A tunnel full waveform inversion gradient optimization method and system

Abstract

The present invention provides a tunnel full waveform inversion gradient optimization method and system. The present invention is aimed at the design and training strategy of a deep neural network for tunnel full waveform inversion gradient optimization. First, the calculated gradient and the current wave velocity model are used as network input to predict the optimization gradient. After the network prediction result is updated to the current wave velocity model, the loss function is calculated with the real wave velocity model to update the gradient optimization network parameters. Finally, the double-loop training of internal full waveform inversion iteration and outer model update is combined to realize the construction of the gradient optimization operator, effectively improving the full waveform inversion gradient morphology and value.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel advance prediction, and in particular relates to a tunnel full waveform inversion gradient optimization method and system. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In order to detect adverse geological structures ahead of the tunnel in advance, the seismic wave method is often used for advanced geological prediction. It has the advantages of long detection distance and sensitive response to changes in wave impedance. It has been widely used in the construction of many tunnels, effectively ensuring the safe and efficient construction of tunnels.

[0004] In tunnel seismic wave advanced geological prediction, accurate acquisition of the velocity distribution ahead of the tunnel is a key factor influencing technicians' assessment of the integrity of the rock mass ahead and also determines the accuracy of the geological structure imaging ahead. Therefore, inferring the velocity distribution ahead of the tunnel is currently a major research topic in tunnel seismic wave advanced geological prediction. Existing methods use methods such as direct waves to replace the rock mass velocity ahead, velocity analysis, and tomography to obtain velocity models. However, the velocity models obtained by these methods are limited in accuracy, hindering high-precision tunnel seismic imaging. In recent years, with the increasing application of full waveform inversion (FWI) in surface seismic exploration, researchers have also begun to use FWI to obtain the velocity distribution ahead of the tunnel. This technique uses an objective function to calculate the error between simulated and observed data, and then calculates the gradient to update the velocity model, ultimately obtaining the velocity distribution ahead. However, due to the limited observation space within the tunnel, the tunnel seismic detection offset is small, and the acquired data contains less valid information. This leads to arc-shaped gradient distortion in the FWI calculation process. Furthermore, the high-frequency data collected makes it difficult to accurately update the velocity, severely limiting its application in tunnel observations. Therefore, how to optimize the gradient of full waveform inversion is the key to the effective application of full waveform inversion method in tunnel earthquake advance prediction.

[0005] In recent years, with the widespread application of deep learning algorithms, the field of geophysics has also begun to use this algorithm to solve traditional problems. By constructing massive data sets to train the parameters of deep neural networks, a more accurate mapping from input to output can be achieved. For full waveform inversion, deep learning algorithms have also been applied to improve the effect of re-parameterized models, improve the objective function of inversion, and improve the inversion gradient. However, the existing gradient optimization methods based on deep learning are mainly based on supervised training, that is, multiple pairs of calculated gradients and actual gradients are prepared in advance, and the gradients are optimized by training the two. Although this method can alleviate the problem of gradient distortion, it is still difficult to alleviate the more accurate update of the wave velocity values ​​within the interface, and it is easy to introduce cumulative errors in the inversion process, affecting the accuracy of the inversion results. In general, there is currently a lack of deep learning methods that can effectively optimize the gradient of full waveform inversion of tunnel seismic.

[0006] There are still two challenges in implementing full waveform inversion gradient optimization based on deep learning:

[0007] First, existing deep learning-based full-waveform inversion gradient optimization methods often use pre-calculated calculated gradients and expected optimization gradients for network training. These training sets are relatively fixed, making it difficult to cope with errors and changes that occur during the inversion iteration process when embedded in full-waveform inversion, leading to inaccurate inversion results. Therefore, existing gradient optimization network training methods that use expected optimization gradients as labels are difficult to implement effectively.

[0008] Second, existing gradient optimization methods often use only calculated gradients as network input, which contains limited information. This makes it difficult for the gradient optimization network to accurately predict the expected optimization gradient, particularly affecting generalization to other datasets. Therefore, existing deep neural networks that use only calculated gradients as input struggle to achieve good generalization, hindering their widespread application to a wider range of models. Summary of the Invention

[0009] In order to solve the above problems, the present invention proposes a tunnel full waveform inversion gradient optimization method and system, which can realize the construction of gradient optimization operator and effectively improve the tunnel full waveform inversion gradient morphology and value.

[0010] According to some embodiments, the present invention adopts the following technical solutions:

[0011] A tunnel full waveform inversion gradient optimization method comprises the following steps:

[0012] Construct a wave velocity model for tunnel earthquake advance geological prediction, and calculate the corresponding forward modeling data through wave equation forward simulation to obtain the initial model;

[0013] Constructing a full waveform inversion gradient optimization network, the network takes the calculated gradient and the current wave velocity model as input and outputs the predicted gradient;

[0014] Select the full waveform inversion iterative process parameters, start with the initial model to perform full waveform inversion calculations, calculate the error between the observed data and the forward data, obtain the calculation gradient, iteratively update the current velocity model, obtain the updated velocity model, calculate the error between the updated velocity model and the true velocity model using the loss function, perform gradient backpropagation based on this error, and update the gradient optimization network parameters until the iteration conditions are met;

[0015] The updated full waveform inversion gradient optimization network is embedded in the full waveform inversion network. After the calculated gradient is obtained by full waveform inversion, it is used together with the current velocity model as the network input to predict the gradient and update the velocity model. The full waveform inversion gradient iteration is then continued until the iteration conditions are met to obtain the final inversion result.

[0016] As an optional implementation, in constructing a wave velocity model for advanced geological prediction of tunnel earthquakes, the wave velocity model is a plurality of layered media with different wave velocities, and the layer interfaces are simplified to inclined or curved interfaces.

[0017] As an optional implementation, the full waveform inversion gradient optimization network includes two groups of encoders and one group of decoders. The two encoders process the calculated gradient and the current wave velocity model respectively, and then sort out the features obtained by the two to obtain the predicted gradient through a group of encoders.

[0018] As a further implementation, the encoder includes 11 layers of convolutional neural network blocks, each of which includes a convolution layer, a regularization layer, and an activation function layer. By increasing the number of channels and the convolution step size, the input single-channel data is channel expanded and size compressed to obtain data features, and then the features of the two encoders are spliced ​​together.

[0019] As a further implementation, the decoder includes six layers of Conv-block-A and six layers of Conv-block-B, which intersect to form the encoder. Each Conv-block-A includes a convolutional layer, a regularization layer, and an activation function layer, which only increases the channel data of the feature without changing the data size; each Conv-block-B includes an upsampling layer, a convolutional layer, and an activation function layer, which only increases the data size without changing the channel size, ultimately restoring the feature to the gradient size of a single channel.

[0020] As an optional implementation, the loss function for calculating the error between the updated wave velocity model and the true wave velocity model includes the sum of the mean square error and the multi-scale structural similarity.

[0021] As an optional implementation, the iteration condition in the full waveform inversion iteration process is that the number of iterations meets a set value.

[0022] As an optional implementation, when updating the wave velocity model, the calculated predicted gradient is multiplied by a set step size to update the current wave velocity model.

[0023] A tunnel full waveform inversion gradient optimization system, comprising:

[0024] An initial model building module is configured to build a wave velocity model for tunnel earthquake advance geological prediction and calculate corresponding forward modeling data through wave equation forward simulation to obtain an initial model;

[0025] a full waveform inversion gradient optimization network construction module configured to construct a full waveform inversion gradient optimization network, the network taking the calculated gradient and the current wave velocity model as inputs and outputting the predicted gradient;

[0026] The gradient optimization network optimization iteration module is configured to select full waveform inversion iterative process parameters, perform full waveform inversion calculation starting with the initial model, calculate the error between the observation data and the forward data, obtain the calculation gradient, iteratively update the current velocity model, obtain the updated velocity model, calculate the error between the updated velocity model and the true velocity model using the loss function, perform gradient backpropagation based on this error, and update the gradient optimization network parameters until the iteration conditions are met;

[0027] The full waveform inversion iteration module is configured to embed the updated full waveform inversion gradient optimization network into the full waveform inversion network. After the full waveform inversion obtains the calculated gradient, it is used together with the current velocity model as the network input to predict the gradient, update the velocity model, and then continue the full waveform inversion gradient iteration until the iteration conditions are met to obtain the final inversion result.

[0028] An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps in the above method are completed.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] This paper addresses the problems of large gradient numerical errors and morphological distortion during tunnel seismic full waveform inversion calculations by improving the gradient optimization network structure design and training strategy. First, the input of the gradient optimization network is improved to include both the calculated gradient and the current wave velocity model. This provides more effective information, effectively improving the gradient optimization effect and enhancing the generalization of the gradient optimization network.

[0031] To address the problem of difficult training of gradient optimization networks, the present invention improves the fitting target of conventional gradient optimization networks from the gradient itself to the difference between the wave velocity after the gradient is updated and the true model. In this way, the error between the wave velocity models is calculated and the gradient optimization network is updated, making it play a more stable role in the full waveform inversion calculation process. In addition, by restarting the full waveform inversion and re-performing iterative calculations with the initial model, the gradient optimization network is also able to participate in the entire full waveform inversion iteration process, thereby improving the applicability of the gradient optimization network and enabling it to be effectively applied to the gradient optimization in the subsequent full waveform inversion process.

[0032] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0034] Figure 1 : is a diagram of the gradient optimization network structure of this embodiment;

[0035] Figure 2 This is a flow chart of the combination of full waveform inversion and gradient optimization network training in this embodiment;

[0036] Figure 3 This is the flowchart of the full waveform inversion calculation embedded after the gradient optimization network training is completed in this example;

[0037] Figure 4 3 is a comparison diagram of the gradient optimization effects of a set of typical models in this example, wherein (a)-(h) are the actual model, initial model, first inversion iteration gradient, optimization result of the first inversion gradient, first inversion iteration velocity difference, 15th inversion iteration gradient, optimization result of the 15th inversion gradient, and 15th inversion velocity difference respectively set in this example;

[0038] Figure 5 3 is the full waveform inversion result diagram based on the gradient optimization network improvement in this example, where (a)-(c) are the actual model, initial model, and inversion result set in this embodiment, respectively. DETAILED DESCRIPTION

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0041] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0042] Example 1

[0043] A tunnel full waveform inversion gradient optimization method comprises the following steps:

[0044] In step S1, a data set of a tunnel seismic velocity model, earthquake advance prediction data and an initial model is first constructed by using stochastic modeling and forward calculation simulation.

[0045] Considering the complexity of geological changes ahead of the tunnel and the need for advance prediction, this example focuses on processing acoustic wave velocity and data. The velocity model ahead of the tunnel is simplified into multiple layered media with different velocities, and the layer interfaces are simplified into inclined or curved interfaces.

[0046] This example uses an 80-row, 150-column matrix and simulates an 80m*150m velocity model with a grid size of 1m. The middle 10 rows of the first 50 columns are used to simulate the tunnel, with a velocity of 340m / s. The velocity at other locations ranges from 1500m / s to 4500m / s to simulate the velocity of common rock masses in tunnels. Finally, a tunnel velocity model is constructed by setting random interface position, morphology, and other parameters. Ricker wavelets are calculated using a main frequency range of 150Hz to 350Hz. The source is excited at each of the two tunnel side walls and received by six geophones arranged at each side wall. The sampling interval is 0.125ms and the step size is 1600, that is, seismic data is recorded for 0.2s. Ultimately, tunnel earthquake advance prediction data is obtained. Then, the surrounding rock velocity replaces the tunnel front velocity model. Low-pass filtering is used to extract the 0-100Hz frequency band signal of the seismic data. 50 rounds of low-frequency full waveform inversion are performed, and the final result is used as the initial model. Through the above implementation content, a data set for this method was constructed.

[0047] Step S2, such as Figure 1 As shown, a gradient optimized deep neural network is constructed and the network structure is designed according to the gradient optimization requirements;

[0048] In this embodiment, the gradient calculated during the full waveform inversion process and the current wave velocity model are used as inputs to the gradient optimization network to predict the actual gradient, such as Figure 1 As shown. Two encoders with the same structure are used to process the calculated gradient and current wave velocity. The encoder includes a total of 11 convolutional neural network blocks, each of which includes a convolution layer, a regularization layer, and an activation function layer. The input single-channel calculated gradient and the current wave velocity are respectively expanded and compressed in size to obtain two sets of data features. The features of the two encoders are then spliced ​​together and input to the decoder; the decoder includes 6 layers of Conv-block-A and 6 layers of Conv-block-B, which appear alternately to form an encoder. Each Conv-block-A includes a convolution layer, a regularization layer, and an activation function layer, which only increases the channel data of the feature without changing the data size; each Conv-block-B includes an upsampling layer, a convolution layer, and an activation function layer, which only increases the data size without changing the data channel, and finally restores the feature to the gradient size of a single channel;

[0049] Step S3, such as Figure 2 As shown, start gradient optimization network training;

[0050] In this embodiment, two loops, inner and outer loops of full waveform inversion and network total update iteration, are set. First, the data set designed in step S1 is read, the initial iteration number k is set to 0, and the full waveform inversion process is started. The initial model is used as the current velocity model, and the forward simulation seismic data is calculated and the data error is calculated with the real data to obtain the gradient g. k , g k and the current wave velocity model m k As the input of the gradient optimization network, the predicted gradient g' is obtained k , with g' k Multiply by the set step size λ to update the current wave speed model to obtain the next wave speed model m k+1 , calculate m k+1 With the real wave velocity model m true The error is used to update the gradient optimization network parameters; then it is determined whether the number of inversion iterations k is less than the set N1. If so, k is set to k+1 and the above cycle is continued; when k is greater than N1, it is determined whether the number of network iterations is less than the set N2. If so, k is reset to 0 and the full waveform inversion iteration cycle is restarted; if so, the network training is completed and the final network parameters are obtained.

[0051] In this example, the traditional mean square error is used to calculate the data error. The calculation formula is:

[0052]

[0053] in and D are the forward modeling data and observation data, respectively.

[0054] In this example, the network loss function, that is, the model error, is calculated using mean square error and multi-scale structural similarity. The calculation formula is:

[0055]

[0056] in and M true The k-th step optimization gradient updates the velocity model matrix and the true velocity model matrix respectively. R is the local similarity of the two matrices calculated at different scales, and the structural similarity SSIM is r The calculation formula is:

[0057]

[0058] Among them, H and W are the height and width of the model. It represents a window of size r×r with the k point as the center in the M model, and c1 and c2 are constant terms that stabilize the numerator and denominator.

[0059] The main network parameters and hardware conditions in this embodiment are as follows: the calculation is implemented on four graphics cards with 40G memory. The network is built based on the PyTorch platform, the Adam optimizer batch size is 4, and the learning rate of the gradient optimization network parameters is set to 5×10 -4 , the full waveform inversion step size λ is set to 20, the number of full waveform inversion iterations N1 is set to 15, and the number of network outer loops N2 is set to 40.

[0060] Step S4, after completing the training of the gradient optimization network, the network parameters are obtained and embedded into the calculation process of full waveform inversion to test the inversion effect of the gradient optimization network on the test set, such as Figure 3 As shown;

[0061] Similar to the network training process shown in step S3, set k = 0, and use the given initial model as the starting point to perform forward calculation to obtain seismic data. The gradient g is calculated after calculating the data error with the actual collected data. k , g k and the current wave velocity model m k As the input of the trained gradient optimization network, the predicted gradient g' is obtained k , with g' k Multiply by the set step size λ to update the current wave speed model to obtain the next wave speed model m k+1, update k=k+1, and judge whether k is less than 20. If so, continue the process of forward calculation, data error calculation, gradient calculation, and gradient prediction. When it exceeds 20, the prediction result is obtained. During the loop, the calculation gradient, optimization gradient, wave velocity model and other parameters are saved for result analysis, such as Figure 4 As shown in the figure, it can be seen that the optimized gradient is closer to the wave velocity difference.

[0062] Step S5, start the actual application, similar to step S4, and obtain the final inversion result, such as Figure 5 As shown, it can be seen that by optimizing the gradient through the gradient optimization network, a result closer to the actual model can be obtained.

[0063] Example 2

[0064] The tunnel full waveform inversion gradient optimization method based on deep learning includes the following steps:

[0065] A typical geological velocity model for tunnel earthquake advance geological prediction was constructed, and the corresponding forward modeling data was calculated through wave equation forward simulation to obtain the initial model. The model was divided into a training set and a validation set to construct a full waveform inversion gradient optimization dataset.

[0066] Construct a full waveform inversion gradient optimization network, which is a deep neural network structure that takes the calculated gradient and the current wave velocity model as input and predicts the gradient as output;

[0067] Select appropriate full waveform inversion iterative process parameters and perform full waveform inversion calculation starting with the initial model;

[0068] The error between the observed data and the forward modeled data is calculated using the objective function to obtain the calculated gradient. The current velocity model and the calculated gradient are used as the input of the gradient optimization network. The output gradient is predicted. The output gradient is multiplied by the step size to update the current velocity model to obtain the updated velocity model. The error between the updated velocity model and the true velocity model is calculated using the loss function. The gradient is then propagated back using this error to update the gradient optimization network parameters. Multiple full waveform inversion calculation cycles are performed in this way.

[0069] Restart the full waveform inversion and recalculate the full waveform inversion and gradient optimization process using the initial model as the starting point so that the gradient optimization network can participate in the entire iterative process of full waveform inversion;

[0070] After network training is completed, the gradient optimization network is embedded in the full waveform inversion network. After the calculated gradient is obtained by full waveform inversion, it is used together with the current velocity model as the network input to predict the gradient, update the current velocity model, and then continue the full waveform inversion gradient iteration. After multiple iterations, the inversion result is finally obtained.

[0071] In this embodiment, velocity models such as uniform velocity model, migration velocity analysis results, tomography results, and low-frequency full waveform inversion results can be used as initial models in the subsequent full waveform inversion calculation process and used to construct the data set;

[0072] In this embodiment, the gradient optimization network consists of two sets of encoders and one set of decoders. The two encoders process the calculated gradient and the current wave velocity model respectively, and then sort out the features obtained by the two, and obtain the predicted gradient through a set of encoders;

[0073] The encoder consists of 11 layers of convolutional neural network blocks, each of which includes a convolution layer, a regularization layer, and an activation function layer. By increasing the number of channels and the convolution step size, the input single-channel data is expanded and compressed to obtain data features. The features of the two encoders are then concatenated together. The decoder consists of 6 layers of Conv-block-A and 6 layers of Conv-block-B, which appear alternately to form the encoder. Each Conv-block-A includes a convolution layer, a regularization layer, and an activation function layer, which only increases the channel data of the feature without changing the data size. Each Conv-block-B includes an upsampling layer, a convolution layer, and an activation function layer, which only increases the data size without changing the channel size, and ultimately restores the feature to the gradient size of a single channel.

[0074] In this embodiment, the loss function for calculating the error between the updated velocity model and the true velocity model may include the sum of the mean square error and the multi-scale structural similarity, and the calculation formula is:

[0075]

[0076] in and M true are the kth optimization gradient update velocity model matrix and the true velocity model matrix, R is the local similarity of the two matrices calculated at different scales, and the structural similarity SSIM r The calculation formula is:

[0077]

[0078] Among them, H and W are the height and width of the model. It represents a window of size r×r with the k point as the center in the M model, and c1 and c2 are constant terms that stabilize the numerator and denominator.

[0079] In summary, this embodiment designs and trains a deep neural network for gradient optimization of full waveform inversion in tunnels. First, the calculated gradient and the current wave velocity model are used as network inputs to predict the optimized gradient. After the network prediction results are used to update the current wave velocity model, the loss function is calculated with the true wave velocity model to update the gradient optimization network parameters. Finally, a double-loop training of internal full waveform inversion iteration and outer model update is combined to realize the construction of the gradient optimization operator, effectively improving the morphology and value of the full waveform inversion gradient.

[0080] Example 3

[0081] A tunnel full waveform inversion gradient optimization system, comprising:

[0082] An initial model building module is configured to build a wave velocity model for tunnel earthquake advance geological prediction and calculate corresponding forward modeling data through wave equation forward simulation to obtain an initial model;

[0083] a full waveform inversion gradient optimization network construction module configured to construct a full waveform inversion gradient optimization network, the network taking the calculated gradient and the current wave velocity model as inputs and outputting the predicted gradient;

[0084] The gradient optimization network optimization iteration module is configured to select full waveform inversion iterative process parameters, perform full waveform inversion calculation starting with the initial model, calculate the error between the observation data and the forward data, obtain the calculation gradient, iteratively update the current velocity model, obtain the updated velocity model, calculate the error between the updated velocity model and the true velocity model using the loss function, perform gradient backpropagation based on this error, and update the gradient optimization network parameters until the iteration conditions are met;

[0085] The full waveform inversion iteration module is configured to embed the updated full waveform inversion gradient optimization network into the full waveform inversion network. After the full waveform inversion obtains the calculated gradient, it is used together with the current velocity model as the network input to predict the gradient, update the velocity model, and then continue the full waveform inversion gradient iteration until the iteration conditions are met to obtain the final inversion result.

[0086] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0087] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts 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, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0088] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0089] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0090] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made by those skilled in the art that fall within the spirit and principles of the present invention and do not require creative effort are intended to be within the scope of protection of the present invention.

Claims

1. A tunnel full waveform inversion gradient optimization method, characterized by: The following steps are involved: Construct a wave velocity model for tunnel earthquake advance geological prediction, and calculate the corresponding forward modeling data through wave equation forward simulation to obtain the initial model; Constructing a full waveform inversion gradient optimization network, the network takes the calculated gradient and the current wave velocity model as input and outputs the predicted gradient; Select the full waveform inversion iterative process parameters, start with the initial model to perform full waveform inversion calculations, calculate the error between the observed data and the forward data, obtain the calculation gradient, iteratively update the current velocity model, obtain the updated velocity model, calculate the error between the updated velocity model and the true velocity model using the loss function, perform gradient backpropagation based on this error, and update the gradient optimization network parameters until the iteration conditions are met; The updated full waveform inversion gradient optimization network is embedded in the full waveform inversion network. After the calculated gradient is obtained by full waveform inversion, it is used together with the current velocity model as the network input to predict the gradient and update the velocity model. The full waveform inversion gradient iteration is then continued until the iteration conditions are met to obtain the final inversion result.

2. A tunnel full waveform inversion gradient optimization method according to claim 1, characterized in that: In constructing a wave velocity model for tunnel earthquake advance geological prediction, the wave velocity model is a layered medium with multiple wave velocities, and the layer interface is simplified to an inclined or curved interface.

3. The tunnel full waveform inversion gradient optimization method according to claim 1, characterized in that: The full waveform inversion gradient optimization network includes two sets of encoders and one set of decoders. The two encoders process the calculated gradient and the current wave velocity model respectively, and then sort out the features obtained by the two to obtain the predicted gradient through the set of encoders.

4. A tunnel full waveform inversion gradient optimization method as claimed in claim 3, characterized in that: The encoder includes 11 layers of convolutional neural network blocks, each of which includes a convolution layer, a regularization layer, and an activation function layer. By increasing the number of channels and the convolution step size, the input single-channel data is channel expanded and the size is compressed to obtain data features, and then the features of the two encoders are spliced ​​together.

5. The tunnel full waveform inversion gradient optimization method according to claim 3, characterized in that: The decoder includes 6 layers of Conv-block-A and 6 layers of Conv-block-B, which appear alternately to form an encoder. Each Conv-block-A includes a convolution layer, a regularization layer, and an activation function layer, which only increases the channel data of the feature without changing the data size; each Conv-block-B includes an upsampling layer, a convolution layer, and an activation function layer, which only increases the data size without changing the channel size, and finally restores the feature to the gradient size of a single channel.

6. A tunnel full waveform inversion gradient optimization method according to claim 1, characterized in that: The loss function for calculating the error between the updated wave velocity model and the true wave velocity model includes the sum of the mean square error and the multi-scale structural similarity.

7. The tunnel full waveform inversion gradient optimization method according to claim 1, characterized in that: The iteration condition in the full waveform inversion iteration process is that the number of iterations meets the set value.

8. The tunnel full waveform inversion gradient optimization method according to claim 1, characterized in that: When updating the wave velocity model, the calculated predicted gradient is multiplied by the set step size to update the current wave velocity model.

9. A tunnel full waveform inversion gradient optimization system, characterized by: include: An initial model building module is configured to build a wave velocity model for tunnel earthquake advance geological prediction and calculate corresponding forward modeling data through wave equation forward simulation to obtain an initial model; a full waveform inversion gradient optimization network construction module configured to construct a full waveform inversion gradient optimization network, the network taking the calculated gradient and the current wave velocity model as inputs and outputting the predicted gradient; The gradient optimization network optimization iteration module is configured to select full waveform inversion iterative process parameters, perform full waveform inversion calculation starting with the initial model, calculate the error between the observation data and the forward data, obtain the calculation gradient, iteratively update the current velocity model, obtain the updated velocity model, calculate the error between the updated velocity model and the true velocity model using the loss function, perform gradient backpropagation based on this error, and update the gradient optimization network parameters until the iteration conditions are met; The full waveform inversion iteration module is configured to embed the updated full waveform inversion gradient optimization network into the full waveform inversion network. After the full waveform inversion obtains the calculated gradient, it is used together with the current velocity model as the network input to predict the gradient, update the velocity model, and then continue the full waveform inversion gradient iteration until the iteration conditions are met to obtain the final inversion result.

10. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the steps of the method according to any one of claims 1 to 8 are completed.

Images