Time sequence seismic data high-resolution processing method based on improved U-shaped network

Abstract

The invention discloses a time sequence seismic data high-resolution processing method based on an improved U-shaped network, and belongs to the technical field of seismic image processing, and the method comprises the steps: generating low-resolution sample data and high-resolution label data through employing a forward modeling method; adding the SE block into the U-Net network, and simultaneously realizing the functions of a residual module and an attention mechanism to obtain an SEU-Net network of the improved U-type network; an MSE-MAE loss function is constructed, and an SEU-Net network is trained based on the sample data, the label data and the MSE-MAE loss function; and obtaining high-resolution three-dimensional seismic data based on the trained SEU-Net network. According to the method, the problems of low resolution and multiple noises of seismic data are solved, the frequency band width is widened, the main frequency of the seismic data is improved, detail texture information is enriched, meanwhile, a good suppression effect on the noises is achieved, and the resolution of the seismic data is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic image processing, and particularly relates to a time-series seismic data high-resolution processing method based on an improved U-shaped network. BACKGROUND

[0002] Due to the continuous deepening of seismic exploration and the limitations of acquisition technology, seismic data often has low resolution and contains a large amount of noise, which brings great challenges to subsequent seismic data interpretation work, and therefore improving the resolution of seismic data is the main problem currently faced. In order to improve the resolution, the traditional methods are mainly divided into three categories: one is the deconvolution method, which widens the frequency band by compressing the wavelet; the second is the inverse Q filtering method, which focuses on amplitude compensation; and the third is the spectral whitening method, which balances the high and low frequency energy in the time-frequency domain to improve the resolution.

[0003] In recent years, with the continuous development of artificial intelligence, the deep learning technology has made obvious progress in the field of image super-resolution processing, and some geophysicists have introduced the super-resolution algorithm into the high-resolution processing of seismic data. According to the different network structures, it can be roughly divided into three categories: convolutional neural network, recurrent neural network and generative adversarial network. These methods learn the effective features and mapping relationship in the seismic data through data-driven way, and at the same time, some can also consider noise suppression, gradually becoming a research hotspot in this field.

[0004] However, these methods are only based on a single network structure, and there are obvious deficiencies in solving the problems of low resolution and multiple noises of seismic data. SUMMARY

[0005] In view of the above deficiencies in the prior art, the time-series seismic data high-resolution processing method based on the improved U-shaped network provided by the present application solves the problems of low resolution and multiple noises of seismic data.

[0006] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows: The present application provides a time-series seismic data high-resolution processing method based on an improved U-shaped network, comprising: S1: generating low-resolution sample data and high-resolution label data by using a forward simulation method; S2: adding a SE block to the U-Net network to realize the functions of the residual module and the attention mechanism, and obtaining an SEU-Net network of the improved U-shaped network; S3: constructing an MSE-MAE loss function based on mean square error and mean absolute error, and training the SEU-Net network based on the low-resolution sample data, the high-resolution label data and the MSE-MAE loss function; S4: Input the low-resolution 3D seismic data into the trained SEU-Net network to obtain high-resolution 3D seismic data.

[0007] Furthermore, the method of generating low-resolution sample data and high-resolution label data using forward modeling includes: S101: Randomly generated reflection coefficient model; S102: Convolution processing is performed using high-frequency Ricker wavelet, low-frequency Ricker wavelet and reflection coefficient model to obtain high-resolution tag data and low-resolution data respectively; The expression for the Reichschild wavelet is:

[0008] In the formula, Represents the Reich wavelet, This represents the exponential function. Indicates time, This indicates the dominant frequency of the Rack wavelet; The calculation formula for the convolution process is as follows:

[0009]

[0010] In the formula, and These represent low-resolution data and high-resolution label data, respectively. and Representing the low-frequency and high-frequency Ricker wavelets, respectively. Represents the reflection coefficient model; S103: Add Gaussian noise of varying degrees to low-resolution data to obtain low-resolution sample data.

[0011] Furthermore, the SEU-Net network includes: an encoding part, a decoding part, a skip connection layer, and an output layer; The encoding part is used for feature extraction and includes three downsampling blocks. Each downsampling block consists of two convolutional layers, two SE blocks, and a pooling layer with a 2×2 kernel. The convolutional layers are used to extract features, the SE blocks extract representative features and reduce feature loss during extraction, and the pooling layer reduces feature size. The convolutional layers include 3×3 convolutional kernels, ReLU activation functions, and batch normalization layers. The input of the convolutional layers is padded around the 3×3 convolutional kernels, and the ReLU activation function is used to output the padded data, enabling the data to obtain non-linear relationships between data. The batch normalization layer is used for anti-fitting processing. The SE blocks include compression and activation parts. The decoding section is used for feature fusion and includes three upsampling blocks. Each upsampling block includes two convolutional layers, two SE blocks, and a deconvolutional layer with a 2×2 kernel. The convolutional layers fuse the extracted features, the SE blocks improve the recognition of representative features and provide fusion speed, and the deconvolutional layer restores the feature size. The convolutional layers include a 3×3 kernel, a ReLU activation function, and a batch normalization layer. Each SE block includes a compression part and an activation part. The skip connection layer is located between each downsampling block and upsampling block, and is used to connect some features in the downsampling block with the features in the corresponding upsampling block; The output layer includes a 1×1 convolutional layer used to output the obtained prediction data.

[0012] Furthermore, the SE block includes: When the SE block is used as a channel attention mechanism, it includes a compression part and an activation part. In the compression part, the SE block uses global pooling to compress the feature map of each channel into a single value. In the activation part, the SE block uses a small fully connected network to learn the weights of each channel. The weights are used to reweight the compressed feature map to obtain a representative feature map. When the SE block is used as a residual module, the feature weights are dynamically adjusted through the SE block to alleviate the gradient vanishing and gradient exploding problems in the SEU-Net network.

[0013] Furthermore, training the SEU-Net network based on low-resolution sample data, high-resolution label data, and the MSE-MAE loss function includes: S301: Input low-resolution sample data into the downsampling block of the SEU-Net network for feature extraction; S302: The upsampling block fuses the extracted features and connects some features in the downsampling block with the corresponding features in the upsampling block through a skip connection layer. The output layer then outputs high-resolution prediction data. S303: Use the MSE-MAE loss function to evaluate the difference between the predicted data and the high-resolution labeled data, and adjust the parameters of the SEU-Net network based on the difference.

[0014] Furthermore, the expression for the MSE-MAE loss function is as follows:

[0015]

[0016]

[0017] in, Mean square error, The mean absolute error, As a weighting factor, The total number of samples, and These are high-resolution labeled data and data predicted by the SEU-Net network, respectively.

[0018] The beneficial effects of this application are: This application provides a high-resolution processing method for time-series seismic data based on an improved U-Net network. The method uses the U-Net network as the main structure and adds SE blocks after the convolutional layers of the U-Net network, simultaneously implementing the functions of residual modules and attention mechanisms. A loss function combining mean squared error and mean absolute error is used for training, thereby constructing a SEU-Net network to improve seismic data resolution and suppress noise. The SEU-Net network is used to establish a nonlinear relationship between low-resolution and high-resolution seismic data. This method can broaden the bandwidth, increase the dominant frequency of seismic data, enrich detailed texture information, effectively protect low-frequency signals, effectively broaden high-frequency signals, and also effectively suppress noise, thus effectively improving the resolution of seismic data. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0020] Figure 1 This is a flowchart illustrating a high-resolution processing method for time-series seismic data based on an improved U-shaped network, as provided in an embodiment of this application.

[0021] Figure 2 This is a schematic diagram of low-resolution sample data and high-resolution tag data provided for an embodiment of this application.

[0022] Figure 3 This is a schematic diagram of a SEU-Net network structure provided in an embodiment of this application.

[0023] Figure 4 This is a schematic diagram of an SE block as a channel attention mechanism provided in an embodiment of this application.

[0024] Figure 5 This is a schematic diagram of an SE block as a residual module provided in an embodiment of this application.

[0025] Figure 6This is a schematic diagram of the variation of the MSE-MAE loss function provided in an embodiment of this application.

[0026] Figure 7 This is a comparison diagram of different network models provided in an embodiment of this application.

[0027] Figure 8 This is a comparison diagram of seismic profiles of four types of data provided in an embodiment of this application.

[0028] Figure 9 A spectrum comparison diagram of a U-Net model and a SEU-Net network provided in an embodiment of this application.

[0029] Figure 10 This is a comparison diagram of seismic time slices of four types of data provided in an embodiment of this application.

[0030] Figure 11 This is a comparison diagram of four types of seismic profile faults provided in an embodiment of this application.

[0031] Figure 12 This is a comparison chart of actual data provided for an embodiment of this application.

[0032] Figure 13 This is a comparison chart of the spectrum of actual data before and after seismic data processing via the SEU-Net network, provided as an embodiment of this application.

[0033] Figure 14 This image shows a comparison between a low-resolution seismic data time slice and a high-resolution seismic data time slice obtained after processing through the SEU-Net network, as provided in an embodiment of this application. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of this application.

[0035] Example 1: Convolutional neural networks extract feature information through convolutional layers for learning. Yuan et al. proposed a sequence convolutional neural network model, using an encoder-decoder structure as the main structure and adding a residual module in the middle of the network to achieve high-resolution processing of one-dimensional seismic data. Li et al. proposed a deep convolutional neural network that simultaneously achieves high-resolution processing and noise suppression of seismic data, and verified it with theoretical and actual data, showing good application results. Sun et al. used the U-Net network for high-resolution processing of seismic data and proposed using a pseudo-reflection coefficient model to create a dataset that is closer to real data. Jo et al. also used the U-Net network for high-resolution processing of seismic data, but they considered the non-stationarity of wavelets when creating the dataset, making the dataset closer to real data. Recurrent neural networks (RNNs) extract feature information through recurrent kernels for learning. Sui et al. used LSTM networks to establish a mapping relationship between synthetic seismic records and fitted seismic records. The fitted seismic records were generated by convolution of a pseudo-reflection coefficient sequence, which features a large weak reflection coefficient and a small strong reflection coefficient, thereby enhancing low-frequency signals. Zhu et al. proposed a bidirectional LSTM network based on LSTM networks to achieve accurate reservoir prediction by improving the resolution of seismic data. Chen et al. went further, introducing a U-Net network into LSTM networks, proposing a new network combining U-Net and LSTM networks, which showed good application results. Generative adversarial networks (GANs) achieve mutual learning through adversarial interaction. Halpert applied GANs to three-dimensional data, and verification showed that this method can effectively improve the resolution of seismic data. Chen et al. used a data-driven approach to generate a large dataset and established a nonlinear relationship between seismic data and strong reflection coefficients through GANs to output strong reflection coefficients.

[0036] Based on this, embodiments of this application provide a high-resolution processing method for time-series seismic data based on an improved U-shaped network, which can be found in [reference needed]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating a high-resolution processing method for time-series seismic data based on an improved U-shaped network, according to an embodiment of this application. The method includes: S1: Use forward modeling to generate low-resolution sample data and high-resolution label data.

[0037] In one embodiment of this application, since deep learning is a data-driven learning method, it requires a large amount of data to train the model so that the model can better learn the inherent patterns and regularities of the data. Therefore, this application requires a large amount of data to train and adjust the model. At the same time, the quality of the dataset directly determines the model's performance. Generally speaking, the more representative the dataset, the better the model's performance, because a high-quality dataset is likely to be closer to the real situation, allowing the model to learn the diversity of data, thereby enhancing the model's generalization ability. This application starts with one-dimensional data to achieve prediction of three-dimensional seismic data volumes. Due to the lower data dimensionality, the computational resources and storage space required for one-dimensional data processing are usually less than those for two-dimensional data processing, reducing computational complexity and effectively improving computational efficiency.

[0038] This application generates low-resolution sample data and high-resolution label data using forward modeling. First, a reflection coefficient model is randomly generated; then, high-frequency and low-frequency Ricker wavelets are convolved with the reflection coefficient model to obtain high-resolution label data and low-resolution data, respectively. To make the synthesized data closer to real data, different levels of Gaussian noise are added to the low-resolution data to obtain low-resolution sample data, thus obtaining training data pairs (low-resolution sample data and high-resolution label data).

[0039] In the specific forward modeling, the Reich wavelet with a dominant frequency of 30Hz and 50Hz was convolved with the reflection coefficient model to obtain low-resolution sample data and high-resolution label data, respectively. The Reich wavelet expression is as follows:

[0040] In the formula, Represents the Reich wavelet, This represents the exponential function. Indicates time, This indicates the dominant frequency of the Reichschild wavelet.

[0041] The formula for calculating the convolution is:

[0042]

[0043] In the formula, and These represent low-resolution data and high-resolution label data, respectively. and Representing the low-frequency and high-frequency Ricker wavelets, respectively. This represents the reflection coefficient model.

[0044] By convolving with different reflection coefficient models, 10,000 low-resolution data points and 10,000 high-resolution label data points were obtained. Different levels of Gaussian noise (mean 0, variances 0.01, 0.05, and 0.1, respectively) were added to the low-resolution data to obtain the low-resolution label data. This generated 10,000 pairs of seismic data, each 256 bytes long. This data will serve as the dataset for this application and will be used for subsequent applications, such as... Figure 2 As shown, the first is low-resolution sample data with Gaussian noise added with a mean of 0 and a variance of 0.01, the second is low-resolution sample data with Gaussian noise added with a mean of 0 and a variance of 0.05, the third is low-resolution sample data with Gaussian noise added with a mean of 0 and a variance of 0.1, and the fourth is high-resolution label data.

[0045] S2: Add the SE block to the U-Net network, and implement the functions of residual module and attention mechanism to obtain the SEU-Net network with improved U-shaped network.

[0046] In one embodiment of this application, U-Net is a convolutional neural network architecture widely used for biomedical image segmentation. It was initially proposed in a 2015 paper by Olaf Ronneberger, Philipp Fischer, and Thomas Brox to address the problem of cell image segmentation in medical imaging. With subsequent development, it has also found wide applications in image super-resolution processing, remote sensing image processing, and autonomous driving. The uniqueness and practicality of the U-Net network lies primarily in its distinctive U-shaped structure, consisting of an encoder path and a symmetrical decoder path, resembling the letter "U". This design helps the network learn features at multiple scales and enables U-Net to perform well even on small datasets. Skip connections between the encoder and decoder pass features from the encoder to the corresponding layer in the decoder, achieving feature connectivity. This design allows the network to retain detailed information at deep layers, promoting accurate reconstruction of details and effectively improving prediction accuracy.

[0047] This application introduces the U-Net network into the field of high-resolution seismic data processing, and adds SE blocks to the network structure to enhance the network's feature representation capabilities, and applies it to the processing of one-dimensional seismic data. The network structure of this application is as follows: Figure 3As shown, it consists of four parts: an encoding part, a decoding part, a skip connection layer, and an output layer. The encoding part is used for feature extraction and mainly includes three downsampling blocks. Each downsampling block includes two convolutional layers and two SE blocks. Each convolutional layer uses a 3×3 kernel, padded with 1s to maintain its size, and then uses the ReLU activation function for output, enabling the network to learn the non-linear relationships between data. Batch normalization layers are used to process the data and prevent overfitting. Each SE block consists of compression and activation parts, which helps the network extract more representative features. Since the SE blocks are all after the convolutional layers, they can also act as residual modules, reducing feature loss during feature extraction. After the two convolutional layers and the two SE blocks, there is a pooling layer with a 2×2 kernel to reduce the size of the features, thus completing one downsampling operation. The decoding section, used for feature fusion, mainly comprises three upsampling blocks. Each upsampling block consists of two convolutional layers and two SE blocks. Each convolutional layer uses a 3×3 kernel, padded with 1s to maintain its size, and then uses the ReLU activation function for output, enabling the network to learn non-linear relationships between data. Batch normalization layers are used to process these relationships and prevent overfitting. Each SE block consists of compression and activation parts, which helps improve the network's recognition of representative features. Since SE blocks are located after convolutional layers, they also act as residual modules, increasing the fusion speed during feature fusion. Immediately following the two convolutional layers and two SE blocks is a 2×2 inverted convolutional layer to restore the feature size, thus completing one upsampling operation. A skip connection connects each downsampling and upsampling block, linking features from the downsampling path to features in the corresponding upsampling block, mitigating the vanishing or exploding gradient problem. The input data undergoes three downsampling operations, three skip connections, and three upsampling operations before finally being output through a 1×1 convolutional layer to obtain the output data.

[0048] The Squeeze-and-Excitation (SE) block is essentially a mechanism for enhancing channel attention in deep neural networks. Originally proposed by Jie Hu et al. in 2018, it achieved significant performance improvements in image classification tasks. Its core idea is to dynamically learn importance weights on each channel so that the network can better focus on the features most relevant to the current task. Due to its wide applicability, it can be easily integrated into existing convolutional neural network architectures to enhance their performance; therefore, this application considers incorporating it into the network.

[0049] The SE block, as a channel attention mechanism, has the following specific structure: Figure 4 As shown, it is mainly divided into two parts: compression (Squeeze): in this part (Figure 4 In the X to U section, the SE block compresses the feature map of each channel into a single value through a global pooling operation (usually global average pooling). This is done to reduce computation and focus attention on the relationships between channels. Excitation: In this section ( Figure 4 From U to In the SE block, a small, fully connected network (typically one or more fully connected layers) is used to learn the weights for each channel. These weights are then used to reweight and compress the feature maps (often using a sigmoid activation function) to produce feature maps with higher responses. This enhances the network's focus on specific channels.

[0050] The SE block, as a residual module, has the following specific structure: Figure 5 As shown, through the residual module, the network output is no longer simply the output F(x) of the convolution, but the superposition of the output of the convolution and the output of the BE block, F(x)+x. This also allows the network to pass through even small errors during backpropagation, thereby avoiding the occurrence of gradient problems.

[0051] This application incorporates SE blocks into the encoding and decoding paths of the U-Net network. By adding SE blocks after the convolutional layers in the encoding and decoding paths, they function as residual modules. These SE blocks dynamically adjust feature weights, mitigating the vanishing and exploding gradient problems and reducing the risk of overfitting. Furthermore, due to the inherent channel attention mechanism of SE blocks, their introduction effectively enhances the network's expressive power, increasing the weights of effective features and decreasing the weights of ineffective or less effective features. Training the model in this way leads to better performance, especially when dealing with complex one-dimensional data, as it better captures the data's feature information and improves the network's prediction accuracy.

[0052] S3: Construct the MSE-MAE loss function based on mean squared error and mean absolute error, and train the SEU-Net network based on low-resolution sample data and high-resolution label data and the MSE-MAE loss function.

[0053] In one embodiment of this application, in regression tasks, loss functions such as MSE, MAE, and Huber loss are generally used. Since the MSE loss function is convex, it means that classic optimization algorithms can be used to find the global minimum, which is beneficial for gradient descent and simplifies the training process. However, due to the use of squared values, MSE may give excessive weight to outliers, potentially reducing the model's generalization ability. Compared to the MSE loss function, the MAE loss function is less sensitive to outliers. However, in some cases, the linear penalty of error by the MAE loss function may make it difficult to find the global minimum during optimization. Furthermore, the gradient of the MAE loss function is constant for all errors (i.e., it does not change with the magnitude of the error), which may cause problems when using optimization algorithms such as gradient descent, especially when the error is close to 0, where the constant gradient may make gradient descent difficult during training. The Huber (Smoothed Mean Absolute Error) loss function is a loss function between mean squared error (MSE) and mean absolute error (MAE), aiming to combine the advantages of both. It introduces a threshold parameter... q We define the threshold as follows: MSE is used when the error is small, and MAE is used otherwise. However, in practice, finding the optimal threshold may require a large number of trials, which can be time-consuming. Furthermore, the optimal threshold may vary for different datasets and tasks. It also requires dynamically switching between the two computation modes based on the error magnitude. This switching can increase computational complexity and is not very beneficial for gradient descent.

[0054] The MSE loss function is defined as follows:

[0055] In the formula, The total number of samples, and These are high-resolution labeled data and data predicted by the SEU-Net network, respectively.

[0056] The MAE loss function is defined as:

[0057] The Huber loss function is defined as:

[0058]

[0059] Therefore, this application considers combining the MSE loss function with the MAE loss function to achieve better results. The MSE-MAE loss function is as follows:

[0060] In the formula, To balance the weighting factors between the MSE and MAE loss functions, based on experiments, Setting it to 0.6 provides the best network performance.

[0061] In the evaluation metrics, the root mean square error (RMSE) and root mean square absolute error (RMAE), which are commonly used in regression tasks, are selected as the evaluation metrics for this application.

[0062] RMSE is defined as:

[0063] RMAE is defined as:

[0064] In one embodiment of this application, the 10,000 pairs of data of length 256 generated above are divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The training set is used to train the model and adjust its weights; the validation set is used to adjust the model's hyperparameters to prevent overfitting; and the test set is used to verify the model's generalization ability to unknown data. Because the generated data is highly random, resulting in significant differences between data points, the data is standardized using the following formula to accelerate model training:

[0065] in, This represents the data after standardization. Represents the original data. and These represent the mean and standard deviation of the original data, respectively.

[0066] After data standardization, the MSE-MAE loss function is selected as the loss function in this application to reduce training error; the Adam optimization algorithm is selected as the optimizer in this application to improve computational efficiency and reduce memory requirements; the initial learning rate of the network is set to 1×10⁻⁶. -4 Due to the Adam optimizer's unique adaptive learning rate mechanism, the learning rate doesn't need excessive adjustment during training. The total number of iterations was 300, the batch size was 64, and the network was trained on a GPU for accelerated performance. The total training time was 65 minutes. The loss function variation graph is shown below. Figure 6 As shown in the figure, after 200 iterations, the loss function values ​​of the training set and the validation set basically converged; in subsequent iterations, the loss function values ​​of the training set and the validation set converged to a stable value, indicating that the model has been trained to its optimal state.

[0067] Example 2: To test the model's effectiveness, synthetic data was used for validation. Two-dimensional seismic profiles and time slices were extracted sequentially from one-dimensional single-channel data and three-dimensional seismic data volumes for verification. Simultaneously, to highlight the effectiveness of the proposed method, the same dataset, parameters, and equipment were used to train the U-Net network, resulting in a trained network model. Test set data was then input into the trained network model to obtain the model's prediction results.

[0068] First, a single data point was randomly selected from the test set for comparison. High-resolution sample data was compared between high-resolution seismic data obtained after processing with the U-Net model and high-resolution seismic data obtained after processing using the method described in this application. The comparison results are as follows: Figure 7 As shown. From Figure 7 As can be seen, the green dashed line and the red line are highly consistent, indicating that the high-resolution seismic data obtained after processing by the method in this application are highly similar to the high-resolution sample data, and the amplitudes are basically the same. The green dashed line and the blue line are mostly consistent, with a small deviation, indicating that the high-resolution seismic data obtained after processing by the U-Net model are basically similar to the high-resolution sample data, but there are some deviations in amplitude. Meanwhile, to avoid randomness, evaluation metrics were used to assess the prediction data of the entire test set, and the specific results are shown in Table 1.

[0069] Table 1 Comparison of Test Set Evaluation Indices

[0070] As can be seen from the table, the RMSE of the data processed by the SEU-Net network of this application decreased by 0.0564 and the RMAE decreased by 0.0441 compared with the U-Net model. This indicates that the evaluation metrics of the SEU-Net network of this application are better than those of the U-Net model. This also shows that the SEU-Net network of this application has better performance and smaller prediction error than the U-Net network.

[0071] Secondly, the trained SEU-Net network model is applied to the prediction of 3D synthetic seismic volumes. First, a reflection coefficient model is used; then, convolutions are performed with the reflection coefficient model using a 30Hz and a 50Hz Ricker wavelet, respectively, to obtain high-resolution and low-resolution 3D sample data volumes. Different levels of Gaussian noise are added to the low-resolution 3D data volume to obtain a further low-resolution 3D sample data volume. The trained model is then used to predict the low-resolution 3D sample data volume, yielding the prediction results.

[0072] To more intuitively represent the prediction results, identical two-dimensional seismic profiles were extracted from the four data volumes for comparison, such as... Figure 8 As shown in the figure, (a) is a low-resolution sample profile, (b) is a high-resolution sample profile, (c) is the prediction result of the U-Net model, and (d) is the prediction result of the SEU-Net network. It can be seen from the figure that the low-resolution sample data contains a large amount of noise, few phase axes, and limited detailed texture information; the high-resolution label data has virtually no noise, more phase axes, and rich detailed texture information; the U-Net model prediction results contain a small amount of noise, indicating that the U-Net model's ability to suppress noise is insufficient, but the increased number of phase axes and rich detailed texture information indicates that the U-Net model has a certain effect on improving the resolution of seismic data; the SEU-Net network prediction results of this application have virtually no noise, more phase axes, and rich detailed texture information, and are highly similar to the high-resolution label data, which also shows that the prediction of the SEU-Net network in this application is relatively accurate and has a good effect on improving the resolution of seismic data and suppressing noise.

[0073] Comparing the low-resolution sample data above, the prediction results of the U-Net model and the prediction results of the SEU-Net network in this application have the following spectra: Figure 9 As shown in the figure, after processing by the U-Net model, the bandwidth of the low-resolution sample data is widened by about 35Hz, and the main frequency is increased by about 20Hz. Low-frequency signals are missing, while high-frequency signals are effectively widened. After processing by the SEU-Net network of this application, the bandwidth of the low-resolution data is widened by about 35Hz, and the main frequency is increased by about 20Hz. Low-frequency signals are effectively protected, and high-frequency signals are effectively widened, further demonstrating the effectiveness of the SEU-Net network of this application.

[0074] Subsequently, at a certain point in time, time slices were extracted from the four data volumes, such as... Figure 10 As shown, (a) is a low-resolution sample slice, (b) is a high-resolution sample slice, (c) is the prediction result of the U-Net model, and (d) is the prediction result of the SEU-Net network. From the time slices, it can be seen that the prediction results of the U-Net model and the SEU-Net network of this application are basically consistent with the time slices of the high-resolution labeled data, showing good continuity. This also indicates that both the U-Net model and the SEU-Net network of this application are effective in improving seismic data, but in terms of accuracy, the SEU-Net network of this application has higher accuracy than the U-Net model.

[0075] To verify the coherence of the data volumes, identical 2D seismic profiles were extracted from each of the four data volumes again, and faults were used to determine the improvement in resolution, such as... Figure 11As shown in the figure, (a) is a low-resolution sample profile, (b) is a high-resolution sample profile, (c) is the prediction result of the U-Net model, and (d) is the prediction result of the SEU-Net network. It can be seen from the figure that the faults in the low-resolution sample profile are not obvious and have poor continuity; the faults in the high-resolution labeled profile are more obvious and have better continuity; the faults in the U-Net model prediction result are more obvious and have better continuity, but there is a lot of noise in the profile; the faults in the SEU-Net network prediction result of this application are more obvious and have better continuity, with virtually no noise. This also shows that as the resolution of seismic data increases, the display of faults becomes clearer, which is more conducive to fault identification.

[0076] Finally, in order to quantitatively analyze the denoising effect of the model, three different levels of noise (mean of 0, variance of 0.01, 0.05 and 0.1 respectively) were selected for comparative experiments, and the results of the comparative experiments were evaluated using the two evaluation indicators mentioned above, as shown in Table 2.

[0077] Table 2 Comparison of Test Values

[0078] As can be seen from the table, the SEU-Net network of this application outperforms the U-Net model in terms of evaluation metrics under three different levels of Gaussian noise, indicating that the SEU-Net network of this application has a better noise suppression effect than the U-Net model in terms of seismic data and has a certain generalization ability.

[0079] Example 3: Compared to synthetic data, real-world data contains more complex noise and is more diverse, increasing the difficulty of resolution improvement. This places higher demands on the fitting and generalization abilities of network models. This application uses a 941×341×400 three-dimensional seismic data volume from a certain region as the real-world data and performs resolution improvement processing on it. A comparison is made between the low-resolution seismic data profile before processing and the high-resolution seismic data profile obtained after processing using the method of this application. Figure 12 As shown in the figure. It can be seen from the figure that the seismic profile before processing (such as...) Figure 12 (a) contains noise and has a coarse phase axis with less detailed texture information; the processed seismic profile (such as...) Figure 12 (b) shows a certain degree of noise suppression and effectively improves resolution, increases the number of in-phase axes, and enhances detailed texture information. This also demonstrates that the network model in this application performs well in practical data applications and has excellent generalization ability.

[0080] Comparing the low-resolution seismic data before processing with the high-resolution seismic data obtained after processing using the SEU-Net network of this application, their spectra are as follows:Figure 13 As shown. From Figure 13 As can be seen, after the actual data is processed by the SEU-Net network of this application, the bandwidth is widened by about 20Hz, the main frequency is increased by about 15Hz, the low frequency signal is effectively protected, and the high frequency signal is effectively widened.

[0081] Comparing the time slices of low-resolution seismic data before processing with the high-resolution seismic data time slices obtained after processing using the SEU-Net network of this application, as shown... Figure 14 As shown in the figure. It can be seen from the figure that the time slice before processing (e.g.) Figure 14 (a) has less detailed texture information; the processed time slice (e.g.) Figure 14 (b) The detailed texture information becomes more abundant. This also demonstrates that the SEU-Net network model of this application can effectively improve the resolution of seismic data and provide a methodological option for high-resolution processing of seismic data.

[0082] Validation using synthetic data shows that seismic data processed by the SEU-Net network of this application exhibits wider bandwidth, higher dominant frequency, and reduced noise. Compared to the U-Net model, the SEU-Net network of this application achieves higher accuracy in improving seismic resolution and better noise suppression. Actual data validation demonstrates that seismic data processed by the SEU-Net network of this application shows reduced noise, significantly enhanced stratigraphic features of seismic profiles, and improved seismic data resolution. However, due to time constraints, the parameters of the SEU-Net network model of this application were not finely tuned; they were roughly set based on past experience, achieving good results. This indicates that there is still room for improvement in the accuracy of the network model. Future research will consider fine-tuning the network model's parameters and adjusting its structure to better extract features, fit data, and thus improve prediction accuracy.

[0083] This application provides a high-resolution processing method for time-series seismic data based on an improved U-Net network. The method uses the U-Net network as its main structure and adds SE blocks after the convolutional layers of the U-Net network, simultaneously implementing the functions of residual modules and attention mechanisms. A loss function combining mean squared error and mean absolute error is used for training, thereby constructing a SEU-Net network to improve seismic data resolution and suppress noise. The SEU-Net network establishes a nonlinear relationship between low-resolution and high-resolution seismic data. This network can broaden the bandwidth, increase the dominant frequency of seismic data, enrich detailed texture information, effectively protect low-frequency signals, effectively broaden high-frequency signals, and also effectively suppress noise, thus effectively improving the resolution of seismic data.

[0084] It should be noted that those skilled in the art will recognize that the embodiments described herein are for the purpose of helping readers understand the principles of this application, and should be understood as not limiting the scope of protection of this application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this application without departing from the essence of this application, and these modifications and combinations are still within the scope of protection of this application.

Claims

1. A high-resolution processing method for time-series seismic data based on an improved U-shaped network, characterized in that, include: S1: Use forward modeling to generate low-resolution sample data and high-resolution label data; S2: Add the SE block to the U-Net network, and implement the functions of residual module and attention mechanism to obtain the SEU-Net network with improved U-shaped network; S3: Construct the MSE-MAE loss function based on mean squared error and mean absolute error, and train the SEU-Net network based on low-resolution sample data and high-resolution label data and the MSE-MAE loss function; S4: Input the low-resolution 3D seismic data into the trained SEU-Net network to obtain high-resolution 3D seismic data.

2. The high-resolution processing method for time-series seismic data based on an improved U-shaped network according to claim 1, characterized in that, The method of generating low-resolution sample data and high-resolution label data using forward modeling includes: S101: Randomly generated reflection coefficient model; S102: Convolution processing is performed using high-frequency Ricker wavelet, low-frequency Ricker wavelet and reflection coefficient model to obtain high-resolution tag data and low-resolution data respectively; The expression for the Reichschild wavelet is: In the formula, Represents the Reich wavelet, This represents the exponential function. Indicates time, This indicates the dominant frequency of the Rack wavelet; The calculation formula for the convolution process is as follows: In the formula, and These represent low-resolution data and high-resolution label data, respectively. and Representing the low-frequency and high-frequency Ricker wavelets respectively, Represents the reflection coefficient model; S103: Add Gaussian noise of varying degrees to low-resolution data to obtain low-resolution sample data.

3. The high-resolution processing method for time-series seismic data based on an improved U-shaped network according to claim 1, characterized in that, The SEU-Net network includes: an encoding part, a decoding part, a skip connection layer, and an output layer; The encoding part is used for feature extraction and includes three downsampling blocks. Each downsampling block consists of two convolutional layers, two SE blocks, and a pooling layer with a 2×2 kernel. The convolutional layers are used to extract features, the SE blocks extract representative features and reduce feature loss during extraction, and the pooling layer reduces feature size. The convolutional layers include 3×3 convolutional kernels, ReLU activation functions, and batch normalization layers. The input of the convolutional layers is padded around the 3×3 convolutional kernels, and the ReLU activation function is used to output the padded data, enabling the data to obtain non-linear relationships between data. The batch normalization layer is used for anti-fitting processing. The SE blocks include compression and activation parts. The decoding section is used for feature fusion and includes three upsampling blocks. Each upsampling block includes two convolutional layers, two SE blocks, and a deconvolutional layer with a 2×2 kernel. The convolutional layers fuse the extracted features, the SE blocks improve the recognition of representative features and provide fusion speed, and the deconvolutional layer restores the feature size. The convolutional layers include a 3×3 kernel, a ReLU activation function, and a batch normalization layer. Each SE block includes a compression part and an activation part. The skip connection layer is located between each downsampling block and upsampling block, and is used to connect some features in the downsampling block with the features in the corresponding upsampling block; The output layer includes a 1×1 convolutional layer used to output the obtained prediction data.

4. The high-resolution processing method for time-series seismic data based on an improved U-shaped network according to claim 3, characterized in that, The SE block includes: When the SE block is used as a channel attention mechanism, it includes a compression part and an activation part. In the compression part, the SE block uses global pooling to compress the feature map of each channel into a single value. In the activation part, the SE block uses a small fully connected network to learn the weights of each channel. The weights are used to reweight the compressed feature map to obtain a representative feature map. When the SE block is used as a residual module, the feature weights are dynamically adjusted through the SE block to alleviate the gradient vanishing and gradient exploding problems in the SEU-Net network.

5. The high-resolution processing method for time-series seismic data based on an improved U-shaped network according to claim 3, characterized in that, The training of the SEU-Net network based on low-resolution sample data, high-resolution label data, and the MSE-MAE loss function includes: S301: Input low-resolution sample data into the downsampling block of the SEU-Net network for feature extraction; S302: The upsampling block fuses the extracted features and connects some features in the downsampling block with the corresponding features in the upsampling block through a skip connection layer. The output layer then outputs high-resolution prediction data. S303: Use the MSE-MAE loss function to evaluate the difference between the predicted data and the high-resolution labeled data, and adjust the parameters of the SEU-Net network based on the difference.

6. The high-resolution processing method for time-series seismic data based on an improved U-shaped network according to claim 5, characterized in that, The expression for the MSE-MAE loss function is as follows: in, Mean square error, The mean absolute error, As a weighting factor, The total number of samples, and These are high-resolution labeled data and data predicted by the SEU-Net network, respectively.

Images