Precise prediction method and system for surrounding rock disaster combining geology semantics and model scheduling

Abstract

This invention discloses a method and system for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling. The method includes: acquiring data related to surrounding rock disasters; extracting multiple dimensional features from the data; normalizing the features to obtain scene feature vectors; selecting multiple surrounding rock disaster prediction models from a model database as candidate models based on the scene feature vectors; determining the prediction performance of each candidate model; selecting the candidate model with the best performance index as the optimal surrounding rock disaster prediction model; and using the optimal surrounding rock disaster prediction model to identify the scene feature vectors to obtain the surrounding rock disaster prediction result. This achieves accurate prediction of surrounding rock disasters.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering technology, and in particular to a method and system for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling. Background Technology

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

[0003] In the field of geotechnical engineering, the safety management of projects such as tunnel excavation and underground cavern construction relies on the accurate prediction of surrounding rock disasters (such as sudden structural instability and cumulative deformation exceeding limits). Existing methods utilize machine learning models for predicting rockfall hazards, but model selection relies heavily on human experience. Current machine learning models exhibit heterogeneity and fragmentation, with different systems corresponding to different tasks—parameter inversion primarily uses optimization algorithms combined with numerical simulation (e.g., PSO-FLAC3D), time series prediction mainly uses recurrent neural networks (RNN / LSTM / GRU), and risk grading primarily uses ensemble learning (XGBoost / random forest). These models differ significantly in input dimension requirements, applicable scenario boundaries, and performance baselines. In practice, model selection relies on manual decision-making by engineers, requiring them to first categorize engineering scenarios based on experience and then test the performance of 3-5 models through trial and error. This process is prone to overlooking optimal models due to experience bias. Furthermore, most models do not consider the geological semantic correlation of geotechnical data, resulting in reduced cross-scenario generalization ability, failing to meet dynamic prediction requirements, and not guaranteeing the accuracy of rockfall hazard prediction results. Summary of the Invention To address the aforementioned problems, this invention proposes a method and system for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling, enabling accurate prediction of surrounding rock disasters in various scenarios.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling is proposed, including: Obtain data related to surrounding rock disasters; Extract multiple dimensional features from data related to surrounding rock disasters; Normalize the features of each dimension to obtain the scene feature vector; Based on the scene feature vector, multiple surrounding rock disaster prediction models are selected from the model database as candidate models; Determine the predictive performance of each candidate model; The candidate model with the best performance index is selected as the optimal surrounding rock disaster prediction model; By using the optimal surrounding rock disaster prediction model, the scene feature vector is identified to obtain the surrounding rock disaster prediction results.

[0005] Furthermore, data related to surrounding rock disasters include engineering monitoring data, geological survey data, and engineering design data; Multiple dimensions of characteristics include geological, engineering, and environmental characteristics.

[0006] Furthermore, the cosine similarity between the scene feature vector and the baseline vector of each surrounding rock disaster prediction model is calculated. Rock hazard prediction models with a cosine similarity greater than or equal to a set similarity threshold are selected as candidate models.

[0007] Furthermore, the surrounding rock disaster prediction model includes a displacement prediction model and a risk level determination model; The displacement prediction model determines the predicted displacement value based on the scene feature vector; The risk level determination model determines the risk level of surrounding rock disaster based on scene feature vectors and displacement prediction values.

[0008] Furthermore, the prediction error rate and prediction time of each candidate model are determined; Based on the prediction error rate and prediction time, the comprehensive score of each candidate model is calculated and determined, and the comprehensive score is used as the performance index of the candidate model.

[0009] Furthermore, the model database stores multiple surrounding rock disaster prediction models, each of which corresponds to a type of scenario feature vector. The process of obtaining the surrounding rock disaster prediction model includes: Acquire multiple training-related data on surrounding rock disasters; Determine the scene feature vector for each training data related to surrounding rock disasters; Cluster the scene feature vectors of all training data related to rock-surrounding disasters to obtain multiple feature clusters; Based on the scene feature vectors in each feature cluster, determine the data preprocessing rules, model parameter constraint rules, and result verification rules; According to the data preprocessing rules, the scene feature vectors in the corresponding feature clusters are preprocessed to obtain the preprocessed vectors; Based on the model parameter constraint rules and result verification rules, the preprocessed vectors of the corresponding feature clusters are used to train the basic model for surrounding rock disaster prediction. After training is completed, the surrounding rock disaster prediction model corresponding to the feature vectors of this type of scenario is obtained.

[0010] Secondly, a precise prediction system for surrounding rock disasters integrating geological semantics and model scheduling is proposed, including: The geological semantic perception and scene feature extraction module is used to acquire data related to surrounding rock disasters; extract multi-dimensional features from the data related to surrounding rock disasters; and normalize the features of each dimension to obtain scene feature vectors. The multi-model intelligent scheduling and adaptation module is used to select multiple surrounding rock disaster prediction models from the model database as candidate models based on the scene feature vector; determine the prediction performance of each candidate model; and select the candidate model with the best performance index as the optimal surrounding rock disaster prediction model. The end-to-end collaborative prediction and decision output module is used to identify scene feature vectors using the optimal surrounding rock disaster prediction model to obtain surrounding rock disaster prediction results.

[0011] Thirdly, a computer device is proposed, the device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling, as proposed in the first aspect.

[0012] Fourthly, a computer-readable storage medium is proposed, which stores a computer program suitable for being loaded and executed by a processor to provide a precise prediction method for surrounding rock disasters that integrates geological semantics and model scheduling, as proposed in the first aspect.

[0013] Fifthly, a computer program product is proposed, which includes a computer program. When the computer program is executed by a processor, it implements the method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling proposed in the first aspect.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a method and system for accurate prediction of rockfall disasters by integrating geological semantics and model scheduling. The method obtains scene feature vectors from relevant rockfall disaster data; based on these feature vectors, it selects multiple rockfall disaster prediction models from a model database as candidate models and determines the performance index of each candidate model; finally, it selects the candidate model with the best performance index as the optimal rockfall disaster prediction model, identifies the scene feature vectors, and obtains the rockfall disaster prediction result; this ensures that the final optimal rockfall disaster prediction model is the best-performing model adapted to the scene, thereby guaranteeing the accuracy of rockfall disaster prediction.

[0015] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0017] Figure 1 This is a flowchart of the method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling, as proposed in this embodiment of the invention. Figure 2 This is a diagram illustrating the overall technical architecture of the accurate prediction method for surrounding rock disasters that integrates geological semantics and model scheduling, as proposed in this embodiment of the invention. Figure 3 This is a schematic diagram illustrating the domain knowledge fusion principle proposed in an embodiment of the present invention. Figure 4 This is a flowchart of the multi-model intelligent scheduling process proposed in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0021] The method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling proposed in this embodiment of the invention is applied to the application scenario of surrounding rock disaster prediction.

[0022] In the field of geotechnical engineering, the safety management of projects such as tunnel excavation and underground cavern construction relies heavily on the accurate prediction of surrounding rock disasters (such as sudden structural instability and cumulative deformation exceeding limits). With the expansion of deep underground space development and engineering construction under complex geological conditions in my country, surrounding rock disasters exhibit characteristics of "multiple coupled causes, nonlinear evolution, high sudden risk, and severe consequences." Traditional technologies are no longer sufficient to meet the requirements for "real-time, accurate, and reliable" prediction. Traditional methods, such as static assessments based on empirical formulas (e.g., deriving stability coefficients from the Hoek-Brown strength criterion), rely on artificial assumptions and cannot characterize dynamic evolution. Numerical simulation-based finite element / discrete element analysis (e.g., FLAC3D, UDEC) requires pre-setting numerous parameters and has low computational efficiency (simulating a single deep tunnel of 100 meters in length takes over 24 hours), making it difficult to adapt to real-time monitoring data updates. Against this backdrop, machine learning, with its advantages in nonlinear mapping, multi-source data fusion, and dynamic prediction, has gradually become a core technology path, forming a multi-type model system covering rock mechanics parameter inversion, time-series deformation prediction, disaster risk classification, and precursor signal identification, providing new support for engineering safety decision-making. However, existing technologies suffer from three major pain points. The primary bottleneck is the contradiction between geotechnical engineering characteristics and model adaptability. Rock mass disasters are controlled by multi-field coupling of "geology-engineering-environment." Geologically, the random distribution of rock mass structural surfaces leads to high dispersion in mechanical parameters, and small sample scenarios often occur in engineering sites, making traditional statistical models prone to overfitting. In terms of engineering disturbances, excavation unloading, support construction, etc., cause the rock mass response to exhibit non-stationary temporal characteristics, which static models cannot capture. Regarding environmental factors, groundwater seepage, temperature changes, and rock mass deformation are complexly coupled nonlinearly, and single-dimensional models (inputting only displacement data) are prone to missing key risk signals. Furthermore, current technologies lack standardized "scenario-model" adaptation logic, requiring customized models for specific scenarios, which limits predictive accuracy. Secondly, there are the issues of experience-dependent model selection and efficiency bottlenecks. Existing machine learning models are characterized by heterogeneity and fragmentation, with different systems corresponding to different tasks. Parameter inversion mainly relies on "optimization algorithms + numerical simulation" (such as PSO-FLAC3D), time series prediction mainly relies on recurrent neural networks (RNN / LSTM / GRU), and risk classification mainly relies on ensemble learning (XGBoost / random forest). The input dimension requirements, applicable scenario boundaries, and performance baselines of various models differ significantly. In application, model selection relies on manual decisions by engineers, who first classify engineering scenarios based on experience and then test the performance of 3-5 types of models through a "trial and error" method. This process is prone to missing the optimal model due to experience bias. At the same time, most models do not consider the "geological semantic correlation" of geotechnical data, resulting in reduced cross-scenario generalization ability and failing to meet the needs of dynamic prediction. Furthermore, there is a disconnect between domain knowledge and model application. Predicting surrounding rock disasters requires deep integration of geomechanical theory (Mohr-Coulomb strength criterion, rock mass quality classification Q system), engineering monitoring experience (displacement rate mutation threshold, support timing standard), and disaster evolution laws (deformation "initial-acceleration-instability" three stages). However, in existing technologies, most geotechnical models focus only on a single task and do not embed domain knowledge rules, resulting in output results lacking engineering semantic support, making it difficult for engineers to judge reliability. At the data processing level, general methods do not consider the "geological correlation" of geotechnical data, which easily leads to data distortion and reduced accuracy. Current solutions for predicting rockfall disasters mainly fall into two categories: "geotechnical-specific models" and "general-purpose machine learning platforms," ​​both of which have significant shortcomings. Geotechnical-specific models often exist as "single-task tools." Parameter inversion models can only identify mechanical parameters, requiring manual export to connect with subsequent prediction tasks, increasing errors and time consumption. While time-series prediction models can output short-term displacement values, they cannot combine geological conditions to classify the reliability of the results, making it difficult to support differentiated prevention and control. Risk warning models rely on manually set weights and cannot dynamically adapt to different scenarios. General-purpose machine learning platforms, while supporting multiple model calls, lack geological semantic adaptation modules, cannot identify the geological attributes of geotechnical data, and are prone to losing key information. Furthermore, they lack engineering decision-making interfaces, and their outputs are mostly abstract probability values ​​or curves, requiring engineers to interpret them into engineering indicators such as "stability coefficients" and "disaster probabilities," failing to directly support on-site decision-making. In summary, existing technologies cannot solve the three core pain points, necessitating the development of a technical solution that integrates "scenario-model intelligent matching, deep embedding of domain knowledge, and full-process collaborative prediction" to improve prediction accuracy and efficiency and ensure the safety of geotechnical engineering.

[0023] To achieve accurate prediction of surrounding rock disasters, this invention proposes a method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling. First, the surrounding rock disaster prediction models are initially screened based on the scenario. Then, based on the predictive performance of the models, the initially screened models are screened again to determine the optimal surrounding rock disaster prediction model. When the optimal surrounding rock disaster prediction model is used to predict surrounding rock disasters in this scenario, the accuracy of the surrounding rock disaster prediction is guaranteed.

[0024] like Figures 1-4 As shown in the embodiments of the present invention, the method for accurate prediction of surrounding rock disasters by integrating geological semantics and model scheduling includes: Obtain data related to surrounding rock disasters; Extract multiple dimensional features from data related to surrounding rock disasters; Normalize the features of each dimension to obtain the scene feature vector; Based on the scene feature vector, multiple surrounding rock disaster prediction models are selected from the model database as candidate models; Determine the predictive performance of each candidate model; The candidate model with the best performance index is selected as the optimal surrounding rock disaster prediction model; By using the optimal surrounding rock disaster prediction model, the scene feature vector is identified to obtain the surrounding rock disaster prediction results.

[0025] After obtaining the scene feature vector of the surrounding rock disaster-related data in this embodiment of the invention, multiple surrounding rock disaster prediction models are selected from the model database as candidate models based on the scene feature vector, and the performance index of each candidate model is determined. Finally, the candidate model with the best performance index is selected as the optimal surrounding rock disaster prediction model, and the scene feature vector is identified to obtain the surrounding rock disaster prediction result. This ensures that the final optimal surrounding rock disaster prediction model is the best performing model that is adapted to the scene, thereby ensuring the accuracy of the surrounding rock disaster prediction.

[0026] In some embodiments, data related to surrounding rock disasters include engineering monitoring data, geological survey data, and engineering design data; Multiple dimensions of characteristics include geological, engineering, and environmental characteristics.

[0027] The engineering monitoring data includes displacement, seepage pressure, and stress.

[0028] Geological exploration data includes core samples, Kv data, and structural planes.

[0029] Engineering design data includes excavation and support parameters, etc.

[0030] For engineering monitoring data, a data parsing script was developed, using Python's Pandas library to implement custom mappings for the data (e.g., "monitoring number → monitoring point ID" and "collection time → timestamp"). An automatic data type validation function was also developed (displacement values ​​are floating-point numbers, and timestamps conform to the "YYYY-MM-DD HH:MM:SS" format). An alarm threshold of ≤0.5% for a single batch of data anomaly rate was set. If this threshold is exceeded, an alarm log is immediately recorded, and subsequent processing of that batch of data is suspended until manual intervention. During the data cleaning phase, missing values ​​are marked as NaN, and duplicate data determined based on monitoring point ID and collection time are directly deleted to ensure data format compliance and valid physical meaning.

[0031] For geological exploration data, a geological exploration data parsing script was developed. The openpyxl library was used to read multi-sheet data such as "borehole information", "lithological stratification", and "mechanical parameters". A cross-sheet data association algorithm based on "borehole ID" was designed to achieve a one-to-one correspondence between lithological stratification and mechanical parameters, thus solving the problem of scattered data storage.

[0032] For engineering design data, develop engineering design data parsing scripts, using the ezdxf library to parse geometric entities such as excavation outlines and support structure locations in CAD drawings, and then using preset coordinate transformation parameters (X... 大地 =a × X CAD +b, Y 大地 =c × Y CAD +d, a, b, c, d (calibrated by engineering surveying) convert the CAD coordinate system to the geodetic coordinate system, extract design parameters such as excavation radius and anchor bolt length, and store them in a structured JSON format.

[0033] In this embodiment of the invention, after obtaining data related to the surrounding rock disaster, missing data was supplemented and abnormal data was deleted to obtain accurate and complete data for subsequent data processing.

[0034] For missing data, a Kriging interpolation algorithm based on lithological similarity was used, introducing lithological similarity weights (0.8 for the same lithology, 0.3 for adjacent lithologies), and an optimization formula was applied. Calculate the interpolation point values ​​to ensure that the interpolation error is reduced compared to traditional methods in areas with frequent lithological changes.

[0035] In the above formula, is the predicted value of the point to be interpolated, referring to the estimated value calculated for the missing location in the surrounding rock disaster-related data; i is the index of the known sample point, representing the sequence number of the adjacent known monitoring or exploration points involved in the calculation; n is the total number of known sample points involved in the interpolation; The Kriging weight coefficient for the i-th sample point is determined by the Kriging interpolation algorithm based on spatial correlation, and the sum of all the weight coefficients is usually 1. This represents the measured value of the i-th known sample point. For example, the original monitoring data such as displacement, stress, or mechanical parameters at that point. As for the lithological similarity weight, the key parameters set by this invention based on geological conditions are: 0.8 when the lithology of the point to be interpolated is the same as that of the known point; and 0.3 when the lithology of the two points is adjacent. This is the average value of the sample area, used to smooth the predicted value by means of regional statistical characteristics when there are large differences in lithology.

[0036] For abnormal data, a dual mechanism of "3σ principle + geological semantic verification" is adopted: first, abnormal values ​​are initially screened by the 3σ principle (mean ± 3 times standard deviation), and then a second verification is performed by combining geological semantic rules (e.g., if the displacement change in the karst cave area corresponds to the construction disturbance record 9, it is judged as normal, and if there is no disturbance, it is judged as abnormal), to ensure that the data integrity reaches more than 95%.

[0037] This invention constructs a three-dimensional feature system of "geology-engineering-environment", extracting geological, engineering, and environmental features from data related to surrounding rock disasters. Among them, the geological features include four features: lithology code, Kv value, structural surface density, and fault influence distance; the engineering features include four features: excavation depth, excavation rate, support stiffness, and monitoring point density; and the environmental features include four features: pore water pressure, temperature change rate, rainfall, and groundwater depth, for a total of 12 core features.

[0038] Twelve core features were quantified to obtain the quantified features. Among them, the lithology code adopts a two-level code of "major category + subcategory" (limestone 0102, sandstone 0201), the Kv value is taken as a continuous value of 0.1~1.0 according to the "Engineering Rock Mass Classification Standard", the structural surface density is counted as "strips / m", the fault influence distance is calculated as "horizontal distance from the fault center (m)", and the remaining features are quantified according to the actual monitoring or design values ​​of the project.

[0039] The Analytic Hierarchy Process (AHP) was used to determine the weights of each feature. Multiple geotechnical engineering experts were invited to score the importance of the features. A judgment matrix was constructed and passed the consistency test (CR<0.1). Finally, the weights of fault influence distance, pore water pressure, etc. were determined. A fusion feature was generated by weighted summation. The fusion feature was used to normalize each core feature to generate a 1×12-dimensional scene feature vector V normalized to [0,1].

[0040] In some embodiments, the cosine similarity between the scene feature vector and the baseline vector of each surrounding rock disaster prediction model is calculated. Rock hazard prediction models with a cosine similarity greater than or equal to a set similarity threshold are selected as candidate models.

[0041] In this embodiment of the invention, multiple surrounding rock disaster prediction models are stored in the model database, and each surrounding rock disaster prediction model corresponds to a type of scenario feature vector; The process of obtaining the surrounding rock disaster prediction model includes: Acquire multiple training-related data on surrounding rock disasters; Determine the scene feature vector for each training data related to surrounding rock disasters; Cluster the scene feature vectors of all training data related to rock-surrounding disasters to obtain multiple feature clusters; Based on the scene feature vectors in each feature cluster, determine the data preprocessing rules, model parameter constraint rules, and result verification rules; According to the data preprocessing rules, the scene feature vectors in the corresponding feature clusters are preprocessed to obtain the preprocessed vectors; Based on the model parameter constraint rules and result verification rules, the preprocessed vectors of the corresponding feature clusters are used to train the basic model for surrounding rock disaster prediction. After training is completed, the surrounding rock disaster prediction model corresponding to the feature vectors of this type of scenario is obtained.

[0042] The embodiments of the present invention construct a data preprocessing rule base (R1 class), a model parameter constraint rule base (R2 class), and a result verification rule base (R3 class).

[0043] Among them, the encoding format of the data preprocessing rule base (R1 class) is R1-XX, such as "R1-03: When Kv<0.4 and pore water pressure>50kPa, enable wavelet threshold denoising with db4 wavelet basis and threshold of 0.05". Each rule is associated with the "applicable lithology type" and "data type" fields to ensure accurate rule matching.

[0044] Model parameter constraint rule library (R2 class), encoded in R2-XX format, categorized by model type, such as "R2-12: Karst area LSTM model parameters: time step 24, number of hidden layer nodes 64, learning rate 0.001, number of iterations 500". The rule parameters are determined based on the training effect optimization of 50+ historical cases.

[0045] The result verification rule base (R3 class) is encoded in the format R3-XX, such as "R3-02: When the surrounding rock stability coefficient is <1.2, trigger cross-verification of displacement rates of 3 adjacent monitoring points. If all are >3mm / d, the risk is valid". The rule priority is divided into high (P1), medium (P2), and low (P3). High priority rules must be executed.

[0046] Develop a rule matching engine. The input is a scene feature vector V. The engine filters the corresponding rules through "feature value matching + applicable scene matching" (e.g., if Kv=0.35<0.4 and pore water pressure=60kPa>50kPa in V, then it matches R1-03). The engine provides an API interface based on the Flask framework of Python to ensure that the response time is ≤100ms.

[0047] Design a model configuration file (JSON format) and associate R2 class rules in the "Parameter Constraint" field (e.g., "param_constraint": "R2-12" in the LSTM configuration file). When the model is loaded, the rule base API will be automatically called to obtain the parameter values, avoiding errors from manual configuration.

[0048] The R3 class rules are transformed into penalty terms in the model loss function. Taking the XGBoost risk grading model as an example, the optimized loss function is as follows: ,in, For logarithmic loss, =0.3 is the penalty coefficient, applied when violating Rule R3. =1), by optimizing parameters through gradient descent, the accuracy of risk classification can be improved.

[0049] Surrounding rock disaster prediction models include displacement prediction models and risk level determination models; The displacement prediction model determines the predicted displacement value based on the scene feature vector; The risk level determination model determines the risk level of surrounding rock disaster based on scene feature vectors and displacement prediction values.

[0050] The displacement prediction model can be based on models such as LSTM. The input of this model is 12 scene feature vectors and known historical displacement data (which can be displacement data from the past 30 days). The output is the displacement prediction value for the next 14 days. The model uses time-series data consisting of 100 sets of scene feature vectors and historical displacement data (sampling frequency 1 hour / time). The Adam optimizer, batch size 32, and dropout=0.2 are used to ensure that the prediction error on the test set is ≤12% and the error on small sample scenes (<10 days of data) is ≤15%. After training, the displacement prediction model is obtained.

[0051] The trained PSO-RF-FLAC3D parameter inversion model can be used to identify scene feature vectors and determine their corresponding displacement data. During the training of the PSO-RF-FLAC3D parameter inversion model, the objective function is to minimize the RMSE between the monitored displacement and the FLAC3D simulated displacement. 80 sets of monitored displacement-known scene feature vectors (training set: test set = 7:3) are used. The number of particles is optimized by grid search with 30 and the number of iterations is 100 to ensure that the RMSE of the test set is ≤0.5mm and the inversion accuracy is ≥85%.

[0052] The risk level determination model can use a base model such as XGBoost. The input of this model is 12 scene feature vectors and displacement prediction values, and the output is the surrounding rock disaster risk level. It uses 120 sets of scene feature vectors and displacement prediction values ​​with known surrounding rock disaster risk levels. The tree depth is optimized to 6, the learning rate is 0.1, and the subsample ratio is 0.8 through 5-fold cross-validation to ensure that the classification accuracy of the test set is ≥88% and the Kappa coefficient is ≥0.82. After training, the risk level determination model is obtained.

[0053] After determining the surrounding rock disaster prediction model corresponding to the feature vector of each scenario, the prediction error rate ε and prediction time t of the surrounding rock disaster prediction model corresponding to each scenario feature vector were also determined. The scenario feature vector and its corresponding prediction error rate and prediction time were used as training data. Multiple sets of training data were used to train a multiple linear regression model to obtain the trained linear regression model corresponding to each surrounding rock disaster prediction model. This model can reflect the relationship between the scenario feature vector and the prediction error rate ε and prediction time t, ensuring that ε prediction RMSE ≤ 3% and t prediction relative error ≤ 10%.

[0054] The surrounding rock disaster prediction model of this invention, after determining the surrounding rock disaster risk level, also determines prevention and control recommendations based on the surrounding rock disaster risk level. These recommendations include the type of measures and their implementation priority. For example, when the risk is "high," monitoring should be intensified to once every 10 minutes.

[0055] In this embodiment of the invention, a RabbitMQ message queue is deployed, creating four dedicated queues: param_queue, displacement_queue, risk_queue, and validation_queue. Message persistence and consumer confirmation mechanisms are set up, and the message format is JSON (including message ID, project ID, data content, etc.), ensuring that the data flow latency is ≤500ms.

[0056] The design incorporates a "multi-level verification" mechanism: the result verification module first performs data integrity verification (such as the continuity of displacement curves), then performs domain rule verification (calling R3 class rules), and finally performs model consistency verification (comparing the risk levels of XGBoost and Random Forest) to ensure the reliability of the results.

[0057] Design an "anomaly feedback and retry" mechanism: When validation fails, generate an instruction containing "anomaly type, location, and retry suggestion", send it to the corresponding module through validation_queue to trigger recalculation, set a maximum of 3 retries, automatically adjust model parameters for each retry, and trigger a manual alarm after 3 failures.

[0058] Develop engineering index conversion function: calculate the surrounding rock stability coefficient K based on the Hoek-Brown criterion (accuracy ±0.05), map the probability of disaster occurrence P through the Sigmoid function (accuracy ±3%), and perform moving average filtering (window 3) on the displacement prediction value output by LSTM to ensure that the displacement error in the next 14 days is ±0.5mm.

[0059] This invention also constructs a "risk level - measure type - implementation priority" mapping table (e.g., when the risk is "high", priority is given to encrypted monitoring up to 10 minutes / time), and generates suggestions containing "measure content, priority, and expected effect" in combination with domain rules, which are then output through the Vue.js web client and the Flutter mobile client.

[0060] In some embodiments, the cosine similarity between the scene feature vector and the baseline vector of each surrounding rock disaster prediction model is calculated. Rock hazard prediction models with a cosine similarity greater than or equal to a set similarity threshold are selected as candidate models.

[0061] Determine the prediction error rate and prediction time for each candidate model; Based on the prediction error rate and prediction time, the comprehensive score of each candidate model is calculated and determined, and the comprehensive score is used as the performance index of the candidate model.

[0062] Among them, the baseline vector for each surrounding rock disaster prediction model is determined by data fitting of the scene feature vector used for training the model.

[0063] Cosine similarity between scene feature vector V and model adaptation baseline vector M The similarity threshold can be set according to needs. For example, a threshold of 70% corresponds to a model adaptation accuracy of ≥90%.

[0064] Rock hazard prediction models with a sim ≥ 70% were selected as candidate models.

[0065] Based on the scene feature vector V, the prediction error rate and prediction time of each candidate model are determined from the trained linear regression models corresponding to each candidate model.

[0066] Therefore, the comprehensive score of each candidate model is calculated and determined. .

[0067] In the formula, The overall score of the candidate models serves as the final performance metric for measuring model performance. A higher score indicates better adaptability of the model to the current scenario. The prediction error rate of the candidate model is obtained by the trained linear regression model based on the current scene feature vector V. The maximum permissible error rate is set as an upper limit, which is used to normalize the prediction error rate; t The prediction time of the candidate model is obtained from the prediction of the trained linear regression model, which reflects the computational speed of the model under the current hardware and data scale. This is the maximum allowed time limit set, used to normalize the predicted time.

[0068] The candidate model with the highest score is selected by sorting by S and used as the optimal surrounding rock disaster prediction model to identify the scene feature vector and obtain the surrounding rock disaster prediction result, ensuring that the scheduling algorithm takes ≤10s.

[0069] The prediction results of the surrounding rock disaster include the surrounding rock disaster level, the type of measures and the implementation priority, and the prediction results are sent to the client for display.

[0070] The pre-mixing method for surrounding rock disasters proposed in this embodiment of the invention has the following effects: (1) Enhanced Adaptability. This invention resolves the contradiction between geotechnical engineering characteristics and model adaptability. Existing technologies suffer from high prediction errors in small-sample scenarios and large short-term prediction deviations in non-stationary time-series scenarios because they cannot capture the geological semantics of geotechnical data. This invention constructs a three-dimensional feature system of "geology-engineering-environment" through a geological semantic perception module, accurately quantifies geological parameters such as lithology and structural surface density, and embeds scenario-based constraints in conjunction with a domain knowledge rule base, thereby improving the adaptability of the model to engineering scenarios. Even in engineering projects with complex geological conditions such as fault fracture zones and deep-buried tunnels, it can still maintain stable prediction accuracy, and its adaptability covers a variety of challenging scenarios such as deep underground engineering, karst area engineering, and active fault zone engineering.

[0071] (2) Efficiency is optimized by orders of magnitude. It breaks through the efficiency bottleneck caused by "experience dependence on model selection". Existing technologies rely on engineers to manually classify scenarios and test models by trial and error. The entire model selection process takes a long time on average and is prone to missing the optimal model due to experience bias. This invention achieves full-process automation through a multi-model intelligent scheduling module: on the one hand, it builds a clearly classified model resource pool and marks the applicable scenario boundaries and performance baselines of each model to avoid the blindness of manual selection; on the other hand, it innovates a two-layer adaptation algorithm of "scenario feature matching degree + model performance ranking". First, it quickly selects the suitable model by cosine similarity (threshold ≥ 0.7) and then predicts the model accuracy and time consumption by pre-trained regression model, and automatically selects the best model according to "accuracy first and time consumption second". In addition, the full-process collaborative prediction module realizes the automatic data flow of "parameter inversion → deformation prediction → risk classification", reducing manual data transmission, secondary processing and other links. The overall prediction efficiency is improved compared with the existing technology, which meets the timeliness requirements of real-time monitoring and dynamic decision-making in engineering.

[0072] (3) Significantly enhanced practicality. Overcoming the barriers to engineering implementation caused by the disconnect between domain knowledge and the model. Existing technologies mostly output abstract probability values ​​or curves, requiring engineers to interpret them in conjunction with domain experience. Furthermore, their cross-scenario generalization ability is poor, making it difficult to directly support on-site decision-making. This invention solves this problem through two innovations: First, it deeply integrates domain knowledge into the entire model process through rule embedding and loss function penalty terms, giving the output results inherent engineering semantics. Second, it designs an engineering decision-making output module that transforms the calculation results into quantitative indicators, visualization results, and implementable prevention and control suggestions, which can directly guide on-site construction without the need for manual secondary interpretation. At the same time, due to the deep integration of geological semantics and domain knowledge, the model's cross-scenario generalization error is reduced, achieving "one-time development, multi-scenario reuse," providing standardized and intelligent technical support for geotechnical engineering safety management and reducing engineering safety risks and management costs.

[0073] This invention also proposes a precise prediction system for surrounding rock disasters that integrates geological semantics and model scheduling, including: The geological semantic perception and scene feature extraction module is used to acquire data related to surrounding rock disasters; extract multi-dimensional features from the data related to surrounding rock disasters; and normalize the features of each dimension to obtain scene feature vectors. The multi-model intelligent scheduling and adaptation module is used to select multiple surrounding rock disaster prediction models from the model database as candidate models based on the scene feature vector; determine the prediction performance of each candidate model; and select the candidate model with the best performance index as the optimal surrounding rock disaster prediction model. The end-to-end collaborative prediction and decision output module is used to identify scene feature vectors using the optimal surrounding rock disaster prediction model to obtain surrounding rock disaster prediction results.

[0074] It should be noted that the above-described embodiment of the accurate prediction system for surrounding rock disasters integrating geological semantics and model scheduling is only illustrated by the division of the functional modules described above. In practical applications, the functions described above can be assigned to different functional modules as needed, that is, the internal structure of the equipment can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the above-described embodiment of the accurate prediction system for surrounding rock disasters integrating geological semantics and model scheduling and the embodiment of the method for accurate prediction of surrounding rock disasters integrating geological semantics and model scheduling belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0075] The present invention also discloses a computer device, the device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling proposed in this embodiment of the invention.

[0076] The present invention also discloses a computer-readable storage medium storing a computer program adapted for loading and execution by a processor of the precise prediction method for surrounding rock disasters that integrates geological semantics and model scheduling proposed in the embodiments of the present invention.

[0077] The present invention also discloses a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling proposed in the embodiments of the present invention.

[0078] The method proposed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0079] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0080] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for accurate prediction of surrounding rock disasters integrating geological semantics and model scheduling, characterized in that, include: Obtain data related to surrounding rock disasters; Extract multiple dimensional features from data related to surrounding rock disasters; Normalize the features of each dimension to obtain the scene feature vector; Based on the scene feature vector, multiple surrounding rock disaster prediction models are selected from the model database as candidate models; Determine the predictive performance of each candidate model; The candidate model with the best performance index is selected as the optimal surrounding rock disaster prediction model; By using the optimal surrounding rock disaster prediction model, the scene feature vector is identified to obtain the surrounding rock disaster prediction results.

2. The method for accurate prediction of surrounding rock disasters integrating geological semantics and model scheduling as described in claim 1, characterized in that, Data related to surrounding rock disasters include engineering monitoring data, geological survey data, and engineering design data; Multiple dimensions of characteristics include geological, engineering, and environmental characteristics.

3. The method for accurate prediction of surrounding rock disasters integrating geological semantics and model scheduling as described in claim 1, characterized in that, Calculate the cosine similarity between the scene feature vector and the baseline vector of each surrounding rock disaster prediction model. Rock hazard prediction models with a cosine similarity greater than or equal to a set similarity threshold are selected as candidate models.

4. The method for accurate prediction of surrounding rock disasters by integrating geological semantics and model scheduling as described in claim 1, characterized in that, Surrounding rock disaster prediction models include displacement prediction models and risk level determination models; The displacement prediction model determines the predicted displacement value based on the scene feature vector; The risk level determination model determines the risk level of surrounding rock disaster based on scene feature vectors and displacement prediction values.

5. The method for accurate prediction of surrounding rock disasters integrating geological semantics and model scheduling as described in claim 1, characterized in that, Determine the prediction error rate and prediction time for each candidate model; Based on the prediction error rate and prediction time, the comprehensive score of each candidate model is calculated and determined, and the comprehensive score is used as the performance index of the candidate model.

6. The method for accurate prediction of surrounding rock disasters integrating geological semantics and model scheduling as described in claim 1, characterized in that, The model database stores multiple surrounding rock disaster prediction models, and each surrounding rock disaster prediction model corresponds to a type of scenario feature vector. The process of obtaining the surrounding rock disaster prediction model includes: Acquire multiple training-related data on surrounding rock disasters; Determine the scene feature vector for each training data related to surrounding rock disasters; Cluster the scene feature vectors of all training data related to rock-surrounding disasters to obtain multiple feature clusters; Based on the scene feature vectors in each feature cluster, determine the data preprocessing rules, model parameter constraint rules, and result verification rules; According to the data preprocessing rules, the scene feature vectors in the corresponding feature clusters are preprocessed to obtain the preprocessed vectors; Based on the model parameter constraint rules and result verification rules, the preprocessed vectors of the corresponding feature clusters are used to train the basic model for surrounding rock disaster prediction. After training is completed, the surrounding rock disaster prediction model corresponding to the feature vectors of this type of scenario is obtained.

7. A precise prediction system for surrounding rock disasters integrating geological semantics and model scheduling, characterized in that: include: The geological semantic perception and scene feature extraction module is used to acquire data related to surrounding rock disasters; Extract multiple dimensional features from data related to surrounding rock disasters; Normalize the features of each dimension to obtain the scene feature vector; The multi-model intelligent scheduling and adaptation module is used to select multiple surrounding rock disaster prediction models from the model database as candidate models based on the scene feature vector; determine the prediction performance of each candidate model; and select the candidate model with the best performance index as the optimal surrounding rock disaster prediction model. The end-to-end collaborative prediction and decision output module is used to identify scene feature vectors using the optimal surrounding rock disaster prediction model to obtain surrounding rock disaster prediction results.

8. An electronic device, characterized in that, The device includes: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted for loading and execution by a processor of the accurate prediction method for surrounding rock disasters that integrates geological semantics and model scheduling as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method for accurate prediction of surrounding rock disasters that integrates geological semantics and model scheduling as described in any one of claims 1-6.

Images