Tunnel unfavorable geological intelligent identification method and system based on ensemble learning

Abstract

The present application belongs to the field of geophysical exploration and tunnel engineering, and provides a tunnel bad geological intelligent identification method and system based on ensemble learning, a three-dimensional geological model containing real geological information is constructed, and different categories of bad geological bodies are randomly set; imaging of different detection methods is performed on the three-dimensional geological model; geological feature parameters are extracted; the three-dimensional geological model, imaging results and geological feature parameters are stored in association, and a multi-source feature map representing the position and shape of the bad geological body is extracted; the multi-source feature map and the geological feature parameters are deeply fused to form a comprehensive feature vector reflecting the spatial shape and geological properties of the bad geological body; the comprehensive feature vector is used to train an ensemble learning model, and the trained ensemble learning model is used to process target detection data to obtain an identification result of the bad geological body. The present application can realize intelligent classification and prediction of the bad geological category and risk level.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of geophysical exploration and tunnel engineering, and particularly relates to a tunnel adverse geology intelligent identification method and system based on ensemble learning. BACKGROUND

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

[0003] With the increasingly complex geological conditions of engineering construction, multiple types of disasters such as fault fracture zone, soft interlayer, karst cave, and water-rich zone are coupled and linked, which can easily induce engineering disasters such as collapse, gushing, sudden mud, and large deformation of surrounding rock, seriously threatening construction safety. When facing complex, superimposed, and multi-scale adverse geological bodies, it is difficult to accurately depict complex geological bodies by relying on a single physical property parameter. In order to improve the prediction reliability, multiple detection methods are gradually used for comprehensive detection to obtain the distribution range of adverse geology, and comprehensive analysis is carried out combined with geological data, field conditions, etc. to obtain the final adverse geology position, category, and other analysis results. However, the current comprehensive interpretation of multi-source and multi-modal data mainly relies on the experience of interpreters, and is mainly based on subjective comparative analysis, which leads to a large space for improvement in the stability and accuracy of adverse geology identification results.

[0004] In recent years, deep learning algorithms have made significant progress in multi-modal data fusion, and are expected to use their nonlinear mapping capabilities and feature extraction and fusion capabilities to realize the identification and interpretation of tunnel multi-source detection data for adverse geology intelligent identification. However, considering the large differences in detection distance, scale, and other aspects of various tunnel detection methods, and the difficulty in effectively introducing multi-source heterogeneous geological information, there are still the following problems in the interpretation of tunnel comprehensive detection results by using the adverse geology intelligent identification method:

[0005] (1) The relationship between geophysical properties and geological data is unclear, and it is difficult to build high-quality simulation data: deep learning is essentially a data-driven method, which needs a large-scale and complete sample data set as support. The existing tunnel advanced detection database mainly uses wave velocity, resistivity, dielectric constant, and other geophysical properties, and lacks systematic modeling and quantitative description of the relationship between geophysical parameters and key geological parameters such as lithology and surrounding rock integrity, thereby restricting the construction of high-credibility simulation data sets for real working conditions;

[0006] (2) Multi-modal data difference is significant, and multi-source information is difficult to efficiently fuse: The multi-source geophysical data has great differences in detection distance, response characteristics and imaging form, and the fusion is difficult. For example, the imaging of seismic wave method is often an arc-shaped anomaly affected by the observation system, and the resistivity method is significantly affected by volume effect. How to effectively align and fuse the results of multiple geophysical methods under the condition of inconsistent spatial scale and abnormal characteristics, and to stably delineate the position, size and shape of the geological body from them, is a key problem in current identification.

[0007] (3) There is a significant difference between the imaging results of geophysical exploration and geological information in representation and data structure, and it is difficult to analyze the adverse geological category and risk level in a unified framework. SUMMARY

[0008] To solve the above problems, the present application provides a tunnel adverse geological intelligent identification method and system based on ensemble learning, which can realize intelligent classification and prediction of adverse geological category and risk level, and significantly improve the accuracy, reliability and engineering applicability of tunnel advanced detection.

[0009] According to some embodiments, the present application adopts the following technical scheme:

[0010] A tunnel adverse geological intelligent identification method based on ensemble learning, comprising the following steps:

[0011] A three-dimensional geological model containing real geological information is constructed, and different categories of adverse geological bodies are randomly set in the three-dimensional geological model;

[0012] Geophysical forward of different detection methods is performed on the three-dimensional geological model to obtain multi-source geophysical data, and inversion imaging is performed to obtain adverse geological imaging results under each detection method;

[0013] The qualitative data in the geological information are structured and quantified to be converted into geological feature parameters;

[0014] The three-dimensional geological model, adverse geological imaging results and geological feature parameters are associatedly stored to form a comprehensive database;

[0015] The seismic imaging, resistivity imaging and geological radar data in the comprehensive database are preprocessed and feature filtered to extract multi-source feature maps representing the position and shape of adverse geology;

[0016] The multi-source feature maps and the geological feature parameters are deeply fused to form a comprehensive feature vector reflecting the spatial shape and geological properties of adverse geology;

[0017] The integrated learning model is trained using the comprehensive feature vector, and the trained integrated learning model is used to process the target detection data to obtain the identification result of the adverse geological body.

[0018] As an alternative embodiment, a three-dimensional geological model containing real geological information is constructed, and the process of randomly setting different categories of adverse geological bodies in the three-dimensional geological model comprises: determining the typical geometric shape, spatial distribution and physical parameter value range of each category of adverse geological body according to the correlation between different categories of adverse geological bodies and the geological information, randomly generating the interface position, dip angle, thickness and the position, size and shape of the adverse geological body in the three-dimensional grid model based on the determined information, and forming a three-dimensional geological model.

[0019] As a further defined embodiment, the process of randomly setting different categories of adverse geological bodies comprises randomly generating a fault fracture zone, defining a normal vector based on the geometric parameters of the fault surface, using the normal vector to describe the strike, tendency and dip angle of the fault, and randomly generating the normal vector component;

[0020] The fault position is generated by randomly selecting an initial cross-sectional position and combining normal distribution disturbance to determine the fault thickness, thereby generating a fault fracture zone with variable width;

[0021] A random cutting mechanism is introduced to simulate the actual situation that the fault does not completely pass through the tunnel, a discontinuous or locally missing fault fracture zone is generated by randomly selecting a cutting area, and the geological authenticity of the model is enhanced.

[0022] As a further defined embodiment, the process of randomly setting different categories of adverse geological bodies comprises randomly generating a water-bearing body, creating three-dimensional grid coordinates according to a given grid shape, randomly determining the center coordinates of an ellipsoid, the semi-axis length and the water-bearing physical property identification value, and realizing multi-scale and arbitrary attitude water-bearing body modeling by implementing single-axis or multi-axis rotation matrix transformation on the ellipsoid coordinates, and realizing multi-ellipsoid splicing extension through sphere center control to complete complex water-bearing body modeling.

[0023] As an alternative embodiment, the process of performing geophysical forward modeling of different detection methods on the three-dimensional geological model to obtain multi-source geophysical data and performing inversion imaging comprises: using observation data to perform forward and inversion imaging for seismic method, and expressing two-dimensional acoustic wave equation in isotropic medium;

[0024] The finite element method is used to perform forward modeling to obtain electrical method data, and a least squares linear inversion method based on smooth constraint is used to perform three-dimensional resistivity inversion;

[0025] Geological radar is used to perform inversion imaging using observation data to obtain the relative permittivity distribution in front of the working face, and to calculate the porosity and water saturation parameters of the medium.

[0026] As an alternative embodiment, the process of structuring and quantifying the qualitative data in the geological information and converting it into the geological feature parameters includes: structuring and quantifying the lithology, surrounding rock grade, rock strength, face state, weathering degree, karst development degree, and fault / groundwater information, and converting them into geological feature parameters;

[0027] According to the correlation between the rock physical parameters and the geological features, the conditional probability constraint is established, the value range and combination of the geological feature parameters are probabilistically described, and the samples are randomly sampled under the conditional probability constraint to ensure that various possible geological scenarios are covered.

[0028] As an alternative embodiment, the process of correlating the three-dimensional geological model, the adverse geological imaging results, and the geological feature parameters to form a comprehensive database includes: standardizing each item of data, converting it into a predetermined storage format, constructing a multi-source database table structure, establishing a unified index and attribute field, and using the unified index and attribute field to manage the data in storage.

[0029] As an alternative embodiment, the process of pre-processing and feature filtering the seismic imaging, resistivity imaging, and geological radar data in the comprehensive database includes: for the seismic and radar imaging results in the comprehensive database, a Gabor filter with learnable parameters is used to extract imaging features, and the expression of the Gabor filter is:

[0030] ;

[0031] Wherein: λ is the wavelength of the sine wave, which determines the frequency characteristics of the filter; θ is the direction of the filter, which is used to capture features at a specific angle; is the phase shift, which affects the waveform of the filtering response; σ is the standard deviation of the Gaussian function, which controls the bandwidth of the filter; γ is the spatial aspect ratio, which adjusts the shape of the filter in the horizontal and vertical directions, , , is the spatial horizontal coordinate, is the spatial vertical coordinate;

[0032] For the resistivity imaging results in the comprehensive database, a variable convolution module is used, and its output is:

[0033] ;

[0034] Wherein, represents the input resistivity imaging, represents the output feature, K is the total number of sampling points of the convolution kernel, is the convolution kernel weight of the kth sampling point, is the fixed sampling position of the traditional convolution kernel, The offset obtained through network learning is used to adjust the sampling position, and the offset The offset of each position and each sampling point is generated by an additional convolution layer, and the input is a current feature map, and the output is an offset.

[0035] As an optional implementation, the process of deeply fusing the multi-source feature map with the geological feature parameter to form a comprehensive feature vector reflecting the spatial form and geological attribute of the adverse geology includes: fusing the three types of features of seismic, resistivity and radar with the geological feature parameter in a unified spatial grid, and realizing adaptive allocation of weights in the spatial dimension and the channel dimension by using a multi-layer convolution attention module.

[0036] The data of different detection methods are transformed by a multi-source spatial registration model to generate multi-resolution three-dimensional raster data, for the input features, the multi-layer convolution attention module calculates channel attention and spatial attention, generates fused features through weighted calculation, calculates the importance weight of each feature by using the multi-layer convolution attention module, and based on the weight, performs weighted summation on each feature to generate a fused feature representation.

[0037] As an optional implementation, the process of training the integrated learning model by using the comprehensive feature vector includes: combining a hybrid encoding architecture formed by a U-Net model and a Transformer model with a LightGBM integrated learning model, first extracting the spatial features of the geophysical data by the U-Net model, then inputting the geophysical data and the geological information into the Transformer model, capturing long-distance dependencies and realizing cross-modal fusion by means of the multi-head attention mechanism, generating a feature tensor containing geological semantic constraints, and finally embedding the fused feature tensor into the LightGBM integrated learning model to complete classification.

[0038] An integrated learning-based tunnel adverse geology intelligent identification system, comprising:

[0039] A three-dimensional geological model construction module configured to construct a three-dimensional geological model containing real geological information, and randomly set different types of adverse geology bodies in the three-dimensional geological model;

[0040] An imaging module configured to perform geophysical forward modeling of different detection methods on the three-dimensional geological model, obtain multi-source geophysical data, perform inversion imaging, and obtain adverse geology imaging results under each detection method;

[0041] A geological feature extraction module configured to structure and quantify the qualitative data in the geological information, and convert the qualitative data into geological feature parameters;

[0042] A comprehensive database construction module configured to store the three-dimensional geological model, the adverse geology imaging results and the geological feature parameters in association to form a comprehensive database;

[0043] a feature extraction module configured to preprocess and feature filter the seismic imaging, resistivity imaging and geological radar data in the integrated database to extract multi-source feature maps representing the position and morphology of adverse geology;

[0044] a deep fusion module configured to perform deep fusion of the multi-source feature maps and geological feature parameters to form an integrated feature vector reflecting the spatial morphology and geological properties of adverse geology;

[0045] a recognition module configured to train an ensemble learning model using the integrated feature vector, and process target detection data using the trained ensemble learning model to obtain a recognition result of adverse geology.

[0046] Compared with the prior art, the present application has the following advantages:

[0047] (1) The present application analyzes the correlation between typical adverse geology and geological information and three-dimensional physical modeling, proposes a complex geological modeling method for intelligent recognition of tunnel adverse geology, and constructs a tunnel multi-source integrated detection data set containing "geological information-physical parameters-multi-source detection results".

[0048] (2) The present application proposes an intelligent recognition strategy of "first circle and then judge" for tunnel adverse geology, establishes an imaging feature fusion strategy for seismic imaging, resistivity inversion and geological radar data, and constructs a multi-source data adverse geology recognition network of "first independent feature extraction and then unified fusion recognition", which realizes automatic delineation of abnormal body features from multi-source detection results.

[0049] (3) The present application proposes an adverse geology classification model based on ensemble learning on the basis of adverse geology delineation, updates and optimizes the model parameters, and finally realizes intelligent recognition and reliable interpretation of adverse geology categories and risk levels in various advanced prediction results.

[0050] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are used for explanation. BRIEF DESCRIPTION OF DRAWINGS

[0051] The drawings accompanying the specification of the present application form a part of the present application, and the schematic embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application.

[0052] Figure 1 is a flow chart of a tunnel adverse geology intelligent identification method based on ensemble learning in an embodiment;

[0053] Figure 2 is a flow chart of geology-property modeling in an embodiment;

[0054] Figure 3 is an overall network architecture diagram of an established intelligent identification in an embodiment;

[0055] Figure 4 is a result map of adverse geology delineation in an embodiment;

[0056] Figure 5 is a result map of adverse geology identification in an embodiment, wherein (a) is a curve diagram of adverse geology identification result, and (b) is a diagram of corresponding identification conclusion of (a). DETAILED DESCRIPTION

[0057] The application will be further described below in conjunction with the accompanying drawings and embodiments.

[0058] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise indicated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the application pertains.

[0059] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit exemplary embodiments according to the present application. As used herein, the singular form is intended to include the plural form unless the context clearly indicates otherwise, and it should also be understood that when the terms "comprise" and / or "include" are used in the specification, there is a presence of a feature, step, operation, device, component and / or combinations thereof.

[0060] The embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0061] Embodiment one

[0062] A tunnel adverse geology intelligent identification method based on ensemble learning, as shown in Figure 1 includes the following steps:

[0063] Step 1, a complex three-dimensional geological model containing geological information is constructed, and various types of adverse geological bodies are randomly arranged in the model;

[0064] Step 2, forward modeling of geophysical data is performed on the three-dimensional geological model to obtain corresponding multi-source geophysical data; inversion and imaging processing are performed on the geophysical data to form adverse geology imaging results under various detection methods;

[0065] Step 3: Structuring and quantifying geological data such as survey, drilling, exposure, and surrounding rock classification, and converting qualitative geological information such as lithology, surrounding rock classification, and degree of weathering into geological feature parameters that can be input into the network, achieving digital expression of geological information.

[0066] Step 4: Correlating and storing complex three-dimensional geological models, multi-source geophysical inversion imaging results, and quantified geological information, and building a comprehensive database containing "geological information-physical property parameters-multi-source imaging data" as the data basis for subsequent deep learning and ensemble learning model training and testing.

[0067] Step 5: Preprocessing seismic imaging, resistivity imaging, and geological radar data in the comprehensive database, and designing corresponding filters based on the imaging characteristics of each type of data; using a multi-layer convolutional neural network to extract features from each type of geophysical data to obtain multi-source feature maps representing the location and shape of adverse geology.

[0068] Step 6: Deeply integrating the multi-source geophysical features obtained in Step 5 and the geological information features obtained in Step 3 to form a comprehensive feature vector that reflects both the spatial shape and geological properties of adverse geology, providing input for subsequent adverse geology classification and risk assessment.

[0069] Step 7: Building an ensemble learning model based on the comprehensive feature vector to identify and classify adverse geology; using multiple base classifiers to distinguish multi-source features from different angles and integrating the results to obtain the classification and risk level of adverse geology.

[0070] The following describes each step in detail.

[0071] In Step 1, the main types of adverse geology in the tunnel are selected, and their relationships with geological information are analyzed to determine the typical geometric shape, spatial distribution, and physical property parameter range of each type of adverse geology. Based on these relationships, a three-dimensional complex geological structure model is generated by randomly placing interfaces and abnormal bodies in the three-dimensional grid model, with random interface positions, inclinations, thicknesses, and abnormal body positions, sizes, and shapes.

[0072] In this embodiment, faults, karst, and water bodies are selected as the main types of adverse geology, and their relationships with lithology, rock strength, surrounding rock classification, fault zones / folds, and groundwater are analyzed.

[0073] In this embodiment, geological structure modeling is performed based on the adverse geology shape and location of the tunnel to achieve a geological shape with interwoven geological interfaces, fault structures, and different shaped water bodies. This is done in two ways.

[0074] On the one hand, the fault fracture zone is randomly set. The normal vector is defined based on the geometric parameters of the fault plane, describing the fault's strike, dip, and dip angle. Spatial diversity of the fault plane is achieved by randomly generating normal vector components. Simultaneously, the fault location is determined by randomly selecting the initial cross-sectional location and combining it with a normal distribution perturbation to determine the fault thickness, thus generating a fault fracture zone with variable width. A random cut-off mechanism is introduced to simulate the actual situation where the fault does not completely cross the tunnel. By randomly selecting the cut-off area, discontinuous or locally missing fault fracture zones are generated, enhancing the geological realism of the model. The fault is determined by the following formula:

[0075] ;

[0076] Where x, y, and z are three-dimensional spatial coordinates. , Indicates the first The offset parameters of the two boundary surfaces of the fault, sgn is the sign function. It is obtained by calculating the rotation matrix R:

[0077] , ;

[0078] in The strike and dip angle (dip) of the fault correspond to the fault attitude, respectively, and are randomly selected from [0, 2π]. Displacements are defined randomly in the local coordinate system. The corresponding displacement is derived in global coordinates using a rotation matrix. have:

[0079] ;

[0080] On the other hand, stochastic modeling of aquifers such as karst is employed. Based on a dynamically adjustable ellipsoid semi-axis generation algorithm and an ellipsoid cluster spatial topology reorganization mechanism, variable-scale and multi-directional modeling of aquifers is achieved. Three-dimensional mesh coordinates are created based on the mesh shape. The three-dimensional mesh coordinates x, y, and z (100×40×40) are created according to the given mesh shape. The ellipsoid center coordinates, semi-axis length (controlling shape and size), and aquifer physical property identifiers such as low resistivity are randomly determined. The ellipsoid calculation formula is as follows:

[0081] ;

[0082] in, , , Here, represents the center coordinates of the ellipsoid, and a, b, and c are the semi-axis lengths along the X, Y, and Z axes, respectively. By performing single-axis or multi-axis rotation matrix transformations on the ellipsoid coordinates, multi-scale, arbitrary-orientation water-bearing body modeling can be achieved.

[0083] ;

[0084] where θ is the rotation angle, similarly, for the case of rotation around Y axis and rotation around Z axis, the rotation matrix can be expressed as:

[0085] ;

[0086] Since the rotation matrix is a linear transformation, when rotation around multiple axes is needed, the individual rotation matrices can be multiplied to obtain the final rotation matrix.

[0087] Unitize the rotated coordinate (divide by the semi-axis length), calculate the distance of the relative coordinate to the origin, and determine whether the grid point is located inside the ellipsoid by the distance <1, the calculation formula is as follows:

[0088] ;

[0089] The multi-ellipsoid splicing extension is realized by the sphere center control, the automatic modeling of the complex water body is completed, and finally a 100x40x40 three-dimensional structure matrix (1-n identifies different structure regions) is output, which provides a basic model for subsequent property parameter assignment.

[0090] In step 2, the three-dimensional geological model is subjected to geophysical forward and inversion and imaging processing to form the imaging results of adverse geology under multiple detection methods, such as shown in Figure 2 ;

[0091] In this embodiment, the observation data of the seismic method is used for forward and inversion imaging, and the two-dimensional acoustic wave equation in isotropic medium can be expressed as:

[0092] ;

[0093] where x and z are spatial coordinates, t is time, u is displacement, v p is the propagation rate of the wave, and the seismic wave forward calculation is obtained by solving the above partial differential equation to obtain the propagation path and arrival time of the seismic wave in the medium, so as to obtain the observation data.

[0094] In this embodiment, the finite element method is used for forward simulation of the electrical method to obtain electrical data, and the specific calculation formula is as follows:

[0095] ;

[0096] where ΔU MN is the potential difference between M point and N point, is the device coefficient determined by the electrode arrangement geometry, is the power supply current.

[0097] The three-dimensional resistivity inversion is carried out by using a least square linear inversion method based on a smooth constraint, and the objective function is as follows:

[0098]

[0099] In the formula, is a difference vector of actual observation data and forward theoretical observation data, is a model update amount, A is a partial derivative matrix (a partial derivative of a model's forward theoretical observation data to a model parameter), also called a sensitivity matrix, C is a smoothness matrix, and λ is a Lagrange constant for adjusting a smooth constraint weight.

[0100] In the embodiment, the geological radar is inversed and imaged by using observation data to obtain a relative permittivity distribution in front of a working face, so that parameters such as porosity and water saturation of the medium can be calculated. The principle is derived from the Maxwell curl equation as follows:

[0101]

[0102]

[0103] In the formula, E is an electric field intensity, D is an electric flux density, H is a magnetic field intensity, B is a magnetic flux density, J is an electric current density, is a magnetic current density.

[0104] In step 3, the geological information needs to be quantitatively processed to realize digital expression of the geological information.

[0105] In the embodiment, the main reference is the Railway Tunnel Design Specification, combined with the Engineering Rock Mass Classification Standard (GB 50218-2014), the Railway Tunnel Advanced Geological Forecast Technical Specification (Q / CR 9217-2015), the Railway Tunnel Intelligent Advanced Geological Forecast Technical Guide (Q / CR 9256-2023) and other standards, the geological data is set as follows, which is used as a reference for data representation and used in subsequent intelligent interpretation.

[0106] (1) Lithology: In the embodiment, the lithology is also divided into 5 categories according to the Railway Tunnel Design Specification, and parameters 1-5 are set to describe the lithology description of the current working face in the preliminary exploration data of the detection area and the working face sketch of the current working face;

[0107] (2) Surrounding rock classification: In the embodiment, 1-6 is used to represent the surrounding rock.

[0108] (3) Rock strength: In the embodiment, the rock strength is divided into 5 categories, i.e., extremely hard rock, hard rock, relatively soft rock, soft rock and extremely soft rock, according to the Railway Tunnel Design Specification, and 1-5 is used to represent the extremely hard rock to the extremely soft rock. ​​​

[0109] (4) Face state: In this embodiment, the face state is divided into stable, front falling, front extrusion, and front instability according to engineering experience, and 1-4 are used to represent the above four states in turn;

[0110] (5) Weathering degree: In this embodiment, the weathering degree is divided into slight weathering, weak weathering, strong weathering, and full weathering, and 1-4 are used to represent the above four states in turn;

[0111] (6) Karst development degree: In this embodiment, the karst development degree is divided into slight development, weak development, medium development, and strong development, and 1-4 are used to represent the above four states in turn;

[0112] (7) Whether there is a fault zone / fold or underground water: In this embodiment, 1 or 2 is used to represent whether there is a fault / fold and underground water in the tunnel advance prediction area in the preliminary survey data;

[0113] In step 4, the geological model, forward and inverse imaging results and geological information of the above steps need to be combined to build a database, as shown in Figure 2 .

[0114] In this embodiment, the data standardization storage format is established by defining and introducing coordinate points through file names, and then the multi-source database table structure is constructed, and the unified data storage management is realized by establishing multi-type data index.

[0115] In step 5, the seismic imaging, resistivity three-dimensional inversion results and geological radar profile in the comprehensive database are respectively input into a multi-encoder parallel structure to fully characterize the spatial response characteristics of adverse geology, as shown in Figure 3 .

[0116] In this embodiment, for feature extraction of seismic and radar imaging results, a parameter-learnable Gabor filter (its core structure is derived from GaborNet) is used to extract seismic wave imaging features. Its mathematical expression is as follows:

[0117] ;

[0118] Wherein: , , λ is the wavelength of the sine wave, which determines the frequency characteristics of the filter; θ is the direction of the filter, which is used to capture the characteristics of a specific angle; is the phase shift, which affects the waveform of the filtering response; σ is the standard deviation of the Gaussian function, which controls the bandwidth of the filter; γ is the spatial aspect ratio, which adjusts the shape of the filter in the horizontal and vertical directions, is the spatial horizontal coordinate, The spatial longitudinal coordinate is denoted as z. The steps of extracting the seismic and radar imaging features using the Gabor filter are as follows: firstly, pre-processing the imaging data, including normalization and denoising, to improve the data quality; then, using the designed Gabor filter, covering the texture features of different directions and scales by adjusting the direction parameter theta and the wavelength lambda; then, applying the filter to the seismic image, generating a multi-channel filter response map through convolution operation, and calculating the amplitude or energy to enhance the feature saliency; finally, fusing the response maps of multiple directions and frequencies to generate a comprehensive texture and structure feature map, which is used for further feature extraction and geological anomaly identification.

[0119] The application designs a variable convolution filter for the imaging results of the resistivity method, which can adaptively focus on the target area and extract the resistivity imaging results. The core idea of variable convolution is to introduce a learnable offset for each sampling point of the convolution kernel, so that the shape of the convolution kernel can be dynamically adjusted according to the local features of the input data. For the input feature map x, the output feature y(p) at position p is:

[0120] ;

[0121] wherein, represents the input resistivity imaging, represents the output feature, K is the total number of sampling points of the convolution kernel, is the convolution kernel weight of the kth sampling point, is the fixed sampling position of the traditional convolution kernel, is the offset learned by the network, used to adjust the sampling position. The offset is generated by an additional convolution layer (Offset Learning Layer), and the input is the current feature map, and the output is the offset of each position and each sampling point.

[0122] In step 6, the features of different methods and different positions need to be weighted and fused in a unified spatial grid to realize the automatic identification and characterization of the position, geometric shape and spatial range of the anomaly body, as shown in Figure 4 , wherein X, Y and Z are the corresponding coordinate axes.

[0123] In this embodiment, a multi-source data feature fusion method based on a multi-layer attention mechanism is proposed, which introduces a multi-layer convolutional attention module (Convolutional Block Attention Module, CBAM) combined with a convolutional network to realize cross-weighted fusion between features, thereby improving the feature fusion efficiency and the accuracy of geological anomaly prediction.

[0124] The multi-source geophysical data is first transformed by a multi-source spatial registration model to generate multi-resolution three-dimensional rasterized data. Second, for the input features F , the CBAM module calculates channel attention and spatial attention :

[0125] ;

[0126] ;

[0127] where σ is the sigmoid activation function, AvgPool and MaxPool are average pooling and maximum pooling respectively, and MLP is a multi-layer perceptron, denotes a 7x7 convolution.

[0128] The fusion features are further generated by weighted calculation:

[0129] ;

[0130] This process calculates the importance weights of each feature through the CBAM attention mechanism, which reflect the contribution of the feature in the fusion process. Then, the features are weighted and summed according to these weights to generate the fused feature representation. This method can effectively retain and strengthen key features while suppressing unimportant features, thereby improving the fusion effect and prediction accuracy.

[0131] In step 7, an ensemble learning model is constructed based on the comprehensive feature vector to identify and classify the adverse geological bodies; multiple base classifiers are used to distinguish the multi-source features from different angles, and the discrimination results are integrated to obtain the adverse geological class and risk level determination results, as shown in Figure 5 (a) and Figure 5 (b).

[0132] In this embodiment, a deep fusion feature extraction framework of U-Net + Transformer + LightGBM integrated classifier is proposed for adverse geological classification: to solve the problem of weak correlation between multi-source data, the spatial features of geophysical data such as seismic wave method and resistivity method are first extracted by U-Net, and then input into Transformer together with geological sketch and drilling information, etc. to capture long-distance dependencies and realize cross-modal fusion through multi-head attention mechanism, generate feature tensor containing geological semantic constraints, and finally embed the fused features into LightGBM to complete classification, which makes up for the defects of traditional feature fusion.

[0133] In the feature fusion stage, the multi-head attention mechanism of the Transformer is used to realize the cross-modal association of geological information and geophysical features. Let the encoding features of the geological information be , and the spatial features of the geophysical data be , both of which are input into the Transformer, and the importance of different modal features is dynamically adjusted through attention weights, and finally the fusion feature tensor is generated, which can be mathematically expressed as:

[0134] ;

[0135] This fusion method effectively captures the deep association between geological information and geophysical features, providing more comprehensive feature expression for subsequent classification.

[0136] In the LightGBM ensemble learning model, the encoding features of the geological information are used as auxiliary input and participate in the construction of the decision tree together with the geophysical features. Through the variance gain criterion, the model automatically learns the influence weight of the geological information on the classification result, so as to fully consider the constraint effect of the geological background in the prediction process. Specifically, let the input feature vector be , and the prediction result of LightGBM be represented as:

[0137] ;

[0138] wherein is the output of the tth decision tree, and T is the total number of decision trees. Through multiple random forest trees, the relationship between multi-class data and the recognition result is analyzed from different angles, and then a comprehensive analysis result is obtained to more accurately predict the category and risk level of the adverse geology.

[0139] The above embodiment builds a complex three-dimensional geological modeling method based on the relationship between geological information and abnormal bodies, establishes a comprehensive database of “geology-property-detection results” covering typical adverse geological working conditions by combining geophysical forward and inversion imaging; focuses on the intelligent identification strategy of “first circle then judge” for adverse geology in tunnels based on deep learning algorithm, and the core idea is to first circle the range of abnormal bodies and then determine the category of abnormal bodies. This strategy is realized through the following two main steps: through the introduction of multi-source filters and multi-layer convolutional neural networks, the multi-source detection data is respectively extracted and unified, the position and shape of the adverse geological body in front of the tunnel are circled, and then the integrated learning model is used and the geological information is introduced to assist in comprehensively identifying and risk assessing the recognition result, realizing the intelligent classification and prediction of the category and risk level of adverse geology, thereby significantly improving the accuracy, reliability and engineering applicability of tunnel advanced detection.

[0140] Embodiment Two

[0141] A tunnel adverse geology intelligent identification system based on ensemble learning, comprising:

[0142] A three-dimensional geological model construction module configured to construct a three-dimensional geological model containing real geological information, and randomly set different categories of adverse geological bodies in the three-dimensional geological model;

[0143] An imaging module configured to perform geophysical forward of different detection methods on the three-dimensional geological model, obtain multi-source geophysical data, perform inversion imaging, and obtain adverse geological imaging results under each detection method;

[0144] A geological feature extraction module configured to structure and quantify qualitative data in the geological information, and convert the qualitative data into geological feature parameters;

[0145] A comprehensive database construction module configured to store the three-dimensional geological model, the adverse geological imaging results, and the geological feature parameters in association, and form a comprehensive database;

[0146] A feature extraction module configured to pre-process and feature filter seismic imaging, resistivity imaging, and geological radar data in the comprehensive database, and extract multi-source feature maps representing adverse geological positions and shapes;

[0147] A deep fusion module configured to perform deep fusion of the multi-source feature maps and the geological feature parameters, and form a comprehensive feature vector reflecting adverse geological spatial shapes and geological properties;

[0148] An identification module configured to train an ensemble learning model using the comprehensive feature vector, process target detection data using the trained ensemble learning model, and obtain an identification result of the adverse geological body.

[0149] Those skilled in the art should understand that embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer usable storage media (including, but not limited to, magnetic disk storage, optical storage, etc.) containing computer usable program code. CD ROM - Figure 1

[0150] ​The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flows and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart or flows and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart or flows and / or block diagram block or blocks.

[0151] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the Figure 1 one or more functions specified in the flowchart or flows and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart or flows and / or block diagram block or blocks.

[0152] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flows and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart or flows and / or block diagram block or blocks. ​ one or more functions specified in the flowchart or flows and / or block diagram block or blocks.

[0153] The application described above is merely preferred embodiments of the application, and the application is not limited thereto. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art without departing from the spirit and principle of the application shall fall within the scope of protection of the application.

Claims

1. A method for intelligent identification of adverse geological conditions in tunnels based on ensemble learning, characterized in that, Includes the following steps: Construct a three-dimensional geological model containing real geological information, and randomly set different types of adverse geological bodies in the three-dimensional geological model; Geophysical forward modeling of the three-dimensional geological model using different detection methods was performed to obtain multi-source geophysical data, and inversion imaging was carried out to obtain adverse geological imaging results under each detection method. Qualitative data in geological information is structured and quantified, and transformed into geological characteristic parameters; The three-dimensional geological model, the imaging results of adverse geological conditions, and the geological feature parameters are linked and stored to form a comprehensive database; Preprocessing and feature filtering are performed on seismic imaging, resistivity imaging and ground-penetrating radar data in the comprehensive database to extract multi-source feature maps that characterize the location and morphology of adverse geological conditions. By deeply integrating multi-source feature maps with geological feature parameters, a comprehensive feature vector reflecting adverse geological spatial morphology and geological attributes is formed. The integrated learning model is trained using comprehensive feature vectors, and the trained integrated learning model is used to process target detection data to obtain the identification results of adverse geological bodies. The process of deeply fusing multi-source feature maps with geological feature parameters to form a comprehensive feature vector that reflects adverse geological spatial morphology and geological attributes includes: fusing three types of features—earthquake, resistivity, and radar—with geological feature parameters within a unified spatial grid, and using a multi-layer convolutional attention module to achieve adaptive weight allocation for spatial and channel dimensions. Data from different detection methods are transformed through a multi-source spatial registration model to generate multi-resolution 3D rasterized data. For input features, a multi-layer convolutional attention module calculates channel attention and spatial attention, and generates fused features through weighted calculation. The importance weight of each feature is calculated using the multi-layer convolutional attention module. Based on the weights, each feature is weighted and summed to generate the fused feature representation. The process of training the ensemble learning model using comprehensive feature vectors includes: using a hybrid coding architecture formed by the U-Net model and the Transformer model combined with the LightGBM ensemble learning model, firstly, the spatial features of the exploration data are extracted through the U-Net model, and then input into the Transformer model along with the geological information. The multi-head attention mechanism is used to capture long-distance dependencies and achieve cross-modal fusion, generating feature tensors with geological semantic constraints. Finally, the fused feature tensors are embedded into the LightGBM ensemble learning model to complete the classification.

2. The method for intelligent identification of adverse geological conditions in tunnels based on ensemble learning as described in claim 1, characterized in that, The process of constructing a three-dimensional geological model containing real geological information and randomly setting different types of adverse geological bodies in the three-dimensional geological model includes: determining the typical geometric shape, spatial distribution and physical property parameter range of each type of adverse geological body according to the correlation between different types of adverse geological bodies and geological information; and randomly generating interface positions, dip angles, thicknesses, as well as the positions, sizes and shapes of adverse geological bodies in the three-dimensional mesh model based on the determined information to form a three-dimensional geological model.

3. The intelligent identification method for adverse geological conditions in tunnels based on ensemble learning as described in claim 2, characterized in that, The process of randomly setting different types of adverse geological bodies includes randomly generating fault fracture zones, defining the geometric parameters of the fault plane based on normal vectors, using normal vectors to describe the strike, dip and dip angle of the fault, and randomly generating normal vector components. The fault location is determined by randomly selecting the initial cross-sectional location and combining it with a normal distribution perturbation to determine the fault thickness, thereby generating a fault fracture zone with variable width; A random excision mechanism is introduced to simulate the actual situation where the fault does not completely cross the tunnel. By randomly selecting the excision area, discontinuous or partially missing fault fracture zones are generated, thereby enhancing the geological realism of the model.

4. The method for intelligent identification of adverse geological conditions in tunnels based on ensemble learning as described in claim 2, characterized in that, The process of randomly setting different types of adverse geological bodies includes randomly generating aquifers, creating three-dimensional grid coordinates based on a given grid shape, randomly determining the center coordinates, semi-axis length, and aquifer property identifiers of the ellipsoid, realizing multi-scale and arbitrary posture aquifer modeling by performing single-axis or multi-axis rotation matrix transformation on the ellipsoid coordinates, and realizing multi-ellipsoid splicing and extension through sphere center control to complete the modeling of complex aquifers.

5. The intelligent identification method for adverse geological conditions in tunnels based on ensemble learning as described in claim 1, characterized in that, The process of performing geophysical forward modeling on the three-dimensional geological model using different detection methods, obtaining multi-source geophysical data, and performing inversion imaging includes: performing forward and inversion imaging using seismic observation data, and expressing the two-dimensional acoustic wave equation in an isotropic medium; The electrical resistivity data were obtained by forward modeling using the finite element method, and then three-dimensional resistivity inversion was performed using the least squares linear inversion method based on smoothness constraints. The ground-penetrating radar is used to perform inversion imaging based on observation data to obtain the relative permittivity distribution in front of the tunnel face and to calculate the porosity and water saturation parameters of the medium.

6. The intelligent identification method for adverse geological conditions in tunnels based on ensemble learning as described in claim 1, characterized in that, The process of structuring and quantifying qualitative data in geological information and converting it into geological characteristic parameters includes: structuring and quantifying information on lithology, surrounding rock grade, rock strength, working face condition, weathering degree, karst development degree, and fault / groundwater information and converting it into geological characteristic parameters; Conditional probability constraints are established based on the correlation between rock physical parameters and geological features. The range of values ​​and combinations of geological feature parameters are probabilistically described, and random sampling is performed under the conditional probability constraints to generate samples to ensure coverage of various possible geological scenarios.

7. The intelligent identification method for adverse geological conditions in tunnels based on ensemble learning as described in claim 1, characterized in that, The process of linking and storing three-dimensional geological models, adverse geological imaging results, and geological feature parameters to form a comprehensive database includes: standardizing various data, converting them into a predetermined data entry format, constructing a multi-source database table structure, establishing unified indexes and attribute fields, and using the unified indexes and attribute fields to manage the data entry.

8. The method for intelligent identification of adverse geological conditions in tunnels based on ensemble learning as described in claim 1, characterized in that, The preprocessing and feature filtering of seismic imaging, resistivity imaging, and ground-penetrating radar data in the integrated database includes: for seismic and radar imaging results in the integrated database, a Gabor filter with learnable parameters is used to extract imaging features. The Gabor filter expression is: ； Where: λ is the wavelength of the sine wave, which determines the frequency characteristics of the filter; θ is the direction of the filter, used to capture features at a specific angle; σ represents the phase shift, affecting the waveform of the filter response; σ is the standard deviation of the Gaussian function, controlling the filter's bandwidth; γ is the aspect ratio, adjusting the filter's shape in both the horizontal and vertical directions. , , The horizontal coordinate is the coordinate of the space. The vertical coordinate is the coordinate in space. The resistivity imaging results from the comprehensive database are used with a variable convolution module, and the output is: ； in, Indicates input resistivity imaging, This represents the output features, where K is the total number of sampling points in the convolution kernel. The weights of the convolution kernel at the k-th sampling point are... For the fixed sampling positions of traditional convolution kernels, The offset is obtained through network learning and is used to adjust the sampling position. It is generated by an additional convolutional layer, with the current feature map as input and the offset of each location and each sampling point as output.

9. A tunnel adverse geological condition intelligent identification system based on ensemble learning, characterized in that, include: The 3D geological model building module is configured to build a 3D geological model containing real geological information, and randomly set different types of adverse geological bodies in the 3D geological model; The imaging module is configured to perform geophysical forward modeling on the three-dimensional geological model using different detection methods, acquire multi-source geophysical data, perform inversion imaging, and obtain adverse geological imaging results under each detection method. The geological feature extraction module is configured to structure and quantify qualitative data in geological information, transforming it into geological feature parameters; The integrated database construction module is configured to associate and store three-dimensional geological models, adverse geological imaging results, and geological feature parameters to form an integrated database; The feature extraction module is configured to preprocess and feature filter the seismic imaging, resistivity imaging and ground-penetrating radar data in the integrated database to extract multi-source feature maps that characterize the location and morphology of adverse geological conditions. The deep fusion module is configured to deeply fuse multi-source feature maps with geological feature parameters to form a comprehensive feature vector that reflects adverse geological spatial morphology and geological attributes. The identification module is configured to train the ensemble learning model using comprehensive feature vectors, and then use the trained ensemble learning model to process the target detection data to obtain the identification results of adverse geological bodies. The process of deeply fusing multi-source feature maps with geological feature parameters to form a comprehensive feature vector that reflects adverse geological spatial morphology and geological attributes includes: fusing three types of features—earthquake, resistivity, and radar—with geological feature parameters within a unified spatial grid, and using a multi-layer convolutional attention module to achieve adaptive weight allocation for spatial and channel dimensions. Data from different detection methods are transformed through a multi-source spatial registration model to generate multi-resolution 3D rasterized data. For input features, a multi-layer convolutional attention module calculates channel attention and spatial attention, and generates fused features through weighted calculation. The importance weight of each feature is calculated using the multi-layer convolutional attention module. Based on the weights, each feature is weighted and summed to generate the fused feature representation. The process of training the ensemble learning model using comprehensive feature vectors includes: using a hybrid coding architecture formed by the U-Net model and the Transformer model combined with the LightGBM ensemble learning model, firstly, the spatial features of the exploration data are extracted through the U-Net model, and then input into the Transformer model along with the geological information. The multi-head attention mechanism is used to capture long-distance dependencies and achieve cross-modal fusion, generating feature tensors with geological semantic constraints. Finally, the fused feature tensors are embedded into the LightGBM ensemble learning model to complete the classification.

Images