subsequent activation function layer B in the upsampling layer of that level for further processing. The output features of the activation function layer B are then transmitted to the next upsampling layer. The output features of the last upsampling layer (level 1) are input to the output layer, which outputs the predicted high-resolution seismic data.

[0070] Step 4.2: Design the loss function required in the U-net model. Unlike other methods, to ensure that the seismic data generated by the diffusion model maintains physical accuracy while improving resolution, this invention innovatively designs a composite loss function to guide the training of the neural network. Total Loss Function It consists of four parts:

[0071] (3)

[0072] in, These are the mean squared error loss, physical constraint loss, sparse constraint loss, and total variation regularization loss, respectively. These are the mean squared error loss weights, physical constraint loss weights, sparse constraint loss weights, and total variation regularization loss weights. These four loss functions will be described in detail below.

[0073] Mean square error loss , is the fundamental loss of the diffusion model, which aims to minimize the Euclidean distance between the high-resolution prediction and the true label, ensuring that the generated data is close to the true label at the sampling point level, thus guaranteeing the basic reconstruction fidelity of the data. Its definition is as follows:

[0074] (4)

[0075] in, For the labeled data, during the first training in step 5, the labeled data... The raw seismic data will be used as the basis for iterative training in step 7, along with the label data. For the previous iteration of the neural network model Output preliminary high-resolution seismic data , For noisy data and low-resolution seismic data Input to neural network model The obtained output; The noisy data generated for its corresponding forward diffusion process; The corresponding low-resolution seismic data, after low-pass filtering, is simultaneously input into the neural network model as a condition. ; The time step is randomly generated for each iteration.

[0076] Physical constraint loss This is the core innovation of this method. This invention introduces a seismic convolution model as a physical constraint, defined as follows:

[0077] (5)

[0078] in, These are known or estimated seismic wavelets, serving as crucial prior knowledge. This represents the convolution operation. The physical meaning of this loss term is: by applying physical constraint loss... The backpropagation gradient acts on the neural network model Network parameters The entire generation process is effectively guided to a physically reasonable solution space, outputting... All of these will conform to the basic physical principles of seismic wave propagation (i.e., the convolution model). It is an L2 norm.

[0079] Sparse constraint loss Guiding neural network models Approaching geological models with sparse reflectance coefficients can lead to better sparsity in the generated results, contributing to sharper in-phase axes and clearer boundaries, thereby improving visual resolution. This is defined as:

[0080] (6)

[0081] in, for Norm.

[0082] Total variational regularization loss It calculates the average amplitude of the difference between adjacent samples in the horizontal direction of the prediction results. It can reduce the negative impact of random noise or anomalous noise on the enhancement of seismic data resolution and helps to enhance the lateral continuity of seismic signals. It is defined as:

[0083] (7)

[0084] in, and Neural network models The output predicts the height and width of high-resolution seismic data. It is a very small positive number used to prevent gradient explosion and numerical instability. In this embodiment, Set as , and for The first in Line number The data in the column and the first Line number The data in the column.

[0085] Combined with mean square error loss Physical constraint loss sparse constraint loss and total variational regularization loss It can enhance the physical reliability, sparsity, lateral continuity, and noise resistance of prediction results while ensuring stable training.

[0086] Step 5: Use the training set generated in Step 3 to train the neural network model constructed in Step 4. Training is performed to obtain a pre-trained neural network model. At this point, the neural network model has been initially trained. It has basic seismic data resolution enhancement capabilities.

[0087] Step 6: Transfer the raw seismic data Input the latest neural network model By using a physical-guided diffusion model to reverse the diffusion process, preliminary high-resolution seismic data can be obtained. Reverse diffusion process as follows:

[0088] (8)

[0089] in, and They are based on time steps Corresponding intermediate high-resolution seismic data The previous time step predicted Intermediate high-resolution seismic data The mean and variance, based on time steps Corresponding intermediate high-resolution seismic data Predicted initial step data As preliminary prediction of high-resolution earthquake data Maximum time step Corresponding intermediate high-resolution seismic data For the maximum time step Corresponding noisy data The expressions for the mean and variance are shown in (9) and (10) respectively:

[0090] (9)

[0091] (10)

[0092] Among them, parameters ,parameter Similarly , It is a predefined noise variance scheduling parameter, in formula (9) For neural network models Predicted time steps Corresponding intermediate high-resolution seismic data The corresponding noise.

[0093] Therefore, the reverse diffusion process can be explicitly expressed as:

[0094] (11)

[0095] Among them, when time step At that time, random noise Satisfies a normal distribution When time steps At that time, random noise ; This represents the sampling variance of the reverse process. Then, according to formula (12), the initial step data corresponding to each time step can be estimated. , Initial step data corresponding to the time To provide preliminary predictions of high-resolution seismic data .

[0096] (12)

[0097] To ensure that the intermediate high-resolution seismic data corresponding to each time step To conform as closely as possible to the physical relationships of convolution, this invention innovatively introduces a physical constraint in each back-diffusion process: calculating preliminary predictions of high-resolution seismic data. Convolution with seismic wavelet Results and raw seismic data The mean squared error (MSE) and its gradient are used to correct the sampling process. This constraint indicates the update direction, as shown in formula (13): (13)

[0098] in, For the new intermediate high-resolution seismic data with added physical constraints, This represents the mean squared error operator. This indicates calculating the gradient. This represents the convolution operation. A scaling factor is used to control the correction amplitude. This correction mechanism ensures that the backdiffusion process is not only conditioned on low-resolution seismic data but also that the generated results conform to the basic physical laws of seismic wave propagation, thereby guiding the neural network model to reconstruct high-resolution details that are both clear and consistent with geophysical realities. New intermediate-resolution seismic data is then utilized. Get time steps Corresponding new intermediate high-resolution seismic data Based on new intermediate high-resolution seismic data The following formula can be used to estimate the time step at each time step. Corresponding initial step data , Initial step data at time Preliminary predictions of high-resolution seismic data are expected. .

[0099] (14)

[0100] in, For time steps Corresponding new intermediate high-resolution seismic data The corresponding noise.

[0101] Step 7: Process the preliminary high-resolution seismic data obtained in Step 6. As new raw earthquake data A low-pass filter is applied to the new raw seismic data to generate new low-resolution seismic data. Then repeat steps 2 and 3 to obtain new noisy data. Based on new raw seismic data New low-resolution seismic data New noisy data Construct a new training set with more significant resolution mapping relationships. Using the new training set For the latest neural network models The network was trained again, resulting in a higher-performing augmentation network. .

[0102] Step 8: Analyze the preliminary high-resolution seismic data output from Step 6. As new raw earthquake data The augmentation network trained in step 7 As a new neural network model Repeat step 6 to obtain the latest preliminary high-resolution seismic data. This allows for multi-level iterative improvements in resolution.

[0103] Step 9: Repeat steps 7 and 8 until the initial prediction of high-resolution seismic data meets the requirements, or until the number of iterations is reached, at which point the training will terminate.

[0104] Step 10: Input the earthquake data to be predicted into the neural network model that has terminated training. This allows us to obtain the final high-resolution earthquake data for prediction.

[0105] In this embodiment, the method of the present invention is verified using the industry standard model data Overthrust and actual exploration seismic data of a certain block. The reflection coefficients of the Ricker wavelet with a dominant frequency of 80 Hz and the Overthrust model are compared. Figure 2 Convolution is performed on (a) to generate synthetic low-resolution seismic data, such as Figure 2 As shown in (b), this synthesized low-resolution seismic data contains 801 traces, each with 186 time sampling points at a 1 ms interval. Wavelet convolution reduces the resolution of the seismic data, making some weak reflective layers in the reflection coefficient model unclear. A typical example is shown below. Figure 2 The first layer in the red rectangular area in the upper left corner of (b) is located in... Figure 2 In (b), the image is almost indistinguishable due to interference from wavelet sidelobes. Furthermore, some minute geological structures cannot be accurately depicted due to insufficient resolution. For example, several structures in the blue rectangular region around 0.1 seconds are shown... Figure 2 In (b), the boundaries are blurred and difficult to define clearly.

[0106] Figures 3 to 4 The resolution enhancement results of high-resolution seismic data processed by three methods—Sparse Pulse Deconvolution (SSI), U-net, and the present invention—are presented at different noise levels.

[0107] Figure 3 The diagram shows the prediction results of noiseless synthetic data. (a) is a schematic diagram of noiseless low-resolution seismic data, and (b), (c), and (d) are schematic diagrams of the final predicted high-resolution seismic data results of SSI, U-net, and the present invention, respectively.

[0108] Figure 4The diagram shows the prediction results of noisy synthetic data, where (a) is a schematic diagram of noisy low-resolution seismic data, and (b), (c), and (d) are schematic diagrams of the final predicted high-resolution seismic data results of SSI, U-net, and the present invention, respectively.

[0109] No noise environment Figure 3 All methods effectively improve resolution. However, in the area marked by the red rectangle, significant wavelet sidelobe interference still exists in the SSI results, resulting in insufficiently clean phase axes and multiple adjacent weak reflection layers merging together and becoming inseparable. Both U-net and the method proposed in this invention can effectively suppress the sidelobes. For the several small geological structures in the blue rectangle area, the SSI method struggles to identify them. U-net and the method proposed in this invention demonstrate strong resolution improvement capabilities in both of these situations. (High-noise environment (noise level 0.5, ...) Figure 4 The differences between methods are becoming increasingly significant. The SSI method is inadequate for handling multiple adjacent weak reflection layers, as shown in the red rectangular area. In the area marked by the blue rectangle (around 0.1 seconds), the boundaries of several small geological structures in the low-resolution seismic data are blurred and difficult to identify. The U-net method outperforms the SSI method in identifying small geological structures, but its noise resistance is weak, and a significant amount of noise remains in the profile. In contrast, the method proposed in this invention can clearly recover the boundaries of these small structures while effectively suppressing noise and maintaining good stratigraphic continuity.

[0110] Table 1 compares the quantitative results of root mean square error (RMSE) and structural similarity index (SSIM) of different methods under different noise levels. At all noise levels, the method of this invention exhibits the lowest RMSE and the highest SSIM, demonstrating its superior overall performance. Under ideal noise-free conditions, the method of this invention performs close to and significantly better than U-net, indicating the inherent advantage of deep learning methods in resolution enhancement tasks. This advantage increases dramatically with increasing noise levels. When the noise level reaches 0.5, the RMSE of SSI and U-net rises to 0.0184 and 0.0194, respectively, while the method of this invention only slightly increases to 0.0145. Regarding the SSIM index, the results of SSI and U-net show a significant decrease, while the method of this invention maintains a high level of 0.8932. This fully demonstrates that the method of this invention has extremely strong noise resistance and stability, and is better suited to the complex noise environment of real seismic data.

[0111] Table 1

[0112]

[0113] To verify the practicality of the proposed method, it was applied to a set of low-resolution seismic data measured in the field, such as... Figure 5 As shown in Figure (a), the amplitude distribution of the data is uneven, with weaker signal strength in both shallow and deep layers, and the phase axis morphology is less clear than in the middle layer. Based on this low-resolution seismic data measured in the field, seismic wavelets were first estimated using commercial software, and then resolution enhancement was performed sequentially using SSI, U-net, and the method of this invention. The results are shown in Figure (a). Figure 5 In (b), (c), and (d) of the figure, all three methods significantly improved the resolution of the original seismic data and brought the amplitude of the phase axes at different depths closer to uniformity. However, there are significant differences in the detail representation among the three methods. The red solid rectangles in the figure mark areas with thinner strata: the SSI and U-net methods cannot accurately distinguish adjacent layers in these areas, and the layer continuity is poor; in contrast, the method of this invention can obtain clear and continuous layer information, demonstrating its stronger ability in recovering information from thin layers. The red dashed rectangles in the figure mark deep and complex areas. The original data in this area has weak amplitude and blurred phase axis morphology. After processing by the three methods, the amplitude of the phase axes in this area is effectively balanced. However, in terms of improving phase axis resolution, the phase axis orientation and morphology obtained by the method of this invention are clearer, indicating that this method has a more significant advantage in handling complex geological structures.

[0114] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0115] Example 2:

[0116] In this embodiment, a computer device is also provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0117] Example 3:

[0118] In this embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0119] Example 4:

[0120] In this embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0121] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for enhancing the resolution of seismic data based on a diffusion model with physical constraints, characterized in that, Includes the following steps: Step 1: Transfer the raw seismic data Low-resolution seismic data is obtained after low-pass filtering. ; Step 2: Transfer the raw seismic data Noisy data is obtained through a forward diffusion process. ; Step 3: Construct the training set ; Step 4: Construct a neural network model Define the total loss function ; Step 5: Use the training set generated in Step 3 to train the neural network model constructed in Step 4. Training is performed to obtain a pre-trained neural network model. ; Step 6: Transfer the raw seismic data Input the latest neural network model By using a physical-guided diffusion model to reverse the diffusion process, preliminary high-resolution seismic data predictions were obtained. ; Step 7: Process the preliminary high-resolution seismic data obtained in Step 6. As new raw earthquake data Construct a new training set For the latest neural network models The network was trained again, resulting in a higher-performing augmentation network. ; Step 8: Analyze the preliminary high-resolution seismic data output from Step 6. As new raw earthquake data The augmentation network trained in step 7 As the latest neural network model Repeat step 6 to obtain new preliminary high-resolution seismic data. ; Step 9: Repeat steps 7 and 8 until the initial prediction of high-resolution seismic data meets the requirements, or until the number of iterations is reached, at which point the training will terminate. Step 10: Input the earthquake data to be predicted into the neural network model that has terminated training. This allows us to obtain the final high-resolution earthquake data for prediction.

2. The method for enhancing the resolution of diffusion model seismic data based on physical constraints according to claim 1, characterized in that, The noise-added data Obtained based on the following formula: ， Among them, parameters , These are noise variance scheduling parameters. ,noise , For time step number, For time steps, This is the initial step data; during the forward diffusion process, the initial step data... Raw earthquake data .

3. The method for enhancing the resolution of diffusion model seismic data based on physical constraints according to claim 1, characterized in that, The neural network model It includes an input layer, an encoder, a decoder, and an output layer. The encoder includes multiple downsampling layers, and the decoder includes upsampling layers corresponding to the downsampling layer levels. There are skip connections between the downsampling layers and the corresponding upsampling layers. Finally, a set number of consecutive downsampling and upsampling layers are embedded with spatial attention modules.

4. The method for enhancing the resolution of diffusion model seismic data based on physical constraints according to claim 3, characterized in that, The downsampling layer with embedded spatial attention module sequentially includes a downsampling module, a first convolutional layer A, a second convolutional layer A, a spatial attention module A, and an activation function layer A. The upsampling layer with embedded spatial attention module sequentially includes an upsampling module, a first convolutional layer B, a second convolutional layer B, a spatial attention module B, and an activation function layer B.

5. The method for enhancing the resolution of diffusion model seismic data based on physical constraints according to claim 1, characterized in that, The total loss function Based on the following formula: ， in, These are the mean squared error loss, physical constraint loss, sparse constraint loss, and total variation regularization loss, respectively. These are the mean squared error loss weight, physical constraint loss weight, sparse constraint loss weight, and total variation regularization loss weight, respectively.

6. The method for enhancing the resolution of diffusion model seismic data based on physical constraints according to claim 5, characterized in that, The mean square error loss Based on the following formula: ， in, For the labeled data, during the first training in step 5, the labeled data... The raw seismic data will be used as the basis for iterative training in step 7, along with the label data. For the previous iteration of the neural network model Output preliminary high-resolution seismic data , For noisy data and low-resolution seismic data Input to neural network model The obtained output, It is an L2 norm; The physical constraint loss Based on the following formula: , in, It is a seismic wavelet. This represents the convolution operation; The sparse constraint loss Based on the following formula: ， The total variation regularization loss Based on the following formula: ， in, and Neural network models The output predicts the height and width of high-resolution seismic data. It is a very small positive number. and for The first in Line number The data in the column and the first Line number The data in the column.

7. The method for enhancing the resolution of diffusion model seismic data based on physical constraints according to claim 6, characterized in that, The physical-guided diffusion model reverse diffusion process is based on the following steps: Calculate new intermediate high-resolution seismic data with added physical constraints. : ， in, For time steps Corresponding intermediate high-resolution seismic data The previous time step predicted Intermediate high-resolution seismic data, This represents the mean squared error operator. This indicates calculating the gradient. This represents the convolution operation. To control the scaling factor of the correction magnitude, For seismic wavelets, Based on new intermediate-high resolution seismic data The following formula can be used to estimate the time step at each time step. Corresponding initial step data , Initial step data at time Preliminary predictions of high-resolution seismic data are expected. , ， in, For time steps Corresponding new intermediate high-resolution seismic data The corresponding noise.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the physical constraint-based diffusion model seismic data resolution enhancement method according to any one of claims 1 to 7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the physical constraint-based diffusion model seismic data resolution enhancement method as described in any one of claims 1 to 7.

Images