An intelligent classification method for dangerous areas after overhanging rock instability

Abstract

The present invention belongs to the technical field of geological disaster monitoring, and specifically discloses a method for intelligently dividing the dangerous area level after the overhang rock becomes unstable. The method scans the overhang rock to generate a DEM and performs correction; samples and performs geotechnical mechanics tests to obtain mechanical parameters and calibrate the distribution range; imports the DEM into the PFC to generate a three-dimensional slope model, freely releases the overhang rock in the three-dimensional slope model to simulate falling, monitors and records key parameters and associates them with risk levels; establishes a hierarchical model, a judgment matrix and an evaluation function, and verifies the consistency of the judgment matrix to ensure that the weight result is credible; sets a risk level division standard; calculates the risk score and matches the risk level through the evaluation function; constructs a data set and a CNN model, and uses the data set to train, verify and test the CNN model; the CNN model receives key parameters, calculates and outputs the corresponding risk level. The present invention has the characteristics of high timeliness, less human intervention and accurate judgment.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster monitoring, and in particular to a method for intelligently grading dangerous areas after overhanging rocks become unstable, which has high timeliness, less human intervention and accurate judgment. Background Art

[0002] Overhang rock is a type of rock mass with potential collapse risk formed on slopes or cliffs due to external forces such as weathering, rainfall, and earthquakes. The instability and collapse of overhang rock not only poses a threat to engineering structures, but may also cause major safety accidents. Therefore, accurately assessing and classifying the level of dangerous areas after the instability of overhang rock has important guiding significance for disaster relief, disaster reduction, and post-disaster reconstruction, and is also an important topic in the current prevention and control of geological disasters.

[0003] At present, the traditional classification of dangerous areas after the instability of overhanging rocks mostly relies on manual inspections and on-site surveys, and then the dangerous areas are divided through expert meetings. Not only is the timeliness poor, but it also relies on human experience, resulting in large differences in judgment results, which seriously restricts disaster relief, disaster reduction and post-disaster reconstruction after instability. For this reason, there are also technical means such as prior surveys and excavations of overhanging rocks and their surroundings to form basic data of overhanging rocks, and quantitative calculations are performed on the basic data. According to the quantitative calculation results, qualitative analysis is performed to form a dangerous area division plan, which is corrected in combination with the on-site survey data after instability, and finally the dangerous area level is divided; although the above method reduces the human impact compared to the traditional method and provides the accuracy of the judgment results, the accuracy and effectiveness are still insufficient, and for overhanging rocks with high distribution positions, steep terrain, and dense surface vegetation, on-site surveys, excavations and other technical means are difficult to effectively apply, which also affects the application.

[0004] In the prior art, there are methods to automatically collect post-disaster data through on-site surveys, radar scanning, and sensors, and then use MATLAB to perform curve fitting on the post-disaster data, and use the BP neural network model for nonlinear mapping, so as to more accurately divide the scope of the dangerous area and the approximate static state of the disaster body, and provide data support for the manual division of the dangerous area level after instability. Compared with pre-disaster surveys and plans, the workload is reduced, and the speed of obtaining post-disaster data is increased, which effectively improves the timeliness. However, since manual division is still required in the later stage, the subjective factors of the division results are relatively large, resulting in poor accuracy of the division results; and there is a lack of pre-disaster data support and only static conditions can be provided after the disaster, so it is impossible to provide basic geological data and dynamic data such as impact force and impulse changes. Therefore, there is a serious lack of data for dividing the dangerous area level, which affects the accuracy of the division results and poses serious safety hazards. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides a method for intelligently grading dangerous areas after overhanging rock instability with high timeliness, less human intervention and accurate judgment.

[0006] The present invention is implemented in the following steps: including the acquisition of digital elevation model, rock and soil mechanics analysis, overhang rock instability simulation, AHP construction evaluation system, CNN model construction, and dangerous area classification steps, the specific contents are:

[0007] A. Acquisition of digital elevation model: Perform high-resolution 3D scanning of the target area of ​​the overhanging rock to obtain point cloud data; then clean the point cloud data and process it at multiple resolutions to generate a high-precision digital elevation model; then perform precision correction on the digital elevation model based on the actual terrain conditions on site;

[0008] B. Geotechnical mechanics analysis: Sampling the target area and conducting geotechnical mechanics tests on the geotechnical samples to obtain mechanical parameters; statistically analyzing the mechanical parameters to determine their distribution form, calculating the mean value, standard deviation and dispersion coefficient of each parameter, and calibrating the distribution range of the mechanical parameters through multiple iterations of the aforementioned geotechnical mechanics tests;

[0009] C. Simulation of overhang rock instability: Import the corrected digital elevation model into the PFC software to generate a three-dimensional slope model, assign the above-mentioned calibrated mechanical parameter range to different areas of the three-dimensional slope model, and perform zoning processing according to the layered structure of the slope; adjust the microscopic parameters of the three-dimensional slope model through multiple iterations of mechanical parameters, and use the trial and error method to make the macroscopic response of the three-dimensional slope model simulation results consistent with the results of geotechnical mechanics tests; then set boundary conditions in the three-dimensional slope model, and simulate the falling process by freely releasing the overhang rock, and use virtual sensors to monitor and record key parameters including movement distance, maximum impact force, accumulation volume, impulse change, and lateral diffusion width in real time; analyze the changes of key parameters under different conditions through multiple simulated falls, and determine the correlation between the changes of key parameters and risk levels;

[0010] D. AHP construction evaluation system: Establish a hierarchical structure model including target layer, criterion layer and scheme layer, in which the target layer is used to evaluate the risk level of the disaster-affected area, the criterion layer is composed of key parameters affecting the instability of overhanging rocks as criteria, and the scheme layer is used to output the risk level of the disaster-affected area; construct a judgment matrix based on the pairwise comparison of key parameters in the criterion layer, and then calculate the weight vector by normalizing the judgment matrix, and check the consistency of the judgment matrix to ensure that the weight calculation result is credible; construct an evaluation function for evaluating the risk level of the disaster-affected area;

[0011] Based on historical data and combined with expert experience, the risk level classification criteria are set; then the key parameters simulated in step C are used to calculate the final risk score through the evaluation function, and the corresponding risk level is matched according to the risk score;

[0012] E. Construct CNN model: Construct a multi-dimensional disaster data set based on the key parameter data obtained by multiple simulations in step C, and divide the data set into training set, validation set and test set; then establish a CNN model including input layer, convolution layer, pooling layer, fully connected layer and output layer, and use the training set, validation set and test set to train, validate and test the CNN model respectively to obtain the optimal CNN model;

[0013] F. Classification of dangerous areas: The CNN model receives the key parameters of the overhanging rock after instability in real time, and quickly calculates and outputs the risk level corresponding to the dangerous area.

[0014] Furthermore, the specific process of step A is as follows:

[0015] A1. Obtaining point cloud data of the target area: Using an aircraft equipped with a high-precision positioning system and a laser radar, perform multiple high-resolution three-dimensional scans of the target area of ​​the overhanging rock at multiple angles to obtain point cloud data of the target area;

[0016] A2. DEM construction: ArcGIS was used to clean the aforementioned point cloud data to remove irrelevant factors, and then the point cloud data was processed with multi-layer resolution using an automatic boundary detection algorithm to extract the features of the overhanging rock and its surrounding terrain in layers to generate a high-precision digital elevation model;

[0017] A3. DEM correction: Combine existing topographic maps or field measurement data to perform multi-level correction on the digital elevation model, and use local high-resolution data processing technology to optimize the digital elevation model to obtain a digital elevation model.

[0018] Furthermore, the specific process of step B is as follows:

[0019] B1. Geotechnical sampling: Geotechnical samples are obtained by stratified random sampling of the overhanging rock mass and its surrounding slope areas;

[0020] B2. Mechanical testing: Conducting geotechnical mechanics testing on the collected rock and soil samples in the laboratory to obtain the mechanical parameters of the rock and soil samples including compressive strength, elastic modulus, Poisson's ratio, shear strength and internal friction angle;

[0021] B3. Parameter analysis: Statistical analysis is performed on each mechanical parameter to determine its distribution form, including normal distribution and log-normal distribution. Then, the mean value, standard deviation and coefficient of dispersion of each mechanical parameter are calculated. Subsequently, the range of values ​​of each mechanical parameter is further analyzed using the Monte Carlo simulation method, and the uncertainty model of each mechanical parameter is derived. Combined with the laboratory mechanical test results, a complete parameter distribution model is constructed.

[0022] B4. Parameter calibration: Combine Bayesian reasoning to perform uncertainty analysis on mechanical parameters, then dynamically update the distribution range of mechanical parameters by combining historical data with experimental results, and calibrate the distribution range of mechanical parameters based on the results of multiple iterative mechanical tests.

[0023] Furthermore, the specific process of step C is as follows:

[0024] C1. PFC modeling: The digital elevation model is imported into the PFC software to generate a three-dimensional slope model. According to the distribution range of mechanical parameters of the calibrated rock and soil samples, the three-dimensional slope model is divided into multiple regions. Then, the Monte Carlo random sampling method is used to extract sample values ​​that conform to the statistical distribution from the parameter distribution. Then, each region is assigned a randomly generated parameter value. The DEM model and the geological survey report are used to determine the hierarchical structure and joint distribution of the rock mass. Then, combined with the PFC particle generation algorithm, different partitions are generated according to different depths and lithology of the slope. Independent particle diameter, bonding strength and mechanical parameters are defined for each partition. Finally, a joint network simulation is applied to the joint development zone, and different contact stiffness and normal bonding force are assigned.

[0025] C2. Parameter adjustment: The three-dimensional slope model is iterated multiple times using the mechanical parameters calibrated in step B by trial and error method, so as to adjust the microscopic parameters of the three-dimensional slope model including contact stiffness and bonding strength to reflect the actual field conditions, so that the difference between the macroscopic response of the stress-strain curve and the laboratory mechanical test results in the simulation results of the three-dimensional slope model is less than 5%, and the three-dimensional slope model containing the most suitable microscopic parameters is obtained;

[0026] C3. Simulation of instability process: Boundary conditions are set in the three-dimensional slope model obtained in C2, and then the overhanging rock is simulated to fall under the action of gravity through free release. Virtual sensors are used to monitor and record key parameters including movement distance, maximum impact force, accumulation volume, impulse change, and lateral diffusion width in real time.

[0027] Furthermore, if the difference between the simulation results including the macroscopic response of the stress-strain curve and the results of the laboratory mechanical test in C2 is not less than 5%, the mesoscopic parameters of the three-dimensional slope model are iteratively adjusted by the trial and error method.

[0028] Furthermore, the specific process of step D is as follows:

[0029] D1. Construction of hierarchical model: A hierarchical model including target layer, criterion layer and scheme layer is established for evaluating the instability disaster of overhanging rock; the target layer is used to evaluate the risk level of the disaster-affected area, the criterion layer is composed of key parameters affecting the instability of overhanging rock as criteria, and the scheme layer is used to output the risk level of the disaster-affected area;

[0030] D2. Construction of judgment matrix: Construct a judgment matrix for pairwise comparison of the key parameters in the criterion layer, and determine the importance of each parameter relative to other parameters based on expert opinions or historical data;

[0031] D3. Calculation of weight vector: The weight vector calculated by normalizing the judgment matrix is ​​to first calculate the normalized value of each column in the judgment matrix, divide each element of the judgment matrix by the sum of its column, and obtain a normalized matrix; then calculate the average value of each row in the judgment matrix to obtain the weight vector of each criterion; the calculation formula is:

[0032] ,

[0033] in: W i For the i The weight vector of the criteria, a ij For the i The factor relative to j The importance ratio of the factors, n is the order of the matrix;

[0034] D4. Consistency test: calculate the consistency index first CI :

[0035] ,

[0036] Where: l max is the maximum eigenvalue of the judgment matrix;

[0037] Then calculate the consistency ratio CR :

[0038] ,

[0039] Where: RI It is a random consistency index obtained by looking up the table according to the order of the judgment matrix;

[0040] If CR<0.1, the consistency of the judgment matrix is ​​considered acceptable and the calculation results of the weight vector are credible; otherwise, the value of the judgment matrix needs to be readjusted;

[0041] D5. Evaluation function construction: Construct an evaluation function for evaluating the risk level of the disaster-affected area. The evaluation function multiplies the weight of each criterion with the actual monitored parameter value to obtain the risk assessment result; the evaluation function is:

[0042] ,

[0043] Where: S The final risk assessment result is W i For the i The weight of the parameters, X i The actual monitored i parameter values;

[0044] D6. Risk level classification: Based on historical data and combined with expert experience, a four-level risk classification standard is set, including low risk, medium risk, high risk and extremely high risk;

[0045] D7. Risk model construction: Construct a risk model that automatically corresponds to the risk assessment results obtained by calculating the evaluation function and the risk level. Input the key parameters simulated in step C into the risk model to obtain the corresponding risk level.

[0046] It should be noted that the risk level in D7: defines static risk standards as the baseline for CNN model training and evaluation; while the risk level in step F: is dynamically evaluated based on real-time data, combined with the baseline provided by the D7 risk model for calibration and verification;

[0047] Application relationship: The risk level in D7 provides the basis for model training and evaluation in step F. Step F predicts and updates the risk level in a more real-time and dynamic manner to achieve precise disaster assessment and early warning.

[0048] Furthermore, it also includes the dynamic update of the risk model: the risk model combines historical data, expert opinions and real-time monitoring data to dynamically adjust the weight vector of each criterion.

[0049] Further, the specific process of step E is as follows:

[0050] E1. Dataset construction: Obtain the data generated by multiple numerical simulations in step C to change the key parameters of the overhanging rock mass and slope, standardize the data, and then construct a multi-dimensional disaster dataset including training set, validation set and test set;

[0051] E2. CNN model construction: Establish a CNN model including input layer, convolution layer, pooling layer, fully connected layer and output layer; among them, the key parameters of the input layer after input standardization; the standardization formula is:

[0052] ,

[0053] Where: X is the original input data, X ' is the standardized data, m and s is the mean and standard deviation of the original data;

[0054] The convolution layer uses multiple convolution kernels to slide on the input data and automatically extract local nonlinear features from the input parameters; the convolution operation formula is:

[0055] ,

[0056] Where: Z i,j For the i The input data features are relative to the j Convolutional layer outputs with different features; W k is the convolution kernel, X i+k-1,j+k-1 is the value of the i+k-1,j+k-1 position of the input data; k is the convolution kernel k Weight index, corresponding to the weight parameter of the convolution kernel W k ; n is the size of the convolution kernel, that is, the total number of weights contained in the convolution kernel; i and j Represents the position index of the input data, representing the rows and columns of the input data matrix, or the rows and columns of the feature map; b is the bias, f ( x ) is the activation function, f ( x )=max(0, x ), ;

[0057] The pooling layer is used to reduce the dimension of the convolutional layer output, retain important features and reduce computational complexity; the commonly used maximum pooling operation is:

[0058] ;

[0059] The fully connected layer associates the features extracted by the convolutional layer with the output risk level through the connection of neurons. The calculation formula is:

[0060] ,

[0061] Where: Z It is the output matrix or tensor, representing the activation value passed from the previous layer; FCThat is, the fully connected layer is used to map the features extracted by the convolutional layer to the final output through linear combination; Z (FC) is the output of the fully connected layer, W (FC) is the weight matrix, b (FC) is the bias, f ( x ) is the activation function;

[0062] The output layer performs multi-classification prediction through the Softmax function, mapping the output of the fully connected layer to the probability of each risk level:

[0063] ,

[0064] Where: To predict the risk level j Probability, K is the number of risk levels, j Represents the position index of the column index in the output matrix. In the convolution operation, j Indexes each column in the output feature map, or the features to be calculated in each convolution window; X It is the original input data, including maximum impact force, accumulation volume, movement distance, impulse change, and lateral diffusion width;

[0065] E3. CNN model training: Use the training set in the multidimensional disaster dataset to train the CNN model, and then use the validation set and test set in the multidimensional disaster dataset to verify and test the CNN model respectively, and calculate the test accuracy. If the test accuracy of the CNN model is less than 0.95, readjust the parameters of the CNN model and train, verify and test again. Otherwise, take this CNN model as the optimal CNN model.

[0066] Furthermore, the specific process of step F is as follows:

[0067] F1. Real-time data input: Through sensors and monitoring systems, key parameters of overhanging rock after instability are obtained in real time and input into the optimal state CNN model;

[0068] F2. Risk level judgment: The optimal CNN model recognizes the complex relationship between parameters based on the input key data, and quickly judges and outputs the risk level corresponding to the current area.

[0069] Furthermore, the F step also includes: the risk level results output by the optimal CNN model are visualized in a three-dimensional topographic map through a GIS platform, and the GIS platform dynamically adjusts the risk level of each area in the visualization according to changes in monitoring data.

[0070] Particle contact stiffness: describes the elastic response of particles when they are in contact.

[0071] Normal bond strength: the maximum tensile force between particles.

[0072] Tangential bond strength: the maximum force between particles resisting shear.

[0073] Particle size distribution: The particle size range and distribution pattern of the generated particles.

[0074] Coefficient of friction: Quantification of the resistance to sliding between particles.

[0075] Density: The mass density of a granular material.

[0076] Beneficial effects of the present invention:

[0077] 1. The present invention performs high-resolution three-dimensional scanning on the target area of ​​the overhanging rock, and generates a high-precision digital elevation model after processing, which not only provides data support for the classification of dangerous areas after instability, thereby improving the accuracy of subsequent classification, but also solves the problem of difficult effective application of traditional surveys and excavations for overhanging rocks with high distribution positions, steep terrain, and dense surface vegetation through high-resolution three-dimensional scanning. In particular, geotechnical mechanics tests are performed on rock and soil samples in overhanging rocks and their surrounding areas, which further lays the foundation for improving the accuracy of intelligent classification of dangerous areas in the future.

[0078] 2. The present invention adopts PFC technology to determine the correlation between changes in key parameters and risk levels; then an evaluation system is constructed through AHP, thereby constructing an evaluation function for evaluating the risk level of the disaster-affected area and setting a risk level classification standard, and the final risk score is calculated through the evaluation function to quantitatively match the corresponding risk level; then a CNN model is constructed, and the key parameters after the overhang rock is instability obtained in real time are received through the CNN model, and the risk level corresponding to the dangerous area is quickly calculated and output, thereby realizing the intelligent risk level classification, which not only effectively solves the problem of low timeliness of the existing risk level classification, but also reduces the subjective factors in the classification process, and can also simulate the instability process by using a three-dimensional slope model with the most suitable mesoscopic parameters before instability, and monitor and record dynamic data such as maximum impact force and impulse change in real time through virtual sensors during the process, so that the later CNN model can combine dynamic and static data when classifying the risk level, effectively improving the accuracy of the risk level classification results, and can eliminate the safety hazards in the disaster relief, disaster reduction and post-disaster reconstruction after instability.

[0079] In summary, the present invention has the characteristics of high timeliness, less human intervention and accurate judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1A schematic diagram of three-dimensional scanning of the drone of the present invention;

[0081] Figure 2 is a flow chart of the present invention;

[0082] Figure 3 This is a flowchart of point cloud data processing in an embodiment of the present invention;

[0083] Figure 4 This is a flow chart of overhang rock instability simulation in an embodiment of the present invention;

[0084] Figure 5 A flowchart of constructing a CNN model and grading dangerous areas in an embodiment of the present invention;

[0085] In the figure: 1- bedrock, 2- dangerous rock mass, 3- slope, 4- drone equipped with lidar, 5- target area. DETAILED DESCRIPTION

[0086] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0087] like Figure 1 As shown, the present invention includes the steps of obtaining a digital elevation model, analyzing rock and soil mechanics, simulating the instability of overhanging rocks, constructing an AHP evaluation system, constructing a CNN model, and classifying dangerous areas. The specific contents are as follows:

[0088] A. Acquisition of digital elevation model: Perform high-resolution 3D scanning of the target area of ​​the overhanging rock to obtain point cloud data; then clean the point cloud data and process it at multiple resolutions to generate a high-precision digital elevation model (DEM); then perform precision correction on the digital elevation model based on the actual terrain conditions on site;

[0089] B. Geotechnical mechanics analysis: Sampling the target area and conducting geotechnical mechanics tests on the geotechnical samples to obtain mechanical parameters; statistically analyzing the mechanical parameters to determine their distribution form, calculating the mean value, standard deviation and dispersion coefficient of each parameter, and calibrating the distribution range of the mechanical parameters through multiple iterations of the aforementioned geotechnical mechanics tests;

[0090] C. Simulation of the instability of overhanging rock: Import the corrected digital elevation model into the PFC (Particle Flow Code) software to generate a three-dimensional slope model. Assign the calibrated mechanical parameter ranges to different regions of the three-dimensional slope model and perform zoning processing according to the layered structure of the slope. Adjust the mesoscopic parameters of the three-dimensional slope model through multiple iterations of mechanical parameters, and use the trial-and-error method to make the macroscopic response of the simulation results of the three-dimensional slope model consistent with the results of geotechnical mechanics tests. Then set the boundary conditions in the three-dimensional slope model and simulate the falling process by freely releasing the overhanging rock. Use virtual sensors to monitor and record the key parameters including the movement distance, maximum impact force, accumulation volume, impulse change, and lateral diffusion width in real time. Analyze the changes of key parameters under different conditions through multiple simulations of the falling process to determine the correlation between the changes of key parameters and the risk level.

[0091] D. Construction of an AHP evaluation system: Establish a hierarchical structure model including an objective layer, a criterion layer, and a scheme layer. The objective layer is used to evaluate the risk level of the disaster-affected area. The criterion layer is composed of the key parameters affecting the instability of the overhanging rock as criteria. The scheme layer is used to output the risk level of the disaster-affected area. Construct a judgment matrix based on the pairwise comparison of key parameters in the criterion layer, and then calculate the weight vector through the normalized judgment matrix and check the consistency of the judgment matrix to ensure the credibility of the weight calculation results. Construct an evaluation function for evaluating the risk level of the disaster-affected area.

[0092] Based on historical data and combined with expert experience, set the classification criteria for risk levels. Then, use the evaluation function to calculate the final risk score for the key parameters obtained from the simulation in step C, and match the corresponding risk level according to the risk score.

[0093] E. Construction of a CNN model: Construct a multi-dimensional disaster dataset based on the key parameter data obtained from multiple simulations in step C, and divide the dataset into a training set, a validation set, and a test set. Then establish a CNN model including an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer, and use the training set, the validation set, and the test set to train, validate, and test the CNN model respectively to obtain the CNN model in the optimal state.

[0094] F. Classification of dangerous area levels: The CNN model receives the key parameters after the instability of the overhanging rock obtained in real time, and quickly calculates and outputs the risk level corresponding to the dangerous area.

[0095] It should be noted that the fuzzy analytic hierarchy process (AHP) is a multi-criteria decision analysis tool that can effectively handle complex decision problems, especially when the decision involves multiple interrelated factors. The basic idea of ​​AHP is to decompose complex problems into several levels, compare each factor pairwise, construct a judgment matrix, calculate the relative importance weight of each factor, and thus form a decision model.

[0096] It should be noted that the convolutional neural network (CNN) is a deep learning model that can automatically extract data features and is suitable for processing multidimensional data. In disaster risk assessment, the CNN model can automatically identify the complex relationship between disaster parameters, thereby effectively assessing the risk level in different scenarios.

[0097] The specific process of step A is as follows:

[0098] A1. Obtaining point cloud data of the target area: Using an aircraft equipped with a high-precision positioning system and a laser radar, perform multiple high-resolution three-dimensional scans of the target area of ​​the overhanging rock at multiple angles to obtain point cloud data of the target area;

[0099] A2. DEM construction: ArcGIS was used to clean the aforementioned point cloud data to remove irrelevant factors, and then the point cloud data was processed with multi-layer resolution using an automatic boundary detection algorithm to extract the features of the overhanging rock and its surrounding terrain in layers to generate a high-precision digital elevation model;

[0100] A3. DEM correction: Combine existing topographic maps or field measurement data to perform multi-level correction on the digital elevation model, and use local high-resolution data processing technology to optimize the digital elevation model to obtain a digital elevation model.

[0101] The specific process of step B is as follows:

[0102] B1. Geotechnical sampling: Geotechnical samples are obtained by stratified random sampling of the overhanging rock mass and its surrounding slope areas;

[0103] B2. Mechanical testing: Conducting geotechnical mechanics testing on the collected rock and soil samples in the laboratory to obtain the mechanical parameters of the rock and soil samples including compressive strength, elastic modulus, Poisson's ratio, shear strength and internal friction angle;

[0104] B3. Parameter analysis: Statistical analysis is performed on each mechanical parameter to determine its distribution form, including normal distribution and log-normal distribution. Then, the mean value, standard deviation and coefficient of dispersion of each mechanical parameter are calculated. Subsequently, the range of values ​​of each mechanical parameter is further analyzed using the Monte Carlo simulation method, and the uncertainty model of each mechanical parameter is derived. Combined with the laboratory mechanical test results, a complete parameter distribution model is constructed.

[0105] B4. Parameter calibration: Combine Bayesian reasoning to perform uncertainty analysis on mechanical parameters, then dynamically update the distribution range of mechanical parameters by combining historical data with experimental results, and calibrate the distribution range of mechanical parameters based on the results of multiple iterative mechanical tests.

[0106] In step B3, the Monte Carlo simulation method is used to further analyze the value range of each mechanical parameter, and the specific process of deriving the uncertainty model of each mechanical parameter is as follows:

[0107] B31.Data preparation:

[0108] Based on the parameter data obtained from laboratory mechanical tests (such as compressive strength, elastic modulus, Poisson's ratio, etc.), preliminarily calculate the mean value, standard deviation, and dispersion coefficient; determine the distribution type of each parameter, such as normal distribution, lognormal distribution, or other suitable distribution;

[0109] B32. Randomly generate samples:

[0110] Using the Monte Carlo method, a large number of sample values ​​(usually more than 10,000 times) are randomly generated based on the above distribution type and statistical parameters (such as mean and standard deviation); the values ​​generated in each simulation reflect the possible actual range of variation of the parameters;

[0111] B33.Uncertainty analysis:

[0112] Perform statistical analysis on the parameter samples generated by simulation and calculate their mean, standard deviation, and confidence interval; use probability density function (PDF) to describe the uncertainty model of the parameters; if necessary, superimpose multi-source data (such as historical observations, other experimental data) to further calibrate the model;

[0113] B34. Result verification:

[0114] Verify that simulation results are consistent with laboratory data (e.g., using the Kolmogorov-Smirnov test); ensure that simulated distributions represent parameter uncertainties.

[0115] The specific process of constructing a complete parameter distribution model in step B3 is as follows:

[0116] B36. Data integration and modeling: Combine laboratory test results with data generated by Monte Carlo simulation, and for each parameter (such as compressive strength and elastic modulus), construct a probability distribution model of the parameter based on random samples; Common distribution types: Normal distribution: suitable for the case where the parameter is symmetrically distributed around the mean; Among them, log-normal distribution: suitable for the case where the parameter has a right-skewed distribution; three-parameter Weibull distribution: suitable for describing the distribution of strength characteristics; Use goodness-of-fit tests (such as chi-square test, Kolmogorov-Smirnov test) to select the best distribution type;

[0117] B37. Distribution parameter extraction: determine the key parameters of the distribution model (such as mean, standard deviation, shape parameter), and use the fitted distribution to generate parameter distribution curves for subsequent applications;

[0118] B38. Parameter assignment: Apply parameter distribution models to numerical simulations (e.g., PFC modeling), randomly extract or assign values ​​based on the distribution range; different model regions (e.g., weathering layer and bedrock layer) use values ​​from different ranges in the distribution;

[0119] B39. Risk assessment: In risk classification, use the values ​​generated by the parameter distribution model to simulate extreme scenarios; predict risks through parameter changes (such as using the AHP model for evaluation);

[0120] B3A. Sensitivity analysis: Based on the distribution model, sensitivity analysis of key parameters (such as compressive strength or friction coefficient) is performed to evaluate the impact of different parameters on slope stability or overhang instability;

[0121] B3B. Model optimization: In dynamic risk assessment, real-time monitoring data and distribution models are combined to dynamically adjust parameter ranges and weights to ensure the accuracy and robustness of prediction results.

[0122] The specific process of step B4 is as follows:

[0123] B41.Data preparation:

[0124] Experimental data: including the mean, standard deviation, and coefficient of dispersion of parameters (such as compressive strength, elastic modulus, Poisson's ratio, etc.) obtained from laboratory mechanical tests;

[0125] Historical data: records of mechanical parameters from similar geological conditions or sites;

[0126] Prior distribution: the initial distribution of the assumed parameter (prior distribution), such as normal distribution or lognormal distribution; the prior distribution can be estimated based on experience, previous literature data, or initial experimental results;

[0127] B42. Uncertainty analysis using Bayesian reasoning:

[0128] Implementation steps: Construct a likelihood function based on experimental data and distribution type, use historical data as prior information, and define prior distribution; combine the likelihood function and prior distribution, and calculate the posterior distribution through the Bayesian formula;

[0129] B43. Dynamic update of mechanical parameter distribution range:

[0130] Use the posterior distribution as the prior distribution for the next step, combine it with new experimental data or field monitoring data, and repeatedly apply Bayesian reasoning; continuously iterate the posterior distribution, gradually calibrate the parameter distribution range, and make it gradually stable;

[0131] B44. Verification of calibration results:

[0132] The calibrated parameter distribution is tested for goodness of fit with the experimental data, for example using the Kolmogorov-Smirnov test or the chi-square test, to ensure that the distribution is consistent with the data characteristics.

[0133] It should be noted that the application of calibrating the distribution range of mechanical parameters in step B4 is:

[0134] 1. In step C1, the calibrated distribution range of mechanical parameters is used as the basis for assigning values ​​to different areas of the 3D slope model; specifically, the geotechnical properties of different areas are randomly sampled or assigned optimal values ​​based on the distribution range, thereby ensuring that the model truly simulates the actual geological characteristics of the site.

[0135] 2. The distribution range of mechanical parameters provides uncertainty boundaries in multiple simulation iterations, allowing the simulation results to cover possible extreme conditions; in the subsequent model adjustment (C2), the trial and error method is used to iterate and ensure that the mesoscopic parameters of the model reflect the actual field conditions.

[0136] 3. In steps D and E, the distribution range of mechanical parameters ensures the adaptability of disaster risk assessment to various parameter values; for example, the mean value, standard deviation and dispersion coefficient of different regions directly affect the setting of evaluation function weights and the diversity of CNN model training data.

[0137] 4. The distribution range calibration of mechanical parameters provides reliable input data for the subsequent risk model (step F), improving the prediction accuracy of the CNN model, especially when the data is updated in real time.

[0138] It should be noted that the different conditions in step C include:

[0139] Topographic conditions: changes in slope, joint distribution, and slope structure;

[0140] External conditions: increased ground motion, water pressure or weathering effects;

[0141] Parameter conditions: adjust normal adhesion and contact stiffness.

[0142] It should be noted that the simulation methods in step C include:

[0143] Change parameters one by one and keep other conditions fixed to test the effect of a single factor on the results;

[0144] Combinations of different conditions: such as high slope + low adhesion to simulate extreme conditions.

[0145] The correlation between the changes in key parameters and the risk level in step C is determined by performing multiple simulations, recording the changes in key parameters (such as impact force and movement distance), and using linear regression or nonlinear analysis to find the relationship between the key parameters and the risk score; then calculating the correlation coefficient to evaluate the contribution rate of the parameters to the risk level.

[0146] In the AHP model, the D step quantifies the correlation between the parameter and the risk level in the form of weight.

[0147] In the E step, key parameters are input as features in the CNN model training to optimize the risk prediction results.

[0148] The specific process of step C is as follows:

[0149] C1. PFC modeling: The digital elevation model is imported into the PFC software to generate a three-dimensional slope model. According to the distribution range of mechanical parameters of the rock and soil samples calibrated above, the three-dimensional slope model is divided into multiple regions (such as the upper layer, middle layer, and bottom layer of the slope). Then, the Monte Carlo random sampling method is used to extract sample values ​​that conform to the statistical distribution from the parameter distribution. Subsequently, each region is assigned a randomly generated parameter value (such as contact stiffness, bond strength, and porosity, etc.). The DEM model and geological survey report are used to determine the hierarchical structure and joint distribution of the rock mass. Then, combined with the PFC particle generation algorithm, different partitions are generated according to different depths and lithology of the slope (such as dividing the model into the top weathering layer, the main rock layer of the slope, and the bottom bedrock layer). Independent particle diameter, bond strength, and mechanical parameters are defined for each partition. Finally, a joint network simulation is applied to the joint development zone, and different contact stiffness and normal bond force are assigned.

[0150] C2. Parameter adjustment: The three-dimensional slope model is iterated multiple times using the mechanical parameters calibrated in step B by trial and error method, so as to adjust the microscopic parameters of the three-dimensional slope model including contact stiffness and bonding strength to reflect the actual field conditions, so that the difference between the macroscopic response of the stress-strain curve and the laboratory mechanical test results in the simulation results of the three-dimensional slope model is less than 5%, and the three-dimensional slope model containing the most suitable microscopic parameters is obtained;

[0151] C3. Simulation of instability process: Boundary conditions are set in the three-dimensional slope model obtained in C2, and then the overhanging rock is simulated to fall under the action of gravity through free release. Virtual sensors are used to monitor and record key parameters including movement distance, maximum impact force, accumulation volume, impulse change, and lateral diffusion width in real time.

[0152] If the difference between the simulation results including the macroscopic response of the stress-strain curve and the results of the laboratory mechanical test in C2 is not less than 5%, the mesoscopic parameters of the three-dimensional slope model are iteratively adjusted using the trial and error method.

[0153] Examples of assigning values ​​to different regions in step C1:

[0154] Upper weathering zone: low strength, low adhesion, high porosity;

[0155] Middle rock mass: medium strength, moderate bonding, medium porosity;

[0156] Lower bedrock: high strength, high adhesion, low porosity.

[0157] The specific process of step C2 is as follows:

[0158] C21. Initial parameter setting:

[0159] According to the macroscopic mechanical parameters (such as compressive strength, elastic modulus, shear strength, etc.) calibrated in step B, they are converted into initial values ​​of microscopic parameters, including: contact stiffness;

[0160] Bond strength, friction coefficient: set the initial friction coefficient according to the rock mass characteristics;

[0161] C22.Simulation and comparison:

[0162] A three-dimensional slope model is set up in PFC (Particle Flow Code), and a load consistent with the experimental conditions (such as uniaxial compression and shear loading) is applied to the model to obtain a simulated stress-strain curve. The simulation results are then compared with the stress-strain curve of the laboratory test, and the error is calculated. If the deviation is greater than 5%, the parameters are adjusted.

[0163] C23. Parameter adjustment:

[0164] Adjust the microscopic parameters step by step according to the error type: if the simulated elastic modulus is lower than the experimental value, increase the normal stiffness and tangential stiffness; if the failure strength is lower than the experimental value, increase the normal bond strength and tangential bond strength; if the plastic stage of the stress-strain curve is not obvious, adjust the friction coefficient; re-simulate after each adjustment and record the parameter changes and corresponding errors.

[0165] C24. Iteration termination condition:

[0166] When the deviation between the stress-strain curve of the simulation result and the experimental result is less than 5%, the iteration is stopped;

[0167] C25. Record and save: Save the micro-parameter values ​​and corresponding simulation results of each iteration to form a record of the entire process of parameter optimization.

[0168] It should be noted that the boundary conditions set in C3 include:

[0169] Foundation constraint: The bottom is fixed to ensure that the bedrock does not move as a whole;

[0170] Free boundary: The sides allow particles to flow to simulate natural collapse;

[0171] Gravity loading: Apply true gravity conditions, taking into account terrain height differences;

[0172] Initial stress field: Apply ground stress to simulate the initial stress state inside the rock mass.

[0173] It should be noted that in C3, the overhanging rock is simulated to fall under the action of gravity by free release. In most cases, the fall can be achieved by releasing the gravity. If the fall needs to be triggered additionally, the following external force can be applied:

[0174] Earthquake loading: applying periodic horizontal vibration to simulate earthquake triggering;

[0175] Water pressure: simulates the increase of water pressure in rock fractures caused by rainfall;

[0176] Among them, the specific size of the applied load is:

[0177] Water pressure: Based on the porosity of rock mass fractures, apply 10 to 50 kPa of water pressure;

[0178] Earthquake acceleration: It is recommended not to exceed 0.2g to match the actual earthquake scale.

[0179] The specific process of step D is as follows:

[0180] D1. Construction of hierarchical model: A hierarchical model including target layer, criterion layer and scheme layer is established for evaluating the instability disaster of overhanging rock; the target layer is used to evaluate the risk level of the disaster-affected area, the criterion layer is composed of key parameters affecting the instability of overhanging rock as criteria, and the scheme layer is used to output the risk level of the disaster-affected area;

[0181] D2. Construction of judgment matrix: Construct a judgment matrix for pairwise comparison of the key parameters in the criterion layer, and determine the importance of each parameter relative to other parameters based on expert opinions or historical data; the form of the judgment matrix is ​​shown in Table 1:

[0182] Table 1 Judgment matrix form

[0183]

[0184] In the matrix: a ij Indicates i The factor relative to j The importance ratio of the factors;

[0185] Importance ratios are usually obtained from expert ratings or historical data. In general, a scale of 1 to 9 is used to represent the comparison level from equally important to extremely important. For example, 1: Both factors are equally important; 3: One factor is slightly more important than the other; 5: One factor is significantly more important than the other; 7: One factor is very important than the other; 9: One factor is extremely important than the other.

[0186] D3. Calculation of weight vector: The weight vector calculated by normalizing the judgment matrix is ​​to first calculate the normalized value of each column in the judgment matrix, divide each element of the judgment matrix by the sum of its column, and obtain a normalized matrix; then calculate the average value of each row in the judgment matrix to obtain the weight vector of each criterion; the calculation formula is:

[0187] ,

[0188] in: W i For the i The weight vector of the criteria, a ij For the i The factor relative to j The importance ratio of the factors, n is the order of the matrix;

[0189] D4. Consistency test: calculate the consistency index first CI :

[0190] ,

[0191] Where: l max is the maximum eigenvalue of the judgment matrix;

[0192] Then calculate the consistency ratio CR :

[0193] ,

[0194] Where: RI It is a random consistency index obtained by looking up the table according to the order of the judgment matrix;

[0195] If CR<0.1, the consistency of the judgment matrix is ​​considered acceptable and the calculation results of the weight vector are credible; otherwise, the value of the judgment matrix needs to be readjusted;

[0196] D5. Evaluation function construction: Construct an evaluation function for evaluating the risk level of the disaster-affected area. The evaluation function multiplies the weight of each criterion with the actual monitored parameter value to obtain the risk assessment result; the evaluation function is:

[0197] ,

[0198] Where: S The final risk assessment result is W i For the i The weight of the parameters, X i The actual monitored i parameter values;

[0199] D6. Risk level classification: Based on historical data and combined with expert experience, a four-level risk classification standard is set, including low risk, medium risk, high risk and extremely high risk;

[0200] D7. Risk model construction: Construct a risk model that automatically corresponds to the risk assessment results obtained by calculating the evaluation function and the risk level. Input the key parameters simulated in step C into the risk model to obtain the corresponding risk level.

[0201] The specific process of re-adjusting the value of the judgment matrix in D4 is as follows:

[0202] D41. Determine the matrix elements that need to be adjusted:

[0203] (1) Calculate the maximum eigenvalue and corresponding eigenvector of the judgment matrix;

[0204] (2) Analyze the deviation between the normalized weight vector of each column and the actual input judgment matrix elements;

[0205] (3) Find the elements with large deviations (i.e., the parts where the eigenvectors do not match the weight relationship calculated by the original matrix);

[0206] D42. Adjust inconsistent elements:

[0207] (1) Re-adjust the elements with large deviations based on expert opinions and the judgment matrix scale (1 to 9) a ij ;

[0208] (2) Ensure that the consistency relationship is met. If not, correct the most inconsistent value so that the adjusted relationship is close to consistency;

[0209] D43. Re-verify consistency.

[0210] The step D also includes dynamic updating of the risk model: the risk model combines historical data, expert opinions and real-time monitoring data to dynamically adjust the weight vector of each criterion.

[0211] The specific process of step E is as follows:

[0212] E1. Dataset construction: Obtain the data generated by multiple numerical simulations in step C to change the key parameters of the overhanging rock mass and slope, standardize the data, and then construct a multi-dimensional disaster dataset including training set, validation set and test set;

[0213] The key parameters include impact force, accumulation volume, movement distance, impulse and lateral diffusion width. The key parameters are simulated multiple times to obtain combinations under different scenarios. A large number of different parameter combinations are used to construct a multi-dimensional disaster data set.

[0214] E2. CNN model construction: Establish a CNN model including input layer, convolution layer, pooling layer, fully connected layer and output layer; among them, the key parameters of the input layer after input standardization; the standardization formula is:

[0215] ,

[0216] Where: X is the original input data, X ' is the standardized data, m and s is the mean and standard deviation of the original data;

[0217] The convolution layer uses multiple convolution kernels to slide on the input data and automatically extract local nonlinear features from the input parameters; the convolution operation formula is:

[0218] ,

[0219] Where: Z i,j For the i The input data features are relative to the j Convolutional layer outputs with different features; W k is the convolution kernel, X i+k-1,j+k-1 is the value of the i+k-1,j+k-1 position of the input data; k is the convolution kernel k Weight index, corresponding to the weight parameter of the convolution kernel W k ; nis the size of the convolution kernel, that is, the total number of weights contained in the convolution kernel; i and j Represents the position index of the input data, representing the rows and columns of the input data matrix, or the rows and columns of the feature map; b is the bias, f ( x ) is the activation function, f ( x )=max(0, x ), ;

[0220] The pooling layer is used to reduce the dimension of the convolutional layer output, retain important features and reduce computational complexity; the commonly used maximum pooling operation is:

[0221] ;

[0222] The fully connected layer associates the features extracted by the convolutional layer with the output risk level through the connection of neurons. The calculation formula is:

[0223] ,

[0224] Where: Z is the output matrix or tensor, which represents the activation value passed from the previous layer; FC is the fully connected layer, which is used to map the features extracted by the convolutional layer to the final output through linear combination; Z (FC) is the output of the fully connected layer, W (FC) is the weight matrix, b (FC) is the bias, f ( x ) is the activation function;

[0225] The output layer performs multi-classification prediction through the Softmax function, mapping the output of the fully connected layer to the probability of each risk level:

[0226] ,

[0227] Where: To predict the risk level j Probability, K is the number of risk levels, j Represents the position index of the column index in the output matrix. In the convolution operation, j Indexes each column in the output feature map, or the features to be calculated in each convolution window; X It is the original input data, including maximum impact force, accumulation volume, movement distance, impulse change, and lateral diffusion width;

[0228] E3. CNN model training: Use the training set in the multidimensional disaster dataset to train the CNN model, and then use the validation set and test set in the multidimensional disaster dataset to verify and test the CNN model respectively, and calculate the test accuracy. If the test accuracy of the CNN model is less than 0.95, readjust the parameters of the CNN model and train, verify and test again. Otherwise, take this CNN model as the optimal CNN model.

[0229] It should be noted that X in the E2 step is the input data of CNN, including maximum impact force, accumulation volume, movement distance, impulse change, and lateral diffusion width.

[0230] It should be noted that the risk level in step E2 is not the static risk level obtained through simulation in step D; the risk level in step D is calculated through historical data and evaluation functions, and is usually a risk level classification based on known data or simulated data; while in step F, the CNN model is used to predict the risk level in real time, and it performs dynamic evaluation based on key parameters input in real time (such as real-time monitoring data); therefore, the risk level in step F is predicted from the input data through the trained CNN model, which predicts the current risk status based on real-time data (such as monitored impact force, accumulation volume, etc.).

[0231] It should be noted that in the E2 step To predict the risk level j Probability involves the forward propagation process of the convolutional neural network (CNN). The following are the specific implementation steps:

[0232] E21. Input data (X): First, pass the input data into the CNN model;

[0233] E22. Convolution layer: The input data is convolved through the convolution kernel to obtain the feature map after convolution. The convolution operation is essentially a weighted summation of the local area in the input data;

[0234] E23. Pooling layer: The pooling operation further reduces the dimension of the feature map or extracts the most significant features (e.g., maximum pooling).

[0235] E24. Fully connected layer (FC): The feature maps extracted by the convolutional layer and the pooling layer are flattened and input into the fully connected layer. After being weighted by the weight (W(FC)) and the activation function (usually ReLU), the final output (such as classification result or risk level) is calculated.

[0236] E25. Final output: Based on the output of the CNN model, a prediction result is given, such as risk level or classification label.

[0237] The specific process of step F is as follows:

[0238] F1. Real-time data input: Through sensors and monitoring systems, key parameters of overhanging rock after instability are obtained in real time and input into the optimal state CNN model;

[0239] F2. Risk level judgment: The optimal CNN model recognizes the complex relationship between parameters based on the input key data, and quickly judges and outputs the risk level corresponding to the current area.

[0240] The F step also includes: the risk level results output by the optimal CNN model are visualized in a three-dimensional topographic map through a GIS platform, and the GIS platform dynamically adjusts the risk level of each area in the visualization according to changes in monitoring data.

[0241] It should be noted that the recognition of complex relationships between parameters in F2 is achieved through the CNN model. The input key data (such as impact force, accumulation volume, etc.) undergoes convolution operations, pooling operations, and fully connected layers, which can automatically recognize and learn the complex nonlinear relationships between these parameters. The multi-layer structure of CNN enables it to capture the local, global, temporal and spatial dependencies between parameters at different levels, and finally derive the corresponding risk level through the output layer.

[0242] Example 1

[0243] S100: Perform high-resolution 3D scanning on the target area of ​​the overhanging rock to obtain point cloud data; then clean the point cloud data and process it at multiple resolutions to generate a high-precision digital elevation model (DEM); then perform precision correction on the digital elevation model based on the actual terrain conditions on site. The specific process is as follows:

[0244] S110: Using drones equipped with high-precision positioning systems such as RTK GPS, inertial measurement units (IMUs) and laser radar (LiDAR), the target area of ​​the overhanging rock is scanned multiple times at multiple angles and with high resolution to obtain point cloud data of the target area. The drone uses multiple flight missions, different heights and angles to make the LiDAR scan multiple angles to fully capture the subtle features of the overhanging rock mass and slopes, ensuring the comprehensiveness of the terrain data.

[0245] S120: ArcGIS was used to clean the aforementioned point cloud data to remove irrelevant factors such as vegetation and loose soil. The point cloud data was then processed with multi-layer resolution using an automatic boundary detection algorithm to extract the features of the overhanging rock and its surrounding terrain in a hierarchical manner to ensure accurate extraction of the overhanging rock boundary. Finally, a high-precision digital elevation model (DEM) was generated.

[0246] S130: Combined with existing topographic maps or field measurement data, the digital elevation model is corrected at multiple levels, and the digital elevation model is optimized using local high-resolution data processing technology, so that the accuracy of the digital elevation model is improved to the sub-meter level, and the detailed features of the overhanging rock such as joint surfaces, fault zones and slope changes are retained to obtain the digital elevation model.

[0247] S200: The specific process of geotechnical mechanics analysis is as follows:

[0248] S210: Conduct systematic geological exploration of the overhanging rock and its surrounding slope areas, design a sampling plan, and then obtain rock and soil samples by stratified random sampling of the overhanging rock mass and its surrounding slope areas according to the sampling plan. When sampling, focus on collecting rock samples near the joint surface of the overhanging rock, as well as rock samples of different heights and different degrees of weathering; the number of samples should be adjusted according to the regional geological characteristics and the specific conditions of the overhanging rock mass to ensure the adequacy and representativeness of the samples.

[0249] S220: The collected rock and soil samples are subjected to rock and soil mechanics tests including uniaxial compression test, triaxial compression test, tensile test and direct shear test in the laboratory to obtain the mechanical parameters of the rock and soil samples including compressive strength, elastic modulus, Poisson's ratio, shear strength and internal friction angle.

[0250] S230: Perform statistical analysis on each mechanical parameter to determine its distribution form including normal distribution and lognormal distribution, then calculate the mean value, standard deviation and coefficient of dispersion of each mechanical parameter, and then use the Monte Carlo simulation method to further analyze the value range of each mechanical parameter, derive the uncertainty model of each mechanical parameter, and build a complete parameter distribution model in combination with the laboratory mechanical test results.

[0251] S240: In order to improve the accuracy of the aforementioned mechanical parameters and the robustness of the model, uncertainty analysis of the mechanical parameters is performed in combination with Bayesian reasoning. Then, the distribution range of the mechanical parameters is dynamically updated by combining historical data with experimental results. The distribution range of the mechanical parameters is calibrated by the results of multiple iterative mechanical tests.

[0252] S300: The specific process of overhanging rock instability simulation is as follows:

[0253] S310: Import the digital elevation model into the PFC (Particle Flow Code) software to generate a three-dimensional slope model, assign the mechanical parameters obtained above to different areas of the three-dimensional slope model, and perform partition processing according to the layered structure of the slope. To ensure the accuracy of the numerical simulation, the different rock layers in the slope and the mesoscopic particles of the PFC must maintain consistent mechanical behavior characteristics.

[0254] S320: The three-dimensional slope model is iterated multiple times using the mechanical parameters calibrated by S200 by trial and error method, so as to adjust the microscopic parameters of the three-dimensional slope model including contact stiffness and bonding strength to reflect the actual field conditions; wherein the multiple iterations are performed until the difference between the macroscopic response in the simulation results of the three-dimensional slope model and the results of the laboratory mechanical test is less than 5%, thereby obtaining a three-dimensional slope model containing the most suitable microscopic parameters.

[0255] S330: Set boundary conditions in the three-dimensional slope model obtained in S320, and then simulate the falling process of the overhanging rock under the action of gravity by free release, observe its movement path, speed change, movement trajectory after collision, and capture the contact process between the overhanging rock and the slope ground, and use virtual sensors to monitor and record key parameters including movement distance, maximum impact force, accumulation volume, impulse change, and lateral diffusion width in real time. In particular, during the falling process of the overhanging rock, the movement trajectory, energy change in the impact area, and ground accumulation morphology are fully monitored and analyzed.

[0256] S400: AHP constructs an evaluation system. The specific process is as follows:

[0257] S410: Establish a hierarchical model including a target layer, a criterion layer and a scheme layer, which are arranged in layers and used to evaluate the instability disaster of overhang rocks; wherein the target layer is the highest layer, which is used to evaluate the risk level of the disaster-affected area; the criterion layer is the middle layer, which is composed of key parameters affecting the instability of overhang rocks as criteria; the scheme layer is the lowest layer, which is used to output the risk level (low, medium, high, and extremely high) of the disaster-affected area.

[0258] Among them, the key parameters include: (1) maximum impact force: the force generated when the overhanging rock collides with the ground or other obstacles when it falls; (2) accumulation volume: the volume of debris accumulation formed after the overhanging rock becomes unstable; (3) movement distance: the length of the movement path of the overhanging rock from the instability point to the end point; (4) impulse change: the change in kinetic energy during the movement of the overhanging rock; (5) lateral diffusion width: the lateral dispersion range of the falling debris on the ground.

[0259] S420: Construct a judgment matrix for pairwise comparison of each parameter in the key parameters of the criterion layer, and determine the importance of each parameter relative to other parameters based on expert opinions or historical data; the judgment matrix is ​​in the form of Table 1:

[0260] Table 1 Judgment matrix form

[0261]

[0262] In the matrix: a ij Indicates i The factor relative to j The importance ratio of the factors.

[0263] Importance ratios are usually obtained from expert ratings or historical data. In general, a scale of 1 to 9 is used to represent the comparison level from equally important to extremely important. For example, 1: Both factors are equally important; 3: One factor is slightly more important than the other; 5: One factor is significantly more important than the other; 7: One factor is very important than the other; 9: One factor is extremely important than the other.

[0264] S430: Calculate the weight vector by normalizing the judgment matrix: first calculate the normalized value of each column in the judgment matrix, divide each element of the judgment matrix by the sum of its column, and obtain the normalized matrix; then calculate the average value of each row in the judgment matrix to obtain the weight vector of each criterion; the calculation formula is:

[0265] ,

[0266] in: W i For the i The weight vector of the criteria, a ij For the i The factor relative to j The importance ratio of the factors, n is the order of the matrix;

[0267] S440: Ensure that the judgment matrix construction is reasonable through consistency test: first calculate the consistency index CI :

[0268] ,

[0269] Where: l max is the maximum eigenvalue of the judgment matrix.

[0270] Then calculate the consistency ratio CR :

[0271] ,

[0272] Where: RI It is a random consistency index obtained by looking up the table according to the order of the judgment matrix.

[0273] If CR<0.1, the consistency of the judgment matrix is ​​considered acceptable and the calculation results of the weight vector are credible; otherwise, the value of the judgment matrix needs to be readjusted.

[0274] S450: Construct an evaluation function for evaluating the risk level of the disaster-affected area. The evaluation function multiplies the weight of each criterion with the actual monitored parameter value to obtain a risk assessment result. The evaluation function is:

[0275] ,

[0276] Where: S The final risk assessment result is W i For the i The weight of the parameters, X i The actual monitored i parameter values.

[0277] S460: Based on the historical data of key parameters such as movement distance, maximum impact force, accumulation volume, impulse change, lateral diffusion width, etc., combined with the experience and opinions of experts, a four-level risk classification standard is set, including low risk, medium risk, high risk and extremely high risk.

[0278] Among them: (1) Low risk - the destructive power caused by the instability of the overhanging rock is small, the impact range is limited, and the response measures required are minor; (2) Medium risk - the fall of the overhanging rock may cause certain damage to the ground, and conventional disaster prevention measures need to be taken; (3) High risk - the fall may cause damage to a large area, posing a great potential threat to personnel and facilities, and emergency preventive measures need to be taken; (4) Extremely high risk - the instability of the overhanging rock may cause large-scale disasters with a wide impact range, and protective and emergency measures must be taken immediately.

[0279] S470: Construct a risk model that automatically corresponds between the risk assessment results calculated based on the evaluation function and the risk level, input the key parameters simulated in step S300 into the risk model, and the risk model obtains the final risk score through the evaluation function calculation, and automatically matches the corresponding risk level according to the risk score.

[0280] S480: The risk model combines historical data, expert opinions and real-time monitoring data to dynamically adjust the weight vector of each criterion, thereby providing a more accurate risk level assessment result; and through multiple simulations and the superposition of historical data, the risk model will continuously self-calibrate and optimize, thereby improving the accuracy and reliability of the risk level assessment results.

[0281] S500: The specific process of building a CNN model is as follows:

[0282] S510: Obtain the data generated by multiple numerical simulations in step S300 to change the key parameters of the overhanging rock mass and the slope. The data is simulated multiple times to obtain combinations under different scenarios. The data is then standardized and a multi-dimensional disaster data set including a training set, a validation set, and a test set is constructed based on the data.

[0283] S520: Establish a CNN model including an input layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer; wherein the input layer inputs the standardized key parameters. The standardized formula is:

[0284] ,

[0285] Where: X is the original input data, X ' is the standardized data, m and s are the mean and standard deviation of the original data.

[0286] The convolution layer uses multiple convolution kernels (filters) to slide on the input data, automatically extracting local nonlinear features from the input parameters. The convolution operation formula is:

[0287] ,

[0288] Where: Z i,j For the i The input data features are relative to the j Convolutional layer outputs with different features; W k is the convolution kernel, X i+k-1,j+k-1 is the value of the i+k-1,j+k-1 position of the input data; k is the convolution kernel k Weight index, corresponding to the weight parameter of the convolution kernel W k ; n is the size of the convolution kernel, that is, the total number of weights contained in the convolution kernel; i and j Represents the position index of the input data, representing the rows and columns of the input data matrix, or the rows and columns of the feature map; b is the bias, f ( x ) is the activation function, f ( x )=max(0, x ), .

[0289] The pooling layer is used to reduce the dimension of the convolutional layer output, retain important features and reduce computational complexity. Commonly used maximum pooling operations are:

[0290] .

[0291] The feature map output by the pooling layer is flattened into a one-dimensional vector and passed to the fully connected layer. The fully connected layer associates the features extracted by the convolution layer with the output risk level through the connection of neurons. The calculation formula is:

[0292] ,

[0293] Where: Z It is the output matrix or tensor, representing the activation value passed from the previous layer; FC That is, the fully connected layer is used to map the features extracted by the convolutional layer to the final output through linear combination; Z (FC) is the output of the fully connected layer, W (FC) is the weight matrix, b (FC) is the bias, f ( x ) is the activation function.

[0294] The output layer performs multi-classification prediction through the Softmax function, mapping the output of the fully connected layer to the probability of each risk level:

[0295] ,

[0296] Where: To predict the risk level j Probability, K is the number of risk levels, j Represents the position index of the column index in the output matrix. In the convolution operation, j Indexes each column in the output feature map, or the features to be calculated in each convolution window; X It is the original input data, including maximum impact force, accumulation volume, movement distance, impulse change, and lateral diffusion width.

[0297] S530: Use the training set in the aforementioned multidimensional disaster dataset to train the CNN model, then use the validation set and test set in the multidimensional disaster dataset to respectively validate and test the CNN model, and calculate the test accuracy. If the test accuracy of the CNN model is less than 0.95, readjust the parameters of the CNN model and perform training, validation and testing again. Otherwise, use this CNN model as the optimal CNN model.

[0298] S600: The specific process of grading hazardous areas is as follows:

[0299] S610: Through sensors and monitoring systems, key parameters of overhanging rock after instability are obtained in real time, and the key parameters are input into the optimal CNN model so that the model can quickly evaluate the current risks.

[0300] S620: The optimal CNN model identifies the complex relationship between parameters based on the input key data, and then quickly determines and outputs the risk level (low, medium, high or extremely high) corresponding to the current area.

[0301] S630: The risk level results output by the optimal CNN model are visualized in a three-dimensional topographic map through the GIS platform (the risk levels of different areas are displayed by color coding: green for low risk, yellow for medium risk, orange for high risk and red for extremely high risk, so as to help decision makers intuitively understand the risk levels of different areas through visualization), and the GIS platform dynamically adjusts the risk level of each area in the visualization according to the changes in monitoring data, thereby reflecting the changes in disaster risk over time and helping decision makers grasp the latest risk situation.

[0302] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for intelligently dividing dangerous areas after overhanging rock instability, characterized in that: It includes the acquisition of digital elevation model, geotechnical mechanics analysis, simulation of overhanging rock instability, AHP evaluation system construction, CNN model construction, and dangerous area classification steps. The specific contents are as follows: A. Acquisition of digital elevation model: Perform high-resolution 3D scanning of the target area of ​​the overhanging rock to obtain point cloud data; then clean the point cloud data and process it at multiple resolutions to generate a high-precision digital elevation model; then perform precision correction on the digital elevation model based on the actual terrain conditions on site; B. Geotechnical mechanics analysis: Sampling the target area and conducting geotechnical mechanics tests on the geotechnical samples to obtain mechanical parameters; statistically analyzing the mechanical parameters to determine their distribution form, calculating the mean value, standard deviation and dispersion coefficient of each parameter, and calibrating the distribution range of the mechanical parameters through multiple iterations of the aforementioned geotechnical mechanics tests; C. Simulation of overhang rock instability: Import the corrected digital elevation model into the PFC software to generate a three-dimensional slope model, assign the above-mentioned calibrated mechanical parameter range to different areas of the three-dimensional slope model, and perform zoning processing according to the layered structure of the slope; adjust the microscopic parameters of the three-dimensional slope model through multiple iterations of mechanical parameters, and use the trial and error method to make the macroscopic response of the three-dimensional slope model simulation results consistent with the results of geotechnical mechanics tests; then set boundary conditions in the three-dimensional slope model, and simulate the falling process by freely releasing the overhang rock, and use virtual sensors to monitor and record key parameters including movement distance, maximum impact force, accumulation volume, impulse change, and lateral diffusion width in real time; analyze the changes of key parameters under different conditions through multiple simulated falls, and determine the correlation between the changes of key parameters and risk levels; D. AHP construction evaluation system: A hierarchical model including target layer, criterion layer and scheme layer is established, in which the target layer is used to evaluate the risk level of the disaster-affected area, the criterion layer is composed of key parameters affecting the instability of overhanging rocks as criteria, and the scheme layer is used to output the risk level of the disaster-affected area; Construct a judgment matrix based on the pairwise comparison of key parameters in the criterion layer, then calculate the weight vector by normalizing the judgment matrix, and check the consistency of the judgment matrix to ensure the credibility of the weight calculation results; construct an evaluation function to assess the risk level of the disaster-affected area; Based on historical data and combined with expert experience, set risk level classification standards; Then, the key parameters obtained by simulation in step C are used to calculate the final risk score through the evaluation function, and the corresponding risk level mentioned above is matched according to the risk score; E. Construct CNN model: Construct a multi-dimensional disaster data set based on the key parameter data obtained by multiple simulations in step C, and divide the data set into training set, validation set and test set; then establish a CNN model including input layer, convolution layer, pooling layer, fully connected layer and output layer, and use the training set, validation set and test set to train, validate and test the CNN model respectively to obtain the optimal CNN model; F. Classification of dangerous areas: The CNN model receives the key parameters of overhanging rock after instability acquired in real time, and quickly calculates and outputs the risk level corresponding to the dangerous area.

2. According to claim 1, the intelligent classification method for dangerous areas after overhang rock instability is characterized by: The specific process of step A is as follows: A1. Obtaining point cloud data of the target area: Using an aircraft equipped with a high-precision positioning system and a laser radar, perform multiple high-resolution three-dimensional scans of the target area of ​​the overhanging rock at multiple angles to obtain point cloud data of the target area; A2. DEM construction: ArcGIS was used to clean the aforementioned point cloud data to remove irrelevant factors, and then the point cloud data was processed with multi-layer resolution using an automatic boundary detection algorithm to extract the features of the overhanging rock and its surrounding terrain in layers to generate a high-precision digital elevation model; A3. DEM correction: Combine existing topographic maps or field measurement data to perform multi-level correction on the digital elevation model, and use local high-resolution data processing technology to optimize the digital elevation model to obtain a digital elevation model.

3. The intelligent classification method for dangerous areas after overhang rock instability according to claim 1 is characterized in that: The specific process of step B is as follows: B1. Geotechnical sampling: Geotechnical samples are obtained by stratified random sampling of the overhanging rock mass and its surrounding slope areas; B2. Mechanical testing: Conducting geotechnical mechanics testing on the collected rock and soil samples in the laboratory to obtain the mechanical parameters of the rock and soil samples including compressive strength, elastic modulus, Poisson's ratio, shear strength and internal friction angle; B3. Parameter analysis: Statistical analysis is performed on each mechanical parameter to determine its distribution form, including normal distribution and log-normal distribution. Then, the mean value, standard deviation and coefficient of dispersion of each mechanical parameter are calculated. Subsequently, the range of values ​​of each mechanical parameter is further analyzed using the Monte Carlo simulation method, and the uncertainty model of each mechanical parameter is derived. Combined with the laboratory mechanical test results, a complete parameter distribution model is constructed. B4. Parameter calibration: Combine Bayesian reasoning to perform uncertainty analysis on mechanical parameters, then dynamically update the distribution range of mechanical parameters by combining historical data with experimental results, and calibrate the distribution range of mechanical parameters based on the results of multiple iterative mechanical tests.

4. The intelligent classification method for dangerous areas after overhang rock instability according to claim 1 is characterized in that: The specific process of step C is as follows: C1. PFC modeling: The digital elevation model is imported into the PFC software to generate a three-dimensional slope model. According to the distribution range of mechanical parameters of the calibrated rock and soil samples, the three-dimensional slope model is divided into multiple regions. Then, the Monte Carlo random sampling method is used to extract sample values ​​that conform to the statistical distribution from the parameter distribution. Then, each region is assigned a randomly generated parameter value. The DEM model and the geological survey report are used to determine the hierarchical structure and joint distribution of the rock mass. Then, combined with the PFC particle generation algorithm, different partitions are generated according to different depths and lithology of the slope. Independent particle diameter, bonding strength and mechanical parameters are defined for each partition. Finally, a joint network simulation is applied to the joint development zone, and different contact stiffness and normal bonding force are assigned. C2. Parameter adjustment: The three-dimensional slope model is iterated multiple times using the mechanical parameters calibrated in step B by trial and error method, so as to adjust the microscopic parameters of the three-dimensional slope model including contact stiffness and bonding strength to reflect the actual field conditions, so that the difference between the macroscopic response of the stress-strain curve and the laboratory mechanical test results in the simulation results of the three-dimensional slope model is less than 5%, and the three-dimensional slope model containing the most suitable microscopic parameters is obtained; C3. Simulation of instability process: Boundary conditions are set in the three-dimensional slope model obtained in C2, and then the overhanging rock is simulated to fall under the action of gravity through free release. Virtual sensors are used to monitor and record key parameters including movement distance, maximum impact force, accumulation volume, impulse change, and lateral diffusion width in real time.

5. The intelligent classification method for dangerous areas after overhanging rock instability according to claim 4 is characterized in that: If the difference between the simulation results including the macroscopic response of the stress-strain curve and the results of the laboratory mechanical test in C2 is not less than 5%, the mesoscopic parameters of the three-dimensional slope model are iteratively adjusted using the trial and error method.

6. The intelligent classification method for dangerous areas after overhang rock instability according to any one of claims 1 to 5 is characterized in that: The specific process of step D is as follows: D1. Construction of hierarchical model: A hierarchical model including target layer, criterion layer and scheme layer is established for evaluating the instability disaster of overhanging rock; the target layer is used to evaluate the risk level of the disaster-affected area, the criterion layer is composed of key parameters affecting the instability of overhanging rock as criteria, and the scheme layer is used to output the risk level of the disaster-affected area; D2. Construction of judgment matrix: Construct a judgment matrix for pairwise comparison of the key parameters in the criterion layer, and determine the importance of each parameter relative to other parameters based on expert opinions or historical data; D3. Calculation of weight vector: The weight vector calculated by normalizing the judgment matrix is ​​to first calculate the normalized value of each column in the judgment matrix, divide each element of the judgment matrix by the sum of its column, and obtain a normalized matrix; then calculate the average value of each row in the judgment matrix to obtain the weight vector of each criterion; the calculation formula is: ， in: W i For the i The weight vector of the criteria, a ij For the i The factor relative to j The importance ratio of the factors, n is the order of the matrix; D4. Consistency test: calculate the consistency index first CI : ， Where: λ max is the maximum eigenvalue of the judgment matrix; Then calculate the consistency ratio CR : ， Where: RI It is a random consistency index obtained by looking up the table according to the order of the judgment matrix; If CR<0.1, the consistency of the judgment matrix is ​​considered acceptable and the calculation results of the weight vector are credible; otherwise, the value of the judgment matrix needs to be readjusted; D5. Evaluation function construction: Construct an evaluation function for evaluating the risk level of the disaster-affected area. The evaluation function multiplies the weight of each criterion with the actual monitored parameter value to obtain the risk assessment result; the evaluation function is: ， Where: S The final risk assessment result is W i For the i The weight of the parameters, X i The actual monitored i parameter values; D6. Risk level classification: Based on historical data and combined with expert experience, a four-level risk classification standard is set, including low risk, medium risk, high risk and extremely high risk; D7. Risk model construction: Construct a risk model that automatically corresponds to the risk assessment results obtained by calculating the evaluation function and the risk level. Input the key parameters simulated in step C into the risk model to obtain the corresponding risk level.

7. The intelligent classification method for dangerous areas after overhanging rock instability according to claim 6 is characterized in that: It also includes dynamic updating of risk models: the risk model combines historical data, expert opinions and real-time monitoring data to dynamically adjust the weight vector of each criterion.

8. The intelligent classification method for dangerous areas after overhanging rock instability according to claim 6 is characterized by: The specific process of step E is as follows: E1. Dataset construction: Obtain the data generated by multiple numerical simulations in step C to change the key parameters of the overhanging rock mass and slope, standardize the data, and then construct a multi-dimensional disaster dataset including training set, validation set and test set; E2. CNN model construction: Establish a CNN model including input layer, convolution layer, pooling layer, fully connected layer and output layer; among them, the key parameters of the input layer after input standardization; the standardization formula is: ， Where: X is the original input data, X ' is the standardized data, μ and σ is the mean and standard deviation of the original data; The convolution layer uses multiple convolution kernels to slide on the input data and automatically extract local nonlinear features from the input parameters; the convolution operation formula is: ， Where: Z i,j For the i The input data features are relative to the j Convolutional layer outputs with different features; W k is the convolution kernel, X i+k-1,j+k-1 is the value of the i+k-1,j+k-1 position of the input data; k is the convolution kernel k Weight index, corresponding to the weight parameter of the convolution kernel W k ; n is the size of the convolution kernel, that is, the total number of weights contained in the convolution kernel; i and j Represents the position index of the input data, representing the rows and columns of the input data matrix, or the rows and columns of the feature map; b is the bias, f ( x ) is the activation function, f ( x )=max(0, x ), ; The pooling layer is used to reduce the dimension of the convolutional layer output, retain important features and reduce computational complexity; the commonly used maximum pooling operation is: ； The fully connected layer associates the features extracted by the convolutional layer with the output risk level through the connection of neurons. The calculation formula is: ， Where: Z It is the output matrix or tensor, representing the activation value passed from the previous layer; FC That is, the fully connected layer is used to map the features extracted by the convolutional layer to the final output through linear combination; Z (FC) is the output of the fully connected layer, W (FC) is the weight matrix, b (FC) is the bias, f ( x ) is the activation function; The output layer performs multi-classification prediction through the Softmax function, mapping the output of the fully connected layer to the probability of each risk level: ， Where: To predict the risk level j The probability of K is the number of risk levels, j Represents the position index of the column index in the output matrix. In the convolution operation, j Indexes each column in the output feature map, or the features to be calculated in each convolution window; X It is the original input data, including maximum impact force, accumulation volume, movement distance, impulse change, and lateral diffusion width; E3. CNN model training: Use the training set in the multidimensional disaster dataset to train the CNN model, and then use the validation set and test set in the multidimensional disaster dataset to verify and test the CNN model respectively, and calculate the test accuracy. If the test accuracy of the CNN model is less than 0.95, readjust the parameters of the CNN model and train, verify and test again. Otherwise, take this CNN model as the optimal CNN model.

9. The intelligent classification method for dangerous areas after overhang rock instability according to claim 6 is characterized in that: The specific process of step F is as follows: F1. Real-time data input: Through sensors and monitoring systems, key parameters of overhanging rock after instability are obtained in real time and input into the optimal state CNN model; F2. Risk level judgment: The optimal CNN model recognizes the complex relationship between parameters based on the input key data, and quickly judges and outputs the risk level corresponding to the current area.

10. The intelligent classification method for dangerous areas after overhang rock instability according to claim 9 is characterized in that: The F step also includes: the risk level results output by the optimal CNN model are visualized in a three-dimensional topographic map through a GIS platform, and the GIS platform dynamically adjusts the risk level of each area in the visualization according to changes in monitoring data.

Images