A tunnel initial support safety early warning method based on multi-source monitoring fusion and time sequence prediction

Abstract

The present application belongs to the technical field of tunnel initial support safety early warning, and discloses a tunnel initial support safety early warning method based on multi-source monitoring fusion and time sequence prediction, comprising: acquiring multi-dimensional monitoring data of surrounding rock pressure, grid steel frame stress and support structure displacement collected by a multi-source sensing monitoring network in real time; performing FLAC3D numerical simulation according to the multi-dimensional monitoring data and surrounding rock physical and mechanical parameters, and calculating the dangerous threshold of the support structure under different working conditions; inputting the multi-dimensional monitoring data into a pre-trained LSTM neural network model, and outputting the evolution prediction trend of the initial support state; and matching and analyzing the evolution prediction trend with the dangerous threshold, and outputting the safety early warning level; the present application solves the problems of single index and lack of timeliness and accuracy of early warning results of the traditional early warning method, realizes accurate early warning based on multi-source data and intelligent algorithms, and improves the safety and stability of tunnel construction.

Description

Technical Field

[0001] This invention belongs to the field of tunnel initial support safety early warning technology, specifically involving a tunnel initial support safety early warning method based on multi-source monitoring fusion and time-series prediction. Background Technology

[0002] During tunnel construction, initial support is the first line of defense to ensure the stability of the surrounding rock and prevent collapse. Accurate sensing and early warning of the stress and deformation state of the initial support are crucial for ensuring construction safety.

[0003] Traditional early warning technologies for initial support mainly rely on preset displacement thresholds or stress limits for alarms. Existing early warning methods for tunnel initial support largely depend on simple linear extrapolation of field measured data or one-way alarms with fixed thresholds. However, these methods have the following major drawbacks: most traditional methods only focus on surrounding rock displacement, failing to simultaneously correlate multi-source mechanical information such as surrounding rock pressure, grid steel frame stress, and shotcrete strain; risk assessment is limited to a single dimension, making it prone to missed alarms due to incomplete information; non-intelligent prediction models rely heavily on empirical criteria, making it difficult to handle non-stationary data fluctuations caused by geological changes, resulting in insufficient prediction accuracy; and the early warning results lack sufficient timeliness and accuracy, failing to meet the risk control needs of dynamic construction environments. Summary of the Invention

[0004] The purpose of this invention is to provide a tunnel initial support safety early warning method based on multi-source monitoring fusion and time-series prediction, which solves the problems of single indicators and lack of timeliness and accuracy of early warning results in traditional early warning methods, and realizes accurate early warning based on multi-source data and intelligent algorithms, thereby improving the safety and stability of tunnel construction.

[0005] The present invention adopts the following technical solution: A method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction includes: Step 1: Acquire multi-dimensional monitoring data of surrounding rock pressure, grid steel frame stress, and support structure displacement collected in real time by the multi-source sensor monitoring network; Step 2: Based on multidimensional monitoring data and surrounding rock physical and mechanical parameters, perform FLAC3D numerical simulation to calculate the danger threshold of the support structure under different working conditions; Step 3: Input the multidimensional monitoring data into the pre-trained LSTM neural network model and output the predicted trend of the evolution of the initial support status; Step 4: Match the predicted evolution trend with the danger threshold and output the safety warning level.

[0006] The beneficial effects of this invention are as follows: The multi-source monitoring network in this invention simultaneously collects three types of data: pressure, stress, and displacement, breaking through the limitations of traditional single-parameter monitoring in risk assessment and achieving comprehensive perception of the stress and deformation state of the support structure; FLAC3D numerical simulation establishes a three-dimensional mechanical model based on the physical and mechanical parameters of the surrounding rock and the constitutive relationship of the support components. By simulating the mechanical response of the support structure under different excavation conditions, it extracts the stress and deformation characteristic values ​​under the critical failure state as a dynamic early warning benchmark, realizing the dynamic updating of the early warning threshold according to the construction conditions; the LSTM neural network model uses forget gates and input gates... By deeply mining the temporal evolution patterns of monitoring data through the output gate mechanism, and combining the blasting response spatiotemporal attention mechanism and construction step sequence encoding vector to enhance the features of the tunnel excavation cycle, advanced perception and trend prediction of support structure failure risk are achieved. By performing real-time matching analysis between LSTM predicted values ​​and FLAC3D dynamic thresholds, and using a combination of analytic hierarchy process and entropy weighting method to determine the comprehensive risk index of displacement and stress, automatic linkage between four-level graded early warning and construction auxiliary decision-making is achieved. This invention enables early perception and accurate early warning of initial support structure failure risk in complex geological sections of drill-and-blast tunnels. Attached Figure Description

[0007] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the sensor layout for the multi-source monitoring section of the initial support in this invention; Figure 3 This is a schematic diagram of the LSTM-based time series data prediction logic in this invention. Detailed Implementation

[0008] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0009] In one embodiment, such as Figure 1 As shown, the present invention provides a method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction, including: acquiring multi-dimensional monitoring data of surrounding rock pressure, grid steel frame stress and support structure displacement collected in real time by a multi-source sensor monitoring network.

[0010] The core system of this invention mainly consists of a field perception module, a data analysis module, and an early warning execution module.

[0011] Regarding the construction of the multi-source monitoring network, specifically, the multi-source sensing monitoring network includes pressure sensors, stress sensors, and displacement sensors. Multi-source sensing equipment is deployed at key monitoring sections of the tunnel, including flexible displacement gauges for measuring surrounding rock convergence, surface strain gauges for measuring the stress on the grid steel frame, and pressure cells for monitoring the pressure of the surrounding rock behind the support. The sampling frequency and data transmission protocol of the sensors are configured to ensure the synchronous acquisition of high-frequency monitoring data.

[0012] The sensor parameters are as follows: the surrounding rock pressure cell adopts a vibrating wire earth pressure cell (model: TY-200), with a range of 0~5MPa, accuracy ≤0.2%FS, and sampling frequency of 1Hz; the steel frame strain gauge adopts a surface-welded vibrating wire strain gauge (model: JMZX-212B), with a range of ±1500με, accuracy ≤0.1%FS, and sampling frequency of 1Hz; the displacement sensor adopts a laser convergence meter and a flexible array displacement sensor (model: WY-100), with a range of 0~500mm, accuracy ±0.1mm, and sampling frequency of 0.5Hz.

[0013] like Figure 2 As shown, the sensor deployment scheme is as follows: the monitoring section spacing for Class III / IV surrounding rock sections is 10m to 15m, and the monitoring section spacing for Class V surrounding rock sections is 5m to 8m; for each monitoring section, one pair of pressure cells and strain gauges are arranged at each of the five measuring points at the arch crown, left arch waist, right arch waist, left side wall, and right side wall; one settlement measuring point is arranged at the arch crown, and two horizontal convergence measuring lines are arranged at the arch waist and side wall positions. Each section has a total of 5 pressure measuring points, 10 stress measuring points, and 3 displacement monitoring lines.

[0014] This also includes: preprocessing multidimensional monitoring data, using a median filtering algorithm to eliminate transient noise pulses caused by blasting vibrations; and employing... Outliers are identified and removed, and then filled using the neighborhood mean. Min-Max normalization is performed on the data after outlier removal.

[0015] After obtaining the raw monitoring data, data smoothing should be performed. A median filtering algorithm should be used to eliminate transient noise pulses caused by various factors such as blasting impacts, construction machinery vibrations, electromagnetic interference, and occasional temperature drift of the sensors. Filtering these outliers improves the smoothness of the input data. Specifically, the average value of the monitoring data acquired within the current sliding window is calculated. Calculate the standard deviation of the monitoring data within this window. ; Determine real-time data Does it meet the requirements? If the condition is not met, it is considered an outlier and the data is either removed or filled using the neighborhood mean.

[0016] In this embodiment, the sliding window size is taken from the most recent 120 consecutive data points (equivalent to the most recent 2 minutes of real-time data stream for a sensor with a sampling frequency of 1Hz); the neighborhood range for the neighborhood mean fill is set to replace the outlier with the arithmetic mean of the three valid data points before and after the outlier (a total of six neighborhood points). After removing outliers, to eliminate the influence of different physical dimensions on model training, the feature data is subjected to Min-Max normalization, linearly mapping the data to the [0, 1] interval. The calculation formula is as follows: , in, These are the monitoring values ​​after linear mapping. These are the original monitoring values. This represents the minimum value of this feature dimension within the sliding window. This represents the maximum value of this feature dimension within the sliding window.

[0017] In one embodiment, FLAC3D numerical simulation is performed based on multidimensional monitoring data and surrounding rock physical and mechanical parameters to calculate the danger threshold of the support structure under different working conditions.

[0018] Specifically, a mechanical model including the surrounding rock and the initial support structure was established using the FLAC3D numerical simulation method to conduct stability analysis of the support system.

[0019] The specific methods for determining the danger threshold in FLAC3D include: inputting the physical and mechanical parameters of the surrounding rock and the constitutive relationship of the support components (including the density, elastic modulus, Poisson's ratio, cohesion, internal friction angle of the surrounding rock, as well as the mechanical parameters of shotcrete, anchor bolts, and steel frames), simulating the entire process of tunnel excavation and support; extracting the stress distribution law when the strength of the support structure reaches the ultimate state, and obtaining the bending moment, axial force, and plastic zone distribution characteristics of the initial support under various working conditions; establishing a dynamic threshold function based on the support safety factor, and determining the displacement limit and stress limit corresponding to different safety levels.

[0020] The boundary conditions, initial in-situ stress field, and constitutive models of support components (shotcrete, anchor bolts, steel frames) are set according to design parameters. Numerical calculations are used to simulate the bending moment distribution, axial force evolution, and plastic zone expansion of the initial support under different excavation conditions. By comparing the simulation results, the critical threshold for the initial support to enter unsteady deformation is determined, serving as a mechanical benchmark for subsequent risk assessment.

[0021] The constitutive selection and parameter settings for the three-dimensional mechanical reference model are as follows: The surrounding rock is modeled using an ideal elastoplastic Mohr-Coulomb constitutive model. Taking Class IV surrounding rock as an example, the density... bulk modulus shear modulus Cohesion internal friction angle Shotcrete (C30) adopts an elastic constitutive model, with an elastic modulus of... Poisson's ratio The system anchor bolts (HRB400, 22mm diameter) adopt the constitutive model of cable structural units, with an elastic modulus of... The tensile yield strength is 150kN; the steel frame (I18 I-beams) adopts Beam structural units, and the elastic modulus is... Yield strength .

[0022] Model dimensions and mesh generation: The 3D model has a clear width of 12m and a clear height of 10m for the excavation section; the overall 3D model has a width of 100m in the X direction (lateral), a width of 60m in the Y direction (longitudinal excavation direction), and a width of 80m in the Z direction (vertical) (45m for the overburden above the tunnel and 25m for the underburden). Boundary conditions are: the bottom of the model is constrained for all translational degrees of freedom, the four sides are constrained for normal translational degrees of freedom, and the top is a free boundary. The model is meshed into 185,000 hexahedral solid elements, with mesh refinement within a 1.5-times-diameter area around the tunnel excavation outline, and the element size is controlled at 0.4m × 0.8m.

[0023] Simulation Results and Key Calculations: In this embodiment, the model simulated the following four typical working conditions: (Condition 1) Full-section excavation without support; (Condition 2) Full-section excavation with initial support (standard design support); (Condition 3) Step excavation with initial support; (Condition 4) CD method excavation with weak surrounding rock and water-rich conditions. Key calculation results show that: Under Condition 1, the surrounding rock undergoes penetrating plastic shear failure, with the plastic zone extending to a depth of 3.8m; under Condition 2, the maximum settlement of the arch crown is 42mm, and the maximum compressive stress of the concrete is 12.4MPa; under Condition 3, the peripheral convergence value is 35mm, and the maximum bending moment of the steel frame is 48kN·m; under Condition 4, the plastic zone of the surrounding rock extends along both sides of the arch foot towards the depth of the shear failure. Under the step method working condition of Class V surrounding rock, the ultimate tensile stress at the crown is concentrated at the inner fold line of the crown (reaching the ultimate value of 2.4 MPa), the maximum compressive stress at the waist reaches 18.5 MPa, and the ultimate settlement displacement limit is 120 mm.

[0024] The specific criterion for determining the critical threshold is: when the shotcrete reaches 80% of its ultimate compressive strength (i.e., the design limit of ultimate concrete stress). The steel frame must reach its yield strength, or as per the relevant specifications. Ultimate displacement calculated by safety factor (e.g., the critical threshold for subsidence of the arch in Class IV surrounding rock) Convergence threshold ), or FLAC3D calculations may fail to converge (unbalanced forces cannot converge to 10). -5The following (and the complete penetration of the plastic zone) serves as the critical threshold for the support structure to enter unsteady deformation.

[0025] Preferably, a dynamic threshold function The mathematical expression is: , in, The initial support ultimate bearing displacement value obtained by FLAC3D (based on the shotcrete reaching its ultimate compressive strength or the steel frame reaching its yield strength). This refers to the allowable compressive strength safety factor for shotcrete. The deformation rate correction coefficient is taken as... ; For real-time deformation acceleration; For monitoring time, this function dynamically adjusts the limit displacement threshold with deformation acceleration and time; that is, the greater the deformation acceleration, the more sensitive the threshold adjustment and the more timely the early warning response.

[0026] The recalculation of the danger threshold is triggered when the tunnel face excavation distance exceeds the preset range, or when it crosses the geological boundary of different surrounding rock levels, or when the displacement change rate monitored in real time on site exceeds the preset range. During recalculation, the real-time monitoring data on site is transmitted to FLAC3D to correct the surrounding rock mechanical parameters.

[0027] Specifically, a recalculation is triggered every 25m of continuous tunnel excavation at the tunnel face, when crossing the geological boundary of different surrounding rock grades, or when the displacement change rate is greater than or equal to 5mm / d as monitored in real time on site. During the recalculation, the actual pressure value of the surrounding rock obtained from real-time on-site monitoring will be used. Actual value of roof slab settlement The surrounding rock deformation modulus is corrected by inputting the data into FLAC3D and using the Levenberg-Marquardt inversion algorithm. With cohesion Parameters such as these.

[0028] The critical displacement and stress values ​​corresponding to the four warning levels (blue / yellow / orange / red) are determined by the ultimate bearing state calculated using FLAC3D simulation combined with a preset safety factor: the blue warning level corresponds to a safety factor... A yellow alert corresponds to a displacement or stress reaching a limit value of 60% ≤ deviation rate < 80%; an orange alert corresponds to a displacement rate reaching a limit value of 80% ≤ deviation rate < 90%; a red alert corresponds to a displacement rate reaching a limit value of ≥ 90%, or a displacement rate exceeding the limit value for 3 consecutive days. In this embodiment, the safety factor is determined according to the "Code for Design of Railway Tunnels" (TB 10003-2016) and the "Code for Design of Highway Tunnels" (JTG 3370.1-2018), and the allowable compressive strength safety factor of shotcrete is taken as... Safety factor values ​​for tensile / compressive strength of steel frame .

[0029] In one embodiment, preprocessed multidimensional monitoring data is input into a pre-trained LSTM neural network model, which outputs a predicted trend of the evolution of the initial support status.

[0030] like Figure 3 As shown, the LSTM neural network model is trained using collected historical monitoring data to predict the evolution of support conditions in future periods. During the training phase, hidden layer neurons with long-range dependency processing capabilities are used to learn the nonlinear characteristics of displacement and stress evolution over time, ensuring that the model can adapt to data fluctuations caused by surrounding rock rheology and changes in construction sequence.

[0031] The predictions made by the LSTM neural network model specifically include: extracting historical displacement change rate and stress growth slope as input feature vectors; automatically filtering long-term and short-term dependencies in historical monitoring data through the forget gate mechanism of the LSTM model; optimizing network weights using the adaptive moment estimation algorithm; and calculating the predicted support status value for the next T time steps in real time.

[0032] Specifically, the model selects "displacement change rate, stress growth rate, and surrounding rock grade (numerically mapped to grades 3-5)" as core input features to represent the deformation dynamics trend, load growth rate, and inherent physical properties of the surrounding rock, respectively. The model incorporates a spatiotemporal attention mechanism for blasting response, constructs a weight matrix for a high-frequency response window from 0 to 6 hours after blasting excavation, and combines this with construction step sequence encoding vectors to enhance the features of the tunnel excavation cycle.

[0033] The network structure employs a 3-layer stacked LSTM architecture. The number of hidden layer neurons is determined by a combination of the input dimension, empirical formulas, and a grid search method, as follows: 128 neurons in the first hidden layer, 64 neurons in the second hidden layer, and 32 neurons in the third hidden layer. A fully connected layer with 32 neurons and an attention weight layer are then connected at the end. A Dropout layer with a dropout rate of 0.2 is placed between any two LSTM layers.

[0034] An online mini-batch gradient update mechanism based on a time-sliding window (window length W=168h, batch size=32) is adopted, with an initial learning rate set to 0.001 and the Adam optimizer used. The algorithm hyperparameters are set as follows: initial learning rate... First-order moment estimation of exponential decay rate Second-order moment estimation of exponential decay rate The learning rate automatically decays to 95% of its original value every 10 epochs of training; the batch size is set to 32, and the maximum epochs are set to 200. Online fine-tuning of the neural network weights is performed every 24 hours of newly acquired monitoring data, with a fine-tuning learning rate set to 0.0001, iterating for 5-10 epochs to avoid single-point noise interference. The number of iterations is set to 200, coupled with Early Stopping (tolerance = 20). The above hyperparameters were determined by comparing the convergence performance of the loss function on the validation set using a grid search method.

[0035] The data used to train the model came from the Taihang Mountain Extra-Long Tunnel Construction Monitoring Database, which contains 150,000 valid high-frequency time-series samples, covering 14 months (approximately 420 days) of continuous monitoring data for typical construction sections of the tunnel in Class III, IV, and V surrounding rock. The dataset was split chronologically into a 70% training set (first 294 days of data), a 15% validation set (middle 63 days of data), and a 15% test set (last 63 days of data).

[0036] Prediction time step of LSTM neural network model This corresponds to the evolution trend over the next 1-7 days. Predicted time step. The range of values ​​is Each time step (corresponding to the evolution trend over the next 1-7 days, with a sampling interval of 1 hour) is determined based on the average engineering response window period of the tunnel support structure from the occurrence of accelerated rheological deformation to the occurrence of local instability and failure. In this embodiment, a time step is set... The model takes historical monitoring data from the past 72 hours as input, calculates and outputs the state evolution trend for the next 24 hours.

[0037] The neural network model was continuously optimized to ensure the generalization ability and timeliness of the prediction results in complex geological transition zones. On the test set, the model's displacement prediction root mean square error (RMSE) was 0.68 mm, mean absolute error (MAE) was 0.45 mm, and mean absolute percentage error (MAPE) was 3.2%; stress prediction RMSE was 0.12 MPa, and MAE was 0.08 MPa. Under Class III / IV surrounding rock conditions, the model's prediction mean absolute error (MAE) was less than 0.8 mm, the RMSE was less than 1.2 mm, and the MAPE was less than 4.5%; under Class V extremely high-stress soft rock conditions, the MAPE was less than 6.2%. Compared with the traditional Autoregressive Integrated Moving Average (ARIMA) model (RMSE=2.15mm, MAE=1.62mm) and the three-layer BP neural network (RMSE=1.42mm, MAE=1.05mm) on the same test set, the RMSE of the model of this invention is reduced by 68.4% compared with ARIMA and by 52.1% compared with BP neural network.

[0038] In one embodiment, the predicted evolution trend is matched with a danger threshold to output a safety warning level.

[0039] Regarding the real-time output of safety warning levels and construction auxiliary decision-making, the LSTM prediction results are matched and analyzed with the hazard thresholds determined by FLAC3D to output the safety warning level in real time and execute the corresponding construction auxiliary decision-making.

[0040] The offset rate is the ratio of the monitored or predicted value to the limit value calculated by FLAC3D. It is used to characterize the degree to which the current or future support state deviates from the safety limit. It includes two forms: displacement offset rate and stress offset rate.

[0041] The formula for calculating the comprehensive risk index of the matching analysis is as follows: , in, The displacement value is predicted by LSTM. The stress value predicted by LSTM. The critical threshold for limit displacement determined by FLAC3D calculation. The critical threshold of ultimate stress determined by FLAC3D calculation. It is the displacement offset rate (i.e., the degree of deviation of the predicted displacement value from its limit threshold). This is the stress offset rate (i.e., the degree of deviation of the predicted stress value from its limit threshold). For displacement weighting coefficients, This is the stress weighting coefficient; it is dynamically adjusted according to the surrounding rock conditions. and The value of , and ;according to The preset range in which the value is located will output the corresponding warning level.

[0042] Specifically, and All values ​​are uniformly taken from the same future time point within the prediction time window (such as the 24th hour in the future) to achieve strict time alignment of cross-channel features; and These represent the deformation and strength design limits determined by the FLAC3D model, respectively. The formulas use ratios... and It achieves dimensionless and standardized processing of displacement and stress, which have different physical dimensions; and The weights were obtained through a combination of the analytic hierarchy process (AHP) and the entropy weighting method; the dynamic adjustment rule is as follows: in sections with high ground stress / rockburst tendency, take... , In the section of soft rock with large deformation, take , ; Regular paragraph selection , .

[0043] The safety warning levels include blue, yellow, orange, and red; blue warning corresponds to... Yellow alert corresponds to Orange alert corresponds to Red alert corresponds to .

[0044] The specific triggering logic for the security alert levels is as follows: Blue alert corresponds to... And the offset rate is less than 60%; a yellow alert is corresponding to... And 60% ≤ offset rate < 80%; Orange alert corresponds to And 80% ≤ offset rate < 90%; Red alert corresponds to Either the offset rate is ≥90%, or the displacement rate is greater than or equal to 10 mm / d for several consecutive days. The offset rate is the percentage of the measured displacement or stress value to the limit value calculated by FLAC3D.

[0045] Based on the safety warning level, corresponding construction auxiliary decisions are executed: under blue and yellow warnings, control parameters are automatically updated and recorded in the background log; under orange warnings, a support reinforcement plan is generated and sent to the project chief engineer's terminal; under red warnings, an audible and visual alarm is triggered and a work stoppage command is output.

[0046] Specifically, blue alert (normal status): Yellow alert indicates the support system is in a stable and elastic state, and excavation can proceed as planned; Yellow alert (monitored status): This indicates localized stress concentration in the support structure; therefore, the monitoring frequency should be increased to twice the original frequency, and patrols should be intensified. Orange alert (alert status): A red alert indicates that the support is approaching its bearing limit, triggering an early warning and initiating local reinforcement measures (such as adding pre-installed anchor bolts). This requires secondary confirmation from the on-site project chief engineer via electronic terminal (response time limited to 2 hours); Red alert (dangerous state): This indicates that the support structure is at risk of instability. Excavation at the working face must be stopped immediately, personnel evacuated, and an emergency reinforcement plan implemented. An audible and visual alarm should be triggered, and the power supply to the working face machinery should be automatically cut off to force a work stoppage and avoid danger. Actual structural reinforcement measures can only be implemented after the project manager and chief engineer have jointly signed off on the physical unlocking on-site.

[0047] The specific measures for generating a support reinforcement plan include the following: (1) Increase the density of anchor bolt support: from the original grid spacing of 1.0m×1.0m (i.e. 1.0 bolts / m) 2 Encrypted to 0.6m × 0.6m (i.e., 2.78 roots / m) 2 Furthermore, the length of the system anchor bolts was extended from 3.5m to 5.0m. This parameter, verified by FLAC3D inversion, can increase the cohesion of the surrounding rock anchoring zone by 35%. (2) Thickened shotcrete: An additional 8cm of steel fiber reinforced concrete is sprayed on top of the original design thickness of 20cm (reaching a total thickness of 28cm). FLAC3D verification shows that the flexural section modulus of the support structure can be increased by 42%, raising the overall safety factor from... Upgraded to .

[0048] In one embodiment, a lightweight 3D digital twin platform is built based on WebGL technology and the Unity3D engine. The platform connects to a cloud database in real time via a RESTful API interface, integrating sensor monitoring and LSTM prediction data. The values ​​are mapped to nodes of the 3D tunnel BIM / FLAC3D model, and the stress evolution and risk distribution of the support structure of each ring segment of the tunnel are dynamically displayed in real time using blue / yellow / orange / red color rendering.

[0049] After reinforcement measures are applied, the system initiates a 72-hour period of high-frequency encrypted monitoring (sampling frequency increased to 2Hz); if the monitoring data shows that the displacement change rate decreases to... Furthermore, the stress growth rate approaches 0, and the LSTM model predicts the next 7 days. If the red / orange warning is lifted, the system will automatically de-escalate the alert to a blue normal state.

[0050] The above steps enable real-time monitoring and intelligent early warning of the initial support status of tunnels, establishing a mechanical safety mirror of the support structure, and completing closed-loop management including risk identification, trend prediction, and disaster prevention. This method eliminates the need for traditional, cumbersome manual extrapolation, offers accurate calculations, and provides rapid response.

Claims

1. A method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction, characterized in that, include: Step 1: Acquire multi-dimensional monitoring data of surrounding rock pressure, grid steel frame stress, and support structure displacement collected in real time by the multi-source sensor monitoring network; Step 2: Based on the multidimensional monitoring data and the physical and mechanical parameters of the surrounding rock, perform FLAC3D numerical simulation to calculate the danger threshold of the support structure under different working conditions; Step 3: Input the multidimensional monitoring data into the pre-trained LSTM neural network model and output the predicted trend of the evolution of the initial support status; Step 4: Match the predicted evolution trend with the danger threshold and output the safety warning level.

2. The method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 1, characterized in that, The multi-source sensing monitoring network includes pressure sensors, stress sensors, and displacement sensors.

3. The method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 1, characterized in that, The recalculation trigger condition for the danger threshold is: the tunnel face excavation distance exceeds the preset range, or it crosses the geological boundary of different surrounding rock levels, or the displacement change rate monitored in real time on site exceeds the preset range. During recalculation, real-time on-site monitoring data is input into FLAC3D to correct the surrounding rock mechanical parameters.

4. The method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 1, characterized in that, The security warning levels include blue, yellow, orange, and red warnings; Blue alert corresponds to Yellow alert corresponds to Orange alert corresponds to Red alert corresponds to ;in, The comprehensive risk index is determined based on the matching analysis results of LSTM prediction values ​​and FLAC3D hazard thresholds.

5. The method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 1, characterized in that, The input features of the LSTM neural network model include the rate of displacement change, the rate of increase of stress, and the surrounding rock grade; the LSTM neural network model contains LSTM stacked layers and an attention mechanism layer.

6. The method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 1, characterized in that, The formula for calculating the comprehensive risk index is as follows: ， in, The displacement value is predicted by LSTM. The stress value predicted by LSTM. The critical threshold for limit displacement determined by FLAC3D calculation. Ultimate stress threshold determined by FLAC3D calculation For displacement offset rate, For stress offset rate, and These are weighting coefficients; they are dynamically adjusted based on the surrounding rock conditions. and The value of , and ;according to The preset range in which the value is located will output the corresponding warning level.

7. The method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 1, characterized in that, Following step 1, the process further includes: preprocessing the multidimensional monitoring data, including: using a median filtering algorithm to eliminate transient noise pulses caused by blasting vibrations; and using... Outliers are identified and removed, and then filled using the neighborhood mean. Min-Max normalization is performed on the data after outlier removal.

8. The method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 1, characterized in that, The prediction time step of the LSTM neural network model This corresponds to the evolution trend over the next 1 to 7 days.

9. A method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 4, characterized in that, The specific triggering logic for the security warning level is as follows: Blue warning corresponds to... And the offset rate is less than 60%; a yellow alert is corresponding to... And 60% ≤ offset rate < 80%; Orange alert corresponds to And 80% ≤ offset rate < 90%; Red alert corresponds to Either the offset rate is ≥90%, or the displacement rate is greater than or equal to 10 mm / d for several consecutive days.

10. A method for early warning of tunnel initial support safety based on multi-source monitoring fusion and time-series prediction according to claim 1, characterized in that, Based on the safety warning level, corresponding construction auxiliary decisions are executed: under blue and yellow warnings, control parameters are automatically updated and recorded in the background log; under orange warnings, a support reinforcement plan is generated and sent to the project chief engineer's terminal. Under a red alert, an audible and visual alarm is triggered, and a shutdown command is issued.

Images