Tunnel advance prediction electromagnetic observation system and detection method

Abstract

The present application belongs to the field of tunnel advanced prediction technology, and discloses a tunnel advanced prediction electromagnetic observation system and a detection method. The observation mode of surface emission, underground reception, multi-source emission and multi-component reception is adopted. In this mode, high-power emission can be conveniently used, the problems of high transportation cost in the tunnel and low signal-to-noise ratio of collected data are improved, the coverage area of observation data can be improved, and more effective information can be obtained. Based on the observation system, a tunnel advanced prediction inversion algorithm based on deep learning is developed, quasi-three-dimensional joint inversion of multi-component, time-domain and frequency-domain signals is realized, the advantages of deep learning algorithm are fully utilized, time-domain and frequency-domain electromagnetic signals are fully utilized, the calculation efficiency and detection precision of inversion are improved, the non-uniqueness in inversion is reduced, and rapid imaging of underground space structure is realized.

Description

Technical Field

[0001] This invention belongs to the field of tunnel advance prediction technology, specifically relating to an electromagnetic observation system and detection method for tunnel advance prediction. Background Technology

[0002] Advanced geological prediction for tunnels refers to the use of drilling and modern geophysical exploration to detect the geological conditions ahead of the excavation face of underground engineering projects such as tunnels, underground powerhouses, etc. It aims to understand the structure, properties, and state of the soil and rock mass ahead of construction, as well as geological information such as the occurrence of groundwater, gas, and in-situ stress. Advanced geological prediction can provide a basis for adjusting and optimizing tunnel design parameters, optimizing tunnel construction organization, and formulating construction safety emergency plans, thereby achieving the goals of safety, quality, schedule, environmental protection, and investment control in tunnel engineering, and directly or indirectly creating significant economic and social benefits. Therefore, advanced prediction of disaster sources in tunnel and underground engineering construction is a major requirement for the development of disaster prevention and mitigation science and technology, and also a huge challenge facing researchers.

[0003] Geophysical electromagnetic exploration technology is one of the important technologies for tunnel prediction, and it mainly includes time-domain electromagnetic methods and frequency-domain electromagnetic methods.

[0004] The time-domain electromagnetic method, also known as the transient electromagnetic method (TEM), is a geophysical exploration method based on the significant difference in conductivity between the underground target and the surrounding rock. It uses a grounded conductor or an ungrounded loop to send pulsed or step currents underground. After shutting down the primary field, the secondary eddy current electromagnetic field generated by the underground anomaly is observed through measuring coils. Since the decay law of the induced secondary field is related to the conductivity of the underground geological body—the better the conductivity, the slower the secondary field decays; conversely, the lower the conductivity, the faster the secondary field decays—studying the change of the transient field over time can help detect underground strata, mined-out areas, and karst distribution. The transient electromagnetic method is characterized by high efficiency, strong adaptability to different sites, and sensitivity to low-resistivity, water-filled, fractured zones. These characteristics make it a promising application in tunnel early warning systems.

[0005] Frequency-domain electromagnetic method (FDEM) transmits a time-harmonic excitation signal, and the receiving coil collects the superposition of the secondary field generated by the eddy currents in the strata and the direct-coupled signal, i.e., the primary field. Typically, the direct-coupled signal is larger than the secondary field signal, which reflects the characteristics of the strata; therefore, the primary field signal needs to be eliminated to obtain the secondary field signal caused by the subsurface medium. Since high-frequency signals have weak penetration depth and mainly reflect shallow information, while low-frequency signals have strong penetration depth and mainly reflect deep information, this characteristic gives FDEM the advantage of greater detection depth, making it suitable for crustal-scale studies. In tunnels, due to limited space, strong electromagnetic coupling occurs when the transmitter and receiving coil are close together. Therefore, the transmitter and receiving coil are usually placed on the surface or in the air for detecting the electrical structure of underground tunnels. However, this method only provides relatively low-precision information about the subsurface medium, and often the observation data cannot be effectively covered due to rugged terrain, vegetation cover, etc. Therefore, FDEM is usually used for overall tunnel electrical structure surveys and rarely used for direct advanced prediction.

[0006] Currently, transient electromagnetic prediction mainly employs two types of devices: one involves observation along the tunnel direction within the excavated space to investigate the surrounding rock conditions at the tunnel's top and bottom; the other involves observation at the tunnel face to explore the geological structure ahead of the face. Both methods involve signal transmission and observation inside the tunnel. High-power generators are typically expensive to transport, and their application in areas like coal mines poses a risk of deflagration. Therefore, in these areas, combined batteries are usually used to power the transmitter. Due to limited transmission power, the signal-to-noise ratio is low. Furthermore, the confined space within the tunnel limits the coverage of observation points, resulting in less effective information. This inevitably leads to stronger non-uniqueness in the inversion compared to surface transient electromagnetic prediction. Secondly, under current operating methods, current tunnel transient electromagnetic technology typically collects the magnetic field component or induced electromotive force perpendicular to the coil plane, neglecting electromagnetic field data in other directions. However, the spatial distribution of electromagnetic fields caused by anomalies is three-dimensional, depending on different receiving and transmitting positions and the underground resistivity structure. The sensitivity of electromagnetic field signals to anomalies varies greatly in different directions. Therefore, if only the vertical component is used to detect underground anomalies, important and effective information will inevitably be lost. Furthermore, current TEM tunnel detection is mainly one-dimensional. Since one-dimensional inversion is based on layered media as the forward model, the inversion results are insufficient to describe the three-dimensional non-uniform underground space. Therefore, the inversion results are prone to false anomalies, are unstable, and have poor lateral continuity. If frequency domain electromagnetic methods are applied to tunnel advance detection, strong electromagnetic coupling can easily occur when the distance between the transmitter and receiver coils is close. In addition, regardless of whether it is a time-domain or frequency-domain electromagnetic method, the inversion results of traditional methods are easily affected by the transmitted current, transmitted waveform, and transmitter size, resulting in unstable inversion results. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a tunnel advanced prediction electromagnetic observation system and detection method. It employs a surface-to-underground transmission and reception, multi-source transmission and multi-component reception observation mode, which allows for convenient use of high-power transmission, improving the problems of high transportation costs within tunnels and low signal-to-noise ratio of acquired data. Furthermore, it increases the coverage area of ​​observation data and acquires more effective information. Based on this observation system, a deep learning-based tunnel advanced prediction inversion method is developed to achieve quasi-three-dimensional joint inversion of multi-component, time-domain, and frequency-domain signals. This aims to fully leverage the advantages of deep learning algorithms and make full use of time-domain and frequency-domain electromagnetic signals, improving computational efficiency and detection accuracy during inversion, reducing non-uniqueness in the inversion process, and enabling rapid imaging of underground spatial structures.

[0008] The tunnel advance prediction electromagnetic observation system of the present invention includes a transmitting source and a receiving coil, wherein the transmitting source is located on the ground surface and the receiving coil is located inside the tunnel;

[0009] When the receiving coil receives the signal, it serves as the observation point. At this time, three different transmitting sources are used. The positions of the three transmitting sources are arbitrarily arranged within the range of x,y∈[-30,30], and their directions are set sequentially as x,y,z, forming multiple combinations.

[0010] The transmitter uses positive and negative square wave current to transmit signals in the time domain, generating pulse electromotive force signals, and the receiving coil collects induced electromotive force signals.

[0011] The transmitter uses a time-harmonic excitation signal in the frequency domain, the receiving coil collects the magnetic field signal, and the time-domain data is converted into magnetic field strength in different frequency domains through a fast Fourier transform.

[0012] Furthermore, each transmission and reception collects nine data streams, including three time-domain data streams, three frequency-domain real-part data streams, and three frequency-domain imaginary-part data streams.

[0013] This invention also discloses a method for advanced electromagnetic detection of tunnels, which is based on the above-mentioned observation system and includes the following steps:

[0014] Step 1: Randomly generate a large number of models containing water-rich, low-resistivity anomalies;

[0015] Step 2: Perform a three-dimensional electromagnetic field numerical simulation using the finite volume method to calculate the induced electromotive force and frequency domain magnetic field signal obtained from the receiving coil.

[0016] Step 3: Using simulated observation data as input and the spatial locations of low-resistivity anomalies of different shapes, resistivity values, quantities, and spatial locations as output, establish a U-Net advanced prediction model and train it.

[0017] Step 4: Use the trained U-Net advance prediction model to perform tunnel electromagnetic advance prediction.

[0018] Furthermore, step 1 specifically involves:

[0019] First, a 3D electrical model is established with resistivity varying linearly from shallow to deep. The tunnel and the area above the ground are filled with air. In each numerical simulation, multiple emission sources are randomly set, and the simulation result corresponding to each emission source is a sample, so that multiple samples can be obtained in each forward model. Then, low-resistivity anomalies are randomly added around the tunnel near the receiving coil. Their resistivity values ​​follow a logarithmic uniform distribution log10(ρ)∈[-1,1]Ω·m.

[0020] Furthermore, the U-Net advance prediction model includes Input, convolutional layers, pooling layers, transposed convolutional layers, and Output. The Input layer comprises six layers, with each pair of layers corresponding to the location of an emission source. The Output layer represents the spatial structure of the anomaly. The network has four layers on each side, totaling 16 convolutional layers, 3 pooling layers, and 3 transposed convolutional layers. The ReLU activation function is applied to the output of the convolutional layers. After the convolutional operation, Batch normalization is used to standardize the data.

[0021] Furthermore, the finite volume method is used to perform three-dimensional forward modeling of the resistivity model to construct the input for the U-Net advanced prediction model, specifically:

[0022] For each sample, the input to the U-Net advance prediction model includes three-component time-domain induced electromotive force data, and three-component frequency-domain magnetic field data, frequency, turn-off time, and spatial location channel; the output of the U-Net advance prediction model is the three-dimensional spatial location of the anomaly.

[0023] Furthermore, the spatial location channel includes: azimuth information from the measurement point to the launch point, represented by a three-dimensional vector, and distance information, represented by multiple neurons.

[0024] Furthermore, the output of the U-Net advance prediction model is the three-dimensional spatial location of the anomaly. To conform to the two-dimensional output structure of the neural network, a spherical coordinate system is first constructed with the anomaly center as the target point and the receiver coil position as the origin. Then, θ is used as the abscissa. Let θ be the ordinate. The spatial location information of the model is constructed using the polar angle and azimuth angle of spherical coordinates. Since the actual predicted location of anomalies always has a deviation and a certain degree of uncertainty, a two-dimensional Gaussian distribution is applied around the location of the generated model, and the location of the anomaly corresponds to the highest point.

[0025] The beneficial effects of this invention are as follows: The tunnel advanced prediction electromagnetic observation system of this invention adopts a surface-to-ground transmission and underground-to-underground reception, multi-source transmission and multi-component reception observation mode. During operation, only the measuring coil needs to be placed inside the tunnel for observation, while the bulky transmission system only needs to be placed on the surface, which greatly improves construction efficiency. Multiple signals complement each other; when the tunnel is far from the surface, the time-domain signal is usually weak; the frequency domain belongs to total field measurement, with a high signal-to-noise ratio and relatively weaker influence from distance, but primary field elimination is difficult. Therefore, combining the two methods can form a complementary effect. Furthermore, single observation data has uncertainty, but this system uses multiple types and multiple components of data to jointly determine anomalies, resulting in better stability; frequency-domain electromagnetic methods typically use total field measurement, which can generate significant electromagnetic coupling noise when the transmitting and receiving coils are close together, affecting the inversion results. This system can avoid electromagnetic coupling, greatly reducing electromagnetic coupling noise. When electromagnetic field signals emitted from the Earth's surface reach the receiving point, their paths may cross other areas of electrical anomalies. These anomalies are not targets but may affect the observation results. This system can change the propagation path of the electromagnetic field by altering the source's location, and comprehensively judges the abnormal structures ahead of the tunnel by combining signals from different sources. Furthermore, this invention uses the UNet model to achieve advanced prediction, realizing quasi-three-dimensional joint inversion of multi-component, time-domain, and frequency-domain signals. It fully leverages the advantages of deep learning algorithms and makes full use of time-domain and frequency-domain electromagnetic signals, improving the computational efficiency and detection accuracy of the inversion, reducing non-uniqueness in the inversion process, and enabling rapid imaging of underground spatial structures. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the observation system described in this invention;

[0027] Figure 2 This is a schematic diagram of the structure of the first layer of the input data for the U-Net advanced prediction model;

[0028] Figure 3 This is a schematic diagram of the U-Net advanced prediction model;

[0029] Figure 4 This is a schematic diagram illustrating the evolution of the loss function during the training of the U-Net advanced prediction model;

[0030] Figure 5 This is a schematic diagram showing the prediction results of the anomaly located at different distances from the tunnel excavation face;

[0031] Figure 6 This diagram illustrates the impact of the distance between the anomaly edge and the excavation face, and the anomaly resistivity value, on the prediction results.

[0032] Figure 7 This is a flowchart of the method described in this invention. Detailed Implementation

[0033] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0034] The tunnel advance prediction electromagnetic observation system described in this invention is as follows: Figure 1 As shown, the system adopts a surface-to-surface transmission and underground-to-underground reception mode. For each observation point, three different transmission sources with different positions and directions are used each time to provide richer observation data. In the time domain, positive and negative square wave current transmission signals are used to generate pulse electromotive force signals, and the receiving coil collects the induced electromotive force signals. In the frequency domain, time-harmonic excitation signals are used, and the receiving coil collects magnetic field signals. The time-domain data is converted into magnetic field strengths in different frequency domains through fast Fourier transform. In order to improve the interpretation of data, a three-component (x, y, z) observation mode is adopted, that is, each transmission and reception can collect nine data channels, including three time-domain data channels, three frequency-domain real part data channels, and three frequency-domain imaginary part data channels.

[0035] Based on the aforementioned electromagnetic observation system, this invention proposes a method for advanced electromagnetic detection of tunnels, such as... Figure 7 As shown. Unlike traditional electromagnetic field inversion methods, the research objective of this invention is not the fine distribution of resistivity in underground space, but rather to explore two core issues in near-range advanced prediction: 1. Whether water-rich geological anomalies exist; 2. The approximate location of the anomalies. This simplification of the inversion objective makes it possible to construct representative training data. Since deep learning methods are purely data-driven, their prediction results depend on the characteristics of the training set rather than on physical processes, thus allowing for flexible setting of multiple input data to achieve joint inversion. Based on this, this invention proposes a multi-component, time-domain, and frequency-domain joint inversion based on deep learning to reduce the non-uniqueness of the inversion and improve its efficiency and stability.

[0036] This invention employs the U-Net model for advanced prediction using the electromagnetic tunneling method. To this end, a large number of models containing water-rich, low-resistivity bodies are first randomly generated. Then, based on a three-dimensional finite volume algorithm, the induced electromotive force and frequency-domain magnetic field signals obtained from the receiving coil are calculated. Next, using simulated observation data as input and the spatial locations of low-resistivity anomalies of different shapes, resistivities, quantities, and spatial positions as output, a neural network model is established and trained. The trained model can then achieve rapid mapping from data to the model.

[0037] This invention employs the finite volume method for three-dimensional electromagnetic field numerical simulation. The finite volume method is used to discretize Maxwell's equations, and the equations are numerically solved in both the frequency and time domains. The induced electromotive force curve of the calculated theoretical model is used as input to ResNet. The equations discretized in the frequency domain are:

[0038] Ce+iωb=S m

[0039]

[0040] Where C is the discrete curl operator, e is the electric field, and b is the magnetic field. S is the inner product operator multiplied by the conductivity σ. m and S e Let represent the magnetic source and the electric source terms, respectively. By eliminating 'b', we obtain the equations that depend only on the electric field:

[0041]

[0042] The linear system can be solved using open-source Python libraries (such as SciPy and Mumps) to obtain a solution that satisfies the field source information.

[0043] Similarly, the time-domain equations have the following form:

[0044]

[0045]

[0046] Where C is the discrete curl operator, b is the magnetic flux, and S is the magnetic flux density. m and S e Let these represent the magnetic field and power supply terms, respectively. The Backward Euler algorithm is used for time discretization, and a single time step is expressed in the following form:

[0047]

[0048] Where Δt k =t k+1 -t k It refers to the time step. In forward simulation, the initial condition b is first... 0 (Depending on the source type and the emitted current waveform) It can be obtained through numerical calculation or analytical solution. Then, the magnetic field b at each time step in the entire space can be calculated iteratively using formula (4). k The distribution of .

[0049] Due to their data-driven nature, the performance of deep learning (DL) methods largely depends on the construction of the training set. Therefore, for DL ​​resistivity model inversion, the design of the training set is crucial to the nonlinear mapping capability of the neural network. A well-designed model can improve the predictive power, generalization ability, and robustness of the neural network. To this end, a resistivity model is first established from shallow ρ... s to depth ρ dA linearly varying 3D electrical model is used, with air inside the tunnel and above the ground. In each numerical simulation, multiple emission sources are randomly set to improve the diversity of data. The simulation result corresponding to each emission source is a sample, so multiple samples can be obtained in each forward model. Next, low-resistivity anomalies are randomly added around the tunnel near the receiving coil. Their resistivity values ​​follow a logarithmic uniform distribution log 10(ρ)∈[-1,1]Ω·m.

[0050] The finite volume method is used to perform a three-dimensional forward modeling of the above model to construct the input of the neural network. For each sample, the input of the neural network includes three-component time-domain induced electromotive force data and three-component frequency-domain magnetic field data, including frequency, turn-off time, and spatial information (channel). Figure 2 The diagram shows the directions of the transmitting source (Tx, Ty, Tz), the induced electromotive force (EMF) of the receiving coil (Vtx, Vty, Vtz), and the frequency domain signal of the receiving coil (Hfx, Hfy, Hfz). Each direction contains a real part and an imaginary part. The spatial position channel includes the azimuth information from the measurement point to the transmitting point, represented by a three-dimensional vector, along with distance information represented by multiple neurons. The output of the neural network is the three-dimensional spatial position of the anomaly. To conform to the two-dimensional output structure of the neural network, a spherical coordinate system is first constructed with the anomaly center as the target point and the receiving coil position as the origin. Figure 1 Then, using θ as the x-coordinate, Using the ordinate as the vertical axis, the spatial location information of the model is constructed. Since the actual predicted location of an anomaly always has some deviation and uncertainty, a two-dimensional Gaussian distribution is applied around the generated model's location, with the anomaly's location corresponding to the highest point. The same method can be used to obtain the two-dimensional data labels for the top-down view. This approach can reduce the impact of data centering errors. In the prediction model stage, the accurate location of the anomaly can be obtained through the peak value of the predicted probability distribution.

[0051] Since the inversion target of this invention is only the three-dimensional spatial location of the low-resistivity anomaly, rather than the resistivity distribution of the entire underground space, this greatly simplifies the prediction parameters, makes it possible to construct sufficiently representative training data, accelerates the convergence speed of the neural network, reduces the non-uniqueness of the inversion, and avoids the low generalization ability of the neural network model caused by the limited coverage of observation data.

[0052] This invention aims to utilize the UNet network to predict the nonlinear mapping between TEM signals and the model space. Compared to traditional neural networks, UNet introduces skip connections to address problems such as vanishing and exploding gradients, allowing the network to learn at a deeper level (Ronneberger et al., 2015). The UNet model to be used in this invention is as follows: Figure 3As shown, the Input layer contains six layers, with each pair of layers corresponding to the location of a transmitter. The structure of the first layer of data in a pair is as follows: Figure 2 As shown, the second layer is the result of taking the absolute value of the first layer and then taking the logarithm, in order to increase the proportion of effective information. The specific number of neurons in each layer is shown. The output is the spatial structure of the anomaly. The network has 4 layers on each side, containing a total of 16 convolutional layers, 3 pooling layers, and 3 transposed convolutional layers. The ReLU activation function is applied to the output of the convolutional layers. After the convolution operation, Batch normalization is used to standardize the data to further prevent gradient vanishing or gradient explosion, and at the same time, it can increase the regularization effect. The stride and filter kernel size can be selected as the optimal hyperparameters based on the results of multiple tests.

[0053] Train the constructed model. Figure 4 The evolution of the loss function during training is shown. It can be seen that the error between the training set and the test set gradually decreases with the increase of the number of iterations and eventually tends to stabilize. This shows that the model can learn the features in the training set very well.

[0054] To investigate the performance of the tunnel electromagnetic advance prediction algorithm proposed in this invention, the following was designed: Figure 5 The model shown above depicts a spherical anomaly placed directly in front of the tunnel. Figure 5 The following shows the statistical average of UNet's predictions when the anomaly is located at different distances from the tunnel face. Specifically, a total of 200 simulations were performed. In each simulation, the anomaly's orientation remained directly in front of the tunnel face, the distance from the tunnel face to the anomaly's edge varied randomly within the range of [3, 35] m, the positions of the three surface emission sources varied randomly within the range of [-30, 30] m in x and y coordinates, and the anomaly's resistivity varied randomly between [1, 10] Ω·m. Figure 5 The prediction results show that when the anomaly is within 15m of the tunnel excavation face, UNet can accurately predict its location, size, and shape with high confidence. When the distance is between 15 and 25m, the position of the anomaly predicted by UNet deviates slightly from the actual model, but the overall result is good. When the distance is greater than 25m, two anomalies appear in the statistical results of the UNet prediction model, indicating that there are multiple instances of large prediction errors. This suggests that the prediction accuracy of the algorithm gradually decreases as the distance between the anomaly and the excavation face increases.

[0055] Figure 6 The study further demonstrates the impact of the distance between the edge of the anomaly and the excavation face, as well as the resistivity value of the anomaly, on the prediction results. Figure 6The midpoint represents the difference between UNet's prediction and the true model under the above conditions, which is calculated using the following formula:

[0056] score=(match_dis / max_dis)*max_val

[0057] Where `match_dis` is the distance between the predicted target location and the actual location; `max_dis` is the maximum projected distance of the target location; `max_val` is the prediction probability; and `score` is the final score. It can be seen that when the anomaly is within 15m of the tunnel face, the prediction scores are mostly high, indicating that the algorithm's prediction results are relatively reliable when the anomaly is within 15m of the tunnel face. When the anomaly is more than 15m away from the tunnel face, the reliability of the prediction results decreases significantly.

[0058] The tunnel advance prediction electromagnetic observation system described in this invention specifically addresses the shortcomings of time-domain and frequency-domain electromagnetic methods in tunnel advance detection. For transient electromagnetic detection, this design improves the low efficiency of tunnel TEM advance prediction exploration and the high transportation cost of high-power transmitting motors, thereby increasing construction efficiency. It also addresses the issue of low data coverage in existing observation systems. For frequency-domain electromagnetic detection, this design avoids the defects of electromagnetic coupling between the frequency-domain transmit and receive coils. Furthermore, the system simultaneously acquires time-domain and frequency-domain electromagnetic data, providing a data foundation for joint inversion of multiple data sources and improving detection accuracy. In addition, this system is not only suitable for tunnel advance prediction but also for situations where transmitting devices cannot reach the detection target, such as complex mountainous terrain and mines, demonstrating a certain degree of adaptability.

[0059] The detection method based on the observation system described in this invention can theoretically achieve all-round three-dimensional detection of tunnels, improving the problem of large uncertainties in the original tunnel TEM advanced detection technology based on layered medium theory, and greatly improving the stability, determinism, and computational efficiency of transient electromagnetic advanced prediction. On the other hand, it realizes multi-component, time-domain, and frequency-domain joint inversion, which can not only give full play to the advantages of deep learning, but also make full use of various data to jointly constrain the inversion model, reducing the non-uniqueness of the inversion and improving the inversion efficiency. In addition, compared with traditional electromagnetic detection algorithms, the method described in this invention is less affected by electromagnetic mutual inductance, emission current, emission waveform, emission source size, etc., and the inversion results are more stable.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to further limit the present invention. All equivalent changes made based on the description and drawings of the present invention are within the protection scope of the present invention.

Claims

1. A method for advanced electromagnetic detection in tunnels, characterized in that, The method is based on a tunnel advance prediction electromagnetic observation system. The tunnel advanced prediction electromagnetic observation system includes a transmitter and a receiver coil. The transmitter is located on the ground surface, and the receiver coil is located inside the tunnel. The receiver coil serves as an observation point when receiving signals. Three different transmitters are used, and their positions are arbitrarily arranged within the range of x, y ∈ [-30, 30], with their directions set sequentially as x, y, z, forming multiple combinations. The transmitter uses positive and negative square wave current to emit signals in the time domain, generating pulsed electromotive force signals, and the receiver coil collects the induced electromotive force signals. The transmitter uses time-harmonic excitation signals in the frequency domain, and the receiver coil collects magnetic field signals. The time-domain data is then converted into magnetic field strengths in different frequency domains using a fast Fourier transform. The method includes the following steps: Step 1: Randomly generate a large number of models containing water-rich, low-resistivity anomalies; Step 2: Perform a three-dimensional electromagnetic field numerical simulation using the finite volume method to calculate the induced electromotive force and frequency domain magnetic field signal obtained from the receiving coil. Step 3: Using simulated observation data as input and the spatial locations of low-resistivity anomalies of different shapes, resistivity values, quantities, and spatial locations as output, establish a U-Net advanced prediction model and train it. Step 4: Use the trained U-Net advance prediction model to perform tunnel electromagnetic advance prediction.

2. The tunnel advance prediction electromagnetic detection method according to claim 1, characterized in that, Each transmission and reception process collects nine data streams, including three time-domain data streams, three frequency-domain real-part data streams, and three frequency-domain imaginary-part data streams.

3. The tunnel advance prediction electromagnetic detection method according to claim 1, characterized in that, Step 1 is as follows: First, a 3D electrical model is established where the resistivity changes linearly from shallow to deep. The tunnel interior and the area above the surface are filled with air. In each numerical simulation, multiple emission sources are randomly selected, with the simulation result corresponding to each source serving as a sample, ensuring multiple samples are obtained for each forward modeling iteration. Then, low-resistivity anomalies are randomly added around the tunnel near the receiving coil; their resistivity values ​​follow a logarithmic uniform distribution. .

4. The tunnel advance prediction electromagnetic detection method according to claim 1, characterized in that, The U-Net advance prediction model includes Input, convolutional layers, pooling layers, transposed convolutional layers, and Output. The Input layer contains six layers, with each pair of layers corresponding to the location of an emission source. The Output layer represents the spatial structure of the anomaly. The network has four layers on each side, comprising a total of 16 convolutional layers, 3 pooling layers, and 3 transposed convolutional layers. The ReLU activation function is applied to the output of the convolutional layers. Batch normalization is used to standardize the data after the convolutional operations.

5. The tunnel advance prediction electromagnetic detection method according to claim 4, characterized in that, The input to the U-Net advanced prediction model is constructed by performing a three-dimensional forward modeling of the resistivity model using the finite volume method, specifically as follows: For each sample, the input to the U-Net advance prediction model includes three-component time-domain induced electromotive force data, and three-component frequency-domain magnetic field data, frequency, turn-off time, and spatial location channel; the output of the U-Net advance prediction model is the three-dimensional spatial location of the anomaly.

6. The tunnel advance prediction electromagnetic detection method according to claim 5, characterized in that, The spatial location channel includes: the orientation information from the measurement point to the launch point, represented by a three-dimensional vector, and the distance information, represented by multiple neurons.

7. The tunnel advance prediction electromagnetic detection method according to claim 5, characterized in that, The output of the U-Net advanced prediction model is the three-dimensional spatial location of the anomaly. To conform to the two-dimensional output structure of the neural network, a spherical coordinate system is first constructed with the anomaly center as the target point and the receiver coil position as the origin. The x-axis is... The vertical axis is , , The spatial location information of the model is constructed using the polar angle and azimuth angle of spherical coordinates. Since the actual predicted location of anomalies always has a deviation and a certain degree of uncertainty, a two-dimensional Gaussian distribution is applied around the location of the generated model, and the location of the anomaly corresponds to the highest point.

Images