Broadband speed model reconstruction method and device based on Res-SKUnet network
By using the broadband speed model reconstruction method of the Res-SKUnet network in seismic velocity modeling, combining the initial velocity model and seismic offset profile data, and using ResNet and attention mechanism modules, the reconstruction from low-frequency to broadband speed model was successfully achieved, solving the challenges of velocity modeling accuracy and computing efficiency in traditional methods.
Patent Information
- Application Number
- CN202311664323.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-06
- Publication Date
- 2025-06-06
AI Technical Summary
Traditional seismic velocity modeling methods are difficult to obtain broadband velocity models with wider frequency domains and more accurate velocities, especially in terms of computing efficiency and data integrity.
A broadband speed model reconstruction method based on Res-SKUnet network is adopted. By combining the initial velocity model and seismic offset profile as input data sets, the neural network is used to extract the velocity feature information, and the ResNet structure block and attention mechanism module are introduced into the U-Net architecture to form the Res-SKUnet architecture to update the velocity feature information.
The reconstruction from the fuzzy low-frequency initial speed model to the accurate broadband speed model is realized. The reconstruction speed layer is clearer, the construction is clearer, the frequency domain is wider, and the information is richer, which improves the accuracy and calculation efficiency of speed modeling.
Smart Images
Figure CN120103467A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of petroleum exploration velocity modeling, and more specifically, to a broadband velocity model reconstruction method and device based on a Res-SKUnet network. Background Art
[0002] In geophysical exploration, underground medium velocity is a very critical parameter. Accurate velocity determination is crucial for observation system design, precise positioning of underground geological targets, structural interpretation and reservoir prediction, and the accuracy of velocity directly affects all aspects of seismic exploration and the final results. The seismic wave propagation velocity parameter runs through the entire seismic exploration process of seismic data acquisition, processing and interpretation. The results of velocity analysis not only affect the imaging effect, but more importantly, the reliability of imaging and interpretation results. Therefore, seismic velocity is one of the most important parameters in seismic exploration. With the gradual increase in the amount of seismic exploration data, conventional seismic velocity modeling methods face challenges in terms of stability, accuracy and computational efficiency.
[0003] In the field of seismic exploration, accurately establishing a broadband velocity model is very important for subsequent data processing and interpretation. However, traditional velocity modeling methods (such as tomography) can only obtain the low-frequency part of the true velocity, while seismic reflection provides medium and high frequency information, and there is a missing frequency band between the two. Full waveform inversion can use reflection and transmission information to fill the missing frequency band and obtain an accurate broadband velocity model, but its actual application is limited by the reflection wave observation system, and it is impossible to obtain full wavelength information, and the calculation cost is high. Summary of the invention
[0004] In view of this, the present invention discloses a broadband velocity model reconstruction scheme based on the Res-SKUnet network, which combines the initial velocity model and the seismic migration profile as input data sets, and can reconstruct the untrained initial velocity model into a velocity model with a wider frequency domain and more accurate velocity through a suitable neural network. The reconstructed velocity horizon is clearer, the structure is clearer, the frequency domain is wider, and the information is richer.
[0005] According to one aspect of the present invention, a broadband speed model reconstruction method based on a Res-SKUnet network is proposed, the method comprising:
[0006] Step 1, determine the input data set;
[0007] Step 2, extracting speed feature information from the input data set through a neural network;
[0008] Step 3, add ResNet structural blocks to the U-Net downsampling and upsampling process to form the Res-SKUnet architecture;
[0009] Step 4: Update the speed feature information through the Res-SKUnet architecture.
[0010] In some embodiments, in step 1, the input data set is configured as a three-dimensional data set including an initial velocity model and a reverse time migration profile.
[0011] In some embodiments, in step 2, the neural network includes a shallow neural network and a deep neural network, wherein the data body of the deep neural network and the data body of the shallow neural network are connected to form a data body of high- and low-level feature fusion.
[0012] In some embodiments, in step 3, an attention mechanism module is introduced on the U-Net connection path.
[0013] In some embodiments, the attention mechanism module includes a convolution layer with two different convolution kernels, a data merging layer, a global average pooling layer, a 1*1 convolution layer, a softmax layer, a weight allocation layer and a data merging layer.
[0014] In some implementations, the specific steps of introducing the attention mechanism module on the U-Net connection path are as follows:
[0015] Step S1, using two convolution kernels to perform convolution operations on the input data of the U-Net connection path respectively, the data volume height and width are H and W, and the number of feature channels is C;
[0016] Step S2, adding corresponding elements of the velocity feature data volumes obtained by different convolution kernels;
[0017] Step S3, performing a global average pooling operation;
[0018] Step S4, using a 1*1 convolution kernel to double the number of global average pooling feature channels, the number of feature channels changes from C to 2C;
[0019] Step S5, divide the data after 1*1 convolution into two parts, and process them respectively using softmax function;
[0020] Step S6, using the output result of the softmax function to multiply the corresponding elements of different convolution channel data volumes, so as to achieve the effect of redistributing the feature channel weights;
[0021] Step S7, add the corresponding elements of the data body after the weights of different convolution channels are assigned, and output the result of the Attention module.
[0022] In some embodiments, in step S3, the Res-SKUnet architecture formed includes an encoding part and a decoding part, wherein the encoding part is composed of four network layer groups including 3*3 convolutional layers, BN layers and ReLU functions, three residual modules and three 2*2 maximum pooling layers; the decoding part is composed of three network layer groups including 3*3 convolutional layers, BN layers and ReLU functions, three residual modules and three deconvolution layers.
[0023] According to one aspect of the present invention, a broadband speed model reconstruction system based on a Res-SKUnet network is also proposed, the system comprising:
[0024] A data input module, used to determine the input data set;
[0025] An information extraction module, used for extracting speed feature information from the input data set through a neural network;
[0026] The network reconstruction module is used to add ResNet structural blocks during the downsampling and upsampling process of U-Net to form the Res-SKUnet architecture;
[0027] The data updating module is used to update the speed characteristic information through the Res-SKUnet architecture.
[0028] According to one aspect of the present invention, an electronic device is also provided, the electronic device comprising:
[0029] A memory storing executable instructions;
[0030] A processor runs the executable instructions in the memory to implement the above-mentioned broadband speed model reconstruction method based on the Res-SKUnet network.
[0031] According to one aspect of the present invention, a computer-readable storage medium is further provided, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned broadband speed model reconstruction method based on the Res-SKUnet network is implemented.
[0032] This technical solution has at least the following advantages:
[0033] The present invention proposes a method and device for reconstruction of broadband velocity model based on deep learning, which combines the initial velocity model and the seismic offset profile as input data sets, and constructs a new network architecture (Res-SKUnet) adapted to the reconstruction of the velocity model. The neural network combining the shallow neural network and the deep neural network can reconstruct the untrained initial velocity model into a velocity model with a wider frequency domain and more accurate velocity, and successfully realizes the reconstruction of the fuzzy low-frequency initial velocity model into an accurate broadband velocity model in a complex model. The velocity layer of the reconstructed velocity model is clearer, the structure is clearer, the frequency domain is wider, and the information is richer, which verifies the correctness, effectiveness and adaptability of the present invention for broadband velocity reconstruction. It also shows that the deep learning broadband velocity model reconstruction method has a good practical application prospect.
[0034] The methods and apparatus of the present invention have other features and advantages that will be apparent from or will be described in detail in the accompanying drawings and subsequent detailed descriptions incorporated herein, which together serve to explain the specific principles of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.
[0036] Figure 1 A flowchart of a broadband speed model reconstruction method based on a Res-SKUnet network according to an embodiment of the present invention is shown;
[0037] Figure 2 A schematic diagram of a basic architecture of a broadband speed model reconstruction neural network according to an embodiment of the present invention is shown;
[0038] Figure 3 A schematic diagram of a residual module in a network according to an embodiment of the present invention is shown;
[0039] Figure 4 A schematic diagram of a SKNet module according to an embodiment of the present invention is shown;
[0040] Figure 5 A schematic diagram of an initial velocity model of actual data according to an embodiment of the present invention is shown;
[0041] Figure 6 shows a schematic diagram of an offset cross-section according to an embodiment of the present invention;
[0042] Figure 7A schematic diagram of a reconstructed model result according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0043] The preferred embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the preferred embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.
[0044] The present invention proposes a broadband speed model reconstruction method based on a Res-SKUnet network, the method comprising:
[0045] Step 1, determine the input data set;
[0046] Step 2, extracting speed feature information from the input data set through a neural network;
[0047] Step 3, add ResNet structural blocks to the U-Net downsampling and upsampling process to form the Res-SKUnet architecture;
[0048] Step 4: Update the speed feature information through the Res-SKUnet architecture.
[0049] In some embodiments, in step 1, the input data set is configured as a three-dimensional data set including an initial velocity model and a reverse time migration profile.
[0050] In some embodiments, in step 2, the neural network includes a shallow neural network and a deep neural network, wherein the data body of the deep neural network and the data body of the shallow neural network are connected to form a data body of high- and low-level feature fusion.
[0051] In some embodiments, in step 3, an attention mechanism module is introduced on the U-Net connection path.
[0052] In some embodiments, the attention mechanism module includes a convolution layer with two different convolution kernels, a data merging layer, a global average pooling layer, a 1*1 convolution layer, a softmax layer, a weight allocation layer and a data merging layer.
[0053] In some implementations, the specific steps of introducing the attention mechanism module on the U-Net connection path are as follows:
[0054] Step S1, using two convolution kernels to perform convolution operations on the input data of the U-Net connection path respectively, the data volume height and width are H and W, and the number of feature channels is C;
[0055] Step S2, adding corresponding elements of the velocity feature data volumes obtained by different convolution kernels;
[0056] Step S3, performing a global average pooling operation;
[0057] Step S4, using a 1*1 convolution kernel to double the number of global average pooling feature channels, the number of feature channels changes from C to 2C;
[0058] Step S5, divide the data after 1*1 convolution into two parts, and process them respectively using softmax function;
[0059] Step S6, using the output result of the softmax function to multiply the corresponding elements of different convolution channel data volumes, so as to achieve the effect of redistributing the feature channel weights;
[0060] Step S7, add the corresponding elements of the data body after the weights of different convolution channels are assigned, and output the result of the Attention module.
[0061] In some embodiments, in step S3, the Res-SKUnet architecture formed includes an encoding part and a decoding part, wherein the encoding part is composed of four network layer groups including 3*3 convolutional layers, BN layers and ReLU functions, three residual modules and three 2*2 maximum pooling layers; the decoding part is composed of three network layer groups including 3*3 convolutional layers, BN layers and ReLU functions, three residual modules and three deconvolution layers.
[0062] The present invention also proposes a broadband speed model reconstruction system based on the Res-SKUnet network, the system comprising:
[0063] A data input module, used to determine the input data set;
[0064] An information extraction module, used for extracting speed feature information from the input data set through a neural network;
[0065] The network reconstruction module is used to add ResNet structural blocks during the downsampling and upsampling process of U-Net to form the Res-SKUnet architecture;
[0066] The data updating module is used to update the speed characteristic information through the Res-SKUnet architecture.
[0067] The present invention further provides an electronic device, comprising:
[0068] A memory storing executable instructions;
[0069] A processor runs the executable instructions in the memory to implement the above-mentioned broadband speed model reconstruction method based on the Res-SKUnet network.
[0070] The present invention also proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned broadband speed model reconstruction method based on the Res-SKUnet network is implemented.
[0071] Example 1
[0072] Figure 1 The flowchart of the broadband speed model reconstruction method based on the Res-SKUnet network according to an embodiment of the present invention is shown. As shown in the figure, the method includes steps 1 to 4.
[0073] Step 1, determine the input data set, the input data set is configured as a three-dimensional data set including the initial velocity model and the reverse time migration profile, the initial seismic velocity field and the migration imaging results are used together, and the low-frequency information of the velocity model and the medium- and high-frequency information of the migration imaging results are used to realize the construction of the broadband velocity model. Among them, the input data set and the general image processing have three different feature channels. Except for the most basic height and width, the number of feature channels is 2 when input to the network, and the network output is a reconstructed broadband high-precision seismic velocity model, whose height and width are consistent with those before reconstruction, but the number of feature channels is 1.
[0074] Step 2: Extract velocity feature information, such as seismic velocity feature information, from the input data set through a neural network.
[0075] Step 3, add ResNet structural blocks in the process of U-Net downsampling and upsampling to form the Res-SKUnet architecture. The combination of U-Net and ResNet modules can make full use of the seismic velocity feature information extracted by the neural network, while ensuring the stability of the network without degradation. Compared with other network reconstruction results, the Res-SKUnet network can achieve higher-precision broadband velocity model reconstruction. In general, the network structure of this technology has important theoretical and practical value for the broadband velocity model reconstruction problem.
[0076] Specifically, U-net is one of the commonly used backbone networks of convolutional neural networks. It has the advantages of complete symmetry and input of any size. At the same time, it integrates shallow network information in the deep network, making the network more powerful. Multiple upsampling also makes the final output clearer. It is a very good basic network for velocity model reconstruction. In the process of neural network learning seismic velocity reconstruction, shallow seismic velocity features can be passed to deeper networks and combined with seismic velocity features extracted by deep networks, making the best use of the information of training seismic data to obtain more refined broadband reconstruction results of velocity models. ResNet can enhance the stability of seismic velocity reconstruction networks and prevent the network from degenerating as the number of layers increases. At the same time, ResNet blocks are lightweight modules that can be added to existing networks as needed.
[0077] Step 4: Update the seismic velocity characteristic information through the Res-SKUnet architecture.
[0078] In one embodiment of the present invention, in step 2, the neural network includes a shallow neural network and a deep neural network, wherein the data body of the deep neural network and the data body of the shallow neural network are connected to form a data body of high- and low-level feature fusion.
[0079] In one embodiment of the present invention, in step 3, an attention mechanism module is introduced on the U-Net connection path; the attention mechanism module includes two convolution layers with different convolution kernels, a data merging layer, a global average pooling layer, a 1*1 convolution layer, a softmax layer, a weight distribution layer and a data merging layer. By adding the attention module of SKNet, it is possible to learn the weight relationship of the velocity feature channels of the input seismic data and update the weights of different velocity features.
[0080] In one embodiment of the present invention, the specific steps of introducing the attention mechanism module on the U-Net connection path are as follows:
[0081] Step S1, use two convolution kernels (3*3 convolution kernel and 5*5 convolution kernel) to perform convolution operations on the input data of the U-Net connection path respectively, the height and width of the data volume are H and W, and the number of feature channels is C;
[0082] Step S2, adding corresponding elements of the velocity feature data volumes obtained by different convolution kernels;
[0083] Step S3, performing a global average pooling operation;
[0084] Step S4, using a 1*1 convolution kernel to double the number of global average pooling feature channels, the number of feature channels changes from C to 2C;
[0085] Step S5, divide the data after 1*1 convolution into two parts, and process them respectively using softmax function;
[0086] Step S6, using the output result of the softmax function to multiply the corresponding elements of different convolution channel data volumes, so as to achieve the effect of redistributing the feature channel weights;
[0087] Step S7, add the corresponding elements of the data body after the weights of different convolution channels are assigned, and output the result of the Attention module.
[0088] In one embodiment of the present invention, in step S3, the Res-SKUnet architecture formed includes an encoding part and a decoding part, wherein the encoding part is composed of four network layer groups including 3*3 convolutional layers, BN layers and ReLU functions, three residual modules and three 2*2 maximum pooling layers; the number of neural network feature channels first changes from 2 to 128 after passing through the convolutional layer, and then doubles each time from 128 to 512, and the height and width of each input data changes from 128*128 to 16*16 after three pooling operations, and the data gradually extracts high-precision features of the speed model during the encoding process. The decoding part is composed of three network layer groups including 3*3 convolutional layers, BN layers and ReLU functions, three residual modules and three deconvolution layers. During this process, the number of neural network channels is gradually reduced from 512 to 128 in multiples of two, and after each deconvolution, a new data body is formed with the data output by the Attention module to perform the next convolution operation. Finally, after the last convolution operation, the number of feature channels changes from 128 to 1, and the height and width of the data are restored from 16*16 at the beginning of decoding to 128*128. Finally, the reconstructed wide-band high-precision speed model is output.
[0089] Specifically, the residual module in the network structure is as follows Figure 3 As shown in the figure, the residual module has two convolutional layers, using a 3*3 convolution kernel. After each convolutional layer, a BN layer is used for normalization to improve the training speed of the network. After that, the ReLU activation function is used. The input data passes through the residual module for a series of training operations, and the results are combined with the short-circuit connection input data above as the residual module output. The residual module in the seismic velocity model broadband reconstruction Res-SKUnet architecture can effectively overcome the network degradation problem caused by the deepening of the network depth. The residual module can learn new velocity features based on the extracted velocity features of the input, which is conducive to improving the stability of the velocity model reconstruction learning network and making the Res-SKUnet reconstruction effect better.
[0090] The deep learning neural network adopted in the present invention is based on U-Net as the basic architecture, and a new Res-SKUnet (Residual Selective Kernel U-Networks) architecture is developed based on the characteristics of input seismic velocity data. Its backbone architecture is U-Net, which can make full use of the seismic velocity feature information extracted by shallow and deep neural networks. ResNet structural blocks are added during the U-Net downsampling and upsampling process to ensure the stability of the network and prevent network degradation. At the same time, the Attention mechanism module of SKNet is introduced on the U-Net connection path, so that it can learn the weight relationship between different seismic velocity feature channels and then update the feature channel weights, so that the network pays more attention to important feature channels and suppresses non-critical feature channels. The neural network architecture is as follows: Figure 2 shown.
[0091] The broadband velocity model reconstruction method based on deep learning of the present invention also constructs a new network architecture (Res-SKUnet) that is adapted to the reconstruction of the velocity model. The method successfully transforms the fuzzy low-frequency initial velocity model in the complex model into an accurate broadband velocity model, and the network trained with simulated seismic data is applied to the initial velocity reconstruction of real seismic data and achieves good results. This shows the effectiveness and versatility of the technology. The present invention uses deep learning technology to solve the problem of velocity model reconstruction for the first time, and constructs a new network structure that is adapted to the reconstruction of the velocity model. The results of its use have important theoretical and application value.
[0092] Example 2
[0093] According to an embodiment of the present invention, a broadband speed model reconstruction system based on a Res-SKUnet network is provided, and the device includes:
[0094] A data input module is used to determine an input data set, wherein the input data set is configured as a three-dimensional data set including an initial velocity model and a reverse time migration profile;
[0095] An information extraction module is used to extract seismic velocity feature information from an input data set through a neural network, wherein the neural network includes a shallow neural network and a deep neural network, wherein the data body of the deep neural network is connected with the data body of the shallow neural network to form a data body with high- and low-level feature fusion;
[0096] The network reconstruction module is used to add ResNet structural blocks in the process of U-Net downsampling and upsampling to form the Res-SKUnet architecture. The combination of U-Net and ResNet modules can make full use of the seismic velocity feature information extracted by the neural network while ensuring the stability and non-degradation of the network. Compared with other network reconstruction results, the Res-SKUnet network can achieve higher-precision broadband velocity model reconstruction. On the other hand, the attention mechanism module is introduced on the U-Net connection path; the attention mechanism module includes two convolution layers with different convolution kernels, a data merging layer, and a global average pooling layer. Layer, 1*1 convolution layer, softmax layer, weight distribution layer and data merging layer; Res-SKUnet architecture includes encoding part and decoding part, where the encoding part consists of four network layer groups including 3*3 convolution layer, BN layer and ReLU function, three residual modules and three 2*2 maximum pooling layers; the number of neural network feature channels first changes from 2 to 128 after passing through the convolution layer, and then doubles from 128 to 512 each time, and each input data height and width changes from 128*128 to 16*16 after three pooling operations. The data gradually extracts high-precision features of the speed model during the encoding process. The decoding part consists of three network layer groups including 3*3 convolution layer, BN layer and ReLU function, three residual modules and three deconvolution layers. In this process, the number of neural network channels is gradually reduced from 512 to 128 in multiples of two. After each deconvolution, the data output by the Attention module is combined into a new data body for the next convolution operation. Finally, after the last convolution operation, the number of feature channels changes from 128 to 1, and the height and width of the data are restored from 16*16 at the beginning of decoding to 128*128. Finally, the reconstructed broadband high-precision speed model is output;
[0097] The data update module is used to update the seismic velocity feature information through the Res-SKUnet architecture. By adding the SKNet attention module, it can learn the weight relationship of the velocity feature channels of the input seismic data and update the weights of different velocity features.
[0098] The deep learning neural network adopted in the present invention is based on U-Net as the basic architecture, and a new Res-SKUnet architecture is developed based on the characteristics of input seismic velocity data. Its backbone architecture is U-Net, which can make full use of the seismic velocity feature information extracted by shallow and deep neural networks. ResNet structural blocks are added during U-Net downsampling and upsampling to ensure the stability of the network and prevent network degradation. At the same time, the Attention mechanism module of SKNet is introduced on the U-Net connection path, so that it can learn the weight relationship of different seismic velocity feature channels and then update the feature channel weights, so that the network pays more attention to important feature channels and suppresses non-critical feature channels.
[0099] Example 3
[0100] According to another aspect of the present invention, there is also provided an electronic device, the electronic device comprising:
[0101] Memory, which stores executable instructions:
[0102] The processor runs the executable instructions in the memory to implement the broadband speed model reconstruction method based on the Res-SKUnet network according to the present invention.
[0103] The method comprises the following steps:
[0104] Step 1, determine the input data set;
[0105] Step 2, extracting seismic velocity characteristic information from the input data set through a neural network;
[0106] Step 3. Add ResNet structural blocks during the U-Net downsampling and upsampling process to form the Res-SKUnet architecture. The Res-SKUnet architecture includes an encoding part and a decoding part. The encoding part consists of four network layer groups including 3*3 convolution layers, BN layers and ReLU functions, three residual modules and three 2*2 maximum pooling layers; the decoding part consists of three network layer groups including 3*3 convolution layers, BN layers and ReLU functions, three residual modules and three deconvolution layers.
[0107] Step 4: Update the seismic velocity characteristic information through the Res-SKUnet architecture.
[0108] In one embodiment of the present invention, in step 1, the input data set is configured as a three-dimensional data set including an initial velocity model and a reverse time migration profile.
[0109] In one embodiment of the present invention, in step 2, the neural network includes a shallow neural network and a deep neural network, wherein the data body of the deep neural network is connected with the data body of the shallow neural network to form a data body of high- and low-level feature fusion. The data body of high- and low-level feature fusion can maximize the use of deep and shallow network feature channel information, and using it in the broadband reconstruction of the velocity model can reconstruct a more refined velocity model.
[0110] In one embodiment of the present invention, in step 3, an attention mechanism module is introduced into the U-Net connection path. The connection path of the U-Net itself directly concatenates the downsampled data body with the upsampled data body without any modification. The present invention introduces the Attention mechanism module of SKNet into the connection path. Figure 4As shown, it includes two convolution layers with different convolution kernels, a data merging layer, a global average pooling layer, a 1*1 convolution layer, a softmax layer, a weight distribution layer and a data merging layer.
[0111] In one embodiment of the present invention, the specific steps of introducing the attention mechanism module on the U-Net connection path are as follows:
[0112] Step S1, use two convolution kernels (3*3 convolution kernel and 5*5 convolution kernel) to perform convolution operations on the input data of the U-Net connection path respectively, the height and width of the data volume are H and W, and the number of feature channels is C;
[0113] Step S2, adding corresponding elements of the velocity feature data volumes obtained by different convolution kernels;
[0114] Step S3, performing a global average pooling operation;
[0115] Step S4, using a 1*1 convolution kernel to double the number of global average pooling feature channels, the number of feature channels changes from C to 2C;
[0116] Step S5, divide the data after 1*1 convolution into two parts, and process them respectively using softmax function;
[0117] Step S6, using the output result of the softmax function to multiply the corresponding elements of different convolution channel data volumes, so as to achieve the effect of redistributing the feature channel weights;
[0118] Step S7, add the corresponding elements of the data body after the weights of different convolution channels are assigned, and output the result of the Attention module.
[0119] Example 4
[0120] According to another aspect of the present invention, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the broadband speed model reconstruction method based on the Res-SKUnet network according to the present invention is implemented.
[0121] The method comprises the following steps:
[0122] Step 1, determining an input data set; the input data set is configured as a three-dimensional data set including an initial velocity model and a reverse time migration profile.
[0123] Step 2, extracting seismic velocity feature information from the input data set through a neural network; the neural network includes a shallow neural network and a deep neural network, wherein the data body of the deep neural network and the data body of the shallow neural network are connected to form a data body with high- and low-level feature fusion.
[0124] Step 3, add ResNet structural blocks in the process of U-Net downsampling and upsampling to form the Res-SKUnet architecture; introduce the attention mechanism module on the U-Net connection path, which includes two convolution layers with different convolution kernels, a data merging layer, a global average pooling layer, a 1*1 convolution layer, a softmax layer, a weight distribution layer and a data merging layer.
[0125] Step 4: Update the seismic velocity characteristic information through the Res-SKUnet architecture.
[0126] In one embodiment of the present invention, the specific steps of introducing the attention mechanism module on the U-Net connection path are as follows:
[0127] Step S1, using two convolution kernels to perform convolution operations on the input data of the U-Net connection path respectively, the data volume height and width are H and W, and the number of feature channels is C;
[0128] Step S2, adding corresponding elements of the velocity feature data volumes obtained by different convolution kernels;
[0129] Step S3, performing a global average pooling operation;
[0130] Step S4, using a 1*1 convolution kernel to double the number of global average pooling feature channels, the number of feature channels changes from C to 2C;
[0131] Step S5, divide the data after 1*1 convolution into two parts, and process them respectively using softmax function;
[0132] Step S6, using the output result of the softmax function to multiply the corresponding elements of different convolution channel data volumes, so as to achieve the effect of redistributing the feature channel weights;
[0133] Step S7, add the corresponding elements of the data body after the weights of different convolution channels are assigned, and output the result of the Attention module.
[0134] In one embodiment of the present invention, in step S3, the Res-SKUnet architecture formed includes an encoding part and a decoding part, wherein the encoding part is composed of four network layer groups including 3*3 convolution layers, BN layers and ReLU functions, three residual modules and three 2*2 maximum pooling layers; the decoding part is composed of three network layer groups including 3*3 convolution layers, BN layers and ReLU functions, three residual modules and three deconvolution layers.
[0135] Example 5
[0136] In order to verify the effect of the broadband velocity model reconstruction solution based on the Res-SKUnet network according to the present invention, this embodiment selects the seismic exploration processing environment for verification. In the actual seismic data application, the neural network trained by simulating the deep domain seismic data in the generalization test phase is used, which still has a good reconstruction effect on the actual velocity model, such as Figures 5 to 7 As shown, the reconstructed velocity horizon is clearer, the structure is clearer, the frequency domain is wider, and the information is richer, which verifies the correctness, effectiveness and adaptability of the invention for broadband velocity reconstruction. It also shows that the deep learning broadband velocity model reconstruction method has a good practical application prospect.
[0137] This example fully demonstrates that the present invention has constructed a new network architecture (Res-SKUnet) that is suitable for velocity model reconstruction, and successfully achieved the reconstruction of the fuzzy low-frequency initial velocity model in the complex model into an accurate broadband velocity model. The application of the network trained with simulated seismic data to the initial velocity reconstruction of real seismic data has also achieved good results, which fully verifies the effectiveness and versatility of this method.
[0138] The method and device for reconstructing a broadband velocity model based on deep learning of the present invention, combined with the initial velocity model and the seismic offset profile as the input data set, can reconstruct the untrained initial velocity model into a velocity model with a wider frequency domain and more accurate velocity by combining a shallow neural network with a deep neural network. In addition, the attention module of SKNet is added to the U-Net connection path, so that it can learn the weight relationship of the velocity feature channels of the input seismic data and update the weights of different velocity features. The reconstructed velocity horizon is clearer, the structure is clearer, the frequency domain is wider, and the information is richer, which verifies the correctness, effectiveness and adaptability of the present invention for broadband velocity reconstruction. It also shows that the deep learning broadband velocity model reconstruction method has a good practical application prospect.
[0139] For other detailed descriptions of this exemplary embodiment, reference may be made to the corresponding descriptions in the aforementioned embodiments, which will not be repeated here.
[0140] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terms used herein are selected to best explain the principles of the embodiments, practical applications, or technical improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.
Claims
1. A broadband speed model reconstruction method based on Res-SKUnet network, It is characterized in that The method comprises: Step 1, determine the input data set; Step 2, extracting speed feature information from the input data set through a neural network; Step 3, add ResNet structural blocks to the U-Net downsampling and upsampling process to form the Res-SKUnet architecture; Step 4: Update the speed feature information through the Res-SKUnet architecture.
2. The broadband speed model reconstruction method based on the Res-SKUnet network according to claim 1, It is characterized in that In step 1, the input data set is configured as a three-dimensional data set including an initial velocity model and a reverse time migration section.
3. The broadband speed model reconstruction method based on the Res-SKUnet network according to claim 1, It is characterized in that In step 2, the neural network includes a shallow neural network and a deep neural network, wherein the data body of the deep neural network and the data body of the shallow neural network are connected to form a data body with high- and low-level feature fusion.
4. The broadband speed model reconstruction method based on the Res-SKUnet network according to claim 1, It is characterized in that In step 3, an attention mechanism module is introduced on the U-Net connection path.
5. The broadband speed model reconstruction method based on the Res-SKUnet network according to claim 4, It is characterized in that The attention mechanism module includes a convolution layer with two different convolution kernels, a data merging layer, a global average pooling layer, a 1*1 convolution layer, a softmax layer, a weight distribution layer and a data merging layer.
6. The broadband speed model reconstruction method based on the Res-SKUnet network according to claim 4, It is characterized in that The specific steps of introducing the attention mechanism module on the U-Net connection path are as follows: Step S1, using two convolution kernels to perform convolution operations on the input data of the U-Net connection path respectively, the data volume height and width are H and W, and the number of feature channels is C; Step S2, adding corresponding elements of the velocity feature data volumes obtained by different convolution kernels; Step S3, performing a global average pooling operation; Step S4, using a 1*1 convolution kernel to double the number of global average pooling feature channels, the number of feature channels changes from C to 2C; Step S5, divide the data after 1*1 convolution into two parts, and process them respectively using softmax function; Step S6, using the output result of the softmax function to multiply the corresponding elements of different convolution channel data volumes, so as to achieve the effect of redistributing the feature channel weights; Step S7, add the corresponding elements of the data body after the weights of different convolution channels are assigned, and output the result of the Attention module.
7. The broadband speed model reconstruction method based on the Res-SKUnet network according to claim 1, It is characterized in that In step S3, the Res-SKUnet architecture formed includes an encoding part and a decoding part, wherein the encoding part is composed of four network layer groups including 3*3 convolution layers, BN layers and ReLU functions, three residual modules and three 2*2 maximum pooling layers; the decoding part is composed of three network layer groups including 3*3 convolution layers, BN layers and ReLU functions, three residual modules and three deconvolution layers.
8. A broadband speed model reconstruction system based on Res-SKUnet network, It is characterized in that include: A data input module, used to determine the input data set; An information extraction module, used for extracting speed feature information from the input data set through a neural network; The network reconstruction module is used to add ResNet structural blocks during the downsampling and upsampling process of U-Net to form the Res-SKUnet architecture; The data updating module is used to update the speed characteristic information through the Res-SKUnet architecture.
9. An electronic device, It is characterized in that The electronic device comprises: A memory storing executable instructions; A processor, wherein the processor runs the executable instructions in the memory to implement the method according to any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, wherein the computer program implements the method according to any one of claims 1 to 7 when executed by a processor.