Residual network-based structural surface strength prediction system and method

Abstract

The invention discloses a residual network-based structural surface strength prediction system and method, relates to the technical field of geological engineering, and aims to solve the technical problems of low efficiency and insufficient precision due to dependence on field measurement and empirical formulas in a current method. Comprising a residual network feature extraction unit, a spatial expansion convolutional neural network and a dynamic DBSCAN clustering unit, wherein the residual network feature extraction unit is used for performing feature extraction on a GPR image by utilizing a residual network; the spatial expansion convolutional neural network is combined with a spatial expansion convolutional enhanced receptive field to capture wider context information; according to the rock mass structural surface strength prediction system and method based on the GPR three-transient characteristics and the residual network, multi-source data are fused, the deep learning technology is applied, the algorithm process is optimized, and the structural surface strength is predicted. Accurate prediction of rock mass structural plane strength and intelligent generation of a structural plane network are realized, and powerful support is provided for design and construction of rock engineering.

Description

Technical Field

[0001] The present invention relates to the technical field of geological engineering, and more specifically, to a structural plane strength prediction system and method based on a residual network. Background Art

[0002] The distribution of structural planes in rock masses, such as joints and fissures, etc., plays a crucial role in the stability of surrounding rocks and excavation surfaces during tunnel construction. Many scholars at home and abroad have used statistical data methods to establish an undisturbed rock mass structural plane network model, and calculated the rock mass stability through methods such as discrete element method and key block theory. The degree of approximation of the structural plane network model to the actual situation determines the accuracy of the rock mass stability analysis results. However, due to the randomness of the spatial distribution of structural planes, it is difficult to construct an accurate model only based on probability theory and statistics. Some scholars have added limiting conditions during the generation process of the structural plane network model by using the known structural plane distribution information on the outcrop surface, which has improved the accuracy of the model to a certain extent. In the actual application process, the traditional manual investigation method of structural planes is time-consuming, has a small coverage area and has a certain degree of danger. And the information provided by the excavation surface is limited, lacking reliable data in the depth direction.

[0003] Traditional methods rely on on-site measurements and empirical formulas, and have problems of low efficiency and insufficient accuracy. With the development of deep learning technology, especially the successful application of residual networks in image processing, it provides a new idea for the automatic and efficient prediction of structural plane strength. In view of this, we propose a structural plane strength prediction system and method based on a residual network. Summary of the Invention

[0004] The purpose of the present invention is to provide a structural plane strength prediction system and method based on a residual network to solve the technical problems that the current methods rely on on-site measurements and empirical formulas and have low efficiency and insufficient accuracy.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: A structural plane strength prediction system based on a residual network, including a residual network feature extraction unit that uses a residual network to extract features from GPR images, a spatial dilation convolutional neural network and a dynamic DBSCAN clustering unit that combines spatial dilation convolution to enhance the receptive field to capture more extensive context information, a structural plane strength prediction unit that extracts the extracted structural plane features to construct a strength prediction model, and a "random-prediction-determination" structural plane network algorithm generation unit that optimizes the description accuracy of the rock mass structural plane network through prediction data;

[0006] The residual network feature extraction unit is connected to the spatially dilated convolutional neural network and the dynamic DBSCAN clustering unit, and the spatially dilated convolutional neural network and the dynamic DBSCAN clustering unit are connected to the structural plane strength prediction unit. The "random - prediction - determination" structural plane network algorithm generation unit is connected to the structural plane strength prediction unit.

[0007] Preferably, the residual network feature extraction unit includes a data pre - processing module for normalizing the input GPR image, a residual network construction module for designing the residual network architecture, and a feature extraction module for extracting feature maps. The data pre - processing module is connected to the residual network construction module, and the residual network construction module is connected to the feature extraction module.

[0008] Preferably, the spatially dilated convolutional neural network and the dynamic DBSCAN clustering unit includes a spatially dilated convolutional module for capturing context features and a dynamic DBSCAN clustering module for identifying different structural plane regions. The spatially dilated convolutional module is connected to the dynamic DBSCAN clustering module.

[0009] Preferably, the structural plane strength prediction unit includes a feature integration module for forming a complete feature vector, a prediction model construction module for learning the mapping relationship between the feature vector and the structural plane strength, and a strength prediction module for calculating the predicted value of the structural plane strength. The feature integration module is connected to the prediction model construction module, and the prediction model construction module is connected to the strength prediction module.

[0010] Preferably, the "random - prediction - determination" structural plane network algorithm generation unit includes a random generation module for generating multiple groups of candidate structural plane network samples, a prediction evaluation module for predicting and evaluating the candidate samples, and an optimal solution determination module for selecting the optimal candidate sample. The random generation module is connected to the prediction evaluation module, and the prediction evaluation module is connected to the optimal solution determination module.

[0011] The rock mass structural plane strength prediction system and method based on GPR three - transient features and residual network proposed by the present invention realizes the accurate prediction of the rock mass structural plane strength and the intelligent generation of the structural plane network by fusing multi - source data, applying deep learning technology and optimizing the algorithm process, providing strong support for the design and construction of rock engineering.

[0012] A method for predicting the strength of a structural plane based on a residual network includes the following steps:

[0013] S1. Data acquisition and sample set construction;

[0014] Collect GPR image data on the rock mass surface or in boreholes using a GPR device, extract the three-instantaneous characteristics of instantaneous amplitude (A), instantaneous frequency (F), and instantaneous phase (Φ) from the GPR image, and then combine the GPR image, its three-instantaneous characteristics, and spatial position information to construct a multi-domain feature vector sample set for training and testing;

[0015] S2. Extraction of multi-domain feature vectors of radar images;

[0016] Fuse the three-instantaneous characteristics (A, F, Φ) of the GPR image with the spatial position information (x, y) to form a multi-domain feature vector V;

[0017] V(t, x, y) = [A(t, x, y), F(t, x, y), Φ(t, x, y), x, y]

[0018] S3. Network structure design and construction of residual network model;

[0019] S301. Residual network design: Design a deep residual network model containing multiple residual units;

[0020]

[0021] where xl is the input of the l-th layer;

[0022] and are the weights of the convolutional layer;

[0023] Conv is the convolutional operation;

[0024] BN is batch normalization;

[0025] ReLU is the activation function;

[0026] S302. Characteristics of residual neural network: Extract high-level features through the residual learning mechanism;

[0027] S4. Numerical simulation settings and deep learning training;

[0028] Set up numerical simulation experiments under different geological conditions, generate a large number of GPR images and their corresponding three-instantaneous characteristic data, and then use the simulated data and part of the measured data to train the residual network model for preliminary prediction of the joint strength;

[0029] S5. Analysis of deep learning results;

[0030] S501. Model evaluation: Evaluate the prediction performance of the model through cross-validation and confusion matrix;

[0031] S502. Feature importance analysis: Analyze the influence of different features on the prediction results, and optimize feature selection and model structure;

[0032] S6. Intelligent extraction algorithm for rock mass structural plane information;

[0033] S601. Spatial multi-scale convolutional point cloud classification: Use a spatial multi-scale convolutional neural network to classify point cloud data;

[0034] S602. Dynamic DBSCAN clustering: Perform dynamic DBSCAN clustering on candidate regions;

[0035] Clusters = DBSCAN(V, ∈, MinPts)

[0036] where V is a set of multi-domain feature vectors;

[0037] ∈ is the neighborhood radius;

[0038] MinPts is the minimum number of points;

[0039] S7. Structural plane segmentation and information extraction;

[0040] S701. Structural plane segmentation: In the structural plane segmentation stage, use an image segmentation algorithm, and adopt threshold-based segmentation, edge detection, or image segmentation technology. Taking threshold-based segmentation as an example, define a threshold function to distinguish the structural plane and non-structural plane regions;

[0041]

[0042] where V (x,y) is the multi-domain feature vector at the point (x, y);

[0043] Threshold is the preset segmentation threshold;

[0044] Mask(x, y) is the binary mask after segmentation;

[0045] 1 represents the structural plane region;

[0046] 0 represents the non-structural plane region;

[0047] S702. Information extraction: After the structural plane is segmented, extract its geometric and physical properties, such as the area A fece , dip angle θ, and dip direction φ;

[0048]

[0049]

[0050] where (x1, y1) and (x2, y2) are the coordinates of adjacent pixels on the structural plane;

[0051] The PixelArea is the actual area represented by each pixel after calibration according to the image resolution and scanning distance;

[0052] The Weight is the weight based on the pixel position or feature importance;

[0053] S8, Structural plane model test and effect analysis;

[0054] S801, Structural plane model test: Construct a physical model or digital model in the laboratory to simulate the structural planes of the actual rock mass; By controlling variables such as the number, spacing, and dip angle of the structural planes, observe and record the behavior of the model, such as strength and deformation;

[0055] S802, Structural plane effect analysis: Use statistical methods to evaluate the influence of different structural plane parameters on the model behavior;

[0056]

[0057] Among them, SS Between is the sum of squares between groups, that is, the difference between different combinations of structural plane parameters;

[0058] SS Total is the total sum of squares;

[0059] S9, Example application and effect analysis;

[0060] S901, Example application: Apply the algorithm to actual rock mass engineering, such as tunnels and slopes, for structural plane identification and strength prediction;

[0061] S902, Example effect analysis: By comparing the prediction results with on-site measured data, such as core drilling and in-situ tests, evaluate the accuracy and reliability of the algorithm;

[0062]

[0063] Among them, Accuracy is the accuracy rate;

[0064] RMSE is the root mean square error;

[0065] N is the number of samples;

[0066] is the predicted value;

[0067] yi is the true value;

[0068] S10, Generation of bidirectionally constrained random structural planes;

[0069] S1001, Stochastic simulation: Generate an initial structural plane network based on geostatistical principles;

[0070] S1002. Two-way Constraint: Constraining the generated structural plane network in combination with geological laws and engineering requirements, using geological structure lines as constraint conditions to ensure that the strike of the structural plane is consistent with the geological structure lines;

[0071] ConstrainedNetwork = RandomNetwork ∩ GeologicalConstraints

[0072] Among them, ConstrainedNetwork is the structural plane network after constraint;

[0073] RandomNetwork is the randomly generated structural plane network;

[0074] GeologicalConstraints is the geological constraint condition;

[0075] S11. Statistical Characteristics of Structural Plane Distribution and "Random - Prediction - Determination" Network Generation;

[0076] Analyze the statistical characteristics of the structural plane distribution, such as density and directionality, and then generate the final structural plane network through random simulation, prediction optimization, and deterministic adjustment;

[0077] S12. In - situ Test Verification;

[0078] S1201. In - situ Test: Conduct drilling and in - situ tests at the rock mass engineering site to obtain actual structural plane data;

[0079] S1202. Result Analysis: Compare the in - situ measured data with the predicted results, analyze the error sources, and propose improvement measures;

[0080] S13. Analysis of the Inversion Results of the Structural Plane Network;

[0081] Based on the in - situ test data, conduct inversion analysis on the structural plane network to verify the accuracy of the prediction model, and then adjust the model parameters and algorithm flow according to the inversion results to continuously improve the prediction accuracy.

[0082] Preferably, the S1 data acquisition and sample set construction further includes the following steps:

[0083] S101. GPR Data Acquisition: Use GPR equipment to collect GPR image data on the rock mass surface or in boreholes;

[0084] S102. Three - transient Feature Extraction: Extract three - transient features of instantaneous amplitude (A), instantaneous frequency (F), and instantaneous phase (Φ) from GPR images;

[0085] A(t,x,y) = Amplitude(I(t,x,y))

[0086] F(t, x, y) = Frequency(I(t, x, y))

[0087] Φ(t, x, y) = Phase(I(t, x, y))

[0088] Wherein, I(t, x, y) is GPR image data;

[0089] t is time;

[0090] x, y are spatial coordinates;

[0091] S103, Sample set construction: Combine the GPR image, its three instantaneous features, and spatial position information to construct a multi-domain feature vector sample set for training and testing.

[0092] Preferably, the S4 numerical simulation setting and deep learning training further include the following steps:

[0093] S401, Numerical simulation: Set numerical simulation experiments under different geological conditions to generate a large number of GPR images and their corresponding three instantaneous feature data;

[0094] S402, Deep learning training: Use the simulated data and part of the measured data to train the residual network model for preliminary prediction of the structural plane strength;

[0095]

[0096] Wherein, N is the number of samples;

[0097] is the predicted value;

[0098] yi is the true value.

[0099] Preferably, the S11 structural plane distribution statistical features and "random - prediction - determination" network generation further include the following steps:

[0100] S1101, Statistical feature analysis: Analyze the distribution features of the structural plane, such as density and direction distribution;

[0101]

[0102] Wherein, Density is the structural plane density;

[0103] Area of Interest is the area of the region of interest;

[0104] N faces is the number of structural planes;

[0105] φ i is the dip of the i-th structural plane;

[0106] S1102, generation of "random - prediction - determination" network;

[0107] S1102a, random simulation: generating an initial network;

[0108] S1102b, prediction optimization: adjusting network parameters based on prediction data, such as GPR images and geological data;

[0109] S1102c, deterministic adjustment: finally determining the network according to geological laws and engineering requirements.

[0110] Preferably, the analysis of the inversion result of the S13 structural plane network further includes the following steps;

[0111] S1301, inversion analysis: based on on - site test data, performing inversion analysis on the structural plane network, that is, inferring the true situation of the underground structural plane according to the observed data;

[0112] S1302, optimization iteration: adjusting model parameters and algorithm processes according to the inversion results to improve prediction accuracy;

[0113]

[0114] Among them, UpdatedParameters are the updated parameters;

[0115] OriginalParameters are the original parameters;

[0116] α is the learning rate;

[0117] is the gradient of the loss function with respect to the parameters.

[0118] Compared with the prior art, the beneficial effects of the present invention are:

[0119] 1. The rock mass structural plane strength prediction system and method based on GPR triple - transient features and residual network proposed by the present invention realize the accurate prediction of the rock mass structural plane strength and the intelligent generation of the structural plane network by integrating multi - source data, applying deep learning technology, and optimizing the algorithm process, providing strong support for the design and construction of rock engineering.

[0120] 2. The present invention also verifies through on-site tests. It conducts tests at multiple actual rock mass engineering sites to verify the practicability and generalization ability of the algorithm and system, compares the on-site measured data with the prediction results, analyzes the error sources, proposes improvement measures, and then conducts inverse analysis on the structural plane network based on the on-site test data to verify the accuracy of the prediction model. Subsequently, it adjusts the model parameters and algorithm process according to the inverse analysis results to continuously improve the prediction accuracy, achieving the verification and improvement of the model through forward and inverse modeling to further improve the accuracy of the prediction of the rock mass structural plane strength.

[0121] 3. The present invention also sets up numerical simulations, deep learning training, and deep learning result analysis. By setting numerical simulation experiments under different geological conditions, a large number of GPR images and their corresponding three-transient characteristic data are generated. Then, the residual network model is trained using the simulation data and some measured data to conduct a preliminary prediction of the structural plane strength. Thus, the prediction performance of the model is evaluated through methods such as cross-validation and confusion matrix, and the influence of different features on the prediction results is analyzed to optimize the feature selection and model structure, so as to provide in-depth learning training for the model and further optimize the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] Figure 1 It is a schematic block diagram of the structural plane strength prediction system based on the residual network in Embodiment 1 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0123] To facilitate the understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will be further described below with reference to the accompanying drawings of the specification.

[0124] Embodiment 1:

[0125] As Figure 1 shown, the present invention provides a structural plane strength prediction system based on the residual network, including a residual network feature extraction unit that extracts features from GPR images using the residual network, a spatial dilation convolutional neural network and a dynamic DBSCAN clustering unit that combine spatial dilation convolution to enhance the receptive field to capture more extensive context information, a structural plane strength prediction unit that extracts the structural plane features to construct a strength prediction model, and a "random - prediction - determination" structural plane network algorithm generation unit that optimizes the description accuracy of the rock mass structural plane network through prediction data;

[0126] The residual network feature extraction unit is connected to the spatial dilation convolutional neural network and the dynamic DBSCAN clustering unit, and the spatial dilation convolutional neural network and the dynamic DBSCAN clustering unit are connected to the structural plane strength prediction unit, and the "random - prediction - determination" structural plane network algorithm generation unit is connected to the structural plane strength prediction unit.

[0127] In an embodiment of the present invention, the residual network feature extraction unit includes a data preprocessing module for normalizing the input GPR image, a residual network construction module for designing the residual network architecture, and a feature extraction module for extracting the feature map. The data preprocessing module is connected to the residual network construction module, and the residual network construction module is connected to the feature extraction module.

[0128] In an embodiment of the present invention, the spatially dilated convolutional neural network and the dynamic DBSCAN clustering unit include a spatially dilated convolutional module for capturing context features and a dynamic DBSCAN clustering module for identifying different structural plane regions. The spatially dilated convolutional module is connected to the dynamic DBSCAN clustering module.

[0129] In an embodiment of the present invention, the structural plane strength prediction unit includes a feature integration module for forming a complete feature vector, a prediction model construction module for learning the mapping relationship between the feature vector and the structural plane strength, and a strength prediction module for calculating the structural plane strength prediction value. The feature integration module is connected to the prediction model construction module, and the prediction model construction module is connected to the strength prediction module.

[0130] In an embodiment of the present invention, the "random - prediction - determination" structural plane network algorithm generation unit includes a random generation module for generating multiple groups of candidate structural plane network samples, a prediction evaluation module for predicting and evaluating the candidate samples, and an optimal solution determination module for selecting the optimal candidate sample. The random generation module is connected to the prediction evaluation module, and the prediction evaluation module is connected to the optimal solution determination module.

[0131] Embodiment 2:

[0132] The present invention provides a method for predicting the strength of structural planes based on a residual network according to Embodiment 1, including the following steps:

[0133] S1. Data acquisition and sample set construction;

[0134] Use a GPR device to collect GPR image data on the rock mass surface or in boreholes, extract the three - transient features of instantaneous amplitude (A), instantaneous frequency (F), and instantaneous phase (Φ) from the GPR image, and then combine the GPR image, its three - transient features, and spatial position information to construct a multi - domain feature vector sample set for training and testing.

[0135] S2. Extraction of multi - domain feature vectors of radar images;

[0136] Fuse the three - transient features (A, F, Φ) of the GPR image with the spatial position information (x, y) to form a multi - domain feature vector V.

[0137] V(t, x, y) = [A(t, x, y), F(t, x, y), Φ(t, x, y), x, y]

[0138] S3. Network structure design and residual network model construction;

[0139] S301. Residual network design: Design a deep residual network model containing multiple residual units;

[0140]

[0141] where xl is the input of the l-th layer;

[0142] and are the weights of the convolutional layer;

[0143] Conv is the convolution operation;

[0144] BN is batch normalization;

[0145] ReLU is the activation function;

[0146] S302. Residual neural network features: Extract high-level features through the residual learning mechanism;

[0147] S4. Numerical simulation settings and deep learning training;

[0148] Set numerical simulation experiments under different geological conditions, generate a large number of GPR images and their corresponding triple-transient feature data, and then use the simulated data and part of the measured data to train the residual network model for preliminary prediction of the structural plane strength;

[0149] S5. Analysis of deep learning results;

[0150] S501. Model evaluation: Evaluate the prediction performance of the model through cross-validation and confusion matrix;

[0151] S502. Feature importance analysis: Analyze the influence of different features on the prediction results, and optimize feature selection and model structure;

[0152] S6. Intelligent extraction algorithm for rock mass structural plane information;

[0153] S601. Spatial multi-scale convolutional point cloud classification: Use a spatial multi-scale convolutional neural network to classify point cloud data;

[0154] S602. Dynamic DBSCAN clustering: Perform dynamic DBSCAN clustering on the candidate regions;

[0155] Clusters = DBSCAN(V, ∈, MinPts)

[0156] Among them, V is a set of multi-domain feature vectors;

[0157] ∈ is the neighborhood radius;

[0158] MinPts is the minimum number of points;

[0159] Through the settings of numerical simulation, deep learning training, and deep learning result analysis, by setting numerical simulation experiments under different geological conditions, a large number of GPR images and their corresponding three instantaneous characteristic data are generated. Then, the residual network model is trained using the simulated data and some measured data to conduct a preliminary prediction of the structural plane strength. Thus, the prediction performance of the model is evaluated through methods such as cross-validation and confusion matrix, and the influence of different features on the prediction results is analyzed to optimize feature selection and model structure, so as to provide in-depth learning training for the model and further optimize the accuracy of the model;

[0160] S7. Structural plane segmentation and information extraction;

[0161] S701. Structural plane segmentation: In the structural plane segmentation stage, an image segmentation algorithm is used, and threshold-based segmentation, edge detection, or image segmentation technology is adopted. Taking threshold-based segmentation as an example, a threshold function is defined to distinguish the structural plane and non-structural plane regions;

[0162]

[0163] Among them, V( x,y ) is the multi-domain feature vector at point ( x,y );

[0164] Threshold is the preset segmentation threshold;

[0165] Mask(x, y) is the binary mask after segmentation;

[0166] 1 represents the structural plane region;

[0167] 0 represents the non-structural plane region;

[0168] S702. Information extraction: After the structural plane is segmented, its geometric and physical properties are extracted, such as the area A fece of the structural plane, dip angle θ, and dip direction φ;

[0169]

[0170] Among them, (x1, y1) and (x2, y2) are the coordinates of adjacent pixels on the structural plane;

[0171] PixelArea is the actual area represented by each pixel after calibration according to the image resolution and scanning distance;

[0172] The Weight is the weight based on pixel position or feature importance;

[0173] S8, Structural plane model test and effect analysis;

[0174] S801, Structural plane model test: Construct a physical model or digital model in the laboratory to simulate the structural planes of the actual rock mass; By controlling variables such as the number, spacing, and dip angle of the structural planes, observe and record the behavior of the model, such as strength and deformation;

[0175] S802, Structural plane effect analysis: Use statistical methods to evaluate the influence of different structural plane parameters on the behavior of the model;

[0176]

[0177] Among them, SS Between is the sum of squares between groups, that is, the difference between different combinations of structural plane parameters;

[0178] SS Total is the total sum of squares;

[0179] S9, Example application and effect analysis;

[0180] S901, Example application: Apply the algorithm to actual rock mass engineering, such as tunnels and slopes, for structural plane identification and strength prediction;

[0181] S902, Example effect analysis: By comparing the prediction results with on-site measured data, such as core drilling and in-situ tests, evaluate the accuracy and reliability of the algorithm;

[0182]

[0183] Among them, Accuracy is the accuracy rate;

[0184] RMSE is the root mean square error;

[0185] N is the number of samples;

[0186] is the predicted value;

[0187] yi is the true value;

[0188] S10, Generation of bidirectionally constrained random structural planes;

[0189] S1001, Stochastic simulation: Generate an initial structural plane network based on geostatistical principles;

[0190] S1002, Bidirectional constraint: Combine geological laws and engineering requirements to constrain the generated structural plane network, use geological tectonic lines as constraint conditions to ensure that the strike of the structural planes is consistent with the geological tectonic lines;

[0191] ConstrainedNetwork = RandomNetwork ∩ GeologicalConstraints

[0192] Among them, ConstrainedNetwork is the network of structural planes after constraint;

[0193] RandomNetwork is the network of structural planes randomly generated;

[0194] GeologicalConstraints are geological constraint conditions;

[0195] S11. Statistical characteristics of structural plane distribution and "random - prediction - determination" network generation;

[0196] Analyze the statistical characteristics of structural plane distribution, such as density and directionality, and then generate the final structural plane network through random simulation, prediction optimization, and deterministic adjustment;

[0197] S12. In - situ test verification;

[0198] S1201. In - situ test: Conduct drilling and in - situ tests at the rock mass engineering site to obtain actual structural plane data;

[0199] S1202. Result analysis: Compare the in - situ measured data with the predicted results, analyze the error sources, and propose improvement measures;

[0200] S13. Analysis of the inversion results of the structural plane network;

[0201] Based on the in - situ test data, conduct inversion analysis on the structural plane network to verify the accuracy of the prediction model, and then adjust the model parameters and algorithm flow according to the inversion results to continuously improve the prediction accuracy.

[0202] The rock mass structural plane strength prediction system and method based on GPR three - transient characteristics and residual network proposed by the present invention realize the accurate prediction of the rock mass structural plane strength and the intelligent generation of the structural plane network by integrating multi - source data, applying deep learning technology, and optimizing the algorithm flow, providing strong support for the design and construction of rock engineering;

[0203] Moreover, through on-site tests and verifications in multiple actual rock mass engineering sites, the practicality and generalization ability of the algorithm and system are tested. By comparing the on-site measured data with the predicted results, the error sources are analyzed, and improvement measures are proposed. Then, based on the on-site test data, an inversion analysis of the structural plane network is carried out to verify the accuracy of the prediction model. Subsequently, according to the inversion results, the model parameters and algorithm flow are adjusted to continuously improve the prediction accuracy. Through forward and inversion, the model is verified and improved to further enhance the accuracy of the prediction of the rock mass structural plane strength.

[0204] As another embodiment of the present invention, the S1 data acquisition and sample set construction further includes the following steps:

[0205] S101. GPR data acquisition: Use GPR equipment to collect GPR image data on the rock mass surface or in boreholes.

[0206] S102. Three-instantaneous feature extraction: Extract three-instantaneous features of instantaneous amplitude (A), instantaneous frequency (F), and instantaneous phase (Φ) from the GPR image.

[0207] A(t, x, y) = Amplitude(I(t, x, y))

[0208] F(t, x, y) = Frequency(I(t, x, y))

[0209] Φ(t, x, y) = Phase(I(t, x, y))

[0210] where I(t, x, y) is the GPR image data;

[0211] t is time;

[0212] x, y are spatial coordinates;

[0213] S103. Sample set construction: Combine the GPR image, its three-instantaneous features, and spatial position information to construct a multi-domain feature vector sample set for training and testing.

[0214] As another embodiment of the present invention, the S4 numerical simulation setting and deep learning training further includes the following steps:

[0215] S401. Numerical simulation: Set numerical simulation experiments under different geological conditions to generate a large amount of GPR images and their corresponding three-instantaneous feature data.

[0216] S402. Deep learning training: Use the simulated data and part of the measured data to train the residual network model for a preliminary prediction of the structural plane strength.

[0217]

[0218] Among them, N is the number of samples;

[0219] is the predicted value;

[0220] yi is the true value.

[0221] As another embodiment of the present invention, the statistical characteristics of the S11 structural plane distribution and the generation of the "random - prediction - determination" network further include the following steps:

[0222] S1101. Statistical feature analysis: Analyze the distribution characteristics of the structural plane, such as density and direction distribution;

[0223]

[0224] Among them, Density is the structural plane density;

[0225] Area of Interest is the area of the region of interest;

[0226] N faces is the number of structural planes;

[0227] φ i is the dip of the i-th structural plane;

[0228] S1102. Generation of the "random - prediction - determination" network;

[0229] S1102a. Random simulation: Generate an initial network;

[0230] S1102b. Prediction optimization: Based on prediction data, such as GPR images and geological data, adjust the network parameters;

[0231] S1102c. Deterministic adjustment: Finally determine the network according to geological laws and engineering requirements.

[0232] As another embodiment of the present invention, the analysis of the S13 structural plane network inversion results further includes the following steps;

[0233] S1301. Inversion analysis: Based on on-site test data, conduct inversion analysis on the structural plane network, that is, infer the true situation of the underground structural plane according to the observed data;

[0234] S1302. Optimization iteration: Adjust the model parameters and algorithm process according to the inversion results to improve the prediction accuracy;

[0235]

[0236] Among them, UpdatedParameters are the updated parameters;

[0237] OriginalParameters are the original parameters;

[0238] α is the learning rate;

[0239] is the gradient of the loss function with respect to the parameters.

[0240] The embodiments disclosed in the present invention are preferred embodiments, but not limited thereto. Those of ordinary skill in the art can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. However, as long as they do not depart from the spirit of the present invention, they are within the protection scope of the present invention.

Claims

1. A structural surface strength prediction system based on residual network, characterized in that: It includes a residual network feature extraction unit that uses a residual network to extract features from GPR images, a spatial dilation convolutional neural network and a dynamic DBSCAN clustering unit that combines spatial dilation convolution to enhance the receptive field to capture a wider range of contextual information, a structural surface strength prediction unit that extracts structural surface features to construct a strength prediction model, and a "random-prediction-deterministic" structural surface network algorithm generation unit that optimizes the accuracy of rock mass structural surface network description through prediction data; The residual network feature extraction unit is connected to the spatial expansion convolutional neural network and the dynamic DBSCAN clustering unit, and the spatial expansion convolutional neural network and the dynamic DBSCAN clustering unit are connected to the structural surface strength prediction unit, and the "random-prediction-deterministic" structural surface network algorithm generation unit is connected to the structural surface strength prediction unit.

2. According to the residual network-based structural surface strength prediction system of claim 1, it is characterized in that: The residual network feature extraction unit includes a data preprocessing module for normalizing the input GPR image, a residual network construction module for designing the residual network architecture, and a feature extraction module for extracting feature maps. The data preprocessing module is connected to the residual network construction module, and the residual network construction module is connected to the feature extraction module.

3. The structural surface strength prediction system based on residual network according to claim 1 is characterized in that: The spatial dilation convolutional neural network and dynamic DBSCAN clustering unit include a spatial dilation convolution module for capturing contextual features and a dynamic DBSCAN clustering module for identifying different structural surface areas, and the spatial dilation convolution module is connected to the dynamic DBSCAN clustering module.

4. The structural surface strength prediction system based on residual network according to claim 1 is characterized in that: The structural surface strength prediction unit includes a feature integration module for forming a complete feature vector, a prediction model construction module for learning the mapping relationship between the feature vector and the structural surface strength, and a strength prediction module for calculating the structural surface strength prediction value. The feature integration module is connected to the prediction model construction module, and the prediction model construction module is connected to the strength prediction module.

5. The structural surface strength prediction system based on residual network according to claim 1 is characterized in that: The "random-prediction-determination" structural surface network algorithm generation unit includes a random generation module for generating multiple groups of candidate structural surface network samples, a prediction and evaluation module for predicting and evaluating candidate samples, and an optimal solution determination module for selecting the optimal candidate samples. The random generation module is connected to the prediction and evaluation module, and the prediction and evaluation module is connected to the optimal solution determination module.

6. A prediction method using the structural surface strength prediction system based on residual network according to claims 1 to 5, characterized in that: The following steps are involved: S1, data acquisition and sample set construction; GPR image data is collected on the rock surface or in the borehole using GPR equipment, and three instantaneous features, namely instantaneous amplitude (A), instantaneous frequency (F) and instantaneous phase (Φ), are extracted from the GPR image. Then, the GPR image and its three instantaneous features and spatial position information are combined to construct a multi-domain feature vector sample set for training and testing. S2, multi-domain feature vector extraction of radar images; The three instantaneous features (A, F, Φ) of the GPR image are fused with the spatial position information (x, y) to form a multi-domain feature vector V; V(t,x,y)=[A(t,x,y),F(t,x,y),Φ(t,x,y),x,y] S3, network structure design and residual network model construction; S301. Residual network design: Design a deep residual network model containing multiple residual units; xl+1=ReLU(BN(Conv(xl,W l 1 ))+Conv(xl,W l 2 ))+xl Among them, xl is the input of the lth layer; W l 1 and W l 2 is the weight of the convolutional layer; Conv is the convolution operation; BN stands for batch normalization; ReLU is the activation function; S302, residual neural network features: extracting high-level features through residual learning mechanism; S4, numerical simulation setup and deep learning training; Numerical simulation experiments under different geological conditions were set up to generate a large number of GPR images and their corresponding three-instantaneous characteristic data. The residual network model was then trained using the simulated data and some measured data to make a preliminary prediction of the structural surface strength. S5. Analysis of deep learning results; S501, Model evaluation: Evaluate the prediction performance of the model through cross-validation and confusion matrix; S502, Feature Importance Analysis: Analyze the impact of different features on prediction results, and optimize feature selection and model structure; S6, intelligent extraction algorithm of rock mass structural surface information; S601, spatial multi-scale convolutional point cloud classification: using spatial multi-scale convolutional neural network to classify point cloud data; S602, dynamic DBSCAN clustering: performing dynamic DBSCAN clustering on the candidate regions; Clusters=DBSCAN(V,∈,MinPts) Among them, V is a set of multi-domain feature vectors; ∈ is the neighborhood radius; MinPts is the minimum number of points; S7, structural surface segmentation and information extraction; S701, structural surface segmentation: In the structural surface segmentation stage, an image segmentation algorithm is used, and threshold-based segmentation, edge detection or image segmentation technology is adopted. Taking threshold-based segmentation as an example, a threshold function is defined to distinguish between structural surface and non-structural surface areas; Among them, V(x,y) is the point (x ,y) at the multi-domain feature vector; Threshold is the preset segmentation threshold; Mask(x,y) is the binary mask after segmentation; 1 indicates the structural surface area; 0 indicates the non-structured surface area; S702, information extraction: After the structural surface is segmented, its geometric and physical properties are extracted, such as the area A of the structural surface. fece , inclination θ and dip φ; Among them, (x1, y1) and (x2, y2) are the coordinates of adjacent pixels on the structure surface; PixelArea is the actual area represented by each pixel after calibration based on image resolution and scanning distance; Weight is a weight based on pixel position or feature importance; S8, structural surface model test and effect analysis; S801. Structural surface model test: construct a physical model or digital model in the laboratory to simulate the structural surface of the actual rock mass; by controlling variables such as the number, spacing, and inclination of the structural surfaces, observe and record the behavior of the model, such as strength and deformation; S802, Structural surface effect analysis: Use statistical methods to evaluate the impact of different structural surface parameters on model behavior; Among them, SS Between is the sum of squares between groups, i.e., the difference between different combinations of structural surface parameters; SS Total is the total sum of squares; S9, Example application and effect analysis; S901. Example application: Apply the algorithm to actual rock mass engineering, such as tunnels and slopes, to identify structural surfaces and predict strength; S902, Example effect analysis: Evaluate the accuracy and reliability of the algorithm by comparing the prediction results with the field measured data, such as drilling coring and in-situ tests; Among them, Accuracy is the accuracy; RMSE is the root mean square error; N is the sample size; is the predicted value; yi is the true value; S10, bidirectional constrained random structure surface generation; S1001, random simulation: Generate the initial structural surface network based on the principles of geostatistics; S1002, bidirectional constraint: constrain the generated structural surface network in combination with geological laws and engineering requirements, use geological structural lines as constraint conditions, and ensure that the direction of the structural surface is consistent with the geological structural line; ConstrainedNetwork=RandomNetwork∩GeologicalConstraints Among them, ConstrainedNetwork is the constrained structural surface network; RandomNetwork is a randomly generated structural surface network; GeologicalConstraints are geological constraints; S11, Statistical characteristics of structural surface distribution and "random-predictive-deterministic" network generation; Analyze the distribution statistical characteristics of the structural surface, such as density and directionality, and then generate the final structural surface network through random simulation, predictive optimization and deterministic adjustment; S12, Field test verification; S1201, Field test: Drill holes and perform in-situ tests at the rock mass engineering site to obtain actual structural surface data; S1202, Result analysis: Compare the field measured data with the predicted results, analyze the source of errors, and propose improvement measures; S13, analysis of structural surface network inversion results; Based on the field test data, the structural surface network is inverted and analyzed to verify the accuracy of the prediction model. The model parameters and algorithm process are then adjusted according to the inversion results to achieve continuous improvement in prediction accuracy.

7. The prediction method according to claim 6, characterized in that: The S1 data acquisition and sample set construction also includes the following steps: S101, GPR data acquisition: Use GPR equipment to collect GPR image data on the rock surface or in the borehole; S102, three-instantaneous feature extraction: extracting three-instantaneous features of instantaneous amplitude (A), instantaneous frequency (F) and instantaneous phase (Φ) from the GPR image; A(t,x,y)=Amplitude(I(t,x,y)) F(t,x,y)=Frequency(I(t,x,y)) Φ(t,x,y)=Phase(I(t,x,y)) Where, I(t,x,y) is the GPR image data; t is time; x, y are spatial coordinates; S103, sample set construction: The GPR image and its three-moment features and spatial position information are combined to construct a multi-domain feature vector sample set for training and testing.

8. The prediction method according to claim 6, characterized in that: The S4 numerical simulation setup and deep learning training also include the following steps: S401, numerical simulation: set up numerical simulation experiments under different geological conditions to generate a large number of GPR images and their corresponding three-instantaneous characteristic data; S402, deep learning training: using simulated data and some measured data to train the residual network model to make a preliminary prediction of the structural surface strength; Where N is the number of samples; is the predicted value; yi is the true value.

9. The prediction method according to claim 6, characterized in that: The S11 structural surface distribution statistical characteristics and the "random-predictive-deterministic" network generation also include the following steps: S1101. Statistical feature analysis: Analyze the distribution characteristics of the structural surface, such as density and directional distribution; Where, Density is the structural surface density; Area of ​​Interest is the area of ​​the region of interest; N faces is the number of structural surfaces; φ i is the tendency of the i-th structural surface; S1102, "random-predictive-deterministic" network generation; S1102a, random simulation: generating an initial network; S1102b, prediction optimization: adjusting network parameters based on prediction data, such as GPR images and geological data; S1102c, Deterministic adjustment: Finalize the network based on geological laws and engineering requirements.

10. The prediction method according to claim 6, characterized in that: The S13 structural surface network inversion result analysis also includes the following steps: S1301, Inversion analysis: Based on the field test data, the structural surface network is inverted for analysis, that is, the real situation of the underground structural surface is inferred from the observed data; S1302, optimization iteration: adjusting model parameters and algorithm flow according to the inversion results to improve prediction accuracy; Among them, UpdatedParameters is the updated parameters; OriginalParameters is the original parameter; α is the learning rate; is the gradient of the loss function with respect to the parameters.

Images