Tunnel joint inversion method and system based on implicit neural network and geological constraint

Abstract

The invention belongs to the field of geophysical exploration, and provides a tunnel joint inversion method and system based on an implicit neural network and geological constraints, and the method comprises the steps: constructing the implicit neural network, taking a three-dimensional space coordinate as an input, and synchronously outputting the normalized model parameters of a wave velocity, resistivity and a dielectric constant, pre-training the network by using prior physical property data such as advanced drilling; calculating data loss of each method by combining earthquake, electrical method and ground penetrating radar forward modeling; a cross gradient loss item is introduced to strengthen structural consistency, and a priori physical property constraint item is fused into geological information; and optimizing the total loss function through the balance weight, and iteratively updating network parameters to obtain high-precision physical property parameter distribution. The problem of inversion nonlinear enhancement caused by a narrow observation environment of a tunnel is effectively relieved, physical property parameter collaborative inversion of multi-resolution data is realized, geological prior physical property constraints and structural constraints are introduced, the imaging precision and the geological interpretation reliability are remarkably improved, and the method is suitable for large-scale popularization and application. The problems that according to an existing scheme, different geophysical methods are difficult to effectively utilize, resolution differences exist, and inversion of physical property parameters is inaccurate are solved.

Description

Technical Field

[0001] This invention belongs to the field of geophysical exploration and tunnel engineering, specifically relating to a tunnel joint inversion method and system based on implicit neural networks and geological constraints. Background Technology

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

[0003] To effectively detect adverse geological structures ahead of tunnels and reveal the structure and physical property distribution of the underground medium, providing a basis for subsequent construction, various geophysical methods such as seismic wave methods, resistivity methods, and ground-penetrating radar methods are commonly used for tunnel advance detection. Existing methods mostly rely on a single geophysical method to invert and obtain information on a specific geological attribute, resulting in a limited perspective. Furthermore, under the complex engineering conditions of tunnels, numerous interference factors and significant environmental noise significantly reduce the quality of the inversion imaging, and each method has its limitations. For example, the tunnel seismic wave method suffers from arc-shaped distortion in the full waveform inversion gradient due to the small detection offset and limited effective information in the acquired data; the tunnel resistivity method, due to the electric field volume effect and small data volume, struggles to accurately characterize interface information in the inversion imaging; and while the tunnel ground-penetrating radar method produces a two-dimensional profile with high imaging accuracy, in actual detection, it is difficult to obtain parameter information and spatial morphology of geological bodies based solely on the profile.

[0004] Joint inversion, by combining various geophysical observation data and utilizing complementary information from different geophysical methods, integrates the advantages of multiple geophysical approaches. This effectively suppresses the ambiguity of inversion and yields more accurate information about the subsurface medium. Joint inversion mainly falls into two categories: structural constraint-based and property constraint-based. The former includes constraints such as cross-gradient constraints and structural consistency constraints, while the latter primarily uses empirical formulas to couple rock-physical relationships between different property parameters. For geophysical inversion problems, especially in tunnel engineering, methods such as advanced drilling and face sketching can provide prior constraints for wave velocity and resistivity inversion, further improving the imaging effect of joint inversion on the subsurface medium.

[0005] Existing multi-source data joint inversion methods in the field of tunnel engineering still have the following problems: Existing multi-source data tunnel joint inversion methods are limited by the narrow observation space of tunnels. The data collected by tunnel advance detection methods contains less effective information and more noise, resulting in stronger inversion nonlinearity. The inversion is prone to getting trapped in local extrema and it is difficult to accurately obtain the distribution of wave velocity, resistivity and dielectric constant in front of the tunnel. Although the joint inversion method can alleviate this inversion nonlinearity problem by fusing multiple types of detection data, it is still difficult to solve the key problem of inaccurate tunnel inversion and further research is needed.

[0006] Existing multi-source data tunneling joint inversion methods typically assume that the different geophysical methods involved in the joint inversion have the same or similar resolution characteristics. However, in practical applications, the resolution differences between methods such as seismic wave methods and ground-penetrating radar are significant. Existing methods often use the method with the lowest resolution as the benchmark for joint inversion, resulting in the underutilization of fine feature information in high-resolution methods, which severely limits the accuracy of inversion imaging and the reliability of geological interpretation.

[0007] Existing multi-source data tunneling joint inversion methods are mostly based on physical property constraints or structural constraints. Physical property constraints usually rely on empirical rock physics relationships, which often have limited applicability under complex geological conditions. While structural constraint methods can improve the consistency of geometric morphology, they are difficult to effectively characterize the complex nonlinear relationships between different physical property parameters. The limitations of relying on only one constraint method result in insufficient adaptability of the inversion results in complex geological environments, making it difficult to accurately reflect the true geological conditions. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a tunnel joint inversion method and system based on implicit neural networks and geological constraints. This invention can achieve joint inversion imaging using multiple geophysical methods, characterize multi-parameter physical property models with the help of implicit neural networks, and effectively improve imaging quality by utilizing high-resolution information and the inherent rock physical relationships within the geological body.

[0009] According to some embodiments, the present invention adopts the following technical solution: A tunnel joint inversion method based on implicit neural networks and geological constraints includes the following steps: An implicit neural network is constructed, which takes the three-dimensional spatial coordinate vector of the inversion space of the probe target as input and the normalized model parameters of wave velocity, resistivity and dielectric constant at the corresponding position as output. By traversing all coordinates of the probe target and predicting the model parameters at the corresponding position, a three-dimensional multi-parameter model of the probe target area can be constructed. Prior physical property parameters are obtained, and sample pairs of spatial coordinates and physical property parameters are constructed. These samples are used as the training set to perform supervised pre-training on a pre-constructed implicit neural network. The mean square error between the normalized model parameters output by the network and the normalized values ​​of the physical property parameters in the sample pairs is used as the target loss function for the implicit neural network pre-training. After pre-training, the weights and bias terms of some structures in the implicit neural network are fixed, while the relationships between the physical property parameters are preserved. As an alternative implementation, the implicit neural network pre-training objective loss function is:

[0010] in, Indicates the number of data categories; These are weighting coefficients used to balance the magnitude differences between different data points; These are partial model parameter values ​​corresponding to prior physical property parameters, output by the implicit neural network and after inverse normalization. These are partial point physical property values ​​obtained through advanced drilling and core testing.

[0011] Based on the normalized model parameters output by the implicit neural network, after inverse normalization processing, forward modeling is performed for seismic, electrical resistivity tomography, and ground penetrating radar respectively to obtain the corresponding prediction data. The loss between the observed data and the prediction data is calculated, and a data loss term is constructed. Calculate the pairwise cross gradients of the three parameter pairs: wave velocity-resistivity, wave velocity-dielectric constant, and resistivity-dielectric constant, and construct the cross gradient loss term accordingly. Based on the rock strata information revealed by advanced drilling and the test data of core samples at corresponding locations, key physical property parameters of some exploration areas are obtained, and prior physical property constraints are constructed. Different weights are set for different losses to balance the magnitude of the data of each loss term. After fusion, the final target loss function is obtained. Backpropagation of the loss is performed to update the weights and bias terms of the implicit neural network. The above process is repeated until the iteration condition is met. By utilizing the updated implicit neural network, the model parameter distribution is obtained, enabling imaging of the wave velocity and resistivity properties of underground media.

[0012] As an alternative implementation, the implicit neural network adopts a fully connected neural network with four hidden layers. The first and fourth hidden layers contain 128 neurons, and the second and third hidden layers contain 256 neurons. Each layer uses the sine activation function to enhance nonlinear expression and prevent gradient vanishing during backpropagation. At the same time, a random deactivation (Dropout) mechanism is added after the hidden layers to improve the network's generalization. Batch normalization is used between layers to accelerate training convergence and stabilize the numerical distribution.

[0013] As an alternative implementation, the implicit neural network further includes an input layer and an output layer. The input layer contains three neurons to receive spatial coordinates x, y, and z, and the output layer contains three neurons to output normalized wave velocity model parameters, resistivity model parameters, and dielectric constant model parameters.

[0014] As an alternative implementation method, the process of performing forward modeling for seismic, electrical resistivity, and ground-penetrating radar (GPR) based on the normalized model parameters output by the implicit neural network, after inverse normalization processing, includes: performing inverse normalization processing on the normalized model parameters obtained by pre-training the implicit neural network to obtain the initial models for forward modeling of seismic, electrical resistivity, and GPR; constructing a certain spatial area in front of the tunnel face as the joint inversion area for seismic, electrical resistivity, and GPR; selecting the joint inversion area according to the effective range of electrical resistivity at the longitudinal scale, and using the same resistivity model, wave velocity model, and dielectric constant model at the transverse scale.

[0015] As an alternative implementation method, the process of setting corresponding weights for different losses to balance the magnitude of the data of each loss term includes: selecting the data loss term as the benchmark, setting the weights of the cross gradient structure constraint term and the prior property constraint term, so that the three losses are of the same or similar magnitude, and balancing the problem that the impact of each loss term on the total loss function due to data differences is greater than the set value.

[0016] As an alternative implementation, the expression for the data loss term is as follows:

[0017]

[0018]

[0019]

[0020] in, Indicates data loss items; The weights for the data loss items of electrical resistivity tomography (EDT) and ground penetrating radar (GPR) are respectively set to ensure that the magnitudes of the three data loss items are the same or similar. , , These are data loss items for seismic, electrical resistivity, and ground-penetrating radar methods, respectively. , , These represent the number of observation data obtained from seismic, electrical resistivity, and ground-penetrating radar, respectively. , , These are the predicted observation data obtained through seismic, electrical resistivity, and ground-penetrating radar forward modeling, respectively, after the network output parameters have been inversely normalized. These are actual observation data from seismic, electrical resistivity tomography, and ground-penetrating radar, respectively.

[0021] As an alternative implementation, the expression for the cross-gradient loss term is as follows:

[0022] For 3D cross gradients:

[0023]

[0024]

[0025] For two-dimensional cross gradients:

[0026]

[0027]

[0028] in, This represents the cross-gradient structure loss term; Indicates the number of data categories; These are weighting coefficients used to balance the magnitude of the cross-gradient loss across the three parameter pairs; The cross gradient loss for a given data pair; and For the gradient of the model parameters of the corresponding data pair, These are the model parameter values ​​for the corresponding data pairs; Representing the current grid and Forward grid in the direction; The current grid is in Side length in the direction, They are respectively The side length of the forward grid in the direction.

[0029] As an alternative implementation, the prior property constraint loss term is:

[0030] in, This represents the prior data loss term; Indicates the number of data categories; These are weighting coefficients used to balance the magnitude differences between different data points; These are partial model parameter values ​​corresponding to prior physical property parameters, output by the implicit neural network and after inverse normalization. These are partial point physical property values ​​obtained through advanced drilling and core testing.

[0031] A tunnel joint inversion system based on implicit neural networks and geological constraints includes: The implicit neural network construction module is configured to construct an implicit neural network, which takes the three-dimensional coordinate vector of the target inversion space as input and the normalized model parameters of wave velocity, resistivity and dielectric constant at the corresponding position as output. The implicit neural network pre-training module is configured to obtain prior physical property parameters through advanced drilling, etc., construct sample pairs of spatial coordinates and physical property parameters, and use them as training sets to perform supervised pre-training on the pre-constructed implicit neural network. The mean square error between the normalized model parameters output by the network and the normalized values ​​of the physical property parameters in the sample pairs is used as the target loss function for the implicit neural network pre-training. After the pre-training is completed, the weights and bias terms of at least some structures in the implicit neural network are fixed and the relationship between physical property parameters is preserved. The forward modeling module is configured to use normalized model parameters based on the output of an implicit neural network. After inverse normalization, forward models are performed for seismic, electrical resistivity, and ground-penetrating radar respectively to obtain the corresponding prediction data. The data loss calculation module is configured to calculate the loss between observed data and predicted data, and to construct a data loss term. The cross-gradient loss calculation module is configured to calculate the pairwise cross-gradients of three parameter pairs: wave velocity-resistivity, wave velocity-dielectric constant, and resistivity-dielectric constant, in order to construct the cross-gradient loss term. The prior physical property loss calculation module is configured to obtain the spatial distribution characteristics of key physical property parameters and construct prior physical property constraints based on the rock strata information revealed by the advance drilling and the core sample test data at the corresponding locations. The objective function loss fusion module is configured to set corresponding weights for different losses to balance the magnitude of the data of each loss term. After fusion, the final objective loss function is obtained. Backpropagation of the loss is performed to update the weights and bias terms of the implicit neural network. The above process is repeated until the iteration condition is met. The implicit neural network iterative module is configured to use the updated implicit neural network to obtain the model parameter distribution, thereby achieving accurate imaging of the wave velocity, resistivity, and dielectric constant properties of the underground medium.

[0032] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement instructions; the computer-readable storage medium being configured to store a plurality of instructions adapted to be loaded by the processor and executed in accordance with the steps of the method described therein.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) In view of the problem that the existing tunnel joint inversion method is more nonlinear due to the narrow space of the tunnel, the present invention adopts the form of implicit neural network and takes advantage of the fact that the solution space of implicit equation can cover complex nonlinear mapping. This mechanism alleviates the problem of enhanced nonlinearity caused by tunnel space limitation. At the same time, the implicit equation tends to the global optimal solution, which can alleviate the problem that the inversion is prone to local minima, thereby improving the inversion effect of tunnel medium physical parameters.

[0034] (2) To address the problem that existing tunnel joint inversion methods cannot effectively utilize the resolution differences of different geophysical methods, this invention establishes a continuous mapping relationship between spatial coordinates and various physical parameters by constructing a unified implicit neural network. This network structure can simultaneously learn multi-scale features provided by different resolution methods, retaining the fine characterization capability of high-resolution methods while integrating the macroscopic background information of low-resolution methods, thus achieving complementary advantages of geophysical data at different resolutions and improving the accuracy of inversion imaging.

[0035] (3) To address the problem of insufficient adaptability of existing constraint methods under complex geological conditions, this invention combines cross-gradient structural constraints with prior physical property constraints to construct a multi-constraint collaborative joint inversion framework. On the one hand, cross-gradient constraints ensure the structural consistency of different physical property models; on the other hand, prior physical property constraints are established using measured data revealed by boreholes, effectively overcoming the limitations of a single constraint method. Simultaneously, the network is pre-trained with prior geological information at the network end, enabling the network to implicitly learn the relationship between different physical property parameters, thus making the inversion results more adaptable and reliable in complex geological environments.

[0036] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0037] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0038] Figure 1 This is a technical roadmap for the tunnel joint inversion method based on implicit neural networks and geological constraints in this embodiment of the invention. Figure 2 This is a schematic diagram of an implicit neural network according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the use of prior information such as advanced drilling for network pre-training in an embodiment of the present invention. Detailed Implementation

[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 description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

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

[0042] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0043] Example 1 In one or more embodiments, a tunnel joint inversion method based on implicit neural networks and geological constraints is disclosed, such as... Figure 1 As shown, the specific process includes the following: (1) Constructing an implicit neural network In this embodiment, the network adopts a fully connected network with four hidden layers, such as... Figure 2 As shown, a fully connected neural network with four hidden layers is used. The first and fourth hidden layers contain 128 neurons, and the second and third hidden layers contain 256 neurons. Each layer uses the sine activation function to enhance nonlinear expressiveness and prevent gradient vanishing during backpropagation. A dropout mechanism is added after each hidden layer to improve network generalization. Batch normalization is used between layers to accelerate training convergence and stabilize numerical distribution. The input layer contains three neurons to accept spatial coordinates x, y, and z, and the output layer contains three neurons, outputting normalized wave velocity model parameters, resistivity model parameters, and dielectric constant model parameters. Prior geological property information obtained through advanced drilling, pilot tunneling, and face excavation is used to construct spatial coordinate-property parameter sample pairs, such as... Figure 3 As shown, the implicit neural network is pre-trained. The mean square error between the normalized model parameters output by the network and the normalized values ​​of the physical property parameters in the sample pair is used as the target loss function for the implicit neural network pre-training. After the pre-training is completed, some neurons in some layers are fixed so that their weights and bias terms remain unchanged, and the relationship between the physical property parameters is preserved.

[0044] (2) Data denormalization and model building The normalized model parameters obtained from the implicit neural network pre-training are denormalized to obtain the initial models for seismic, electrical resistivity, and ground-penetrating radar forward modeling. , , The joint inversion area has a spatial dimension of 30×30×30m. The seismic wave velocity model is divided into a 1×1×1m spatial grid, the electrical resistivity model is divided into a 1×1×1m spatial grid, and the ground-penetrating radar dielectric constant model is divided into a 0.05×0.05m planar grid. A fault model containing karst caves is set up, with a 5×5×5m karst cave located 5m in front of the tunnel face and a 30×7×30m fault fracture zone located 23m in front of the tunnel face. The wave velocity model has a background wave velocity of 3000m / s, a karst cave wave velocity of 2100m / s, and a fault fracture zone wave velocity of 2500m / s. The resistivity model has a background resistivity of 1000Ω·m, a karst cave resistivity of 100Ω·m, and a fault fracture zone resistivity of 500Ω·m. The dielectric constant model has a background relative dielectric constant of 4, a karst cave relative dielectric constant of 10, and a fault fracture zone relative dielectric constant of 7.5.

[0045] (3) Forward modeling of each method and calculation of data loss term Forward modeling was performed on the established real model using seismic, electrical resistivity tomography, and ground-penetrating radar methods to obtain real observation data from the three methods. , , After processing the model parameters output by the implicit neural network, the initial models of the three methods were forward-modeled to obtain the predicted observation data of the three detection methods. , , The data loss term is calculated based on the difference between the predicted data and the actual data. The expression for the data loss term is as follows:

[0046]

[0047]

[0048]

[0049] in, Indicates data loss items; The weights for the data loss items of electrical resistivity tomography (EDT) and ground penetrating radar (GPR) are respectively set to ensure that the magnitudes of the three data loss items are the same or similar. , , These are data loss items for seismic, electrical resistivity, and ground-penetrating radar methods, respectively. , , These represent the number of observation data obtained from seismic, electrical resistivity, and ground-penetrating radar, respectively. , , These are the predicted observation data obtained through seismic, electrical resistivity, and ground-penetrating radar forward modeling, respectively, after the network output parameters have been inversely normalized. These are actual observation data from seismic, electrical resistivity tomography, and ground-penetrating radar, respectively.

[0050] (4) Calculate the cross gradient loss term In this example, the three-dimensional cross-gradient loss is calculated for the wave velocity model-resistivity model data pair, and the two-dimensional cross-gradient loss is calculated for the wave velocity model-dielectric constant model and resistivity model-dielectric constant model data pairs. The expression for the cross gradient loss term is as follows:

[0051] For three-dimensional cross gradients,

[0052]

[0053]

[0054] For two-dimensional cross gradients,

[0055]

[0056]

[0057] in, This represents the cross-gradient structure loss term; This indicates the number of data categories, which is 3 in this example. These are weighting coefficients used to balance the magnitude of the cross-gradient loss across the three parameter pairs; The cross gradient loss for a given data pair; and For the gradient of the model parameters of the corresponding data pair, These are the model parameter values ​​for the corresponding data pairs; Representing the current grid and Forward grid in the direction; The current grid is in Side length in the direction, They are respectively The side length of the forward grid in the direction.

[0058] (5) Calculate the prior physical property constraint loss term In this example, using known point physical property parameters obtained through methods such as advanced drilling and core testing, a priori physical property loss term is constructed by comparing the network's output value at the known point coordinates with the measured true value. The expression is as follows:

[0059] in, This represents the prior data loss term; This indicates the number of data categories, which is 3 in this example. These are weighting coefficients used to balance the magnitude differences between different data points; These are partial model parameter values ​​corresponding to prior physical property parameters, output by the implicit neural network and after inverse normalization. These are partial point physical property parameter values ​​obtained through methods such as advanced drilling and core testing.

[0060] (6) Construct the target loss function This example includes three data loss terms, one structural constraint term, and one physical property constraint term, expressed as follows:

[0061] in, These are the weights of the cross-gradient structure constraint term and the weights of the prior physical property constraint term, respectively, to ensure that the magnitudes of each loss term are the same or similar. , , These are the data loss term, the cross-gradient structure loss term, and the prior property loss term, respectively.

[0062] (7) Network Iterative Optimization The objective loss function is backpropagated to update and optimize the network parameters of the implicit neural network, and the model parameter values ​​are re-output. The previous process is repeated until the total loss objective function meets the termination condition. The normalized model parameters output by the implicit neural network are denormalized and mapped to the three-dimensional space model to obtain the final joint inversion result.

[0063] Example 2 A tunnel joint inversion system based on implicit neural networks and geological constraints includes: The implicit neural network construction module is configured to construct an implicit neural network, which takes the three-dimensional coordinate vector of the target inversion space as input and the normalized model parameters of wave velocity, resistivity and dielectric constant at the corresponding position as output. The implicit neural network pre-training module is configured to obtain prior physical property parameters through advanced drilling, etc., construct sample pairs of spatial coordinates and physical property parameters, and use them as training sets to perform supervised pre-training on the pre-constructed implicit neural network. The mean square error between the normalized model parameters output by the network and the normalized values ​​of the physical property parameters in the sample pairs is used as the target loss function for the implicit neural network pre-training. After the pre-training is completed, the weights and bias terms of at least some structures in the implicit neural network are fixed and the relationship between physical property parameters is preserved. The forward modeling module is configured to perform inverse normalization processing on the normalized model parameters output by the implicit neural network, and to perform forward modeling calculations on the corresponding seismic, electrical resistivity tomography, and ground-penetrating radar prediction data using the wave equation, Poisson equation, and Maxwell equation. The data loss calculation module is configured to calculate the difference between earthquake prediction data and actual data as earthquake data constraints, calculate the difference between electrical resistivity tomography (ERT) prediction data and actual data as ERT data constraints, and calculate the difference between ground penetrating radar (GPR) prediction data and actual data as GPR data constraints. The module balances the significant magnitude differences among the three types of data constraints using weighting factors to construct the data constraints. The cross-gradient loss calculation module is configured to perform denormalization processing on the normalized model parameters output by the implicit neural network, and calculate the pairwise cross-gradient loss of three parameter pairs: wave velocity-resistivity, wave velocity-dielectric constant, and resistivity-dielectric constant through the cross-gradient loss function to construct the cross-gradient loss term. The prior physical property loss calculation module is configured to introduce known point physical property parameters obtained by methods such as advanced drilling and core testing, and construct a prior physical property loss term by comparing the difference between the network's output value at the known point coordinates and the measured true value. The objective function loss assembly module is configured to calculate a weighted sum of three types of data losses as data constraints, cross-gradient loss as structural constraints, and prior property loss as property constraints. The weights of each constraint are set based on the magnitude of the data constraints to form the total objective loss function. The implicit neural network iterative module is configured to backpropagate the total objective loss function, update and optimize the network parameters of the implicit neural network until the total objective loss function satisfies the iteration condition. The normalized model parameters output by the implicit neural network are denormalized and mapped to three-dimensional space to obtain the final joint inversion result.

[0064] Example 3 A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store a plurality of instructions adapted for loading by the processor and executing the steps of the method described in Embodiment 1.

[0065] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0066] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] 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.

[0068] 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.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A joint inversion method for tunnel advance detection based on geological constraints, characterized in that, Includes the following steps: Obtain prior physical property parameters, construct sample pairs of spatial coordinates and physical property parameters, and use them as training sets to perform supervised pre-training on the pre-constructed coordinate mapping network. The coordinate mapping network takes the three-dimensional spatial coordinate vector of the target inversion space as input and the normalized model parameters of wave velocity, resistivity and dielectric constant at the corresponding position as output. After pre-training, fix the weights and bias terms of at least some structures in the coordinate mapping network and retain the relationship between physical property parameters. Based on the normalized model parameters output by the coordinate mapping network, after inverse normalization processing, forward modeling is performed for seismic, electrical resistivity tomography, and ground penetrating radar respectively to obtain the corresponding prediction data. The loss between the observed data and the prediction data is calculated, and a data loss term is constructed. Calculate the pairwise cross gradients of the three parameter pairs: wave velocity-resistivity, wave velocity-dielectric constant, and resistivity-dielectric constant, and construct the cross gradient loss term accordingly. Based on the rock strata information revealed by advanced drilling and the test data of core samples at corresponding locations, the spatial distribution characteristics of key physical property parameters are obtained, and prior physical property constraints are constructed. Different weights are set for different losses to balance the magnitude of the data of each loss term. After fusion, the final target loss function is obtained. Backpropagation of the loss is performed to update the weights and bias terms of the coordinate mapping network. The above process is repeated until the iteration condition is met. By using the updated coordinate mapping network, the model parameter distribution is obtained, enabling imaging of the wave velocity and resistivity properties of the subsurface medium.

2. The joint inversion method for tunnel advanced detection based on geological constraints as described in claim 1, characterized in that, The coordinate mapping network also includes an input layer and an output layer. The input layer contains three neurons to receive spatial coordinates x, y, and z. The output layer contains three neurons to output normalized wave velocity model parameters, resistivity model parameters, and dielectric constant model parameters.

3. The joint inversion method for tunnel advanced detection based on geological constraints as described in claim 1, characterized in that, Based on the normalized model parameters output by the coordinate mapping network, after inverse normalization processing, the process of performing forward modeling for seismic, electrical resistivity, and ground-penetrating radar (GPR) methods includes: performing inverse normalization processing on the normalized model parameters obtained from the pre-training of the coordinate mapping network to obtain the initial models for seismic, electrical resistivity, and GPR forward modeling; constructing a certain spatial region in front of the tunnel face as the joint inversion region for seismic, electrical resistivity, and GPR methods; selecting the joint inversion region according to the effective range of electrical resistivity at the longitudinal scale, while using the same resistivity model, wave velocity model, and dielectric constant model at the transverse scale.

4. The tunnel advanced detection joint inversion method based on geological constraints as described in claim 1, characterized in that, The process of setting appropriate weights for different losses to balance the magnitude of the data of each loss term includes: selecting the data loss term as the benchmark, setting the weights of the cross gradient structure constraint term and the prior physical property constraint term, so that the three losses are of the same or similar magnitude, and balancing the problem that the impact of each loss term on the total loss function due to data differences is greater than the set value.

5. The joint inversion method for tunnel advanced detection based on geological constraints as described in claim 1, characterized in that, The expression for the data loss term is as follows: in, The weights for the data loss terms of electrical resistivity tomography (EDT) and ground penetrating radar (GPR) are respectively set to ensure that the magnitudes of the three data loss terms are the same or similar. , , These represent the amounts of observational data obtained from seismic, electrical resistivity, and ground-penetrating radar, respectively. These are data loss items for seismic, electrical resistivity tomography, and ground penetrating radar, respectively.

6. The joint inversion method for tunnel advanced detection based on geological constraints as described in claim 1, characterized in that, The expression for the cross-gradient loss term is as follows: For 3D cross gradients: For two-dimensional cross gradients: Wherein represents the number of data categories; These are weighting coefficients used to balance the magnitude of the cross-gradient loss across the three parameter pairs; The cross gradient loss for a given data pair; and For the gradient of the model parameters of the corresponding data pair, These are the model parameter values ​​for the corresponding data pairs; s, b, c, and d represent the current grid and the forward grid in the x, y, and z directions, respectively. , , These represent the side lengths of the current grid in the x, y, and z directions, respectively. , , These are the side lengths of the forward grid in the x, y, and z directions, respectively.

7. The joint inversion method for tunnel advanced detection based on geological constraints as described in claim 1, characterized in that, The prior property constraint loss term is: in, Indicates the number of data categories. These are weighting coefficients used to balance the differences in magnitude between different data points. These are partial model parameter values ​​output by the coordinate mapping network and after inverse normalization, corresponding to the prior physical property parameters. These are partial point physical property values ​​obtained through advanced drilling and core testing.

8. The joint inversion method for tunnel advanced detection based on geological constraints as described in claim 1, characterized in that, The process of backpropagating the loss and updating the weights and biases of the coordinate mapping network includes: backpropagating the target loss function, updating and optimizing the network parameters of the implicit neural network, re-outputting the model parameter values, repeating the previous process until the total loss objective function meets the termination condition, and then denormalizing the normalized model parameters output by the implicit neural network and mapping them to the three-dimensional space model to obtain the final joint inversion result.

9. A joint inversion system for tunnel advanced detection based on geological constraints, characterized in that, include: The coordinate mapping network construction and training module is configured to obtain prior physical property parameters, construct sample pairs of spatial coordinates and physical property parameters, and use them as training sets to perform supervised pre-training on the pre-constructed coordinate mapping network. The coordinate mapping network takes the three-dimensional spatial coordinate vector of the target inversion space as input and the normalized model parameters of wave velocity, resistivity and dielectric constant at the corresponding position as output. After the pre-training is completed, the weights and bias terms of at least some structures in the coordinate mapping network are fixed and the relationship between physical property parameters is preserved. The forward modeling module is configured to use normalized model parameters output by the coordinate mapping network. After inverse normalization, forward models are performed for seismic, electrical resistivity, and ground-penetrating radar respectively to obtain the corresponding prediction data. The data loss calculation module is configured to calculate the loss between observed data and predicted data, and to construct a data loss term. The cross-gradient loss calculation module is configured to calculate the pairwise cross-gradients of three parameter pairs: wave velocity-resistivity, wave velocity-dielectric constant, and resistivity-dielectric constant, in order to construct the cross-gradient loss term. The prior physical property loss calculation module is configured to obtain the spatial distribution characteristics of key physical property parameters and construct prior physical property constraints based on the rock strata information revealed by the advance drilling and the core sample test data at the corresponding locations. The objective function loss fusion module is configured to set corresponding weights for different losses to balance the magnitude of the data of each loss term. After fusion, the final objective loss function is obtained. Backpropagation of the loss is performed to update the weights and bias terms of the coordinate mapping network. The above process is repeated until the iteration condition is met. The coordinate mapping network iteration module is configured to use the updated coordinate mapping network to obtain the model parameter distribution and realize imaging of the wave velocity, resistivity, and dielectric constant properties of the underground medium.

10. A terminal device, characterized in that, It includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store a plurality of instructions adapted to be loaded by the processor and executed as steps in the method of any one of claims 1-8.

Images