A wide fracture zone tunnel construction risk regulation method based on real-time prediction

Abstract

This invention discloses a risk control method for tunnel construction in wide fault zones based on real-time prediction, relating to the field of underground engineering. The invention first collects multi-source data such as water pressure and rock mass integrity coefficient in real time, and then fuses them into a comprehensive data vector after standardization processing. Next, it uses a GBDT-LSTM hybrid model to predict construction risks in real time, combined with an accuracy rate used to measure the accuracy of early warnings. Pr Recall rate, which measures the comprehensiveness of early warnings Re And comprehensively reflecting the early warning performance of the model F A warning threshold is determined by a risk value; if the risk value exceeds the threshold, a warning is triggered. Subsequently, based on real-time geological conditions and risk prediction results, an optimal combination of construction parameters is dynamically generated through a multi-objective optimization algorithm. Finally, the parameters are issued to guide construction, and new data is continuously collected to update the model. This invention integrates risk prediction and construction parameter control, overcoming the blindness and lag of traditional construction methods, and improving the safety and efficiency of tunnel construction.

Description

Technical Field

[0001] This invention relates to the field of underground engineering, specifically to a method for risk control during the construction of tunnels in wide fault zones based on real-time prediction. Background Technology

[0002] With the rapid development of infrastructure projects such as water conservancy, tunnel engineering inevitably needs to traverse complex and adverse geological sections. Among them, wide fault zones are a typical and extremely hazardous type of adverse geological body. These zones typically consist of faults, breccia, and fractured rocks, and are usually over 100 meters wide. The rock mass is extremely fragmented, with low strength and poor self-stability. Furthermore, fault zones often serve as conduits for groundwater accumulation and migration, leading to significant risks during construction, such as high water pressure, water inrush, and mudslides. In addition, the geostress state around fault zones is often complex and concentrated, further exacerbating the possibility of surrounding rock instability.

[0003] In traditional tunnel construction methods, dealing with wide fault zones relies primarily on advanced geological forecasting and engineers' experience. A static design principle of "drainage-oriented" or "plugging-oriented" is typically adopted, with pre-defined support parameters such as support type, grouting pressure and grout mix ratio, and excavation advance. However, this method has the following significant drawbacks: Insufficient risk foresight: Traditional geological forecasting methods, such as ground-penetrating radar and TSP, suffer from multiple interpretations and time lags in assessing geological conditions ahead, making it difficult to accurately and in real-time quantify construction risks. When the actual geological conditions revealed differ from the forecasts, the pre-set construction plan may no longer be applicable, easily leading to safety accidents.

[0004] Fixed parameters and lack of adaptability: Pre-set support and excavation parameters are static and cannot be dynamically adjusted according to real-time changes in water pressure, rock integrity, and stress state. This can lead to wasted materials and time when the risk is low, while inadequate measures can fail to effectively control the disaster when the risk increases sharply.

[0005] Decision-making relies heavily on experience, and its scientific rigor needs improvement: Key decisions during construction depend heavily on engineers' personal experience, lacking unified and quantifiable scientific basis. When facing complex and ever-changing wide fault zones, the limitations and instability of human experience-based judgment may lead to decision-making errors.

[0006] Isolated analysis of various factors lacks synergistic optimization: factors such as water pressure, rock mass integrity, and stress state are interrelated and jointly affect the stability of the surrounding rock. Traditional methods are insufficient for comprehensively analyzing these multi-source and heterogeneous key indicators and for synergistically optimizing multiple construction stages such as support, grouting, and excavation.

[0007] Chinese invention patent CN120314934A discloses an advanced geological prediction method and system, employing ground-penetrating radar (GPR), a traditional geological prediction method, to achieve advanced geological prediction. Chinese invention patent CN119375937A discloses a TSP (Technical Separation Point) advanced geological prediction system. However, existing technologies suffer from multiple interpretations and delays in assessing geological conditions ahead, making it difficult to accurately and in real-time quantify construction risks. When the actual revealed geological conditions differ from the prediction, the pre-set construction plan may no longer be applicable, easily leading to safety accidents.

[0008] Therefore, there is an urgent need for a technology that can predict construction risks in real time and accurately, and can intelligently and dynamically adjust the construction plan based on the prediction results, in order to overcome the blindness and lag in the existing technology and achieve safe, efficient and refined construction of tunnels when crossing wide fault zones. Summary of the Invention

[0009] The purpose of this invention is to provide a method for risk control in tunnel construction in wide fault zones based on real-time prediction. This method integrates real-time risk prediction with dynamic intelligent control of construction parameters during tunnel construction, overcoming the blindness and lag of traditional construction methods and improving construction safety and efficiency.

[0010] To achieve the above objectives, the technical solution of this invention is as follows: A method for risk control in tunnel construction in wide fault zones based on real-time prediction includes the following steps: S1: Data Acquisition and Fusion S11: Real-time acquisition and fusion of multi-source data during tunnel construction, including upstream water pressure, rock mass integrity coefficient, surrounding rock stress state, tunneling parameters, and support stress. Data originates from a sensor network deployed within the tunnel, such as stress sensors, water pressure gauges, displacement monitoring points, and advanced geological prediction equipment.

[0011] S12: Data Standardization Processing: Standardize the multi-source data collected in step S11 to eliminate the influence of different data sources' dimensions. The standardized data value is equal to the original data value minus the minimum historical data value, divided by the difference between the maximum and minimum historical data values. The standardization formula is as follows: (1) in, x The original data value; and These are the minimum and maximum values ​​of historical data. The standardized data values ​​(range [0,1]).

[0012] S13: Merge the data standardized in step S12 into a comprehensive data vector. This is used for subsequent risk prediction. (Comprehensive data vector) The vector is generated by sequentially analyzing six parameters: water pressure, rock mass integrity coefficient, surrounding rock stress state, tunneling speed, support stress, and displacement monitoring values, all monitored in real time at the front. Defined as: (2) in: This is for real-time monitoring of water pressure at the front line, in MPa. The rock mass integrity coefficient is expressed as a percentage. The stress state of the surrounding rock is expressed in MPa. The tunneling speed is expressed in m / h. Support stress, unit MPa; These are displacement monitoring values, in mm.

[0013] S2: Machine Learning Risk Prediction This invention employs a hybrid prediction model architecture for machine learning risk prediction: it uses a machine learning model to predict construction risks in real time and outputs a comprehensive risk level or risk probability value. Based on a hybrid model combining structured feature extraction and temporal feature extraction, it utilizes a hybrid model of Gradient Boosting Decision Tree (GBDT) and Long Short-Term Memory (LSTM) networks, fully leveraging the advantages of both. The GBDT component extracts structured features, and the LSTM component extracts temporal features. The two types of features are concatenated and input into a fully connected layer to calculate the risk probability.

[0014] GBDT component: Processes structured tabular data, sets tree depth to 6-10 levels and learning rate to 0.05-0.1, can efficiently capture the complex interaction relationships and nonlinear mappings between multiple features such as water pressure, rock mass integrity, and stress, and automatically sorts and assesses the importance of key risk factors.

[0015] LSTM Component: As a variant of Recurrent Neural Network (RNN), it sets the hidden layer dimension to 64-128 and the time step to 5-10. It is used to process sequential data, memorize and utilize long-term dependencies in historical data, such as data from previous construction cycles, and identify trends in risk evolution, such as accelerated displacement or slow stress accumulation. Fusion Method: The high-order feature vector extracted by GBDT is concatenated with the temporal feature vector extracted by LSTM to form a fused feature vector. This fused feature vector is then input into a fully connected neural network layer with 32 neurons and the Sigmoid activation function, outputting the final risk probability value. In the fields of machine learning and neural networks, the Sigmoid function is one of the most basic and widely used activation functions, a common technique, and will not be elaborated upon here.

[0016] Fusion method: The high-order features extracted by GBDT are concatenated with the temporal features extracted by LSTM, and then input together into a fully connected neural network layer for final risk probability calculation. This architecture considers both snapshots of the current state parameters and the evolution of historical states, making the prediction more comprehensive and accurate.

[0017] Machine learning risk prediction specifically includes the following steps: S21: Model Training and Validation Model training and validation includes dataset construction and partitioning, as well as the partitioning of training, validation, and test sets.

[0018] Dataset Construction and Partitioning: The dataset is derived from completed tunnel engineering cases that traverse similar wide fault zones.

[0019] Data Samples: Each sample consists of two parts: a. Feature Vector: a standardized composite data vector at a specific moment or construction progress segment. Wherein, the composite data vector at time t Recorded as b. Risk Label: Whether a definable engineering risk event, such as a large water inrush, significant deformation, or support failure, has occurred since this point. The label is uniformly assigned by experts after the event based on monitoring records, construction logs, and geological reports, such as 0 - no risk, 1 - risky.

[0020] The division of training, validation, and test sets: Because the construction process is highly time-dependent, the dataset must be divided strictly in chronological order to ensure the objectivity and generalization ability of the model evaluation.

[0021] Training set: Construction data from the first 70% to 80% of the time series. This data is used to teach the model to learn the mapping pattern from multi-source data to risk outcomes.

[0022] The loss function uses binary cross-entropy: (3) in, This represents the number of training samples; For the first i The true risk label of each sample (0 indicates no risk, 1 indicates risk); The risk probability value predicted by the model. Loss function. Q Used to quantify the risk probability value predicted by the model. Compared with the true risk label of the sample The deviation provides a target direction for model parameter optimization, ultimately enabling the model to accurately learn the mapping pattern between multi-source construction data and risk outcomes. During model training, QThe values ​​need to gradually decrease and stabilize, and convergence is achieved when the values ​​no longer change significantly with the number of iterations. In this case, the GBDT-LSTM hybrid model parameters are optimal.

[0023] Validation set: Construction data from the first 10%–15% of the time series immediately following the training set. This dataset is used during training but is not used for parameter updates.

[0024] Test set: Construction data from the last 10% to 15% of the time series. This data is completely unavailable until the final model training and tuning are completed. It is used to simulate the final performance evaluation of the model when facing new and unknown construction sections in real-world applications, providing the most reliable estimates of metrics such as accuracy and recall.

[0025] S22: Risk Prediction and Verification Risk prediction and verification includes risk prediction calculation and early warning thresholds. The determination and triggering of early warnings and verification.

[0026] Regarding the risk prediction calculation: the trained GBDT-LSTM hybrid model serves as an ensemble prediction function. Receives real-time fused and standardized data vectors As input, the model ultimately outputs a scalar value between 0 and 1. This refers to the real-time comprehensive risk probability value. Its mathematical expression is: (4) in, This includes all trained parameters of the model, such as the tree structure parameters of GBDT, the weights and biases of LSTM, and the parameters of fully connected layers. This represents the risk prediction value for real-time construction scenarios. For real-time integrated data vectors; through trained... The input composite data vector Mapped to risk probability value . The closer the value is to 1, the higher the probability that the system predicts that in the next excavation cycle or in the next few hours, the risk of danger such as rock instability, water inrush, and mudslide will occur. The closer the value is to 0, the safer the current environment is.

[0027] The prediction process is continuous, updating dynamically with each tunneling cycle or at fixed time intervals such as hourly, forming a risk probability curve. This allows for proactive risk assessment of the construction process.

[0028] Regarding the aforementioned warning threshold The method for determining the warning threshold is as follows: The warning threshold is determined by comprehensively optimizing the model's performance on the validation set and considering engineering risk tolerance. The main steps are as follows: a. Model Performance Evaluation: Using validation set data, the standardized composite data vector of the validation set is input into the trained GBDT-LSTM hybrid model. The risk probability value output by the model is the risk prediction value for each sample in the validation set. The true label for each sample is known. (0 or 1) can be used to plot the precision-recall curve.

[0029] b. Key indicator calculation: Candidate threshold A discrete set is generated by uniformly sampling within the interval [0,1] at fixed step sizes such as 0.01 or 0.001, and all candidate thresholds are traversed. ,Will ≥ Samples that meet the criteria are classified as "positive for warning," while those that do not are classified as "negative for warning." This is combined with the actual label of the sample. Count the number of true positive samples ( TP ), number of false positive samples ( FP ), number of true negative samples ( TN ), number of false negative samples ( FN ), forming a confusion matrix: ①True positive: The prediction is risky and the risk actually exists.

[0030] ② False positive: A false positive is a prediction of risk when there is actually no risk.

[0031] ③ True negative: The prediction is risk-free and there is actually no risk.

[0032] ④ False negative: The prediction is that there is no risk, but there is actually a risk, i.e., the false negative is missed.

[0033] Then calculate: (5) (6) (7) in, For model accuracy; For model recall; F Precision is the harmonic mean of precision and recall, and it is a good indicator of the overall performance of a model at a specific threshold.

[0034] c. Threshold optimization selection: Select a threshold that optimizes the validation set. F To reach the maximum As This aims to find the optimal balance between accurate early warning and comprehensive risk assessment. FNear the optimal value, decision-makers can adjust the threshold based on the level of importance placed on safety in the specific project. If the project has extremely high safety requirements and can tolerate some false alarms but absolutely cannot miss alarms, then the threshold can be appropriately lowered to improve the recall rate. If the project is extremely sensitive to construction efficiency and wants to minimize false alarm interference, then the threshold can be appropriately raised to improve the accuracy rate.

[0035] Regarding the aforementioned triggering warning and verification: During real-time construction, the system continuously calculates... The conditions for triggering the warning are: > Once triggered, the "Dynamic Optimization and Decision" process is initiated, generating suggestions for adjusting construction parameters.

[0036] S3: Dynamic Optimization and Decision Making A genetic algorithm, an intelligent iterative search algorithm, is used to simulate the engineering decision-making optimization process of "generating solutions - evaluating merits - survival of the fittest - recombination and innovation." The population size is set to 50-100, the number of iterations to 30-50 generations, the crossover probability to 0.7-0.8, and the mutation probability to 0.01-0.05. This algorithm automatically finds the optimal combination of construction parameters under the current risk condition. When the risk prediction value... Exceeding the warning threshold At this time, the system switches from "monitoring and early warning" mode to "dynamic control" mode. The core of this step is to dynamically generate a set of optimal construction parameters for the current working conditions by solving a multi-objective optimization problem based on real-time geological conditions and risk prediction, aiming to synergistically ensure construction safety and efficiency. Specifically, this includes: ① Optimize the objective function Define a comprehensive cost function C As a minimization objective, this function quantifies the weighted sum of safety deviation and efficiency loss under specific construction parameters: (8) in, The tunneling speed is (m / h). The maximum allowable displacement (mm) represents the maximum deformation that the surrounding rock can withstand within a safe range. The minimum allowable displacement (mm) represents the minimum reasonable deformation under stable rock conditions. Both are determined based on engineering specifications, the ultimate deformation capacity of the support structure, and the rock grade. The maximum feasible tunneling speed (m / h) represents the maximum tunneling efficiency achievable under the existing mechanical equipment performance and construction organization capabilities. The minimum feasible tunneling speed (m / h) represents the minimum tunneling speed threshold that ensures construction quality and safety. Both are determined by the performance of mechanical equipment and the construction organization capability. and These are the safety weight coefficient and the efficiency weight coefficient, respectively, satisfying... + =1. Needs to be adjusted based on real-time risk level. Dynamic adjustment: when When the value is high, assign a higher weight to security (e.g., set...). =0.8, =0.2); when When the value is close to or low enough to the threshold, the efficiency weight can be appropriately increased (e.g., setting the threshold). =0.4, =0.6).

[0037] The predicted displacement (mm) represents the maximum possible displacement of the surrounding rock under the proposed construction parameters. This displacement is determined using a pre-trained displacement prediction model. calculate.

[0038] It is an independent machine learning model whose function is to establish a nonlinear mapping relationship from "geological conditions + construction parameters" to "surrounding rock displacement". It has a different division of labor from the aforementioned "risk prediction model": the risk model predicts the probability of "accident", while the displacement model predicts the specific amount of "deformation" to accurately quantify safety costs.

[0039] The overall solution method is as follows: A displacement prediction regression model based on GBDT-LSTM is established, which uses standardized integrated data vectors. X Input: Predicted displacement value The model is trained using historical displacement monitoring data and used in dynamic optimization to evaluate expected displacements under different combinations of construction parameters, thereby supporting optimization decisions.

[0040] The specific method is as follows: Similarly, the standardized composite data vector is used. Establish displacement data labels to represent the maximum displacement value actually monitored within a certain period after the corresponding time (such as the next construction cycle or several hours). (mm). Similarly, it is divided into training set (70%~80%), validation set (10%~15%), and test set (10%~15%) in chronological order.

[0041] Since this is a regression problem, the loss function is usually the mean squared error (MSE) or the mean absolute error (MAE): (9) in, N When the number of training samples; For the model to the first i Predicted displacement for each sample; For the first i The actual maximum displacement of each sample.

[0042] The model parameters are iteratively optimized using the training set data, and the model's generalization ability is monitored using the validation set data. Training stops when the model's MSE on the validation set shows no decrease for five consecutive iterations, resulting in a fully trained displacement prediction model. : (10) in, The trained parameters of the displacement prediction model.

[0043] When it is necessary to evaluate different combinations of construction parameters (especially tunneling speed) v When this is done, the proposed construction parameters (such as the new tunneling speed) will be used. v 0 ) and other current real-time monitoring data Combined into a new comprehensive data vector The detection data should be standardized according to formula (1) and then combined into a new comprehensive data vector. ,Will Input the pre-trained displacement prediction model Output predicted displacement ,Will Substitute into the comprehensive cost function C In this process, safety deviations are calculated, and the overall cost of the set of construction parameters is then assessed.

[0044] ②Constraints Water pressure constraint: ,in, The safe water pressure threshold is determined according to engineering safety standards; Rock mass integrity constraints: ,in, The critical rock mass integrity; Stress constraints: ,in, This represents the allowable stress value of the surrounding rock.

[0045] Variables to be optimized Advanced processing T : Discrete variable, the range of values ​​is determined by the engineering plan, and is an integer (0 represents no support, 1 represents pipe roof, 2 represents small pipe, 3 represents pre-grouting). When optimizing the selection of pretreatment methods that include grouting (e.g., T=1, 2, or 3), the following parameters need to be optimized simultaneously: grouting pressure. (MPa), slurry water-cement ratio (Dimensionless) Grouting volume (m) 3 ); Excavation step length (m).

[0046] Grouting pressure With water pressure The grouting pressure can be dynamically adjusted according to the water pressure. (11) in, and These are the coefficients determined through optimization.

[0047] and The method for determining it is as follows: Collecting historical engineering projects under different water pressures The grouting pressure that was ultimately verified to have successfully blocked the water inrush without causing rock mass damage was determined. Data pairs. Using methods such as linear regression, these ( , The data points are fitted to obtain the best-fit line. coefficients in k and b The training results k and b It is incorporated as a fixed parameter into the subsequent dynamic optimization algorithm.

[0048] ④ Dynamic optimization algorithm An intelligent iterative search algorithm is used to simulate the engineering decision-making optimization process of "generating a solution - evaluating its merits and demerits - eliminating the inferior and reorganizing and innovating", and automatically finds the best combination of construction parameters under the current risk conditions.

[0049] (1) Generate multiple candidate construction schemes Based on the current geological and risk conditions, the system automatically and randomly generates N sets (e.g., 50 sets) of feasible construction parameter combinations, forming the "first-generation" candidate solution library. Each solution (called an "individual") contains a specific set of parameter values, represented as a decision vector: (12) in, No. i The support method of the scheme; For the first i The water-cement ratio of the solution; For the first i The grouting volume of the solution; For the first i Excavation step length of the proposed scheme. Support method. The value rules are as follows: 0 represents no support, 1 represents pipe roof, 2 represents small pipe, and 3 represents pre-grouting.

[0050] For schemes that require grouting, the grouting pressure is calculated using formula (11).

[0051] Safety and cost simulation assessments were conducted for each solution: For the i The system sets up a solution and assigns its decision vector. Compared with other data currently being monitored in real time (such as water pressure) Rock mass integrity (etc.) are combined to form a new, complete standardized state vector. .Will The surrounding rock displacement prediction model is input into the trained formula (10). In the process, the maximum deformation of the surrounding rock that may be caused by adopting this scheme is quickly predicted. The cost of the scheme is calculated according to formula (8). C i It also checks whether each solution meets the constraints; if not, it assigns an extremely high cost. C i This is marked as "infeasible". Therefore, the comprehensive cost... C i The optimal solution is the set of construction parameters that minimizes the risk while satisfying all safety constraints.

[0052] S4: Execution and Feedback The optimized construction parameters are sent to on-site construction equipment and personnel to guide the current construction operations, and the model learns online through continuous monitoring.

[0053] During construction, new monitoring data are collected in real time and added to the training dataset. Through periodic retraining or incremental learning, the model can adapt to changes in geological conditions and improve prediction accuracy. Data updates adopt an incremental learning approach, periodically supplementing the dataset with newly collected monitoring data in chronological order, updating the dataset in an equivalent manner according to the original training, validation, and test set partitioning ratios, and iteratively fine-tuning the GBDT-LSTM hybrid model.

[0054] The technical solution of this invention has the following technical effects: 1. Existing technologies, when tunneling through wide fault zones, mainly rely on static design and experience-based decision-making. They cannot provide real-time, quantitative risk warnings for dynamically changing geological conditions ahead of construction, resulting in a lack of adaptability in support and excavation parameters. Existing technologies struggle to accurately predict real-time risks arising from changes in key indicators such as high water pressure and rock fracturing. Furthermore, existing construction schemes cannot dynamically and collaboratively self-optimize based on risk prediction results. This invention enables an integrated method for real-time risk prediction and dynamic intelligent control of construction parameters, overcoming the blindness and lag of traditional construction methods and improving construction safety and efficiency.

[0055] 2. This invention enables real-time, accurate early warning and quantitative assessment of construction risks. By integrating multi-source real-time monitoring data and employing machine learning models for dynamic prediction, the system can proactively and quantitatively identify the probability of combined risks such as surrounding rock instability and water inrush, effectively overcoming the lag and ambiguity of traditional geological forecasting. This allows construction teams to take targeted measures in advance, significantly improving the initiative and scientific rigor of risk prediction and reducing the likelihood of safety accidents at the source.

[0056] 3. This invention breaks through the limitations of traditional static design, realizing intelligent dynamic optimization of construction parameters and multi-objective collaboration. The method of this invention adaptively adjusts countermeasures based on real-time changes in the geological conditions ahead, avoiding the hidden dangers of insufficient measures in high-risk conditions and preventing over-support and resource waste in low-risk conditions, thereby improving construction efficiency and economic benefits while ensuring safety.

[0057] 4. This invention constructs a closed-loop intelligent control system with continuous evolution capabilities. By introducing an online learning mechanism, the model can be updated and optimized in real time using new monitoring data generated during construction, gradually enhancing its predictive and decision-making capabilities as the project progresses. This self-improving characteristic significantly enhances the system's adaptability to complex and variable geological conditions, providing long-term and reliable technical support for tunnels traversing wide fault zones. Attached Figure Description

[0058] Figure 1 This is a flowchart of the risk control method for tunnel construction in a wide fault zone based on real-time prediction, as described in this invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific examples described herein are only some embodiments of this invention, not all embodiments, and are not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0060] This invention provides a detailed description of the risk control method for tunnel construction in a wide fault zone based on real-time prediction, using the example of tunnel construction in a water conservancy project.

[0061] S1: Data Acquisition and Fusion S11: Data Acquisition: A sensor network including stress sensors, water pressure gauges, and displacement monitoring points is deployed inside the tunnel, along with advanced geological forecasting equipment to collect real-time data on water pressure ahead. Rock mass integrity coefficient Stress state of surrounding rock tunneling speed Support stress Displacement monitoring values Data from multiple sources. For example, the raw data collected at a certain moment might be: =2.5MPa, =65%, =18MPa =0.8m / h, =12MPa, =8mm.

[0062] S12: Data Standardization Processing: Historical data includes real-time water pressure monitoring at the front end. minimum value =0.5MPa, maximum value =5MPa. Rock mass integrity coefficient. The minimum value is 30%, and the maximum value is 90%. Surrounding rock stress state. The minimum value is 10 MPa, and the maximum value is 30 MPa. Tunneling speed. The minimum value is 0.2 m / h, and the maximum value is 1.5 m / h. Support stress The minimum value is 5 MPa, and the maximum value is 20 MPa. Displacement monitoring values. The minimum value is 2mm and the maximum value is 15mm.

[0063] According to the standardized formula Calculate the standardized data: Real-time water pressure monitoring at the front Standardized values: Similarly, the results are obtained sequentially based on the standardized formulas described above. The standardized value is 0.58. The standardized value is 0.4. The standardized value is 0.46. The standardized value is 0.47. The standardized value is 0.46.

[0064] S13: Multi-source data fusion: fusing standardized data into a comprehensive data vector. [0.4, 0.58, 0.4, 0.46, 0.47, 0.46].

[0065] S2: Machine Learning Risk Prediction Hybrid prediction model architecture: Construct a GBDT-LSTM hybrid model. The GBDT component processes the structured features in the standardized comprehensive data vector, and the LSTM component processes the sequence features formed by historical construction data. The features extracted from both are concatenated and input into a fully connected neural network layer to calculate the risk probability.

[0066] S21: Model Training and Validation Dataset Construction: Monitoring data, construction logs, and geological reports from completed tunnel construction projects in similar wide fault zones were collected to construct a dataset. Each sample contains a feature vector (a standardized composite data vector at a specific moment) and a risk label (0 - no risk, 1 - risky).

[0067] Dataset partitioning: Assuming the dataset contains 1000 time series samples, the training set is divided into the first 800 samples (80%), the validation set is the 801-900 samples (10%), and the test set is the 901-1000 samples (10%).

[0068] Model training: The GBDT-LSTM hybrid model is trained using the training set data. The loss function is binary cross-entropy. The model parameters are optimized through iterative training until the loss function value of the model on the training set converges.

[0069] S22: Risk Prediction and Verification Risk prediction calculation: This involves using the current moment's comprehensive data vector. X =[0.4, 0.58, 0.4, 0.46, 0.47, 0.46] Input the trained hybrid model to obtain the real-time comprehensive risk probability value. =0.65.

[0070] Warning threshold Determination: Use validation set data to evaluate model performance, plot precision-recall curves, and iterate through candidate thresholds. Calculate different corresponding , and F Value. Assuming when When =0.6, F The value reaches its maximum, therefore the warning threshold is determined. =0.6.

[0071] (3) Triggering an early warning: Current risk probability value =0.65> =0.6, triggering an early warning and initiating dynamic optimization and decision-making processes.

[0072] S3: Dynamic Optimization and Decision Making Objective function determination: Based on engineering specifications, the ultimate deformation capacity of the support structure, and the surrounding rock grade, determine the allowable displacement range. =12mm, =4mm; Determine the feasible range of tunneling speed based on the performance of mechanical equipment and construction organization capabilities. =1.2m / h, =0.3m / h. Current risk level. =0.65 is too high, set a safety weight coefficient. =0.8, efficiency weighting coefficient =0.2.

[0073] Application of displacement prediction model: Combine the current comprehensive data vector with different construction parameters to form a new state vector, and input it into the pre-trained displacement prediction model. The predicted displacement is obtained. For example, the tunneling speed of a candidate construction scheme. v =0.5m / h, and the standardized data of other parameters are combined and input into the displacement prediction model to obtain =10mm.

[0074] Comprehensive cost calculation: based on the comprehensive cost function

[0075] Calculate the cost of this plan: max(10-12,0)+max(4-10,0)=0+0=0, safety deviation term: =0; Efficiency loss item: ≈0.78; Comprehensive cost =0.8×0 + 0.2×0.78≈0.16; Constraint settings: Based on engineering safety requirements, set a safe water pressure threshold. =3MPa, currently =2.5MPa< Satisfying water pressure constraints; critical rock mass integrity =50%, current =65%≥RQD0, satisfying the rock mass integrity constraint; allowable stress value of surrounding rock. =25MPa, currently =18MPa≤ It satisfies stress constraints.

[0076] Variables to be optimized and dynamic optimization algorithm: Look-ahead processing form of variables to be optimized T The value of is in the range of 0-3. Assuming 50 candidate construction schemes are generated, the decision vector of one of the schemes is . =[2, 0.8, 15, 0.5], meaning the support method is small catheter ( T =2), slurry water-cement ratio =0.8, grouting volume =15m 3 Excavation step length =0.5m.

[0077] Grouting pressure calculation: Based on historical engineering data fitting, k=1.2, b=0.3, currently... =2.5MPa, therefore the grouting pressure =1.2×2.5+0.3=3.3MPa.

[0078] Scheme Evaluation: The decision vector of this scheme is combined with the current real-time monitoring data to form a new standardized state vector, which is then input into the displacement prediction model to obtain... =9mm, calculate the overall cost: ≈0.16, which satisfies all constraints.

[0079] Optimal Solution Selection: Each of the 50 candidate solutions is evaluated, and the solution with the lowest overall cost that meets the constraints is selected as the optimal construction solution. Assuming the above solution has the lowest overall cost, it is determined to be the optimal solution.

[0080] S4: Execution and Feedback Implementation plan: The optimal construction plan parameters are: support form with small guide pipes, grout water-cement ratio of 0.8, and grouting volume of 15m³. 3 The grouting pressure of 3.3MPa and the excavation step length of 0.5m were issued to the on-site construction equipment and personnel to guide the construction operation.

[0081] Data Feedback and Model Updates: During construction, new monitoring data, such as post-construction water pressure, rock mass integrity, and displacement, are collected in real time and added to the training dataset. The GBDT-LSTM risk prediction model and displacement prediction model are periodically retrained to update model parameters and improve the model's prediction accuracy and adaptability.

[0082] In this embodiment, the data related to risk prediction accuracy, construction efficiency, and safety accident incidence were all obtained through parallel comparative tests under the same working conditions during the same period. Two construction sections with geological conditions (rock mass fracture degree, water pressure range, and ground stress level) consistent with the wide fault zone of this tunnel were selected. Section A adopted the method of this invention, while Section B adopted a combination of traditional experience judgment and static construction parameters. The entire construction process of both sections was tracked, monitored, and statistically analyzed. Statistics showed that 120 construction risk warning nodes were identified during the construction of Section A. Based on subsequent construction verification, the prediction results for 110 nodes were consistent with the actual working conditions, resulting in a risk prediction accuracy of 110 ÷ 120 ≈ 91.7%. In Section B, 118 risk warning nodes were identified during the same period, and 77 nodes were verified to be accurate during construction. The risk prediction accuracy of the traditional experience judgment method was calculated to be 77 ÷ 118 ≈ 65.3%. The accuracy improvement of the method of this invention compared to the traditional method was calculated to be (91.7% - 65.3%) ÷ 65.3% ≈ 40.4%. In terms of construction efficiency, the average time to complete the same tunneling progress in Section A was 28.5 days, while in Section B it was 35 days. The efficiency improvement of the method of this invention is calculated as (35-28.5)÷35=18.6%, meaning that the construction efficiency is 18.6% higher than that of the traditional static method. Regarding safety accident statistics, no safety accidents such as water inrush, mudslides, or surrounding rock instability occurred during the entire construction process in Section A, resulting in a safety accident rate of 0. In Section B, a total of 4 minor safety accidents occurred during the entire construction process, covering 3.2% of the total tunneling progress traversing the wide fault zone. This corresponds to a 3.2% impact rate of safety accidents under the traditional method in the same geological conditions. Compared to traditional methods that rely on experience-based judgment, this invention achieves three core advantages through a closed-loop control mechanism of real-time risk prediction, dynamic parameter optimization, and continuous model iteration: the accuracy of risk prediction is improved by over 40 percentage points, effectively avoiding the misjudgment and omission problems of traditional methods; construction efficiency is improved by nearly 19 percentage points, avoiding the waste of efficiency caused by over-support; and zero safety accidents occur during construction in wide fault zones, completely solving the pain point of difficult safety risk control under complex geological conditions in traditional methods, fully verifying the significant superiority of the method of this invention in both safety and efficiency dimensions.

[0083] The above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for risk control in tunnel construction in wide fault zones based on real-time prediction, characterized in that, Includes the following steps: S1: Data Acquisition and Fusion: Real-time acquisition of multi-source data during tunnel construction. This multi-source data includes at least three of the following: water pressure data, rock mass integrity coefficient data, surrounding rock stress data, tunneling speed data, support stress data, and displacement monitoring data. The multi-source data is standardized, and the standardized data values ​​are then fused into a comprehensive data vector. This provides input data for subsequent risk prediction; S2: A hybrid model combining structured feature extraction and temporal feature extraction: This model is based on a hybrid of Gradient Boosting Decision Tree (GBDT) and Long Short-Term Memory (LSTM) network. The GBDT component extracts structured features, and the LSTM component extracts temporal features. The two types of features are concatenated and input into a fully connected layer to calculate the risk probability, using the aforementioned comprehensive data vector. The system takes real-time predicted construction risks as input and outputs a real-time comprehensive risk probability value. Determine the warning threshold ,when > Timely triggering of warnings; S3: Dynamic Optimization and Decision-Making: When an early warning is triggered, based on the real-time geological conditions and the risk prediction results, the optimal combination of construction parameters is dynamically generated by solving a multi-objective optimization problem with a weighted combination of safety cost and efficiency cost as the objective function and water pressure, rock mass integrity, and surrounding rock stress as constraints. S4: Execution and Feedback: The optimal construction parameters are distributed to on-site construction equipment and personnel to guide construction operations. At the same time, new monitoring data is continuously collected and added to the multi-source data in step S1 to achieve data updates.

2. The method for risk control in tunnel construction in wide fault zones based on real-time prediction as described in claim 1, characterized in that: The standardized data value in step S1 is equal to the original data value minus the minimum value of historical data, and then divided by the difference between the maximum value of historical data and the minimum value of historical data.

3. The method for risk control in tunnel construction in wide fault zones based on real-time prediction as described in claim 1, characterized in that: The integrated data vector in step S1 It includes at least the parameters of real-time monitoring of water pressure, rock mass integrity coefficient, surrounding rock stress state, tunneling speed, support stress, and displacement monitoring values, with each parameter forming a vector in sequence.

4. The method for risk control in tunnel construction in wide fault zones based on real-time prediction as described in claim 1, characterized in that... In step S2, the GBDT component of the hybrid model is set to a tree depth of 6-10 layers and a learning rate of 0.05-0.1, the LSTM component is set to a hidden layer dimension of 64-128 and a time step of 5-10, and the fully connected neural network layer contains 32 neurons and the activation function is Sigmoid.

5. The method for risk control in tunnel construction in wide fault zones based on real-time prediction as described in claim 1, characterized in that, The model training and validation in step S2 includes the following steps: Construct a dataset where each data sample includes a feature vector and a risk label. The feature vector is a comprehensive data vector representing a specific moment or a specific construction progress segment. The composite data vector at time t Recorded as Risk labels are uniformly assigned based on monitoring records, construction logs, and geological reports; Based on the constructed dataset, the training set, validation set, and test set are divided in chronological order. The training set uses the first 70% to 80% of the time series construction data, the validation set uses the 10% to 15% of the time series construction data immediately following the training set, and the test set uses the last 10% to 15% of the time series construction data. For the training set, binary cross-entropy is used as the loss function. The deviation between the risk probability value predicted by the quantitative model and the true risk label of the sample is calculated using the binary cross-entropy as the loss function. （3） in, This represents the number of training samples; For the first i The true risk label of each sample; This represents the risk probability value predicted by the model.

6. The method for risk control in tunnel construction in wide fault zones based on real-time prediction as described in claim 1, characterized in that, The warning threshold in step S2 The method for determining it includes the following steps: Model performance evaluation: Using validation set data, the model outputs a risk prediction value for each sample. Combined with the real label corresponding to each sample , The value is 0 or 1, where 0 indicates no risk and 1 indicates risk. This is used to plot the precision-recall curve, providing basic data for subsequent threshold evaluation. Key indicator calculation: Risk prediction value based on each sample. and real labels Iterate through the candidate thresholds between 0 and 1 ,Will ≥ The sample is judged as a positive warning sample, and the sample is judged as a negative warning sample. The number of true positive samples is calculated. TP ), number of false positive samples ( FP ), number of true negative samples ( TN ), number of false negative samples ( FN The accuracy is calculated using the following formula. Pr Recall rate Re and F value: （5） （6） （7） Threshold optimization selection: based on each of the candidate thresholds. corresponding F Value, select to make the validation set F The value reaches its maximum As .

7. The method for risk control in tunnel construction in wide fault zones based on real-time prediction as described in claim 1, characterized in that, In step S3, the objective function to be optimized is the comprehensive cost function. C The formula is: （8） in, This refers to the tunneling speed; This represents the maximum allowable displacement. This is the minimum allowable displacement. This represents the maximum feasible tunneling speed. This represents the minimum feasible tunneling speed. For safety weighting coefficients, For efficiency weighting coefficients, satisfying + =1; For displacement prediction models based on pre-trained GBDT-LSTM The output is the predicted displacement.

8. The method for risk control in tunnel construction in wide fault zones based on real-time prediction as described in claim 7, characterized in that, The predicted displacement Calculate the displacement prediction model based on pre-training Specifically, it includes: Constructing a displacement prediction model The training dataset is derived from historical monitoring data of tunnels that have been constructed across wide fault zones. Each data sample contains input features and label data, with the input features being a standardized composite data vector. , For real-time monitoring of water pressure at the front line; This is the rock mass integrity coefficient; The surrounding rock stress state; This refers to the tunneling speed; For support stress; The data represents the displacement monitoring value; the label data represents the maximum displacement value of the surrounding rock actually monitored in the next construction cycle after the corresponding input feature acquisition time. ; The model is divided into training, validation, and test sets in chronological order. The training set uses the first 70%–80% of the time series data, the validation set uses the next 10%–15% of the time series data, and the test set uses the last 10%–15% of the time series data. A GBDT-LSTM hybrid architecture is used to build a regression model, with mean squared error (MSE) as the loss function. The model parameters are iteratively optimized using the training set data, and the model's generalization ability is monitored using the validation set data. Training stops when the MSE on the validation set shows no further decrease over consecutive iterations, resulting in a trained displacement prediction model. Its mathematical expression is: （10） in, The trained parameters of the displacement prediction model; The real-time monitoring data of the current construction scene is obtained, processed according to the standardized formula (1), and combined with the construction parameters to be evaluated to form a new comprehensive data vector. ;Will Input the completed displacement prediction model The model output is the predicted displacement under the current combination of construction parameters. .

9. The method for risk control in tunnel construction in wide fault zones based on real-time prediction as described in claim 1, characterized in that, The dynamic optimization algorithm in step S3 includes the following steps: N sets of feasible construction parameter combinations are randomly generated to form a candidate solution library. The decision vector for each solution is: (12) in, No. i The support structure of the solution, For the first i The water-cement ratio of the solution; For the first i The grouting volume of the solution; For the first i The excavation step length of the scheme; For each scheme, the decision vector is combined with other data currently being monitored in real time to form a new standardized state vector. Input surrounding rock displacement prediction model Predict the maximum deformation of the surrounding rock and calculate the cost of the proposed solution. C i It then checks whether the constraints are met; if the constraints are not met, an extremely high cost is imposed. C i Mark as infeasible, select comprehensive cost C i The construction parameter scheme with the lowest cost and that satisfies all safety constraints is considered the optimal scheme.

10. The method for risk control in tunnel construction in wide fault zones based on real-time prediction according to claim 1, characterized in that, The data update in step S4 adopts an incremental learning approach, periodically supplementing the dataset with newly collected monitoring data in chronological order. The dataset is updated in an equivalent manner, using the first 70% to 80% of the time series data for the training set, the 10% to 15% of the time series data immediately following the training set for the validation set, and the last 10% to 15% of the time series data for the test set. The GBDT-LSTM hybrid model is then iteratively fine-tuned.

Images