Tunnel surrounding rock grading method and system

Abstract

The invention relates to the technical field of tunnel engineering, in particular to a tunnel surrounding rock grading method and system, comprising intelligent sensing and data acquisition, multi-source data fusion and modeling, hybrid model dynamic grading, real-time decision and support optimization, online learning and dynamic feedback, and risk early warning and emergency response. Compared with the prior art that a geological data acquisition mode combining manual drilling coring and low-resolution geophysical prospecting is adopted, efficiency is low, subjective errors are large, and a complex geological structure is difficult to cover, unmanned aerial vehicle LiDAR scanning, intelligent rock core image analysis and a high-density IoT sensor network work cooperatively, and the working efficiency is greatly improved. Real-time dynamic acquisition of full-section geological information is achieved, manual intervention errors are eliminated in combination with a multi-source data fusion algorithm, the automation level and three-dimensional space representation precision of data acquisition are remarkably improved, and a high-resolution holographic data base is provided for surrounding rock classification.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel engineering, and in particular to a tunnel surrounding rock grading method and a system thereof. Background Art

[0002] In tunnel engineering, surrounding rock classification is a key technical step in ensuring construction safety and design rationality. Traditional methods, which rely primarily on manual experience and static classification systems (such as the Q system and RMR method), present significant technical bottlenecks.

[0003] Traditional geological exploration relies primarily on coring and manual recording, which is inefficient and susceptible to subjective errors. For example, the deviation rate of manually calculated joint density often exceeds 20%. While geophysical methods (such as seismic waves and electromagnetic waves) can supplement some data, their resolution is limited to the meter level, making it difficult to accurately identify microscopic geological features such as faults and fissures. Furthermore, the parameters of existing classification models are rigid and cannot dynamically adapt to changes in complex geological conditions, such as the release of ground stress after excavation and changes in seepage paths. This results in a disconnect between the classification results and the actual state of the surrounding rock. Numerical simulation techniques (such as finite element method (FEM)) are often limited to single-field analysis and lack the ability to simulate multi-field coupling effects (such as stress-seepage-chemical corrosion), further exacerbating the classification lag problem.

[0004] Traditional support design methods rely on empirical formulas or manuals, lacking real-time data-driven optimization. This often leads to overdesign (increasing costs by 30% to 50%) or inadequate support (increasing the risk of collapse). Adjustments to these plans rely on manual intervention, resulting in slow response times and a lack of capacity to address sudden geological risks. Existing risk warning systems are often based on threshold-based alarms (e.g., exceeding deformation limits), lack the ability to proactively predict time-series data, and are inefficient in generating emergency plans. This makes them unable to meet the rapid response requirements in high-risk areas such as water-rich fault zones. While recent research has attempted to improve data collection capabilities by introducing sensor networks or discrete element modeling (DEM), bottlenecks remain, such as isolated multi-source data, static models, and low automation. For example, geological, construction, and monitoring data lack a unified spatiotemporal benchmark, making integration difficult. Machine learning model parameters are rigid and cannot be dynamically optimized. Support execution relies on manual intervention.

[0005] Therefore, there is an urgent need for a surrounding rock classification system that integrates intelligent perception, dynamic modeling, real-time optimization and automated execution to break through the systematic limitations of traditional methods in terms of efficiency, accuracy, real-time performance and safety. Summary of the Invention

[0006] In order to overcome the problems raised in the above background technology, the present invention proposes a tunnel surrounding rock classification method and system.

[0007] The technical solution of the present invention is: a method for grading tunnel surrounding rock, comprising the following steps: S11: Intelligent sensing and data acquisition, using a variety of data acquisition equipment and sensor technologies to collect geological data in real time. This intelligent sensing and data acquisition, using a variety of data acquisition equipment and sensor technologies, enables real-time and comprehensive acquisition of geological data, providing rich and accurate raw data for subsequent analysis and classification, reducing classification deviations caused by missing or inaccurate data, and helping to more accurately understand the actual conditions of the tunnel surrounding rock. S12: Multi-source data fusion and modeling: Integrate the various collected data to construct a millimeter-level 3D geological BIM model. This high-precision 3D model can intuitively and clearly present the spatial distribution and structural characteristics of the tunnel surrounding rock, providing a more intuitive and detailed reference for the subsequent classification and support plan formulation, thereby improving the scientificity and accuracy of decision-making. S13: Hybrid model dynamic grading integrates deep learning into the Q system, using the improved Q system to modify parameters in real time and simulate multi-field coupling effects to output the surrounding rock grade. This method fully considers the dynamic changes of tunnel surrounding rock in complex environments, adjusts the grading results in real time, makes them more consistent with the actual surrounding rock conditions, improves the accuracy and reliability of the grading, and provides strong support for the timely implementation of appropriate support measures. S14: Real-time decision-making and support optimization. The digital twin platform drives the NSGA-II algorithm to generate the optimal support plan based on the surrounding rock grade. Utilizing advanced algorithms and digital twin technology, it comprehensively considers multiple factors, such as surrounding rock grade, project conditions, and economic costs, to quickly generate scientific, reasonable, and economical support plans, improving construction efficiency and safety while reducing project costs. S15: Online learning and dynamic feedback: real-time monitoring of data including deformation and stress, triggering online learning and dynamic optimization of model parameters. This allows the entire grading and support plan formulation process to be self-learning and self-improving. As the project progresses and data accumulates, the model can be continuously optimized, improving the accuracy and adaptability of grading and support plan formulation, and better addressing various complex situations that may arise in the tunnel surrounding rock. S16: Risk Warning and Emergency Response: This system uses AI algorithms to predict surrounding rock instability risks and automatically generates emergency plans. This system can detect potential surrounding rock instability risks in advance, issue early warnings, and automatically generate targeted emergency plans. This helps to take timely measures to avoid or reduce the occurrence of surrounding rock instability accidents, ensuring the safety of construction workers, the smooth progress of the project, and reducing accident losses.

[0008] Preferably, when collecting geological data in real time using a variety of data acquisition equipment and sensor technologies, the following are specifically included: A11: Drone + LiDAR Scanning: Drones equipped with LiDAR and cameras perform 3D scanning of tunnel faces, generating millimeter-level precision point cloud data and a real-time 3D geological model. Joint occurrence, crack density, and spatial distribution are automatically annotated. Drones equipped with LiDAR and cameras perform 3D scanning of tunnel faces, generating millimeter-level precision point cloud data and a real-time 3D geological model. Joint occurrence, crack density, and spatial distribution are automatically annotated, providing a highly accurate and intuitive 3D model for geological analysis, greatly improving the accuracy and visualization of geological information. A12: Intelligent core analysis uses the YOLO model to automatically analyze core images, identify fractures, fillings, and RQD values, and then identify and mark high-risk areas. This intelligent analysis of core images using the YOLO model can quickly identify fractures, fillings, and RQD values, and mark high-risk areas. This automates and efficiencies core analysis, helping to promptly identify potential geological risks. A13: IoT sensor network monitoring uses pre-buried fiber optic sensors, piezometers, and microseismic instruments to monitor ground stress, groundwater pressure, and surrounding rock deformation in real time, generating dynamic data streams. By building an IoT sensor network with pre-buried fiber optic sensors, piezometers, and microseismic instruments, key geological parameters such as ground stress, groundwater pressure, and surrounding rock deformation can be monitored in real time, generating dynamic data streams that provide reliable data support for continuous tracking and early warning of geological conditions.

[0009] As a preference, when integrating the various collected data to construct a millimeter-level three-dimensional geological BIM model, the following steps are specifically included: S21: 3D geological-BIM model construction. The LiDAR point cloud, core quality report, and dynamic data stream are unified into the same coordinate system through the GeoSLAM algorithm. The data splicing errors are eliminated based on the ICP algorithm and coordinate transformation matrix. The 3D geological model is then constructed using BIM software. Fault zones, water-rich areas, and rockburst risk areas are annotated. The spatial distribution of rock mass parameters is integrated. Finally, the consistency of geophysical and drilling data is cross-validated. The LiDAR point cloud, core quality report, and dynamic data stream are unified into the same coordinate system through the GeoSLAM algorithm. The data splicing errors are eliminated based on the ICP algorithm and coordinate transformation matrix. This effectively integrates data from multiple sources and different types, ensuring accurate spatial alignment of the data. This lays a solid foundation for building an accurate 3D geological model, avoids model distortion caused by inconsistent data coordinates or splicing errors, and improves the authenticity and reliability of the model. S22: Machine Learning Interpolation and Prediction: This method normalizes the rock mass parameters and geological characteristics at known exploration points to eliminate dimensional differences. A label spreading algorithm is then used to combine a small amount of annotated borehole data with a large amount of unlabeled geophysical and LiDAR-derived features. Random forests are used to mine nonlinear relationships, and kriging is used to capture spatial autocorrelations. This method trains a rock mass parameter prediction model. After inputting the spatial coordinates of the entire tunnel, the model outputs parameter distributions for unexplored areas, automatically annotates high-risk areas, and calculates kriging variance to assess prediction confidence, prompting additional exploration in areas with low reliability. The rock mass parameters and geological characteristics at known exploration points are normalized to eliminate dimensional differences, enabling different types of data to be analyzed and processed at the same scale. This fully utilizes the limited available exploration data and improves data utilization efficiency. Furthermore, model training combines a small amount of annotated borehole data with a large amount of unlabeled geophysical and LiDAR-derived features, fully leveraging the information value of the unlabeled data, reducing reliance on large amounts of labeled data and reducing the cost and workload of data collection and annotation.

[0010] As a preferred method, when deep learning is integrated into the Q system, the improved Q system is used to correct parameters in real time and simulate multi-field coupling effects to output the surrounding rock grade, specifically including: S31: Physical-data hybrid driven calculation, using the Q system formula to calculate the initial surrounding rock grade, combined with CNN image recognition to dynamically modify joint parameters, and cross-validate the Q value and RMR results; S32: Multi-field coupled numerical simulation, using FEM and DEM coupled models to simulate the interaction effects of ground stress, seepage and deformation, predict high-risk areas and trigger dynamic degradation.

[0011] As a preference, when calculating the initial surrounding rock grade through the Q system formula, dynamically correcting the joint parameters by combining CNN image recognition, and cross-validating the Q value and RMR results, the following steps are specifically included: S41: Initial Q value calculation. Based on the Q system formula, input parameters including RQD, Jn and Jr to calculate the initial Q value and output the preliminary determination result of the surrounding rock grade. The Q system formula is: ; in, is the rock quality index, is the number of joint groups, is the joint roughness, is the degree of joint alteration, is the joint water reduction coefficient, is the stress reduction factor; the Q system formula is used to calculate the initial surrounding rock grade. Based on a mature theoretical system, it can provide a solid theoretical basis and reliable initial results for surrounding rock grade assessment, ensuring that the assessment results are professional and accurate. S42: CNN image correction uses a pre-trained ResNet-50 model to analyze tunnel face images, identify joint density errors, and dynamically correct manually entered Jn values. This dynamic correction of joint parameters, combined with CNN image recognition technology, fully leverages the powerful capabilities of deep learning in image processing and data feature extraction. Through CNN's precise recognition and analysis of joint images, it can timely capture changes in joint characteristics, more accurately reflecting the impact of joints on surrounding rock stability and making surrounding rock grade assessments more relevant to actual project conditions. S43: Dynamic Correction and Cross-Validation: The Q value is recalculated based on the revised Jn value, and the rock mass grade is adjusted based on the updated Q value. This is then cross-validated with the RMR score, with a conservative approach used to ensure the rock mass grade. By cross-validating the Q value and RMR results, the results of the two different assessment systems are cross-checked. This multi-system validation approach can effectively identify potential deviations or limitations of a single assessment method, further improving the reliability and accuracy of rock mass grade assessment results and providing a more solid basis for subsequent engineering decisions.

[0012] As a preference, when using the FEM and DEM coupled model to simulate the interaction effect of geostress, seepage and deformation, predict high-risk areas and trigger dynamic degradation, the following steps are specifically included: S51: Model Construction: Based on the geological BIM model and excavation parameters, a coupled FEM-DEM model was established to simulate the release of geostress and changes in seepage paths after excavation. The coupled FEM (Finite Element Method) and DEM (Discrete Element Method) model simulates the interaction between geostress, seepage, and deformation, comprehensively considering the interactions between multiple complex geological factors. FEM is suitable for simulating the mechanical behavior of continuous media, while DEM excels at handling the movement and interactions of discrete particles or discontinuities. The coupled FEM model can more comprehensively and accurately simulate the mechanical response and deformation of surrounding rock under different geological conditions, providing a powerful tool for in-depth understanding of surrounding rock stability mechanisms. S52: Simulation and risk prediction: Through stress release simulation and seepage-stress coupling analysis, high-risk area predictions are output. Multi-field coupled numerical simulations can predict high-risk areas in the surrounding rock in advance. This predictive capability allows engineers to take targeted preventive measures for potential high-risk areas before actual construction begins, such as strengthening support and optimizing construction plans, thereby effectively reducing the probability of engineering accidents and ensuring the safety of personnel and property during construction and operation. S53: Dynamic grading adjustment triggers rock degradation based on simulation results. When simulation results indicate changes in rock mass conditions that could potentially cause engineering safety issues, a dynamic degradation mechanism is triggered. This mechanism allows for timely adjustment of rock mass grades based on actual rock mass conditions, providing real-time, accurate information for engineering decision-making. This allows engineering measures to quickly adapt to rock mass changes, ensuring the project remains safe and controllable, and improving the project's ability and efficiency to cope with complex geological conditions.

[0013] As a preferred method, when the NSGA-II algorithm is driven by the digital twin platform to generate the optimal support scheme according to the surrounding rock grade, the following steps are specifically included: S61: Data twin synchronization: Based on real-time monitoring data and updated geological models, the digital twin platform constructs a virtual tunnel model that simultaneously maps the actual state of the physical tunnel. This ensures a high degree of consistency between the virtual model and the physical tunnel, providing an accurate data foundation for subsequent decision-making, ensuring that decisions are based on the most realistic tunnel conditions and improving the scientific nature and accuracy of decision-making. S62: Dynamic Decision-Making: Utilizing real-time rendering with the Unity engine, the platform dynamically updates surrounding rock grades and generates a graded map, marking sections requiring support adjustments. This dynamic decision-making mechanism promptly identifies changes in tunnel surrounding rock and accurately locates areas requiring support adjustments, avoiding the issues of under- or over-support caused by information lags or inaccurate judgments in traditional methods. This improves the timeliness and pertinence of support decisions. S63: Support scheme generation, using the NSGA-II algorithm, with safety factor and cost control as the core objectives, optimizes the support parameter combination. Among them, the NSGA-II algorithm first randomly generates multiple groups of schemes, evaluates safety performance through finite element simulation, combines material and labor costs to calculate economic efficiency, uses Pareto frontier analysis to screen non-inferior solutions, and finally outputs the optimal solution; the NSGA-II algorithm is used to optimize the support parameter combination with safety factor and cost control as the core objectives. This algorithm randomly generates multiple groups of schemes, evaluates safety performance through finite element simulation, combines material and labor costs to calculate economic efficiency, and then uses Pareto frontier analysis to screen non-inferior solutions, and finally outputs the optimal solution. This multi-objective optimization method can reduce costs as much as possible while ensuring support safety, achieve a balance between safety and economy, avoid the limitations of traditional single-objective optimization methods, and improve the comprehensive benefits of support schemes; S64: AR Verification and Automated Execution: After the optimized solution is simulated and verified on the digital twin platform, it is projected onto the construction site using AR technology. This visual verification method allows construction personnel to intuitively see the effects of the support solution in the actual scenario, identify potential problems in advance, and make adjustments to ensure the feasibility and effectiveness of the solution, avoiding rework and delays caused by the solution not meeting the actual site conditions.

[0014] Preferably, when real-time monitoring of data including deformation and stress is performed, online learning is triggered, and model parameters are dynamically optimized, specifically including: S71: Data acquisition and preprocessing: Real-time monitoring data is cleaned and standardized to extract key features and generate a structured time series dataset. This data cleaning and standardization effectively removes noise and outliers, improving data accuracy and consistency. Extracting key features and generating a structured time series dataset makes the data easier to analyze and process, providing a high-quality data foundation for subsequent model training, thereby improving model performance and prediction accuracy. S72: Incremental learning and model updating, based on the Bayesian network to dynamically adjust parameter weights, using streaming gradient descent to locally update the model. Using the Bayesian network to dynamically adjust parameter weights allows for flexible adjustment of model parameters based on real-time data changes, making the model more adaptable and generalizable. This streaming gradient descent method enables rapid iteration and optimization of model parameters, avoiding the high computational cost and time delays associated with global retraining, and improving the efficiency of model updates. S73: Dynamic classification and feedback: Updated parameters are input into the classification model, surrounding rock grade is recalculated, and early warning instructions are pushed to the digital twin platform. Updated parameters are input into the classification model, surrounding rock grade is recalculated, and early warning instructions are pushed to the digital twin platform, providing timely and accurate basis for engineering decision-making. Pushing early warning instructions to the digital twin platform enables rapid transmission and visualization of early warning information, helping managers to quickly take countermeasures and reduce engineering risks. S74: Closed-loop verification and stability control. This system compares model predictions with measured data after construction, triggering secondary incremental learning when deviations exceed limits. This enables real-time evaluation and verification of model performance. Triggering secondary incremental learning when deviations exceed limits creates a closed-loop feedback mechanism, ensuring the model remains optimal and improving system stability and reliability. This closed-loop verification mechanism helps promptly detect and correct errors in model predictions, ensuring the accuracy and effectiveness of project monitoring.

[0015] As a preferred method, when predicting the risk of surrounding rock instability through AI algorithms and automatically generating emergency plans, the following are specifically included: S81: Data acquisition and preprocessing: real-time monitoring of data including surrounding rock deformation rate, stress change, microseismic frequency, and seepage pressure. This data is then decomposed and standardized using STL to eliminate noise. This data is then fed into the LSTM model. This ensures data quality and lays a solid foundation for subsequent accurate predictions. This processing method effectively improves data accuracy and reliability, helping the model capture more realistic patterns of surrounding rock state changes. S82: Risk prediction and early warning. Using the LSTM model to analyze deformation rate trends, predict collapse probability, and trigger early warnings when thresholds are exceeded. Leveraging the LSTM model's advantages in processing time series data, we analyze deformation rate trends and predict collapse probability. The LSTM model can capture long-term dependencies in data, enabling more accurate predictions of surrounding rock instability risks and providing strong support for timely early warnings and countermeasures. S83: Knowledge graph retrieval, based on the Neo4j graph database, matches current risk characteristics with historical cases to quickly retrieve the optimal emergency plan. This approach fully utilizes the value of historical data and, by matching similar cases, provides reference solutions for current risks, improving the efficiency and accuracy of emergency response. S84: Plan optimization: Integrate existing resource data, adjust parameters based on existing resource constraints, and generate construction instructions after verification by the rule engine. This step ensures that the emergency plan is not only scientific and reasonable, but also meets the actual resource conditions, improving the feasibility and operability of the plan. By optimizing the plan, resources can be more effectively utilized and emergency response costs can be reduced. S85: Feedback Verification: Real-time data collection and calculation of construction results are performed to verify the effectiveness of the plan. This closed-loop feedback mechanism helps to promptly identify problems during plan implementation and provides a basis for dynamic adjustment and optimization of the plan. By continuously verifying and improving the plan, the effectiveness of emergency response can be improved and construction safety can be ensured.

[0016] The classification system for tunnel surrounding rock includes: Intelligent sensing and data acquisition module for real-time collection of high-precision geological data through drone LiDAR scanning, intelligent core analysis, and IoT sensor networks; A multi-source data fusion and modeling module integrates LiDAR point clouds, borehole data, and geophysical exploration results, uses the GeoSLAM algorithm to eliminate stitching errors, constructs millimeter-level 3D geological-BIM models, and uses random forest and Kriging interpolation to predict rock mass parameters in unexplored areas and mark high-risk areas. A hybrid model dynamic grading module is used to calculate the initial surrounding rock grade based on the Q system formula, dynamically modify joint parameters in combination with CNN image recognition, and simulate the interaction effect of ground stress, seepage and deformation through the FEM and DEM coupling model to achieve dynamic grading driven by multi-field coupling; A real-time decision-making and support optimization module, which uses the digital twin platform to update the surrounding rock grade in real time, using the NSGA-II algorithm to balance safety and cost, and generate the optimal support parameter combination; An online learning and dynamic feedback module is used to trigger Bayesian network parameter updates based on real-time monitoring data, dynamically optimize model weights through incremental learning, and verify model reliability in a closed-loop manner using post-construction measured data. The risk warning and emergency response module is used to predict landslide risks using the LSTM model, generate emergency plans by matching historical cases with the Neo4j knowledge graph, and provide synchronous feedback for construction effect verification.

[0017] Beneficial effects of the present invention: 1. Compared with the existing geological data collection method that combines manual coring with low-resolution geophysical exploration, which has the disadvantages of low efficiency, large subjective errors, and difficulty in covering complex geological structures, this invention uses drone LiDAR scanning, intelligent core image analysis (YOLO model) and a high-density IoT sensor network to achieve real-time dynamic collection of full-section geological information. Combined with a multi-source data fusion algorithm to eliminate human intervention errors, it significantly improves the automation level of data collection and the accuracy of three-dimensional spatial representation, providing a high-resolution, fully informationized data base for surrounding rock classification. 2. Compared with the existing technology that relies on static Q system formulas and single physical field numerical simulations for surrounding rock classification, which has the defects of model rigidity and inability to reflect the dynamic effects of multi-field coupling and real-time changes in geological parameters, this invention innovatively embeds deep learning (CNN image recognition) into the traditional Q system. By dynamically correcting joint parameters (such as Jn values) and integrating FEM / DEM multi-field coupling simulation (in-situ stress-seepage-deformation interaction analysis), a physical-data hybrid driven model is constructed to achieve real-time iteration of classification parameters and advanced prediction of risk evolution. This overcomes the limitations of traditional methods in responding to complex geological dynamics and significantly improves the timeliness and engineering applicability of classification results. 3. Compared with the existing technology that uses empirical support design and manual adjustment schemes, which have the disadvantages of poor economy, insufficient safety redundancy, and difficulty in dynamically adapting to surrounding rock changes, this invention relies on the digital twin platform to integrate real-time monitoring data, drive the NSGA-II multi-objective optimization algorithm to globally search for support parameter combinations, and combine it with AR / VR interactive verification technology to achieve multi-dimensional collaborative optimization of safety and economy of support schemes. It is precisely executed through the robotic construction queue, forming a closed loop of "dynamic perception-intelligent decision-making-automatic execution", which completely changes the empirical and lagging mode of traditional support design; 4. Compared with the existing technology that relies on manual experience for risk warning and emergency response, which has the defects of delayed warning, low disposal efficiency, and dependence on human intervention, the present invention predicts the risk of surrounding rock instability through the LSTM time series model, combines the Neo4j knowledge graph to intelligently match the historical case library to generate emergency plans, and relies on support robots to automatically execute reinforcement processes (such as precise layout of advanced pipe racks), building an "AI prediction-knowledge drive-machine execution" full-link emergency response system, realizing unmanned and high-precision operation of the entire process from risk identification to disposal, greatly reducing the risk of personnel exposure and improving the active prevention and control capabilities of tunnel construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Shown is a schematic flow chart of the tunnel surrounding rock classification method of the present invention; Figure 2 Shown is a schematic diagram of the structure of the grading system of the tunnel surrounding rock of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described below with reference to the accompanying drawings and examples.

[0020] See also Figure 1-2 The present invention provides an embodiment: a method for grading tunnel surrounding rock, comprising the following steps: Step 1: Intelligent Perception and Data Collection Real-time geological data is collected through a variety of data acquisition devices and sensor technologies. The methods used include: Drone + LiDAR scanning: Drones equipped with LiDAR and cameras perform 3D scanning of the tunnel face, generating millimeter-level precision point cloud data and a real-time 3D geological model, automatically annotating joint occurrence, crack density, and spatial distribution. Intelligent core analysis: Using the YOLO model, core images are automatically analyzed to identify cracks, fillings, and RQD values, obtaining and marking high-risk areas. IoT sensor network monitoring: Pre-buried fiber optic sensors, piezometers, and microseismometers monitor ground stress, groundwater pressure, and surrounding rock deformation in real time, generating dynamic data streams. Step 2: Multi-source data fusion and modeling Integrate the various collected data to build a millimeter-level 3D geological BIM model, including: 3D geological-BIM model construction: The GeoSLAM algorithm is used to unify LiDAR point clouds, core quality reports, and dynamic data streams into the same coordinate system. Data splicing errors are eliminated based on the ICP algorithm and coordinate transformation matrix. BIM software is then used to construct a 3D geological model, and fault zones, water-rich areas, and rockburst risk areas are annotated. The spatial distribution of rock mass parameters is integrated, and finally, the consistency of geophysical and borehole data is cross-validated. Machine learning interpolation and prediction are used to normalize the rock mass parameters and geological characteristics of known exploration points to eliminate dimensional differences. The Label Spreading algorithm is used to combine a small amount of borehole annotation data with a large amount of unlabeled geophysical and LiDAR-derived features. Random forest is used to mine nonlinear relationships, and Kriging is used to capture spatial autocorrelation to train the rock mass parameter prediction model. After inputting the spatial coordinates of the entire tunnel, the model outputs the parameter distribution of the unexplored area, automatically annotates high-risk areas, and calculates the Kriging variance to assess the prediction confidence, prompting additional exploration in low-reliability areas. Step 3: Dynamic classification of hybrid models Integrate deep learning into the Q system, use the improved Q system to modify parameters in real time and simulate multi-field coupling effects to output the surrounding rock grade, including: The initial surrounding rock grade is calculated using the Q system formula, and the joint parameters are dynamically corrected in combination with CNN image recognition. The Q value and RMR results are cross-validated. Specifically, based on the Q system formula, the parameters including RQD, Jn, and Jr are input to calculate the initial Q value, and the preliminary determination result of the surrounding rock grade is output. The Q system formula is: ;in, is the rock quality index, is the number of joint groups, is the joint roughness, is the degree of joint alteration, is the joint water reduction coefficient, is the stress reduction factor; the pre-trained ResNet-50 model is then used to analyze the tunnel face image, identify joint density errors, and dynamically correct the manually entered Jn value. Finally, the Q value is recalculated based on the corrected Jn value, and the surrounding rock grade is adjusted based on the updated Q value. This is then cross-validated with the RMR score, and a conservative result is taken to ensure the surrounding rock grade. The FEM and DEM coupling model is used to simulate the interaction between geostress, seepage, and deformation, predict high-risk areas, and trigger dynamic degradation. Specifically, the following steps are performed: First, based on the geological BIM model and excavation parameters, an FEM-DEM coupling model is established to simulate the release of geostress and changes in seepage paths after excavation. Then, through stress release simulation and seepage-stress coupling analysis, a prediction of high-risk areas is output. Finally, surrounding rock degradation is triggered based on the simulation results. Step 4: Real-time decision-making and support optimization The digital twin platform drives the NSGA-II algorithm to generate the optimal support solution based on the surrounding rock grade, including: Data twin synchronization: Based on real-time monitoring data and geological model update results, the digital twin platform constructs a virtual tunnel model, synchronously mapping the actual status of the physical tunnel; dynamic decision-making: Through real-time rendering in the Unity engine, the platform dynamically updates the surrounding rock grade and generates a graded map, marking the sections that require support adjustment; support scheme generation: The NSGA-II algorithm is used to optimize the support parameter combination with safety factor and cost control as the core goals. The NSGA-II algorithm first randomly generates multiple sets of schemes, evaluates safety performance through finite element simulation, calculates economic efficiency based on material and labor costs, and uses Pareto frontier analysis to screen non-inferior solutions, ultimately outputting the optimal solution; AR verification and automated execution: After the optimization scheme is simulated and verified by the digital twin platform, it is projected to the construction site using AR technology; Step 5: Online learning and dynamic feedback Real-time monitoring of data including deformation and stress triggers online learning and dynamic optimization of model parameters, including: Data collection and preprocessing: Real-time monitoring data is cleaned and standardized to extract key features and generate a structured time series data set; incremental learning and model updating: Dynamically adjust parameter weights based on a Bayesian network and use streaming gradient descent to locally update the model; dynamic classification and feedback: Input updated parameters into the classification model, recalculate the surrounding rock grade, and push early warning instructions to the digital twin platform; closed-loop verification and stability control: Compare model predictions with measured data after construction, and trigger secondary incremental learning when the deviation exceeds the limit; Step 6: Risk Warning and Emergency Response AI algorithms are used to predict the risk of surrounding rock instability and automatically generate emergency plans, including: Data collection and preprocessing: real-time monitoring of data including surrounding rock deformation rate, stress change, microseismic frequency and seepage pressure value, and standardization after STL decomposition to eliminate noise, and input into the LSTM model; risk prediction and early warning: using the LSTM model to analyze deformation rate trends, predict collapse probability, and trigger early warning when the threshold is exceeded; knowledge graph retrieval: matching current risk characteristics with historical cases based on the Neo4j graph database to retrieve the optimal emergency plan; plan optimization: integrating existing resource data, and then adjusting parameters based on existing resource data constraints, and generating construction instructions after verification by the rule engine; feedback verification: real-time collection of data during construction, real-time calculation of construction effects, and verification of plan effectiveness.

[0021] The classification system for tunnel surrounding rock includes: Intelligent sensing and data acquisition module for real-time collection of high-precision geological data through drone LiDAR scanning, intelligent core analysis, and IoT sensor networks; A multi-source data fusion and modeling module integrates LiDAR point clouds, borehole data, and geophysical exploration results, uses the GeoSLAM algorithm to eliminate stitching errors, constructs millimeter-level 3D geological-BIM models, and uses random forest and Kriging interpolation to predict rock mass parameters in unexplored areas and mark high-risk areas. A hybrid model dynamic grading module is used to calculate the initial surrounding rock grade based on the Q system formula, dynamically modify joint parameters in combination with CNN image recognition, and simulate the interaction effect of ground stress, seepage and deformation through the FEM and DEM coupling model to achieve dynamic grading driven by multi-field coupling; A real-time decision-making and support optimization module, which uses the digital twin platform to update the surrounding rock grade in real time, using the NSGA-II algorithm to balance safety and cost, and generate the optimal support parameter combination; An online learning and dynamic feedback module is used to trigger Bayesian network parameter updates based on real-time monitoring data, dynamically optimize model weights through incremental learning, and verify model reliability in a closed-loop manner using post-construction measured data. The risk warning and emergency response module is used to predict landslide risks using the LSTM model, generate emergency plans by matching historical cases with the Neo4j knowledge graph, and provide synchronous feedback for construction effect verification.

[0022] Example 1: Dynamic classification and support optimization of high-stress soft rock tunnels Application scenario: A deep soft rock tunnel passes through a high-stress area. After excavation, the large deformation rate of the surrounding rock exceeds the limit. The traditional grading method causes the support plan to lag due to parameter rigidity, causing the risk of local collapse.

[0023] Implementation steps: Intelligent perception and data collection: The drone LiDAR scanned the tunnel face, generating a millimeter-scale point cloud model to identify the joint density (Jn=4) and spatial distribution. IoT sensors monitored the surrounding rock deformation rate (5→8 mm / d) and ground stress (peak 22 MPa) in real time. The intelligent core analyzer used the YOLO model to identify the core RQD=65%, marking it as a high-risk area.

[0024] Multi-source data fusion and modeling: Integrate LiDAR, sensor data, and drilling reports to construct a 3D geological-BIM model, annotating high-stress areas and potential plasticity zones; and use random forest interpolation to predict rock mass parameters in unexcavated sections (Kv=0.35→Class V surrounding rock).

[0025] Dynamic classification of the hybrid model: the Q system was initially calculated as Q=10 (Level III), and the CNN image was corrected as Jn=4→Q=6 (Level IV). The FEM / DEM coupling simulation showed that the release of ground stress led to the expansion of cracks, triggering a downgrade to Level V.

[0026] Real-time decision-making and support optimization: The digital twin platform drives the NSGA-II algorithm to optimize support parameters: anchor bolt spacing from 0.8m to 0.6m, and steel frame spacing from 1.2m to 0.8m. After the AR glasses verify the solution, the support robot performs double-layer steel frame installation and radial grouting.

[0027] Online learning and feedback: After grouting, the deformation rate dropped to 2 mm / d, the Bayesian network updated the ground stress weight (0.2→0.35), and the model was iteratively optimized.

[0028] Technical effect: The response time for surrounding rock classification has been shortened from the traditional 24 hours to 30 minutes; the support scheme adjustment cycle has been compressed from 3 manual days to 2 hours, reducing the risk of landslides by 90%; material waste has been reduced by 25%, and the construction period has been shortened by 15%.

[0029] Example 2: Risk Warning and Emergency Response for Tunnels in Water-Rich Fault Zones Application scenario: A tunnel passes through a water-rich fault zone. After excavation, sudden seepage causes the surrounding rock to soften. Traditional methods cannot predict the risk of landslides in a timely manner.

[0030] Implementation steps: Intelligent perception and data collection: The fiber optic sensor monitored a sudden increase in seepage pressure (0.3→0.8MPa), and the microseismometer captured the crack expansion signal; the LiDAR scan showed that the fault zone displacement was 10mm (exceeding the limit by 8mm).

[0031] Risk warning and emergency response: The LSTM model analyzed deformation rate time series data and predicted an 85% probability of collapse within three hours. The Neo4j knowledge graph matched historical cases and generated an emergency plan: advanced pipe shed (Φ42 steel pipes, 0.4m spacing) + radial grouting (cement slurry water-cement ratio 0.8:1). The rule engine verified the inventory materials (200 steel pipes, 50 tons of cement) and issued instructions to the support robot.

[0032] Automated execution and verification: The support robot drilled holes along the preset path (depth 4m, error ±2cm) and simultaneously grouting (pressure 0.6MPa); after grouting, the seepage pressure value dropped to 0.4MPa, the deformation rate stabilized at 1mm / h, and the system marked the risk was eliminated.

[0033] Dynamic grading and model updating: The digital twin platform updated the surrounding rock grade to Grade V, triggering the NSGA-II algorithm to optimize the support parameters of subsequent sections (anchor density increased by 20%). The incremental learning module also updated the seepage-stress coupling weight (0.15→0.25), improving subsequent prediction accuracy.

[0034] Technical effect: The advance time for landslide warning has been extended from the traditional 0.5 hours to 3 hours; the entire emergency response process is automated, and the handling time is shortened from 8 hours manual work to 1.5 hours; the utilization rate of grouting materials has increased by 30%, and the number of times people enter high-risk areas has been reduced by 95%.

[0035] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the scope of knowledge of those skilled in the art without departing from the spirit of the present invention.

Claims

1. Tunnel surrounding rock classification method; characterized by: The following steps are involved: S11: Intelligent sensing and data acquisition, real-time geological data collection through a variety of data acquisition devices and sensor technologies; S12: Multi-source data fusion and modeling: integrating various collected data to construct millimeter-level 3D geological BIM models; S13: Hybrid model dynamic classification, which integrates deep learning into the Q system, uses the improved Q system to modify parameters in real time and simulate multi-field coupling effects to output the surrounding rock grade; S14: Real-time decision-making and support optimization, using the digital twin platform to drive the NSGA-II algorithm and generate the optimal support plan based on the surrounding rock grade; S15: Online learning and dynamic feedback: real-time monitoring of data including deformation and stress, triggering online learning and dynamically optimizing model parameters; S16: Risk warning and emergency response: predict the risk of surrounding rock instability through AI algorithms and automatically generate emergency plans.

2. The method for grading tunnel surrounding rock according to claim 1, characterized in that: When collecting geological data in real time through a variety of data acquisition equipment and sensor technologies, specifically: A11: Drone + LiDAR scanning: A drone equipped with LiDAR and a camera performs 3D scanning of the tunnel face, generating point cloud data with millimeter-level accuracy, creating a real-time 3D geological model, and automatically annotating joint occurrence, crack density, and spatial distribution. A12: Intelligent core analysis: Uses the YOLO model to automatically analyze core images, identify fractures, fillings, and RQD values, and mark high-risk areas. A13: IoT sensor network monitoring uses pre-buried fiber optic sensors, piezometers, and microseismometers to monitor ground stress, groundwater pressure, and surrounding rock deformation in real time, generating dynamic data streams.

3. The method for grading tunnel surrounding rock according to claim 2, characterized in that: When integrating the various collected data to construct a millimeter-level 3D geological BIM model, the following are specifically involved: S21: 3D geological-BIM model construction. Using the GeoSLAM algorithm, LiDAR point clouds, core quality reports, and dynamic data streams are unified into the same coordinate system. Data splicing errors are eliminated using the ICP algorithm and coordinate transformation matrix. BIM software is then used to construct a 3D geological model, annotating fault zones, water-rich areas, and rockburst risk areas. The spatial distribution of rock mass parameters is also integrated. Finally, the consistency of geophysical and drilling data is cross-validated. S22: Machine learning interpolation and prediction: normalizes the rock mass parameters and geological characteristics of known exploration points to eliminate dimensional differences. A Label Spreading algorithm is used to combine a small amount of borehole annotation data with a large amount of unlabeled geophysical and LiDAR-derived features. Random forests are used to mine nonlinear relationships, and kriging is used to capture spatial autocorrelations. The rock mass parameter prediction model is trained. After inputting the spatial coordinates of the entire tunnel, the model outputs the parameter distribution of unexplored areas, automatically annotates high-risk areas, and calculates the kriging variance to assess prediction confidence, prompting additional exploration in low-reliability areas.

4. The method for grading tunnel surrounding rock according to claim 3, characterized in that: When integrating deep learning into the Q system, the improved Q system is used to correct parameters in real time and simulate multi-field coupling effects to output the surrounding rock grade, specifically including: S31: Physical-data hybrid driven calculation, using the Q system formula to calculate the initial surrounding rock grade, combined with CNN image recognition to dynamically modify joint parameters, and cross-validate the Q value and RMR results; S32: Multi-field coupled numerical simulation, using FEM and DEM coupled models to simulate the interaction effects of ground stress, seepage and deformation, predict high-risk areas and trigger dynamic degradation.

5. The method for grading tunnel surrounding rock according to claim 4, characterized in that: When calculating the initial surrounding rock grade through the Q system formula, dynamically correcting the joint parameters in combination with CNN image recognition, and cross-validating the Q value and RMR results, the following are specifically included: S41: Initial Q value calculation: Based on the Q system formula, input parameters including RQD, Jn and Jr to calculate the initial Q value, and output the preliminary determination result of the surrounding rock grade; S42: CNN image correction, which uses a pre-trained ResNet-50 model to analyze tunnel face images, identify joint density errors, and dynamically correct manually entered Jn values; S43: Dynamic correction and cross-validation: recalculate the Q value according to the corrected Jn value, adjust the surrounding rock grade according to the updated Q value, and cross-validate with the RMR score to take a conservative result to ensure the surrounding rock grade.

6. The method for grading tunnel surrounding rock according to claim 5, characterized in that: When using the FEM and DEM coupled model to simulate the interaction between ground stress, seepage and deformation, predict high-risk areas and trigger dynamic degradation, the following are specifically included: S51: Model construction: Based on the geological BIM model and excavation parameters, a FEM-DEM coupling model is established to simulate the release of ground stress and the change of seepage path after excavation; S52: Simulation and risk prediction, outputting high-risk area prediction through stress release simulation and seepage-stress coupling analysis; S53: Dynamic classification adjustment, triggering surrounding rock degradation based on simulation results.

7. The method for grading tunnel surrounding rock according to claim 6, characterized in that: When the NSGA-II algorithm is driven by the digital twin platform to generate the optimal support scheme based on the surrounding rock grade, the following steps are specifically included: S61: Data twin synchronization: Based on real-time monitoring data and geological model update results, the digital twin platform builds a virtual tunnel model that simultaneously maps the actual status of the physical tunnel; S62: Dynamic decision-making, through real-time rendering in the Unity engine, the platform dynamically updates the surrounding rock grade and generates a graded map, marking the sections that require support adjustment; S63: Support scheme generation uses the NSGA-II algorithm, with safety factor and cost control as the core objectives, to optimize support parameter combinations. The NSGA-II algorithm first randomly generates multiple sets of schemes, evaluates safety performance through finite element simulation, calculates economic efficiency based on material and labor costs, and uses Pareto frontier analysis to screen non-inferior solutions, ultimately outputting the optimal solution. S64: AR verification and automated execution: After the optimization plan is simulated and verified on the digital twin platform, it is projected to the construction site through AR technology.

8. The method for grading tunnel surrounding rock according to claim 7, characterized in that: When real-time monitoring of data including deformation and stress is carried out, online learning is triggered, and model parameters are dynamically optimized, specifically including: S71: Data collection and preprocessing: real-time monitoring data is cleaned and standardized to extract key features and generate a structured time series data set; S72: Incremental learning and model updating, dynamically adjusting parameter weights based on Bayesian networks, and using streaming gradient descent to locally update the model; S73: Dynamic classification and feedback: input the updated parameters into the classification model, recalculate the surrounding rock grade, and push warning instructions to the digital twin platform; S74: Closed-loop verification and stability control, comparing model predictions with measured data after construction, triggering secondary incremental learning when the deviation exceeds the limit.

9. The method for grading tunnel surrounding rock according to claim 8, characterized in that: When predicting the risk of surrounding rock instability through AI algorithms and automatically generating emergency plans, the following are specifically included: S81: Data acquisition and preprocessing: real-time monitoring of data including surrounding rock deformation rate, stress change, microseismic frequency, and seepage pressure. After decomposition and noise elimination by STL, the data is standardized and input into the LSTM model. S82: Risk prediction and early warning, using the LSTM model to analyze deformation rate trends, predict collapse probability, and trigger an early warning when the threshold is exceeded; S83: Knowledge graph retrieval, matching current risk characteristics with historical cases based on the Neo4j graph database to retrieve the optimal emergency plan; S84: Scheme optimization: integrating existing resource data, adjusting parameters based on existing resource data constraints, and generating construction instructions after verification by the rule engine; S85: Feedback verification, real-time collection of construction data, real-time calculation of construction effects, and verification of the effectiveness of the plan.

10. A tunnel surrounding rock grading system, used in the tunnel surrounding rock grading method according to any one of claims 1 to 9, characterized in that: include: Intelligent sensing and data acquisition module for real-time collection of high-precision geological data through drone LiDAR scanning, intelligent core analysis, and IoT sensor networks; A multi-source data fusion and modeling module integrates LiDAR point clouds, borehole data, and geophysical exploration results, uses the GeoSLAM algorithm to eliminate stitching errors, constructs millimeter-level 3D geological-BIM models, and uses random forest and Kriging interpolation to predict rock mass parameters in unexplored areas and mark high-risk areas. A hybrid model dynamic grading module is used to calculate the initial surrounding rock grade based on the Q system formula, dynamically modify joint parameters in combination with CNN image recognition, and simulate the interaction effect of ground stress, seepage and deformation through the FEM and DEM coupling model to achieve dynamic grading driven by multi-field coupling; A real-time decision-making and support optimization module, which uses the digital twin platform to update the surrounding rock grade in real time, using the NSGA-II algorithm to balance safety and cost, and generate the optimal support parameter combination; An online learning and dynamic feedback module is used to trigger Bayesian network parameter updates based on real-time monitoring data, dynamically optimize model weights through incremental learning, and verify model reliability in a closed-loop manner using post-construction measured data. The risk warning and emergency response module is used to predict landslide risks using the LSTM model, generate emergency plans by matching historical cases with the Neo4j knowledge graph, and provide synchronous feedback for construction effect verification.

Images