A method and system for detecting water-bearing structures in front of a tunnel face

Abstract

The present application belongs to the technical field of geophysical inversion, and provides a tunnel face front water-bearing structure detection method and system, obtains potential observation data in front of the tunnel face, and carries out pretreatment; a double-branch collaborative inversion network is constructed and trained, including a physical field branch for predicting potential distribution and a geometric boundary branch for predicting level set function; through a differentiable interface function, the physical field branch and the geometric boundary branch are coupled; a joint loss function is constructed to train the double-branch collaborative inversion network, and based on the pretreated potential detection data, the trained double-branch collaborative inversion network is used for processing to obtain the spatial distribution of the water-bearing structure. The present application fundamentally solves the problem of exponential increase of calculation cost with resolution, and meets the real-time requirement of high-precision imaging in tunnel construction site.

Description

Technical Field

[0001] This invention belongs to the field of geophysical inversion technology, specifically relating to a method and system for detecting water-bearing structures ahead of a tunnel face. 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] Accurate detection of water-bearing structures ahead of the tunnel face is a core element in ensuring safety during tunnel construction. Sudden water inrushes are among the most common and extremely dangerous geological hazards. As national infrastructure extends into complex geological areas, particularly with the large-scale construction of plateau tunnels, the geological conditions faced by these projects are becoming increasingly complex, with frequent occurrences of unfavorable geological formations such as karst caves, water-rich faults, and high-pressure fracture zones. These water-bearing structures can easily induce catastrophic water inrushes, surrounding rock instability, and even surface collapse, causing severe economic losses and casualties, posing an extreme threat to tunnel construction safety.

[0004] Advanced geological forecasting is crucial for mitigating this threat. Among these methods, electrical resistivity methods (such as resistivity and induced polarization methods) are exceptionally sensitive to resistivity differences in water and aquifer structures, making them the mainstream technique for detecting aquifers ahead. They can effectively identify potential hazards such as groundwater, water-filled caves, and fault fracture zones, providing a basis for developing treatment plans in advance, preventing disasters, and fundamentally ensuring construction safety. Therefore, advanced electrical resistivity surveying of tunnels is a lifeline for early warning of water hazards and determining construction plans.

[0005] However, current electrical resistivity tomography (EDT) methods for detecting and retrieving aquifers, such as least squares inversion, Bayesian inversion, and fully connected neural network inversion, while capable of characterizing the shape and location of aquifers to some extent, generally suffer from insufficient resolution and blurred boundary identification. Currently, the detection of aquifers ahead of tunnel faces three major technical bottlenecks: The challenge of multiple anomalies: The scale of aquifer structures varies greatly, and traditional fixed-grid inversion methods cannot clearly characterize millimeter-scale fractures and meter-scale karst caves in a single model at the same time, resulting in a contradiction between resolution and stability.

[0006] Blurred boundary identification: Traditional inversion relies on smooth constraints, which causes the inversion results of multiple connected hydrous structures to merge into a fuzzy anomaly, making it impossible to accurately characterize the sharp boundaries between them, resulting in large geometric errors.

[0007] Computational efficiency bottleneck: Traditional gradient-based methods require repeated solving of the inverse matrix during the optimization process, which leads to an exponential increase in computational cost as resolution increases. Summary of the Invention

[0008] To address the aforementioned problems, this invention proposes a method and system for detecting water-bearing structures ahead of the tunnel face. This invention fundamentally solves the problem of exponentially increasing computational costs with increasing resolution, and meets the real-time requirements for high-precision imaging at tunnel construction sites.

[0009] According to some embodiments, the present invention adopts the following technical solution: A method for detecting water-bearing structures ahead of a tunnel face includes the following steps: Acquire potential observation data in front of the tunnel face and perform preprocessing; A two-branch cooperative inversion network is constructed and trained, the two-branch cooperative inversion network including a physical field branch for predicting the potential distribution and a geometric boundary branch for predicting the level set function; By using a differentiable interface function, the level set function output by the geometric boundary branch is mapped to a spatial conductivity distribution, and then substituted into a preset physical control equation to couple the physical field branch with the geometric boundary branch. A joint loss function is constructed to train the dual-branch collaborative inversion network. The joint loss function includes at least a physical constraint loss based on the physical control equation and a data fitting loss based on the potential observation data. During the training process of the dual-branch collaborative inversion network, the shape parameters of the interface function are dynamically adjusted so that the boundary representation of the water-bearing structure gradually changes from fuzzy to clear. Based on the preprocessed potential detection data, the spatial distribution of hydrous structures is obtained by using a trained bi-branch collaborative inversion network.

[0010] As an alternative implementation method, the process of acquiring and preprocessing potential observation data in front of the tunnel face includes: setting up an observation system at the tunnel face and the tunnel wall behind it; using power supply electrodes to inject a stable DC current into the surrounding rock medium to establish an artificial electric field; synchronously acquiring potential response signals through a measuring electrode array; and preprocessing the acquired raw potential data to remove distortion points and noise interference, perform standardization processing, and construct a potential dataset.

[0011] As an alternative implementation, the two branches of the dual-branch collaborative inversion network are set in parallel, and both branches adopt a fully connected neural network structure. The physical field branch serves as a solver for the physical control equations and is used to fit the Poisson equation of the underground electric field to ensure that the inversion results strictly satisfy the physical conservation law. The geometric branch is used to implicitly characterize the geometric shape of the water-bearing structure.

[0012] As an alternative implementation, the physical control equation is: ; in, For Hamiltonian operators, For potential distribution, For conductivity distribution, For point power intensity, For the Dirac function, For spatial coordinates, Location of the power supply point; The boundary conditions are: ; ; These respectively represent the insulation boundary The normal derivative is zero at the known potential boundary. Potential is a fixed value .

[0013] As an alternative implementation, the differentiable interface function is a discontinuous conductivity mapping model, specifically: ; in, The electrical conductivity of the surrounding rock, For the electrical conductivity of water-bearing structures, This is a level-set-based Heaviside step function used to achieve smooth transitions and differentiable coupling between different dielectric conductivities. For level set functions, define It is a hydrous structural region. The surrounding rock area To construct the boundary.

[0014] As a further defined implementation, a Heaviside step function based on level sets... The function is defined using a differentiable approximation based on the Sigmoid form: ; Used to distinguish different geological media, providing a physical medium switching function, when the level set function When it is a positive value, The conductivity at that location approaches 1. Approaching the conductivity of water-bearing structures ;when When it is negative, The conductivity at that location approaches 0. Approaching the background electrical conductivity of the surrounding rock ; To introduce a sharpening factor that decays exponentially with the number of rounds.

[0015] As a further limiting embodiment, the sharpening coefficient decreases exponentially with the number of rounds. for: ; in, The initial smoothing coefficients are... For attenuation rate, parameter Defined as a sharpening factor that controls the width of the transition band at the interface, it is used to provide dynamic sharpening control. In the early stages of network training, it is set to be less than a threshold. The value blurs the interface, which facilitates the propagation of gradient information across the boundary and prevents the inversion from getting trapped in local minima; as training progresses, by gradually increasing... Value, making By gradually approaching the ideal step function, accurate and clear imaging of the boundary of aquifer structures can be achieved.

[0016] As an alternative implementation, the process of mapping the level set function output by the geometric boundary branch to a spatial conductivity distribution and substituting it into a preset physical control equation includes: mapping the coordinates of the spatial sampling points... The input network undergoes multi-level scaling and logarithmic transformation to generate high-dimensional feature vectors. These high-dimensional feature vectors are then fed simultaneously to both the physics and geometry branches for parallel computation to obtain the predicted potential under the current parameters. and prediction level set function ; will predict Converted to conductivity at the current location This completes the numerical transfer from the geometric field to the physical parameter field, and comprehensively predicts... , , Given the coordinate information, we substitute it into the joint loss function to calculate the PDE residual, boundary error, and data mismatch value, providing gradient basis for parameter updates of the dual-branch collaborative inversion network.

[0017] As an alternative implementation, the joint loss function includes physical constraint loss, data fitting loss, boundary condition loss, and regularization constraint loss. The physical constraint loss is used to ensure that the output of the two-branch collaborative inversion network satisfies the physical control equation. The boundary condition loss includes known potential boundaries and insulation boundaries. The data fitting loss is used to fit measured potential data, with the integration domain being the observation surface. The regularization constraint loss includes level set unit gradient constraints, boundary repulsion constraints, and conductivity range constraints to ensure geological rationality.

[0018] As an alternative implementation, during the training process of the dual-branch collaborative inversion network, the optimization objective is to minimize the joint loss function. The parameter gradients of the physical field branch and the geometric branch are calculated through the backpropagation algorithm, and the network weights of the two branches of the dual-branch collaborative inversion network are iteratively updated.

[0019] A system for detecting water-bearing structures ahead of a tunnel face includes: The data acquisition module is configured to acquire potential observation data in front of the tunnel face and perform preprocessing. The network model construction and training module is configured to construct and train a two-branch cooperative inversion network, which includes a physical field branch for predicting potential distribution and a geometric boundary branch for predicting level set functions. A differentiable interface function maps the level set function output by the geometric boundary branch to a spatial conductivity distribution, and substitutes it into a preset physical control equation to couple the physical field branch and the geometric boundary branch. A joint loss function is constructed to train the two-branch cooperative inversion network, and the joint loss function includes at least a physical constraint loss based on the physical control equation and a data fitting loss based on the potential observation data. During the training process of the two-branch cooperative inversion network, the shape parameters of the interface function are dynamically adjusted to gradually change the boundary representation of the hydrous structure from fuzzy to clear. The identification module is configured to process the pre-processed potential detection data using a trained bi-branch collaborative inversion network to obtain the spatial distribution of hydrous structures.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention combines the principle of resistivity method with a dual-branch neural network architecture to simultaneously invert the potential field distribution and the geometric boundary of aquifer structures. It innovatively introduces a dynamic level set function neural network branch to represent multi-connected aquifer structures. Combined with multiple regularization constraints and a hybrid optimization strategy, it employs an unsupervised learning method to solve the problem of detecting multiple anomalies in existing technologies. Addressing the difficulty of traditional methods in simultaneously handling millimeter-level fractures and meter-level karst caves, and the resulting contradiction between resolution and stability, this invention breaks through the limitations of single-scale analysis by constructing multi-scale feature inputs containing multi-level scaling factors and logarithmic moduli. This mechanism enhances the network's ability to capture strong gradient fields near the source, achieving simultaneous and detailed characterization of microscopic fractures and macroscopic structures within the same inversion model.

[0021] This invention overcomes the shortcomings of ambiguous boundary identification of anomalies. In response to the problem that traditional inversion relies on smooth constraints, which leads to the fusion of adjacent structures and the inability to characterize sharp boundaries, this invention establishes a dual-branch architecture that coordinates physical and geometric fields. By utilizing the implicit evolution capability of the level set of the geometric branch, combined with the discontinuous conductivity mapping mechanism, it directly drives the "steep" changes of physical parameters, thereby accurately restoring the topological morphology and sharp boundaries of complex underground water-bearing bodies and avoiding the artificial fusion of anomalies.

[0022] This invention overcomes the computational efficiency bottleneck of high-resolution imaging. Addressing the challenge of traditional gradient methods requiring repeated inverse matrix calculations, where computational costs increase exponentially with resolution, this invention employs a meshless deep learning inversion framework, completely avoiding large-scale matrix operations. Combined with a dynamic sharpening strategy that adaptively decays during training, the network can rapidly converge from the global contour to fine boundaries, significantly reducing computational complexity and time costs while maintaining high-resolution imaging quality.

[0023] 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

[0024] 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.

[0025] Figure 1 A method flowchart is provided for one embodiment; Figure 2 A schematic diagram of a dual-branch collaborative inversion network architecture provided in one embodiment; Figure 3 A flowchart of training and parameter optimization for a dual-branch collaborative inversion network is provided for one embodiment. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] 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.

[0028] 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.

[0029] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0030] Example 1 A method for detecting water-bearing structures ahead of the tunnel face, such as Figure 1 As shown, it includes the following steps: Acquire potential observation data in front of the tunnel face and perform preprocessing; A two-branch cooperative inversion network is constructed and trained, the two-branch cooperative inversion network including a physical field branch for predicting the potential distribution and a geometric boundary branch for predicting the level set function; By using a differentiable interface function, the level set function output by the geometric boundary branch is mapped to a spatial conductivity distribution, and then substituted into a preset physical control equation to couple the physical field branch with the geometric boundary branch. A joint loss function is constructed to train the dual-branch collaborative inversion network. The joint loss function includes at least a physical constraint loss based on the physical control equation and a data fitting loss based on the potential observation data. During the training process of the dual-branch collaborative inversion network, the shape parameters of the interface function are dynamically adjusted so that the boundary representation of the water-bearing structure gradually changes from fuzzy to clear. Based on the preprocessed potential detection data, the spatial distribution of hydrous structures is obtained by using a trained bi-branch collaborative inversion network.

[0031] In this embodiment, the process of acquiring and preprocessing the potential observation data in front of the tunnel face includes: setting up an observation system at the tunnel face and behind the tunnel wall; injecting a stable DC current into the surrounding rock medium using power supply electrodes to establish an artificial electric field; simultaneously acquiring potential response signals through a measuring electrode array; and preprocessing the acquired raw potential data to remove distortion points and noise interference, thereby constructing an observation potential dataset. This dataset contains the physical field response information of the target body (such as aquifer) to be inverted underground, which serves as a data-driven constraint for subsequent deep learning network training.

[0032] The process of constructing and training a bi-branch co-inversion network involves constructing a bi-branch deep neural network that coordinates physical fields and geometric boundaries. This network does not rely on specific geological prior models, but automatically learns geological structures through the co-constraints of physical equations and geometric level sets.

[0033] Dual-branch network architecture: such as Figure 2 As shown, the network consists of two parallel branches. The physical field branch input... Output the predicted potential distribution This branch is responsible for fitting the physical governing equations (i.e., the Poisson equation). The geometric boundary branch takes a high-dimensional feature vector as input and outputs the predicted level set function. This branch is used to implicitly characterize the geometric boundaries of geological bodies.

[0034] Conductivity mapping model: To couple the physical field and the geometric field, a Heaviside function in the form of a sigmoid is used. Constructing the conductivity mapping relationship: ; in , These are the baseline values ​​for electrical conductivity of the surrounding rock and the aquifer, respectively.

[0035] Specifically, the process of the physical governing equation (i.e., the Poisson equation) is as follows: Solving the global domain We establish the Poisson equation, which describes the distribution of the underground electric field, as the governing equation: ; in, For Hamiltonian operators, For potential distribution, For conductivity distribution, For point power intensity, For the Dirac function, For spatial coordinates, The location of the power supply point is defined. Boundary conditions are also defined: ; ; These respectively represent the insulation boundary ( The normal derivative is zero, at the known potential boundary ( The potential is a fixed value. .

[0036] To characterize the geometry of aquifer structures, a level set function is introduced. ,definition It is a hydrous structural region. The surrounding rock area To construct the boundary, a discontinuous conductivity mapping model is established based on this: ; in, The electrical conductivity of the surrounding rock, For the electrical conductivity of water-bearing structures, This is a level-set-based Heaviside step function. A smooth Heaviside step function is used to achieve smooth transitions and differentiable coupling between different dielectric conductivities.

[0037] To ensure the gradient differentiability of the neural network during backpropagation, this invention abandons the traditional discrete step function and adopts a differentiable approximation function definition based on the Sigmoid form: ; To achieve progressive edge sharpening, a differentiable sigmoid form is defined, and a sharpening coefficient that decays exponentially with the number of epochs is introduced during training. : ; in, This is the initial smoothing coefficient (e.g., 0.1m). This represents the attenuation rate (e.g., 0.98).

[0038] In the definition of a differentiable approximation function, the specific physical meaning and function of each parameter are as follows: Function of physical medium switch: Used to distinguish different geological media. When the level set function... When the value is positive (i.e., inside an aquifer), The conductivity at that location approaches 1. Approaching the conductivity of water-bearing structures ;when When the value is negative (i.e., in the surrounding rock area), The conductivity at that location approaches 0. Approaching the background electrical conductivity of the surrounding rock .

[0039] Dynamic sharpening control function: parameters Defined as the sharpening factor that controls the width of the transition band in the interface. In the early stages of network training, a smaller value is set. The value blurs the interface, which facilitates the propagation of gradient information across the boundary, thus preventing the inversion from getting trapped in local minima; as training progresses, by gradually increasing... Value, making By gradually approaching the ideal step function, accurate and clear imaging of the boundary of aquifer structures can be achieved.

[0040] In this embodiment, the two branches of the dual-branch cooperative inversion network undertake different inversion tasks: Physics branch: Employing a fully connected neural network ,enter Output the predicted potential distribution .

[0041] ; in, For network parameters; The input is a spatial coordinate vector.

[0042] This branch acts as a solver for the physical equations, responsible for fitting the Poisson equation for the underground electric field and ensuring that the inversion results strictly satisfy the physical conservation laws.

[0043] Geometric branch: Employs a fully connected neural network ,enter Output the predicted level set function .

[0044] ; For network parameters; This branch is used to implicitly characterize the geometry of aquifer structures and has the ability to automatically evolve complex topological structures.

[0045] To achieve synergistic optimization of the two branches, this embodiment establishes a coupling relationship between the constructed smooth Heaviside function and the discontinuous conductivity mapping model. Specifically, the level set function output by the geometric branch is... Substituting into the mapping model above, we can smoothly transform it into a spatial conductivity distribution. In this way, changes in geometric shape are mapped to physical parameters in real time, which are then directly substituted into physical equations for calculation, achieving a strong coupling between "form" and "number".

[0046] The complete forward inference process of a two-branch cooperative inversion network is as follows: Step 1: Feature Encoding. This involves encoding the coordinates of the spatial sampling points. The input network undergoes multi-level scaling and logarithmic transformation to generate high-dimensional feature vectors. Step 2: Parallel Inference. The high-dimensional feature vector is simultaneously fed into both the physical field branch and the geometric branch, and the predicted potential under the current parameters is calculated in parallel. and prediction level set function ; Step 3: Parameter Coupling. Using the defined mapping relationships, the predicted parameters are coupled... Converted to conductivity at the current location This completes the numerical transfer from the geometric field to the physical parameter field; Step 4: Residual Calculation. Comprehensive Prediction , , Given the coordinate information, substitute it into the total loss function defined in step two to calculate the PDE residual, boundary error, and data mismatch value, providing a gradient basis for network parameter updates.

[0047] During the training phase, the total loss function (i.e., the joint loss function) is defined. Used for network parameter optimization, this function consists of the following four weighted parts: ; , , and These are the weights for the corresponding losses.

[0048] Physical constraint loss: The output of the constraint network satisfies the Poisson equation, and the integration domain is the global domain. : ; in, Indicates the entire computational domain Expected value within; It is a divergence operator; To predict the spatial gradient of the potential; Electrical conductivity; This is a point source stimulus term.

[0049] Boundary condition loss: including known potential boundary ( ) and insulation boundary ( ): ; in, and These represent the expected values ​​at known potential boundaries and insulation boundaries, respectively; Predict potentials for the network; The given known boundary potential is a fixed value; This represents the partial derivative of the potential along the direction of the outward normal to the boundary.

[0050] Data fitting loss: used to fit measured potential data (integration domain is the observation surface). ): ; in, This indicates the observation surface where the detection electrodes are arranged. Expected value; The potential value predicted by the network on this observation surface; These are actual observed potential data collected on-site.

[0051] Regularization constraint loss: Includes triple physical constraints to ensure geological rationality, as shown in Table 1: ; Table 1. Triple Regularization Constraint System

[0052] The gradient of the level set function; It is a linear rectification activation function; To predict electrical conductivity; and These represent the maximum and minimum electrical conductivity extremes under a reasonable geological model.

[0053] The specific training process, such as Figure 3 As shown, it includes: S1 parameter initialization Physics field branch network parameters and geometric branch network parameters Perform initialization.

[0054] S2 sets dynamic parameters Set the initial values ​​for the dynamic sharpening parameters, including the initial sharpening coefficient. and attenuation rate .

[0055] S3Adam Global Optimization (Main Loop) Set the maximum number of iterations Enter the main loop training (For each k = 1… ): S3.1 Random Sampling In the solution domain Random spatial sampling is performed within the sample area to obtain a set of sampling points. D .

[0056] S3.2 Calculate Physical Loss

[0057] Substitute the sampling points into the network and calculate the Poisson equation residuals. This yields the physical loss. .

[0058] S3.3 Calculate Boundary Loss

[0059] Based on the Neumann and Dirichlet conditions, the boundary errors are calculated, and the boundary losses are obtained. .

[0060] S3.4 Calculate data loss

[0061] Calculation of predicted potential and measured observation data The mean square error between them. This yields the data loss. .

[0062] S3.5 Calculate Regularization Loss To ensure the geological validity of the solution, the unit gradient loss of the level set is calculated separately. Boundary repulsion loss and conductivity range loss .

[0063] S3.6 parameter update Calculate the gradient of the total loss with respect to the parameters. Update network parameters using the Adam optimizer , .

[0064] S3.7 Dynamic Sharpening Update At the end of each iteration, according to the formula Update the sharpening factor. As training progresses... The function decreases exponentially, causing the Heaviside function to gradually approach the step function, thereby automatically sharpening the boundaries of the aquifer.

[0065] S4 L-BFGS Local Fine-tuning After completing Adam training, switch to the L-BFGS optimizer. Utilize its second-order optimization properties to fine-tune the network parameters, further reducing physical residuals and improving inversion accuracy.

[0066] S5 Result Output Output the final predicted potential after training is complete. u and hydrous structural level set function .

[0067] The key parameter settings in the above process are as follows: To ensure algorithm performance, the key parameter values ​​in this embodiment are set according to Table 2 as follows: Table 2 Typical parameter values

[0068] During training, the constructed total loss function is used as the optimization objective. The parameter gradients of the physics branch and the geometry branch are calculated using the backpropagation algorithm, and the network weights are iteratively updated. and The training process follows an evolutionary pattern of "first the overall outline, then the fine details": in the early stages of training, the larger... The value allows for rapid topological changes in the geometric boundaries; as training progresses, Exponential decay leads to gradual sharpening and locking of the boundaries, ultimately resulting in high-resolution hydrous structural morphology.

[0069] To quantitatively evaluate the physical field fitting accuracy and geometric boundary identification effect of the inversion network, this embodiment introduces root mean square error (RMSE) and intersection-over-union ratio (IoU) as evaluation indicators.

[0070] The physical field fit evaluation uses the root mean square error (RMSE) to assess the degree of agreement between the predicted potential field and the observed data, defined as follows: ; in,N This represents the total number of observation points; For the network at the observation point The predicted potential value at the location; This corresponds to the actual observed potential value. The smaller the RMSE value, the higher the accuracy of the physical field branch in reconstructing the underground electric field.

[0071] The cross-union ratio (CUNR) is used to evaluate the overlap accuracy between the inverted hydrous structure geometry and the actual geological structure in anomaly boundary identification assessment. The CUNR is defined as follows: ; in For the predicted aquifer region, This represents the real area. The closer the IoU value is to 1, the more accurate the identification of the location and morphology of the aquifer.

[0072] Example 2 A system for detecting water-bearing structures ahead of a tunnel face includes: The data acquisition module is configured to acquire potential observation data in front of the tunnel face and perform preprocessing. The network model construction and training module is configured to construct and train a two-branch cooperative inversion network, which includes a physical field branch for predicting potential distribution and a geometric boundary branch for predicting level set functions. A differentiable interface function maps the level set function output by the geometric boundary branch to a spatial conductivity distribution, and substitutes it into a preset physical control equation to couple the physical field branch and the geometric boundary branch. A joint loss function is constructed to train the two-branch cooperative inversion network, and the joint loss function includes at least a physical constraint loss based on the physical control equation and a data fitting loss based on the potential observation data. During the training process of the two-branch cooperative inversion network, the shape parameters of the interface function are dynamically adjusted to gradually change the boundary representation of the hydrous structure from fuzzy to clear. The identification module is configured to process the pre-processed potential detection data using a trained bi-branch collaborative inversion network to obtain the spatial distribution of hydrous structures.

[0073] 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 be implemented in 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.).

[0074] 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, as well as 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.

[0075] 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.

[0076] 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.

[0077] 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 method for detecting water-bearing structures ahead of a tunnel face, characterized in that, Includes the following steps: Acquire potential observation data in front of the tunnel face and perform preprocessing; A two-branch cooperative inversion network is constructed and trained, the two-branch cooperative inversion network including a physical field branch for predicting the potential distribution and a geometric boundary branch for predicting the level set function; By using a differentiable interface function, the level set function output by the geometric boundary branch is mapped to a spatial conductivity distribution, and then substituted into a preset physical control equation to couple the physical field branch with the geometric boundary branch. A joint loss function is constructed to train the dual-branch collaborative inversion network. The joint loss function includes at least a physical constraint loss based on the physical control equation and a data fitting loss based on the potential observation data. During the training process of the dual-branch collaborative inversion network, the shape parameters of the interface function are dynamically adjusted so that the boundary representation of the water-bearing structure gradually changes from fuzzy to clear. Based on the preprocessed potential detection data, the spatial distribution of hydrous structures is obtained by using a trained bi-branch collaborative inversion network.

2. The method for detecting water-bearing structures ahead of a tunnel face as described in claim 1, characterized in that, The process of acquiring and preprocessing potential observation data in front of the tunnel face includes: setting up an observation system at the tunnel face and behind the tunnel wall; injecting a stable DC current into the surrounding rock medium using power supply electrodes to establish an artificial electric field; synchronously acquiring potential response signals through a measuring electrode array; preprocessing the acquired raw potential data to remove distortion points and noise interference; standardizing the data; and constructing a potential dataset.

3. The method for detecting water-bearing structures ahead of a tunnel face as described in claim 1, characterized in that, The two branches of the dual-branch collaborative inversion network are set in parallel, and both branches adopt a fully connected neural network structure. The physical field branch serves as a solver for the physical control equations and is used to fit the Poisson equation of the underground electric field to ensure that the inversion results strictly satisfy the physical conservation law. The geometric branch is used to implicitly characterize the geometric morphology of the water-bearing structure.

4. The method for detecting water-bearing structures ahead of a tunnel face as described in claim 1, characterized in that, The physical control equation is: ； in, For Hamiltonian operators, For potential distribution, For conductivity distribution, For point power intensity, For the Dirac function, For spatial coordinates, Location of the power supply point; The boundary conditions are: ； ； These respectively represent the insulation boundary The normal derivative is zero at the known potential boundary. Potential is a fixed value .

5. The method for detecting water-bearing structures ahead of a tunnel face as described in claim 1, characterized in that, The differentiable interface function is a discontinuous conductivity mapping model, specifically: ； in, The electrical conductivity of the surrounding rock, For the electrical conductivity of water-bearing structures, This is a level-set-based Heaviside step function used to achieve smooth transitions and differentiable coupling between different dielectric conductivities. For level set functions, define It is a hydrous structural region. The surrounding rock area To construct the boundary.

6. The method for detecting water-bearing structures ahead of a tunnel face as described in claim 5, characterized in that, Heaviside step function based on level set The function is defined using a differentiable approximation based on the Sigmoid form: ； Used to distinguish different geological media, providing a physical medium switching function, when the level set function When it is a positive value, The conductivity at that location approaches 1. Approaching the conductivity of water-bearing structures ;when When it is negative, The conductivity at that location approaches 0. Approaching the background conductivity of the surrounding rock ; To introduce a sharpening factor that decays exponentially with the number of rounds.

7. The method for detecting water-bearing structures ahead of a tunnel face as described in claim 6, characterized in that, along with... Sharpening factor with exponential decay of rounds for: ； in, The initial smoothing coefficients are... For attenuation rate, parameter Defined as a sharpening factor that controls the width of the transition band at the interface, it is used to provide dynamic sharpening control. In the early stages of network training, it is set to be less than a threshold. The value blurs the interface, which facilitates the propagation of gradient information across the boundary and prevents the inversion from getting trapped in local minima; as training progresses, by gradually increasing... Value, making By gradually approaching the ideal step function, accurate and clear imaging of the boundary of aquifer structures can be achieved.

8. The method for detecting water-bearing structures ahead of a tunnel face as described in claim 1, characterized in that, The process of mapping the level set function output by the geometric boundary branch to a spatial conductivity distribution and substituting it into a preset physical control equation includes: mapping the coordinates of the spatial sampling points... The input network undergoes multi-level scaling and logarithmic transformation to generate high-dimensional feature vectors. These high-dimensional feature vectors are then fed simultaneously to both the physics and geometry branches for parallel computation to obtain the predicted potential under the current parameters. and prediction level set function ; will predict Converted to conductivity at the current location This completes the numerical transfer from the geometric field to the physical parameter field, and comprehensively predicts... , , Given the coordinate information, we substitute it into the joint loss function to calculate the PDE residual, boundary error, and data mismatch value, providing gradient basis for parameter updates of the dual-branch collaborative inversion network.

9. The method for detecting water-bearing structures ahead of a tunnel face as described in claim 1, characterized in that, The joint loss function includes physical constraint loss, data fitting loss, boundary condition loss, and regularization constraint loss. The physical constraint loss is used to ensure that the output of the two-branch collaborative inversion network satisfies the physical control equation. The boundary condition loss includes known potential boundaries and insulation boundaries. The data fitting loss is used to fit measured potential data, with the integration domain being the observation surface. The regularization constraint loss includes level set unit gradient constraints, boundary repulsion constraints, and conductivity range constraints to ensure geological rationality. During the training process of the dual-branch collaborative inversion network, the optimization objective is to minimize the joint loss function. The parameter gradients of the physical field branch and the geometric branch are calculated through the backpropagation algorithm, and the network weights of the two branches of the dual-branch collaborative inversion network are iteratively updated.

10. A system for detecting water-bearing structures ahead of a tunnel face, characterized in that, include: The data acquisition module is configured to acquire potential observation data in front of the tunnel face and perform preprocessing. The network model construction and training module is configured to construct and train a two-branch cooperative inversion network, which includes a physical field branch for predicting potential distribution and a geometric boundary branch for predicting level set functions. A differentiable interface function maps the level set function output by the geometric boundary branch to a spatial conductivity distribution, and substitutes it into a preset physical control equation to couple the physical field branch and the geometric boundary branch. A joint loss function is constructed to train the two-branch cooperative inversion network, and the joint loss function includes at least a physical constraint loss based on the physical control equation and a data fitting loss based on the potential observation data. During the training process of the two-branch cooperative inversion network, the shape parameters of the interface function are dynamically adjusted to gradually change the boundary representation of the hydrous structure from fuzzy to clear. The identification module is configured to process the pre-processed potential detection data using a trained bi-branch collaborative inversion network to obtain the spatial distribution of hydrous structures.

Images