A Multi-Level, Air-Ground Integrated Detection Method and System for Karst Tunnels in Complex Terrain

Abstract

This invention belongs to the technical field of tunnel exploration. To address the shortcomings of existing high-resolution tunnel detection methods, it proposes a multi-level detection method and system integrating air and ground in karst tunnels in complex terrain. By jointly inverting airborne electromagnetic data and ground electrical resistivity data, and based on the joint inversion results and the airborne electromagnetic data and ground electrical resistivity data, the detection and identification results are obtained through a detection and identification model. The inversion method can effectively distinguish between deep conductors and shallow low-resistivity interference, reduce the ambiguity of single methods, and achieve high-resolution detection of the actual geological conditions of the tunnel. In the model training, physical constraints are added to improve the accuracy of detection and identification.

Description

Technical Field

[0001] This invention belongs to the technical field of tunnel exploration, and in particular relates to a multi-level air-ground integrated detection method and system for karst tunnels in complex terrain. Background Technology

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

[0003] During tunnel construction in karst areas, frequent occurrences of water inrush, associated mudslides, and secondary disasters pose significant challenges and potential safety risks. Common geological hazards in tunnel construction include fault fracture zones, landslides and sliding bodies, soft rock, expansive rock, karst collapse columns, debris flows, rock bursts, karst collapses, and water and mud inrushes. These geological problems have already triggered numerous accidents, severely impacting the normal progress of tunnel construction and even leading to casualties and economic losses. Therefore, it is essential to employ appropriate exploration methods to ensure the safety of tunnel construction in high-risk karst areas and mitigate the risks and losses caused by water and mud inrushes.

[0004] However, current traditional tunnel exploration technologies cannot achieve detailed detection of the entire tunnel section. Traditional semi-airborne transient electromagnetic detection is limited by the power of the transmitting source and the sensitivity of the receiving coil, resulting in a significant decrease in resolution in deep karst exploration, making it difficult to accurately identify the boundaries of hidden caves and underground rivers, and prone to false anomalies or missed detections. Ground-based high-density electrical resistivity tomography (EDT) inverts the resistivity distribution of underground media by injecting current underground and measuring the surface potential difference. Its accuracy is better than that of semi-airborne transient electromagnetic detection, but due to its high detection cost and the dependence of electrode deployment on terrain, it is not suitable for large-area exploration in practical engineering. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, this invention provides a multi-level air-ground integrated detection method and system for karst tunnels in complex terrain. Through inversion, it can effectively distinguish between deep conductors and shallow low-resistivity interference, reduce the ambiguity of single methods, and achieve refined detection of the actual geological conditions of the tunnel. In model training, physical constraints are added to improve the accuracy of detection and identification.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a multi-level, integrated air-ground detection method for karst tunnels in complex terrain, including: The airborne electromagnetic data and ground electrical resistivity data of the area to be detected are collected, and a joint inversion is performed based on the airborne electromagnetic data and ground electrical resistivity data of the area to be detected to obtain the joint inversion results; Based on the joint inversion results, as well as the airborne electromagnetic data and ground electrical resistivity data of the area to be detected, the detection and identification results are obtained using the trained detection and identification model. The physical constraint function, constructed from the current density field and source current density obtained by inverting electromagnetic data from the training data, is used as the loss function to train the detection and identification model.

[0007] Secondly, this invention provides a multi-layered air-ground integrated detection system for karst tunnels in complex terrain, including: The inversion module is configured to: collect airborne electromagnetic data and ground electrical resistivity data of the area to be detected, and perform joint inversion based on the airborne electromagnetic data and ground electrical resistivity data of the area to be detected to obtain joint inversion results; The detection and identification module is configured to obtain detection and identification results based on the joint inversion results, as well as the airborne electromagnetic data and ground electrical resistivity data of the area to be detected, using a trained detection and identification model. The physical constraint function, constructed from the current density field and source current density obtained by inverting electromagnetic data from the training data, is used as the loss function to train the detection and identification model.

[0008] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0009] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0010] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0011] The above one or more technical solutions have the following beneficial effects: In this invention, by jointly inverting airborne electromagnetic data and ground electrical resistivity data, and based on the joint inversion results and the airborne electromagnetic data and ground electrical resistivity data, the detection and identification results are obtained through a detection and identification model. The inversion method can effectively distinguish between deep conductors and shallow low-resistivity interference, reduce the ambiguity of a single method, and achieve refined detection of the actual geological conditions of the tunnel. In the model training, physical constraints are added to improve the accuracy of detection and identification.

[0012] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is a flowchart of the multi-level refined detection system for complex terrain karst tunnels, integrating air and ground, in Embodiment 1 of the present invention. Figure 2 This is a diagram of the data joint inversion architecture in Embodiment 1 of the present invention. Detailed Implementation

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0017] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0018] Example 1 This embodiment discloses a multi-level, air-ground integrated detection method for karst tunnels in complex terrain, including: The airborne electromagnetic data and ground electrical resistivity data of the area to be detected are collected, and a joint inversion is performed based on the airborne electromagnetic data and ground electrical resistivity data of the area to be detected to obtain the joint inversion results; Based on the joint inversion results, as well as the airborne electromagnetic data and ground electrical resistivity data of the area to be detected, the detection and identification results are obtained using the trained detection and identification model. Among them, the physical constraint function constructed by inverting the current density field and the source current density from the electromagnetic data in the training data is used as the loss function to train the detection and recognition model.

[0019] In this embodiment, the hardware system consists of a ground-based transmitter, a UAV pod receiving system, and a ground-based high-density electrical resistivity tomography (EDT) vehicle. The ground-based transmitter employs a high-power transient electromagnetic transmitter equipped with a high-frequency noise suppression circuit; the transmission line source uses an 80m×40m rectangular array, and multiple transmitters are phase-locked through a GPS clock synchronization unit with a time deviation of <50ns. The UAV is equipped with a dual-band three-component magnetic sensor and an integrated INS / GNSS navigation module, achieving a positioning accuracy of 5cm and an attitude angle error of <0.1°. It uses a six-rotor platform to perform grid scanning at a height of 30m above the ground at a speed of 5m / s, with a flight trajectory pre-set with 10 survey lines covering the target area.

[0020] During the comprehensive survey, the electromagnetic acquisition system continuously recorded full-waveform data at a sampling rate of 256 kSPS, and used a 24-bit Δ-Σ ADC to eliminate 50 Hz power frequency interference. The measured data were processed using real-time wavelet denoising (dB6 wavelet basis, 8 decomposition levels) to generate a three-dimensional conductivity gradient model. When an abnormal resistivity gradient change >18% / m was detected, the ground-based high-density electrical resistivity subsystem was triggered. The ground-based exploration vehicle, equipped with a quadrupedal walking mechanism and an automatic electrode insertion / removal device, deployed a 0.5m × 0.5m grid electrode array in the marked anomaly area. A Winner-Schlumberger combined device (maximum electrode spacing 50m) was used to focus on covering the UAV-detected anomaly area, analyzing the electrical structure at depths of 50-200m, obtaining apparent resistivity profiles, and utilizing adaptive current technology (current dynamic range 0.1-10A) to ensure a signal-to-noise ratio ≥60dB for the measurement signal.

[0021] The transmitting unit includes a transmitter, generator, grounding wire source, and current synchronization unit. It is a ground-based transmitter using a high-power transmitter. The receiving unit includes a receiver and receiving coil. It employs a full-time domain data acquisition method for high-speed, continuous signal acquisition. The acquisition unit is a high-sampling-rate, full-waveform transient electromagnetic acquisition system based on a dual-CPU architecture. This design utilizes a 256ksps sampling rate for high-quality data acquisition and features a 24-bit ADC sampling depth, improving overall data sampling accuracy.

[0022] The drone is a hexacopter equipped with a low-frequency transmitting coil. The transmission current can be adjusted according to the real-time flight altitude. Optionally, a hexacopter drone is used, with a preset flight altitude of 30m to ensure the pod sensor maintains a stable detection distance from the geological body. The survey line spacing is 50m, and the single scan coverage radius is greater than or equal to 500m.

[0023] The ground-based high-density electrical resistivity subsystem employs a multi-parameter detection electrode module, integrating resistivity measurement with an accuracy of ±0.5%, natural potential compensation with an error of <0.5mV, and seismic wave pickup capabilities within a frequency band of 0.1-100Hz. A quadrupedal mobile platform equipped with an automatic electrode insertion and removal mechanism is used to adaptively refine the electrode grid based on the coordinates of the anomaly area initially determined by the semi-airborne TEM, with a minimum grid size of 0.5m × 0.5m.

[0024] The ground-based detection vehicle must strictly cover the anomaly areas marked by the UAV for detection. The maximum electrode spacing of the high-density electrical resistivity tomography (EDT) electrodes is 50m, and the minimum electrode spacing is 0.5m. The minimum electrode spacing is only used in locally encrypted areas.

[0025] The data acquisition device is defined as a comprehensive survey mode, employing a semi-airborne system with preset scanning parameters such as a center frequency of 0.1-10Hz and a flight line spacing of 50-100m for semi-airborne detection. The aerial TEM data is filtered by wavelet transform to generate a conductivity gradient map, which drives the ground-based adaptive transmitter parameter controller to adjust the power supply current, ensuring AB / 2 ≥ 50m. Ground-based synchronous verification of anomaly thresholds, such as resistivity differences > 15%, automatically marks key anomaly areas. High-density electrical resistivity tomography is used in anomaly areas to obtain the spatial distribution information of the apparent resistivity of the subsurface medium, thereby achieving multi-level, refined detection.

[0026] This system employs a combined air-ground detection approach, utilizing a step-by-step strategy of "comprehensive survey - anomaly detection - detailed exploration." First, a comprehensive air-based survey is conducted to analyze potential hazards and adverse geological conditions ahead of the tunnel on a large scale. When an abnormal response is detected due to resistivity gradient changes, a high-density electrical resistivity method is triggered to accurately predict detailed adverse geological bodies. Through multi-modal data joint inversion, combined with a geological risk area intelligent identification model trained using CNN-LSTM deep learning, accurate identification is achieved. This effectively distinguishes between deep conductors and shallow low-resistivity interference, reducing the ambiguity of single methods and enabling detailed exploration of the actual geological conditions of the tunnel. This overcomes the limitations of traditional methods and is suitable for multiple application scenarios.

[0027] The following is a detailed description of the multi-level, air-ground integrated detection method for karst tunnels in complex terrain proposed in this embodiment: For data collected by UAVs, wavelet transform multi-resolution fusion is used to decompose the data using db6 wavelet. For the apparent resistivity data of high-density electrical resistivity methods in the abnormal region, Kriging interpolation is used to generate a continuous apparent resistivity field. The UAV conductivity gradient and the ground apparent resistivity are used as multi-physics field inputs for data fusion.

[0028] In the multimodal data fusion section, wavelet transform multi-resolution fusion is adopted to decompose the high-frequency, low-resolution airborne electromagnetic data into db6 wavelet decomposition with 8 decomposition layers. Cross-validation is performed with low-frequency, high-resolution ground electrical resistivity data to align the low-frequency part of the semi-airborne data with the low-frequency part of the high-density electrical resistivity data and verify consistency. Subsequently, the joint response matrix is ​​reconstructed by weighting the coefficients using the fusion weighting formula.

[0029] The formula for the fusion weight is as follows:

[0030] in, Variance of airborne electromagnetic data. Variance of ground electrical resistivity data; The fusion weights for airborne electromagnetic data; The fusion weights are for ground electrical resistivity data.

[0031] Joint response matrix:

[0032] in, For airborne electromagnetic data; This is ground-based electrical resistivity tomography data.

[0033] Based on the joint response matrix obtained from reconstruction The grid subdivision results of the area to be detected and the background electrical parameters are used to generate an initial three-dimensional electrical model for joint inversion. The joint response matrix is ​​then used as the observation data input for joint inversion to obtain the three-dimensional resistivity distribution and / or current density field of the area to be detected.

[0034] In the joint inversion part, the airborne electromagnetic data after db6 wavelet denoising and the ground electrical resistivity data after gridding are uniformly registered and scaled to construct a joint response matrix. The response values ​​in the joint response matrix are used as the observation data items for joint inversion, and at least one multiphysics parameter from the resistivity three-dimensional distribution, conductivity gradient, and current density field is output through joint inversion.

[0035] The core joint inversion formula is as follows:

[0036] in, The joint inversion objective function; The three-dimensional electrical model to be inverted; is the observed response value of the i-th observation data point in the airborne electromagnetic data; This is the i-th airborne electromagnetic prediction response value calculated based on the current three-dimensional electrical model m; Let be the observed response value of the j-th observation data point in the ground electrical resistivity data; This is the j-th predicted ground electrical resistivity tomography response value calculated based on the current three-dimensional electrical model m; and These are the standard deviations of the data for airborne electromagnetic data and ground-based electrical resistivity data, respectively. and These are the data weighting coefficients corresponding to airborne electromagnetic data and ground-based electrical resistivity data, respectively. The regularization coefficient is used. Three-dimensional electrical model Spatial gradient; This represents the total number of data points for space-based electromagnetic observations. This represents the total number of ground-based electrical resistivity tomography (EDT) observation data points.

[0037] The joint response matrix is ​​used to uniformly characterize the electrical responses of airborne electromagnetic data and ground-based electrical resistivity data on the same spatial grid or corresponding observation locations. During joint inversion, each response value in the joint response matrix is ​​treated as an observed data item, and the data weights are determined based on the data type, data variance, and detection scale corresponding to each response value. By minimizing the weighted residuals between the observed responses and the corresponding forward responses in the joint response matrix, and combining this with regularization constraints of the three-dimensional electrical model, the three-dimensional resistivity distribution, conductivity gradient, and / or current density field of the region to be detected are obtained.

[0038] By taking on-site photographs of geological risk areas at different levels, a large number of typical photos of these areas were collected. The image data was preprocessed and labeled to create corresponding annotation files, forming a dataset. The risk level labels, ranging from level one to four, were manually annotated, and PointNet++ was used to capture the microstructure of the terrain. The dataset was then input into a CNN-LSTM hybrid network model, and the parameters of the CNN-LSTM hybrid network model were adjusted based on the results to train a model for detecting and identifying geological risk areas.

[0039] This embodiment integrates CNN, LSTM, attention mechanism, physically constrained neural network, and hierarchical feature distillation technology to achieve intelligent identification of complex risk factors. Specifically, it incorporates geophysical parameters into the CNN-LSTM architecture. In the CNN convolutional layers, current density, resistivity gradient, etc., are converted into physical features and fed into the network along with the original input data. This forces the convolutional kernels to prioritize learning features consistent with physical laws, thereby constraining the kernel weight allocation as prior knowledge. The attention mechanism involves setting a feature layer between the CNN and LSTM layers, generating channel weights through global average pooling to automatically capture and focus on high-risk areas for identification.

[0040] Specifically, the input to the CNN includes the resistivity 3D distribution, conductivity gradient, and / or current density field obtained by joint inversion based on the joint response matrix. The CNN part uses a 3D convolutional layer with a kernel size of 5×5×3 to extract spatial electrical features, including resistivity 3D distribution features, conductivity gradient features, and geophysical anomaly features related to karst development. A bidirectional LSTM unit with 256 hidden nodes is used as input, taking transient electromagnetic decay curves and multiple detection data (i.e., observation sequence data of transient electromagnetic and ground electrical methods) to capture time-varying responses. The spatial electrical features extracted by the CNN and the time-varying response features extracted by the LSTM are input into the feature layer, which then serves as input data for the attention mechanism, focusing on the boundaries of high-risk areas. During the training of the detection and identification model, a physical constraint loss function based on the current density field and source current density ensures that the output of the detection and identification model conforms to geophysical laws, avoiding misjudgments due to non-physical risks.

[0041] The attention mechanism in the system consists of two parallel parts: spatial attention and channel attention. Spatial attention calculates the average and maximum values ​​of the feature maps in the spatial dimension, and then generates spatial weights through 7×7 convolutions to reweight the spatial features. Channel attention generates channel weights through global average pooling and max pooling, followed by processing through a multilayer perceptron with shared weights, thereby emphasizing the features of important channels. These two attention mechanisms complement each other to comprehensively capture risk features.

[0042] CNN uses the ResNet architecture, which includes an input layer, pooling layers, convolutional layers, and fully connected layers. There are 146 convolutional layers, all connected to ReLU layers, and each convolutional kernel has different weight parameters, calculated as follows:

[0043] Where Y is the output feature, which serves as the input to the next layer; a is the ReLU activation function; and X is the input feature, which serves as the output of the previous layer. is the weight of the i-th convolutional kernel; E is the number of channels.

[0044] LSTM can implement the decoder function. In this embodiment, the number of memory units and gating units is set to 256, the number of hidden layers is 256, the learning rate parameter is 0.01, the optimizer type is gradient descent algorithm, and the gating unit is Tanh. To prevent overfitting, this module randomly shuts down neurons between LSTM layers, and the dropout rate is set to 0.3.

[0045] The CNN-LSTM hybrid network model training adopts a transfer learning strategy. By embedding geomechanical equations and resistivity gradient priors into the pre-training stage, approximately 1.2×10^6 adjustable parameters are pre-trained in the network, enabling the model to learn physically interpretable features in the initial stage, thereby solving the small sample problem. Furthermore, a lightweight model is deployed using an edge computing device with a computing power of 4 TOPS and a power consumption of 15W.

[0046] This example first converts the terrain data into a 10m resolution digital elevation model, sets the slope standard deviation threshold to 15°, and simultaneously overlays gravity and magnetic gradient tensor data. The vertical second derivative calculation accuracy is 0.05nT / m², and finally outputs a four-color risk level map, which effectively identifies whether there is a risk. The model identification effect is good.

[0047] A hybrid training method using Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) was employed. Multimodal data preprocessing was used to rasterize the spatial data, converting LiDAR point cloud data such as terrain undulation and slope features into 10m×10m raster.

[0048] In the model training of this embodiment, hierarchical feature annotation is used to divide the risk area into coarse-grained regions. By taking on-site photos of geological risk areas of different levels, a large number of typical photos of geological risk areas are collected. The image data is preprocessed and annotated to create corresponding annotation files and form a dataset.

[0049] Hierarchical feature annotation mainly involves extracting topographic microstructure features and electrical conductivity features from the aforementioned joint inversion data. The joint inversion results include multi-physical field parameter data such as resistivity three-dimensional distribution and electrical conductivity gradient. The inversion results are used as input for hierarchical feature annotation. Risk areas are divided and labeled according to the four-level risk area division principle. The annotation is based on the aforementioned four-level topographic features and electrical characteristics, thereby achieving the function of dividing finely explored areas.

[0050] The hierarchical labeling consists of three layers: the first layer divides the entire exploration space into blocks with different risk levels; the second layer identifies spatially connected anomalies within the same risk level, such as cave systems and underground river channels; and the third layer accurately marks the geometric boundaries of the anomalies and extracts microscopic sensitive indicators inside the anomalies as auxiliary monitoring signals.

[0051] Specifically, the first layer: based on the three-dimensional resistivity distribution, conductivity gradient, topographic data, regional geological maps, and known hazard locations obtained through joint inversion, the geological risk level is classified according to the intensity of karst development, tectonic activity, hydrogeological conditions, and the probability of potential hazards. Specifically: Level 1 risk area (extremely high risk): Karst development intensity ≥ Level 3, i.e., underground river channel density > 5 lines / km², karst cave volume ratio > 15%; and located at the intersection of the main structural fault zone, i.e., fault fracture width > 50m; Level II risk area (high risk): moderately developed karst, i.e., cave volume ratio of 8%-15%; and within 50-200m of the main fault; Level 3 risk area (medium risk): dissolution fracture zone is well developed, i.e., linear karst rate is 5-8%; and the joint density in the area is 3-5 joints / m². Level 4 risk area (low risk): micro-karstification zone, i.e., karstification rate <3%; and rock mass integrity coefficient >0.75.

[0052] The second layer: For high-risk and extremely high-risk areas in the first-level annotation, a 26-neighbor 3D connected component analysis is used to classify spatially adjacent voxels (i.e., those sharing a face, edge, or vertex) as the same connected component.

[0053] The second level of annotation identifies spatially connected anomalous targets at the same risk level, such as a single cave or connected karst conduits, to guide the model in learning the completeness of the target.

[0054] The third layer: Based on the annotation of the second layer, the micro-sensitive indicators inside the anomaly body are calculated, including the slope aspect standard deviation, the frequency of historical disasters, the density of dissolution fissures and the resistivity anomaly gradient. These micro-sensitive indicators are used as auxiliary supervision signals for model training to enhance the model's ability to perceive risk details.

[0055] Design a multi-stage loss function, including a classification loss function. Physical constraint loss function Consistency loss function with risk gradient The multi-objective loss design employs the region classification loss method to address the class imbalance problem. The multi-objective loss function formula is as follows: +β

[0056] in, , β and β are the weight coefficients corresponding to the classification loss function, physical constraint loss function, and risk gradient consistency loss function, respectively. For example, =2, β=1, =1; In actual training process , β can be adjusted according to the sample category distribution, physical constraint strength, and risk boundary continuity requirements.

[0057] Among them, the classification loss function for:

[0058] in, For the first i The true label of each sample The probability of the model predicting a level four risk area; This represents the number of samples.

[0059] Risk gradient consistency loss function for:

[0060] in, This is a reference risk level field obtained based on joint inversion results or manual annotation; The risk level field predicted by the model; and These are the spatial gradients of the reference risk level field and the predicted risk level field, respectively; by minimizing the difference between the two, the model is constrained to ensure that the changes in the predicted risk gradient are consistent with the actual risk gradient trend.

[0061] Physical constraint loss function based on geomechanical equations and resistivity gradient prior. for:

[0062] Where J is the current density field, obtained by electromagnetic data inversion; q is the source current density, with known emission parameters; Ω is the model calculation domain, such as a range of 500m on both sides of the axis of the test area; The volume of the computational domain Ω or the total volume of the discrete mesh elements is used to normalize the physical constraint residuals. This represents the divergence.

[0063] In this embodiment, geological and physical laws are embedded into the model training, thereby improving the model's scientific rigor, generalization ability, and engineering applicability.

[0064] Hierarchical distillation is a technique that optimizes model performance by progressively transferring knowledge from higher to lower levels. Its core idea is to gradually pass abstract semantic information from higher-level features to lower levels, enhancing the model's ability to perceive complex patterns. Specifically, it employs an attention mechanism to extract global risk features.

[0065] As the model trains, the weights of each loss component are dynamically adjusted. Initially, physical constraints are emphasized to ensure the model's rationality, while later, the classification loss weights are gradually increased to improve prediction accuracy. After iterative loss and CNN-LSTM analysis, risk level labels are output, thus achieving data stratification. The data stratification process follows the workflow of "raw data → risk prediction → stratification threshold division → hierarchical labeling → visualization".

[0066] As an optional implementation, Transformer encoding and 3D convolutional networks are used to fuse temporal data and geological LiDAR point clouds, and PointNet++ is used to capture the microstructure of the terrain. A fine-tuning strategy is employed, specifically optimizing the multi-objective loss function using geological datasets, performing distillation using a teacher-student model, setting the pre-trained cloud model to a lightweight mesh, minimizing the multi-objective loss function, adjusting the model input resolution, compressing multi-level features into a lightweight model, and deploying it on edge computing devices.

[0067] This embodiment performs joint inversion on airborne electromagnetic data and ground electrical resistivity data. Based on the joint inversion results and the airborne electromagnetic data and ground electrical resistivity data, the detection and identification results are obtained through a detection and identification model. The inversion method can effectively distinguish between deep conductors and shallow low-resistivity interference, reduce the ambiguity of a single method, and achieve refined detection of the actual geological conditions of the tunnel. In the model training, physical constraints are added to improve the accuracy of detection and identification.

[0068] Example 2 The purpose of this embodiment is to provide an integrated air-ground multi-level detection system for karst tunnels in complex terrain, including: The inversion module is configured to: collect airborne electromagnetic data and ground electrical resistivity data of the area to be detected, and perform joint inversion based on the airborne electromagnetic data and ground electrical resistivity data of the area to be detected to obtain joint inversion results; The detection and identification module is configured to obtain detection and identification results based on the joint inversion results, as well as the airborne electromagnetic data and ground electrical resistivity data of the area to be detected, using a trained detection and identification model. Among them, the physical constraint function constructed by inverting the current density field and the source current density from the electromagnetic data in the training data is used as the loss function to train the detection and recognition model.

[0069] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0070] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0071] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0072] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0073] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0074] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0075] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0076] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0077] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0078] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0079] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-level, air-ground integrated detection method for karst tunnels in complex terrain, characterized in that: include: The airborne electromagnetic data and ground electrical resistivity data of the area to be detected are collected, and a joint inversion is performed based on the airborne electromagnetic data and ground electrical resistivity data of the area to be detected to obtain the joint inversion results; Based on the joint inversion results, as well as the airborne electromagnetic data and ground electrical resistivity data of the area to be detected, the detection and identification results are obtained using the trained detection and identification model. The physical constraint function, constructed from the current density field and source current density obtained by inverting electromagnetic data from the training data, is used as the loss function to train the detection and identification model.

2. The multi-level air-ground integrated detection method for karst tunnels in complex terrain as described in claim 1, characterized in that, The physical constraint function for: Where J is the current density field obtained from electromagnetic data inversion, q is the source current density, and Ω is the computational domain of the model. To calculate the volume of the domain Ω or the total volume of the discrete mesh elements for the model, This represents the divergence.

3. The multi-level air-ground integrated detection method for karst tunnels in complex terrain as described in claim 1, characterized in that, Based on the joint inversion results, as well as the airborne electromagnetic data and ground electrical resistivity data of the area to be detected, the detection and identification results are obtained using the trained detection and identification model, specifically: Based on the joint inversion results, CNN is used to capture spatial electrical features; Based on airborne electromagnetic data and ground electrical resistivity data of the area to be detected, a two-way LSTM is used to capture the time-varying response. Based on spatial electrical characteristics and time-varying response, the detection and identification results are obtained by focusing on the boundary of high-risk areas through an attention mechanism.

4. The multi-level air-ground integrated detection method for karst tunnels in complex terrain as described in claim 1, characterized in that, The training of the detection and recognition model is specifically as follows: Based on karst development intensity, tectonic activity, hydrogeological conditions, and potential disaster probability, the exploration space is divided into four risk level labels, completing the first layer of labeling; Based on the first layer of annotation, identify and annotate connected anomalies within each risk area to complete the second layer of annotation; Based on the second layer of annotation, the geometric boundaries of the anomaly are marked, and micro-sensitive indicators are extracted; the micro-sensitive indicators include the slope aspect standard deviation, the frequency of historical disasters, the density of dissolution fissures, and the resistivity anomaly gradient; The detection and recognition model is trained using hierarchically labeled training data.

5. The multi-level air-ground integrated detection method for karst tunnels in complex terrain as described in claim 1, characterized in that, A joint inversion was performed based on airborne electromagnetic data and ground-based electrical resistivity data of the area to be detected, yielding the following joint inversion results: The fusion weights of the airborne electromagnetic data and the fusion weights of the ground electrical resistivity data are determined based on the variances of the airborne electromagnetic data and the ground electrical resistivity data, respectively. A joint response matrix is ​​constructed based on the fusion weights of airborne electromagnetic data and ground electrical resistivity data. The response values ​​in the joint response matrix are used as the observation data for joint inversion. At least one physical field parameter from the three-dimensional resistivity distribution, conductivity gradient, and current density field is output through joint inversion.

6. The multi-level air-ground integrated detection method for karst tunnels in complex terrain as described in claim 1, characterized in that, The loss function of the detection and identification model includes a classification loss function, a risk gradient consistency loss function, and a physical constraint loss function.

7. A multi-level air-ground integrated detection system for karst tunnels in complex terrain, characterized in that: include: The inversion module is configured to: collect airborne electromagnetic data and ground electrical resistivity data of the area to be detected, and perform joint inversion based on the airborne electromagnetic data and ground electrical resistivity data of the area to be detected to obtain joint inversion results; The detection and identification module is configured to obtain detection and identification results based on the joint inversion results, as well as the airborne electromagnetic data and ground electrical resistivity data of the area to be detected, using a trained detection and identification model. The physical constraint function, constructed from the current density field and source current density obtained by inverting electromagnetic data from the training data, is used as the loss function to train the detection and identification model.

8. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-6.

Images