An acoustic wave impedance inversion system and method

Abstract

This invention belongs to the field of geophysical exploration technology and relates to an acoustic impedance inversion system and method, comprising: a model building module for building an acoustic impedance inversion model, which includes a neural network model and a transfer learning network model, wherein the output of the neural network model is connected to the input of the transfer learning network model, and the output of each convolutional layer of the neural network model is connected to a filter response normalization layer; a model training module for training the acoustic impedance inversion model, using a seismic trace dataset to train the neural network model while keeping the parameters of the trained neural network model unchanged, and using well logging data to train the transfer learning network model; and an output module for performing acoustic impedance inversion on seismic traces of the entire work area based on the trained acoustic impedance inversion model. This eliminates the dependence on batch size in network training, resulting in more stable processing results.

Description

Technical Field

[0001] This invention relates to an acoustic impedance inversion system and method based on an improved fully convolutional residual network and transfer learning, belonging to the field of geophysical exploration technology. Background Technology

[0002] Acoustic impedance inversion is a common technique for predicting stratigraphic structure and estimating rock properties using seismic wave reflection data. Due to the lack of low- and high-frequency components in seismic data, the unknown seismic wavelet, and the influence of various noises, the inverse problem involved in this inversion process is ill-posed. Acoustic impedance inversion is essentially an optimization process with multiple constraints, typically solved using traditional model-based gradient optimization methods, but the selection of constraints is often highly subjective. In recent years, deep learning has flourished. Convolutional Neural Networks (CNNs), as a popular method, have shown excellent performance in various fields. CNNs have received increasing attention due to their excellent ability to extract and learn complex features. Fully Convolutional Neural Networks (FCNs) are a special type of CNN without fully connected layers, capable of accepting input data of arbitrary size. Currently, CNNs and FCNs have been used to solve some geophysical problems. To make CNNs easier to train, some researchers have added skip connections (shoutcuts) to convolutional networks to form residual blocks, preventing network degradation during training and mitigating the gradient vanishing problem caused by increasing depth in deep neural networks. This type of network is called a Residual Network (ResNet). To further enhance the generalization ability of networks and enable them to adapt to other datasets (different from the training set), transfer learning strategies have been widely used.

[0003] Currently, common inversion methods based on convolutional neural networks and transfer learning have the following shortcomings: Most mainstream deep learning models employ Batch Normalization (BN) layers to accelerate training and improve performance; for CNN models, BN has become standard. However, BN layers require calculating intermediate statistics on batches during training, an operation heavily reliant on the batch size. When the batch size is too small, BN operations become inaccurate in providing statistical information across the batch dimension, leading to poor network performance. Conversely, blindly increasing the batch size can easily result in a sharp minimum, reducing network generalization. Therefore, network performance is overly dependent on the batch size parameter.

[0004] Traditional transfer learning methods first determine the structure of the neural network and then train it using a training set. To improve the network's generalization ability, a small amount of labeled data from a new dataset is used as new samples to fine-tune the network parameters. However, for seismic data, well logging data is very scarce and often has poor consistency. Therefore, using a limited number of labels to fine-tune the original network parameters can easily damage the network performance and reduce its generalization ability.

[0005] Traditional neural network optimization algorithms mainly employ stochastic gradient descent (SGD) and adaptive moment estimation (Adam). While SGD can converge to better minima, it often requires manual setting of the learning rate and has a slow convergence speed. Adam-like algorithms, on the other hand, can adaptively calculate the learning rate and converge quickly in the early stages of training, but tend to converge to local minima in the later stages. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide an acoustic impedance inversion system and method based on an improved fully convolutional residual network and transfer learning. This system eliminates the dependence on batch size during network training, and the processing results are more stable than those of traditional transfer learning methods that directly adjust the original network weight parameters during the transfer learning process.

[0007] To achieve the above objectives, the present invention proposes the following technical solution: an acoustic impedance inversion system, comprising: a model building module, a model training module, and an output module; the model building module is used to build an acoustic impedance inversion model, the acoustic impedance inversion model including a neural network model and a transfer learning network model, the output of the neural network model being connected to the input of the transfer learning network model, and the output of each convolutional layer of the neural network model being connected to a filter response normalization layer; the model training module is used to train the acoustic impedance inversion model, using a seismic trace dataset to train the neural network model while keeping the parameters of the trained neural network model unchanged, and using well logging data to train the transfer learning network model; the output module is used to perform acoustic impedance inversion on the seismic traces of the entire work area based on the trained acoustic impedance inversion model.

[0008] Furthermore, the seismic trace dataset includes multiple seismic wavelets with different dominant frequencies and phases, and the formulas for these seismic wavelets are as follows:

[0009]

[0010] Where t is time, f M s(t) is the dominant frequency of the wavelet, and s(t) is the time-domain seismic wavelet.

[0011] Furthermore, the optimal size parameters of the convolution kernel in the neural network model are determined by using a partial sample set from the seismic trace dataset, and the neural network model is trained using the Swats algorithm on the seismic trace dataset.

[0012] Furthermore, the filter response normalization layer consists of a normalization function and a threshold linear unit activation function.

[0013] Furthermore, the normalization formula for the normalization function is:

[0014]

[0015] x represents the input sample data, and ε is a positive constant; It is the standardized sample data;

[0016]

[0017] N = W × H, where W is the width of the sample data and H is the height of the sample data; the affine transformation of the normalized result is given by the following formula:

[0018]

[0019] Where γ and β are parameters to be learned, and y is the sample data. The result after affine transformation.

[0020] Furthermore, the formula for the threshold linear unit activation function is:

[0021] z = max(y, τ)

[0022] Where z is the output of the activation function and τ is the threshold.

[0023] Furthermore, the well logging data is processed into square waves so that the sparsity of the reflection coefficient sequence of the processed wave impedance curve is consistent with the training dataset in the seismic trace data.

[0024] Furthermore, the square wave processing method is as follows: low-pass filtering is applied to the logging data; the inflection point positions of the well wave impedance curve of the logging data are extracted; and the impedance values ​​between each inflection point are modified to the weighted average value of the corresponding interval.

[0025] Furthermore, the optimization objective function of the neural network model and the transfer learning network model is:

[0026]

[0027] Where J(W) is the loss function, and w is the weight parameter of the neural network. It's a neural network, dobs Here, λ is the input seismic data, and λ is the regularization coefficient. It is the square of the L2 norm.

[0028] This invention also discloses an acoustic impedance inversion method, comprising the following steps: establishing an acoustic impedance inversion model, the acoustic impedance inversion model including a neural network model and a transfer learning network model, the output of the neural network model being connected to the input of the transfer learning network model, and the output of each convolutional layer of the neural network model being connected to a filter response normalization layer; training the acoustic impedance inversion model by using a seismic trace dataset to train the neural network model while keeping the parameters of the trained neural network model unchanged, and using well logging data to train the transfer learning network model; and performing acoustic impedance inversion on the seismic traces of the entire work area based on the trained acoustic impedance inversion model.

[0029] This invention, by adopting the above technical solutions, has the following advantages: First, this invention uses a Filter Response Normalization (FRN) layer instead of the Batch Normalization (BN) layer in traditional convolutional networks, so that the training process of the pre-trained network no longer depends on the influence of the batch size parameter, reducing the training difficulty of the network and improving its convergence. Second, traditional transfer learning uses a limited number of samples obtained from well logging data to directly fine-tune the parameters of the pre-trained network. If the consistency of the well logging data is poor, it can easily damage the performance of the pre-trained network. This invention adopts a scheme of sequentially connecting a transfer learning network to the output of the pre-trained network. During the transfer learning training process, only the network parameters of the sequentially connected part are updated, making the prediction performance of the network more stable. Finally, in order to enable the network to converge to a better solution quickly during training, this invention uses the Swats (Switching from Adam to SGD) optimization algorithm to update the network parameters. This algorithm provides the convergence speed of the Adam algorithm in the early stage of network training and the convergence accuracy of the SGD algorithm in the later stage of training, improving prediction accuracy while ensuring training speed. Attached Figure Description

[0030] Figure 1 This is a structural diagram of an acoustic impedance inversion system according to an embodiment of the present invention;

[0031] Figure 2 This is a zero-phase synthetic seismic profile of a seismic trace dataset in one embodiment of the present invention. Figure 2 (a) is a zero-phase synthetic seismic profile with a dominant frequency of 30Hz. Figure 2 (b) is a zero-phase synthetic seismic profile with a dominant frequency of 40Hz. Figure 2 (c) is a zero-phase synthetic seismic profile with a dominant frequency of 50Hz. Figure 2(d) is a zero-phase synthetic seismic profile with a dominant frequency of 60Hz;

[0032] Figure 3 This is a schematic diagram of the acoustic impedance inversion model in one embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of the structure of the filter response normalization layer in one embodiment of the present invention;

[0034] Figure 5 This is a zero-phase synthetic seismic profile of an overthrust model according to one embodiment of the present invention. Figure 5 (a) is a zero-phase synthetic seismic profile with a dominant frequency of 30Hz. Figure 5 (b) is a zero-phase synthetic seismic profile with a dominant frequency of 40Hz. Figure 5 (c) is a zero-phase synthetic seismic profile with a dominant frequency of 50Hz. Figure 5 (d) is a zero-phase synthetic seismic profile with a dominant frequency of 60Hz;

[0035] Figure 6 This is a diagram showing the acoustic impedance inversion result of an overthrust model in one embodiment of the present invention. Figure 6 (a) is the actual acoustic impedance curve for channel 200. Figure 6 (b) is the 200th acoustic impedance curve of each profile obtained by the method of the present invention;

[0036] Figure 7 This is a graph showing the acoustic impedance inversion result of actual data in one embodiment of the present invention. Figure 7 (a) is a profile of the original seismic record. Figure 7 (b) is a wave impedance profile obtained by the method in this invention;

[0037] Figure 8 This is a flowchart of an acoustic impedance inversion method in one embodiment of the present invention. Detailed Implementation

[0038] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention is described in detail through specific embodiments. However, it should be understood that the specific embodiments are provided only for a better understanding of the present invention and should not be construed as limiting the present invention. In the description of the present invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0039] To address the challenges of training high network complexity caused by the batch normalization layer's limitation by the batch size parameter in existing technologies, this invention proposes an acoustic impedance inversion system and method based on an improved fully convolutional residual network and transfer learning. It replaces the traditional batch normalization layer (BN) in convolutional networks with a filter response normalization layer (FRN), freeing the pre-trained network from the influence of the batch size parameter and reducing training difficulty. Furthermore, traditional transfer learning uses a new small sample set to further update the original network parameters, which can easily degrade the performance of the pre-trained network. This invention employs a method of sequentially connecting a transfer learning network to the output of the pre-trained network, updating only the additional network weights during transfer learning training, thus making the network's prediction performance more stable. Finally, to enable the network to quickly converge to a better solution during training, this invention uses the Swats optimization algorithm, which can improve prediction accuracy while maintaining training speed. The following detailed description, in conjunction with the accompanying drawings and embodiments, illustrates the method of this invention in detail.

[0040] Example 1

[0041] This embodiment discloses an acoustic impedance inversion system, such as Figure 1 As shown, it includes: a model building module, a model training module, and an output module;

[0042] The model building module is used to build an acoustic impedance inversion model, which includes a neural network model and a transfer learning network model. The output of the neural network model is connected to the input of the transfer learning network model. The output of each convolutional layer of the neural network model is connected to the filter response normalization layer (FRN) to obtain an improved fully convolutional residual network. The transfer learning network model is mainly used to correct the low frequency of the output results of the pre-trained network.

[0043] The model training module is used to train the acoustic impedance inversion model. It uses seismic trace datasets to train the neural network model, keeps the parameters of the trained neural network model unchanged, and uses well logging data to train the transfer learning network model.

[0044] The output module is used to perform acoustic impedance inversion on the seismic traces of the entire work area based on the trained acoustic impedance inversion model.

[0045] This embodiment effectively overcomes the shortcomings of traditional methods. It replaces the batch normalization layer (BN) in the traditional fully convolutional neural network with a more advanced filter response normalization layer (FRN) to eliminate the dependence on batch size during network training. During transfer learning, the original pre-trained network structure and parameters are kept unchanged, followed by a simple convolutional network. During transfer learning training, only the weight parameters of the additional network are adjusted, which is equivalent to a secondary fine-tuning of the pre-trained network's output. This approach yields more stable results than traditional transfer learning methods that directly adjust the original network weight parameters. Finally, to ensure training speed and improve prediction accuracy, this embodiment employs the Swats optimization algorithm, providing the convergence speed of the Adam algorithm in the early stages of training and the convergence accuracy of the SGD algorithm in the later stages.

[0046] like Figure 2 As shown, this embodiment uses the Marmousi2 model dataset from seismic exploration as its foundation. The seismic trace dataset includes multiple seismic wavelets with different dominant frequencies and phases, serving as the network's pre-training dataset. The Marmousi2 model dataset is a Marmousi2 acoustic impedance model. The impedance model after thinning the Marmousi2 acoustic impedance model contains 1700 traces, each with 700 sampling points, and a time sampling rate of 2ms. To generate more training samples, this embodiment generates 16 Ricker wavelets with different dominant frequencies and phases. Figure 2 (a) is a zero-phase synthetic seismic profile with a dominant frequency of 30Hz. Figure 2 (b) is a zero-phase synthetic seismic profile with a dominant frequency of 40Hz. Figure 2 (c) is a zero-phase synthetic seismic profile with a dominant frequency of 50Hz. Figure 2 (d) is a zero-phase synthetic seismic profile with a dominant frequency of 60Hz; Figure 2 In (a), the phase is -30°. Figure 2 In (b), the phase is -10°. Figure 2 In (c), the phase is 0°. Figure 2In (d), the phase is 10°. By convolving the reflection coefficients of each trace in the Marmousi2 acoustic impedance model, 16 seismic profiles with a total of 27,200 traces are obtained. In this embodiment, the first 30% of each profile, totaling 8,160 traces, is used as the test set. Of the remaining 19,040 traces, 2,000 are randomly selected as the validation set, and the remaining 17,040 traces are used as the training set, with a batch size of 40. Since the sample data contains various wavelet morphologies, when the network converges, it already possesses a preliminary ability to predict relative wave impedance, which is absolute wave impedance for the Marmousi2 acoustic impedance model. When predicting other datasets, only a small amount of labeled data is needed for transfer learning to adjust the low-frequency output impedance of the network to obtain the absolute wave impedance.

[0047] The formula for the seismic wavelet is as follows:

[0048]

[0049] Where t is time, f M s(t) is the dominant frequency of the wavelet, and s(t) is the time-domain seismic wavelet. After obtaining zero-phase wavelets with different dominant frequencies, their phases are rotated:

[0050]

[0051] in, It is a seismic wavelet after phase rotation. It is the phase rotation angle, imag[·] is the imaginary part of the complex number, and hilbert(·) is the Hilbert transform.

[0052] Experiments were conducted using a small subset of samples from the seismic trace dataset to determine the optimal kernel size parameters for the neural network model. The improved fully convolutional parametric network was trained using the Swats algorithm on the seismic trace dataset.

[0053] The Swats algorithm is a combination of the classic Adam optimization algorithm and the SGD algorithm. In the early stage of network training, this algorithm provides the convergence speed of the Adam algorithm and the convergence accuracy of the SGD algorithm in the later stage of training.

[0054] The well logging data was processed into square waves, and the experimental parameters were repeatedly adjusted to ensure that the sparsity of the reflection coefficient sequence of the processed wave impedance curve was consistent with the centralized training dataset of the seismic trace data. The square wave processing method was as follows: low-pass filtering was applied to the well logging data; the inflection point positions of the well wave impedance curve of the well logging data were extracted; and the impedance values ​​between each inflection point were modified to the weighted average of the corresponding intervals.

[0055] like Figure 3As shown, the input layer of the improved fully convolutional residual network is connected to the first convolutional layer, which consists of 16 convolutional kernels of size 300×1. This is followed by three residual blocks, each containing two convolutional layers: the first layer consists of 16 convolutional kernels of size 300×1, and the second layer consists of 16 convolutional kernels of size 3×1. These three residual blocks form a residual network, which is then connected to another convolutional network with 3×1 kernels. Since the improved fully convolutional residual network is entirely composed of convolutional layers, it has no limitation on the length of the input data and can accept seismic data of varying lengths. The shortcut connections in the three residual blocks effectively prevent network degradation and alleviate the gradient vanishing problem during the training process of deep networks. The stride of the convolutional layers is set to 1, and zero-padding is used for each convolutional layer. After each convolutional layer, there is a Filter Response Normalization (FRN) layer. Figure 4 As shown, the Filter Response Normalization (FRN) layer consists of a normalization function and a Threshold Linear Unit (TLU) activation function. The addition of the FRN layer eliminates the dependence on batch size parameters during training, accelerating network convergence. The kernel size parameters in each convolutional layer were selected based on extensive experimental results.

[0056] The normalization formula for the normalization function is:

[0057]

[0058] x represents the input sample data, and ε is a very small positive constant to prevent the denominator from being zero; It is the standardized sample data;

[0059]

[0060] N = W × H, where W is the width of the sample data and H is the height of the sample data. An affine transformation is performed on the normalized result to prevent the network from being affected by denormalization. The formula for the affine transformation is:

[0061]

[0062] Where γ and β are parameters to be learned, and y is the sample data. The result after affine transformation.

[0063] The formula for the threshold linear unit activation function is:

[0064] z = max(y, τ)

[0065] Where z is the output of the activation function and τ is the threshold.

[0066] The transfer learning network model consists of four simple convolutional layers connected in series. The first and third convolutional layers each consist of 16 convolutional kernels of size 300×1, the second convolutional layer consists of 16 convolutional kernels of size 3×1, and the last convolutional layer consists of a single convolutional kernel of size 3×1.

[0067] The optimization objective function for neural network models and transfer learning network models is:

[0068]

[0069] Where J(W) is the loss function of the network, and w is the weight parameter of the neural network. It's a neural network, d obs Here, λ is the input seismic data, and λ is the regularization coefficient. It is the square of the L2 norm.

[0070] Figure 5 This is a zero-phase synthetic seismic profile of the overthrust model in this embodiment. The overthrust model is used as a new dataset to test the method proposed in this invention. The model contains 400 traces, each with 700 points, and a sampling rate of 2ms. Figure 5 (a) is a zero-phase synthetic seismic profile with a dominant frequency of 30Hz. Figure 5 (b) is a zero-phase synthetic seismic profile with a dominant frequency of 40Hz. Figure 5 (c) is a zero-phase synthetic seismic profile with a dominant frequency of 50Hz. Figure 5 (d) is a zero-phase synthetic seismic profile with a dominant frequency of 60Hz. The seismic profile is obtained by convolving Ricker wavelets with dominant frequencies of 30Hz, 40Hz, 50Hz, and 60Hz with the reflection coefficient profile of the thrust model. One trace is extracted every 70 traces from each seismic profile, for a total of 20 traces, as a pseudo-well dataset for transfer learning training of the network.

[0071] Figure 6 This is a diagram showing the acoustic impedance inversion results of the overthrust model in this embodiment. Figure 6 (a) is the actual acoustic impedance curve for channel 200. Figure 6 (b) shows the 200th acoustic impedance curve of each profile obtained by the method of this invention; the predicted results of each channel are in high agreement with the actual impedance curves, and the correlation coefficients are all greater than 0.92. This test result verifies the effectiveness of the method of this invention.

[0072] Figure 7 This is a graph showing the acoustic impedance inversion results of actual data in this embodiment. Figure 7(a) is the original seismic record profile, showing a two-dimensional seismic survey line with 1001 traces, each with 551 sampling points and a sampling rate of 2ms. First, keeping the parameters of the pre-trained fully convolutional residual network unchanged, the parameters of the additional transfer learning network are updated using well logging data as labels. After complete convergence, all seismic traces within the survey line are input into the network for prediction. Figure 7 (b) is a wave impedance profile obtained by the method of the present invention. Figure 7 (b) The white box represents the embedded filtered logging impedance curve. The results show that the predicted impedance of the present invention matches the logging curve well, proving the practicality of the method of the present invention.

[0073] Example 2

[0074] Based on the same inventive concept, this embodiment discloses an acoustic impedance inversion method, such as... Figure 8 As shown, it includes the following steps:

[0075] An acoustic impedance inversion model is established, which includes a neural network model and a transfer learning network model. The output of the neural network model is connected to the input of the transfer learning network model, and the output of each convolutional layer of the neural network model is connected to the filter response normalization layer.

[0076] The acoustic impedance inversion model was trained using a seismic trace dataset to train the neural network model. The parameters of the trained neural network model were kept constant, and well logging data was used to train the transfer learning network model.

[0077] Based on the trained acoustic impedance inversion model, acoustic impedance inversion is performed on the seismic traces of the entire work area.

[0078] This embodiment uses a WeChat mini program as an example for illustration. However, this embodiment only illustrates one method of creating a computer program, but it is not limited to this. In addition to being set on a mobile phone, the computer program can also be set on any kind of mobile communication device, such as a laptop, tablet or smartwatch.

[0079] This computer program first requires registering a mini-program on the official WeChat platform and downloading the WeChat Developer Tools. A mini-program project is created, with each sub-page consisting of JS (page logic), JSON (page configuration), WXML (common styles), and WXSS (page styles) files, developed in JavaScript. The backend API is written in Java, primarily using the Spring Boot framework.

[0080] The mini-program's functional logic is based on Examples 1 to 3. After the program starts, it is set to periodically store time and location information in the corresponding local cache according to a set time, and to clear expired cache data older than 14 days daily. Simultaneously, it automatically retrieves daily official information, compares local travel records with officially released risk information, and issues risk pop-up warnings for cases of overlapping travel times and locations. This aims to achieve the goal of automatically and accurately recording travel itineraries and conducting epidemiological investigations.

[0081] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific embodiments of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention. The above content is only a specific embodiment of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. An acoustic impedance inversion system, characterized in that, include: The model building module, model training module, and output module are all included. The model building module is used to build an acoustic impedance inversion model, which includes a neural network model and a transfer learning network model. The output of the neural network model is connected to the input of the transfer learning network model, and the output of each convolutional layer of the neural network model is connected to a filter response normalization layer. The model training module is used to train the acoustic impedance inversion model, train the neural network model using seismic trace dataset, keep the parameters of the trained neural network model unchanged, and train the transfer learning network model using well logging data. The output module is used to perform acoustic impedance inversion on the seismic traces of the entire work area based on the trained acoustic impedance inversion model. Each convolutional layer output of the neural network model is connected to a Filter Response Normalization (FRN) layer to obtain an improved fully convolutional residual network. The input layer of the improved fully convolutional residual network is connected to a first convolutional layer consisting of 16 convolutional kernels of size 300×1. After that, three residual blocks are connected, each containing two convolutional layers. The first convolutional layer consists of 16 convolutional kernels of size 300×1, and the second convolutional kernel consists of 16 convolutional kernels of size 3×1. The three residual blocks form a residual network, which is then connected to a convolutional network with 3×1 kernels. The stride of the convolutional layers is set to 1, and zero-padding is used for each convolutional layer.

2. The acoustic impedance inversion system as described in claim 1, characterized in that, The seismic trace dataset includes multiple seismic wavelets with different dominant frequencies and phases. The formulas for these seismic wavelets are as follows: ； Where t is time, f M It is the dominant frequency of the wavelet. It is a time-domain seismic wavelet.

3. The acoustic impedance inversion system as described in claim 2, characterized in that, The optimal size parameters of the convolution kernel in the neural network model are determined by using a partial sample set from the seismic trace dataset, and the neural network model is trained using the Swats algorithm on the seismic trace dataset.

4. The acoustic impedance inversion system as described in claim 1, characterized in that, The filter response normalization layer consists of a normalization function and a threshold linear unit activation function.

5. The acoustic impedance inversion system as described in claim 4, characterized in that, The normalization formula for the normalization function is: ； x represents the input sample data. It is a positive integer; It is the standardized sample data; ； N = W × H, where W is the width of the sample data and H is the height of the sample data; the affine transformation of the normalized result is calculated using the following formula: ； in, γ and β All of these are parameters that need to be learned, and y is the sample data. The result after affine transformation.

6. The acoustic impedance inversion system as described in claim 5, characterized in that, The formula for the threshold linear unit activation function is: ； Where z is the output of the activation function. It is a threshold.

7. The acoustic impedance inversion system according to any one of claims 1-6, characterized in that, The well logging data is processed into square waves so that the sparsity of the reflection coefficient sequence of the processed wave impedance curve is consistent with the training dataset in the seismic trace data.

8. The acoustic impedance inversion system as described in claim 7, characterized in that, The method for square wave conversion is as follows: The well logging data is low-pass filtered; Extract the inflection point location of the well wave impedance curve of the well logging data; The impedance values ​​between each inflection point are modified to the weighted average of the corresponding intervals.

9. The acoustic impedance inversion system according to any one of claims 1-6, characterized in that, The optimization objective function for the neural network model and the transfer learning network model is: ； in, is the loss function of the network, and w is the weight parameter of the neural network. It's a neural network. It is the input earthquake data. It is the regularization coefficient. It is the square of the L2 norm.

10. A method for acoustic impedance inversion, characterized in that, The acoustic impedance inversion system as described in any one of claims 1-9 includes the following steps: An acoustic impedance inversion model is established, which includes a neural network model and a transfer learning network model. The output of the neural network model is connected to the input of the transfer learning network model, and the output of each convolutional layer of the neural network model is connected to a filter response normalization layer. The acoustic impedance inversion model is trained using a seismic trace dataset to train the neural network model. The parameters of the trained neural network model are kept constant, and well logging data is used to train the transfer learning network model. Based on the trained acoustic impedance inversion model, acoustic impedance inversion is performed on the seismic traces of the entire work area.

Images