Seismic data distinguishing method based on knowledge distillation

Abstract

The invention provides a seismic data resolution method based on knowledge distillation, relates to the technical field of seismic exploration, and can effectively improve the resolution of low-resolution seismic data. Experimental verification on synthetic data and actual seismic data shows that high-resolution prior information contained in a teacher network is transmitted to a student network through a knowledge distillation technology, and the precision and stability of seismic data high-resolution processing can be remarkably improved. Experimental results show that the method can effectively compensate the high-frequency component of the seismic data, improve the dominant frequency and frequency bandwidth of the data, and maintain the integrity of the low-frequency component at the same time. Through the implementation of the technology, the resolution of a thin layer structure and detail features is improved, reservoir description and stratigraphic structure analysis can be more accurately carried out, the generalization ability and adaptability of a high-resolution model are remarkably enhanced, and important technical support is provided for high-resolution imaging of seismic data.

Description

Technical Field

[0001] The present invention relates to the field of seismic exploration technology, and in particular to a seismic data resolution method based on knowledge distillation. Background Art

[0002] In seismic exploration, the quality of seismic data directly affects the analysis and interpretation of underground geological structures. High-resolution seismic data can provide detailed underground structural information, such as faults, folds, pore distribution, etc., thus providing support for the exploration and development of oil and gas reservoirs. However, in actual exploration, due to factors such as formation absorption attenuation, noise interference, and data acquisition costs, it is often difficult to obtain high-resolution seismic data. In order to reduce costs and improve efficiency, only low-resolution seismic data can usually be collected. These data often cannot effectively reveal the underground structure, resulting in inaccurate exploration results. Therefore, how to recover high-resolution data from low-resolution seismic data through technical means has become a key issue that needs to be solved in seismic data processing.

[0003] Conventional high-resolution seismic data processing methods include spectral whitening, deconvolution, and inverse Q filtering. Spectral whitening aims to compensate for frequency attenuation by expanding the amplitude spectrum of the signal. Traditional spectral whitening methods, such as wavelet spectrum whitening and Hilbert spectrum whitening, can enhance high-frequency information to a certain extent, but may introduce noise and affect the true reflection of the data under low signal-to-noise ratio conditions. Deconvolution methods reduce the filtering effect caused by seismic wave propagation and compress seismic wavelets to restore high-resolution data close to the formation reflection coefficient. Although methods such as predictive deconvolution and homomorphic deconvolution have certain effects, these methods rely on the accurate extraction of seismic wavelets and have limitations in practical applications. Inverse Q filtering compensates for amplitude and frequency attenuation and improves data resolution by estimating the Q value (quality factor) during seismic wave propagation. However, the uncertainty of Q value estimation leads to limited application of inverse Q filtering, especially under high noise conditions.

[0004] In recent years, with the development of seismic exploration technology, the limitations of traditional methods have gradually been exposed, and they cannot meet the needs of processing massive seismic data. Therefore, methods based on deep learning have gradually become an emerging technical path for high-resolution processing of seismic data. Deep learning, especially convolutional neural networks (CNN) and generative adversarial networks (GAN), has made significant progress in the field of seismic data processing. These methods can effectively improve the spatial resolution of seismic data and the detailed expression of underground structures by learning the mapping relationship between low-resolution and high-resolution seismic data. U-Net is a fully convolutional neural network structure designed for pixel-level segmentation tasks. The architecture is characterized by its encoding-decoding symmetric structure and jump connection, which can achieve efficient and accurate image segmentation in a variety of tasks. The U-Net network structure consists of two parts: encoder and decoder. The network presents a symmetrical U-shaped structure. The input data is gradually downsampled by the encoder to extract features, and gradually upsampled by the decoder to reconstruct the image. The characteristic of the U-net network is the use of "jump connection", that is, a feature layer with the same resolution is directly connected between the encoder and the decoder. The jump connection directly transfers the shallow features to the decoding stage to avoid the loss of high-resolution information in downsampling. Through skip connections and symmetric design, U-Net naturally achieves multi-scale information fusion.

[0005] However, existing deep learning methods still have shortcomings in high-resolution processing of seismic data, mainly manifested in the lack of guidance from prior information, which limits the restoration of high-frequency details. Similar to the prior information guidance method in face image processing, the introduction of prior information in the field of seismic data processing can significantly improve the effect of high-resolution restoration. High-resolution restoration of face images usually adopts low-resolution prior, high-resolution prior and prior methods based on knowledge distillation. The low-resolution prior method improves high-resolution performance by extracting information from low-resolution images, but its accuracy is limited; the high-resolution prior obtains prior information by preliminarily enhancing the low-resolution image, thereby optimizing the result; the prior method based on knowledge distillation transfers high-resolution prior information to the student network through the teacher network, avoiding the difficulty of directly estimating prior information and being able to more effectively improve high-resolution restoration performance.

[0006] In order to solve the problem of difficulty in recovering high-frequency details due to the lack of prior information in high-resolution processing of seismic data, the present invention constructs a high-resolution method for seismic data based on knowledge distillation. The inventive method constructs a teacher network and uses real high-resolution seismic data prior information for training, so that the teacher network can more clearly characterize the underground structure and reservoir characteristics, and then uses knowledge distillation technology to transfer the prior knowledge in the teacher network to the student network, so that the student network can directly recover high-resolution data from low-resolution seismic data without estimating prior information. Compared with traditional deep learning methods, the high-resolution method for seismic data based on knowledge distillation solves the problem of missing prior information in the recovery of low-resolution seismic data and improves the accuracy and effect of high-resolution seismic data reconstruction. Summary of the invention

[0007] In view of the above problems in the prior art, this application proposes a seismic data resolution method based on knowledge distillation, comprising the following steps:

[0008] S1. Use low-resolution seismic data as input data and high-resolution seismic data as labels, randomly select low-frequency wavelets and intermediate frequency wavelets of high-frequency wavelets for convolution as high-resolution prior information, and obtain synthetic seismic data; perform data enhancement on the normalized synthetic seismic data; slice the synthetic data in the form of a sliding window to obtain slice data of appropriate size;

[0009] S2. Build a high-resolution network for seismic data based on knowledge distillation. The network framework includes a teacher network and a student network. The knowledge distillation technology is used to transfer the high-resolution prior in the teacher network to the student network.

[0010] S3. After training the teacher network and the student network, the network is fine-tuned through the transfer learning method, and some parameters of the pre-trained network are updated using a small amount of actual work area data.

[0011] In one embodiment, data augmentation includes randomly flipping upside down or rotating 90 degrees or 180 degrees counterclockwise.

[0012] In one embodiment, the teacher network is trained with low-resolution seismic data and prior information, and the student network distills high-resolution seismic prior knowledge from the teacher network.

[0013] In one implementation, both the teacher network and the student network adopt a U-net structure.

[0014] In one embodiment, during the training process, the teacher network characterizes the underground structure details through real high-resolution data, while the student network is guided by soft target loss and feature space loss, so that it can recover real high-resolution details from low-resolution seismic data.

[0015] In one embodiment, the teacher network extracts features from the high-resolution prior through convolutional layers, while extracting features from the low-resolution seismic data through convolutional layers and upsampling layers.

[0016] In one embodiment, the convolution layer is used to extract deep features, and the connection layer is used for feature fusion; the deep features extracted from the high-resolution prior and low-resolution data are concatenated to form a joint feature tensor as the input of the U-net network.

[0017] In one embodiment, a residual module is used so that the high-resolution output is superimposed on itself through a residual layer, so as to learn more high-frequency information from the input data and improve the ability to express details.

[0018] In one embodiment, the high-resolution prediction results output by the teacher network are regarded as soft labels, and the real high-resolution data are regarded as hard labels; the soft labels include different degrees of fuzziness and uncertainty.

[0019] In one embodiment, for the teacher network, the convolutional layers extract the prior data P and the low-resolution data I LR The features of the convolutional layer are concatenated to form a joint feature tensor as the input of the teacher network; during the teacher training process, the intermediate layer features F of each convolutional layer are T Store and return to facilitate intermediate layer supervision in knowledge distillation; finally, the high-resolution output is superimposed on itself through the residual layer.

[0020] The above technical features can be combined in various suitable ways or replaced by equivalent technical features as long as the purpose of the present invention can be achieved.

[0021] The seismic data resolution method based on knowledge distillation provided by the present invention has at least the following beneficial effects compared with the prior art:

[0022] The present invention proposes a high-resolution method for seismic data based on knowledge distillation, which can effectively improve the resolution of low-resolution seismic data. Experimental verification of synthetic data and actual seismic data shows that the accuracy and stability of high-resolution processing of seismic data can be significantly improved by passing the high-resolution prior information contained in the teacher network to the student network through the knowledge distillation technology. Experimental results show that the method of the present invention can effectively compensate for the high-frequency components of seismic data, improve the main frequency and bandwidth of the data, and maintain the integrity of the low-frequency components. Through the implementation of this technology, not only the resolution of thin-layer structures and detailed features is improved, which helps to more accurately describe reservoirs and analyze stratigraphic structures, but also the generalization ability and adaptability of high-resolution models are significantly enhanced, providing important technical support for high-resolution imaging of seismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The present invention will be described in more detail below based on embodiments and with reference to the accompanying drawings, wherein:

[0024] Figure 1 The schematic diagram of the U-net network structure is shown;

[0025] Figure 2 A schematic diagram of the seismic high-resolution teacher network architecture is shown;

[0026] Figure 3 A schematic diagram of the seismic high-resolution student network architecture is shown;

[0027] Figure 4 The overall architecture of the high-resolution network for seismic data based on knowledge distillation is shown;

[0028] Figure 5 A flow chart of the model training process is shown;

[0029] Figure 6 The Marmousi2 model reflection coefficients are shown;

[0030] Figure 7 The teacher network loss function curves corresponding to different loss weights are shown. (a) The teacher network loss function obtained with a loss weight of 0.7; (b) The teacher network loss function obtained with a loss weight of 0.9; (c) The teacher network loss function obtained with a loss weight of 1.0;

[0031] Figure 8 Shows (a) loss function curve during teacher network training; (b) SSIM curve; (c) PSNR curve;

[0032] Fig. 9The prediction results of the teacher network of the Marmousi2 model are shown. (a) Low-resolution synthetic data; (b) High-resolution synthetic data; (c) High-resolution prediction results of the teacher network; (d)-(f) are local enlargements of (a)-(c);

[0033] Fig.10 Shows (a) loss function curve during student network training; (b) SSIM curve; (c) PSNR curve;

[0034] Fig.11 The prediction results of the student network of the Marmousi2 model are shown. (a) The high-resolution prediction result of the student network; (b) is a local enlargement of (a);

[0035] Fig.12 The high-resolution prediction results of actual earthquake data are shown (a) low-resolution original data; (b) high-resolution prediction results of the student network;

[0036] Fig.13 The spectrum comparison of actual seismic data before and after high-resolution processing is shown. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings.

[0038] The U-net network structure used in the present invention is as follows Figure 1 As shown in the figure, the number of channels starts from 64 and increases to 512 in multiples of 2. In the encoder part of U-net, each layer contains two convolutional layers with a kernel size of , and a step size of 1 to keep the size unchanged. A ReLU activation function is set after each convolutional layer, followed by a maximum pooling layer for downsampling and doubling the number of feature channels. The pooling layer size is 2×2, and higher-level features are extracted layer by layer, but detailed information is gradually lost. In the decoder part, each layer is upsampled by a transposed convolutional layer to restore the size of the data. The upsampled feature map is jump-connected with the feature map of the corresponding encoder layer and spliced ​​through the channel dimension. After splicing, the feature map is further processed by two convolutional layers with a kernel size of 3×3 and a ReLU activation function. As the data size is gradually restored, the number of feature channels decreases layer by layer. The last layer uses a convolutional layer with a kernel size of 1×1 to generate the output image.

[0039] Knowledge distillation is a teacher-student network structure that transfers the knowledge from a complex and high-performance teacher model to a smaller or simpler student model to reduce the computational cost and memory usage of the model while ensuring performance. The core of knowledge distillation is that the student model not only learns the labels of real data (hard labels), but also learns the additional information in the teacher model's prediction of the input data (soft labels). The loss function commonly used in knowledge distillation is a weighted combination of hard labels and soft targets:

[0040] L=αL hard +(1-α)L soft (1)

[0041] Where L hard is the hard label loss, L soft is the soft label loss, and α is an adjustable high parameter used to weigh the impact of hard labels and soft labels.

[0042] The main methods of knowledge distillation can be divided into four categories: output-based distillation, intermediate-layer-based distillation, relationship-based distillation, and self-distillation. Output-based distillation uses the final output of the teacher model (such as category probability distribution) as a soft label, and the student model learns knowledge by imitating these distributions; intermediate-layer-based distillation allows the student model to learn more structural knowledge by matching the intermediate-layer features of the student model with the intermediate-layer features of the teacher model; relationship-based distillation allows the student model to learn the relationship or pattern between different samples of the teacher model; self-distillation only has a student model, without a clear teacher model, so that the student model is trained through self-enhancement (such as learning from its own past output).

[0043] Two commonly used quantitative metrics are used to evaluate the high-resolution results: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).

[0044] Peak signal-to-noise ratio is a commonly used indicator to measure the quality of image reconstruction. It is calculated based on the ratio of the maximum possible signal power to the signal noise power. Peak signal-to-noise ratio is usually used to evaluate the quality of images after high-resolution processing. The higher the value, the smaller the difference between the reconstructed image and the original image, and the higher the image quality. The calculation formula of peak signal-to-noise ratio is:

[0045]

[0046] Where R is the maximum possible value of the image pixel value. For an 8-bit image, R = 255; m×n represents the size of the seismic image, I SR (i, j) represents the pixel value of the high-resolution prediction result, I HR (i, j) represents the pixel value of the high-resolution label data.

[0047] The structural similarity index is used to measure the structural similarity between the high-resolution prediction results and the high-resolution label data, rather than just the difference between pixels. Compared with the peak signal-to-noise ratio, the value of the structural similarity index is closer to the human eye's evaluation of image quality. The closer the structural similarity index value is to 1, the higher the structural similarity between the high-resolution prediction results and the high-resolution label data. Its definition is as follows:

[0048]

[0049] Among them, x and y represent two earthquake data, μ x and μ y represent the mean of earthquake data, and They represent the standard deviation of seismic data, σ xy represents the covariance between x and y, C 1 and C 2 is a stabilizing constant used to prevent the denominator from being zero.

[0050] The high resolution of seismic data is crucial for accurately analyzing underground geological structures. On the one hand, existing high-resolution processing methods for seismic data, such as spectral whitening, deconvolution, inverse Q filtering and other traditional methods, have some limitations, especially in the case of low signal-to-noise ratio, which may introduce noise and destroy the authenticity of the data, making it difficult to meet the needs of large-scale seismic data processing. On the other hand, most deep learning high-resolution methods directly learn the mapping relationship between low-resolution data and high-resolution data in an end-to-end network framework. In the absence of guidance from prior information of seismic data, it is difficult to restore the high-frequency detail information of the real seismic data. Therefore, in order to solve the problem that low-resolution seismic data is difficult to restore high-resolution underground details and the errors caused by inaccurate prior estimation, the present invention proposes a high-resolution method for seismic data based on knowledge distillation, which uses knowledge distillation technology to transfer high-resolution prior information in the teacher network to the student network, so that the student network can directly restore high-resolution data from low-resolution seismic data without prior information. In actual work, using a trained student network to achieve high-resolution processing of seismic data can avoid the problem of lack of prior information in actual seismic data. This method effectively avoids the impact of inaccurate prior estimates, can more clearly restore underground structures and reservoir characteristics, and improve the spatial resolution of seismic data, thereby providing more accurate support for oil and gas exploration and geological analysis.

[0051] The present invention proposes a high-resolution method for seismic data based on knowledge distillation, which uses low-resolution seismic data and high-resolution seismic prior information as input to train the network architecture constructed by the present invention. The network architecture of the present invention includes a teacher network and a student network, both of which use a U-net network structure. Among them, the teacher network is trained with low-resolution seismic data and prior information, and the student network distills high-resolution seismic prior knowledge from the teacher network. During the training process, the teacher network depicts the underground structure details through real high-resolution data, and the student network is guided by soft target loss and feature space loss, so that it can recover real high-resolution details from low-resolution seismic data. Finally, the high-resolution results are evaluated by two indicators, PSNR and SSIM.

[0052] In one embodiment, the reflection coefficient of the Marmousi2 model is convolved with the low-frequency wavelet and the high-frequency wavelet respectively to construct low-resolution seismic data as input data, and high-resolution seismic data as labels, and then the intermediate frequency band wavelets of the low-frequency wavelet and the high-frequency wavelet are randomly selected for convolution as high-resolution prior information. After all the synthetic seismic data are obtained, in order to enhance the generalization ability of the deep learning model and prevent the overfitting of the model, the synthetic data is normalized and randomly flipped up and down, rotated 90 degrees counterclockwise, rotated 180 degrees, and other data enhancement operations are performed. In order to obtain data suitable for the input size of the deep learning model, the synthetic data is sliced ​​in the form of a sliding window to obtain slice data of appropriate size.

[0053] The network framework used in the present invention consists of two networks: a teacher network and a student network, both of which adopt a U-net structure. The teacher network takes low-resolution seismic data and high-resolution prior information as input, and can clearly depict the underground structure under the guidance of real high-resolution prior information; while the student network only inputs low-resolution seismic data. In order to enable the student network to introduce prior information without estimating prior information, the knowledge distillation technology is used to transfer the high-resolution prior in the teacher network to the student network. In the model training process of the present invention, we used L1 loss, feature space loss, soft label loss and SSIM loss.

[0054] The teacher network takes as input the low-resolution data and the high-resolution prior concatenation and is defined as:

[0055]

[0056] where P is the high-resolution seismic prior, I LR is the low-resolution seismic data, f Teacher represents the function of the teacher network, is the high-resolution seismic data generated by the teacher network, and [·,·] is the concatenation operator. The teacher network is only used in the training phase, when it is trained with synthetic data, so accurate high-resolution prior information can be obtained.

[0057] The specific architecture of the earthquake high-resolution teacher network is as follows Figure 2As shown in the figure, the teacher network extracts features from high-resolution priors through convolutional layers, and extracts features from low-resolution seismic data through convolutional layers and upsampling layers. This is because the resolution of low-resolution seismic data is low, and direct use for subsequent feature fusion may lead to insufficient seismic structure information, thereby limiting the model's ability to reconstruct the target high-resolution structure. Therefore, it is necessary to compensate for the insufficient spatial resolution of low-resolution seismic data through upsampling layers to ensure that the feature map does not lose information due to resolution differences in subsequent processing. Convolutional layers are used to extract deep features, and connection layers are used for feature fusion. The deep features extracted from high-resolution priors and low-resolution data are concatenated to form a joint feature tensor as the input of the U-net network. In addition, the residual module is used to allow the high-resolution output to be superimposed on itself through the residual layer, so as to learn more high-frequency information from the input data and improve the ability to express details.

[0058] In order to constrain the teacher network to generate high-resolution seismic data, L 1 The loss and structural similarity loss (SSIM) are used as its loss function. SSIM loss focuses on the texture details and structural similarity of seismic data. After adding SSIM loss, the model will be more inclined to retain detailed information, such as the reflection characteristics of seismic waves and geological boundaries, which is conducive to improving the fidelity of seismic reflection information. 1 The loss is defined as follows:

[0059]

[0060] in, is the high-resolution seismic data output by the teacher network, I HR It is high-resolution earthquake label data.

[0061] SSIM is used to evaluate the similarity between the predicted image and the target image in terms of structure, brightness, and contrast. The SSIM range is [0,1], and the closer the value is to 1, the more similar the two images are. It is defined as follows:

[0062]

[0063] Among them, x and y represent two earthquake data, μ x and μ y represent the mean of earthquake data, and They represent the standard deviation of seismic data, σ xy represents the covariance between x and y, C 1 and C 2 is a stable constant used to avoid the denominator being zero. SSIM loss takes 1-SSIM(x,y) as the optimization objective to minimize the structural differences between seismic data. SSIM Defined as:

[0064]

[0065] Therefore, the loss function of the teacher network is defined as:

[0066] L Teacher =α×L Voxel +(1-α)×L SSIM (8)

[0067] α is the weight of the loss function.

[0068] The teacher network effectively obtains geological structure information through accurate high-resolution priors during the training phase. However, it is difficult to obtain accurate and reliable prior information when processing real seismic data during the inference phase. In order to utilize high-resolution priors during the inference phase, the present invention adopts a prior knowledge distillation strategy to distill the prior knowledge in the teacher network equipped with high-resolution priors into the student network that does not rely on priors. In this way, the student network can fully utilize and obtain high-resolution priors, without the need to directly input prior information or estimate priors.

[0069] The specific architecture of the earthquake high-resolution student network is as follows Figure 3 As shown in Figure 1, the structure of the student network is similar to that of the teacher network, except that the input of the student network only includes low-resolution seismic data. Therefore, in the inference stage, the student network can use low-resolution seismic data to generate high-resolution seismic data. The loss function of the student network is defined as follows:

[0070] L Student =α*L Pixel +β*L Feature +γ*L Soft +κ*L SSIM (9)

[0071] Among them, α, β, γ and κ are L Pixel , L Feature , L Soft and L SSIM In order to transfer high-resolution earthquake priors from the teacher network to the student network, the present invention introduces a soft label loss L Soft and feature space loss L Feature In addition, relying solely on pixel-level differences may not be sufficient to obtain important structural features in seismic data, so it is necessary to combine the SSIM loss function L SSIM To improve the structural fidelity of the generated data and ensure that the high-resolution data is consistent with the high-resolution labeled data in terms of seismic characteristics. Pixel Yes and I HR Applied pixel-by-pixel L 1The loss function is as follows:

[0072]

[0073] in, is the high-resolution result of the student network, I HR It is high-resolution earthquake label data.

[0074] In the knowledge distillation technology, we can regard the high-resolution prediction results output by the teacher network as a soft label, and the real high-resolution data is called a hard label. Soft labels can include different degrees of fuzziness and uncertainty, making the learning goals of the student network broader, thereby improving the adaptability and generalization ability of unknown earthquake data. In addition, soft labels do not set overly clear and difficult goals for the student network, but provide a smoother learning direction to help students gradually move closer to the teacher network. Therefore, soft labels can provide simpler but more effective supervision for the student network. The soft label loss is based on L 1 The loss is implemented to measure the difference between the output of the student network and the output of the teacher network, and the formula is as follows:

[0075]

[0076] in, is the high-resolution result of the student network, is the prediction result of the teacher network, i.e., the soft label, composed of I LR and P. Soft labels contain rich high-resolution prior information and are easier to learn than true targets, enabling the student network to learn from the teacher network more efficiently, thereby significantly improving model performance.

[0077] The teacher network uses high-resolution prior information during training, and its intermediate features contain rich prior information. When the student network has no prior information during the inference phase, it can learn these features through feature space loss to make up for the missing information. Therefore, in order to explicitly constrain the intermediate features of the teacher network and the student network, the feature space loss is used to calculate the difference between the output features of the teacher network and the student network at certain specific layers (usually intermediate layers or hidden layers). The formula is as follows:

[0078]

[0079] Among them, F i S and F i T They represent the intermediate layer features of the student network and the teacher network respectively, which are usually the output tensors of the intermediate layers of the network, such as the activation features after the convolutional layer, and n is the number of selected intermediate feature layers.

[0080] The overall structure of the high-resolution network for seismic data based on knowledge distillation is shown in Figure 2. Figure 4 As shown. For the teacher network, the convolutional layer extracts prior data P and low-resolution data I respectively LR The characteristic f a and f b , concatenate the features to form a joint feature tensor as the input of the teacher network; during the teacher training process, the intermediate layer features F of each convolutional layer are T Store and return to facilitate the supervision of the intermediate layer in knowledge distillation; finally, the high-resolution output is superimposed on itself through the residual layer to improve the ability to express details. The teacher network outputs the intermediate layer feature F T and high-resolution prediction results The student network is used for training. The student network uses the low-resolution data features of the convolutional layer as the input of the student network and loads the teacher network weights to provide the intermediate layer features F T and high resolution output During the training process, the intermediate layer features F of each convolutional layer are also S Store and return, and finally further learn high-frequency detail information through the residual layer.

[0081] Based on the synthetic data construction and network building process mentioned above, the model training process of the present invention can be divided into the following steps, as shown in the flowchart: Figure 5 shown.

[0082] Step 1: Generate earthquake data

[0083] Firstly, the reflection coefficient model is used to convolve with Ricker wavelet to synthesize low-resolution input data I LR , high-resolution prior data P and high-resolution label data I HR , and perform data enhancement on the synthetic seismic data.

[0084] Step 2: Train the teacher network

[0085] The feature f is extracted from the high-resolution prior through the convolutional layer a , extracting features f from low-resolution seismic data through convolutional layers and upsampling layers b , f a and f b Perform feature concatenation as the input of the U-net network. Adjust the weight α 1 and α 2 , using L Teacher The loss trains the teacher network and outputs the intermediate layer features F T and high-resolution prediction data The prediction results are evaluated by PSNR and SSIM.

[0086] Step 3: Train the student network

[0087] Load the model parameters of the trained teacher network, including the intermediate features of the teacher network and the high-resolution prediction results. Extract features from the low-resolution seismic data through the convolutional layer as the input of the U-net network. Student Loss training student network: calculate L1 loss, the difference between the intermediate features of the student and teacher networks, the difference between the student and teacher outputs, and SSIM loss. Dynamically adjust L during training Student Loss weight: The distillation loss has a larger weight in the early stage of training, and as the training progresses, the distillation loss weight gradually decreases. Output student network high-resolution results And it is evaluated by PSNR and SSIM.

[0088] Step 4: Transfer Learning

[0089] When the network trained directly with synthetic seismic data is tested on actual work area data, the test results may be inaccurate due to the direct differences between seismic data and seismic wavelets, and the data of the actual work area is not sufficient to retrain the network. Therefore, the network is fine-tuned through transfer learning method, and some parameters of the pre-trained network are updated using a small amount of actual work area data, so that the model can better adapt to the characteristics of the actual work area data.

[0090] Freeze the parameters of the first few layers (feature extraction layers) of the pre-trained network, because these layers have learned general features such as edges and textures. Only the last two convolutional layers and the residual layer of the network are trained and their weights are updated so that they can adapt to the specific features of the actual work area data.

[0091] In one embodiment, Figure 6 The reflection coefficients of some Marmousi2 models are shown. The reflection coefficients of the Marmousi2 model are convolved with low-frequency wavelets and high-frequency wavelets respectively to construct low-resolution seismic data as input data and high-resolution seismic data as labels. Then, the intermediate frequency band wavelets of low-frequency wavelets and high-frequency wavelets are randomly selected for convolution as high-resolution prior information. In order to enhance the generalization of the network, the low-frequency wavelet is a randomly selected Ricker wavelet with a main frequency of 20-30Hz, the high-frequency wavelet selects a Ricker wavelet with a main frequency of 60Hz, and the intermediate frequency band wavelet selects a Ricker wavelet with a main frequency of 40Hz. The sampling time is 1ms, and 30 sets of high-resolution data with a size of 1000×300 and the corresponding low-resolution data and high-resolution prior are constructed. By setting an appropriate sliding step size, the 1000×300 synthetic data is divided into 7400 synthetic data with a size of 64×64, and data enhancement operations such as upside down flipping, counterclockwise rotation of 90 degrees, and rotation of 180 degrees are randomly performed. The ratio of the training set to the validation set is set to 9:1.

[0092] First use L Teacher Loss training teacher network, Figure 7 Shows L Teacher Table 1 shows the corresponding PSNR and SSIM values ​​for the teacher network loss function curves when the loss weight α in is different. It can be seen that the loss weight affects the rate of decrease of the loss function. When the loss weight is 1.0, that is, L Teacher When the weight of SSIM loss is 0, the loss function drops too fast, which shows the importance of SSIM loss. Therefore, the training parameters of the teacher network are set as follows: loss weight is 0.7, learning rate is 0.00001, optimizer is Adam optimizer, iteration number is 30, and batch size is 16. Figure 8 The loss curves on the training set and validation set during the teacher network training process, as well as the PSNR and SSIM values ​​on the validation set.

[0093] Table 1 SSIM and PSNR corresponding to different loss weights α

[0094]

[0095] In one embodiment, the entire Marmousi2 model is tested using the trained teacher network. Fig. 9 (a) is the low-resolution synthetic seismic data obtained by convolution of 30Hz Ricker wavelet and reflection coefficient. Fig. 9 (b) is the high-resolution synthetic data obtained by convolution of 60Hz Ricker wavelet. Fig. 9 (c) is the high-resolution prediction result using the teacher network. Fig. 9 (d)-(f) show Fig. 9 The local magnification of the black box part in (a)-(c). It can be seen that the high-resolution results of the teacher network are very close to the high-resolution labeled data, the local structural information is clearly discernible, and the resolution of the data has been effectively improved. The results show that the teacher network with the introduction of high-resolution prior has a significant effect on the high-resolution processing of seismic data.

[0096] In one embodiment, the student network is trained using the trained teacher network, the teacher network parameters are frozen, and the intermediate layer features and high-resolution output of the teacher network are used to guide the training of the student network. The loss function is L Student , where L Pixel Weight α = 1.4, L Soft Weight γ = 0.1, L SSIM The weight of L is κ = 0.4. Feature The weights need to be adjusted dynamically because the student network has not yet fully learned the basic features at the beginning of training. FeatureIf the weight of feature space loss is too low, the student network may deviate from the target provided by the teacher network. Therefore, a higher weight needs to be given to force the student network to be as close to the characteristics of the teacher network as possible. As the training progresses, the student network gradually learns some features and establishes preliminary representation capabilities. At this time, if the weight of feature space loss is kept too high, it may limit the freedom of the student network and make it difficult to further optimize. Therefore, it is necessary to gradually reduce the weight of feature space loss to relax the constraints, so that the student network can flexibly adjust under the guidance of the teacher network to better learn the details of specific tasks. Feature The weight β is defined as:

[0097]

[0098] epoch is the number of iterations in the student network training, and d is the threshold. When training the Marmousi2 model, d=5. The Adam optimizer is also selected, the number of iterations is 30, and the batch size is 16.

[0099] The loss function, SSIM curve and PSNR curve of the student network during training are as follows: Fig.10 As shown in the figure, it can be seen that the loss decreases significantly with the increase of the number of training times. The loss decreases significantly in the early stage of training and gradually stabilizes after 10 times. The trends of the validation set and the training set are basically consistent, and the SSIM curve and PSNR curve of the student network are very similar to the results of the teacher network. Fig.11 The high-resolution results of the student network are shown, which are very close to those of the teacher network. This shows that the knowledge distillation technique successfully transfers the high-resolution priors in the teacher network to the student network, and the student network can reconstruct high-resolution data directly from low-resolution data without inputting prior information.

[0100] Actual seismic data from a certain work area were selected for testing. The seismic profile size was 500×400 and the sampling rate was 4ms. Due to the difference between the seismic wavelets of the Marmousi2 model and the actual work area data, the prediction results could not meet expectations when the network trained by the Marmousi2 model was directly used to test the actual data. Therefore, the transfer learning method was used. First, the actual seismic data was divided into large and small data blocks, 20 of which were randomly selected as low-resolution inputs, and then the spectrum continuation method was used to construct high-resolution data from these selected data, obtaining 20 pairs of high-resolution and low-resolution data pairs. These data were used to fine-tune the trained network and then test the entire seismic profile. Fig.12 The high-resolution results of the original seismic data and the fine-tuned student network are shown. It can be seen that the overall resolution of the processed seismic data has been significantly improved, and the thin layers that were originally difficult to identify have become clearer (black box part), showing more detailed information, which is conducive to more accurate reservoir characterization and stratigraphic structure analysis.

[0101] We further selected the 250th trace of the actual seismic data profile and compared the spectra before and after high-resolution processing, as shown in Figure 2. Fig.13 It can be seen that the student network high-resolution method compensates for the high-frequency part of the data, and the main frequency and bandwidth are improved to a certain extent without damaging the low-frequency part.

[0102] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims

1. A seismic data resolution method based on knowledge distillation, characterized in that: The following steps are involved: S1. Use low-resolution seismic data as input data and high-resolution seismic data as labels, randomly select low-frequency wavelets and intermediate frequency wavelets of high-frequency wavelets for convolution as high-resolution prior information, and obtain synthetic seismic data; perform data enhancement on the normalized synthetic seismic data; slice the synthetic data in the form of a sliding window to obtain slice data of appropriate size; S2. Build a high-resolution network for seismic data based on knowledge distillation. The network framework includes a teacher network and a student network. The knowledge distillation technology is used to transfer the high-resolution prior in the teacher network to the student network. S3. After training the teacher network and the student network, the network is fine-tuned through the transfer learning method, and some parameters of the pre-trained network are updated using a small amount of actual work area data.

2. The seismic data resolution method based on knowledge distillation according to claim 1, characterized in that: Data augmentation includes random upside-down flipping and 90 or 180 degree counterclockwise rotation.

3. The seismic data resolution method based on knowledge distillation according to claim 1, characterized in that: The teacher network is trained with low-resolution seismic data and prior information, while the student network distills high-resolution seismic prior knowledge from the teacher network.

4. The seismic data resolution method based on knowledge distillation according to claim 3, characterized in that: Both the teacher network and the student network adopt the U-net structure.

5. The seismic data resolution method based on knowledge distillation according to claim 3, characterized in that: During the training process, the teacher network characterizes the underground structure details through real high-resolution data, while the student network is guided by soft target loss and feature space loss, enabling it to recover real high-resolution details from low-resolution seismic data.

6. The seismic data resolution method based on knowledge distillation according to claim 5, characterized in that: The teacher network extracts features from high-resolution priors through convolutional layers, while extracting features from low-resolution seismic data through convolutional and upsampling layers.

7. The seismic data resolution method based on knowledge distillation according to claim 6, characterized in that: The convolution layer is used to extract deep features, and the connection layer is used for feature fusion; the deep features extracted from the high-resolution prior and low-resolution data are concatenated to form a joint feature tensor as the input of the U-net network.

8. The seismic data resolution method based on knowledge distillation according to claim 7, characterized in that: The residual module is used to make the high-resolution output superimposed on itself through the residual layer, so as to learn more high-frequency information from the input data and improve the ability to express details.

9. The seismic data resolution method based on knowledge distillation according to claim 8, characterized in that: The high-resolution prediction results output by the teacher network are regarded as soft labels, and the real high-resolution data are regarded as hard labels; the soft labels include different degrees of ambiguity and uncertainty.

10. The seismic data resolution method based on knowledge distillation according to claim 8, characterized in that: For the teacher network, the convolutional layers extract prior data P and low-resolution data I respectively. LR The features of the convolutional layer are concatenated to form a joint feature tensor as the input of the teacher network; during the teacher training process, the intermediate layer features F of each convolutional layer are T Store and return to facilitate intermediate layer supervision in knowledge distillation; finally, the high-resolution output is superimposed on itself through the residual layer.

Images