Slope stability classification method for rock engineering

Abstract

The invention relates to the technical field of geological disaster prevention and control, in particular to a rock engineering slope stability classification method. According to the technical scheme, the method comprises the following steps: obtaining spatial distribution characteristics of a slope structural plane through combined detection of three-dimensional laser scanning and a geological radar, and establishing a quantitative index system containing occurrence, density and connectivity of the structural plane; the method comprises the following steps: acquiring a fracture signal in a slope in real time by adopting a micro-seismic monitoring system, and extracting a micro-seismic event energy release rate, a dominant frequency offset and a seismic source mechanism parameter through time-frequency analysis; and constructing a geologic structure evolution inversion model, reconstructing a geologic structure evolution process based on the regional geologic database and the field drilling data, and calculating the distribution characteristics of the residual stress field of the structure. According to the method, comprehensive, accurate and real-time monitoring and classified evaluation of the rock engineering slope stability are realized, potential risks can be effectively predicted, a scientific and accurate decision basis is provided for slope support and treatment, and the safety and stability of the rock engineering slope are greatly guaranteed.

Description

Technical Field

[0001] The invention relates to the technical field of geological disaster prevention and control, and in particular to a rock engineering slope stability classification method. Background Art

[0002] The stability classification of rock engineering slopes is crucial in all types of geotechnical engineering construction and is related to the safety and stability of the project. The traditional slope stability classification method has many limitations:

[0003] In terms of obtaining slope structural surface information, the early days relied on manual measurement, which was not only inefficient and labor-intensive, but also affected by the terrain and subjective factors of the surveyors. It was difficult to accurately obtain the spatial distribution characteristics of the structural surface, resulting in inaccurate determination of quantitative indicators such as the structural surface attitude, density and connectivity, and unable to provide a reliable basis for slope stability analysis.

[0004] In terms of microseismic monitoring, previous technologies found it difficult to collect real-time and comprehensive internal rupture signals of the slope, and the signal processing methods were relatively simple, making it impossible to effectively extract key dynamic characteristics such as the energy release rate of microseismic events, main frequency offset, and focal mechanism parameters. This resulted in an insufficient understanding of the expansion of rock ruptures within the slope, making it difficult to issue timely warnings when potential dangers appeared on the slope.

[0005] In the past, there was a lack of effective inversion models in the study of geological structural evolution, which made it impossible to fully utilize regional geological databases and field drilling data to reconstruct the geological structural evolution process, and thus it was impossible to accurately calculate the distribution characteristics of the structural residual stress field, ignoring the important impact of geological structural history on slope stability.

[0006] Existing slope stability classification models are mostly single factor or simple combination factor analysis, which makes it difficult to comprehensively consider multiple factors such as static structural parameters, dynamic microseismic characteristics, and tectonic stress field parameters. Moreover, the models often lack adaptive capabilities and cannot dynamically adjust the classification results according to real-time monitoring data. When faced with complex and changeable slope engineering environments, it is difficult to meet the actual needs of the project and cannot provide accurate and effective decision support for slope support and management. Therefore, this application proposes a rock engineering slope stability classification method. Summary of the invention

[0007] The purpose of the present invention is to propose a rock engineering slope stability classification method to address the problem that the existing slope stability classification models in the background technology are mostly single factor or simple combination factor analysis, which is difficult to comprehensively consider multiple factors such as static structural parameters, dynamic microseismic characteristics and tectonic stress field parameters.

[0008] The technical solution of the present invention is a rock engineering slope stability classification method, comprising the following steps:

[0009] S1. The spatial distribution characteristics of the slope structural surface are obtained through the joint detection of 3D laser scanning and geological radar, and a quantitative index system including the occurrence, density and connectivity of the structural surface is established;

[0010] S2. Use the microseismic monitoring system to collect the internal rupture signal of the slope in real time, and extract the energy release rate, main frequency offset and focal mechanism parameters of the microseismic event through time-frequency analysis;

[0011] S3. Construct a geological structural evolution inversion model, reconstruct the geological structural evolution process based on the regional geological database and field drilling data, and calculate the distribution characteristics of the structural residual stress field;

[0012] S4, establishing a multi-objective classification model integrating machine learning, taking the static structural parameters, dynamic microseismic characteristics and tectonic stress field parameters obtained in steps S1-S3 as input variables, and ranking the feature importance by using an improved random forest algorithm;

[0013] S5. Design a dynamic feedback correction mechanism. When the cumulative energy of the microseismic events monitored in real time exceeds the preset threshold, it automatically triggers the iterative update of the classification model parameters and simultaneously corrects the boundary conditions for determining the slope stability level.

[0014] S6. Generate three-dimensional visual classification results, and integrate the stability grade zoning, critical sliding surface prediction results and support optimization suggestions on the digital twin platform.

[0015] Optionally, the three-dimensional laser scanning in step S1 adopts multi-site cloud registration technology, and the automatic calibration of the scanning coordinate system is realized by introducing an inertial navigation unit. The structural surface recognition algorithm integrates point cloud curvature analysis and regional growing method to automatically extract the structural surface spacing d, trace length L and roughness JRC parameters.

[0016] Optionally, the microseismic signal processing in step S2 adopts a wavelet packet-Hilbert transform joint analysis method to establish a multi-dimensional feature matrix of energy-frequency-time, wherein:

[0017] The main frequency offset Δf is calculated by the instantaneous frequency estimation method, which reflects the rock mass fracture propagation speed;

[0018] The focal mechanism parameters include the anisotropy index AI and the rupture surface normal vector obtained by moment tensor decomposition;

[0019] The energy release rate is calculated using an improved S-transform time-frequency energy density integration method to eliminate the high-frequency aliasing error of the traditional method.

[0020] Optionally, the geological structure evolution inversion model in step S3 is constructed by using a finite element-discrete element coupling algorithm through the following steps:

[0021] Establish an initial tectonic stress field based on the regional geological map;

[0022] Invert the tectonic movement period using the statistics of drill core joints;

[0023] The reverse time marching method is used to simulate the tectonic movement process and iteratively calculate the residual stress distribution;

[0024] The parameters of rock damage constitutive equation are calibrated through acoustic emission experiments.

[0025] Optionally, the improved random forest algorithm in step S4 includes:

[0026] Introducing a dynamic feature weight mechanism to adjust the split node selection strategy based on the mutual information between features;

[0027] Design a multi-objective optimization function to minimize classification error and feature dimension;

[0028] Use Bayesian optimization algorithm to automatically adjust the decision tree depth and the minimum number of samples for leaf nodes;

[0029] The SHAP value interpretation module is integrated to output quantitative indicators of the contribution of each characteristic parameter to the stability classification.

[0030] Optionally, the dynamic feedback correction mechanism in step S5 includes:

[0031] Set dual threshold trigger conditions: the cumulative microseismic energy E on the day d ≥E threshold1 And the main frequency offset Δf ≥ Δf threshold When the first level correction is initiated;

[0032] When E is met for 3 consecutive days d ≥0.7E threshold1 The second level correction is initiated when

[0033] The correction process uses an online learning algorithm to update the model parameters while retaining the memory factor of historical data;

[0034] The prediction performance of the updated model was evaluated by K-fold time series cross validation.

[0035] Optionally, the digital twin platform in step S6 integrates the following modules:

[0036] Real-time monitoring data cockpit, showing the heat map of the spatial distribution of microseismic events;

[0037] Stability level 3D shading rendering module, using HSV color space to map stability coefficients;

[0038] Critical sliding surface probability cloud map generation module, which calculates the probability of sliding surface occurrence based on Monte Carlo simulation;

[0039] The support structure optimization suggestion module automatically matches the support parameter database according to the stability level.

[0040] Optionally, the structural surface recognition algorithm further includes the following steps:

[0041] Construct a point cloud anomaly filtering network based on deep learning, and use a three-dimensional convolutional neural network to identify and remove interference point clouds caused by vegetation coverage and temporary support structures;

[0042] Design a multi-scale structural surface matching algorithm, quantify the structural surface morphological similarity through Fourier descriptors, and automatically associate the matching structural surfaces of adjacent scanning sites;

[0043] The robust estimation theory is introduced to correct the structural surface attitude parameters and the following compensation model is established:

[0044]

[0045] Where Δθ is the correction value of the occurrence angle, k is the rock weathering coefficient, JRC i is the roughness coefficient of each sub-region, d is the distance between the structural surfaces;

[0046] The water seepage status of the structural surface is inverted by the laser echo intensity, and the infrared thermal imaging auxiliary verification module is triggered when the echo intensity attenuation rate is greater than 15%.

[0047] Optionally, the microseismic signal processing further includes:

[0048] Deploy an adaptive ambient noise base library and separate construction vibration and rock fracture signals through matching pursuit algorithm;

[0049] A collaborative verification mechanism for array sensors is designed. When a single sensor detects a microseismic event, three adjacent sensors are activated to form a temporary detection array. A valid event is confirmed only when the following conditions are met:

[0050]

[0051] Among them, C coh is the waveform coherence coefficient between sensors;

[0052] Establish a temperature-wave velocity compensation model to correct the earthquake source location coordinates in real time:

[0053] v(T)=v 0 ·[1-α(TT 0 )] 1 / 2

[0054] Where v(T) is the wave velocity at the current temperature, α is the thermal expansion coefficient of the rock mass, and T 0To calibrate the temperature; a transfer learning strategy is adopted, and the time-frequency characteristics of historical microseismic events are used as a pre-training data set to improve the feature recognition accuracy under small sample conditions.

[0055] Optionally, the dynamic feedback correction mechanism further includes a model robustness enhancement module:

[0056] Design a feature drift detector to monitor changes in input data distribution by calculating KL divergence, and activate the model protection mechanism when the KL value is >2.0;

[0057] Establish an adversarial sample generation network, inject labeled noise data before model updating, and test the model's anti-interference ability;

[0058] Adopt the Elastic Weight Consolidation (EWC) algorithm to retain the important weight history memory when updating parameters to prevent catastrophic forgetting;

[0059] Set up a model rollback mechanism. When the validation set accuracy drops by more than 5% for 5 consecutive iterations, it will automatically revert to the previous stable version.

[0060] Compared with the prior art, the present invention has at least one of the following beneficial technical effects:

[0061] The combined detection of 3D laser scanning and geological radar, combined with multi-site cloud registration and advanced structural surface recognition algorithms, can obtain the spatial distribution characteristics of structural surfaces with high precision, and extract parameters such as occurrence, density, and connectivity more accurately, laying a solid data foundation for stability analysis. The microseismic monitoring system cooperates with the wavelet packet-Hilbert transform joint analysis method to collect and process fracture signals in real time, obtain key dynamic characteristics such as energy release rate and main frequency offset, accurately reflect changes inside the rock mass, and promptly discover potential hidden dangers.

[0062] The geological tectonic evolution inversion model reconstructs the geological history based on regional geological data and drilling information, accurately calculates the tectonic residual stress field, and fully considers the impact of geological history on slope stability, making the analysis more scientific and comprehensive.

[0063] By integrating the multi-objective classification model of machine learning and using the improved random forest algorithm to process multi-source data, the feature importance is sorted and the classification is more accurate. The dynamic feedback correction mechanism automatically updates the model parameters and determines the boundary conditions based on real-time data such as microseismic energy. It has strong adaptive capabilities and can accurately reflect changes in slope stability.

[0064] The three-dimensional visualization classification results are integrated and displayed on the digital twin platform, and the stability grade zoning, critical sliding surface prediction results and support optimization suggestions are intuitively presented, which makes it convenient for engineering personnel to quickly understand the slope conditions, make scientific decisions and ensure project safety.

[0065] The present invention realizes comprehensive, accurate and real-time monitoring and classified evaluation of the stability of rock engineering slopes, can effectively predict potential risks, and provide a scientific and accurate decision-making basis for slope support and management, thus greatly ensuring the safety and stability of rock engineering slopes. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 This is a structural schematic diagram of a rock engineering slope stability classification method. DETAILED DESCRIPTION

[0067] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0068] Example 1

[0069] like Figure 1 As shown, the present invention proposes a rock engineering slope stability classification method, and the steps of the method are described in detail below.

[0070] (I): Obtaining the spatial distribution characteristics of the slope structure surface

[0071] The slope was scanned using a 3D laser scanner, and multi-site cloud registration technology was used to set up scanning sites at different locations to ensure a comprehensive scan of the entire slope. An inertial navigation unit was introduced to achieve automatic calibration of the scanning coordinate system, improving the accuracy and consistency of the scan data.

[0072] Geological radar is used to detect the interior of the slope and complement it with three-dimensional laser scanning data to obtain more comprehensive structural surface information.

[0073] A structural surface recognition algorithm that integrates point cloud curvature analysis and region growing method is adopted: first, point cloud curvature analysis is performed on the scanned point cloud data, and the area where the structural surface may exist is determined by calculating the curvature of each point on the point cloud surface; then, the region growing method is used, with points with obvious curvature changes as seed points, and the structural surface area is gradually expanded according to certain growth criteria, thereby automatically extracting the structural surface spacing d, trace length L and roughness JRC parameters.

[0074] We can further construct a point cloud anomaly filtering network based on deep learning, use a three-dimensional convolutional neural network to identify and remove interference point clouds caused by vegetation coverage and temporary support structures, and improve data quality; design a multi-scale structural surface matching algorithm, quantify the morphological similarity of the structural surface through Fourier descriptors, and automatically associate the matching structural surfaces of adjacent scanning sites; introduce the robust estimation theory to correct the structural surface occurrence parameters, and use the formula

[0075]

[0076] Calculate the occurrence angle correction, where Δθ is the occurrence angle correction, k is the rock weathering coefficient, JRC i is the roughness coefficient of each sub-region, and d is the distance between the structural surfaces. The water seepage status of the structural surface is inverted by the laser echo intensity. When the echo intensity attenuation rate is greater than 15%, the infrared thermal imaging auxiliary verification module is triggered to further obtain the water seepage status of the structural surface.

[0077] The problem of accumulated deviations in structural surface parameter extraction under complex surface interference has been solved. The parameter distortion caused by vegetation / artificial object occlusion has been solved through deep learning point cloud filtering and multi-scale matching. The robust estimation model has broken through the error-sensitive defects of the traditional least squares method. Laser-infrared fusion detection has achieved accurate inversion of the water seepage state of the structural surface. The measurement error of the structural surface attitude has been reduced from ±8° to ±2°, thereby improving the accuracy of water seepage state identification.

[0078] (II): Collecting and processing internal slope rupture signals

[0079] Deploy a microseismic monitoring system and rationally arrange multiple sensors inside and on the surface of the slope to ensure that the internal rupture signals of the slope can be fully collected.

[0080] The wavelet packet-Hilbert transform joint analysis method is used to process the microseismic signals: first, the microseismic signals are decomposed into different frequency bands using wavelet packet decomposition, and then each frequency band signal is subjected to Hilbert transform to establish a multi-dimensional feature matrix of energy-frequency-time.

[0081] The main frequency offset Δf is calculated by the instantaneous frequency estimation method according to the instantaneous frequency change of the microseismic signal. This parameter reflects the rock mass fracture propagation speed.

[0082] Calculate the focal mechanism parameters, including the anisotropy index AI and the rupture surface normal vector obtained by moment tensor decomposition.

[0083] The improved S-transform time-frequency energy density integration method is used to calculate the energy release rate, eliminating the high-frequency aliasing error of the traditional method and obtaining the energy release of microseismic events more accurately.

[0084] An adaptive environmental noise base library can be further deployed to separate construction vibration and rock fracture signals through matching pursuit algorithm to improve signal purity; an array sensor collaborative verification mechanism is designed. When a single sensor detects a microseismic event, three adjacent sensors are activated to form a temporary detection array.

[0085]

[0086] The valid event is confirmed only when C coh is the waveform coherence coefficient between sensors, and the temperature-wave velocity compensation model v(T)=v 0 ·[1-α(TT 0 )] 1 / 2 , where v(T) is the wave velocity at the current temperature, α is the thermal expansion coefficient of the rock mass, and T 0 In order to calibrate the temperature and correct the source location coordinates in real time, a transfer learning strategy is adopted, and the time-frequency characteristics of historical microseismic events are used as a pre-training data set to improve the feature recognition accuracy under small sample conditions.

[0087] (III) Construction of an inversion model for geological structural evolution

[0088] Based on the regional geological map, the regional geological structure characteristics are analyzed and the initial tectonic stress field is established to provide initial conditions for subsequent simulations.

[0089] Collect statistical results of on-site drill core joints, analyze the distribution, occurrence and other characteristics of the joints, invert the tectonic movement periods, and determine the geological tectonic movement history experienced by the slope.

[0090] The reverse time marching method is used to simulate the tectonic movement process. According to the sequence of geological tectonic movement, the simulation is carried out step by step from the present to the past, and the residual stress distribution is iteratively calculated. The mechanical response of the rock in the tectonic movement is considered to accurately obtain the distribution characteristics of the tectonic residual stress field.

[0091] Through acoustic emission experiments, the failure process of rock under different stress conditions is simulated, the parameters of the rock damage constitutive equation are calibrated, more accurate rock mechanics parameters are provided for the model, and the simulation accuracy of the model is improved.

[0092] (IV): Establish a multi-target classification model integrating machine learning

[0093] The static structural parameters, dynamic microseismic characteristics and tectonic stress field parameters obtained above are used as input variables to construct the input data set.

[0094] An improved random forest algorithm is used: a dynamic feature weight mechanism is introduced, and the split node selection strategy is adjusted according to the mutual information between features, so that the algorithm can more reasonably select features for splitting and improve classification efficiency; a multi-objective optimization function is designed to minimize the classification error and feature dimension, while ensuring classification accuracy and reducing the complexity of the model; a Bayesian optimization algorithm is used to automatically adjust the decision tree depth and the minimum number of leaf nodes, and optimize the structure and parameters of the model; a SHAP value interpretation module is integrated to output a quantitative indicator of the contribution of each feature parameter to the stability classification, which facilitates understanding of the decision-making process of the model and the importance of each parameter.

[0095] (V) Design a dynamic feedback correction mechanism

[0096] Set dual threshold trigger conditions: the cumulative microseismic energy E on the day d ≥E threshold1 And the main frequency offset Δf ≥ Δf threshold When the first level correction is initiated;

[0097] When E is met for 3 consecutive days d ≥0.7E threshold1 The second level correction is initiated.

[0098] The correction process uses an online learning algorithm to update the model parameters and retains the memory factor of historical data during the update process, so that the model can make full use of new data without losing important information in the historical data.

[0099] The prediction performance of the updated model was evaluated through K-fold time series cross validation to ensure that the model still had good prediction accuracy and stability after the update.

[0100] The model robustness enhancement module can be further added: design a feature drift detector to monitor the changes in the input data distribution by calculating the KL divergence, and start the model protection mechanism when the KL value is >2.0; establish an adversarial sample generation network, inject labeled noise data before the model is updated, and test the model's anti-interference ability; use the elastic weight consolidation algorithm (EWC) to retain important weight history memory when updating parameters to prevent catastrophic forgetting; set up a model rollback mechanism, and automatically restore to the previous stable version when the verification set accuracy drops by more than 5% for 5 consecutive iterations.

[0101] (VI): Generate 3D visualization classification results

[0102] Integrate relevant modules on the digital twin platform: establish a real-time monitoring data cockpit, display the spatial distribution heat map of microseismic events in real time, and intuitively show the occurrence location and intensity distribution of microseismic events.

[0103] The stability grade 3D shading rendering module is used to map the stability coefficient using the HSV color space to perform a 3D visualization of the slope stability grade. Different colors represent different stability grades, which facilitates intuitive judgment of the slope stability.

[0104] Through the critical sliding surface probability cloud map generation module, the probability of sliding surface occurrence is calculated based on Monte Carlo simulation, and the possible location of the critical sliding surface is predicted, providing an important reference for slope support.

[0105] With the help of the support structure optimization suggestion module, the support parameter database is automatically matched according to the stability level, and optimization suggestions are provided for slope support, guiding engineering personnel to take reasonable support measures to ensure the stability of the slope.

[0106] It is worth noting that in this embodiment, the three-dimensional laser scanning and geological radar joint detection are combined with multi-site cloud registration technology, inertial navigation unit calibration and advanced structural surface recognition algorithm to obtain the spatial distribution characteristics of the slope structural surface with high precision, accurately extract the structural surface occurrence, density, connectivity, spacing, trace length, and roughness parameters, and provide a reliable static data basis for stability analysis. Through the microseismic monitoring system combined with the wavelet packet-Hilbert transform joint analysis method, the internal rupture signal of the slope is collected and accurately processed in real time, and the key dynamic characteristics of the microseismic event energy release rate, main frequency offset and focal mechanism parameters are obtained, and the rock mass rupture extension is reflected in a timely manner, so as to realize the effective monitoring of the real-time state of the slope and facilitate the timely discovery of potential dangers.

[0107] Construct a geological structural evolution inversion model, reconstruct the geological structural evolution process based on the regional geological database and field drilling data, accurately calculate the distribution characteristics of the structural residual stress field, and fully consider the impact of geological history on slope stability, making the stability analysis more comprehensive and scientific. Integrate the multi-objective classification model of machine learning, and use the improved random forest algorithm to sort the feature importance of multi-source data to improve classification accuracy. At the same time, design a dynamic feedback correction mechanism to automatically update model parameters and determine boundary conditions based on real-time monitoring of microseismic event energy and other data, enhance the model's adaptive ability, and more accurately reflect changes in slope stability.

[0108] In addition, three-dimensional visualization classification results are generated, and the stability grade zoning, critical sliding surface prediction results and support optimization suggestions are integrated and displayed on the digital twin platform, which intuitively displays the slope stability status and provides engineering personnel with a clear decision-making basis, facilitating timely and effective support measures to ensure the safety of slope projects.

[0109] Example 2

[0110] On the basis of Example 1, in order to further improve the accuracy and practicality of the rock engineering slope stability classification method, the following progressive technical solutions are added:

[0111] Introduce multi-source satellite remote sensing data fusion analysis: Combine optical satellite images and radar satellite data, and use image fusion algorithms to obtain high-resolution, multi-spectral slope image data. Extract vegetation coverage change information from optical images. If the vegetation coverage decreases by more than 10% in a short period of time, it may indicate shallow sliding or loose rock and soil on the slope; use the interferometric measurement technology (InSAR) of radar satellite data to obtain micro-deformation information on the slope surface with an accuracy of up to millimeter level. Fuse this information with three-dimensional laser scanning and geological radar detection data to more comprehensively understand the deformation trend of the slope from the macro and micro levels, supplement the spatial distribution feature data of the structural surface, and provide a richer basis for stability analysis.

[0112] Optimize the layout algorithm of microseismic monitoring network: Optimize the layout of microseismic sensors based on genetic algorithm. With the objective function of maximizing the coverage of sensors on the internal fracture signals of the slope and minimizing signal interference, the optimal position of the sensors is determined by considering the topography and geomorphology of the slope, the characteristics of geological structure and the prediction results of the potential fracture area. Through this optimization algorithm, the effective acquisition rate of microseismic signals can be increased by more than 20%, the monitoring blind spots can be reduced, microseismic events can be captured more accurately, and the monitoring accuracy of parameters such as the energy release rate and main frequency offset of microseismic events can be improved.

[0113] Improve the boundary condition processing of the geological structure evolution inversion model: Based on the finite element-discrete element coupling algorithm, the dynamic boundary conditions of regional plate movement are considered. Collect historical data and real-time monitoring data of regional plate movement to establish a plate movement trend prediction model. The stress changes caused by plate movement are used as the dynamic boundary input of the geological structure evolution inversion model, so that the model can more realistically reflect the influence of external stress during the geological structure evolution process, thereby more accurately calculating the distribution characteristics of the structural residual stress field and improving the accuracy of the analysis of the relationship between geological structure history and slope stability.

[0114] Strengthen the generalization ability of the multi-objective classification model that integrates machine learning: adopt a method that combines transfer learning and meta-learning. Pre-train on historical data of different types of rock slopes to build a general slope stability classification model framework. When applied to new slope projects, use a small amount of sample data from new projects for fine-tuning. At the same time, based on the meta-learning algorithm, learn the optimal strategy and hyperparameter settings for model training from multiple existing slope stability classification tasks, quickly adapt to new engineering environments, improve the generalization ability of the model under different geological conditions and construction conditions, reduce the training time and data requirements of the model in new scenarios, and increase the classification accuracy of the model under small samples and complex geological conditions by more than 15%.

[0115] Improve the risk warning classification of the dynamic feedback correction mechanism: On the basis of the dual threshold trigger conditions, a risk warning classification system is constructed according to the cumulative energy of microseismic events, the main frequency offset, and the changes in the tectonic stress field. When the cumulative energy of microseismic events reaches 80%-100% of the preset threshold and the main frequency offset continues to increase, and at the same time, the tectonic stress field changes abnormally, an orange warning is issued to remind engineering personnel to strengthen monitoring and prepare to take emergency measures; when the preset threshold is reached or exceeded, a red warning is issued and emergency plans are immediately initiated, such as evacuation of personnel and cessation of related construction activities. By refining the risk warning classification, the timeliness and effectiveness of the response to changes in slope stability can be improved to ensure engineering safety.

[0116] Upgrade the interactive function of the digital twin platform: Add augmented reality (AR) and virtual reality (VR) interactive modules to the digital twin platform. Engineers can use AR devices to view the stability level zoning of the slope, critical sliding surface prediction results and other information in real time on site, and overlay them with the actual slope scene to facilitate on-site evaluation and decision-making; using VR technology, engineers can immerse themselves in the virtual slope environment, observe the slope structure from different angles, simulate the slope deformation under different working conditions, and more intuitively understand the slope stability status, providing more intuitive support for the formulation and implementation of support optimization proposals.

[0117] Through the above progressive technical solutions, Example 2 is improved on the basis of Example 1 in terms of multi-source data fusion, monitoring network optimization, model improvement, risk warning and interactive function upgrade, further improving the technical effect of the rock engineering slope stability classification method, enhancing the monitoring, analysis and prediction capabilities of slope stability, and providing more powerful support for the safety of rock engineering slopes.

[0118] Example 3

[0119] This embodiment discloses a rock engineering slope stability classification method, which is described in detail below.

[0120] (I): Obtaining the spatial distribution characteristics of the slope structure surface

[0121] The slope was scanned using a 3D laser scanner, and multi-site cloud registration technology was used to set up scanning sites at different locations to ensure a comprehensive scan of the entire slope. An inertial navigation unit was introduced to achieve automatic calibration of the scanning coordinate system, improving the accuracy and consistency of the scan data.

[0122] Geological radar is used to detect the interior of the slope and complement it with three-dimensional laser scanning data to obtain more comprehensive structural surface information.

[0123] A structural surface recognition algorithm that integrates point cloud curvature analysis and region growing method is adopted: first, point cloud curvature analysis is performed on the scanned point cloud data, and the area where the structural surface may exist is determined by calculating the curvature of each point on the point cloud surface; then, the region growing method is used, with points with obvious curvature changes as seed points, and the structural surface area is gradually expanded according to certain growth criteria, thereby automatically extracting the structural surface spacing d, trace length L and roughness JRC parameters.

[0124] We can further construct a point cloud anomaly filtering network based on deep learning, use a three-dimensional convolutional neural network to identify and remove interference point clouds caused by vegetation coverage and temporary support structures, and improve data quality; design a multi-scale structural surface matching algorithm, quantify the morphological similarity of the structural surface through Fourier descriptors, and automatically associate the matching structural surfaces of adjacent scanning sites; introduce the robust estimation theory to correct the structural surface occurrence parameters, and use the formula

[0125]

[0126] Calculate the occurrence angle correction, where Δθ is the occurrence angle correction, k is the rock weathering coefficient, JRC i is the roughness coefficient of each sub-region, and d is the distance between the structural surfaces. The water seepage status of the structural surface is inverted by the laser echo intensity. When the echo intensity attenuation rate is greater than 15%, the infrared thermal imaging auxiliary verification module is triggered to further obtain the water seepage status of the structural surface.

[0127] This embodiment also includes oblique photogrammetry-assisted positioning based on drones and structural surface filling material composition analysis, which will be described in detail below.

[0128] Among them, the oblique photogrammetry-assisted positioning based on drones uses drones equipped with high-definition cameras to perform oblique photogrammetry on the slopes. The drone obtains image data of the slope from multiple angles, processes it through photogrammetry software, and generates a high-precision three-dimensional model of the slope. This model can not only show the macroscopic topography of the slope, but also be compared and integrated with three-dimensional laser scanning data. For areas that are difficult to cover with laser scanning, such as steep cliffs, oblique photogrammetry data can be effectively supplemented to ensure the integrity of the structural surface information acquisition. Through the feature point matching algorithm, the structural surface features in the oblique photogrammetry model are matched with the structural surface features in the laser scanning point cloud data, further improving the accuracy of structural surface positioning, so that the measurement error of quantitative indicators such as structural surface occurrence, density and connectivity is reduced by 10%-15%.

[0129] In addition, during the structural surface identification process, spectral analysis technology is used to analyze the composition of the structural surface filling. By collecting structural surface filling samples on site, using a portable spectrometer to obtain its spectral characteristics, and comparing it with the known mineral spectrum database, the composition of the filling is determined. Fillers with different compositions have different effects on slope stability. For example, fillers with a high clay mineral content may lead to a decrease in the shear strength of the structural surface. Incorporating filler composition information into the quantitative index system as an important parameter for evaluating slope stability makes the stability classification results more in line with actual conditions.

[0130] (II): Collecting and processing internal slope rupture signals

[0131] Deploy a microseismic monitoring system and rationally arrange multiple sensors inside and on the surface of the slope to ensure that the internal rupture signals of the slope can be fully collected.

[0132] The wavelet packet-Hilbert transform joint analysis method is used to process the microseismic signals: first, the microseismic signals are decomposed into different frequency bands using wavelet packet decomposition, and then each frequency band signal is subjected to Hilbert transform to establish a multi-dimensional feature matrix of energy-frequency-time.

[0133] The main frequency offset Δf is calculated by the instantaneous frequency estimation method according to the instantaneous frequency change of the microseismic signal. This parameter reflects the rock mass fracture propagation speed.

[0134] Calculate the focal mechanism parameters, including the anisotropy index AI and the rupture surface normal vector obtained by moment tensor decomposition.

[0135] The improved S-transform time-frequency energy density integration method is used to calculate the energy release rate, eliminating the high-frequency aliasing error of the traditional method and obtaining the energy release of microseismic events more accurately.

[0136] An adaptive environmental noise base library can be further deployed to separate construction vibration and rock fracture signals through matching pursuit algorithm to improve signal purity; an array sensor collaborative verification mechanism is designed. When a single sensor detects a microseismic event, three adjacent sensors are activated to form a temporary detection array.

[0137]

[0138] The valid event is confirmed only when C coh is the waveform coherence coefficient between sensors, and the temperature-wave velocity compensation model v(T)=v 0 ·[1-α(TT 0 )] 1 / 2 , where v(T) is the wave velocity at the current temperature, α is the thermal expansion coefficient of the rock mass, and T 0In order to calibrate the temperature and correct the source location coordinates in real time, a transfer learning strategy is adopted, and the time-frequency characteristics of historical microseismic events are used as a pre-training data set to improve the feature recognition accuracy under small sample conditions.

[0139] (III) Construction of an inversion model for geological structural evolution

[0140] Based on the regional geological map, the regional geological structure characteristics are analyzed and the initial tectonic stress field is established to provide initial conditions for subsequent simulations.

[0141] Collect statistical results of on-site drill core joints, analyze the distribution, occurrence and other characteristics of the joints, invert the tectonic movement periods, and determine the geological tectonic movement history experienced by the slope.

[0142] The reverse time marching method is used to simulate the tectonic movement process. According to the sequence of geological tectonic movement, the simulation is carried out step by step from the present to the past, and the residual stress distribution is iteratively calculated. The mechanical response of the rock in the tectonic movement is considered to accurately obtain the distribution characteristics of the tectonic residual stress field.

[0143] Through acoustic emission experiments, the failure process of rock under different stress conditions is simulated, the parameters of the rock damage constitutive equation are calibrated, more accurate rock mechanics parameters are provided for the model, and the simulation accuracy of the model is improved.

[0144] (IV): Establish a multi-target classification model integrating machine learning

[0145] The static structural parameters, dynamic microseismic characteristics and tectonic stress field parameters obtained in the above steps are used as input variables to construct the input data set.

[0146] An improved random forest algorithm is used: a dynamic feature weight mechanism is introduced, and the split node selection strategy is adjusted according to the mutual information between features, so that the algorithm can more reasonably select features for splitting and improve classification efficiency; a multi-objective optimization function is designed to minimize the classification error and feature dimension, while ensuring classification accuracy and reducing the complexity of the model; a Bayesian optimization algorithm is used to automatically adjust the decision tree depth and the minimum number of leaf nodes, and optimize the structure and parameters of the model; a SHAP value interpretation module is integrated to output a quantitative indicator of the contribution of each feature parameter to the stability classification, which facilitates understanding of the decision-making process of the model and the importance of each parameter.

[0147] In this embodiment, it also includes establishing a multi-target classification model that integrates machine learning, and the model includes an integrated multi-model fusion strategy and feature cross-combination and deep feature mining.

[0148] It is worth noting that in addition to the improved random forest algorithm, the integrated multi-model fusion strategy introduces multiple machine learning models such as support vector machine (SVM) and gradient boosted decision tree (GBDT). These models are used to train the input variables respectively to obtain the classification results of different models. Then the model fusion is carried out by weighted voting, and the weight of each model is determined according to its performance on the validation set. For example, in multiple experiments, if the accuracy of the improved random forest algorithm on the validation set is 85%, SVM is 80%, and GBDT is 78%, their weights in the fusion model can be determined according to the accuracy ratio. Through multi-model fusion, the advantages of different models are fully utilized, the limitations of a single model are reduced, and the accuracy of the classification model is improved by 8%-12% under complex geological conditions.

[0149] Furthermore, feature cross-combination and deep feature mining are performed on the input static structural parameters, dynamic microseismic features and tectonic stress field parameters before model training. For example, the structural surface spacing and energy release rate are cross-operated to generate new features to explore the potential relationship between different features. At the same time, deep learning technologies such as autoencoders are used to perform deep feature mining on the original features and cross-combination features to extract higher-level and more representative features. These newly generated features can more accurately reflect the inherent laws of slope stability and provide richer and more effective information for the classification model, thereby improving the classification accuracy of the model.

[0150] (V) Design a dynamic feedback correction mechanism

[0151] Set dual threshold trigger conditions: the cumulative microseismic energy E on the day d ≥E threshold1 And the main frequency offset Δf ≥ Δf threshold When the first level correction is initiated;

[0152] When E is met for 3 consecutive days d ≥0.7E threshold1 The second level correction is initiated.

[0153] The correction process uses an online learning algorithm to update the model parameters and retains the memory factor of historical data during the update process, so that the model can make full use of new data without losing important information in the historical data.

[0154] The prediction performance of the updated model was evaluated through K-fold time series cross validation to ensure that the model still had good prediction accuracy and stability after the update.

[0155] The model robustness enhancement module can be further added: design a feature drift detector to monitor the changes in the input data distribution by calculating the KL divergence, and start the model protection mechanism when the KL value is >2.0; establish an adversarial sample generation network, inject labeled noise data before the model is updated, and test the model's anti-interference ability; use the elastic weight consolidation algorithm (EWC) to retain important weight history memory when updating parameters to prevent catastrophic forgetting; set up a model rollback mechanism, and automatically restore to the previous stable version when the verification set accuracy drops by more than 5% for 5 consecutive iterations.

[0156] (VI): Generate 3D visualization classification results

[0157] Integrate relevant modules on the digital twin platform: establish a real-time monitoring data cockpit, display the spatial distribution heat map of microseismic events in real time, and intuitively show the occurrence location and intensity distribution of microseismic events.

[0158] The stability grade 3D shading rendering module is used to map the stability coefficient using the HSV color space to perform a 3D visualization of the slope stability grade. Different colors represent different stability grades, which facilitates intuitive judgment of the slope stability.

[0159] Through the critical sliding surface probability cloud map generation module, the probability of sliding surface occurrence is calculated based on Monte Carlo simulation, and the possible location of the critical sliding surface is predicted, providing an important reference for slope support.

[0160] With the help of the support structure optimization suggestion module, the support parameter database is automatically matched according to the stability level, and optimization suggestions are provided for slope support, guiding engineering personnel to take reasonable support measures to ensure the stability of the slope.

[0161] The above specific embodiments are only several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A rock engineering slope stability classification method, characterized in that: The following steps are involved: S1. The spatial distribution characteristics of the slope structural surface are obtained through the joint detection of 3D laser scanning and geological radar, and a quantitative index system including the occurrence, density and connectivity of the structural surface is established; S2. Use the microseismic monitoring system to collect the internal rupture signal of the slope in real time, and extract the energy release rate, main frequency offset and focal mechanism parameters of the microseismic event through time-frequency analysis; S3. Construct a geological structural evolution inversion model, reconstruct the geological structural evolution process based on the regional geological database and field drilling data, and calculate the distribution characteristics of the structural residual stress field; S4, establishing a multi-objective classification model integrating machine learning, taking the static structural parameters, dynamic microseismic characteristics and tectonic stress field parameters obtained in steps S1-S3 as input variables, and ranking the feature importance by using an improved random forest algorithm; S5. Design a dynamic feedback correction mechanism. When the cumulative energy of the microseismic events monitored in real time exceeds the preset threshold, it automatically triggers the iterative update of the classification model parameters and simultaneously corrects the boundary conditions for determining the slope stability level. S6. Generate three-dimensional visual classification results, and integrate the stability grade zoning, critical sliding surface prediction results and support optimization suggestions on the digital twin platform.

2. A rock engineering slope stability classification method according to claim 1, characterized in that: In step S1, the three-dimensional laser scanning adopts multi-site cloud registration technology, and the automatic calibration of the scanning coordinate system is realized by introducing an inertial navigation unit. The structural surface recognition algorithm integrates point cloud curvature analysis and regional growing method to automatically extract the structural surface spacing d, trace length L and roughness JRC parameters.

3. A rock engineering slope stability classification method according to claim 1, characterized in that: The microseismic signal processing in step S2 adopts the wavelet packet-Hilbert transform joint analysis method to establish a multi-dimensional feature matrix of energy-frequency-time, where: The main frequency offset Δf is calculated by the instantaneous frequency estimation method, which reflects the rock mass fracture propagation speed; The focal mechanism parameters include the anisotropy index AI and the rupture surface normal vector obtained by moment tensor decomposition; The energy release rate is calculated using an improved S-transform time-frequency energy density integration method to eliminate the high-frequency aliasing error of the traditional method.

4. A rock engineering slope stability classification method according to claim 1, characterized in that: The geological structure evolution inversion model in step S3 is constructed by using the finite element-discrete element coupling algorithm through the following steps: Establish an initial tectonic stress field based on the regional geological map; Invert the tectonic movement period using the statistics of drill core joints; The reverse time marching method is used to simulate the tectonic movement process and iteratively calculate the residual stress distribution; The parameters of rock damage constitutive equation are calibrated through acoustic emission experiments.

5. A rock engineering slope stability classification method according to claim 1, characterized in that: The improved random forest algorithm in step S4 includes: Introducing a dynamic feature weight mechanism to adjust the split node selection strategy based on the mutual information between features; Design a multi-objective optimization function to minimize classification error and feature dimension; Use Bayesian optimization algorithm to automatically adjust the decision tree depth and the minimum number of samples for leaf nodes; The SHAP value interpretation module is integrated to output quantitative indicators of the contribution of each characteristic parameter to the stability classification.

6. A rock engineering slope stability classification method according to claim 1, characterized in that: The dynamic feedback correction mechanism in step S5 includes: Set dual threshold trigger conditions: the cumulative microseismic energy E on the day d ≥E threshold1 And the main frequency offset Δf ≥ Δf threshold When the first level correction is initiated; When E is met for 3 consecutive days d ≥0.7E threshold1 The second level correction is initiated when The correction process uses an online learning algorithm to update the model parameters while retaining the memory factor of historical data; The prediction performance of the updated model was evaluated by K-fold time series cross validation.

7. A rock engineering slope stability classification method according to claim 1, characterized in that: The digital twin platform in step S6 integrates the following modules: Real-time monitoring data cockpit, showing the heat map of the spatial distribution of microseismic events; Stability level 3D shading rendering module, using HSV color space to map stability coefficients; Critical sliding surface probability cloud map generation module, which calculates the probability of sliding surface occurrence based on Monte Carlo simulation; The support structure optimization suggestion module automatically matches the support parameter database according to the stability level.

8. A rock engineering slope stability classification method according to claim 2, characterized in that: The structural surface recognition algorithm also includes the following steps: Construct a point cloud anomaly filtering network based on deep learning, and use a three-dimensional convolutional neural network to identify and remove interference point clouds caused by vegetation coverage and temporary support structures; Design a multi-scale structural surface matching algorithm, quantify the structural surface morphological similarity through Fourier descriptors, and automatically associate the matching structural surfaces of adjacent scanning sites; The robust estimation theory is introduced to correct the structural surface attitude parameters and the following compensation model is established: Where Δθ is the correction value of the occurrence angle, k is the rock weathering coefficient, JRC i is the roughness coefficient of each sub-region, d is the distance between the structural surfaces; The water seepage status of the structural surface is inverted by the laser echo intensity, and the infrared thermal imaging auxiliary verification module is triggered when the echo intensity attenuation rate is greater than 15%.

9. A rock engineering slope stability classification method according to claim 3, characterized in that: The microseismic signal processing further includes: Deploy an adaptive ambient noise base library and separate construction vibration and rock fracture signals through matching pursuit algorithm; A collaborative verification mechanism for array sensors is designed. When a single sensor detects a microseismic event, three adjacent sensors are activated to form a temporary detection array. A valid event is confirmed only when the following conditions are met: Among them, C coh is the waveform coherence coefficient between sensors; Establish a temperature-wave velocity compensation model to correct the earthquake source location coordinates in real time: v(T)=v0·[1-α(T-T0)] 1 / 2 Among them, v(T) is the wave velocity at the current temperature, α is the thermal expansion coefficient of the rock mass, and T0 is the calibration temperature; A transfer learning strategy is adopted to use the time-frequency characteristics of historical microseismic events as a pre-training dataset to improve the feature recognition accuracy under small sample conditions.

10. A rock engineering slope stability classification method according to claim 6, characterized in that: The dynamic feedback correction mechanism also includes a model robustness enhancement module: Design a feature drift detector to monitor changes in input data distribution by calculating KL divergence, and activate the model protection mechanism when the KL value is >2.0; Establish an adversarial sample generation network, inject labeled noise data before model updating, and test the model's anti-interference ability; Adopt elastic weight consolidation algorithm to retain important weight history memory when updating parameters to prevent catastrophic forgetting; Set up a model rollback mechanism. When the validation set accuracy drops by more than 5% for 5 consecutive iterations, it will automatically revert to the previous stable version.

Images