A method for real-time location of microseismic sources in deep-buried tunnels

Abstract

This invention discloses a real-time microseismic source location method for deep-buried tunnels, including system initialization and geological velocity modeling, travel time field prediction based on a physical information neural network, online real-time monitoring and signal processing, microseismic source location calculation, and dynamic adaptive updates to adapt to tunnel excavation. This invention uses the equations describing the physical laws of wave propagation as mandatory constraints, embedding them into a deep neural network training framework to construct an "implicit" travel time field model that strictly adheres to physical principles while efficiently calculating the travel time at any point. This method combines the reliability of a physical model with the flexibility of a data-driven model and features a specially designed dynamic adaptive mechanism to address the unique challenges of frequent sensor layout adjustments and complex, time-varying geological conditions in tunnel engineering.

Description

Technical Field

[0001] This invention relates to the field of seismic source monitoring technology, and in particular to a method for real-time location of microseismic sources in deeply buried tunnels. Background Technology

[0002] In deep-buried tunnel construction, rockburst, as a typical deep-seated dynamic disaster, is characterized by its suddenness, destructive power, and short warning window, seriously threatening the safety of construction personnel and the progress of the project. To achieve dynamic perception and early warning of rockburst risks, microseismic monitoring technology has been widely applied in deep-buried tunnel projects in railways, water conservancy, and energy sectors. This technology uses a high-sensitivity sensor array to capture elastic wave signals released by microfractures in the surrounding rock in real time, thereby retrieving the location, energy, and frequency of microseismic events to assess the evolution trend of surrounding rock damage and the distribution of high-risk areas. Microseismic source location, as the core link in the microseismic monitoring process, directly determines the reliability of disaster early warning through its efficiency and accuracy. Currently, the mainstream location methods can be divided into two main categories: physical model-based methods and data-driven methods.

[0003] 1. Traditional physical model localization method

[0004] Currently, most mainstream methods in engineering are based on travel time difference models, which utilize the time difference of P-wave arrival at different sensors and combine it with the wave propagation speed in the medium to infer the spatial coordinates of the seismic source. The theoretical basis of this type of method originates from the travel time equation in seismology.

[0005] (1) Geiger method and its improvements: such as Geiger linear iteration method, double difference method and grid search method. These methods usually assume that the wave propagates in a straight line in a homogeneous or simple layered medium, and the velocity model is simplified.

[0006] (2) Numerical methods for complex paths: In order to improve the accuracy of velocity modeling under complex geological conditions, researchers have introduced numerical methods such as the fast travel method and Dijkstra's shortest path algorithm in recent years. By solving the equation of the equation, the propagation path of elastic waves in non-uniform media is characterized, so as to deal with the diffraction phenomenon caused by obstacles such as cavities and faults.

[0007] 2. Data-driven deep learning methods

[0008] With the development of artificial intelligence, end-to-end localization methods based on deep learning such as convolutional neural networks and recurrent neural networks have been proposed. These methods aim to learn the mapping relationship of the source location directly from the original waveform or features without explicitly building a physical model of wave propagation.

[0009] The disadvantages are:

[0010] 1. Traditional physical modeling methods struggle to balance accuracy and efficiency:

[0011] (1) Poor accuracy of uniform / simplified velocity model: Simplified methods such as Geiger method based on uniform velocity model directly assume that the wave speed is constant and the path is a straight line, completely ignoring the deflection (diffraction, refraction) of wave propagation path caused by cavities, joints, and velocity change interfaces that are common in deep buried tunnels, resulting in the positioning results being seriously deviated from the true location near complex structures.

[0012] (2) High-precision numerical methods are inefficient: Although methods such as the rapid travel method can simulate the propagation of waves in non-uniform media by solving the equation of the process function, they must first construct a high-resolution three-dimensional velocity model and perform iterative calculations at each grid point. The computational complexity is high and the time consumption is long, which cannot meet the real-time early warning requirements of minutes or even seconds in tunnel construction.

[0013] 2. Purely data-driven deep learning methods have poor engineering applicability:

[0014] (1) Strong dependence on a large amount of labeled data: Convolutional recurrent networks, recurrent neural networks and other models generally rely on a large number of labeled (i.e., known accurate location) microseismic event samples for supervised training.

[0015] (2) Data is scarce and dynamic in the deep-buried tunnel scenario: Deep-buried tunnels are typical linear extension structures, and sensors need to be frequently adjusted as the tunnel face advances, resulting in the monitoring system being in a state of continuous dynamic change; at the same time, effective microseismic events are sparse and spatially unevenly distributed, making it difficult to accumulate training data of sufficient scale and diversity. Therefore, existing data-driven models have weak generalization ability and high deployment costs in this scenario, limiting their practicality.

[0016] 3. Weak dynamic adaptability:

[0017] During tunnel excavation, the sensor placement locations and the velocity structure of the surrounding rock (such as low-velocity cavities formed in newly excavated areas) are constantly changing. Traditional physical methods require remodeling and resolving; data-driven methods require recollecting data and training the model, which cannot be quickly deployed and updated, seriously affecting the continuity of the monitoring system and its engineering practicality.

[0018] Therefore, it is necessary to develop a real-time microseismic source location method for deep-buried tunnels to solve the above problems. Summary of the Invention

[0019] The purpose of this invention is to design a real-time positioning method for microseismic sources in deeply buried tunnels in order to solve the above-mentioned problems.

[0020] The present invention achieves the above objectives through the following technical solutions:

[0021] A method for real-time location of microseismic sources in deeply buried tunnels includes:

[0022] Obtain relevant basic data for training, including geological survey reports, borehole data, seismic wave test results, and the precise spatial coordinates of the deployed microseismic sensor array in the area where the tunnel is located;

[0023] Using basic data, a three-dimensional velocity model V(x,y,z) and a three-dimensional computational domain for the tunnel area are first constructed. Then, the physical information neural network PINN is trained, specifically training a dedicated PINN for each microseismic sensor. i PINN i The training process uses the spatial coordinates of any point in the three-dimensional computational domain as input and the theoretical P-wave travel time from that point to the corresponding microseismic sensor as output. The training process adopts an unsupervised mode and uses physical constraints, boundary conditions and initial conditions for training.

[0024] The current microseismic waveform data is obtained by real-time acquisition of the sensor array and then filtered and denoised.

[0025] An automatic P-wave first arrival picking algorithm is used to identify the P-wave first arrival time t corresponding to each sensor based on the current microseismic waveform data. i obs Then call all the trained PINNs i By combining the double-difference objective function and global grid search, the spatial coordinates of the current seismic source are obtained.

[0026] Furthermore, the real-time location method for microseismic sources in deeply buried tunnels also includes the following: when the tunnel face advances a certain distance, some sensors need to be moved forward because they are too far from the face, or when the three-dimensional velocity model V(x,y,z) has been significantly modified based on new basic data; the PINN trained under the original microseismic sensor positions or similar three-dimensional velocity models V(x,y,z) is then used. i The network weights serve as the initial weights for the new network; based on the new sensor coordinates or the updated 3D velocity model V(x,y,z), the network is fine-tuned iteratively.

[0027] Preferably, the three-dimensional velocity model V(x,y,z) reflects large-scale structures, lithological interfaces, and the known locations of cavities or solution spaces.

[0028] Preferably, the double-difference objective function is:

[0029] ;

[0030] in, and When sensor waveforms i and j are observed, respectively, For the corresponding observation travel time difference, and PINN i and PINNj The predicted theoretical timeout, Let X = (x, y, z) be the theoretical travel time difference corresponding to the prediction, and let X = (x, y, z) be the source coordinates to be determined.

[0031] Specifically, the global mesh search is as follows: a three-dimensional search mesh is generated within the three-dimensional computational domain, and for each mesh node X... k Call all trained PINNs i Calculate T i (X k Then, the objective function value F(X) is calculated. k ), T i (X k ) represents grid point X k The theoretical timekeeping of sensor i.

[0032] Specifically, determining the epicenter location includes finding the grid node X that minimizes F(X). * This is the estimated location of the microseismic event.

[0033] Furthermore, in X * Sub-mesh accuracy can be obtained by performing secondary interpolation or local optimization in the vicinity.

[0034] Preferably, the physical constraints are functional equations obtained by factorization.

[0035] The beneficial effects of this invention are:

[0036] 1. Achieve the unity of high-precision physical modeling and efficient computation: Get rid of the dependence on explicit three-dimensional velocity model mesh construction and iterative numerical solution, and directly learn the travel time field from any spatial point to each sensor through physical information neural network, reducing the computational cost from "hours" to "minutes" in traditional numerical methods, and meeting real-time requirements.

[0037] 2. Embedding physical laws to ensure the rationality of the solution: The fundamental physical law of elastic wave propagation—the equation of equations—is embedded as a hard constraint in the loss function of the neural network, ensuring that the travel time prediction results strictly conform to the physical mechanism. Even without labeled data, a reasonable solution can be obtained through physical constraints, overcoming the dependence of purely data-driven methods on a large amount of labeled data.

[0038] 3. Enhanced adaptability to complex geological structures: By solving the equations, the propagation behavior of waves around anomalous bodies such as gradient velocity layers, layered interfaces, and cavities can be automatically captured, including refraction and diffraction. This results in higher positioning accuracy in models containing these complex structures than uniform velocity models and fast travel methods.

[0039] 4. Achieve rapid dynamic adaptation: By introducing a transfer learning mechanism, when tunnel excavation causes sensor displacement or velocity field updates, the model weights pre-trained on similar conditions can be loaded as initial values. Fine-tuning can be performed with only a small number of new samples or short training time, reducing the model reconstruction and adaptation time from "hours" to "minutes".

[0040] 5. Achieve meter-level high-precision positioning: Combining the above advantages, the positioning accuracy of microseismic sources is finally achieved at the meter level, so that the spatial distribution of microseismic events can accurately reflect the surrounding rock fracture zone, providing a reliable basis for the delineation and early warning of high-risk rockburst areas. Attached Figure Description

[0041] Figure 1 This is a diagram of the PINN network architecture in this invention;

[0042] Figure 2 This is a schematic diagram of the dynamic update mechanism in this invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0044] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0045] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0046] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0047] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0048] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, terms such as "set" and "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0049] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0050] A method for real-time location of microseismic sources in deeply buried tunnels includes:

[0051] Obtain relevant basic data for training, including geological survey reports, borehole data, seismic wave test results (such as blasting vibration test) of the tunnel area, and the precise spatial coordinates of the deployed microseismic sensor array (high-precision microseismic sensor arrays are deployed in key locations such as behind the tunnel face according to the monitoring plan, and the signals are synchronously and continuously acquired and transmitted through the data acquisition station).

[0052] Using basic data, a three-dimensional velocity model V(x,y,z) and a three-dimensional computational domain Ω for the tunnel area are first constructed. Then, the physical information neural network PINN is trained. The three-dimensional velocity model V(x,y,z) reflects large-scale structures, lithological interfaces, and the locations of known cavities or solution spaces. The computational domain is defined as follows: the three-dimensional computational domain Ω is defined as: based on the monitoring range, it includes all potential seismic source areas and sensor locations, and records the precise spatial coordinates S of all sensors. i =(x i ,y i ,z i ), i=1,2,...,N;

[0053] Among these, a dedicated physical information neural network PINN is trained for each microseismic sensor. i The network learns a mapping: f i :(x,y,z) → T i (x,y,z), PINN iThe training process uses the spatial coordinates of any point within the three-dimensional computational domain Ω as input and the theoretical P-wave travel time from that point to the corresponding microseismic sensor as output. The training is unsupervised, utilizing physical constraints, boundary conditions, and initial conditions, without requiring any real microseismic event data. This mode is suitable for the initial deployment phase of the system or when events are scarce. After training, a set of efficient travel time field predictors {PINN1,PINN2,...,PINN} is obtained. N It can instantly calculate the travel time from any point within the domain to each sensor;

[0054] The current microseismic waveform data is obtained by real-time acquisition from the sensor array and then filtered (e.g., bandpass filtering to highlight the P-wave frequency band) and denoised. An automatic P-wave first arrival time (t) is identified for each sensor using an automatic P-wave first arrival acquisition algorithm (e.g., based on the STA / LTA ratio or a deep learning pickup). i obs ;

[0055] Then call all the trained PINNs i By combining the double-difference objective function and global grid search, the current spatial coordinates of the earthquake source are obtained.

[0056] The objective function is constructed here based on the principle of travel time difference positioning. To eliminate the unknown absolute time t0 of the earthquake source occurrence, a double-difference objective function is adopted:

[0057] ;

[0058] in, and When sensor waveforms i and j are observed, respectively, For the corresponding observation travel time difference, and PINN i and PINN j The predicted theoretical timeout, Let X = (x, y, z) be the theoretical travel time difference corresponding to the prediction, and let X = (x, y, z) be the source coordinates to be determined.

[0059] Global mesh search is defined as: generating a 3D search mesh within a 3D computational domain, and for each mesh node X... k Call all trained PINNs i Calculate T i (X k Then, the objective function value F(X) is calculated. k ), T i (X k ) represents grid point X k The theoretical timekeeping of sensor i.

[0060] Determining the epicenter location includes finding the grid node X that minimizes F(X). * This is the estimated location of the microseismic event. In X... * Sub-mesh accuracy can be obtained by performing secondary interpolation or local optimization in the vicinity.

[0061] like Figure 2 As shown, a dynamic adaptive update is implemented here to address tunnel excavation. Specifically, this includes: when the tunnel face advances a certain distance, some sensors become too far from the face and need to be moved forward, or when significant corrections are made to the 3D velocity model V(x,y,z) based on new baseline data (mainly); the update process is automatically triggered, updating the PINN sensor trained under the original microseismic sensor positions or similar 3D velocity models V(x,y,z). i Network weights serve as the initial weights for the new network (corresponding to the new sensor location or the new 3D velocity model); Fast retraining: Based on the new sensor coordinates or the updated 3D velocity model V(x,y,z), the network is fine-tuned using a small number of iterations (e.g., only 10%-15% of the initial training rounds). Since the initial point is already near the solution space, convergence is extremely fast. Seamless switching: After fine-tuning, the system immediately switches to the new model for subsequent event localization, achieving continuous and uninterrupted monitoring.

[0062] like Figure 1 The diagram shows the design architecture of the physical information neural network, which fully considers numerical stability, training efficiency, and physical consistency.

[0063] Specifically, it includes:

[0064] 1. Input processing module:

[0065] Input: The three-dimensional coordinates (x, y, z) of the sampling points within the three-dimensional computational domain Ω.

[0066] Coordinate normalization: A linear transformation is performed on all coordinates so that the numerical range of each dimension is mapped to the interval [-1, 1]. This step can significantly improve the stability and convergence speed of neural network training.

[0067] 2. Neural Network Backbone Modeling Module:

[0068] Network type: Fully connected feedforward neural network.

[0069] Input layer: 3 neurons, corresponding to the normalized (x,y,z).

[0070] Hidden layer configuration: 6 to 8 hidden layers are used, each containing 20 to 30 neurons. This deep structure aims to enhance the network's ability to represent complex nonlinear travel-time fields.

[0071] Activation function: The hidden layer uses the hyperbolic tangent function as the activation function. The hyperbolic tangent function is centered at zero, and its gradient is larger near the origin, which helps to alleviate the gradient vanishing problem. Compared with functions such as ReLU, it is smoother and more stable in handling changes in sign.

[0072] Residual connections: Residual connections are introduced between adjacent hidden layers. That is, the output of layer l is not only passed to layer l+1, but also directly added to the activation function of layer l+1 through a shortcut path. This design can effectively solve the gradient decay problem in deep networks, ensuring that the error signal can be smoothly passed to the shallow layers during backpropagation, greatly improving the stability and convergence speed of training.

[0073] 3. Physical constraint embedding and functional equation factorization module:

[0074] Singularity handling: The standard equation ||∇T(x)|| = 1 / V(x) at the source point x s The existence of singularities (infinite gradient) at a certain point makes it difficult and unstable to directly teach the neural network T(x).

[0075] Factoring technique: Introduce a known singularity factor τ0(x) = ||x - x s || / V(x s The total travel time field is decomposed into: T(x) = τ0(x) * u(x). Here, u(x) is a correction term with a value of 1 at the source point that varies smoothly throughout the entire computational domain.

[0076] Network output design: The output layer of the neural network is set to a single neuron, directly predicting the smoothing correction term u(x). The final travel time is calculated through post-processing: T pred (x) = τ0(x) * u pred (x).

[0077] Physical loss construction: Substitute the factorized equation into the loss function. Use automatic differentiation to calculate the gradient ∇u of the network output u(x) with respect to the input x, and then obtain ∇T = ∇(τ0 * u).

[0078] 4. Adaptive Multi-Task Loss Balancing Module:

[0079] The total loss function is composed of:

[0080] L total = λ phy * L phy + λ data * L data + λ bc * L bc+ λ init * L init

[0081] L phy Physical loss. It is defined as the loss at each batch sampling point {x}. j The mean square error on the upper side forces the network to predict the travel time field in all locations to satisfy the basic physical laws of wave propagation.

[0082] L_data: Data loss. When there is known travel time data T obs When, calculate (T) pred - T obs ) 2 The mean.

[0083] L_bc: Boundary loss. Dirichlet or Neumann boundary conditions can be applied to the boundaries of the computational domain.

[0084] L_init: Initial condition loss. Forced at the epicenter x. s At, u(x) s ) = 1 (i.e., T(x) s =0).

[0085] Adaptive Weights (λ): Traditional physical information neural networks use fixed weights λ, which makes it difficult to balance the magnitude of different loss terms and convergence speed. This patent employs an adaptive weight adjustment strategy. For example, the SoftAdapt method based on the relative rate of change of each loss term, or dynamically updating λ according to the proportion of the current values ​​of each loss term. This mechanism allows the model to focus on quickly satisfying the physical equations and boundary conditions (λ) in the early stages of training. phy , λ bc (relatively large), in the later stages of training, as the physical residual decreases, the accuracy of data fitting (λ) is gradually increased. data The pursuit of dynamic equilibrium avoids the tedious manual parameter tuning and automatically finds the optimal convergence path, significantly improving the overall performance and generalization ability of the model.

[0086] 5. Transfer Learning and Dynamic Adaptation Module:

[0087] Problem Background: Tunnels are excavated daily, and sensors need to be moved and repositioned every few days. Traditional methods require complete remodeling and recalculation, which is inefficient. This patent introduces the concept of transfer learning into a physical information neural network framework. When the sensor moves from position S... i Move to new location S i 'At that time, the new fashion show T i '(x) and the old time field T iThe physical equations (function equations) followed by (x) are of the same form, changing only in the boundary conditions (sensor point coordinates). Therefore, the knowledge about the velocity field structure and physical laws learned in the network weights is transferable. The specific operation process is as follows.

[0088] Weight inheritance: Create a new network PINN i Its structure is similar to the original PINN. i Exactly the same. (PinN) i All weight parameters (W, b) are copied to PINN i 'As initialization'.

[0089] Condition update: Set the coordinates S of the new sensor i 'As a new initial condition (u(S) i The input training process is the point corresponding to ') = 1.

[0090] Short-term fine-tuning: Under the new configuration, short-term training (e.g., 1000-3000 epochs) is performed at a high learning rate. Since the network already has good physical priors, it can adjust parameters very quickly to adapt to the new sensor location.

[0091] 6. Seismic source calculation and visualization output module:

[0092] Efficient Search: Utilizing a Trained PINN i Forward propagation is performed, and its computational speed far exceeds that of traditional numerical methods for iterative solutions on the mesh. This makes it possible to perform fast traversal of high-resolution 3D meshes.

[0093] Parallel computing: all PINN i The calculations for grid points can be performed in parallel, further accelerating the solution process.

[0094] Output result: Final output of the three-dimensional coordinates of the seismic source (x) s ,y s ,z s It can be overlaid in real time onto the tunnel's BIM (Building Information Modeling) or geological 3D model, and dynamically displayed as red dots. The system can also simultaneously output information such as positioning error estimation and event energy.

[0095] In this application, the equation is embedded as a physical constraint into the neural network loss function to achieve travel time field prediction without explicit velocity model, and the source singularity is handled by factorization to improve numerical stability.

[0096] This application includes a microseismic monitoring workflow that integrates physical constraints and data-driven approaches. The workflow includes real-time data acquisition and arrival time picking, travel time field modeling based on physical information neural networks, dynamic model updating based on transfer learning, source location calculation and visualization output, and is applicable to dynamic safety monitoring and disaster early warning in deep-buried tunnel construction.

[0097] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for real-time positioning of microseismic sources in deeply buried tunnels, characterized in that, include: Obtain relevant basic data for training, including geological survey reports, borehole data, seismic wave test results, and the precise spatial coordinates of the deployed microseismic sensor array in the area where the tunnel is located; Using basic data, a three-dimensional velocity model V(x,y,z) and a three-dimensional computational domain for the tunnel area are first constructed. Then, the physical information neural network PINN is trained, specifically training a dedicated PINN for each microseismic sensor. i PINN i The training process uses the spatial coordinates of any point in the three-dimensional computational domain as input and the theoretical P-wave travel time from that point to the corresponding microseismic sensor as output. The training process adopts an unsupervised mode and uses physical constraints, boundary conditions and initial conditions for training. The current microseismic waveform data is obtained by real-time acquisition of the sensor array and then filtered and denoised. An automatic P-wave first arrival picking algorithm is used to identify the P-wave first arrival time corresponding to each sensor based on the current microseismic waveform data. Then call all the trained PINNs i By combining the double-difference objective function and global grid search, the spatial coordinates of the current seismic source are obtained.

2. The method for real-time positioning of microseismic sources in deep-buried tunnels according to claim 1, characterized in that, The real-time location method for microseismic sources in deeply buried tunnels also includes the following: when the tunnel face advances a certain distance and some sensors are too far from the face and need to be moved forward, or when the three-dimensional velocity model V(x,y,z) has been significantly modified based on new basic data; the PINN trained under the original microseismic sensor positions or similar three-dimensional velocity models V(x,y,z) is then used. i Network weights, used as the initial weights for a new network; The network is fine-tuned iteratively based on the new sensor coordinates or the updated 3D velocity model V(x,y,z).

3. The method for real-time positioning of microseismic sources in deep-buried tunnels according to claim 1, characterized in that, The three-dimensional velocity model V(x,y,z) reflects large-scale structures, lithological interfaces, and the location of known cavities or solution spaces.

4. The method for real-time positioning of microseismic sources in deep-buried tunnels according to claim 1, characterized in that, The double-difference objective function is: ； in, and When sensor waveforms i and j are observed, respectively, For the corresponding observation travel time difference, and PINN i and PINN j The predicted theoretical timeout, Let X = (x, y, z) be the theoretical travel time difference corresponding to the prediction, and let X = (x, y, z) be the source coordinates to be determined.

5. The method for real-time positioning of microseismic sources in deep-buried tunnels according to claim 4, characterized in that, Global mesh search is defined as: generating a 3D search mesh within a 3D computational domain, and for each mesh node X... k Call all trained PINNs i Calculate T i (X k Then, the objective function value F(X) is calculated. k ), T i (X k ) represents grid point X k The theoretical timekeeping of sensor i.

6. The method for real-time positioning of microseismic sources in deep-buried tunnels according to claim 5, characterized in that, Determining the epicenter location includes finding the grid node X that minimizes F(X). * This is the estimated location of the microseismic event.

7. The method for real-time positioning of microseismic sources in deep-buried tunnels according to claim 6, characterized in that, In X * Sub-mesh accuracy can be obtained by performing secondary interpolation or local optimization in the vicinity.

8. The method for real-time positioning of microseismic sources in deep-buried tunnels according to claim 1, characterized in that, The physical constraints are functional equations obtained after factorization.

Images