A method and system for combined electromagnetic scanning detection of tunnels

Abstract

This invention belongs to the field of tunnel advanced prediction technology, and discloses a tunnel electromagnetic joint scanning detection method and system. It introduces a novel tunnel electromagnetic detection system, called TEJS, which realizes three-dimensional joint inversion of multi-component, time-domain, and frequency-domain signals, forming a tunnel joint scanning image for predicting low-resistivity anomalies ahead of the tunnel. The method employs a surface-to-subsurface transmission and reception mode, with multi-source transmission and multi-component reception. During detection, a parallel transmission source moves along the x-axis to scan the subsurface medium, performing surface and subsurface imaging until the entire target area is covered. Based on this observation system, a large number of stochastic models are constructed and numerical simulations are performed. A large training dataset is built using the simulation data to train the UNet model. This model can achieve real-time and rapid imaging of the location of low-resistivity anomalies in three-dimensional space. The algorithm uses surface imaging and subsurface imaging to form a dual-checking mechanism, jointly constraining the three-dimensional spatial location of the anomaly to prevent misjudgment.

Description

Technical Field

[0001] This invention belongs to the field of tunnel advance prediction technology, and particularly relates to a tunnel electromagnetic joint scanning detection method, system, equipment and medium. Background Technology

[0002] Transient electromagnetic (TEM) methods are widely used to identify low resistivity areas before tunnel construction in order to predict severe geological hazards such as water inrush, mudslides, and landslides. This method is based on the significant difference in conductivity between the underground target and the surrounding rock, and studies the changes in the transient field over time to detect underground strata, mined-out areas, and karst distribution.

[0003] Currently, transient electromagnetic prediction mainly employs a central loop device for observation at the tunnel face, transmitting and observing signals inside the tunnel. High-power generators are typically expensive to transport, and their use in areas like coal mines poses a risk of deflagration. Therefore, in these areas, combined batteries are usually used to power the transmitter. However, the limited transmission power results in a low signal-to-noise ratio. Furthermore, the confined space within tunnels restricts the coverage of observation points, leading to less effective information. This inevitably results in stronger non-uniqueness in the inversion compared to surface transient electromagnetic prediction. Secondly, 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 the receiving and transmitting positions and the underground resistivity structure. The sensitivity of electromagnetic field signals to anomalies varies significantly depending on the direction. Therefore, using only the vertical component to detect underground anomalies will inevitably result in the loss of crucial and effective information. In addition, current TEM tunnel detection is mainly one-dimensional. Since one-dimensional inversion is based on layered media as the basis for forward modeling, the inversion results are insufficient to describe the three-dimensional non-uniform underground space. Therefore, the inversion results are prone to false anomalies, and the inversion is unstable with poor lateral continuity.

[0004] Based on the above analysis, the problems and defects of the existing technology are as follows: (1) The existing tunnel advance prediction electromagnetic observation system has low construction efficiency, small observation data coverage area, and low signal-to-noise ratio; (2) Due to the limited tunnel space, the observation data has a weak constraint on the inversion model, and the inversion is mainly based on a one-dimensional layered model, which is insufficient to describe the three-dimensional non-uniform underground space, and is prone to false anomalies. Moreover, the inversion is unstable and has poor lateral continuity. Therefore, the tunnel TEM advance prediction has a large degree of uncertainty and non-uniqueness. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a method, system, device, and medium for tunnel electromagnetic joint scanning detection. This invention aims to introduce a new tunnel electromagnetic detection system, called Tunnel Electromagnetic Joint Scanning (TEJS), which achieves three-dimensional joint inversion of multi-component, time-domain, and frequency-domain signals, forming a tunnel joint scanning imaging mode for predicting low-resistivity bodies ahead of the tunnel.

[0006] This invention is implemented as follows: a tunnel electromagnetic joint scanning detection method, the tunnel electromagnetic joint scanning detection method comprising:

[0007] Step 1: Using the new tunnel electromagnetic detection system, a joint scanning imaging mode for the tunnel is formed. The mode of surface transmission and underground reception, multi-source transmission and multi-component reception is adopted. During the detection process, the parallel transmission source moves along the x-axis to scan the underground medium and perform surface and underground imaging until the entire target area is covered.

[0008] Step 2: Based on the observation system, establish a large number of tunnel resistivity stochastic models, simulate time-domain and frequency-domain electromagnetic field data based on the finite volume method, use the simulated data as the input of the neural network, and use the spatial location of low-resistivity anomalies in the resistivity model as the output to establish training and test sets.

[0009] Step 3: Construct UNet, establish a loss function, preprocess the training data, train the network, adjust the parameters based on the prediction accuracy, and obtain the optimal prediction model.

[0010] Step 4: Import the measured data into UNet to quickly predict the spatial location of low-resistivity water-bearing bodies near the tunnel.

[0011] Furthermore, the tunnel electromagnetic joint scanning detection method uses three parallel surface emission sources to sequentially emit electromagnetic field signals, while a receiving coil operates near the excavation face; the transmitting coil alternately emits two types of signals.

[0012] The first type is a step current signal, used to generate a pulsed electromagnetic field for TEM detection;

[0013] The second type is harmonic signals, and the received signals are converted into frequency magnetic field signals through fast Fourier transform; a three-component (x, y, z) observation mode is adopted.

[0014] Furthermore, in each scanning step of step one, the three emission sources are randomly set in the separated sections (y∈[-30, -10], [-10, 10], [10, 30]) along the parallel direction of the X-axis. Alternatively, one emission source can be used to scan the three separated sections sequentially. The simulated observation data corresponding to the three emission sources in the tunnel electromagnetic joint scanning detection method is used as a training sample.

[0015] Furthermore, the imaging of the low-resistivity water-bearing body in step one is completed based on a deep learning algorithm. The surface imaging refers to the projection of the low-resistivity anomaly into a rectangular area composed of three parallel surface emission sources. The subsurface imaging projects the anomaly onto a two-dimensional coordinate system composed of azimuth and polar angles. By integrating the results of surface imaging and subsurface imaging, the three-dimensional spatial location of the anomaly is finally determined.

[0016] Furthermore, the tunnel electromagnetic joint scanning detection method establishes a deep learning model. First, a three-dimensional resistivity model training set is established, in which the resistivity changes linearly from shallow to deep layers, the tunnel is filled with air, and the ground surface is also filled with air. Low resistivity anomalies are randomly introduced around the tunnel, with the distance from the tunnel excavation face varying between 10 meters and 35 meters. The resistivity value of the anomaly follows a logarithmic uniform distribution log10(ρ)∈[-1,1]Ω·m. The finite volume method is used to perform time-domain and frequency-domain forward modeling calculations to construct a training dataset. The simulated dataset is used to construct and train the UNet model, and the model is optimized by iteratively adjusting the model parameters.

[0017] Furthermore, the input of the neural network of the tunnel electromagnetic joint scanning detection method described in step two includes three-component time-domain induced electromotive force data and three-component frequency-domain magnetic field data, as well as frequency, turn-off time and spatial information channels. The spatial information channel is vector information from the measurement point to the emission source, represented by azimuth and polar angle, and distance in spherical coordinates, represented as three one-dimensional vectors. Two processing methods are applied to the time-domain and frequency-domain data, and the processed data is used as two separate channels: (1) standardization and (2) taking the absolute value and then taking the logarithm. When each emission source emits a signal, data from two channels are collected to improve the proportion of effective information. Therefore, each sample includes a total of six channels of input data.

[0018] Furthermore, the output of the neural network of the tunnel electromagnetic joint scanning detection method described in step two includes two channels, and the output result is the inversion imaging result: (1) underground imaging, a spherical coordinate system is established with the center of the anomaly as the target point and the position of the receiving coil as the origin; then, θ is used as the horizontal coordinate, As a vertical coordinate, it represents the spatial location information of the model. A two-dimensional Gaussian distribution is applied around the generated model location, and the highest point represents the location of the anomaly. In the prediction stage, the precise location of the anomaly is determined by identifying the peak of the prediction probability distribution. (2) Surface imaging: The Gaussian distribution of the underground low-resistivity anomaly projected onto the surface, and the projection area is a rectangular area of ​​the surface composed of three emission sources.

[0019] Furthermore, in step three, the construction and training of the neural network for the tunnel electromagnetic joint scanning detection method utilizes a UNet network to predict the nonlinear mapping between electromagnetic signals and the spatial distribution of low-resistivity anomalies. The model employs the ReLU activation function, applied to the output of the convolutional layer. After the convolution operation, batch normalization is applied to standardize the data. The Sigmoid function is applied to the last layer, limiting the output results to the range [0, 1]. The optimal hyperparameters for stride and filter kernel size will be determined based on multiple tests. The cross-entropy function is used to measure the closeness between the predicted variable q_i(x) and the corresponding label p_i(x).

[0020] H(p,q)=H1(p,q)+H2(p,q)

[0021]

[0022]

[0023] i=1 and i=2 represent the presence or absence of low resistivity anomalies, respectively; x represents the spatial location of the output neuron; and H1 and H2 correspond to the output channels of surface imaging and subsurface imaging, respectively.

[0024] Another object of the present invention is to provide an electromagnetic joint scanning detection system, the tunnel electromagnetic joint scanning detection system comprising:

[0025] The tunnel electromagnetic observation module is used to collect effective electromagnetic field information;

[0026] Training and testing sets are used to train and test stochastic models of tunnel resistivity.

[0027] The prediction model optimization module is used to optimize the model to obtain the optimal prediction model;

[0028] The rapid prediction module is used to quickly predict the spatial location of low-resistivity water-bearing bodies near the tunnel.

[0029] Another object of the present invention is to provide a computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the tunnel electromagnetic joint scanning detection method.

[0030] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the tunnel electromagnetic joint scanning detection method.

[0031] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0032] First, this invention presents a deep learning-based parameter inversion method to improve upon traditional tunnel electromagnetic imaging methods. This method involves a tunnel electromagnetic joint scan (TEJS) observation system, utilizing surface transmission and underground reception, as well as multi-source transmission and multi-component reception modes. This configuration facilitates the use of high-power transmission, thereby improving the coverage and signal-to-noise ratio of the observation data. Based on this system, this invention provides a deep learning-based parameter inversion method for predicting water-bearing structures ahead of the tunnel. The trained neural network can quickly and accurately predict the three-dimensional spatial location of low-resistivity anomalies. Furthermore, the deep learning model provides anomaly imaging from two different perspectives, forming a self-checking mechanism for the prediction, which is particularly important for evaluating the inversion results.

[0033] Secondly, the dual imaging mechanism of this invention can check the reliability of the inversion and identify false anomalies caused by inversion non-uniqueness. This invention introduces a new tunnel electromagnetic detection system called TEJS, and based on this, proposes a parameterized inversion algorithm based on deep learning. It realizes three-dimensional joint inversion of multi-component, time-domain, and frequency-domain signals, forming a joint tunnel scanning imaging mode for predicting low-resistivity bodies ahead of the tunnel. Compared with traditional algorithms, it improves the computational efficiency and detection accuracy of the inversion. In addition, this method forms a self-checking mechanism from surface imaging and subsurface imaging, which can effectively identify false anomalies that may be caused by inversion ambiguity.

[0034] This invention employs a surface-to-underground transmission and underground-to-underground reception observation mode, utilizing multi-source transmission and multi-component reception to achieve omnidirectional three-dimensional tunnel detection. Its advantages include: ① greater convenience and efficiency; ② complementary signals with good stability; ③ avoidance of electromagnetic coupling; ④ cross-verification through multiple transmissions; and ⑤ a dual-imaging inspection mechanism: the 3D spatial location of an anomaly can only be determined when the two predicted imaging results are spatially perfectly matched. This improves prediction accuracy and avoids misjudgments.

[0035] Third, the expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0036] 1. Improve tunnel construction safety: Tunnel electromagnetic scanning detection technology can effectively detect the internal structure and geological conditions of tunnels, providing accurate geological surveys and assessments for tunnel construction and preventing geological disasters during tunnel construction.

[0037] 2. Reduce tunnel construction costs: The application of tunnel electromagnetic joint scanning detection technology can reduce downtime and rework in tunnel construction, avoid unreasonable tunnel structural design, and thus reduce tunnel construction costs.

[0038] 3. Improve tunnel construction efficiency: Tunnel electromagnetic scanning detection technology can achieve rapid and accurate detection of the internal structure and geological conditions of tunnels, shorten the tunnel exploration and design cycle, and improve tunnel construction efficiency.

[0039] The technical solution of this invention solves a long-standing technical problem that has remained unsolved: Essentially, it proposes a novel transient electromagnetic observation system and a parameter inversion method based on deep learning to improve upon the shortcomings of traditional tunnel electromagnetic imaging methods, particularly addressing the drawbacks of TEM tunnel advance detection, such as multiple solutions, instability, and poor continuity. Furthermore, it significantly improves computational efficiency, laying the foundation for real-time tunnel advance prediction.

[0040] Fourth, the following are the significant technological advancements brought about by each claim:

[0041] 1) A brand-new tunnel electromagnetic detection system has been introduced. This system adopts the method of transmitting from the ground and receiving from underground. Compared with the traditional detection method, this detection method, which combines multi-component time-domain and frequency-domain electromagnetic field information, provides more comprehensive and higher-resolution underground information.

[0042] 2) By using three surface emission sources to emit electromagnetic field signals sequentially and alternately emitting two types of signals, the depth and breadth of the detection were enhanced, providing a richer source of data.

[0043] 3) The method of randomly setting emission sources is mentioned. This random strategy can increase the diversity of data, provide more comprehensive training data for subsequent deep learning models, and improve the generalization ability of the models.

[0044] 4) It emphasizes the use of deep learning algorithms for imaging, combining detection technology with state-of-the-art algorithms to make imaging results more accurate and intuitive.

[0045] 5) The process of building a deep learning model was clarified, including the creation of the resistivity model and the simulation of anomalies, which provided a solid data foundation for subsequent neural network training.

[0046] 6) The input structure and data preprocessing methods of the neural network are described in detail, and the data input is optimized to ensure that the network can capture more effective information from the data.

[0047] 7) The content and format of the neural network output are clarified, providing clearer spatial location information of anomalies, making it easier for end users to interpret the results.

[0048] 8) The specific architecture and training method of the neural network are described, ensuring the performance and stability of the model and making the entire detection method more reliable.

[0049] Each claim innovates and optimizes upon the existing technology, bringing significant technological advancements to tunnel electromagnetic detection and improving its accuracy and efficiency. Attached Figure Description

[0050] Figure 1 This is a flowchart of the tunnel electromagnetic joint scanning detection method provided in the embodiments of the present invention;

[0051] Figure 2 This is a schematic diagram of the TEJS observation system provided in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the first layer channel design in the neural network input data provided in this embodiment of the invention;

[0053] Figure 4 The images show the prediction results of the synthetic model UNet provided in this embodiment of the invention. The lower left shows the relative position of the real model and the tunnel. The upper image is a surface scan image, and the lower right image is an underground tunnel image.

[0054] Figure 5 This is a statistical analysis diagram of the prediction results provided in the embodiments of the present invention; (a) correct prediction results; (b) prediction results containing false anomalies; (c) prediction results on average over 100 trials; (d) standard deviation of prediction results. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0056] like Figure 1 As shown, the tunnel electromagnetic joint scanning detection method provided in this embodiment of the invention includes the following steps:

[0057] S101 utilizes a new tunnel electromagnetic detection system to form a tunnel joint scanning imaging mode. It adopts a mode of surface transmission and underground reception, multi-source transmission and multi-component reception. During the detection process, the parallel transmission source moves along the x-axis to scan the underground medium and perform surface and underground imaging until the entire target area is covered.

[0058] S102, based on the observation system, establish a large number of tunnel resistivity stochastic models, simulate time-domain and frequency-domain electromagnetic field data based on the finite volume method, use the simulated data as the input of the neural network, and use the spatial location of low-resistivity anomalies in the resistivity model as the output to establish training and test sets.

[0059] S103, construct UNet, establish the loss function, preprocess the training data, train the network, adjust the parameters based on the prediction accuracy, and obtain the optimal prediction model;

[0060] S104 imports the measured data into UNet to quickly predict the spatial location of low-resistivity water-bearing bodies near the tunnel.

[0061] Each step provides a specific implementation plan:

[0062] 1. **Step One: Joint Scanning Imaging of the Tunnel Electromagnetic Detection System**

[0063] Select appropriate electromagnetic transmitting equipment and place it on the ground to ensure that it can emit stable and powerful electromagnetic signals.

[0064] Multi-component receivers are installed in tunnels or underground to receive electromagnetic signals emitted from the surface from different directions and angles.

[0065] The transmitting device on the ground is controlled to translate along the x-axis at a fixed speed and step distance to ensure that the entire target area is scanned.

[0066] The electromagnetic response data at each point is recorded in the receiver and transmitted to the data processing center in real time.

[0067] 2. **Step Two: Establishment and Simulation of the Resistivity Model**

[0068] Computer-generated random resistivity models were used to represent different subsurface structures and resistivity values.

[0069] The electromagnetic responses of these resistivity models in the time and frequency domains are simulated using the finite volume method.

[0070] Based on the simulation results, training and testing data are prepared for the neural network. The simulated electromagnetic response data is used as input, while the corresponding resistivity distribution (especially the spatial location of low-resistivity anomalies) is used as output.

[0071] 3. **Step Three: Construction and Training of the UNet Neural Network**

[0072] The UNet network structure can be built using existing deep learning frameworks such as TensorFlow or PyTorch.

[0073] Define an appropriate loss function, such as mean squared error loss, to measure the accuracy of the network's predictions.

[0074] Perform necessary preprocessing on the training data, such as normalization and augmentation.

[0075] Train the network using an optimizer (such as Adam or SGD) and use validation data to detect overfitting. Adjust and optimize parameters based on prediction accuracy.

[0076] 4. **Step Four: Predicting Actual Data**

[0077] The actual measured electromagnetic response data is input into the already trained UNet model.

[0078] The network will output the predicted spatial location of the low-resistivity anomaly.

[0079] Based on this location information, the location of water-bearing bodies or other low-resistivity bodies near the tunnel can be determined, thus providing important information for tunnel construction and maintenance.

[0080] This implementation scheme provides a specific technical path for the tunnel electromagnetic joint scanning detection method, but in practical applications, further optimization and adjustment are still needed based on specific geological conditions, equipment performance, and actual needs.

[0081] The tunnel electromagnetic joint scanning detection method provided in this embodiment of the invention includes the following steps:

[0082] Step 1, Observation System: A novel tunnel electromagnetic detection system, called Tunnel Electromagnetic Joint Scanning (TEJS), is introduced. This system achieves three-dimensional joint inversion of multi-component, time-domain, and frequency-domain signals, forming a joint scanning imaging mode for predicting low-resistivity bodies ahead of the tunnel. The method employs a surface-to-subsurface transmission and reception mode, using multi-source transmission and multi-component reception. During the detection process, parallel transmission sources move along the x-axis to scan the subsurface medium, performing surface and subsurface imaging until the entire target area is covered. In each scanning step, three transmission sources are randomly positioned in separate segments (y∈[-30, -10], [-10, 10], [10, 30]) parallel to the x-axis; alternatively, a single transmission source can be used to scan the three segments sequentially.

[0083] Step 2, numerical simulation to establish training set: establish a large number of tunnel resistivity stochastic models, conduct numerical simulation based on the new observation system, use the finite volume method to simulate the electromagnetic field response in the time domain and frequency domain, use the simulation data as the input of the neural network, and use the spatial location of low resistivity anomalies in the resistivity model as the output, including underground imaging and surface imaging, to establish training set and test set.

[0084] Step 3: Construct UNet and establish the loss function; preprocess the training data, then train the network, adjust the parameters based on the prediction accuracy, and obtain the optimal prediction model; import the measured data into UNet to quickly predict the spatial location of possible low-resistivity water-bearing bodies near the tunnel.

[0085] The following details the construction of the training set and the construction, training, and prediction of UNet.

[0086] 1. Training set construction

[0087] First, a three-dimensional resistivity model is established, where the resistivity varies linearly from shallow to deep layers, and the tunnel is filled with air, as is the area above the surface. Then, low-resistivity anomalies are randomly introduced around the tunnel, varying in distance from the tunnel face between 10 and 35 meters. The resistivity values ​​of these anomalies follow a log-uniform distribution log10(ρ) ∈ [-1, 1] Ω·m. Subsequently, time-domain and frequency-domain forward modeling is performed using the finite volume method (Heagy et al., 2017; 2020) to construct a training dataset. Based on this, a UNet model (Ronneberger, 2015) is constructed and trained using the generated dataset. Figure 3 The model is optimized by iteratively adjusting its parameters.

[0088] The neural network input includes three-component time-domain induced electromotive force data and three-component frequency-domain magnetic field data, as well as frequency, turn-off time, and spatial information channels. Figure 3 The spatial information channel comprises vector information from the measurement point to the emission source, represented by azimuth and polar angles, and distance in spherical coordinates, expressed as three one-dimensional vectors where each element of the vector is equal. To reduce the difference in magnitude and retain important information, two processing methods were applied to the time-domain and frequency-domain data. The processed data were then used as two separate channels: (1) normalization and (2) taking the absolute value followed by the logarithm. Considering the presence of three emission sources, each sample data includes a total of six channels ( Figure 3 ).

[0089] Figure 3 The image above shows the channel design of the first layer in the input data of a neural network. Each frequency domain data contains both real and imaginary information. The coordinates of the emission source relative to the observation point are spherical coordinates. The figure below shows the model design of UNet for tunnel advance prediction. The input contains six layers, with each pair of layers corresponding to the location of an emission source. The structure of the first layer of data is shown in the figure above. The second layer is the result of taking the absolute value of the first layer and then taking the logarithm, in order to improve the proportion of effective information.

[0090] The output of the neural network includes two channels: (1) underground imaging. To match the two-dimensional output structure of the network, a spherical coordinate system was established, with the center of the anomaly as the target point and the position of the receiving coil as the origin. Figure 2 Then, using θ as the horizontal coordinate, As vertical coordinates, they represent the spatial location information of the model. Considering the inherent bias and uncertainty of the predicted actual location of the anomaly, a two-dimensional Gaussian distribution is applied around the generated model location, with the highest point representing the location of the anomaly (Zhu and Beroza, 2018). This method effectively mitigates the impact of location errors in the dataset. In the prediction stage, the precise location of the anomaly can be determined by identifying the peak of the probability distribution in the prediction results. (2) Surface imaging: Subsurface imaging only provides the spatial location of the anomaly and lacks information on the distance between the anomaly and the observation point. To overcome this limitation, surface imaging is used to project the location of the anomaly onto the surface, with the imaging location corresponding to the location of three parallel emission sources. By combining surface scanning with subsurface imaging, the spatial location of the anomaly can be determined.

[0091] 2. Construction and Training of Neural Networks

[0092] This invention aims to utilize the UNet network to predict the nonlinear mapping between TEM signals and the model space. Figure 3 The model employs the ReLU activation function, applied to the output of the convolutional layers. After the convolution operation, batch normalization is applied to standardize the data, further preventing gradient vanishing or exploding problems while enhancing regularization. The Sigmoid function is applied to the last layer, restricting the output to the range [0, 1]. The optimal hyperparameters for stride and filter kernel size will be determined through multiple tests. To evaluate the performance of the UNet network, the cross-entropy function (Zhang and Sabuncu, 2018) is used to measure the closeness between the predicted variable q_i(x) and the corresponding label p_i(x).

[0093] H(pq) = H1(pq) + H2(pq)

[0094]

[0095]

[0096] Here, i = 1 and i = 2 represent the presence or absence of low resistivity anomalies, respectively, and x represents the spatial location of the output neuron. H_1 and H_2 correspond to the output channels of surface imaging and subsurface imaging, respectively.

[0097] 3. Model Predictions

[0098] Based on step 2, an effective deep learning prediction model is established. Then, the designed observation system is deployed in the actual tunnel to collect data. The collected data is then imported into the input of the deep learning model to obtain the prediction results of whether there are low-resistivity water-bearing bodies, thus providing important information for judging whether there are safety hazards ahead of the tunnel.

[0099] The following are four specific embodiments and implementation schemes, covering the specific steps of the tunnel electromagnetic joint scanning detection method and the novel tunnel electromagnetic observation system in claim 1:

[0100] Example 1: Tunnel Joint Scanning Detection Method

[0101] A novel observation system combining surface emission and underground observation is designed. Based on time-domain and frequency-domain data obtained from numerical simulation, a deep learning model is established to predict low-resistivity water-bearing anomalies in the three-dimensional space in front of and around the tunnel.

[0102] Example 2: Dual Imaging Mutual Inspection Mechanism

[0103] Neural networks can simultaneously perform subsurface and surface imaging. Subsurface imaging determines the orientation of the anomaly, while surface scanning imaging determines its planar spatial location. Combining the two can accurately estimate the three-dimensional spatial location of the anomaly. If the two imaging results contradict each other, it indicates significant uncertainty in the prediction, requiring adjustments to instrument parameters (emission current, coil size, etc.) or the position of the scanning source's y-coordinate for re-prediction.

[0104] Example 3: Theoretical Model Testing of the Prediction Model

[0105] Designed as follows Figure 4 The model shown in the lower left corner has a spherical anomaly placed directly in front of the tunnel at coordinates [13.5, 4, -3.3] with a resistivity of 1.6 Ω·m. Three parallel transmitting coils are placed on the ground and scan from left to right. Receiving coils are placed at ye[30,10][10,10] and [10,30] to achieve joint detection of the anomaly. Figure 4The image above shows the surface imaging results when the transmitting coil is located in different positions. It can be seen that when the transmitting coil is close to directly above the anomalous object, the anomalous object gradually appears on the right side of the prediction result, and the anomalous object is exactly in the middle at a distance of 15m directly above the anomalous object; as the transmitting coil moves further away from the anomalous object, the anomalous object in the prediction result will eventually disappear. Figure 4 The lower right corner displays the underground imaging results of UNet, and the specific meanings of its horizontal and vertical axes are as follows: Figure 2 As shown, the prediction results indicate that the anomaly is located approximately 12m directly in front of the tunnel, slightly deviating towards the positive y-axis, which is consistent with the actual model. Combining the two prediction results from UNet, this invention can obtain the three-dimensional spatial location of the anomaly. The above results demonstrate that the tunnel electromagnetic advance prediction algorithm proposed in this invention accurately and reliably determines the three-dimensional spatial location of the anomaly by utilizing the prediction results in the XY plane and the orientation information of the low-resistivity anomaly in three-dimensional space.

[0106] Example 4: Uncertainty Test of Prediction Model

[0107] A key advantage of data-driven deep learning inversion methods is their ability to achieve rapid model predictions. This capability facilitates statistical analysis of the inversion results, making it easier to assess their uncertainty. To achieve this, for a model with an anomaly at the location [-8.3, 8.4, -47], this invention simulates 100 different sets of observation data by varying the positions of the emission sources. Three emission sources share a common x-coordinate of 10 meters but have different y-coordinates, different emission currents, and different sizes of emission coils. The tunnel depth is set to 33 meters. Furthermore, this invention applies 5% Gaussian noise to the simulated data. Subsequently, 100 predictions are generated using UNet. Two of these predictions are shown below. Figure 5 As shown in a and b of 5, the white circles represent the locations of the actual anomalies. It is clear that due to the inherent non-uniqueness of inversion, the prediction results differ. It is worth noting that, compared to... Figure 5 Compared to a, Figure 5 The value of 'b' indicates an additional anomalous body on the left. Furthermore, Figure 5 The surface imaging results in b show a shift in the negative Y-axis direction, consistent with the subsurface imaging results. Therefore, it is difficult to determine... Figure 5 Is the anomaly in the negative Y-axis direction of b a real anomaly?

[0108] To address this problem, the present invention averages 100 prediction results, such as... Figure 5As shown in c, the surface imaging results tend to gravitate towards the positive Y-axis direction. This indicates that anomalies along the negative Y-axis direction are spurious anomalies, as they cannot generate surface anomalies in the positive Y-axis direction. Therefore, by performing multiple predictions and averaging them, combined with a self-checking mechanism, spurious anomalies can be effectively identified, mitigating the impact of inversion non-uniqueness and improving the interpretability of the inversion results. Furthermore, Figure 5 The 'd' value represents the standard deviation of 100 predicted outcomes, which is an indicator of the uncertainty of the predictions. This suggests that both true and false anomalies can present uncertainty. However, by considering the average statistical results and using a self-checking mechanism, the presence of true anomalies can be more easily identified.

[0109] The above embodiments and implementation schemes describe the specific steps of the tunnel electromagnetic joint scanning detection method and the novel tunnel electromagnetic observation system. These steps enable scanning of the underground medium and surface imaging, detecting potential low-resistivity water-bearing bodies near tunnels, and providing a more effective and high-precision method for tunnel engineering surveys and geological exploration.

[0110] Based on the tunnel electromagnetic joint scanning detection method of the present invention, the following are specific embodiments and implementation schemes:

[0111] Example 1: Tunnel Electromagnetic Detection Based on a Mobile Vehicle-Mounted Platform

[0112] 1) Vehicle-mounted platform: Uses an automated vehicle-mounted platform with tires, equipped with three ground-based transmitters.

[0113] 2) Transmission signal control: The transmitter is equipped with a microcontroller to control the transmitter source and transmit step current signals and harmonic signals in turn according to the method described above.

[0114] 3) Data reception and transmission: After the receiving coil collects the data, it sends it to the data processing center through the wireless module.

[0115] 4) Real-time deep learning analysis: The data processing center receives data in real time and uses the trained UNet model to make predictions, quickly identifying the spatial location of low-resistivity anomalies.

[0116] 5) Anomaly Alert System: When an anomaly is detected, the vehicle platform will display the location of the anomaly in real time and issue an audible and visual alarm.

[0117] This embodiment provides two different implementation methods for tunnel electromagnetic detection, which can be selected and adjusted according to actual application scenarios and needs.

[0118] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for combined electromagnetic scanning detection of tunnels, characterized in that, include: Step 1: Using the tunnel electromagnetic detection system, a joint scanning imaging mode for the tunnel is formed. The mode of surface transmission and underground reception, multi-source transmission and multi-component reception is adopted. During the detection process, the parallel transmission source moves along the x-axis to scan the underground medium and perform surface and underground imaging until the entire target area is covered. Step 2: Based on the observation system, establish a large number of tunnel resistivity stochastic models, simulate time-domain and frequency-domain electromagnetic field data based on the finite volume method, use the simulated data as the input of the neural network, and use the spatial location of low-resistivity anomalies in the resistivity model as the output to establish training and test sets. Step 3: Construct UNet, establish a loss function, preprocess the training data, train the network, adjust the parameters based on the prediction accuracy, and obtain the optimal prediction model. Step 4: Import the measured data into UNet to quickly predict the spatial location of low-resistivity water-bearing bodies near the tunnel; The specific implementation method of step one is as follows: Select appropriate electromagnetic transmitting equipment and place it on the ground to ensure that it can emit stable and powerful electromagnetic signals; Install multi-component receivers in tunnels or underground to receive electromagnetic signals emitted from the surface from different directions and angles; The ground-based transmitting equipment is controlled to translate along the x-axis at a fixed speed and step size to ensure that the entire target area is scanned. The electromagnetic response data at each point is recorded in the receiver and transmitted to the data processing center in real time. The output of the neural network of the tunnel electromagnetic joint scanning detection method described in step two includes two channels, and the output result is the inversion imaging result: (1) Underground imaging: a spherical coordinate system is established with the center of the anomaly as the target point and the position of the receiving coil as the origin; then, θ is used as the horizontal coordinate and φ is used as the vertical coordinate to represent the spatial position information of the model. A two-dimensional Gaussian distribution is applied around the generated model position, and the highest point represents the position of the anomaly; in the prediction stage, the precise position of the anomaly is determined by identifying the peak of the prediction probability distribution; (2) Surface imaging: the Gaussian distribution of the underground low-resistivity anomaly projected on the surface, and the projection area is a rectangular area of ​​the surface composed of three emission sources; The construction and training of the neural network for the tunnel electromagnetic joint scanning detection method described in step three utilizes the UNet network to predict the nonlinear mapping between electromagnetic signals and the spatial distribution of low-resistivity anomalies. The model employs the ReLU activation function, applied to the output of the convolutional layer. After the convolution operation, batch normalization is applied to standardize the data. The Sigmoid function is applied to the last layer to restrict the output results to the range [0, 1]. The optimal hyperparameters for stride and filter kernel size will be determined based on multiple tests. The cross-entropy function is used to measure the closeness between the predicted variable q_i(x) and the corresponding label p_i(x). i=1 and i=2 represent the presence or absence of low resistivity anomalies, respectively; x represents the spatial location of the output neuron; and H1 and H2 correspond to the output channels of surface imaging and subsurface imaging, respectively.

2. The tunnel electromagnetic joint scanning detection method as described in claim 1, characterized in that, The specific implementation method for step two is as follows: Computer-generated random resistivity models were used to represent different subsurface structures and resistivity values. The electromagnetic responses of these resistivity models in the time and frequency domains are simulated using the finite volume method. Based on the simulation results, training and testing data are prepared for the neural network; the simulated electromagnetic response data is used as input, and the corresponding resistivity distribution is used as output.

3. The tunnel electromagnetic joint scanning detection method as described in claim 1, characterized in that, The specific implementation method for step three is as follows: Utilize existing deep learning frameworks to build the UNet network structure; Define an appropriate loss function, such as mean squared error loss, to measure the accuracy of the network's predictions; Perform necessary preprocessing on the training data; The network is trained using an optimizer, and validation data is used to detect overfitting of the model; parameters are adjusted and optimized based on prediction accuracy.

4. The tunnel electromagnetic joint scanning detection method as described in claim 1, characterized in that, The specific implementation method for step four is as follows: The actual measured electromagnetic response data is input into the already trained UNet model; The network will output the predicted spatial location of the low-resistivity anomaly; Based on this location information, the location of water-bearing bodies or other low-resistivity bodies near the tunnel can be determined, thus providing important information for tunnel construction and maintenance.

5. The tunnel electromagnetic joint scanning detection method as described in claim 1, characterized in that, Three parallel surface transmitters sequentially emit electromagnetic field signals, while a receiving coil operates near the excavation face; the transmitting coil alternately emits two types of signals. The first type is a step current signal, used to generate a pulsed electromagnetic field for TEM detection; The second type is harmonic signals. The received signal is converted into a frequency domain magnetic field signal through a fast Fourier transform and an observation mode with three components (x, y, z) is adopted. In each scanning step of step one, three emission sources are randomly set along the X-axis parallel direction in the separated segments y∈([-30, -10], [-10, 10], [10, 30]). Alternatively, one emission source can be used to scan the three separated segments sequentially. The simulated observation data corresponding to the three emission sources in the tunnel electromagnetic joint scanning detection method is used as a training sample.

6. The tunnel electromagnetic joint scanning detection method as described in claim 1, characterized in that, The neural network input of the tunnel electromagnetic joint scanning detection method described in step two includes three-component time-domain induced electromotive force data and three-component frequency-domain magnetic field data, as well as frequency, turn-off time and spatial information channels. The spatial information channel is vector information from the measurement point to the transmitter, represented by azimuth and polar angle, and distance in spherical coordinates, represented as three one-dimensional vectors. Two processing methods are applied to the time-domain and frequency-domain data. The processed data is used as two separate channels: (1) standardization and (2) taking the absolute value and then taking the logarithm. When each transmitter emits a signal, data from two channels are collected to improve the proportion of effective information. Therefore, each sample data includes a total of six channels of input data.

7. The tunnel electromagnetic joint scanning detection system according to any one of claims 1 to 6, wherein the tunnel electromagnetic joint scanning detection system comprises: Tunnel electromagnetic observation module, used for predicting low-resistivity bodies ahead of the tunnel; Training and testing sets are used to train and test stochastic models of tunnel resistivity. The prediction model optimization module is used to optimize the model to obtain the optimal prediction model; The rapid prediction module is used to quickly predict the spatial location of low-resistivity water-bearing bodies near the tunnel.

Images