A Method for Predicting the Multi-Task Learning Speed Model

Abstract

The present invention relates to a method for predicting a multi-task learning speed model, belonging to a method for predicting a speed model based on borehole seismic exploration DAS-VSP data. A velocity model set of a quasi-real geological structure and a vertical seismic profile data set are constructed, a multi-task learning network is constructed and trained, and the velocity model prediction is realized by using DAS-VSP data. The advantages of the present invention are that it effectively solves the problem of lack of samples in the network training process, constructs a multi-task learning network with an encoder-dual decoder architecture, uses sub-tasks to perform structural constraints on the main task, enhances the accuracy of the prediction results, and introduces an attention mechanism to improve the key feature recognition ability of DAS-VSP data. The present invention effectively overcomes the dependence on prior information of traditional methods, significantly improves the calculation efficiency and prediction accuracy, provides more accurate data support for subsequent imaging and interpretation work, and enhances the influence of DAS-VSP technology in the field of oil and gas resource exploration.

Description

Technical Field

[0001] The present invention belongs to a method for predicting a velocity model of DAS-VSP data in borehole seismic exploration, and in particular, to a prediction method for accurately predicting a formation structure and velocity values under multi-task learning. Background Art

[0002] Seismic exploration infers the lithology and structure of a formation by analyzing the responses of artificial seismic waves, including wave velocity, travel time, reflection, and refraction angles. As an important means of modern seismic exploration, Vertical Seismic Profile (VSP) has the advantages of high resolution and accurate depth detection. In recent years, Distributed Acoustic Sensing (DAS) has gradually been applied to the field of VSP seismic exploration due to its advantages such as low cost, high sensitivity, high precision, large array, and repeatable acquisition.

[0003] An accurate velocity model is a key prerequisite for high-resolution seismic imaging techniques such as reverse time migration and prestack depth migration. Currently, seismic exploration is gradually developing towards more complex subsurface structures and oil and gas reservoirs. Traditional methods such as Full Waveform Inversion (FWI) have high accuracy and strong interpretability, but the step-by-step iterative reconstruction of the velocity model requires setting a large number of parameters, which is time-consuming and costly in calculation, and the setting of the initial velocity model will also affect the results. Moreover, it is extremely difficult to predict the velocity model of DAS_VSP data using FWI. Therefore, exploring an accurate and effective method for predicting the velocity model applicable to DAS-VSP data is an urgent problem to be solved in the current field of seismic exploration.

[0004] In recent years, with the rapid development of computing power, many Deep Learning (DL) models trained using large-scale datasets have emerged to directly predict the velocity model from prestack surface seismic data. However, currently, no researcher has used deep learning methods to achieve accurate prediction from prestack DAS-VSP data to the velocity model. There are mainly the following problems: the velocity model set required for deep learning network training is scarce; realizing a complex mapping from DAS-VSP data to the velocity model is often interfered by various irrelevant information; without appropriate formation constraints, the network cannot accurately reconstruct a fine subsurface structure. Summary of the Invention

[0005] The present invention provides a method for predicting a multitask learning speed model to solve the problems faced by current traditional speed model prediction methods, namely, the complex engineering background and the difficulty in predicting the speed model from DAS-VSP data. At the same time, aiming at the lack of speed model sets required for training deep learning methods, and the fact that realizing complex mappings from DAS-VSP data to speed models is often interfered by various irrelevant information and lacks appropriate formation constraints, and the network cannot accurately reconstruct the fine subsurface structure.

[0006] The technical solution adopted by the present invention includes the following steps:

[0007] Step 1: Construct a pseudo-real geological structure speed model set and a DAS-VSP data set;

[0008] (1) Construct a pseudo-real geological structure speed model set V;

[0009] The process of constructing the speed model includes drawing formation curves and filling speed values. When drawing the formation structure curve, a linear function and a trigonometric function are used to simulate the continuous and fluctuating formation structure;

[0010] The curve equation of the m-th layer can be expressed as follows:

[0011]

[0012] where a j , T j , θ j , j = 1, 2, 3 are the amplitude, period, and phase parameters of the trigonometric function, used to simulate formation fluctuations in different situations, r m is the formation inclination degree parameter, c m is the layer depth parameter:

[0013] Construct a horizontal speed model, an inclined speed model, a fluctuating speed model, and a fault speed model;

[0014] According to the above speed model construction method, p randomly independent speed models v i , i ∈ [1, p] are generated in batches, and the speed model set V = {v i | i = 1, 2,... p};

[0015] (2) Construct a DAS-VSP data set D, and obtain DAS-VSP data by forward modeling of the speed model;

[0016] Step 2: Construct and train a multitask learning network;

[0017] (1) Construct a multitask learning network

[0018] The multi-task learning network adopts an encoder-dual decoder architecture. The main decoder predicts the main task of the velocity model, and the sub-decoder reconstructs the sub-task of the velocity model structure.

[0019] (2) Train the multi-task learning network;

[0020] Establish a training set. After training, the multi-task learning network can predict the velocity model v given the DAS-VSP data d i ; i ;

[0021] Step 3: Perform velocity model prediction on DAS-VSP data;

[0022] The output result of the main task decoder is the result of velocity model prediction for DAS-VSP data

[0023] In the construction of the virtual true geological structure velocity model set V in step 1 of the present invention, the horizontal velocity model is constructed according to the following steps:

[0024] 1) According to formula (1), set a j = 0, j = 1, 2, 3, r 1 = 0, c 1 > 0 to generate a straight line as the top interface of the velocity model;

[0025] 2) According to formula (1), set a j = 0, j = 1, 2, 3, r m = 0, c m > c m-1 Iteratively generate the lower interface until all formation interfaces are generated;

[0026] 3) Follow the principle that the propagation velocity value is larger in deeper media, set the top layer velocity value as q 1 , ensure that under the condition that the velocity value q of the mth layer m - q m-1 > q min , iteratively add velocity values to each layer.

[0027] In the construction of the virtual true geological structure velocity model set V in step 1 of the present invention, the inclined velocity model is constructed according to the following steps:

[0028] 1) According to formula (1), set a j = 0, j = 1, 2, 3, r 1 ≠ 0, c 1 > 0 to generate a slant line as the top interface of the velocity model;

[0029] 2) According to formula (1), set a j= 0, j = 1, 2, 3, r m ≠ 0, (y m ) min >(y m-1 ) max Iteratively generate the lower-layer interface until all formation interfaces are generated;

[0030] 3), Follow the principle that the propagation speed value increases with the depth of the medium, and set the top-layer speed value to q 1 , Ensure that the speed value q of the m-th layer m -q m-1 >q min Under the condition, iteratively add speed values to each layer.

[0031] In the step (1) of step 1 of the present invention, in constructing the set V of velocity models of the pseudo-real geological structure, the wavy velocity model is constructed according to the following steps:

[0032] 1), According to formula (1), set a min <a j <a max , j = 1, 2, 3, T min <T j <T max , j = 1, 2, 3, θ min <θ j <θ max , j = 1, 2, 3, r 1 ≠ 0, c 1 > 0 to generate a wavy curve as the top interface of the velocity model;

[0033] 2), According to formula (1), set a min <a j <a max , j = 1, 2, 3, T min <T j <T max , j = 1, 2, 3, θ min <θ j <θ max , j = 1, 2, 3, r m ≠ 0, (y m ) min >(y m-1 ) max Iteratively generate the lower-layer interface until all formation interfaces are generated;

[0034] 3), Follow the principle that the propagation speed value increases with the depth of the medium, and set the top-layer speed value to q 1 , Ensure that the speed value q of the m-th layer m -q m-1 >q min Under the condition, iteratively add speed values to each layer.

[0035] In the step 1 (1) of the present invention, in constructing the velocity model set V of the pseudo-real geological structure, based on the wave-like velocity model, the fault velocity model is constructed according to the following steps:

[0036] 1), Select the constructed wave-like velocity model, randomly set the fault starting point at the top interface, set the fault line slope z < 0, and generate the fault line;

[0037] 2), Select the velocity model on the right side of the fault line and move it downward along the fault line by h, where h < c 1 ;

[0038] 3), To avoid the lack of velocity values caused by the movement of the velocity model, under the condition of ensuring that the velocity value q of the m-th layer m -q m-1 >q min Iteratively add velocity values to each layer where velocity values are missing.

[0039] In the step 1 (2) of the present invention, constructing the DAS-VSP dataset D is specifically as follows:

[0040] Select the velocity model v from the velocity model set V = {v i |i = 1, 2,... p}, set appropriate observation parameters and layouts to construct a two-dimensional observation system, ensure that the DAS-VSP data covers all parts of the velocity model, and can capture the subtle changes of the underground medium. Use the elastic wave forward modeling method to simulate the actual observation process, and simulate the wavelet form received by the geophone with a single Ricker wavelet i , where A is the amplitude, f

[0041]

[0042] is the main frequency of the Ricker wavelet, t 0 is the wavelet delay time. Simulate the propagation process of seismic waves in the formation by the wave equation, and generate the corresponding DAS-VSP forward modeling data d 0 by the finite difference method i ,i∈[1,p],

[0043] Traverse all the velocity models in the forward modeling velocity model set V to form the DAS-VSP dataset D = {d i |i = 1, 2,... p}.

[0044] In the step 2 (1) of the present invention, constructing the multi-task learning network includes:

[0045] The encoder consists of m 1 Conv+BN+ReLu and m 2 feature extraction modules, and each feature extraction module consists of m3 one Max pooling and m 4 consisting of several Conv+BN+ReLu, which realizes the extraction of high-level features of DAS-VSP data. Among them, multiple convolutional layers can fully extract the features in DAS-VSP data, and the activation function can ensure the non-linear relationship between layers, improving the mapping ability of the network from DAS-VSP data to high-level features;

[0046] The decoder consists of a main task decoder and a sub-task decoder. The main task decoder consists of n 1 DeConv, n 2 feature reconstruction modules and n 3 Conv+BN+ReLu. Each feature reconstruction module consists of n 4 Conv+BN+ReLu and n 5 DeConv. The main task decoder is used to convert high-level features into the prediction result of the velocity model. Among them, multiple deconvolution layers can decode and reconstruct high-level features into a fine velocity model, and the activation function can improve the mapping ability of the network from high-level features to the velocity model; The sub-task decoder also consists of n 1 DeConv, n 2 feature reconstruction modules and n 3 Conv+BN+ReLu. Each feature reconstruction module consists of n 4 Conv+BN+ReLu and n 5 DeConv. The sub-task decoder is only used to convert high-level features into the result of the velocity model structure. Among them, multiple deconvolution layers can decode and reconstruct high-level features into a fine velocity model formation structure, and the activation function can improve the mapping ability of the network from high-level features to the velocity model structure;

[0047] Perform skip connections between the results of each Max pooling in the encoder and the results of DeConv of the same-sized feature maps in the main decoder and the sub-decoder, weaken the loss of DAS-VSP feature information in the downsampling process, and fully retain the feature information related to velocity reconstruction; Introduce an attention module at the end of the network skip connection to filter out the redundant information irrelevant to the velocity model reconstruction extracted by the encoder, highlighting the key velocity information and structural high-level features; Adopt a multi-task learning network architecture, and use the structural information predicted by the sub-task decoder as the structural loss to constrain the main task decoder to achieve high-precision velocity model reconstruction.

[0048] In step (2) of step 2 of the present invention, train the multi-task learning network as follows:

[0049] Randomly select groups of paired data from the constructed set V of velocity models of pseudo-real geological structures and the DAS-VSP data set D Construct a training set and

[0050] DAS-VSP data d i and velocity model v i The forward relationship between them is expressed as d i = F(v i ). For velocity model prediction, the corresponding velocity model needs to be inferred from the DAS-VSP data d i That is: Namely:

[0051]

[0052] where G 1 represents the mapping function represented by the main task decoder, and θ represents the model parameters of the multi-task network.

[0053] During the velocity model prediction process, the loss function is used to update the parameters.

[0054] The loss function and parameter update process used in the present invention are as follows:

[0055] 1), Mean squared error loss L MSE

[0056] The mean squared error (MSE) loss function is defined as follows:

[0057]

[0058] where m represents the number of pixels of the velocity model, and k represents the amount of training data;

[0059] 2), Cross-entropy loss L contour

[0060] The cross-entropy loss function is defined as follows:

[0061]

[0062] where C(v i ) is the contour structure of the velocity model v i , is the inversion of C(v i ), and s i represents the result of performing a softmax operation on two channels of the sub-task decoder:

[0063]

[0064] where is the result of the nth, n = 1, 2 channels of the sub-task decoder;

[0065] 3) Overall loss and update of model parameters θ

[0066] The overall loss function is defined as follows:

[0067] L total = a 1 L MSE + a 2 L contour (7)

[0069] where α 1 and α 2 respectively represent the weights of L MSE and L contour in the overall loss. The loss function is used to calculate the deviation between the true value and the predicted value of the velocity model in the training set. Different model parameters θ correspond to different losses. The goal of model training is to find a set of model parameters that minimize the loss At this time, it is considered that the network model reaches the optimal performance. Using the paired data in the training set to train the multi-task learning network, and using the Adam optimizer to accelerate the convergence of the training process. The trained multi-task learning network can predict the velocity model v i given the DAS-VSP data d i .

[0070] The prediction method in step 3 of the present invention is as follows:

[0071] Select the remaining group of paired data {v i , d i} from the constructed set of true-like geological structure velocity models V and the DAS-VSP data set D to form a test set and

[0072] Use the multi-task learning network to perform velocity model prediction tests on the DAS-VSP data, that is, select the DAS-VSP data test from the test set D as the network test input, and the output result of the main task decoder is the result of the velocity model prediction for the DAS-VSP data

[0073] The present invention proposes a solution for an intelligent DAS-VSP prediction velocity model. By automatically constructing a set of velocity models of pseudo-real geological structures, it provides a large amount of data support for network training. At the same time, a multi-task network model is constructed and trained, and an attention gating mechanism is introduced to prompt the network to focus on features highly relevant to velocity model prediction and weaken the interference of irrelevant information. In addition, the sub-task of extracting formation structures is used to constrain the main task of velocity model prediction for high-precision velocity model reconstruction. The velocity model prediction solution provided by the present invention aims to provide a solid foundation for high-precision imaging of subsequent DAS-VSP data.

[0074] The advantages of the present invention are as follows: The multi-task learning network proposed by the present invention realizes an intelligent DAS-VSP data prediction velocity model solution. First, it automatically constructs a sufficient set of velocity models of pseudo-real geological structures, which can effectively solve the problem of the lack of velocity model sets. Second, a multi-task learning network is constructed and trained, and an attention mechanism is introduced to prompt the network to focus on features highly relevant to velocity model prediction. In addition, the sub-task of extracting formation structures is used to constrain the main task of velocity model prediction, reducing the instability of prediction results and improving the prediction accuracy of velocity models at the same time. The current traditional methods for velocity model prediction require a large amount of computing resources and are highly dependent on the initial model. It is more difficult to use traditional methods to achieve velocity model prediction for DAS-VSP data. The present invention provides a low-cost, high-efficiency and high-precision velocity model prediction method, which is crucial for DAS-VSP imaging quality, geological formation interpretability and oil and gas exploitation efficiency, and provides strong technical support for DAS technology in the field of seismic exploration. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a schematic diagram of a velocity model and a 2D observation system constructed, including horizontal, inclined, wavy and fault structures. The black inverted triangles are the geophones arranged by the DAS system, and the white circles are the seismic sources arranged by the DAS system;

[0076] Figure 2 is Figure 1 Four DAS-VSP record diagrams obtained by using the elastic wave forward modeling method for the four velocity models in

[0077] Figure 3 It is a multi-task learning network structure diagram, which consists of an encoder - double decoder structure and is in an overall Y-shaped structure. The main task decoder is mainly used for high-precision fitting of velocity values, and the sub-task decoder assists the main decoder in smooth fitting of layer boundaries. It also includes a feature extraction module and a feature reconstruction module. The feature extraction module consists of 1 Max pooling and 2 Conv+BN+ReLu, and the feature reconstruction module consists of 2 Conv+BN+ReLu and 1 DeConv;

[0078] Figure 4 It is a comparison graph of the prediction results and the ground truth of the multi-task learning network processing the undulating velocity model, where a) is the ground truth and b) is the prediction result of the trained multi-task learning network;

[0079] Figure 5 It is a comparison graph of the prediction results and the ground truth of the multi-task learning network processing the fault velocity model, where a) is the ground truth and b) is the prediction result of the trained multi-task learning network;

[0080] Figure 6 It corresponds to Figure 4 The velocity-depth curve at the position of the offset of 240 m of the undulating velocity model, which intuitively shows the difference between the prediction result and the ground truth;

[0081] Figure 7 It corresponds to Figure 5 The velocity-depth curve at the position of the offset of 240 m of the fault velocity model, which intuitively shows the difference between the prediction result and the ground truth;

[0082] Figure 8 It is a graph of the comparison experiment results between the multi-task learning network and the network without sub-task constraints, where a) is the ground truth, b) is the prediction result of the multi-task learning network, and c) is the prediction result of the network without sub-task constraints. Specific implementation method

[0083] It includes the following steps:

[0084] Step 1: Construct a set of velocity models of pseudo-real geological structures and a DAS-VSP data set;

[0085] (1) Construct a set of velocity models of pseudo-real geological structures V;

[0086] The velocity model construction process includes drawing a formation curve and filling velocity values. Drawing the formation structure curve requires using a linear function and a trigonometric function to simulate a continuous and undulating formation structure;

[0087] The curve equation of the m-th layer can be expressed as follows:

[0088]

[0089] where a j , T j , θ j , j = 1, 2, 3 are the amplitude, period, and phase parameters of the trigonometric function, used to simulate different situations of formation fluctuations; rm is the formation dip parameter, c m is the layer depth parameter;

[0090] Construct a horizontal velocity model according to the following steps:

[0091] 1), Set a j = 0, j = 1, 2, 3, r 1 = 0, c 1 > 0 to generate a straight line as the top interface of the velocity model;

[0092] 2), Set a according to formula (1) j = 0, j = 1, 2, 3, r m = 0, c m > c m-1 Iteratively generate the lower interface until all formation interfaces are generated;

[0093] 3) Follow the principle that the propagation velocity value is larger in the deeper medium, and set the top layer velocity value as q 1 , Ensure that the velocity value q of the m-th layer m - q m-1 > q min Under the condition, iteratively add velocity values to each layer;

[0094] Construct an inclined velocity model according to the following steps:

[0095] 1), Set a according to formula (1) j = 0, j = 1, 2, 3, r 1 ≠ 0, c 1 > 0 to generate an oblique line as the top interface of the velocity model;

[0096] 2), Set a according to formula (1) j = 0, j = 1, 2, 3, r m ≠ 0, (y m ) min > (y m-1 ) max Iteratively generate the lower interface until all formation interfaces are generated;

[0097] 3), Follow the principle that the propagation velocity value is larger in the deeper medium, and set the top layer velocity value as q 1 , Ensure that the velocity value q of the m-th layer m - q m-1 > q min Under the condition, iteratively add velocity values to each layer;

[0098] Construct a wavy velocity model according to the following steps:

[0099] 1), Set a according to formula (1)min <a j <a max , j = 1, 2, 3, T min <T j <T max , j = 1, 2, 3, θ min <θ j <θ max , j = 1, 2, 3, r 1 ≠0, c 1 > 0 generates a wave curve as the top interface of the velocity model;

[0100] 2), According to formula (1), set a min <a j <a max , j = 1, 2, 3, T min <T j <T max , j = 1, 2, 3, θ min <θ j <θ max , j = 1, 2, 3, r m ≠0, (y m ) min >(y m-1 ) max Iteratively generate the lower interface until all formation interfaces are generated;

[0101] 3), Follow the principle that the propagation velocity value is larger in deeper media, set the top layer velocity value as q 1 , ensure that the velocity value q of the mth layer m -q m-1 >q min Under the condition, iteratively add velocity values to each layer.

[0102] On the basis of the wave-shaped velocity model, construct the fault velocity model according to the following steps:

[0103] 1), Select the constructed wave-shaped velocity model, randomly set the fault starting point at the top interface, set the fault line slope z < 0, and generate the fault line;

[0104] 2), Select the velocity model on the right side of the fault line and move it down h along the fault line, h < c 1 ;

[0105] 3), To avoid the lack of velocity values caused by the movement of the velocity model, ensure that the velocity value q of the mth layer m -q m-1 >q min Under the condition, iteratively add velocity values to each layer where the velocity value is missing;

[0106] Generate p randomly independent velocity models v in batches according to the above velocity model construction method i , i ∈ [1, p], to form the velocity model set V = {v i | i = 1, 2, … p};

[0107] (2) Construct the DAS-VSP dataset D, and obtain the DAS-VSP data by forward modeling of the velocity model;

[0108] Select the velocity model v i from the velocity model set V = {v i , i ∈ [1, p], set appropriate observation parameters and layouts to construct a 2D observation system to ensure that the DAS-VSP data covers all parts of the velocity model and can capture the subtle changes of the underground medium. Use the elastic wave forward modeling method to simulate the actual observation process. The wavelet form received by the geophone is simulated by a single Ricker wavelet (see

[0109] Equation 2):

[0110]

[0111] where A is the amplitude, f 0 is the main frequency of the Ricker wavelet, t 0 is the wavelet delay time. Simulate the propagation process of seismic waves in the formation by the wave equation, and generate the corresponding DAS-VSP forward modeling data d i , i ∈ [1, p];

[0112] Traverse all the velocity models in the forward modeling velocity model set V to form the DAS-VSP dataset D = {d i | i = 1, 2, … p};

[0113] Step 2: Construct and train a multi-task learning network;

[0114] (1) Construct a multi-task learning network

[0115] The multi-task learning network adopts an encoder-dual decoder architecture. The main decoder performs the main task of velocity model prediction, and the sub-decoder performs the sub-task of velocity model structure reconstruction;

[0116] The encoder is composed of m 1 Conv+BN+ReLu and m 2 feature extraction modules. Each feature extraction module consists of m 3 Max pooling and m 4It consists of a Conv + BN + ReLu to achieve advanced feature extraction of DAS - VSP data. Among them, multiple convolutional layers can fully extract features in DAS - VSP data, and the activation function can ensure the non - linear relationship between layers, improving the mapping ability of the network from DAS - VSP data to advanced features;

[0117] The decoder consists of a main - task decoder and a sub - task decoder. The main - task decoder consists of n 1 DeConvs, n 2 feature reconstruction modules and n 3 Conv + BN + ReLus. Each feature reconstruction module consists of n 4 Conv + BN + ReLus and n 5 DeConvs. The main - task decoder is used to convert advanced features into speed model prediction results. Among them, multiple de - convolutional layer decodings can reconstruct advanced features into a fine - grained speed model, and the activation function can improve the mapping ability of the network from advanced features to the speed model; The sub - task decoder also consists of n 1 DeConvs, n 2 feature reconstruction modules and n 3 Conv + BN + ReLus. Each feature reconstruction module consists of n 4 Conv + BN + ReLus and n 5 DeConvs. The sub - task decoder is only used to convert advanced features into speed model structure results. Among them, multiple de - convolutional layer decodings can reconstruct advanced features into a fine - grained speed model formation structure, and the activation function can improve the mapping ability of the network from advanced features to the speed model structure;

[0118] Perform skip connections between each Max pooling result of the encoder and the DeConv results of feature maps of the same size in the main decoder and the sub - decoder, weaken the loss of DAS - VSP feature information in the down - sampling process, and fully retain the feature information related to speed reconstruction; Introduce an attention module at the end of the network skip connection to filter out redundant information unrelated to speed model reconstruction extracted by the encoder, highlighting key speed information and structural advanced features; Adopt a multi - task learning network architecture, use the structural information predicted by the sub - task decoder as the structural loss to constrain the main - task decoder to achieve high - precision speed model reconstruction;

[0119] (2) Train the multi - task learning network

[0120] Randomly select groups of paired data from the constructed set V of class - true geological structure speed models and the DAS - VSP data set D to form the training set and

[0121] The forward relationship between DAS-VSP data d i and the velocity model v i is expressed as d i = F(v i ). For velocity model prediction, the corresponding velocity model needs to be deduced from the DAS-VSP data d i . That is:

[0122]

[0123] where G 1 represents the mapping function represented by the main task decoder, and θ represents the model parameters of the multi-task network.

[0124] The loss function and parameter update process used in the velocity model prediction process of the present invention are as follows:

[0125] 1), Mean square error loss L MSE

[0126] The mean square error (MSE) loss function is defined as follows:

[0127]

[0128] where m represents the number of pixels of the velocity model, and k represents the amount of training data;

[0129] 2), Cross-entropy loss L contour

[0130] The cross-entropy loss function is defined as follows:

[0131]

[0132] where C(v i ) is the contour structure of the velocity model v i , is the inversion of C(v i ), and s i represents the result of performing a softmax operation on two channels of the sub-task decoder:

[0133]

[0134] where is the result of the nth, n = 1, 2 channels of the sub-task decoder;

[0135] 3), Overall loss and model parameter θ update;

[0136] The overall loss function is defined as follows:

[0137] Ltotal = a 1 L MSE + a 2 L contour (7)

[0139] where α 1 and α 2 represent the weights of L MSE and L contour in the overall loss respectively. The loss function is used to calculate the deviation between the true value and the predicted value of the velocity model in the training set. Different model parameters θ correspond to different losses. The goal of model training is to find a set of model parameters that minimize the loss At this time, it is considered that the network model reaches the optimal performance. Using the paired data in the training set to train the multi-task learning network, and using the Adam optimizer to accelerate the convergence of the training process. After training, the multi-task learning network can predict the velocity model v i given the DAS-VSP data d i ;

[0140] Step 3: Predict the velocity model for DAS-VSP data

[0141] Select the remaining paired data {v , d i , d i} from the constructed set V of velocity models of the pseudo-real geological structure and the DAS-VSP data set D to form the test set and Use the multi-task learning network to predict the velocity model for the DAS-VSP data, that is, select the DAS-VSP data test from the test set D as the network test input, and the output result of the main task decoder is the result of predicting the velocity model for the DAS-VSP data

[0142] The following further illustrates the present invention through experimental examples

[0143] In this experimental example, according to the geological conditions and actual requirements, a large number of velocity models are automatically drawn, covering different stratigraphic structures and velocity distributions. A two-dimensional observation system is constructed according to the DAS system layout and forward modeling formation parameters in Table 1 to ensure that the DAS-VSP data covers all parts of the velocity model and can capture the subtle changes of the underground medium. In addition, the elastic wave forward modeling method is used to simulate the DAS-VSP data generated by the actual observation system, and the multi-task learning network is trained and tested

[0144] Step 1: Construct a set of velocity models of pseudo-real geological structures and a DAS-VSP data set;

[0145] (1). Construct a set of velocity models of pseudo-real geological structures V

[0146] Referring to the real geological structure, set the size of each velocity model to 500×100, the grid size to 5m×5m, the simulated formation structure size to 2500m×500m, and set the velocity parameter q 1＝ 2000,q min = 200, the formation dip parameter -1 < r m < 1, the layer depth 10 < c m < 20, for the wavy velocity model, set a min = -1, a max = 1, T min = -2, T max = 2, θ min = -π, θ max = π, and batch draw 8000 velocity models v including horizontal, inclined, wavy and fault velocity models in the ratio of 2:3:10:5 to form the velocity model set V = {v i |i = 1,2,…8000};

[0147] Figure 1 The schematic diagrams of four structural velocity models are shown, and these schematic diagrams include Figure 1 a) horizontal, Figure 1 b) inclined, Figure 1 c) wavy and Figure 1 d) faults and other different geological structures, providing rich training data and test samples for subsequent velocity model prediction tasks;

[0148] (2). Construct a DAS-VSP data set D

[0149] Select the velocity model v from the velocity model data set V and deploy the two-dimensional DAS observation system as in Figure 1 , and the relevant forward simulation formation parameters are shown in Table 1.

[0150] Table 1. DAS system layout and forward simulation formation parameters

[0151]

[0152] A single-shot seismic source is deployed at an offset of 500 m, and 500 geophones are arranged in a well with a depth of 2500 m at intervals of 5 m. The DAS-VSP data forward simulation is carried out by using the sixth-order finite-difference method, and the PML (Perfectly Matched Layer) boundary absorption condition is used to absorb the non-physical reflections from each boundary. In the simulation, a Ricker wavelet with a dominant frequency of 35 Hz close to the actual DAS system acquisition record is adopted, and the DAS-VSP data is simulated at a sampling frequency of 2000 Hz, and the observation duration of each record is 2 seconds. Traverse the velocity models v in the velocity model set V, forward generate DAS-VSP data and form a seismic record data set D = {d i | i = 1, 2, … 8000};

[0153] Figure 2 Shows the DAS-VSP records corresponding to Figure 1 the four velocity models in Figure 2 a) is the DAS-VSP data corresponding to the horizontal velocity model, Figure 2 b) is the DAS-VSP data corresponding to the inclined velocity model, Figure 2 c) is the DAS-VSP data corresponding to the wavy velocity model, Figure 2 d) is the DAS-VSP data corresponding to the fault velocity model;

[0154] Step 2: Construct and train a multi-task learning network

[0155] (1) Construct a multi-task learning network

[0156] The multi-task learning model in this example is as Figure 3 shown. The multi-task learning network adopts an encoder-dual decoder architecture. The main decoder performs the main task of predicting the velocity model, and the sub-decoder performs the sub-task of reconstructing the velocity model structure;

[0157] The encoder is composed of 2 Conv+BN+ReLu and 4 feature extraction modules. Each feature extraction module is composed of 1 Maxpooling and 2 Conv+BN+ReLu. It realizes the high-level feature extraction of DAS-VSP data. Among them, multiple convolutional layers can fully extract the features in DAS-VSP data, and the activation function can ensure the non-linear relationship between layers and improve the mapping ability of the network from DAS-VSP data to high-level features;

[0158] The decoder consists of a main task decoder and a sub-task decoder. The main task decoder consists of 1 DeConv, 3 feature reconstruction modules, and 2 Conv+BN+ReLu. Each feature reconstruction module consists of 2 Conv+BN+ReLu and 1 DeConv. The main task decoder is used to convert high-level features into speed model prediction results. Among them, multiple deconvolution layers can decode and reconstruct high-level features into a fine speed model, and the activation function can improve the mapping ability between high-level features of the network and the speed model; The sub-task decoder also consists of 1 DeConv, 3 feature reconstruction modules, and 2 Conv+BN+ReLu. Each feature reconstruction module consists of 2 Conv+BN+ReLu and 1 DeConv. The sub-task decoder is only used to convert high-level features into speed model structure results. Among them, multiple deconvolution layers can decode and reconstruct high-level features into a fine speed model formation structure, and the activation function can improve the mapping ability between high-level features of the network and the speed model structure;

[0159] Perform skip connections between the Max pooling results of each encoder and the DeConv results of feature maps of the same size in the main decoder and the sub-decoder, weaken the loss of DAS-VSP feature information during the downsampling process, and fully retain the feature information related to speed reconstruction; Introduce an attention module at the end of the network skip connection to filter out the redundant information irrelevant to speed model reconstruction extracted by the encoder, and highlight the key speed information and structural high-level features; Adopt a multi-task learning network architecture, use the structural information predicted by the sub-task decoder as the structural loss, and constrain the main task decoder to achieve high-precision speed model reconstruction.

[0160] Table 2 Multi-task learning network parameter settings

[0161] Encoder Main Decoder Auxiliary Decoder Conv(1, 64, 3), BN, ReLu deconv(1024, 512, 2), concat deconv(1024, 512, 2), concat Conv(64, 64, 3), BN, ReLu Conv(1024, 1024, 3), BN, ReLu Conv(1024, 1024, 3), BN, ReLu max pooling(2, 2) Conv(512, 512, 3), BN, ReLu Conv(512, 512, 3), BN, ReLu Conv(64, 128, 3), BN, ReLu deconv(512, 256, 2), concat deconv(512, 256, 2), concat Conv(128, 128, 3), BN, ReLu Conv(512, 256, 3), BN, ReLu Conv(512, 256, 3), BN, ReLu max pooling(2, 2) Conv(256, 256, 3), BN, ReLu Conv(256, 256, 3), BN, ReLu Conv(128, 256, 3), BN, ReLu deconv(256, 128, 2), concat deconv(256, 128, 2), concat Conv(256, 256, 3), BN, ReLu Conv(256, 128, 3), BN, ReLu Conv(256, 128, 3), BN, ReLu max pooling(2, 2) Conv(128, 128, 3), BN, ReLu Conv(128, 128, 3), BN, ReLu Conv(256, 512, 3), BN, ReLu deconv(128, 64, 2), concat deconv(128, 64, 2), concat Conv(512, 512, 3), BN, ReLu Conv(128, 64, 3), BN, ReLu Conv(128, 64, 3), BN, ReLu max pooling(2, 2) Conv(64, 64, 3), BN, ReLu Conv(64, 64, 3), BN, ReLu Conv(512, 1024, 3), BN, ReLu Conv(64, 1, 1) Conv(64, 2, 1) Conv(1024, 1024, 3), BN, ReLu

[0162] (2), Train the multi-task learning network

[0163] Manually initialize the model weight parameter α 2 / α 1 = 10 6 . In the constructed speed model set and the paired DAS-VSP dataset, randomly select 6400 pairs of paired data and construct the training set V train ={v i |i = 1, 2, …6400} and D train ={d i |i = 1, 2, …6400}. Take every 16 pairs as a batch and send them into the model for training batch by batch. Use formulas (4) and (5) to calculate the loss functions L MSE and L contour, the gradient of the loss function is backpropagated, and the Adam optimizer is used to continuously correct the parameters of each layer of the network by using the forward propagation and backpropagation algorithms to alternately update the network parameters. After training for 200 epochs, the overall loss L in formula (7) total is minimized and tends to converge. At this time, the multi-task learning network reaches the expected performance, and this network will be used for velocity model prediction.

[0164] Step 3, Perform velocity model prediction on DAS-VSP data

[0165] When the multi-task learning network model training is completed, use the DAS-VSP data to perform a velocity model prediction performance test on this model. In this example, data other than the training data is selected to construct the test set V test ={v j |j = 1, 2, …1600} and D test ={d j |j = 1, 2, …1600}, and paired data is selected for velocity model prediction. The DAS-VSP records in the paired data are input into the trained network model, and the velocity model prediction results output by the main task decoder are compared with the labeled velocity model in the paired data to analyze the velocity model prediction performance of the multi-task learning network. The prediction results are as Figure 4 and Figure 5 shown. Two representative wavy and fault velocity models are selected for visual comparison. Among them, Figure 4 a) and Figure 4 b) are the prediction results and the ground truth of the wavy velocity model respectively, Figure 5 a) and Figure 5 b) are the prediction results and the ground truth of the fault velocity model respectively. It can be seen that this network can accurately predict the velocity value and formation structure of each layer, and can restore the micro-structures in the formation. To further refine the difference between the prediction results and the ground truth, in Figure 6 and Figure 7 respectively, the curves of velocity vs. depth at an offset of 240 meters for the wavy velocity model and the fault velocity model in Figure 4 and Figure 5 are plotted. It can be seen that the method of the present invention does not depend on the initial velocity model, but can accurately fit the velocity value of each layer, indicating the accuracy of this model. To further verify the performance of the multi-task learning network, a network without sub-task constraints is trained and a comparative test is carried out. Figure 8 a) is the ground truth of the velocity model, Figure 8 b) is the prediction result of the multi-task network, Figure 8c) is the prediction result of the constraint network without sub-tasks. It can be seen that the multi-task network has obvious advantages in both structural characterization and deep velocity value prediction, indicating the stability of the model. To quantify the prediction effect of the method of the present invention, the following four indicators are selected: MAE (Mean Absolute Error), MSE (Mean Squared Error), UQI (Universal Quality Index), and LPIPS (Learned Perceptual Image Patch Similarity). The results are shown in Table 3.

[0166] Table 3. Statistical indicators of the prediction results of the present invention

[0167]

[0168] It can be observed that MAE and LPIPS reach 10 -2 and MSE reaches 10 -3 while UQI is close to 1. The results show that the model has high reliability and demonstrates the superior performance of the method of the present invention. Through the above qualitative and quantitative analyses, it is verified that the multi-task learning network model can not only accurately predict the formation velocity distribution but also effectively capture the subtle structural changes in the formation, providing high-precision technical support for the DAS-VSP data in the process of velocity model prediction.

[0169] In summary, the current traditional methods for velocity model prediction require a large amount of computing resources and are highly dependent on the initial model. It is more difficult to use traditional methods to achieve velocity model prediction for DAS-VSP data. The present invention proposes an intelligent multi-task learning scheme for predicting the velocity model using DAS-VSP data. First, a sufficient amount of realistic geological structure velocity model sets are automatically constructed to solve the problem of the lack of velocity model sets. Second, a multi-task learning network model is constructed and trained, and an attention mechanism is introduced to prompt the network to focus on the features highly relevant to velocity model prediction. In addition, the sub-task of extracting the formation structure is used to constrain the main task of velocity model prediction, reducing the instability of the prediction results and improving the reconstruction accuracy of the velocity model.

[0170] The present invention automatically extracts high-level features from DAS-VSP data through a multi-task learning network to achieve high-precision velocity model prediction, providing a low-cost, high-efficiency, and high-precision velocity model prediction method, which is crucial for DAS-VSP imaging quality, geological formation interpretation, and oil and gas extraction efficiency, and provides strong technical support for DAS technology in the field of seismic exploration.

Claims

1. A multi-task learning speed model prediction method, characterized in that: The following steps are involved: Step 1: Construct a velocity model set of quasi-real geological structures and a DAS-VSP dataset; (1) Construct a set of velocity models V that resemble real geological structures; The velocity model construction process includes drawing formation curves and filling velocity values. Drawing formation structure curves requires the use of linear and trigonometric functions to simulate continuous and fluctuating formation structures. The curve equation of the mth layer can be expressed as follows: where a j ,T j ,θ j , j = 1, 2, 3 are the trigonometric function amplitude, period, and phase parameters, which are used to simulate formation fluctuations in different situations. m is the formation inclination parameter, c m is the layer depth parameter: Construct horizontal velocity model, inclined velocity model, wave velocity model and fault velocity model; According to the above velocity model construction method, p random independent velocity models v are generated in batches. i , i∈[1,p], forming a velocity model set V={v i |i=1,2,…p}; (2) Construct the DAS-VSP dataset D and obtain DAS-VSP data by forward modeling of the velocity model; Step 2: Build and train a multi-task learning network; (1) Constructing a multi-task learning network The multi-task learning network adopts an encoder-dual decoder architecture. The main decoder performs the main task of velocity model prediction, and the sub-decoder performs the sub-task of velocity model structure reconstruction. (2) Train the multi-task learning network; Establish a training set. After training, the multi-task learning network can i The velocity model v is predicted under the condition of i ; Step 3: DAS-VSP data is used to make velocity model predictions; The output of the main task decoder is the result of velocity model prediction based on DAS-VSP data.

2. A multi-task learning speed model prediction method according to claim 1, characterized in that: In the step 1 (1), in constructing a velocity model set V of a similar real geological structure, a horizontal velocity model is constructed according to the following steps: 1) According to formula (1), set a j =0, j=1,2,3, r1=0, c1>0 to generate a straight line as the top interface of the velocity model; 2) According to formula (1), set a j =0,j=1,2,3,r m =0, c m >c m-1 Iteratively generate the lower layer interface until all the stratum interfaces are generated; 3) Following the principle that the deeper the medium, the greater the propagation velocity, the top layer velocity value is set to q1, ensuring that the mth layer velocity value is q m -q m-1 >q min Under these conditions, iteratively add velocity values ​​at each layer.

3. A multi-task learning speed model prediction method according to claim 1, characterized in that: In the step 1 (1), in constructing a velocity model set V of a similar real geological structure, an inclined velocity model is constructed according to the following steps: 1) According to formula (1), set a j =0, j=1,2,3, r1≠0, c1>0 generates a slant line as the top interface of the velocity model; 2) According to formula (1), set a j =0,j=1,2,3,r m ≠0, (y m ) min >(y m-1 ) max Iteratively generate the lower layer interface until all the stratum interfaces are generated; 3) Following the principle that the deeper the medium, the greater the propagation velocity, the top layer velocity value is set to q1, ensuring that the mth layer velocity value is q m -q m-1 >q min Under these conditions, iteratively add velocity values ​​at each layer.

4. A multi-task learning speed model prediction method according to claim 1, characterized in that: In the step 1 (1), in constructing a velocity model set V of a similar real geological structure, a wave velocity model is constructed according to the following steps: 1) According to formula (1), set a min j max ,j=1,2,3,T min <T j <T max ,j=1,2,3,​​ θ min <θ j <θ max ,j=1,2,3,r1≠0,c1>0 to generate a fluctuation curve as the top interface of the velocity model; 2) According to formula (1), set a min j max ,j=1,2,3,T min <T j <T max ,j=1,2,3,θ min <θ j <θ max ,j=1,2,3,r m ≠0, (y m ) min >(y m-1 ) max Iteratively generate the lower layer interface until all the stratum interfaces are generated;​​ 3) Following the principle that the deeper the medium, the greater the propagation velocity, the top layer velocity value is set to q1, ensuring that the mth layer velocity value is q m -q m-1 >q min Under these conditions, velocity values ​​are iteratively added to each layer.

5. A multi-task learning speed model prediction method according to claim 4, characterized in that: In the step 1 (1), in constructing a velocity model set V of a similar real geological structure, a fault velocity model is constructed based on the wave velocity model according to the following steps: 1) Select the constructed wave velocity model, randomly set the fault starting point at the top interface, set the fault line slope z<0, and generate the fault line; 2) Select the velocity model on the right side of the fault line and move down along the fault line h, h <c1; 3) To avoid the loss of speed value caused by the movement of the speed model, the speed value q of the mth layer is guaranteed. m -q m-1 >q min Under these conditions, iteratively add speed values ​​to each speed value missing layer.

6. A multi-task learning speed model prediction method according to claim 1, characterized in that: The step 1 (2) of constructing the DAS-VSP dataset D is as follows: From the velocity model set V = {v i |i=1,2,…p} select velocity model v i ,i∈[1,p], set appropriate observation parameters and layout to build a two-dimensional observation system, ensure that DAS-VSP data covers all parts of the velocity model and can capture subtle changes in the underground medium, use elastic wave forward modeling to simulate the actual observation process, and use a single Ricker wavelet to simulate the wavelet form received by the detector. Where A is the amplitude, f0 is the main frequency of the Ricker wavelet, and t0 is the wavelet delay time. The wave equation is used to simulate the propagation process of seismic waves in the stratum, and the corresponding DAS-VSP forward modeling data d is generated by the finite difference method. i ,i∈[1,p], Traverse all velocity models in the forward velocity model set V to form a DAS-VSP data set D = {d i |i=1,2,…p}.

7. A multi-task learning speed model prediction method according to claim 1, characterized in that: The step 2 (1) of constructing a multi-task learning network includes: The encoder is composed of m1 Conv+BN+ReLu and m2 feature extraction modules. Each feature extraction module is composed of m3 Maxpooling and m4 Conv+BN+ReLu, which can realize the high-level feature extraction of DAS-VSP data. Multiple convolutional layers can fully extract the features in DAS-VSP data. The activation function can ensure the nonlinear relationship between layers and improve the mapping ability of the network from DAS-VSP data to high-level features. The decoder is composed of a main task decoder and a subtask decoder. The main task decoder is composed of n1 DeConv, n2 feature reconstruction modules and n3 Conv+BN+ReLu. Each feature reconstruction module is composed of n4 Conv+BN+ReLu and n5 DeConv. The main task decoder is used to convert high-level features into velocity model prediction results, wherein multiple deconvolution layer decoding can reconstruct high-level features into a fine velocity model, and the activation function can improve the mapping ability between the network's high-level features and the velocity model; the subtask decoder is also composed of n1 DeConv, n2 feature reconstruction modules and n3 Conv+BN+ReLu. Each feature reconstruction module is composed of n4 Conv+BN+ReLu and n5 DeConv. The subtask decoder is only used to convert high-level features into velocity model structure results, wherein multiple deconvolution layer decoding can reconstruct high-level features into a fine velocity model stratigraphic structure, and the activation function can improve the mapping ability between the network's high-level features and the velocity model structure; A jump connection is made between each Max pooling result of the encoder and the DeConv results of the same-sized feature maps in the main decoder and sub-decoder to weaken the loss of DAS-VSP feature information in the downsampling process and fully retain the feature information related to speed reconstruction; an attention module is introduced at the end of the network jump connection to filter out redundant information irrelevant to speed model reconstruction extracted by the encoder and highlight key speed information and high-level structural features; a multi-task learning network architecture is adopted to use the structural information predicted by the sub-task decoder as the structural loss to constrain the main task decoder to achieve high-precision speed model reconstruction.

8. A multi-task learning speed model prediction method according to claim 1, characterized in that: In step 2 (2), training the multi-task learning network is as follows: Randomly selected from the constructed quasi-real geological structure velocity model set V and the DAS-VSP dataset D Group Paired Data Constructing the training set and DAS-VSP data i and velocity model v i The forward relationship between them is expressed as d i =F(v i ), for velocity model prediction, it is necessary to obtain the DAS-VSP data d i The corresponding velocity model is derived from Right now: Where G1 represents the mapping function represented by the main task decoder, θ represents the model parameters of the multi-task network, The loss function is used to update parameters in the velocity model prediction process.

9. A multi-task learning speed model prediction method according to claim 8, characterized in that: The loss function and parameter update process used are as follows: 1) Mean square error loss L MSE The mean square error (MSE) loss function is defined as follows: Where m represents the number of pixels of the velocity model, and k represents the amount of training data; 2) Cross entropy loss L contour The cross entropy loss function is defined as follows: Where C(v i ) is the velocity model v i The outline structure, is C(v i ), s i Represents the result of the softmax operation on the two channels of the subtask decoder: in The result of the nth, n=1,2th channel of the subtask decoder; 3) Overall loss and model parameter θ update The overall loss function is defined as follows: L total =a1L MSE +a2L contour (7) where α1 and α2 represent L MSE and L contour The weight occupied in the overall loss. The loss function is used to calculate the deviation between the true value and the predicted value of the velocity model in the training set. Different model parameters θ correspond to different losses. The goal of model training is to find a set of model parameters that minimize the loss. At this point, the network model is considered to have achieved optimal performance, using the paired data in the training set The multi-task learning network is trained and the Adam optimizer is used to accelerate the convergence of the training process. After training, the multi-task learning network can i The velocity model v is predicted under the condition of i .

10. A multi-task learning speed model prediction method according to claim 1, characterized in that: The prediction method of step 3 is as follows: Select the remaining ones from the constructed quasi-realistic geological structure velocity model set V and the DAS-VSP dataset D Group paired data {v i ,d i }, forming a test set and The multi-task learning network is used to test the velocity model prediction on DAS-VSP data, that is, from the test set D test Select DAS-VSP data As the network test input, the main task decoder output is the result of the velocity model prediction of DAS-VSP data.

Images