A method for seismic data resolution based on knowledge distillation

Abstract

The application provides a kind of seismic data resolution method based on knowledge distillation, it is related to the technical field of seismic exploration, can effectively improve the resolution of low resolution seismic data.The experimental verification of synthetic data and actual seismic data shows that: through the knowledge distillation technology, the high-resolution prior information contained in the teacher network is transmitted to the student network, which can significantly improve the accuracy and stability of high-resolution processing of seismic data.The experimental results show that the method of the application can effectively compensate the high-frequency component of seismic data, improve the main frequency and bandwidth of the data, while maintaining the integrity of the low-frequency component.Through the implementation of the technology, not only the resolution of thin layer structure and detail feature is improved, which is helpful for more accurate reservoir description and stratigraphic structure analysis, but also the generalization ability and adaptability of high-resolution model are significantly enhanced, which provides important technical support for high-resolution imaging of seismic data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of seismic exploration, and particularly relates to a seismic data resolution method based on knowledge distillation. BACKGROUND

[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 information about underground structures, such as faults, folds, and pore distribution, thereby providing support for oil and gas exploration and development. However, in actual exploration, it is often difficult to obtain high-resolution seismic data due to factors such as stratum absorption attenuation, noise interference, and data acquisition costs. In order to reduce costs and improve efficiency, only low-resolution seismic data can be collected, which 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 problem 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 spectral whitening and Hilbert spectral whitening, can enhance high-frequency information to some extent, but in low signal-to-noise ratio conditions, they may introduce noise and affect the true reflection of the data. Deconvolution methods recover high-resolution data close to the stratum reflection coefficient by reducing the filtering effect caused by seismic wave propagation and compressing the seismic wavelet. Although methods such as predictive deconvolution and homomorphic deconvolution have some effect, these methods rely on accurate extraction of the seismic wavelet and have limitations in practical applications. Inverse Q filtering compensates for amplitude and frequency attenuation by estimating the Q value (quality factor) during seismic wave propagation to improve data resolution. However, the estimation of the Q value is uncertain, which limits the application effect of inverse Q filtering, especially in high-noise conditions.

[0004] With the development of seismic exploration technology in recent years, the limitations of traditional methods have gradually emerged, and they cannot meet the demand for processing massive seismic data. Therefore, methods based on deep learning have gradually become a new technology 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 detail performance 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. This architecture is characterized by its encoding-decoding symmetric structure and skip connections, which can achieve efficient and accurate image segmentation in various tasks. The U-Net network structure consists of two parts: the encoder and the decoder. The network presents a symmetric U-shaped structure, and the input data is gradually down-sampled to extract features through the encoder, and then gradually up-sampled to reconstruct the image through the decoder. The U-net network is characterized by the use of "skip connections", which are designed to connect the same resolution feature layers directly between the encoder and the decoder. The skip connection directly transmits the shallow features to the decoding stage, avoiding the loss of high-resolution information in the down-sampling process. Through skip connections and symmetric design, U-Net naturally realizes multi-scale information fusion.

[0005] However, existing deep learning methods still have limitations in high-resolution processing of seismic data, mainly manifested in the lack of prior information guidance, which limits the recovery effect of high-frequency details. Similar to the prior information guidance method in face image processing, introducing prior information in the field of seismic data processing can significantly improve the effect of high-resolution recovery. High-resolution recovery of face images usually uses low-resolution prior, high-resolution prior, and prior methods based on knowledge distillation. Low-resolution prior methods improve high-resolution performance by extracting information from low-resolution images, but their accuracy is limited; high-resolution prior methods obtain prior information by preliminarily enhancing low-resolution images to optimize results; prior methods based on knowledge distillation pass high-resolution prior information from the teacher network to the student network, avoiding the difficulty of directly estimating prior information and more effectively improving high-resolution recovery performance.

[0006] In order to solve the problem that the high-frequency details are difficult to be restored due to the lack of prior information in the high-resolution processing of seismic data, the application constructs a high-resolution method of seismic data based on knowledge distillation. The method trains the teacher network by constructing the teacher network and using the real high-resolution seismic data prior information, so that the teacher network can clearly depict the underground structure and reservoir characteristics. Then, the prior knowledge in the teacher network is transferred to the student network by using the knowledge distillation technology, so that the student network can directly restore the high-resolution data from the low-resolution seismic data without estimating the prior information. Compared with the traditional deep learning method, the high-resolution method of seismic data based on knowledge distillation solves the problem of the lack of prior information in the restoration of low-resolution seismic data, and improves the accuracy and effect of the reconstruction of high-resolution seismic data. SUMMARY

[0007] In view of the problems in the prior art, the application provides a seismic data resolution method based on knowledge distillation, which comprises the following steps:

[0008] S1, using low-resolution seismic data as input data and high-resolution seismic data as label, randomly selecting a middle frequency band wavelet of a low-frequency wavelet and a high-frequency wavelet for convolution, as high-resolution prior information, to obtain synthetic seismic data; performing data enhancement on the normalized synthetic seismic data; slicing the synthetic data in the form of a sliding window to obtain slice data of a proper size;

[0009] S2, building a high-resolution network of seismic data based on knowledge distillation, the network framework comprising a teacher network and a student network, and using a knowledge distillation technology 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, fine-tuning the network by using a transfer learning method, and updating part of the parameters of the pre-trained network by using a small amount of actual work area data.

[0011] In one embodiment, the data enhancement includes randomly performing up-down flipping, counterclockwise rotation by 90 degrees or 180 degrees.

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

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

[0014] In an embodiment, in 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 the real high-resolution details from low-resolution seismic data.

[0015] In an embodiment, the teacher network extracts features from high-resolution priors through convolution layers, and extracts features from low-resolution seismic data through convolution layers and up-sampling layers.

[0016] In an embodiment, the convolution layers are used for deep feature extraction, and the connection layers are used for feature fusion; the deep features extracted from the high-resolution priors and the low-resolution data are spliced to form a joint feature tensor, which is used as the input of the U-net network.

[0017] In an embodiment, the high-resolution output is superimposed on itself through a residual layer by using a residual module, so as to learn more high-frequency information from the input data and improve the detail performance.

[0018] In an embodiment, the high-resolution prediction result output by the teacher network is regarded as a soft label, and the real high-resolution data is regarded as a hard label; the soft label includes different degrees of ambiguity and uncertainty.

[0019] In an embodiment, for the teacher network, the convolution layers extract features of the prior data P and the low-resolution data I LR respectively, and the features are spliced to form a joint feature tensor, which is used as the input of the teacher network; in the teacher training process, the intermediate layer features F T of each convolution layer are stored and returned, so as to facilitate the intermediate layer supervision in knowledge distillation; finally, the high-resolution output is superimposed on itself through a 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 application can be achieved.

[0021] The method provided by the application has at least the following beneficial effects compared with the prior art:

[0022] The application provides a high-resolution method for seismic data based on knowledge distillation, which can effectively improve the resolution of low-resolution seismic data. Experimental verification on synthetic data and actual seismic data shows that the high-resolution prior information contained in the teacher network can be transmitted to the student network through the knowledge distillation technology, which can significantly improve the accuracy and stability of high-resolution processing of seismic data. The experimental results show that the method 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 the technology, not only the resolution of thin layer structure and detail features is improved, which is helpful for more accurate reservoir description and stratigraphic structure analysis, but also the generalization ability and adaptability of the high-resolution model are significantly enhanced, which provides important technical support for high-resolution imaging of seismic data. BRIEF DESCRIPTION OF DRAWINGS

[0023] The application will be described in more detail below based on the embodiments and with reference to the accompanying drawings. Among them:

[0024] Figure 1 A U-net network structure schematic diagram is shown;

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

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

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

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

[0029] Figure 6 A Marmousi2 model reflectance is shown;

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

[0031] Figure 8 Teacher network training process (a) loss function curve; (b) SSIM curve; (c) PSNR curve is shown;

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

[0033] Figure 10 The student network training process is shown. (a) loss function curve; (b) SSIM curve; (c) PSNR curve;

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

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

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

[0037] The application will be further described below with reference to the accompanying drawings.

[0038] The U-net network structure used by the application is as shown in Figure 1 The number of channels starts from 64 and increases by 2 to 512. In the encoder part of the 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 max pooling layer for downsampling operation while doubling the number of feature channels. The pooling layer size is 2x2, and higher layer features are extracted layer by layer, but the detailed information is gradually lost. In the decoder part, each layer is upsampled by a transpose convolutional layer to restore the size of the data. The upsampled feature map is connected to the corresponding encoder layer feature map through a channel dimension splicing. After splicing, the feature map is further processed by two convolutional layers with a kernel size of 3x3 and a ReLU activation function. As the data size gradually recovers, the number of feature channels decreases layer by layer. The last layer uses a convolutional kernel size of 1x1 to generate an output image.

[0039] Knowledge distillation is a teacher-student network structure that transfers the knowledge of a complex and high-performance teacher model to a smaller or simpler student model to reduce the computational cost and memory occupancy 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 additional information in the teacher model's prediction (soft labels). The commonly used loss function in knowledge distillation is a weighted combination of hard labels and soft targets:

[0040] L = aL hard + (1 - a)L soft (1)

[0041] where L hard is the hard label loss, L soft is the soft label loss, and a is an adjustable hyper-parameter to balance the influence of hard label and soft label.

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

[0043] Two commonly used quantitative indicators 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 image reconstruction quality, which is calculated based on the ratio of the maximum possible power of the signal to the power of the signal noise. Peak signal-to-noise ratio is usually used to evaluate the quality of high-resolution processing of images, and the higher the value, the smaller the difference between the reconstructed image and the original image, and the higher the image quality. The formula for calculating the peak signal-to-noise ratio is:

[0045]

[0046] where R is the maximum possible value of the image pixel value, R = 255 for an 8-bit image; m x 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] Structural similarity index is used to measure the structural similarity between high-resolution prediction results and high-resolution label data, rather than just the difference between pixels. Compared with peak signal-to-noise ratio, the value of structural similarity index is closer to the evaluation of image quality by the human eye. The closer the value of structural similarity index is to 1, the higher the structural similarity between the high-resolution prediction result and the high-resolution label data. Its definition is as follows:

[0048]

[0049] where x and y represent two seismic data respectively, μ x and μ y represent the mean of seismic data respectively, and represent the standard deviation of seismic data respectively, σ xy represents the covariance between x and y, C1 and C2 are stable constants for avoiding zero denominator.

[0050] High resolution of seismic data is crucial for accurate analysis of underground geological structure. On the one hand, existing seismic data high resolution processing methods, 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, and it is difficult to meet the demand of large-scale seismic data processing. On the other hand, most of the deep learning high resolution methods directly learn the mapping relationship between low resolution data and high resolution data in the end-to-end network framework, and it is difficult to recover the real high frequency detail information of seismic data in the absence of prior information guidance. Therefore, in order to solve the problem of low resolution seismic data difficult to recover high resolution underground details and the error caused by inaccurate prior estimation, the present application proposes a seismic data high resolution method 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 recover high resolution data from low resolution seismic data without prior information. In practical work, the trained student network is used to realize the high resolution processing of seismic data, which can avoid the problem of lack of prior information of actual seismic data. This method effectively avoids the influence of inaccurate prior estimation, can more clearly recover the underground structure and reservoir characteristics, and improves the spatial resolution of seismic data, thereby providing higher precision support for oil and gas exploration and geological analysis.

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

[0052] In one embodiment, 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, and then a middle-frequency wavelet between the low-frequency wavelets and the high-frequency wavelets is randomly selected for convolution as high-resolution prior information. After obtaining all the synthetic seismic data, in order to enhance the generalization ability of the deep learning model and prevent overfitting of the model, the synthetic data is normalized and then randomly subjected to data enhancement operations such as up-down flipping, counterclockwise rotation by 90 degrees, and rotation by 180 degrees. 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 a suitable size.

[0053] The network framework used in the present application is composed 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 underground structures under the guidance of real high-resolution prior information. The student network only inputs low-resolution seismic data, and 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 application, L1 loss, feature space loss, soft label loss, and SSIM loss are used.

[0054] The teacher network takes low-resolution data and high-resolution prior as input, which 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 a splicing operator. The teacher network is only used in the training stage, and at this time, since synthetic data is used for training, accurate high-resolution prior information can be obtained.

[0057] The specific architecture of the seismic high-resolution teacher network is as follows: Figure 2As shown, the teacher network extracts features from high-resolution priors through convolution layers, and extracts features from low-resolution seismic data through convolution layers and up-sampling layers. This is because the resolution of low-resolution seismic data is low, and directly using it for subsequent feature fusion may result in insufficient seismic structural information, thereby limiting the model's ability to reconstruct the target high-resolution structure, so it is necessary to compensate for the lack of spatial resolution of low-resolution seismic data through up-sampling layers to ensure that the feature maps do not lose information due to resolution differences in subsequent processing. Convolution 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 spliced to form a joint feature tensor, which is used as the input of the U-net network. In addition, the high-resolution output is stacked through the residual layer to itself using the residual module, so as to learn more high-frequency information from the input data and improve the detail performance.

[0058] To constrain the teacher network to generate high-resolution seismic data, L1 loss and structural similarity loss (SSIM) are used as the loss function of the teacher network. SSIM loss focuses on the texture details and structural similarity of seismic data. After adding SSIM loss, the model will tend to preserve more detailed information, such as seismic wave reflection characteristics and geological boundaries, which is beneficial to improve the fidelity of seismic reflection information. The L1 loss is defined as follows:

[0059]

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

[0061] SSIM is used to evaluate the similarity of the predicted image and the target image in 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. Its definition is as follows:

[0062]

[0063] wherein, x and y represent two seismic data, respectively, x and μ y represent the mean of the seismic data, and represent the standard deviation of the seismic data, σ xy represents the covariance between x and y, and C1 and C2 are stability constants to avoid zero denominator. The SSIM loss takes 1-SSIM(x, y) as the optimization goal to minimize the structural difference between the seismic data, L SSIM is defined as:

[0064]

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

[0066] L Teacher = a * L Voxel + (1-a) * L SSIM (8)

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

[0068] The teacher network effectively obtains geological structure information in the training phase through accurate high-resolution priors. However, it is difficult to obtain accurate and reliable prior information when processing real seismic data in the inference phase. To utilize high-resolution priors in the inference phase, the present application 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, so that it is not necessary to directly input prior information or estimate priors.

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

[0070] L Student = a * L Pixel + b * L Feature + g * L Soft + k * L SSIM (9)

[0071] Where a, b, g and k are the weights of L Pixel , L Feature , L Soft and L SSIM , respectively. In order to transfer high-resolution seismic priors from the teacher network to the student network, the present application introduces prior knowledge distillation losses such as soft label loss L Soft and feature space loss L Feature . In addition, relying only on pixel-level differences may not be enough 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 label data in terms of seismic features. L Pixel is a pixel-wise L1 loss function applied to and I HR , and its formula is as follows:

[0072]

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

[0074] In knowledge distillation, the high-resolution predictions output by the teacher network can be considered as soft labels, while the actual high-resolution data is called hard labels. Soft labels can include varying degrees of fuzziness and uncertainty, broadening the learning objectives of student networks and improving their adaptability and generalization ability to unknown seismic data. Furthermore, soft labels do not set overly explicit and difficult goals for student networks, but rather provide a smoother learning path, helping students gradually align with the teacher network. Therefore, soft labels can provide simpler yet more effective supervision for student networks. The soft label loss, based on L1 loss, is used to measure the difference between the student network output and the teacher network output, as shown in the following formula:

[0075]

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

[0077] The teacher network uses high-resolution prior information during training, and its intermediate features contain rich prior information. Since the student network lacks prior information during the inference phase, it can learn these features through feature space loss to compensate for the missing information. Therefore, to explicitly constrain the intermediate features of the teacher and student networks, feature space loss is used to calculate the difference between the output features of the teacher and student networks at certain specific layers (usually intermediate or hidden layers), as shown in the following formula:

[0078]

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

[0080] The overall structure of a high-resolution seismic data network based on knowledge distillation is as follows: Figure 4 As shown. For the teacher network, the convolutional layers extract prior data P and low-resolution data I, respectively. LR Features fa and f b , and the feature splicing is performed to form a joint feature tensor as the input of the teacher network; the intermediate layer features F T of each convolutional layer are stored and returned during the teacher training process, facilitating the intermediate layer supervision in knowledge distillation; and finally, the high-resolution output is superimposed on itself through a residual layer to improve the detail performance. The teacher network outputs the intermediate layer features F T and the high-resolution prediction result for the student network. The student network takes the low-resolution data features of the convolutional layer as the input of the student network, loads the teacher network weights to provide the intermediate layer features F T and the high-resolution output During the training process, the intermediate layer features F S of each convolutional layer are also stored and returned, and finally, the residual layer is used to further learn the high-frequency detail information.

[0081] Based on the above-mentioned synthetic data construction and network building process, the model training process of the present application can be divided into the following steps, and the flow chart is shown in Figure 5 .

[0082] Step 1: Generating seismic data

[0083] First, the low-resolution input data I LR , the high-resolution prior data P and the high-resolution label data I HR are synthesized by using the reflectivity model and the Ricker wavelet convolution, and the synthetic seismic data is subjected to data enhancement processing.

[0084] Step 2: Training the teacher network

[0085] The features f a are extracted from the high-resolution prior through the convolutional layer, the features f b are extracted from the low-resolution seismic data through the convolutional layer and the upsampling layer, the f a and f b are subjected to feature splicing as the input of the U-net network. The weights a1 and a2 are adjusted, the L Teacher loss is used to train the teacher network, and the intermediate layer features F T and the high-resolution prediction data are output, and the prediction results are evaluated by PSNR and SSIM.

[0086] Step 3: Training the student network

[0087] The model parameters of the trained teacher network are loaded, including the intermediate features and the high-resolution prediction results of the teacher network. The features are extracted from the low-resolution seismic data through the convolutional layer as the input of the U-net network. The LStudent Loss training student network: calculate L1 loss, difference of intermediate features of student and teacher network, difference of student and teacher output and SSIM loss. Dynamically adjust the weight of L Student Loss: the weight of distillation loss is larger at the beginning of training, and gradually decreases as the training proceeds. Output high-resolution results of student network And evaluated by PSNR and SSIM.

[0088] Step4: transfer learning

[0089] When the network trained by synthetic seismic data is directly used to test the actual work area data, due to the difference between seismic data and seismic wavelet, the test result may not be accurate enough, and the actual work area data is not enough to retrain the network, so the transfer learning method is used to fine-tune the network, and a small amount of actual work area data is used to update part of the parameters of the pre-trained network, 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 residual layers of the network are trained to update their weights so that they can adapt to specific features of the actual work area data.

[0091] In one embodiment, Figure 6 The reflectivity of part of the Marmousi2 model is shown. The reflectivity of the Marmousi2 model is convolved with low-frequency wavelet and high-frequency wavelet respectively to construct low-resolution seismic data as input data and high-resolution seismic data as label, and then a middle-frequency wavelet is randomly selected to convolve with the low-frequency wavelet and the high-frequency wavelet as high-resolution prior information. In order to enhance the generalization of the network, the low-frequency wavelet is a random Ricker wavelet with a main frequency of 20-30Hz, the high-frequency wavelet is a Ricker wavelet with a main frequency of 60Hz, and the middle-frequency wavelet is a Ricker wavelet with a main frequency of 40Hz. The sampling time is 1ms. 30 groups of high-resolution data with a size of 1000x300 and corresponding low-resolution data and high-resolution prior are constructed. By setting appropriate sliding step, the 1000x300 synthetic data is divided into 7400 64x64 synthetic data, and random data enhancement operations such as up-down flipping, counterclockwise rotation by 90 degrees, rotation by 180 degrees, etc. are performed. The ratio of training set and validation set is set to 9:1.

[0092] First, L Teacher Loss training teacher network, Figure 7 L Teachercorresponding to different loss weights α in FIG. 2, and Table 1 shows the corresponding PSNR and SSIM values. It can be seen that the loss weight affects the speed of the loss function, and when the loss weight is 1.0, i.e. L Teacher When the weight of the SSIM loss in L is 0, the loss function drops too fast, which shows the importance of the SSIM loss. Therefore, the training parameters of the teacher network are set as follows: the loss weight is 0.7, the learning rate is 0.00001, the optimizer is the Adam optimizer, the number of iterations is 30, and the batch size is 16. Figure 8 FIG. 3 shows the loss curves of the teacher network on the training set and the validation set during the training process, and the PSNR and SSIM values on the validation set.

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

[0094]

[0095] In an embodiment, the trained teacher network is used to test the entire Marmousi2 model, Figure 9 (a) of FIG. 4 is low-resolution synthetic seismic data obtained by convolution of a 30Hz Ricker wavelet and reflection coefficients, Figure 9 (b) of FIG. 4 is high-resolution synthetic data obtained by convolution of a 60Hz Ricker wavelet, Figure 9 (c) is the high-resolution prediction result using the teacher network, Figure 9 (d)-(f) of FIG. 4 show the Figure 9 (a)-(c) of FIG. 4 are local enlargements of the black boxes. It can be seen that the high-resolution result of the teacher network is very close to the high-resolution label data, the local structural information is clear, the resolution of the data is effectively improved, and the result shows that the teacher network with high-resolution prior has a significant effect on high-resolution processing of seismic data.

[0096] In an embodiment, the trained teacher network is used to train the student network, the parameters of the teacher network 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, and the loss function is L Student wherein L Pixel the weight α = 1.4, L Soft the weight γ = 0.1, L SSIM the weight κ = 0.4. L Feature The weight needs to be dynamically adjusted, because at the beginning of the training, the student network has not fully learned the basic features, and if L FeatureThe weight of the feature space loss is too low, and the student network may deviate from the target provided by the teacher network, so a higher weight is needed to force the student network to be as close as possible to the characteristics of the teacher network. As the training progresses, the student network gradually learns some features and establishes a preliminary representation ability. At this time, if the weight of the feature space loss remains too high, it may limit the freedom of the student network, making it difficult to further optimize. Therefore, the weight of the feature space loss needs to be gradually reduced to relax the constraints and allow the student network to flexibly adjust under the guidance of the teacher network to better learn the details of the specific task. Feature The weight β is defined as:

[0097]

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

[0099] The loss function, SSIM curve, and PSNR curve of the student network during training are shown in Figure 10 It can be seen that the loss decreases significantly with the increase of the number of training times, and the loss decreases rapidly in the early stage of training, and gradually stabilizes after 10 times. The trends of the training set and the validation set are basically the same, and the SSIM curve and the PSNR curve of the student network are very similar to the results of the teacher network. Figure 11 The high-resolution results of the student network are shown, which are very close to the results of the teacher network. This indicates that the knowledge distillation technology successfully transfers the high-resolution priori in the teacher network to the student network, and the student network can directly reconstruct high-resolution data from low-resolution data without inputting priori information.

[0100] Actual seismic data from a certain work area are selected for testing, with a seismic profile size of 500x400 and a sampling rate of 4ms. Due to the differences between the seismic wavelet of the Marmousi2 model and the actual work area data, the network trained by the Marmousi2 model cannot achieve the expected results when directly used to test the actual data. Therefore, the transfer learning method is used. First, the actual seismic data are divided into data blocks of a certain size, and 20 of them are randomly selected as low-resolution inputs. Then, high-resolution data are constructed from these selected data using the spectral continuation method to obtain 20 pairs of high-resolution and low-resolution data. The network that has been trained is fine-tuned using these data, and then the entire seismic profile is tested. Figure 12 The original seismic data and the high-resolution results of the student network after fine-tuning 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 previously difficult to identify are clearer (black square part), revealing more detailed information and being beneficial to more accurate reservoir characterization and stratigraphic structure analysis.

[0101] Further, the 250th trace of the actual seismic data profile is selected, and the spectra before and after high resolution processing are compared, as shown in FIG. 6. Figure 13 As can be seen, the student network high resolution method compensates for the high frequency portion of the data, and the dominant frequency and frequency bandwidth are improved to some extent, without damaging the low frequency portion.

[0102] While the application has been described herein with reference to particular embodiments thereof, a latitude of modification, various changes and substitutions are intended in the foregoing disclosures, and in some instances some features of the present application will be employed without a corresponding use of other features. Accordingly, it is intended that the application be construed broadly and construed as including all such modifications, various features, and substitutions, and that the application is intended to cover by the appended claims any and all adaptations, modifications, equivalent processes, and sub-combinations of the elements of the systems and methods described herein. It is intended that each of the individual elements, features, compositions, articles, equipment, and / or methods that are described herein can be utilized to the exclusion of one or more other individual elements, features, compositions, articles, equipment, and / or methods, unless otherwise stated. It is also contemplated that some embodiments can consist of, consist essentially of, or consist of, any of the compositions, articles, equipment, and / or methods described herein.

Claims

1. A method for seismic data resolution based on knowledge distillation, characterized in that, The method comprises the following steps: S1, using low-resolution seismic data as input data, high-resolution seismic data as label, randomly selecting middle frequency wavelet of low frequency wavelet and high frequency wavelet for convolution, taking the high-resolution prior information as synthetic seismic data; the normalized synthetic seismic data is subjected to data enhancement; the synthetic data is subjected to slicing in the form of sliding window to obtain slice data with proper size; S2, a high-resolution seismic data network based on knowledge distillation is built, the network framework comprises a teacher network and a student network, and the knowledge distillation technology is used to transfer the high-resolution prior in the teacher network to the student network; S3, after the teacher network and the student network are trained, the network is fine-tuned through the transfer learning method, and a small amount of actual work area data is used to update part of the parameters of the pre-trained network; The teacher network is trained by 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 describes the underground structure details through real high-resolution data, and the student network is guided through soft target loss and feature space loss, so that it can restore real high-resolution details from low-resolution seismic data; The teacher network extracts features from high-resolution prior through a convolution layer, and extracts features from low-resolution seismic data through a convolution layer and an up-sampling layer; The convolution layer is used for extracting deep features, and the connection layer is used for feature fusion; the deep features extracted from the high-resolution prior and the low-resolution data are spliced to form a joint feature tensor, which is used as the input of the U-net network; 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 detail performance; The high-resolution prediction result output by the teacher network is regarded as a soft label, and the real high-resolution data is regarded as a hard label; the soft label includes different degrees of ambiguity and uncertainty; For the teacher network, the convolution layer extracts the features of the prior data and the low-resolution data respectively, splices the features to form a joint feature tensor, which is used as the input of the teacher network; the intermediate layer features of each convolution layer are stored and returned during the teacher training process, so as to facilitate the intermediate layer supervision in the knowledge distillation; finally, the high-resolution output is superimposed on itself through the residual layer.

2. The method of claim 1, wherein, The data enhancement includes random up-down flipping, 90-degree or 180-degree counterclockwise rotation.

3. The method of claim 1, wherein, The teacher network and the student network both adopt the U-net structure.

Images