Low-frequency model construction method based on time-space gating circulation unit fusion network

Abstract

The invention discloses a low-frequency model construction method based on a time-space gating circulation unit fusion network, relates to the field of geophysical exploration of petroleum, and provides a low-frequency model prediction neural network framework oriented to small sample data. The framework takes a GRU as a basic unit to fully consider a long-term dependency relationship of a low-frequency model time sequence; the integrity of long wavelength information in the longitudinal direction of the constructed low-frequency model is guaranteed, in order to improve the continuity of the low-frequency model in the transverse direction, a spatial-temporal feature fusion module is designed to convert multi-channel data into potential vectors containing spatial information, meanwhile, envelope information and an initial model serve as prior information to guide a neural network, and the neural network is optimized; the method helps the network to generate stable low-frequency model prediction and reduce the nonlinearity of prediction, further enhances the generalization ability of the network, and generates spatial disturbance on original data to generate label data with more abundant combination modes and more types in the longitudinal direction.

Description

Technical Field

[0001] The invention belongs to the field of petroleum geophysical exploration, and specifically is a low-frequency model construction method based on a spatiotemporal gated cyclic unit fusion network. Background Art

[0002] The traditional seismic inversion method aims to use the forward algorithm to construct a target functional with the smallest error between synthetic seismic records and actual seismic records, and continuously iterate and update the initial model through the optimization algorithm to obtain the optimal solution. However, the actual seismic records lack low-frequency components. Based on the problems existing in the construction of conventional initial models, the superior mapping ability and flexibility of neural networks are used to construct a reasonable low-frequency model, laying the foundation for high-precision seismic inversion. However, in practical applications, neural networks are unstable and have poor generalization ability, making it difficult to be widely promoted and applied in production. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects in the prior art. In view of the problems of complex structure and sparse logging in actual work areas, the present invention provides a low-frequency model construction method based on a spatiotemporal gated recurrent unit fusion network to realize information mining of spatial lateral continuity and temporal long-term dependence, and improve the generalization ability of neural networks through multi-information fusion strategies and data enhancement methods.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solution: a method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network, comprising the following steps:

[0005] S1: Carry out data enhancement of wave impedance. On the basis of the original data set, the elastic transformation algorithm is introduced to enhance the data. The original data includes seismic data, m-well logging data constrained interpolation wave impedance model and wave impedance initial model. The initial model is obtained through logging constrained interpolation algorithm, full waveform inversion and seismic velocity. Elastic transformation is used on the wave impedance model and the initial model to obtain the spatial disturbance in the vertical direction.

[0006] S2: Establish a low-frequency model sample library, where seismic records are characterized by the convolution of seismic wavelets and formation reflection coefficients;

[0007] S3: Neural network construction and parameter training, building a neural network model including a temporal feature extraction module and a spatiotemporal feature fusion module, and building a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model;

[0008] S4: Neural network test, calculate the envelope of seismic data and use the seismic data, initial model and seismic envelope to form three-channel data as input. Similarly, the predicted data is standardized and the processed data is used to carry out neural network test.

[0009] Preferably, the elastic transformation process is mainly divided into three steps. First, a new spatial coordinate is obtained based on the original data coordinate space through affine transformation. Then, the new coordinate space is smoothed by Gaussian function to obtain a random displacement field. Finally, the original data is mapped to the new coordinate space to obtain the disturbed data. The elastic transformation process is expressed in the following form:

[0010]

[0011] Where D represents the generated random displacement field, T represents the time series, α controls the disturbance degree of the random displacement field, δ is the Gaussian filter kernel size, and controls the smoothness of the random displacement field. The larger α is and the smaller δ is, the more severe the spatial disturbance is. Conversely, the more moderate the spatial disturbance is. The elastic transformation algorithm changes the distribution and thickness of the reservoir parameters in the original sample, thereby increasing the diversity of the sample.

[0012] Preferably, the seismic record is represented by the convolution of the seismic wavelet and the formation reflection coefficient, and the relationship between the wave impedance and the reflection coefficient is as follows:

[0013]

[0014] Among them, t n represents the nth time sampling point, Δt represents the time increment, R represents the reflection coefficient, and Z represents the wave impedance.

[0015] The convolution model process is expressed as follows:

[0016] S=R*W+N (3)

[0017] Among them, S is seismic data, R is reflection coefficient, W is Ricker wavelet, N is noise, and the expression of Ricker wavelet W is as follows:

[0018]

[0019] where f m is the main frequency of the seismic wavelet, t represents time. In order to further expand the diversity of samples, the main frequency of the seismic data is determined by time-frequency analysis. Five main frequency information are obtained by adding and subtracting at intervals of 5 Hz based on the main frequency of the actual seismic data.

[0020] Preferably, five main frequency information is obtained by adding and subtracting at intervals of 5 Hz based on the main frequency of the actual seismic data, and the seismic record is subjected to Hilbert transform using the following formula:

[0021]

[0022] Among them, y(t) represents the earthquake record, y H(t) represents the Hilbert transform of the seismic record, τ represents the integral variable, and the synthetic seismic record of the augmented wave impedance model is obtained by combining formula (5) and the convolution model S=R*W+N.

[0023] Preferably, a high-pass filter is used to filter out the low-frequency components of the synthetic data according to the original data frequency band to simulate the lack of low-frequency components of seismic data in actual situations. The envelope of the band-limited seismic data still carries the long-wavelength information of the underground medium. The envelope operator is a nonlinear demodulation operator. The envelope of the seismic signal is obtained by obtaining the amplitude of the analytical signal. The process can be written as:

[0024]

[0025] Where e represents the envelope of the seismic signal. The final sample library contains seismic data, seismic envelope, initial model and low-frequency model. The seismic data, seismic envelope and initial model are used as input, the low-frequency model is the training label, and the input data is 5 channels beside the well to form 101 pairs of two-dimensional data to enhance the spatial representation ability of the prediction results. The label data is one-dimensional data. Finally, the data is standardized to reduce the difficulty of neural network training. The low-frequency model is standardized using the mean and variance of the initial model.

[0026] Preferably, a single gated recurrent unit (GRU) improves a simple recurrent neural network by coupling a reset gate and an update gate, and the process can be expressed as follows:

[0027]

[0028] Among them, u t 、r t , X t , and h t They represent the update gate vector, reset gate vector, input sample, candidate hidden state and hidden state at the t-th time point respectively. w and b represent the weight parameter and bias term respectively. Sigmoid and tanh represent different activation functions. * represents the dot product operator.

[0029] Preferably, the bidirectional gated recurrent network (Bi-GRU) unit builds a back propagation layer based on the GRU, taking into account the forward and reverse time dependencies. The temporal feature extraction module is composed of multiple parallel 3-series Bi-GRUs. The number of parallel blocks depends on the value of i in the S4 step, which is set to 2, that is, there are 5 parallel blocks. The spatiotemporal feature fusion module is composed of a Bi-GRU and a fully connected layer.

[0030] Preferably, the parameter training step of the neural network includes: initializing the network weight parameter w and the bias term b, inputting the two-dimensional seismic data, the initial model and the seismic envelope in the sample library into the neural network, outputting a one-dimensional predicted low-frequency model, and then constructing a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model. The loss function is expressed as follows:

[0031]

[0032] Where N is the number of sampling points per channel, f is the network model, and Z L is the actual low-frequency model, S is the earthquake record, Z init is the initial model, e is the seismic envelope, w and b represent the weight parameter and bias term of the network respectively, the first term of the loss function is the wave impedance low-frequency model loss term, the second term is the initial model loss term, and λ is the weight used to balance the two losses and control the similarity between the prediction results and the initial model.

[0033] Preferably, the derivatives corresponding to each step of the input data through the neural network are recorded, and the partial derivatives of the loss function with respect to the weight parameters and bias terms are obtained using the chain rule. Subsequently, the Adam optimizer based on variable step-size gradient descent is used to calculate the update amount of the weight parameters and bias terms. The learning rate is set to 0.05, and the weight parameters and bias terms are updated. This process is repeated until the loss function error is minimized, and finally the network weight parameters and bias terms are saved.

[0034] A low-frequency model construction method based on a spatiotemporal gated recurrent unit fusion network disclosed in the present invention also includes a neural network testing method, comprising the following steps:

[0035] (1) When constructing the low-frequency model of the test data, the envelope of the seismic data is first calculated and the seismic data, the initial model and the seismic envelope are taken as the input to form three-channel data. The predicted data is also standardized.

[0036] (2) Load the trained network weight parameters and bias items to construct a low-frequency model for the test data, extract the variance σ and mean μ of the initial model, and perform denormalization on the predicted low-frequency model. The denormalization is expressed as follows:

[0037] x=σX+μ (9)

[0038] Where X is the normalized data and x is the denormalized data.

[0039] Beneficial effect: The low-frequency model construction method based on the space-time gated recurrent unit fusion network disclosed in the present invention is a new joint inversion structure coupling idea compared with the existing known technology. It uses singular values ​​to characterize the spatial distribution characteristics of multiple physical properties of the model, constructs an objective function containing low-rank structural constraints, realizes the joint inversion of multiple geophysical fields, and improves the correlation of multiple physical property models. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0041] In the attached picture:

[0042] Figure 1 It is a flow chart of a method for constructing a low-frequency model based on data enhancement of the present invention;

[0043] Figure 2 It is a schematic diagram of the spatiotemporal gated recurrent fusion network of the present invention;

[0044] Figure 3 It is the original data of Marmousi2 of the present invention;

[0045] Figure 4 It is the disturbance effect of different α and δ combinations of the data set enhancement method based on elastic transformation of the present invention;

[0046] Figure 5 These are the prediction results and absolute errors of different network models of the present invention. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The following text is only used to describe an implementation method of a low-frequency model construction method based on a spatiotemporal gated recurrent unit fusion network of the present invention, and does not strictly limit the scope of protection specifically requested by the present invention.

[0048] In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the fact that they can be implemented by ordinary technicians in the field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the invention.

[0049] Embodiment 1:

[0050] like Figure 1 , Figure 2 and Figure 4As shown, the present invention discloses a method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network, comprising the following steps:

[0051] S1: Carry out data enhancement of wave impedance. On the basis of the original data set, the elastic transformation algorithm is introduced to enhance the data. The original data includes seismic data, m-well logging data constrained interpolation wave impedance model and wave impedance initial model. The initial model is obtained through logging constrained interpolation algorithm, full waveform inversion and seismic velocity. Elastic transformation is used on the wave impedance model and the initial model to obtain the spatial disturbance in the vertical direction.

[0052] S2: Establish a low-frequency model sample library, where seismic records are characterized by the convolution of seismic wavelets and formation reflection coefficients;

[0053] S3: Neural network construction and parameter training, building a neural network model including a temporal feature extraction module and a spatiotemporal feature fusion module, and building a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model;

[0054] S4: Neural network test, calculate the envelope of seismic data and use the seismic data, initial model and seismic envelope to form three-channel data as input. Similarly, the predicted data is standardized and the processed data is used to carry out neural network test.

[0055] In the first embodiment, the elastic transformation process is mainly divided into three steps. First, a new spatial coordinate is obtained based on the original data coordinate space through affine transformation, then the new coordinate space is smoothed by Gaussian function to obtain a random displacement field, and finally the original data is mapped to the new coordinate space to obtain the disturbed data. The elastic transformation process is expressed in the following form:

[0056]

[0057] Where D represents the generated random displacement field, T represents the time series, α controls the disturbance degree of the random displacement field, δ is the Gaussian filter kernel size, and controls the smoothness of the random displacement field. The larger α is and the smaller δ is, the more severe the spatial disturbance is. Conversely, the more gentle the spatial disturbance is. The elastic transformation algorithm changes the distribution and thickness of the reservoir parameters in the original sample, increasing the diversity of the sample, such as Figure 4 As shown, Figure 4 is the perturbation result of different α and δ combinations of data enhancement method based on elastic transformation, where Figure 4 (a) α = 1300, δ = 200, Figure 4 (b) α = 1300, δ = 130, Figure 4 (c) α = 2000, δ = 130, Figure 4In (d), α=130, δ=30. The present invention selects α=1300, δ=130, and the number of elastic transformations is set to 20. In addition to the original wave impedance model, 20 expanded wave impedance models that can characterize the characteristics of the original model and are more diverse are obtained through this process.

[0058] In the first embodiment, the seismic record is represented by the convolution of the seismic wavelet and the formation reflection coefficient. The relationship between the wave impedance and the reflection coefficient is as follows:

[0059]

[0060] Among them, t n represents the nth time sampling point, Δt represents the time increment, R represents the reflection coefficient, and Z represents the wave impedance.

[0061] The convolution model process is expressed as follows:

[0062] S=R*W+N (3)

[0063] Among them, S is seismic data, R is reflection coefficient, W is Ricker wavelet, N is noise, and the expression of Ricker wavelet W is as follows:

[0064]

[0065] where f m is the main frequency of the seismic wavelet, t represents time. In order to further expand the diversity of samples, the main frequency of the seismic data is determined by time-frequency analysis. Five main frequency information are obtained by adding and subtracting at intervals of 5 Hz based on the main frequency of the actual seismic data.

[0066] Preferably, five main frequency information is obtained by adding and subtracting at intervals of 5 Hz based on the main frequency of the actual seismic data, and the seismic record is subjected to Hilbert transform using the following formula:

[0067]

[0068] Among them, y(t) represents the earthquake record, y H (t) represents the Hilbert transform of the seismic record, τ represents the integral variable, and the synthetic seismic record of the augmented wave impedance model is obtained by combining formula (5) and the convolution model S=R*W+N.

[0069] In the first embodiment, a high-pass filter is used to filter out the low-frequency components of the synthetic data according to the original data band to simulate the lack of low-frequency components of seismic data in actual situations. The envelope of the band-limited seismic data still carries the long-wavelength information of the underground medium. The envelope operator is a nonlinear demodulation operator. The envelope of the seismic signal is obtained by obtaining the amplitude of the analytical signal. The process can be written as:

[0070]

[0071] Where e represents the envelope of the seismic signal. The final sample library contains seismic data, seismic envelope, initial model and low-frequency model. The seismic data, seismic envelope and initial model are used as input, and the low-frequency model is the training label. The input data is 5 channels beside the well to form 101 pairs of two-dimensional data to enhance the spatial representation ability of the prediction results. The label data is one-dimensional data. The data is standardized to reduce the difficulty of neural network training. The low-frequency model is standardized using the mean and variance of the initial model.

[0072] In the first embodiment, Figure 2 As shown in (a), a single gated recurrent unit (GRU) improves a simple recurrent neural network by coupling a reset gate and an update gate. The process can be expressed as follows:

[0073]

[0074] Among them, u t 、r t , X t , and h t They represent the update gate vector, reset gate vector, input sample, candidate hidden state and hidden state at the t-th time point respectively. w and b represent the weight parameter and bias term respectively. Sigmoid and tanh represent different activation functions. * represents the dot product operator.

[0075] In the first embodiment, Figure 2 As shown in (b), the bidirectional gated recurrent network (Bi-GRU) builds a back propagation layer on the basis of GRU, taking into account both the forward and reverse time dependencies. The temporal feature extraction module is composed of multiple parallel 3-series Bi-GRUs. The number of parallel blocks depends on the value of i in step S4. It is set to 2, that is, there are 5 parallel blocks, as shown in Figure 2 (c) The spatiotemporal feature fusion module consists of a Bi-GRU and a fully connected layer.

[0076] In the first embodiment, the parameter training step of the neural network includes: initializing the network weight parameter w and the bias term b, inputting the two-dimensional seismic data, the initial model and the seismic envelope in the sample library into the neural network, outputting a one-dimensional predicted low-frequency model, and then constructing a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model. The loss function is expressed as follows:

[0077]

[0078] Where N is the number of sampling points per channel, f is the network model, and Z L is the actual low-frequency model, S is the earthquake record, Z initis the initial model, e is the seismic envelope, w and b represent the weight parameter and bias term of the network respectively, the first term of the loss function is the wave impedance low-frequency model loss term, the second term is the initial model loss term, and λ is the weight used to balance the two losses and control the similarity between the prediction results and the initial model.

[0079] In Example 1, the derivatives corresponding to each step of the input data through the neural network are recorded, and the partial derivatives of the loss function with respect to the weight parameters and bias terms are obtained using the chain rule. Subsequently, the Adam optimizer based on variable step-size gradient descent is used to calculate the update amount of the weight parameters and bias terms. The learning rate is set to 0.05, and the weight parameters and bias terms are updated. This process is repeated until the loss function error is minimized, and finally the network weight parameters and bias terms are saved.

[0080] A low-frequency model construction method based on a spatiotemporal gated recurrent unit fusion network disclosed in the present invention also includes a neural network testing method, comprising the following steps:

[0081] (1) When constructing the low-frequency model of the test data, the envelope of the seismic data is first calculated and the seismic data, the initial model and the seismic envelope are taken as the input to form three-channel data. The predicted data is also standardized.

[0082] (2) Load the trained network weight parameters and bias items to construct a low-frequency model for the test data, extract the variance σ and mean μ of the initial model, and perform denormalization on the predicted low-frequency model. The denormalization is expressed as follows:

[0083] x=σX+μ (9)

[0084] Where X is the normalized data and x is the denormalized data.

[0085] In the second embodiment, Figure 3 The second embodiment shown uses the original data of Marmousi2, where: Figure 3 (a) is the wave impedance, Figure 3 (b) is the synthetic seismic record, Figure 3 (c) is the low frequency model, Figure 4 (d) is the initial model, which has a total of 2266 channels, each with 406 sampling points, and a time sampling rate of 6ms. Figure 3 The dotted lines in the figure are the selected pseudo well data.

[0086] In Example 2, the low-frequency model construction method based on the spatiotemporal gated recurrent unit fusion network of the present invention (ST-GRU) is compared with three neural network models with significant characteristics. The comparison models include CNN-GRU, TCN and UNet networks. CNN-GRU uses GRU as the basic unit and adds an extended convolution layer to extract features of different scales. TCN uses extended convolution as the basic unit. Different expansion rates can obtain different receptive fields. UNet is one of the most classic network frameworks. It gradually increases the receptive field through the encoding-decoding structure of down-up sampling. Figure 5 shows the performance of different network models on the marmousi2 dataset. Figure 5 (a) Figure 5 (c) Figure 5 (e) and Figure 5 (g) is the prediction result of CNN-GRU, TCN, UNet and ST-GRU. Figure 5 (b) Figure 5 (d) Figure 5 (f) and Figure 5 (h) represents the absolute error of CNN-GRU, TCN, UNet and ST-GRU predictions, respectively. Figure 5 The results show that the low-frequency model construction method based on the spatiotemporal gated recurrent unit fusion network (ST-GRU) has good spatial continuity, complete structural expression, inconspicuous background noise, and clearer depiction of the complex structure in the middle.

[0087] The above only describes the present invention and its implementation methods, which is not restrictive. What is shown in the accompanying drawings is only one of the implementation methods of the present invention, and the actual structure is not limited to this. Therefore, if ordinary technicians in this field are inspired by it and do not deviate from the purpose of the invention, they can creatively design structural methods and embodiments similar to the technical solution, which should all fall within the scope of protection of the present invention.

Claims

1. A method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network, characterized in that: The following steps are involved: S1: Carry out data enhancement of wave impedance. On the basis of the original data set, the elastic transformation algorithm is introduced to enhance the data. The original data includes seismic data, m-well logging data constrained interpolation wave impedance model and wave impedance initial model. The initial model is obtained through logging constrained interpolation algorithm, full waveform inversion and seismic velocity. Elastic transformation is used on the wave impedance model and the initial model to obtain the spatial disturbance in the vertical direction. S2: Establish a low-frequency model sample library, where seismic records are characterized by the convolution of seismic wavelets and formation reflection coefficients; S3: Neural network construction and parameter training, building a neural network model including a temporal feature extraction module and a spatiotemporal feature fusion module, and building a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model; S4: Neural network test, calculate the envelope of seismic data and use the seismic data, initial model and seismic envelope to form three-channel data as input. Similarly, the predicted data is standardized and the processed data is used to carry out neural network test.

2. The method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network according to claim 1, characterized in that: The elastic transformation process is mainly divided into three steps. First, the new spatial coordinates are obtained based on the original data coordinate space through affine transformation, and then the new coordinate space is smoothed by Gaussian function to obtain the random displacement field. Finally, the original data is mapped to the new coordinate space to obtain the perturbed data. The elastic transformation process is expressed in the following form: Where D represents the generated random displacement field, T represents the time series, α controls the disturbance degree of the random displacement field, and δ is the Gaussian filter kernel size, which controls the smoothness of the random displacement field.

3. The method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network according to claim 1, characterized in that: Seismic records are represented by the convolution of seismic wavelets and stratum reflection coefficients. The relationship between wave impedance and reflection coefficient is as follows: Among them, t n represents the nth time sampling point, Δt represents the time increment, R represents the reflection coefficient, and Z represents the wave impedance. The convolution model process is expressed as follows: S=R*W+N (3) Among them, S is seismic data, R is reflection coefficient, W is Ricker wavelet, N is noise, and the expression of Ricker wavelet W is as follows: where f m is the main frequency of the seismic wavelet, and t represents time.

4. The method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network according to claim 3, characterized in that: Based on the main frequency of the actual seismic data, five main frequency information is obtained by adding and subtracting at intervals of 5 Hz, and the seismic record is Hilbert transformed using the following formula: Among them, y(t) represents the earthquake record, y H (t) represents the Hilbert transform of the seismic record, τ represents the integral variable, and the synthetic seismic record of the augmented wave impedance model is obtained by combining formula (5) and the convolution model S=R*W+N.

5. The method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network according to claim 4, characterized in that: According to the original data frequency band, a high-pass filter is used to filter out the low-frequency components of the synthetic data to simulate the lack of low-frequency components of seismic data in actual situations. The envelope of the band-limited seismic data still carries the long-wavelength information of the underground medium. The envelope operator is a nonlinear demodulation operator. The envelope of the seismic signal is obtained by calculating the amplitude of the analytical signal. The process can be written as: Where e represents the envelope of the seismic signal.

6. The method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network according to claim 1, characterized in that: A single gated recurrent unit improves a simple recurrent neural network by coupling a reset gate and an update gate. The process can be expressed as follows: Among them, u t 、r t , X t , and h t They represent the update gate vector, reset gate vector, input sample, candidate hidden state and hidden state at the t-th time point respectively. w and b represent the weight parameter and bias term respectively. Sigmoid and tanh represent different activation functions. * represents the dot product operator.

7. The method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network according to claim 6, characterized in that: The bidirectional gated recurrent network unit builds a back propagation layer on the basis of a single gated recurrent unit, taking into account the forward and reverse time dependencies. The temporal feature extraction module consists of multiple parallel three-series bidirectional gated recurrent networks. The number of parallel blocks depends on the value of i in the S4 step, which is set to 2, that is, there are 5 parallel blocks. The spatiotemporal feature fusion module consists of a bidirectional gated recurrent network and a fully connected layer.

8. The method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network according to claim 1, characterized in that: The parameter training steps of the neural network include: initializing the network weight parameter w and the bias term b, inputting the two-dimensional seismic data, the initial model and the seismic envelope in the sample library into the neural network, outputting a one-dimensional predicted low-frequency model, and then constructing a loss function to measure the distance between the predicted low-frequency model and the actual low-frequency model. The loss function is expressed as follows: Where N is the number of sampling points per channel, f is the network model, and Z L is the actual low-frequency model, S is the earthquake record, Z init is the initial model, e is the seismic envelope, w and b represent the weight parameter and bias term of the network respectively, the first term of the loss function is the wave impedance low-frequency model loss term, the second term is the initial model loss term, and λ is the weight used to balance the two losses and control the similarity between the prediction results and the initial model.

9. The method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network according to claim 8, characterized in that: Record the derivatives corresponding to each step of the input data through the neural network, use the chain rule to obtain the partial derivatives of the loss function with respect to the weight parameters and bias terms, then use the Adam optimizer based on variable step size gradient descent to calculate the update amount of the weight parameters and bias terms, set the learning rate to 0.05, update the weight parameters and bias terms, repeat this process until the loss function error is minimized, and finally save the network weight parameters and bias terms.

10. A method for constructing a low-frequency model based on a spatiotemporal gated recurrent unit fusion network for use in any one of claims 1 to 9, further comprising a neural network testing method, characterized in that: The following steps are involved: (1) When constructing the low-frequency model of the test data, the envelope of the seismic data is first calculated and the seismic data, the initial model and the seismic envelope are taken as the input to form three-channel data. The predicted data is also standardized. (2) Load the trained network weight parameters and bias items to construct a low-frequency model for the test data, extract the variance σ and mean μ of the initial model, and perform denormalization on the predicted low-frequency model. The denormalization is expressed as follows: x=σX+μ (9) Where X is the normalized data and x is the denormalized data.

Images