Tunnel unfavorable geological physical field-hydrological field fusion holographic detection method and system

Abstract

The present application belongs to the technical field of underground engineering unfavorable geological disaster prediction and intelligent control, and provides a tunnel unfavorable geological physical field-hydrological field fusion holographic detection method and system, which uses multi-physical field detection information such as induced electric field, seismic electric field, natural electric field and hydrological field as fusion data source, establishes a cross gradient inversion as a coupling objective function with physical constraint conditions, constructs a CNN-GNN-Transformer hybrid network for multi-level feature fusion, dynamically balances the contribution of the physical constraint conditions in the coupling objective function and the deep learning data-driven feature learning ability of the hybrid network through adaptive weight, realizes multi-field holographic interpretation based on the hybrid deep learning model, improves the fusion degree of multi-source heterogeneous data, can realize three-dimensional holographic imaging and water gushing prediction of sudden water gushing disaster source, and improves the accuracy of unfavorable geological disaster detection and the reliability of intelligent decision-making.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of underground engineering unfavorable geological disaster prediction and intelligent control, and particularly relates to a tunnel unfavorable geological physical field-hydrological field fusion holographic detection method and system. BACKGROUND

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

[0003] In underground engineering such as tunnels, mines and water conservancy, water and mud inrush disasters have the characteristics of strong suddenness, great destructive power and difficult prediction, which seriously threaten construction safety. The formation mechanism of such disasters is complex, involving the coupling of multiple physical processes such as groundwater seepage, rock-soil mass failure and stress redistribution.

[0004] Existing detection methods mainly include resistivity method, induced polarization method, transient electromagnetic method, geological radar method, etc., which mainly use single physical field for anomaly recognition, each has its own advantages but has significant limitations: the spatial resolution of the induced electric field is limited, the seismic electric field is easily disturbed by noise, the quantitative degree of the natural electric field is not high, and the coverage range of the hydrological field is limited.

[0005] With the continuous expansion of underground engineering and the increasingly complex geological conditions, unfavorable geological disaster sources show strong concealment, multiple types of superposition and spatio-temporal evolution, and single information has been difficult to support accurate prediction, so it is urgent to rely on comprehensive analysis and intelligent processing of multi-source information.

[0006] However, the existing comprehensive methods still have problems:

[0007] On the one hand, the data format, dimension and sampling frequency of different physical fields (such as induced electric field, seismic electric field, natural electric field and hydrological field) are different, and direct connection is difficult; some have high resolution but small range (such as induced electric field), and some have large coverage but low precision (such as hydrological field), which leads to inconsistency after fusion. Moreover, different physical fields are disturbed in different ways and have large differences in noise characteristics, which can easily cause errors during fusion.

[0008] On the other hand, there is a lack of effective fusion mechanism: traditional methods generally use simple superposition or parallel processing, lack a unified holographic detection framework, the internal relationship between physical fields (such as the correspondence between electric field anomalies and seepage conditions) is not fully modeled, and the coupling is insufficient; for boundary fuzzy and complex structure abnormal bodies, spatial positioning and shape description are not accurate enough.

[0009] Finally, a single model is difficult to capture local features, topological relationships and global dependencies at the same time, feature extraction is not sufficient, and it is difficult to balance prediction accuracy and applicability, and some technical solutions rely on complex algorithms, although they improve prediction accuracy, but they are time-consuming and difficult to apply to the field. SUMMARY

[0010] In order to solve the above problems, the present application provides a tunnel adverse geological physical field-hydrological field fusion holographic detection method and system, which uses induced electric field, seismic electric field, natural electric field, hydrological field and other physical field detection information as fusion data source, realizes multi-field holographic interpretation and intelligent decision method with hybrid deep learning model as carrier, improves the fusion degree of multi-source heterogeneous data, can realize three-dimensional holographic imaging and water gushing prediction of water gushing disaster source, and improves the accuracy of adverse geological disaster detection and the reliability of intelligent decision.

[0011] According to some embodiments, the present application adopts the following technical scheme:

[0012] A tunnel adverse geological physical field-hydrological field fusion holographic detection method, comprising the following steps:

[0013] Obtain geological data and drilling data of the detection area, and perform induced electric field, seismic electric field, natural electric field and hydrological field detection on the potential water gushing area in the potential water gushing area, to obtain multi-source detection data;

[0014] Preprocess the multi-source detection data, and in the preprocessing process, introduce a time-space synchronization control method to align the multi-source detection data to a unified time-space coordinate system to form a standardized time-space data structure;

[0015] Establish a cross-gradient inversion as a coupled objective function of physical constraint conditions;

[0016] Construct a CNN-GNN-Transformer hybrid network for multi-level feature fusion, wherein the CNN is used to extract local nonlinear response features and process spatial changes of multiple physical fields, the GNN is used to model the spatial topological relationship between physical field sources and reflect the structural constraints of cross-gradient, and the Transformer is used to capture long-range dependencies and global features;

[0017] Balance the contribution of the physical constraint conditions in the coupled objective function and the deep learning data-driven feature learning ability of the CNN-GNN-Transformer hybrid network through adaptive weights;

[0018] Construct the gradient correlation between different physical fields, learn the nonlinear mapping relationship between different physical fields through the CNN-GNN-Transformer hybrid network, and make predictions;

[0019] The prediction results of the CNN-GNN-Transformer hybrid network are compared with the calculation results of the coupled objective function. When the deviation between the two is greater than a set value, a local adjustment is triggered to correct the abnormal region, update the parameters of the CNN-GNN-Transformer hybrid network, and adjust the adaptive weights until the prediction results meet the requirements.

[0020] Based on the final prediction results, a three-dimensional spatial model of the geological structure is created, and structural visualization and risk zoning are performed in combination with physical property parameters to construct a spatial distribution model of disaster sources.

[0021] As an alternative implementation, the multi-source detection data includes corresponding detection data of induced electric field, seismic electric field, natural electric field and hydrological field.

[0022] As an alternative implementation, the preprocessing of the multi-source detection data includes: noise filtering and outlier removal of the multi-source detection data, using a streaming normalization integration method, and eliminating the differences in dimensions and scales of different detection data through standardization processing.

[0023] As an alternative implementation, the physical constraints are: ;

[0024] in A Let be the forward operator matrix of a certain physical field. To account for the difference between observed data and theoretical response, This represents the correction amount for the geophysical model parameters. For regularization terms, For the cross-gradient constraint term of the induced electric field and the hydrological field, The cross-gradient constraint term of the seismic electric field and the natural electric field, For multi-field integrated cross-gradient constraint terms, Used to represent the difference in cross gradients associated with the initial geophysical model;

[0025] The aforementioned physical constraints are integrated into the coupled objective function as basic constraints, and the synergistic utilization of different detection information is achieved through multi-physics joint inversion.

[0026] As an alternative implementation, the coupling objective function is:

[0027] ;

[0028] in, d For the observed data vector, f(m) This represents the forward response, where m is the model parameter. L 1 、L 2 represents the smoothness constraint operator, and λ1, λ2, λ3, λ4, and λ5 are regularization parameters.B 1 、B 2 、B 3 is a cross-gradient constraint operator matrix of different physical field combinations: B 1 is a cross-gradient operator matrix of the induced polarization field-hydrological field, B 2 is a cross-gradient operator matrix of the seismic-electric field-natural electric field, B 3 is a multi-field integrated cross-gradient operator matrix.

[0029] As an optional implementation, the process of dynamically balancing the contribution of the physical constraint conditions in the objective function and the deep learning data-driven feature learning ability of the CNN-GNN-Transformer hybrid network by adaptive weights includes:

[0030] F_fusion = α1F_CNN + α2F_GNN + α3F_Transformer;

[0031] Wherein, F_CNN is a local nonlinear response feature vector extracted by CNN, mainly capturing the spatial distribution law and local anomaly information of each physical field detection data; F_GNN is a spatial topological relationship feature vector extracted by GNN, modeling the spatial connection relationship between physical field sources, reflecting the structural constraint of cross-gradient; F_Transformer is a long-range dependence and global feature vector extracted by Transformer, capturing the global correlation of large-scale, multi-temporal and spatial span data, α1, α2, α3 are used to dynamically adjust the weight coefficients, satisfying α1+α2+α3= 1, and are adaptively adjusted according to the training process and data characteristics, to realize the optimal balance between physical constraints and data-driven.

[0032] As an optional implementation, the process of modifying the abnormal area includes modifying using a loss function, and the loss function L is:

[0033] L = L_pred + α·L_residual + β·L_consistency;

[0034] Wherein, L_pred is a prediction loss, L_residual is a residual loss, L_consistency is a consistency loss, α and β are weight coefficients, and the adaptive update of parameters is realized by a gradient descent algorithm.

[0035] As an optional implementation, according to the final prediction result, the process of three-dimensional spatial modeling of geological structure and structure visualization and risk zoning combined with physical parameters includes: presetting different threshold values of physical parameters, when the corresponding physical parameter is greater than the threshold value of the corresponding level, it is considered that there is a risk of the corresponding level, and different levels of risk adopt different construction schemes.

[0036] A tunnel bad geological physical field-hydrological field fusion holographic detection system comprises:

[0037] A multi-source detection data acquisition module is configured to acquire geological data and drilling data of a detection area, and perform induced polarization field, seismic-electric field, natural electric field and hydrological field detection on a potential gushing water area in a coverage airfield, a hole and a tunnel scene to acquire multi-source detection data.

[0038] A preprocessing module is configured to preprocess the multi-source detection data, and in the preprocessing process, a time-space synchronization control method is introduced to align the multi-source detection data to a unified time-space coordinate system to form a standardized time-space data structure.

[0039] A target function construction module is configured to establish a coupling target function with cross-gradient inversion as a physical constraint condition.

[0040] A hybrid network construction module is configured to construct a CNN-GNN-Transformer hybrid network for multi-level feature fusion, wherein the CNN is used to extract local nonlinear response features and process spatial changes of multiple physical fields, the GNN is used to model spatial topological relationships between physical field sources and reflect structural constraints of cross-gradient, and the Transformer is used to capture long-range dependencies and global features.

[0041] A weight configuration module is configured to dynamically balance the contribution of the physical constraint condition in the coupling target function and the deep learning data-driven feature learning ability of the CNN-GNN-Transformer hybrid network through adaptive weights.

[0042] A deep learning prediction module is configured to construct gradient correlations between different physical fields, learn nonlinear mapping relationships between different physical fields through the CNN-GNN-Transformer hybrid network, and perform prediction.

[0043] A hybrid network correction module is configured to compare the prediction results of the CNN-GNN-Transformer hybrid network and the calculation results of the coupling target function, trigger local adjustment when the deviation between the two is greater than a set value, correct abnormal areas, update the parameters of the CNN-GNN-Transformer hybrid network, adjust the adaptive weights, and continue until the prediction results meet the requirements.

[0044] A risk zoning module is configured to perform three-dimensional space modeling of geological structures according to the final prediction results, combine physical parameters to perform structure visualization and risk zoning, and construct a disaster source spatial distribution model.

[0045] Compared with the prior art, the present application has the following beneficial effects:

[0046] This invention proposes a multi-field holographic detection and intelligent decision-making method for underground engineering. By using a fusion mechanism of CNN, GNN, and Transformer hybrid networks and cross-gradient inversion, the detection results of multiple physical fields such as induced electric field, seismic electric field, natural electric field, and hydrological field are used as fusion data sources. This forms a multi-field holographic interpretation and intelligent decision-making method based on a deep learning model, realizing three-dimensional holographic imaging of the source of sudden water inrush in front of the tunnel face. In particular, it can achieve accurate identification of the boundary of the sudden water inrush source and intelligent construction decision-making.

[0047] This invention proposes the idea of ​​using multiple physical field detection information as training data for deep learning models, realizing cross-modal fusion of multi-field holographic data and solving the problem of severe multiple solutions in traditional single physical field detection.

[0048] This invention utilizes a hybrid network architecture combining CNN, GNN, and Transformer, along with adaptive weight coefficients, to form a holographic interpretation method that integrates physical constraints and data-driven approaches, thereby improving the accuracy of detecting adverse geological hazards.

[0049] This invention solves the key problem of accurate three-dimensional holographic imaging identification of sudden water inrush sources in front of the tunnel face by extracting multi-dimensional feature vectors through deep learning networks and combining them with the physical constraint mechanism of cross-gradient inversion. It also establishes a residual feedback-driven self-learning correction mechanism, which dynamically corrects the model parameters through on-site monitoring data, thereby achieving real-time optimization and updating of model parameters and ensuring high accuracy and robustness of prediction results.

[0050] This invention constructs a complete closed-loop intelligent system of detection, interpretation, identification, correction, and decision-making, realizing the generation of standardized engineering decisions directly driven by the results of multi-field holographic detection.

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0052] 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.

[0053] Figure 1 This is a flowchart of a multi-physics field fusion holographic detection method for adverse geological conditions in tunnels, as described in one embodiment.

[0054] Figure 2 This is a schematic diagram of a multi-field holographic detection CNN-GNN-Transformer hybrid network structure in one embodiment;

[0055] Figure 3is a training method flowchart of a CNN-GNN-Transformer hybrid network in an embodiment. DETAILED DESCRIPTION

[0056] The application will be further described below in connection with the drawings and embodiments.

[0057] It should be noted that the following detailed description is illustrative only, and is intended to provide further description of the application. Unless otherwise defined, 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 application belongs.

[0058] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments in accordance with the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, components, and / or groups thereof, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0059] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0060] Embodiment One

[0061] A tunnel bad geological physical field-hydrological field fusion holographic detection method, as shown in Figure 1 includes the following steps:

[0062] A. First, collect and analyze geological data and drilling data in the detection area; for key high-risk gushing water areas, implement multi-physical field detection data collection of induced electric field, seismic electric field, natural electric field, hydrological field, etc., covering air, ground, hole and tunnel scenes.

[0063] B. Preprocess the multi-field detection data by noise filtering and outlier rejection, etc., adopt a streaming normalization integration method, eliminate dimension and scale differences through Z-score standardization or Min-Max scaling; introduce time-space synchronization control technology, align multi-source data to a unified time-space coordinate system, and form a standardized time-space data structure.

[0064] C. Establish a cross-gradient inversion as a coupling objective function of physical constraint conditions:

[0065] wherein the physical constraint condition is: ;

[0066] wherein A : the forward operator matrix corresponding to a certain physical field (here, the induced electric field data related to the Cole model), : difference between observed data and theoretical response, : correction of model parameters. : regularization term (control the smoothness of the model), : cross-gradient constraint term between induced electric field and hydrological field (ensure the structural consistency between resistivity p, polarizability η and permeability K, porosity φ), : cross-gradient constraint term between seismic electric field and natural electric field (ensure the structural consistency between seismic electric response and natural electric potential), : cross-gradient constraint term between multiple fields (ensure the overall structural consistency between the four physical field models), : represents the cross-gradient difference related to the initial model.

[0067] The coupling objective function is:

[0068] ;

[0069] where, d is the observation data vector, f(m) is the forward response, and m is the model parameter, L 1 、L 2is the smoothing constraint operator, and λ1, λ2, λ3, λ4, λ5 are regularization parameters, B 1 、B 2 、B 3is the cross-gradient constraint operator matrix of different physical field combinations: B 1is the cross-gradient operator matrix between induced electric field and hydrological field (gradient cross-constraint between resistivity p, polarizability η and permeability K, porosity φ), B 2is the cross-gradient operator matrix between seismic electric field and natural electric field (gradient cross-constraint between seismic electric response and natural electric potential), B 3is the cross-gradient operator matrix between multiple fields (comprehensive gradient cross-constraint between the four physical fields).

[0070] D. A CNN-GNN-Transformer hybrid network is constructed, which includes CNN modules, GNN modules and Transformer modules connected in turn, realizing multi-level feature fusion: the CNN module is used to extract local nonlinear response features and process spatial changes of multiple physical fields; the GNN module is used to model the spatial topological relationship between physical field sources and reflect the structural constraints of cross-gradient; the Transformer module is used to capture long-range dependencies and global features.

[0071] As Figure 2As shown, the mixed network structure is designed for the characteristics of multi-source heterogeneous data such as induced electric field, seismic electric field, natural electric field and hydrological field, adopts a double-path parallel processing architecture, comprehensively utilizes the advantages of different deep learning models, realizes the unification of local feature extraction, spatial relationship modeling and global dependence analysis, and provides reliable support for three-dimensional imaging and intelligent decision-making of complex adverse geological disasters.

[0072] Based on the cross-gradient constraint condition established in step C, the physical inversion processing is performed on the induced electric field, seismic electric field, natural electric field and hydrological field, the internal correlation between different physical fields is utilized to provide prior constraint information conforming to the geophysical law, and the physical rationality of the inversion result is ensured.

[0073] The network includes CNN module, GNN module and Transformer module connected in sequence:

[0074] The CNN module adopts the structure of convolution layer combined with ReLU nonlinear activation function, which is used for extracting local nonlinear response features and can effectively process the spatial distribution law of multi-physical field detection data and identify local abnormal information. The introduction of ReLU activation function enhances the expression ability of the network for complex nonlinear geological structure.

[0075] The GNN module is used for modeling the spatial topological relationship between physical field sources through the graph convolution network structure, and combines the cross-gradient constraint mechanism to make the structural features of different physical fields be coupled and expressed, and effectively model the spatial adjacency relationship.

[0076] The Transformer module is used to capture global dependence and long-range correlation features through the multi-head self-attention mechanism, solve the shortcomings of traditional methods in processing large-scale, multi-temporal and spatial span data, and effectively model the long-distance spatial correlation.

[0077] Through the adaptive fusion mechanism, the physical constraint results of the multi-field holographic inversion mechanism and the data-driven features of the CNN-GNN-Transformer mixed network are fused according to the dynamic weight. This mechanism can automatically adjust the contribution proportion of physical prior knowledge and deep learning features to form a unified deep feature expression, realizing the deep combination of physical constraint and data-driven.

[0078] The output after fusion contains the three-dimensional spatial structure of the disaster body, the distribution of physical parameters and the risk level evaluation, which provides comprehensive information support for subsequent intelligent decision-making.

[0079] The core advantage of the mixed network structure is that on the one hand, the geological rationality of the result is guaranteed through physical inversion, and on the other hand, the extraction ability of complex features is improved through deep learning network, and the organic combination of the two significantly improves the accuracy and reliability of adverse geological disaster detection.

[0080] E. By dynamically balancing the contribution of traditional cross-gradient physical prior constraints and deep learning data-driven feature learning ability through adaptive weights, the formula is:

[0081] F_fusion= a1F_CNN + a2F_GNN + a3F_Transformer;

[0082] Where F_CNN is the local nonlinear response feature vector extracted by the CNN module, mainly capturing the spatial distribution of multi-physical field detection data and local anomaly information; F_GNN is the spatial topological relationship feature vector extracted by the GNN module, modeling the spatial connection relationship between physical field sources, reflecting the structural constraints of cross-gradient; F_Transformer is the long-range global feature vector extracted by the Transformer module, capturing the global correlation of large-scale, multi-temporal and spatial span data. a1, a2, a3: dynamic adjustment weight coefficients, satisfying a1+a2+a3=1, self-adaptively adjusting according to the training process and data characteristics, realizing the optimal balance between physical constraints and data-driven.

[0083] F. Based on the CNN-GNN-Transformer hybrid network constructed in step D, the gradient correlation between different physical fields is constructed, and the complex nonlinear mapping relationship between induced electric field, seismic electric field, natural electric field and hydrological field is learned through the network, realizing the organic combination of deep feature extraction and physical constraints.

[0084] The specific process includes:

[0085] Using the preprocessed multi-source physical field data as network input;

[0086] Extracting local spatial features of each physical field through CNN module;

[0087] Modeling the spatial topological relationship between different physical field measuring points through GNN module;

[0088] Capture the long-range dependence relationship between multiple physical fields through the Transformer module;

[0089] Construct the cross-gradient correlation matrix between physical fields to quantify the spatial consistency of different physical field responses;

[0090] Learn the nonlinear mapping function from multi-physical field observation data to the spatial distribution of underground anomalies.

[0091] G. Through the residual feedback driving mechanism, the dynamic fusion of physical constraints and deep learning is realized: during the running process of the CNN-GNN-Transformer hybrid network, the predicted results P_ MLwith the physical constraint calculation result P_ Physics , when the residual |P_ ML | is greater than a set threshold, an automatic local adjustment mechanism is triggered to focus on correcting abnormal areas, and the hybrid network parameters and adaptive adjustment fusion weights a1, a2, a3 are updated through the gradient descent algorithm. The whole process follows the principle of dynamic weight adjustment: in the initial stage, the physical constraint weight is larger to ensure geological rationality, in the learning stage, the deep learning weight is gradually increased to exert the feature capturing ability, and in the convergence stage, the contributions of the two are balanced to form an optimal fusion model.

[0092] The whole process follows the principle of dynamic weight adjustment, and the balance weight λ_ physics between physical constraints and deep learning is introduced.

[0093] Final fusion result = λ_ physics × physical constraint result + (1-λ_ physics ) × deep learning result.

[0094] In the initial stage, λ_ physics is larger (such as 0.7-0.8) to ensure that the result conforms to the basic geological law; in the learning stage, the value of λ_ physics is gradually reduced (such as to 0.3-0.5) to allow deep learning to better capture complex features; in the convergence stage, λ_ physics is adjusted to an optimal value (such as 0.4-0.6) to balance the contributions of the two, and finally a fusion model is formed that conforms to the physical law and has high prediction accuracy.

[0095] As shown in Figure 3 , the training of the CNN-GNN-Transformer hybrid network adopts a multi-stage optimization strategy: first, data preparation and network architecture setting, initialization of hybrid network parameters and setting of hyperparameters; in the training process, multi-GPU parallel computing architecture is used to improve large-scale data processing efficiency, dynamic learning rate adjustment strategy (high learning rate in the early stage to accelerate convergence, and small learning rate in the later stage for fine tuning) and performance evaluation checkpoint mechanism are used to prevent overfitting; finally, the model generalization ability and prediction accuracy are improved through result optimization tuning and independent training set fusion ensemble learning strategy.

[0096] The features output by CNN, GNN and Transformer are fused through a dynamic weight mechanism, and the prediction error is fed back using the loss function L = L_ pred + a·L_ residual + β·L_ consistency , where L_ pred is the prediction loss, L_ residual is the feature loss, and L_ consistency is the constraint loss.For residual loss, L_ consistency The consistency loss is represented by α and β, which are weighting coefficients. During training, residual feedback and gradient descent methods are used to progressively optimize the network parameters. Training is terminated when the validation loss no longer decreases or reaches the preset convergence accuracy after several consecutive epochs. This ensures that the model achieves the optimal balance between physical constraints and data-driven approaches during the convergence phase, thereby obtaining high-precision 3D imaging and risk prediction capabilities.

[0097] H. Based on the final prediction results of the CNN-GNN-Transformer hybrid network and the dynamic fusion of physical constraints, a three-dimensional spatial model of geological structures is realized. Combined with typical physical property parameters such as resistivity, polarizability, permeability, and porosity, structural visualization and risk zoning are performed to construct an accurate spatial distribution model of disaster sources.

[0098] I. Based on the risk level classification results, the system automatically matches the corresponding construction treatment plan (advance grouting, drainage hole layout, reinforced support, segmented excavation, etc.) through an intelligent mapping mechanism of risk map - treatment template library - construction suggestions, and generates standardized engineering suggestions to achieve closed-loop control from detection and identification to construction decision-making.

[0099] A four-level risk assessment system is established based on pre-defined risk grading standards: Level I risk (extremely high risk) is defined as polarization rate > 6% and penetration rate > 1×10⁻⁶. -4 m / s, porosity >0.15; Level II risk (high risk) is polarizability 4%~6%, permeability 1×10 -5 ~1×10 - 4 m / s, porosity 0.10~0.15; Level III risk (medium risk) is polarizability 2%~4%, permeability 1×10 -6 ~1×10 -5 m / s, porosity 0.05~0.10; Level IV risk (low risk) is polarizability <2% and permeability <1×10 -6 When any physical property parameter of the detection area reaches the corresponding level threshold, it is judged as the risk level of the corresponding level. When considering multiple parameters, the highest risk level is adopted as the final risk rating of the area, and risk maps of levels I to IV are automatically generated to provide a quantitative basis for subsequent decision-making.

[0100] The specific implementation of the intelligent decision mapping mechanism is as follows: the construction plan is automatically matched according to the risk level. The comprehensive treatment plan of advanced grouting + guide hole + reinforced support is adopted for the Level I risk area; the advanced grouting + conventional support plan is adopted for the Level II risk area; the guide hole + conventional support plan is adopted for the Level III risk area; and the conventional excavation support plan is adopted for the Level IV risk area.

[0101] Example 2

[0102] A tunnel bad geological physical field-hydrological field fusion holographic detection system, comprising:

[0103] A multi-source detection data acquisition module is configured to acquire geological data and drilling data of a detection area, and perform induced polarization field, seismic-electric field, natural electric field and hydrological field detection on a potential gushing water area in a coverage airfield, a hole and a tunnel scene to acquire multi-source detection data.

[0104] A preprocessing module is configured to preprocess the multi-source detection data, and in the preprocessing process, a time-space synchronization control method is introduced to align the multi-source detection data to a unified time-space coordinate system to form a standardized time-space data structure.

[0105] A target function construction module is configured to establish a coupled target function with cross-gradient inversion as a physical constraint condition.

[0106] A hybrid network construction module is configured to construct a CNN-GNN-Transformer hybrid network for multi-level feature fusion, wherein the CNN is used to extract local nonlinear response features and process spatial changes of multiple physical fields, the GNN is used to model spatial topological relationships between physical field sources and reflect structural constraints of cross-gradient, and the Transformer is used to capture long-range dependencies and global features.

[0107] A weight configuration module is configured to dynamically balance the contribution of the physical constraint condition in the coupled target function and the deep learning data-driven feature learning ability of the CNN-GNN-Transformer hybrid network through adaptive weights.

[0108] A deep learning prediction module is configured to construct gradient correlations between different physical fields, learn nonlinear mapping relationships between different physical fields through the CNN-GNN-Transformer hybrid network, and perform prediction.

[0109] A hybrid network correction module is configured to compare the prediction results of the CNN-GNN-Transformer hybrid network with the calculation results of the coupled target function, and when the deviation between the two is greater than a set value, trigger local adjustment, correct abnormal areas, update the parameters of the CNN-GNN-Transformer hybrid network, adjust the adaptive weights, and continue until the prediction results meet the requirements.

[0110] A risk zoning module is configured to perform three-dimensional spatial modeling of geological structures according to the final prediction results, combine with physical parameters to perform structure visualization and risk zoning, and construct a disaster source spatial distribution model.

[0111] The above merely describes the preferred embodiments of the present application and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art without departing from the spirit and principle of the present application shall fall within the protection scope of the present application.

Claims

1. A tunnel bad geological physical field-hydrological field fusion holographic detection method, characterized in that, Includes the following steps: Obtain geological and drilling data of the detection area, and conduct induced electric field, seismic electric field, natural electric field and hydrological field detection covering the entire scene of open ground, boreholes and tunnels in potential water inrush areas to obtain multi-source detection data; The multi-source detection data is preprocessed. During the preprocessing process, a spatiotemporal synchronization control method is introduced to align the multi-source detection data to a unified spatiotemporal coordinate system, forming a standardized spatiotemporal data structure. Establish a coupled objective function with cross-gradient inversion as a physical constraint; A CNN-GNN-Transformer hybrid network is constructed to perform multi-level feature fusion. The CNN is used to extract local nonlinear response features and handle the spatial changes of multi-physics fields. The GNN is used to model the spatial topological relationship between physical field sources and reflect the structural constraints of cross gradients. The Transformer is used to capture long-range dependencies and global features. The contribution of adaptive weight dynamic balancing to the physical constraints in the objective function and the deep learning data-driven feature learning capability of the CNN-GNN-Transformer hybrid network. Construct gradient correlations between different physical fields, and learn nonlinear mapping relationships between different physical fields through a CNN-GNN-Transformer hybrid network for prediction; The prediction results of the CNN-GNN-Transformer hybrid network are compared with the calculation results of the coupled objective function. When the deviation between the two is greater than a set value, a local adjustment is triggered to correct the abnormal region, update the parameters of the CNN-GNN-Transformer hybrid network, and adjust the adaptive weights until the prediction results meet the requirements. Based on the final prediction results, a three-dimensional spatial model of the geological structure is created, and structural visualization and risk zoning are performed in combination with physical property parameters to construct a spatial distribution model of disaster sources.

2. The method according to claim 1, wherein the method is characterized by, The multi-source detection data includes corresponding detection data of induced electric field, seismic electric field, natural electric field and hydrological field.

3. The holographic detection method for adverse geological physical field-hydrological field fusion in tunnels as described in claim 1, characterized in that, The preprocessing of the multi-source detection data includes: noise filtering and outlier removal of the multi-source detection data, using a streaming normalization integration method, and eliminating the differences in dimensions and scales of different detection data through standardization processing.

4. The holographic detection method for unfavorable geological physical field-hydrological field fusion in tunnels as described in claim 1, characterized in that, As an alternative embodiment, the physical constraint condition: ; wherein A is a forward operator matrix for a certain physical field, is a difference between observed data and theoretical response, is a correction of geophysical model parameters, is a regularization term, is a cross-gradient constraint term for the induced polarization field and hydrological field, is a cross-gradient constraint term for the seismic electric field and natural electric field, is a cross-gradient constraint term for the multi-field integration, is used to represent the cross-gradient difference related to the initial geophysical model; The aforementioned physical constraints are integrated into the coupled objective function as basic constraints, and the synergistic utilization of different detection information is achieved through multi-physics joint inversion.

5. The holographic detection method for unfavorable geological physical field-hydrological field fusion in tunnels as described in claim 1, characterized in that, The coupling objective function is: ； wherein, d is the observed data vector, f(m) is the forward response, m is the model parameter, L 1 、L 2is a smoothing constraint operator, λ1, λ2, λ3, λ4, λ5are regularization parameters, B 1 、B 2 、B 3is a cross-gradient constraint operator matrix of different physical field combinations: B 1is a cross-gradient operator matrix of the induced polarization field and the hydrological field, B 2is a cross-gradient operator matrix of the seismic electric field and the natural electric field, B 3is a multi-field integrated cross-gradient operator matrix.

6. The holographic detection method for adverse geological physical field-hydrological field fusion in tunnels as described in claim 1, characterized in that, The process by which the contribution of the deep learning data-driven feature learning capability of the CNN-GNN-Transformer hybrid network to the dynamic balancing of physical constraints in the coupled objective function through adaptive weights includes: F_fusion = α1F_CNN + α2F_GNN + α3F_Transformer; Wherein, F CNN is the local nonlinear response feature vector extracted by CNN, mainly capturing the spatial distribution law and local anomaly information of each physical field detection data; F GNN is the spatial topological relationship feature vector extracted by GNN, modeling the spatial connection relationship between physical field sources, and embodying the structural constraint of cross gradient; F Transformer is the long-range dependence and global feature vector extracted by Transformer, capturing the global correlation of large-scale, multi-temporal and spatial span data, and a1, a2 and a3 are respectively used for dynamically adjusting the weight coefficients, satisfying a1+a2+a3=1, and realizing the optimal balance between physical constraint and data-driven according to the adaptive adjustment of training process and data characteristics.

7. The holographic detection method for adverse geological physical field-hydrological field fusion in tunnels as described in claim 6, characterized in that, By dynamically balancing the contributions of physical constraints in the coupling objective function and the deep learning data-driven feature learning ability of the CNN-GNN-Transformer hybrid network through adaptive weights, the balance weight λ between physical constraints and deep learning is introduced physics : Final fusion result = λ physics × physical constraint result + (1 - λ physics ) × deep learning result; In the initial stage, λ physics is greater than a set value to ensure that the results are consistent with basic geological rules; In Learning stage, gradually reduce the value of λ physics , let the deep learning better play the capture ability to complex characteristics; in the convergence stage, adjust λ physics to the optimal value, balance the contribution of both, and finally form a fusion model that conforms to the physical law and has high prediction accuracy.

8. The holographic detection method for adverse geological physical field-hydrological field fusion in tunnels as described in claim 1, characterized in that, The process of modifying the abnormal area includes modifying by using a loss function, and the loss function L is: L = L_pred + a L_residual + b L_consistency; Wherein, L_pred is a prediction loss, L_residual is a residual loss, L_consistency is a consistency loss, and a and b are weight coefficients, and the adaptive update of parameters is realized by a gradient descent algorithm.

9. The holographic detection method for adverse geological physical field-hydrological field fusion in tunnels as described in claim 1, characterized in that, According to the final prediction result, the process of three-dimensional space modeling of geological structure and structure visualization and risk zoning combined with physical parameters includes: presetting different threshold values of physical parameters, and when the corresponding physical parameter is greater than the threshold value of the corresponding level, it is considered that there is a risk of the corresponding level, and different levels of risk adopt different construction schemes.

10. A tunnel bad geological physical field-hydrological field fusion holographic detection system, characterized in that, Including: A multi-source detection data acquisition module configured to acquire geological data and drilling data of a detection area, and to perform induced electric field, seismic electric field, natural electric field and hydrological field detection of the whole scene of the coverage space, hole and tunnel in the potential gushing water area of the detection area, to acquire multi-source detection data; A preprocessing module configured to preprocess the multi-source detection data, and in the preprocessing process, a time-space synchronization control method is introduced to align the multi-source detection data to a unified time-space coordinate system to form a standardized time-space data structure; A target function construction module configured to establish a coupling target function of cross gradient inversion as a physical constraint condition; A hybrid network construction module configured to construct a CNN-GNN-Transformer hybrid network for multi-level feature fusion, wherein the CNN is used to extract local nonlinear response features and process spatial changes of multiple physical fields, the GNN is used to model the spatial topological relationship between physical field sources and embody the structural constraint of cross gradient, and the Transformer is used to capture long-range dependence and global features; A weight configuration module configured to dynamically balance the contribution of the physical constraint condition in the coupling target function and the deep learning data-driven feature learning ability of the CNN-GNN-Transformer hybrid network through adaptive weights; A deep learning prediction module configured to construct gradient correlation between different physical fields, learn the nonlinear mapping relationship between different physical fields by the CNN-GNN-Transformer hybrid network, and perform prediction; The hybrid network correction module is configured to compare the prediction result of the CNN-GNN-Transformer hybrid network with the calculation result of the coupling target function, trigger local adjustment when the deviation of the two is greater than a set value, correct the abnormal area, update the parameters of the CNN-GNN-Transformer hybrid network, and adjust the adaptive weight until the prediction result meets the requirements; The risk zoning module is configured to perform three-dimensional space modeling of the geological structure according to the final prediction result, combine the physical property parameters to perform structure visualization and risk zoning, and construct a disaster source space distribution model.

Images