Large-view seismic signal high-resolution inversion method

Abstract

The invention discloses a large-view seismic signal high-resolution inversion method, and the method comprises the steps: S1, building a training data set: carrying out the preprocessing of geological data, and intercepting the data to form two groups of training samples; s2, constructing and training a model: designing and constructing an inversion network model framework by taking Transform as a basic module; performing pre-training and re-training by using the two groups of training samples to obtain an inversion network model; and S3, prediction result output: based on the trained inversion network model, using the seismic data as input, outputting a prediction result corresponding to the seismic data, and processing the prediction result to obtain a high-resolution inversion result. According to the inversion method, the geological laws in the seismic data can be extracted on a large scale by using the global view characteristics of the Transform, and the formation parameters are predicted by using more comprehensive formation characteristic information, so that the accuracy and resolution of the inversion parameters are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of oil and gas exploration, and particularly relates to a large field of view seismic signal high resolution inversion method. BACKGROUND

[0002] Seismic exploration is one of the fastest developing and most widely applied oil and gas exploration technologies in modern times, and has the advantages of wide detection range and high imaging precision. Seismic signal inversion uses seismic data to obtain various elastic parameters of the underground medium by means of inversion method, and then predicts reservoir parameters and oil and gas distribution. The conversion method of seismic signal to stratum parameter used in seismic inversion realizes accurate description of the underground world, and provides a scientific basis for the detection and development of oil and gas resources, and has important strategic significance for ensuring national economic stability and energy security.

[0003] Traditional seismic inversion methods use mathematical equations to simulate the propagation process of seismic waves, associate seismic wave responses with underground medium properties based on mathematical models, and calculate stratum parameters such as velocity, density and porosity by minimizing the difference between synthetic seismic data and actual seismic data. With the continuous consumption of oil and gas resources and the increasing difficulty of exploitation, oil and gas exploration and development are facing more and more complex challenges. Traditional stratum parameter inversion methods often have problems such as insufficient precision and low efficiency when dealing with complex geological backgrounds, massive data and nonlinear problems. In order to effectively cope with these challenges and improve the precision and efficiency of stratum inversion, intelligent learning methods based on data driving have emerged as the times require, and in particular, deep learning and neural network technologies have become key means to improve inversion precision and reliability.

[0004] Methods based on neural networks rely on the processing power of computers for large-scale seismic data, learn patterns and features in the data by analyzing a large amount of data, establish a nonlinear mapping relationship from input data to output data based on a large number of network parameters, and apply this knowledge to new data sets. The generalization of neural networks makes them exhibit significant advantages in seismic inversion problems. However, both traditional seismic inversion and inversion methods based on neural networks are in the form of fitting the corresponding relationship between a single seismic gather and a single stratum parameter. This fitting mode results in the limited view of the neural network, and the information extracted from the input data is limited. Adding additional constraint information or increasing the view of the neural network can help the network model find the inherent correlation structure in the training data, thereby improving the generalization performance of the network.

[0005] Many methods have been developed by researchers to improve the geological features extracted by neural networks, which can be divided into three categories. The first category is to add preset constraint rules in the loss function to guide the training process of the neural network and make the neural network integrate these constraint information. For example, adding L1 norm constraint in the loss function can make the network output results with obvious interface or blocky features. If L2 norm constraint is added to the loss function, the network output can be smooth. Researchers have also developed some mixed norm constraint information for work areas with diverse geological structures. These pre-set constraint information works better in areas with relatively simple sedimentary environment and clear geological structure. However, in complex work areas with faults and strong heterogeneity, this method still has great limitations. The second category of methods takes the attribute data extracted from seismic data as additional input of the neural network, and integrates the features contained in the seismic attribute into the network parameters during the training process of the network. For example, embedding relative geologic time (RGT) attribute in the input data can enhance the geological time continuity of the inversion results. If seismic reflection coefficient is used as the input of the neural network, the neural network can better fit the reflection pattern of seismic wave and improve the consistency of the inversion results with the seismic data. This method enriches the geological features extracted by the neural network to a certain extent, but also loses the ability to further explore and integrate the geological features in the seismic data itself. The third category of methods considers expanding the field of view of the neural network, so that the neural network can extract more complete geological features from the seismic data itself. For example, adding global pooling, larger convolution kernel, and other modules such as dilated convolution to improve the ability of the model to capture long-range dependencies and global information, and increase the field of view of the network model in the input data. These methods can to some extent alleviate the limitation of the field of view of the neural network, but the ability to capture complex patterns and long-range dependencies is insufficient, resulting in low resolution and accuracy of the inversion results in complex geological structure work areas.

[0006] A semi-supervised seismic inversion method based on spatiotemporal sequence residual model was proposed by researchers in 2022, which can fully exploit the spatiotemporal features of the data. This method takes the initial model as the input of the network model, and makes the network learn the correction value of the initial model and the residual between the corrected initial model and the true formation parameters during the training process. The overall framework of this method is as follows Figure 1As shown, the inversion network mainly includes two parts: (1) The inversion network uses the initial model data of the i-th trace and the seismic data of the i-th trace and its adjacent 2k traces to predict the formation parameters of the i-th trace. Then the predicted formation parameters of the i-th trace are input into the forward network to synthesize the corresponding seismic data. The mean square error between the synthesized seismic data and the input seismic data is calculated to adjust the parameters of the inversion network and the forward network. (2) The well trace seismic data and its adjacent 2k seismic data and the initial model are input into the inversion network to predict the formation parameters at the logging position. The mean square error between the predicted formation parameters and the logging data is calculated to update the parameters in the inversion network. According to the actual logging formation parameters, the forward network is used to calculate the synthesized seismic data at the position. The mean square error between the well trace seismic data and the synthesized seismic data is calculated to update the parameters in the forward network.

[0007] The pre-stack seismic data is collected according to the propagation law of seismic waves, and contains rich geological information. In practical application, whether it is a traditional seismic inversion or an inversion method based on neural network, the corresponding relationship between a single seismic trace set and a single formation parameter is fitted. This fitting mode causes the field of view of the neural network to have limitations, and the information extracted from the input data is limited. The above technical solution takes the 2k adjacent seismic data at the target position as input, which can alleviate the problem of limited field of view of the neural network. However, the module for extracting data features in this scheme only has a convolution layer, and the convolution module is insufficient in capturing complex patterns and long-range dependencies in data, which restricts the ability of the network model to deal with complex geological structures. In the area with complex formation structure, limited by the lower data quality and the complexity of the formation deposition process, the characteristics of the inversion result may not be accurate, and the inversion needs to be combined with the geological law in a larger range. The introduction of more extensive data information will help the network model to invert more accurate formation parameters. SUMMARY

[0008] The purpose of the present application is to provide a large field of view seismic signal high resolution inversion method, which uses the global field of view characteristics of the Transformer to extract the geological law in the seismic data on a large scale, and uses more comprehensive formation feature information to predict the formation parameters, effectively improving the accuracy and resolution of the inversion parameters.

[0009] In order to solve the above-mentioned technical problems, the present application specifically adopts the following technical scheme: the large field of view seismic signal high resolution inversion method specifically includes the following steps:

[0010] S1 Construction of training data set: pre-process the geological data, and intercept the data to form two groups of training samples;

[0011] S2, constructing and training a model: designing and constructing a model framework of inversion network based on a Transformer as a basic module; then pre-training and re-training by using two groups of training samples respectively, so as to obtain an inversion network model;

[0012] S3, predicting a result and outputting: based on the trained inversion network model, using seismic data as input, outputting a predicted result corresponding to the seismic data, and processing the predicted result to obtain a high-resolution inversion result.

[0013] By adopting the technical solution, the global dependency relationship of data is efficiently modeled based on the self-attention mechanism in the Transformer module, the network model is pre-trained and re-trained by using two different sample data sets, the network model has a basic inversion capability from seismic data to formation parameters, the re-training enhances the inversion capability of the network model; the network model can realize extraction and fusion of global features, avoiding the limitation of single-channel data features; finally, based on the trained network model, the seismic data is used as input to obtain a predicted result containing global features; the method uses the global view characteristics of the Transformer to extract geological rules in seismic data on a large scale, and obtains more comprehensive formation feature information, so that more comprehensive formation feature information is used to predict formation parameters, effectively improving the accuracy and resolution of the inversion parameters. The method can make the neural network model consider the global dependency relationship of seismic data when predicting formation parameters, effectively avoiding the phenomenon that the inversion result is inaccurate due to the geological information deviation in single-channel seismic data, and effectively improving the ability of the model to obtain global features.

[0014] Preferably, the geological data of step S1 includes seismic data, logging data and horizon data, and the specific steps of step S1 are:

[0015] S11, pre-processing the seismic data, logging data and horizon data, normalizing the values of the seismic data and the logging data; and interpolating the horizon data according to the range of the seismic data to complete the missing trace set and horizon data;

[0016] S12, intercepting corresponding seismic data and logging data according to the position of the target horizon as training data of the neural network; that is, determining the target horizon (segment) of the intercepted seismic data and logging data according to the horizon data;

[0017] S13, obtaining a low-frequency initial model of formation parameters based on seismic data similarity interpolation by using logging data and corresponding well trace seismic data; and establishing a sample database by using well trace seismic data, logging data and the initial model;

[0018] S14: The extracted wellside seismic data and the low-frequency model data at the corresponding position are taken as a set of training samples, i.e., as pre-training sample data; the wellside seismic data and the well logging data at the corresponding position are taken as a set of training samples, i.e., as re-training sample data.

[0019] Preferably, the stratum parameters in the seismic data and the well logging data are normalized to [0, 1] in the step S11, and the formula is: ; ; wherein, S input represents the pre-stack seismic data input to the neural network model; S represents the input seismic data; S min represents the minimum value of the input seismic data; S max represents the maximum value of the input seismic data; M label represents the label data for training the neural network model; M min represents the minimum value of the well logging data; M max represents the maximum value of the well logging data, and M represents the input well logging data. The normalization processing of the input data and the label data is conducive to the convergence of the neural network model.

[0020] Preferably, the specific step of generating the initial model in the step S13 is:

[0021] S131: The similarity of the target trace seismic data to all wellside seismic data is calculated, and then a similarity coefficient ;

[0022] S132: According to the quality of the seismic data of the work area, the well logging data with a similarity greater than a certain threshold value can be selected, and the threshold value is preferably set to 0.7; a too large threshold value will lead to poor lateral continuity of the initial model interpolated in the end, and a too small threshold value will lead to poor geological rules of the initial model;

[0023] S133: The selected well logging data and the corresponding similarity coefficient are used for interpolation to obtain the stratum parameters of the target trace; the interpolation coefficient of the i-th well logging data is calculated in the following manner: ;

[0024] wherein, N is the number of well logging data selected by the threshold value, is the similarity of the target trace seismic data to the j-th wellside seismic data, wherein j = 1, 2, 3, …, N; is the interpolation coefficient of the i-th well logging data;

[0025] The formula of the interpolation result of the target trace is: ;

[0026] wherein, represents the i-th trace of the well logging data, wherein i = 1, 2, 3, …, N; represents the interpolation result of the target trace, and the formation parameters contained are the P-wave velocity , the S-wave velocity and the density .

[0027] Preferably, the specific steps of the step S2 are:

[0028] S21: constructing an inversion network model: a Transformer module is added to the neural network model, the seismic signal is modeled through a self-attention mechanism, a Transformer architecture is constructed, and a convolution module is combined to extract feature information of different dimensions, the high-dimensional features of the target trace are fused with the high-dimensional features of adjacent data, and the parameters are updated; a Transformer is used as a basic module to design and construct an inversion network model framework;

[0029] S22: training the model: first, pre-training is performed on the pre-trained sample data, the initial model data is used as a label sample during pre-training, and a pre-trained model is obtained; then, the pre-trained model is retrained using retrained sample data, so as to obtain an inversion network model containing the space-time correlation information in the seismic data.

[0030] Preferably, the specific steps of constructing the Transformer architecture in the step S21 are:

[0031] When the prestack seismic data is input, the signal value of each time in the seismic trace is regarded as an input feature, it is assumed that the length of the time sequence of the seismic signal is T and the length of the space sequence is S, and then the seismic data can be represented as a matrix;

[0032] In order to be able to input the seismic data into the Transformer model, the matrix is expanded into a one-dimensional vector sequence as the input of the Transformer model; for the input seismic data , wherein T is the time dimension and S is the space dimension;

[0033] After linear mapping, the sequence input into the Transformer is , wherein is the total number of data points, and D is the feature dimension of each data point;

[0034] Then, the seismic signal is modeled by a self-attention mechanism; the core idea of the self-attention mechanism is to dynamically adjust the representation of each element according to the correlation of each element of the input data with other elements; the calculation formula of the self-attention mechanism is:

[0035] ;

[0036] Wherein, Q is a query matrix (Query), K is a key matrix (Key), V is a value matrix (Value), is the dimension of the key; The attention score matrix is represented; Softmax is a normalization function, and the weight distribution is obtained; the final weighted summation value matrix (V) is obtained by attention output;

[0037] By calculating the similarity between the query matrix and the key matrix, and using the soft maximum function to obtain the weight coefficient, the value is weighted and summed to obtain the weighted representation of each data point;

[0038] The Mix-FNN module (mixed feature neural network module) is used to provide position information to the entire data; in this way, the Transformer can capture the spatio-temporal relationship between different seismic traces, so as to extract more comprehensive feature information;

[0039] When processing complex seismic data, a single attention head may not be able to fully capture the spatio-temporal dependence relationship of different scales and different levels; the multi-head self-attention mechanism in the Transformer introduced by the application can calculate multiple attention heads by mapping the query, key and value matrices to multiple subspaces respectively, and each attention head can capture the spatio-temporal dependence relationship of the input data from different angles;

[0040] Finally, the outputs of multiple attention heads are spliced together to form the final representation; the calculation formula of the multi-head self-attention mechanism is:

[0041]

[0042] Wherein, h is the number of attention heads, is the output linear transformation matrix, is the output of the i-th attention head, i=1, 2, 3……h; is the weight matrix used to map the query, key and value.

[0043] Through multi-head self-attention, the model can learn different hierarchical features of seismic data in multiple subspaces in parallel, thereby improving the expression ability of the model. In seismic data, especially in complex geological backgrounds, the nonlinear relationship and noise interference of seismic signals are very complex. The multi-head self-attention mechanism can model seismic signals from different dimensions, thereby improving the inversion accuracy and stability.

[0044] Preferably, the inversion network model framework designed in the step S21 with the Transformer as the basic module comprises a Transformer module, a convolution module, a regularization and an activation function module, and a residual connection is applied to the corresponding layer in the middle of the network to prevent overfitting of the network model; the network model takes prestack seismic data as input and outputs stratum parameters ; the specific steps of feature extraction and fusion combined with the convolutional neural network are as follows:

[0045] The network model embedded with the Transformer module can effectively extract high-level seismic signal features;

[0046] In each layer of the encoder, the network model transforms the input spatio-temporal data through the self-attention mechanism and the feedforward neural network, and gradually improves the understanding of the deep-level spatio-temporal features of the seismic data;

[0047] Under the design of the inversion network model framework with the Transformer as the basic module in a large field of view, the high-dimensional features of the target trace and the high-dimensional features of the adjacent data are fused after dimension reduction; after extracting the high-dimensional features, in order to improve the calculation efficiency and reduce the memory occupation, the principal component analysis (PCA) dimension reduction technology is used to compress the features; through the PCA technology, the high-dimensional features can be compressed to a lower dimension while retaining important information as much as possible, thereby improving the training speed and inference efficiency of the model on a large-scale data set; the principal component analysis dimension reduction technology is used to calculate the conversion matrix W for the multi-dimensional data x with a dimension of d, so as to convert it into data with a dimension of k .

[0048] Preferably, the principal component analysis dimension reduction technology is used to calculate the conversion matrix W for the multi-dimensional data x with a dimension of d, so as to convert it into data with a dimension of k , and the formula is as follows:

[0049] ;

[0050] ;

[0051] wherein x1, x2, x3, … x d represent data with dimensions of 1, 2, 3, … d, respectively, denotes the set of all real number vectors; denotes the set of all vectors composed of d real numbers; W denotes the weight matrix after linear transformation from d-dimensional space to k-dimensional space; denotes the dimension of the weight matrix W, d rows and k columns; The specific steps of the above formula are as follows: S211: Centering the input data, i.e. removing the mean value, the mean value of the data is ; ; S212: Calculate the covariance matrix, where and are the mean values of samples j and samples m, respectively, since the data has been centered, i.e. the mean value of each sample is 0, the formula is: ; where denotes the i-th value in the vector ; n is the number of elements in the vector; denotes the transpose of the vector ; S213: Use the Singular Value Decomposition (SVD) eigenvalue decomposition method to find the eigenvalues of the covariance matrix , and the eigenvectors v; S214: Determine the target dimension k for dimension reduction; arrange the eigenvectors in rows from top to bottom according to the size of the corresponding eigenvalues, and calculate the proportion of each eigenvalue in the total number of eigenvalues , where i=1, 2, 3, …, d; set the principal component eigenvalue proportion threshold t, and when the sum of the proportions of the current k eigenvalues exceeds t, determine k as the target dimension for dimension reduction, and take the first k rows of eigenvectors to form the dimension reduction conversion matrix ; ; ; S215: Dimension reduction conversion, use data z to represent the data after dimension reduction with dimension k; the formula is: ; On the basis of efficient feature extraction, by introducing an optimization algorithm based on gradient descent method, further optimization and adjustment of the parameters in the inversion process are carried out to ensure the efficient operation of the model in large-scale data; assuming that the objective function of inversion is L( ), where is the model parameter, the parameter is updated by gradient descent method, and the formula is: ; wherein, is a learning rate; is the model parameter to be updated at the current time t, is the updated model parameter; is the gradient of the objective function, through the above optimization, further improve the performance of the model on large-scale seismic data, extract high-dimensional features of seismic data in large field of view.

[0052] Preferably, the specific steps of the step S22 are: S221 pre-training: using pre-training sample data to pre-train the network model until the mean square error between the output result of the network model and the label data is lower than the set pre-training threshold or the set training times are reached, i.e. stop training; During pre-training, the initial model data is used as a label sample, the prediction output of the network is compared with the label, and the implicit relationship between pre-stack seismic data and formation parameters is established, and the mean square error is used as a loss function; the loss function MSE of pre-training is defined as: ; wherein, denotes the formation parameter, including the longitudinal wave velocity , the transverse wave velocity and the density three parameters; denotes the formation parameter output by the network; is the sample data; denotes the number of samples used at each iteration. When the pre-training process is completed, the network model basically has the ability to map seismic data to formation parameters; S222 retraining: inheriting the pre-trained network parameters in step S221, then using the sample set composed of well logging data and well seismic data, i.e. retraining sample data, to retrain the network model; until the mean square error between the output result of the network model and the corresponding well logging data label is lower than the set retraining threshold or the set training times are reached, i.e. stop training; When the two training of the neural network model are completed, the inversion network model is obtained, due to the influence of the Transformer module and the well logging data, the inversion network model contains the space-time correlation information in the seismic data, and accurately maps the pre-stack seismic data to the formation parameters.

[0053] Preferably, in the step S3, after the training is completed, the pre-stack seismic data is directly used as the input data, the output result of the network model is the normalized value of the formation parameter, and the real value of the formation parameter can be obtained based on the inverse normalization function, and the formula is:

[0054] ;

[0055] wherein, M output represents the stratum parameter value output by the network model; M max represents the maximum value of the input stratum parameter during preprocessing; represents the minimum value of the input stratum parameter during preprocessing; represents the stratum parameter value finally obtained.

[0056] Compared with the prior art, the technical scheme of the present application has the beneficial technical effects that:

[0057] (1) The neural network model is constructed based on the Transformer module, the Transformer architecture automatically learns the space-time relationship between different seismic traces by virtue of its self-attention mechanism, and can capture the global features in the seismic signal, thereby obtaining the global features of the seismic data;

[0058] (2) The network model is pre-trained and re-trained using two different sample data sets, so that the network model has the basic inversion capability from seismic data to stratum parameters, and the re-training enhances the inversion capability of the network model;

[0059] (3) The method can effectively improve the ability of the model to obtain global features, and improve the resolution and accuracy of the inversion result. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical scheme in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0061] Figure 1 is a schematic diagram of the inversion framework of the prior art;

[0062] Figure 2 is a flowchart of the high-resolution inversion method of the wide-view seismic signal of the present application;

[0063] Figure 3 is a schematic diagram of the well logging data interpolation algorithm based on the similarity of seismic data in the high-resolution inversion method of the wide-view seismic signal of the present application, wherein the interpolation coefficient is obtained by normalizing the similarity of the target trace seismic data and the well trace seismic data;

[0064] Figure 4Seismic data in the large field of view seismic signal high resolution inversion method of the application; wherein (a) is post-stack seismic data; (b) is near offset pre-stack seismic data; (c) is medium offset pre-stack seismic data; (d) is far offset pre-stack seismic data;

[0065] Figure 5 The Transformer module network structure diagram in the large field of view seismic signal high resolution inversion method of the application;

[0066] Figure 6 The large field of view neural network structure diagram in the large field of view seismic signal high resolution inversion method of the application;

[0067] Figure 7 The data position schematic diagram in the large field of view seismic signal high resolution inversion method of the application; wherein, (a) is the target data position; (b) is the adjacent data position;

[0068] Figure 8 The large field of view feature fusion module schematic diagram in the large field of view seismic signal high resolution inversion method of the application;

[0069] Figure 9 The large field of view seismic signal high resolution inversion method training mode in the large field of view seismic signal high resolution inversion method of the application;

[0070] Figure 10 The P-wave velocity inverted by the neural network model in the large field of view seismic signal high resolution inversion method of the application; wherein, (a) is the inversion result of the prior art scheme; (b) is the inversion result of the application;

[0071] Figure 11 The S-wave velocity inverted by the neural network model in the large field of view seismic signal high resolution inversion method of the application; wherein, (a) is the inversion result of the prior art scheme; (b) is the inversion result of the application;

[0072] Figure 12 The density inverted by the neural network model in the large field of view seismic signal high resolution inversion method of the application; wherein, (a) is the inversion result of the prior art scheme; (b) is the inversion result of the application;

[0073] Figure 13 The P-S wave velocity ratio inverted by the neural network model in the large field of view seismic signal high resolution inversion method of the application; wherein, (a) is the inversion result of the prior art scheme; (b) is the inversion result of the application;

[0074] Figure 14The inversion result single channel display of the prior art scheme in the large field of view seismic signal high resolution inversion method of the present application; wherein (a) P-wave velocity; (b) S-wave velocity; (c) density; (d) P-S wave velocity ratio;

[0075] Figure 15 The inversion result single channel display in the large field of view seismic signal high resolution inversion method of the present application in the large field of view seismic signal high resolution inversion method of the present application; wherein, (a) P-wave velocity; (b) S-wave velocity; (c) density; (d) P-S wave velocity ratio. DETAILED DESCRIPTION

[0076] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0077] Embodiment: As shown in the figure, the large field of view seismic signal high resolution inversion method specifically comprises the following steps: Figure 2

[0078] S1: Construction of training data set: pre-process the geological data, and intercept the data to form two groups of training samples;

[0079] The geological data of the step S1 comprises seismic data, logging data and horizon data, and the specific steps of the step S1 are as follows:

[0080] S11: Pre-process the seismic data, logging data and horizon data, normalize the values of the seismic data and the logging data, and interpolate the horizon data according to the range of the seismic data to complete the missing trace set and horizon data;

[0081] In the step S11, the stratum parameters in the seismic data and the logging data are normalized to [0, 1], and the formula is as follows:

[0082] ;

[0083] ;

[0084] Wherein, S input represents the pre-stack seismic data input by the neural network model; S represents the input seismic data; S min represents the minimum value of the input seismic data; S max represents the maximum value of the input seismic data; M label ​The label data used for training the neural network model; M min The minimum value of the logging data; M max The maximum value of the logging data represented by M, and M represents the input logging data; Normalizing the input data and the label data facilitates the convergence of the neural network model.

[0085] As Figure 4 shown, Figure 4 (a) in (a) is post-stack seismic data, Figure 4 (b) in (b) is near-offset pre-stack seismic data, Figure 4 (c) in (c) is medium-offset pre-stack seismic data, Figure 4 (d) in (d) is far-offset pre-stack seismic data; the pre-stack seismic data volume covers an area of 100km 2 , including 501 main lines and 631 contact lines, and each pre-stack gather has 9 different offset seismic data; when making training data, all pre-stack data are stacked according to the near-offset, medium-offset and far-offset parts, and then the post-stack seismic data are used as the input of the neural network together;

[0086] S12: According to the position of the target layer, the corresponding seismic data and logging data are intercepted as the training data of the neural network; that is, for training and obtaining the low-frequency initial model in step S13;

[0087] S13: Obtain the low-frequency initial model of the formation parameter based on the similarity interpolation of the seismic data by using the logging data and the corresponding well trace seismic data; at the same time, establish a sample database by using the well trace seismic data, the logging data and the initial model;

[0088] As Figure 3 shown is a schematic diagram of generating an initial model, and the specific steps of generating an initial model in step S13 are as follows:

[0089] S131: Calculate the similarity of the target trace seismic data and all well trace seismic data, and then obtain the similarity coefficient ;

[0090] S132: According to the quality of the seismic data in the work area, the logging data with a similarity greater than a certain threshold value can be selected, and the threshold value set in this embodiment is 0.7; if the threshold value is too large, the initial model obtained by interpolation will have poor lateral continuity, and if the threshold value is too small, the geological rule of the initial model will be poor;

[0091] S133: Use the selected logging data and the corresponding similarity coefficient to perform interpolation to obtain the formation parameter of the target trace; the interpolation coefficient of the i-th well trace logging data is calculated as follows:

[0092] ;

[0093] Wherein, N is the number of well logging data screened out by threshold, is the similarity of the target trace seismic data and the jth trace well offset seismic data, wherein j = 1, 2, 3 … N; is the interpolation coefficient of the ith trace well logging data;

[0094] The formula of the interpolation result of the target trace is:

[0095]

[0096] Wherein, represents the ith trace well logging data, wherein i = 1, 2, 3 … N; represents the interpolation result of the target trace, and the formation parameters contained are P-wave velocity , S-wave velocity and density ;

[0097] S14: Take the extracted well offset seismic data and the low frequency model data at the corresponding position as a set of training samples, that is, take them as pre-training sample data; Take the well offset seismic data and the well logging data at the corresponding position as a set of training samples, that is, take them as re-training sample data;

[0098] S2: Construct and train the model: design and construct the inversion network model framework with Transform as the basic module; Then pre-train and re-train respectively by using the two sets of training samples, so as to obtain the inversion network model;

[0099] The specific steps of the step S2 are:

[0100] S21: Construct the inversion network model: add the Transform module in the neural network model, model the seismic signal through the self-attention mechanism, construct the Transform architecture, combine the convolution module, extract the feature information of different dimensions, reduce and fuse the high-dimensional features of the target trace and the adjacent data, and then update the parameters; Design and construct the inversion network model framework with Transform as the basic module;

[0101] The specific steps of constructing the Transform architecture in the step S21 are:

[0102] The seismic data is essentially a multi-dimensional signal with space-time characteristics, and its information not only includes the change on the time sequence, but also includes the dependence on the spatial position. The traditional convolutional neural network usually limits to local receptive field when processing such data, and cannot fully capture the global space-time dependence relationship. As shown in Figure 5 , the Transform architecture can efficiently model the long-range dependence relationship of the data through the self-attention mechanism;​

[0103] In inputting pre-stack seismic data, the signal value of each time in the seismic trace is regarded as an input feature, assuming that the length of the time sequence of the seismic signal is T and the length of the space sequence is S, the seismic data can be represented as a matrix;

[0104] In order to be able to input seismic data into the Transformer model, the matrix is expanded into a one-dimensional vector sequence as the input of the Transformer model; for the input seismic data , wherein T is the time dimension, and S is the space dimension;

[0105] After linear mapping, the sequence input into the Transformer is , wherein is the total number of data points, and D is the feature dimension of each data point;

[0106] Then, the seismic signal is modeled by the self-attention mechanism; the core idea of the self-attention mechanism is to dynamically adjust the representation of each element according to the correlation between each element and other elements of the input data; the calculation formula of the self-attention mechanism is:

[0107] ;

[0108] , wherein Q is the query matrix (Query), K is the key matrix (Key), V is the value matrix (Value), is the dimension of the key; is the attention score matrix; Softmax is a normalization function to obtain the weight distribution; the final weighted sum value matrix (V) is obtained by weighting and summing the output of the attention;

[0109] By calculating the similarity between the query matrix and the key matrix, and using the soft-max function to obtain the weight coefficient, the value is weighted and summed to obtain the weighted representation of each data point;

[0110] Figure 5 The Mix-FNN module in the above formula is used to provide position information to the entire data; in this way, the Transformer can capture the spatio-temporal relationship between different seismic traces, so as to extract more comprehensive feature information;

[0111] When processing complex seismic data, a single attention head may not be able to comprehensively capture the spatio-temporal dependence relationship of different scales and different levels; the multi-head self-attention mechanism in the Transformer introduced by the present application can calculate multiple attention heads by mapping the query, key and value matrices to multiple subspaces respectively, and each attention head can capture the spatio-temporal dependence relationship of the input data from different angles;

[0112] The outputs of the plurality of attention heads are finally spliced together to form a final representation; the calculation formula of the multi-head self-attention mechanism is as follows:

[0113]

[0114] wherein h is the number of attention heads, is an output linear transformation matrix, is the output of the i-th attention head, i = 1, 2, 3, …, h; is a weight matrix for mapping queries, keys, and values;

[0115] Through the multi-head self-attention mechanism, the model can learn different levels of features of the seismic data in multiple subspaces in parallel, thereby improving the expression ability of the model. In seismic data, especially in a complex geological background, the nonlinear relationship of the seismic signal and the noise interference are very complex; the multi-head self-attention mechanism can model the seismic signal from different dimensions, thereby improving the inversion accuracy and stability.

[0116] In the design of the inversion framework, in some actual applications, appropriate network parameters need to be selected according to the characteristics of the data. In this embodiment, the method proposed in the present application is used to perform seismic inversion on the actual work area data set in the northwest region of China to verify the effect. Figure 6 A neural network structure diagram with a wide field of view is shown; the input data is prestack seismic data, and the output data is a formation parameter; the network structure includes a Transformer module, a convolution module, a regularization and activation function module, and a residual connection is applied to the corresponding layers in the middle of the network to prevent overfitting of the network model; the network model uses the Transformer module and the convolution module as the basic structure, and the middle includes a residual connection and a feature fusion module, and finally uses a fully connected layer to output the predicted parameters;

[0117] The inversion network model framework designed in the step S21 with the Transformer as the basic module includes a Transformer module, a convolution module, a regularization and activation function module, and a residual connection is applied to the corresponding layers in the middle of the network to prevent overfitting of the network model; the network model takes prestack seismic data as input and outputs formation parameters ;

[0118] Table 1 provides detailed parameter information of the network model, including input channels (Ic), output channels (Oc), convolution kernel size (Ks), data block size (Bs), and input data height (Ih). Generally, Ih corresponds to the length of each channel of the prestack seismic data, and Ic refers to the number of angles in the prestack seismic angle gather. Since the inversion result includes , the final Oc is set to 3.

[0119] Table 1 Specific parameter information of network model

[0120] Operation layer Channel (Ic / Oc) Convolution kernel (Ic x Oc x Ks) Output (Ih x Oc x Bs) Input data … … Ih x 4 x 8 Transformer module 1 4Ih / 4Ih … Ih x 4 x 8 Convolution module 1 4 / 5 4×5×3 Ih x 5 x 8 Transformer module 2 5Ih / 5Ih … Ih x 5 x 8 Convolution module 2 5 / 6 5×6×3 Ih x 6 x 8 Feature fusion 46Ih / 7Ih … Ih x 7 x 8 Transformer module 3 7Ih / 7Ih … Ih x 7 x 8 Convolution module 3 7 / 6 7×6×3 Ih x 6 x 8 Transformer module 4 6Ih / 6Ih … Ih x 6 x 8 Convolution module 4 6 / 5 6×5×3 Ih x 5 x 8 Transformer module 5 5Ih / 5Ih … Ih x 5 x 8 Convolution module 5 5 / 4 5×4×3 Ih x 4 x 8 Fully connected layer 4Ih / 3Ih … 3Ih x 8 Output data … … Ih x 3 x 8

[0121] The design of the inversion network model framework taking the Transformer as a basic module expands the view of the neural network model on the seismic data, including a feature extraction and fusion module, in some embodiments, the specific steps of the feature extraction and fusion combined with the convolutional neural network for feature extraction and fusion are as follows:

[0122] The network model embedded with the Transformer module can effectively extract high-level seismic signal features;

[0123] In each layer of the encoder, the network model transforms the input spatio-temporal data through the self-attention mechanism and the feedforward neural network, gradually improving the understanding of the deep spatio-temporal features of the seismic data;

[0124] The neural network framework of the present application not only takes the target trace seismic data as input, but also extracts the geological features of the adjacent trace seismic data, so as to further expand the view of the network model. The positions of the target trace data and the adjacent trace data in the seismic data are as shown in Figure 7 , wherein Figure 7 (a) in the figure is the position of the target data; Figure 7 (b) in the figure is the position of the adjacent data, wherein the adjacent data contains 8 traces of data around the target trace; the feature fusion module is as shown in Figure 8 ;

[0125] Under the large-view inversion network model framework designed taking the Transformer as a basic module, the high-dimensional features of the target trace and the high-dimensional features of the adjacent data are fused in dimension reduction; after extracting the high-dimensional features, in order to improve the calculation efficiency and reduce the memory occupation, the principal component analysis (PCA) dimension reduction technology is used to compress the features; through the PCA technology, the high-dimensional features can be compressed to a lower dimension while retaining important information as much as possible, so as to improve the training speed and inference efficiency of the model on a large-scale data set; the principal component analysis dimension reduction technology is used to calculate the conversion matrix W for the multi-dimensional data x with a dimension of d, so as to convert it into data with a dimension of k;

[0126] In the step S21, the principal component analysis dimension reduction technology is used to calculate the conversion matrix W for the multi-dimensional data x with a dimension of d, so as to convert it into data with a dimension of k, and the formula is:

[0127] ;

[0128] ;

[0129] wherein x1, x2, x3, …x d respectively represent data of dimensions 1, 2, 3, …d, represent a set of all real number vectors; represent a set of all vectors composed of d real numbers; W represents a weight matrix after linear transformation from a d-dimensional space to a k-dimensional space; represent the dimension of the weight matrix W, d rows and k columns;

[0130] The specific steps to obtain the above formula are as follows:

[0131] S211: Centering the input data, i.e. removing the average value, whose average value is:

[0132] ;

[0133] S212: Calculate the covariance matrix, wherein and are the average values of sample j and sample m, since the data has been centered, i.e. the average value of each sample is 0, the formula is:

[0134] ;

[0135] wherein, represents the i-th value in the vector ; n is the number of elements in the vector; represents the transpose of the vector ;

[0136] S213: Use the singular value decomposition (SVD) eigenvalue decomposition method to solve the eigenvalue and the eigenvector v in the covariance matrix;

[0137] S214: Determine the dimension k of the dimension reduction target; arrange the eigenvectors in rows from top to bottom according to the corresponding eigenvalue size, and calculate the proportion of each eigenvalue in the total number of eigenvalues , wherein i=1, 2, 3……d; set the principal component eigenvalue proportion threshold t, when the sum of the proportions of the current k eigenvalues exceeds t, it is determined that k is the dimension reduction target dimension, and the first k rows of eigenvectors are taken to form the dimension reduction conversion matrix ;

[0138] ;

[0139] ;

[0140] S215: Dimension reduction conversion, data z represents the data with dimension k after dimension reduction; the formula is:

[0141] ;

[0142] On the basis of efficient feature extraction, an optimization algorithm based on gradient descent method is introduced to further optimize and adjust the parameters in the inversion process, ensuring efficient operation of the model in large-scale data; assuming that the objective function of inversion is L( ), wherein is the model parameter, and the parameter is updated by the gradient descent method, and the formula is:

[0143] ;

[0144] wherein, is the learning rate; is the model parameter to be updated at the current time t, is the updated model parameter; is the gradient of the objective function, and through the above optimization, the performance of the model on large-scale seismic data is further improved, and the high-dimensional features of seismic data in a large field of view are extracted;

[0145] S22: Training model: first, pre-training sample data is used for pre-training, and the initial model data is used as a label sample during pre-training to obtain a pre-training model; then, the pre-training model is retrained using retraining sample data, thereby obtaining an inversion network model containing the space-time correlation information in the seismic data;

[0146] As shown in Figure 9 , the specific steps of step S22 are:

[0147] S221: Pre-training: the network model is pre-trained using pre-training sample data until the mean square error between the output of the network model and the label data is lower than the set pre-training threshold or the set training times are reached, i.e. stop training; the pre-training threshold is 0.1%, and the training rounds are 20;

[0148] During pre-training, the initial model data is used as a label sample, the prediction output of the network is compared with the label, the implicit relationship between pre-stack seismic data and formation parameters is established, and the mean square error is used as a loss function; the loss function MSE of pre-training is defined as:

[0149] ;

[0150] wherein, denotes the formation parameter, including the longitudinal wave velocity , the transverse wave velocity , and the density three parameters; representing the formation parameters of the network output; is sample data; is the mean square error between the formation parameters of the network output and the sample; representing the number of samples used at each iteration; after the pre-training process is completed, the network model basically has the ability to map seismic data to formation parameters;

[0151] S222 retraining: inherit the network parameters pre-trained in step S221, and then retrain the network model using the sample set composed of well logging data and well seismic data, i.e. retraining sample data; until the mean square error between the output result of the network model and the corresponding well logging data label is lower than the set retraining threshold or reaches the set training times, i.e. stop training; the training threshold value is 0.1%, and the training round is 10;

[0152] When the two training of the neural network model is completed, i.e. the inversion network model is obtained, due to the influence of the Transformer module and the well logging data, the inversion network model contains the space-time correlation information in the seismic data, which accurately maps the pre-stack seismic data to the formation parameters;

[0153] S3 prediction result and output: based on the trained inversion network model, using seismic data as input, output the prediction result corresponding to the seismic data, and process the prediction result to obtain a high-resolution inversion result;

[0154] In the step S3, after the training is completed, the pre-stack seismic data is directly used as the input data, the output result of the network model is the normalized value of the formation parameters, and the real value of the formation parameters can be obtained based on the inverse normalization function, and the formula is:

[0155] ;

[0156] wherein, M output represents the formation parameter value output by the network model; M max represents the maximum value of the input formation parameters during pre-processing; represents the minimum value of the input formation parameters during pre-processing; represents the final obtained formation parameter value.

[0157] In a specific application, there are 64 available well logging data in the working area, of which 62 are used as known data for the entire inversion process, and 2 are reserved data for verifying the accuracy of the inversion effect. Figures 10-12 The pre-stack inversion effect of the formation parameters is shown in the following figure: Figures 10-12 (a) in the figure is the inversion result of the prior art scheme, Figures 10-12(b) shows the inversion results of this invention; the differences in inversion effects between the prior art and the method of this invention are highlighted with black boxes. The comparison shows that the stratigraphic parameter resolution obtained by the method of this invention is significantly higher than that of the prior art. Preliminary geological analysis revealed a strong correlation between the hydrocarbon characteristics of the target strata and the P-wave / S-wave velocity ratio of the stratigraphic parameters. In the inversion results of the P-wave / S-wave velocity ratio, the prior art can only invert one relatively thick layer, while the method of this invention can obtain two thin layers, such as... Figure 13 As shown, Figure 13 (a) in the figure represents the inversion result of the existing technical solution. Figure 13 (b) in the figure represents the inversion result of the present invention. The method proposed in this invention exhibits better performance in terms of resolution of the inversion results. To quantitatively illustrate the accuracy of the method of the present invention, this embodiment selects data from one retained well to verify the consistency between the inversion results and the actual well logging data. Figure 14 This is a single-channel demonstration of the inversion results from existing technical solutions. Figure 14 In the figure, (a) represents the longitudinal wave velocity. Figure 14 In the equation (b), the transverse wave velocity is... Figure 14 (c) in the figure represents density. Figure 14 In this context, (d) represents the ratio of P-wave to S-wave velocity; Figure 15 This is a single-channel demonstration of the inversion results of the method of the present invention. Figure 15 In the figure, (a) represents the longitudinal wave velocity; Figure 15 In the equation (b), the transverse wave velocity is... Figure 15 (c) in the figure represents density. Figure 15 In the figure, (d) represents the P-wave to S-wave velocity ratio; the red curve represents the actual well logging data, and the black curve represents the inversion result. The figure also shows that the formation parameters inverted by the method of this invention have high resolution. Furthermore, this embodiment also calculates the correlation between the inversion result and the actual well logging data curves. The correlations between the four formation parameter curves inverted by the prior art and the actual well logging data are 0.807, 0.857, 0.822, and 0.818, respectively. The correlations between the four formation parameter curves inverted by the method of this invention and the actual well logging data are 0.905, 0.908, 0.873, and 0.885, respectively. The single-channel display and quantitative similarity data demonstrate that the method of this invention significantly improves the resolution and accuracy of the inversion results compared to the prior art.

[0158] Special note, Figure 4 , Figures 10-13 In the inline seismic data, the line number is represented by Inline, the trace number by Xline, Time (ms) represents the time depth of the strata, and Amplitude represents the normalized amplitude of the seismic data.

[0159] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative, but not restrictive, and a person of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, and these all belong to the protection of the present application.

Claims

1. A method of seismic signal high resolution inversion with large field of view, characterized in that, Specifically comprising the following steps: S1 construction of a training data set: the geological data is preprocessed, and the data is intercepted to form two groups of training samples; S2 model construction and training: a transformation is used as a basic module to design and construct an inversion network model framework; then the two groups of training samples are used for pre-training and re-training respectively, so as to obtain an inversion network model; S3 prediction result and output: based on the trained inversion network model, the seismic data is used as input, and the prediction result corresponding to the seismic data is output; then the prediction result is processed to obtain a high-resolution inversion result.

2. The method of wide-angle seismic signal high-resolution inversion according to claim 1, characterized in that, The geological data of step S1 includes seismic data, logging data and horizon data, and the specific steps of step S1 are: S11: the seismic data, logging data and horizon data are preprocessed, that is, the values of the seismic data and logging data are normalized; the horizon data is interpolated according to the range of the seismic data to complete the missing trace set and horizon data; S12: the corresponding seismic data and logging data are intercepted according to the position of the target horizon as the training data of the neural network; S13: the low-frequency initial model of the stratum parameter is obtained based on the seismic data similarity interpolation by using the logging data and the corresponding well trace seismic data; at the same time, the sample database is established by using the well trace seismic data, logging data and initial model; S14: the extracted seismic data and the low-frequency model data at the corresponding position are taken as a group of training samples, that is, they are taken as pre-training sample data; the well trace seismic data and the logging data at the corresponding position are taken as a group of training samples, that is, they are taken as re-training sample data.

3. The method of wide-angle seismic signal high-resolution inversion according to claim 2, characterized in that, In step S11, the stratum parameters in the seismic data and logging data are normalized to [0, 1], and the formula is: ； ； Wherein, S input represents the pre-stack seismic data input of the neural network model; S represents the input seismic data; S min represents the minimum value of the input seismic data; S max represents the maximum value of the input seismic data; M label represents the label data used for training the neural network model; M min represents the minimum value of the logging data; M max represents the maximum value of the logging data; M represents the input logging data.

4. The method of wide-angle seismic signal high-resolution inversion according to claim 2, characterized in that, The specific steps of generating the initial model in step S13 are: S131: Calculate the similarity between the target trace and all the well-adjacent trace seismic data, and then obtain the similarity coefficient ; S132: according to the quality of the seismic data of the work area, the logging data with a similarity greater than a certain threshold value can be selected; S133: the stratum parameters of the target trace are obtained by interpolation using the selected logging data and the corresponding similarity coefficient; The interpolation coefficient calculation method of the i-th well logging data is: ； Wherein, N is the number of well logging data screened by threshold value, is the similarity of target trace seismic data and jth trace well side trace seismic data, wherein j = 1, 2, 3…N; is the interpolation coefficient of ith trace well logging data; The formula of the interpolation result of the target trace is: ； wherein, represents the i-th trace of the well logging data, wherein i = 1, 2, 3, …, N; represents the interpolation result of the target trace, containing the formation parameters of P-wave velocity , S-wave velocity , and density .

5. The method of wide-angle seismic signal high-resolution inversion according to claim 2, characterized in that, The specific steps of step S2 are: S21 construction of an inversion network model: a transformation module is added to the neural network model, the seismic signal is modeled through a self-attention mechanism, a transformation architecture is constructed, and a convolution module is combined to extract different dimensional feature information, the high-dimensional features of the target trace and the high-dimensional features of the adjacent data are fused after dimension reduction, and the parameters are updated; a transformation is used as a basic module to design and construct an inversion network model framework; S22 model training: the pre-training sample data is used for pre-training, and the initial model data is used as a label sample during pre-training to obtain a pre-training model; the re-training sample data is used for re-training of the pre-training model, so as to obtain an inversion network model containing the space-time correlation information in the seismic data.

6. The method of wide-angle seismic signal high-resolution inversion according to claim 5, characterized in that, The specific steps of constructing the transformation architecture in step S21 are: When inputting pre-stack seismic data, the signal value of each time in the seismic trace is regarded as an input feature. Assuming that the length of the time series of the seismic signal is T and the length of the spatial series is S, the seismic data can be represented as a matrix of ​ The matrix is unfolded into a one-dimensional vector sequence as the input of the Transformer model; for the input seismic data where T is the time dimension and S is the spatial dimension; The sequence input to the Transformer after linear mapping wherein is the total number of data points, and D is the feature dimension of each data point. Next, the seismic signal is modeled by a self-attention mechanism; the core idea of the self-attention mechanism is to dynamically adjust the representation of each element according to the correlation between each element and other elements of the input data; the calculation formula of the self-attention mechanism is: ； where Q is the query matrix, K is the key matrix, and V is the value matrix, is the dimension of the key; denotes the attention score matrix; Softmax is a normalization function that gets a weight distribution; the final weighted sum value matrix V gets the attention output; The similarity between the query matrix and the key matrix is calculated, and the weight coefficient is obtained by using the soft-max function, so as to weight the sum of the values and obtain the weighted representation of each data point; Finally, the outputs of multiple attention heads are spliced together to form the final representation; the calculation formula of the multi-head self-attention mechanism is: where h is the number of attention heads, is an output linear transformation matrix, is the output of the i-th attention head, i = 1, 2, 3,... h; is a weight matrix for mapping queries, keys, values.

7. The method of wide-angle seismic signal high-resolution inversion according to claim 6, characterized in that, The inversion network model framework designed in the step S21 with the Transformer as a basic module comprises a Transformer module, a convolution module, a regularization and an activation function module, and a residual connection is applied to a corresponding layer in the middle of the network to prevent overfitting of the network model; the network model takes pre-stack seismic data as input and outputs stratum parameters The specific steps of feature extraction and fusion in combination with the convolutional neural network are as follows: In each layer of the encoder, the network model transforms the input spatio-temporal data through the self-attention mechanism and the feedforward neural network, and gradually improves the understanding of the deep spatio-temporal features of the seismic data; Under the framework of the inversion network model designed with the Transformer as the basic module in a large field of view, the high-dimensional features of the target road and the high-dimensional features of the adjacent data are fused in dimension reduction; after the high-dimensional features are extracted, the principal component analysis dimension reduction technology is used to compress the features; the principal component analysis dimension reduction technology converts the multi-dimensional data x with a dimension of d into data with a dimension of k by calculating the conversion matrix W .

8. The method of wide-angle seismic signal high-resolution inversion according to claim 7, characterized in that, In the step S21, the principal component analysis dimension reduction technique is used to calculate the conversion matrix W for converting the multi-dimensional data x with the dimension d into data with the dimension k The formula is: ； ； where x1, x2, x3,... x d represents a data set of dimensions 1, 2, 3,... d, respectively, represents a set of all real number vectors; represents a set of all vectors consisting of d real numbers; W represents a weight matrix after linear transformation from a d-dimensional space to a k-dimensional space; represents the dimension of the weight matrix W as d x k, d rows, k columns; The specific steps of the above formula are: S211: Centering the input data, i.e. de-meaning it with the mean of the data ； S212: Compute the covariance matrix, where and are the average values of sample j and sample m, respectively, since the data has been centered, i.e., the average value of each sample is 0, and the formula is: ； wherein represents the i-th value in the vector ; n is the number of elements in the vector; represents the transpose of the vector ; S213: Eigenvalue decomposition of the covariance matrix with singular value decomposition with the eigenvector v; S214: Determine the dimensionality reduction target dimension k; arrange the feature vectors in rows from top to bottom according to the corresponding eigenvalue size, and calculate the proportion of each eigenvalue in the total number of eigenvalues , where i = 1, 2, 3…d; set the principal component eigenvalue proportion threshold value as t, and when the sum of the proportions of the current k eigenvalues exceeds t, determine k as the dimensionality reduction target dimension, and take the first k rows of feature vectors to form the dimensionality reduction conversion matrix ; ； ； S215: Dimensionality reduction conversion, using data z to represent the data with dimension k after dimensionality reduction; the formula is: ； On the basis of efficient feature extraction, an optimization algorithm based on gradient descent is introduced to further optimize and adjust the parameters in the inversion process; Assume the objective function of inversion is L( ), where is the model parameter, and the parameter is updated by gradient descent method, and the formula is: ； wherein, is a learning rate; is a model parameter to be updated at a current time t, is an updated model parameter; is a gradient of the objective function.

9. The method of wide-angle seismic signal high-resolution inversion according to claim 7, characterized in that, The specific steps of the step S22 are: S221 pre-training: the network model is pre-trained using pre-training sample data until the mean square error between the output of the network model and the label data is lower than the set pre-training threshold or the training reaches the set number of times, that is, the training is stopped; During pre-training, the initial model data is used as a label sample, the prediction output of the network is compared with the label, the implicit relationship between the pre-stack seismic data and the formation parameters is established, and the mean square error is used as a loss function; the loss function MSE of pre-training is defined as: ； wherein, represents formation parameters, including P-wave velocity S-wave velocity and density three parameters; represents the formation parameters output by the network; is the sample data; represents the number of samples used at each iteration; S222 retraining: the network parameters pre-trained in step S221 are inherited, and the network model is retrained using the sample set composed of well logging data and well seismic data, that is, retraining sample data; until the mean square error between the output of the network model and the corresponding well logging data label is lower than the set retraining threshold or the training reaches the set number of times, that is, the training is stopped; After the two training of the neural network model are completed, the inversion network model is obtained, which contains the spatio-temporal correlation information in the seismic data, and accurately maps the pre-stack seismic data to the formation parameters.

10. The method of wide-angle seismic signal high-resolution inversion according to claim 9, characterized in that, In the step S3, after the training is completed, the pre-stack seismic data is directly used as the input data, the output result of the network model is the normalized value of the formation parameter, and the real value of the formation parameter can be obtained based on the inverse normalization function, and the formula is: ； wherein M output represents a formation parameter value output by the network model; M max represents a maximum value of the input formation parameter at the preprocessing time; represents a minimum value of the input formation parameter at the preprocessing time; represents a final obtained formation parameter value.

Images