Active prevention and control method and system for high-stress tunnel rockburst section

Abstract

The invention provides an active prevention and control method and system for a high-stress tunnel rockburst section, and relates to the field of disaster prevention and control. The method comprises the following steps: acquiring a multi-source monitoring data set, identifying a rock mass surface temperature abnormal region according to thermal imaging data, and marking a high-stress risk point; and performing feature level fusion on the multi-source monitoring data set, extracting space-time correlation features, inputting the space-time correlation features into a physical information embedded neural network, and generating an initial crustal stress field inversion result. And analyzing a historical stress evolution law by using a space-time diagram convolutional network, predicting future window crustal stress field distribution, correcting a stress prediction result, and outputting a dynamically updated stress concentration coefficient and a risk early warning level. And dynamically adjusting a rockburst judgment threshold value based on the stress concentration coefficient, calculating a comprehensive risk index, and judging the occurrence probability of rockburst. And inputting the risk early warning grade and the rockburst occurrence probability into a multi-objective optimization model to generate an optimal regulation and control scheme. The technical problems of incomplete monitoring data coverage and lack of physical interpretability in crustal stress inversion and prediction are solved.

Description

Technical Field

[0001] This application relates to the field of disaster prevention and control, and in particular to an active prevention method and system for rockburst sections in high-stress tunnels. Background Technology

[0002] In high-stress scenarios such as deep-buried tunnels, mining operations, and hydropower projects, rockburst hazards are a core risk factor threatening construction safety and hindering project progress. Accurately acquiring the distribution of the geostress field, predicting stress evolution trends, and implementing scientific prevention and control are crucial prerequisites for rockburst hazard management. Existing methods for high-stress tunnel rockburst sections mostly rely on single-type equipment such as traditional microseismic sensors, stress gauges, and convergence meters. In some scenarios, manual drilling and conventional numerical simulation are used for monitoring. This results in insufficient overall monitoring coverage, weak resistance to electromagnetic interference, and microseismic positioning accuracy of only ±10m. Monitoring of surrounding rock deformation is lagging, and manual screening of risk points is inefficient, prone to omissions, and unable to provide comprehensive and accurate multi-dimensional basic data. Furthermore, when using traditional techniques such as hydraulic fracturing and stress relief methods, or AI models based on single data sources for data processing and judgment, the AI ​​inversion model exhibits certain black-box characteristics, leading to large errors in the magnitude and direction of geostress inversion, and a lagging inversion process, failing to meet the needs of real-time dynamic analysis. Summary of the Invention

[0003] This application provides an active prevention method and system for rockburst sections in high-stress tunnels, which solves the technical problems of incomplete monitoring data coverage, weak anti-interference ability, insufficient accuracy, lack of multi-source data fusion, and lack of physical interpretability, accuracy and foresight in geostress inversion and prediction.

[0004] To achieve the above objectives, this application adopts the following technical solution: Firstly, a proactive prevention method for rockburst sections in high-stress tunnels includes: collecting multi-source monitoring datasets and identifying abnormal temperature zones on the rock mass surface based on thermal imaging data, marking high-stress risk points. The multi-source monitoring datasets include acoustic signals, strain data, image features, and thermal imaging data. Feature-level fusion is performed on the multi-source monitoring datasets, spatiotemporal correlation features are extracted and input into a neural network based on physical information, and combined with the rock mass constitutive relationship and energy conservation equation to generate an initial geostress field inversion result. Based on the initial geostress field inversion result, a spatiotemporal graph convolutional network is used to analyze historical stress evolution patterns, predict the distribution of the geostress field in future windows, and, combined with advanced borehole radar detection data, correct the stress prediction results, outputting dynamically updated stress concentration coefficients and risk warning levels. The rockburst discrimination threshold is dynamically adjusted based on the stress concentration coefficient, and a comprehensive risk index is calculated by combining the microseismic energy accumulation index and the surrounding rock deformation rate to determine the probability of rockburst occurrence. The risk warning level and rockburst probability are input into a multi-objective optimization model, and, combined with decompression costs, support costs, and project progress, an optimal control scheme is generated.

[0005] Based on the above technical solutions, the active prevention method for rockburst sections in high-stress tunnels provided in this application constructs a multi-source monitoring system and combines it with infrared thermal imaging to accurately identify rock mass temperature anomaly zones. This not only improves the efficiency of locating high-stress risk points but also solves the problems of insufficient coverage of traditional single monitoring and high false negative rates in manual screening. Simultaneously, by extracting spatiotemporal correlation features through feature-level fusion and using a PINN network embedded with engineering physics equations to invert the initial geostress field, the method balances model physical interpretability and data coupling, reducing geostress inversion errors and directional errors, thus overcoming the bottlenecks of traditional AI inversion's black box nature and insufficient accuracy. Furthermore, by utilizing ST-GCN to mine historical stress evolution patterns and integrating advanced borehole radar data to correct prediction results, the accuracy of stress prediction is improved, effectively solving the problems of traditional prediction lag and lack of deep data support, allowing sufficient time for prevention and control during construction. At the same time, adjusting the rockburst discrimination threshold based on the dynamic stress concentration coefficient and combining it with a multi-index calculation of the comprehensive risk index also achieves precise discrimination based on the specific rock mass, reducing the risk of false positives and false negatives in rockburst detection. Finally, by balancing cost and schedule through a multi-objective optimization model, the optimal control scheme is generated. This avoids the drawbacks of traditional experience-based control, improves construction efficiency, reduces the rock burst rate, resolves the contradiction between safety control and cost and schedule, and forms a closed-loop intelligent control system for the entire process, providing an efficient and safe intelligent solution for high ground stress tunnel construction.

[0006] In conjunction with the first aspect mentioned above, one possible implementation involves feature-level fusion of multi-source monitoring datasets to extract spatiotemporal correlation features. Specifically, this includes: analyzing acoustic signals, strain data, image features, and thermal imaging data from the multi-source monitoring datasets and establishing a four-dimensional spatiotemporal coordinate system; performing synchronous registration through timestamp alignment and spatial coordinate transformation; extracting time-frequency domain features from the registered acoustic signals to generate an acoustic energy distribution feature map; extracting strain abrupt change features from the registered strain data to obtain the location signal of the stress concentration area; extracting the geometric features of the borehole wall fractures from the registered image features to obtain fracture density, opening, and azimuth data; and extracting temperature gradient features from the registered thermal imaging data to obtain the contour of the rock mass surface temperature anomaly area. A multi-head attention mechanism is then constructed to weightedly fuse the acoustic energy distribution feature map, the stress concentration area location signal, the fracture density, opening, and azimuth data, and the contour of the rock mass surface temperature anomaly area to generate spatiotemporal correlation features.

[0007] In conjunction with the first aspect mentioned above, one possible implementation involves generating the initial geostress field inversion result by combining the rock mass constitutive relation and the energy conservation equation. Specifically, this includes: constructing a neural network model, which comprises a feature encoding layer, a physical equation solving layer, and a residual optimization layer. Spatiotemporal correlation features are input into the feature encoding layer of the neural network model to obtain a high-dimensional feature vector, which is then fed back to the physical equation solving layer. The physical equation solving layer uses the rock mechanics governing equations as physical constraints and calculates the predicted stress and strain fields output by the network through automatic differentiation techniques. The residual optimization layer calculates the fitting loss based on the predicted stress and strain fields and iterates the network parameters of the neural network model using a backpropagation algorithm to generate the initial geostress field inversion result.

[0008] In conjunction with the first aspect mentioned above, one possible implementation involves using a spatiotemporal graph convolutional network to analyze historical stress evolution patterns and predict future stress field distribution based on the initial stress field inversion results. Specifically, this includes: constructing a spatiotemporal graph convolutional network to map the initial stress field inversion results into high-dimensional feature vectors and converting advanced borehole radar data into binary mask images; acquiring historical stress window sequences and corresponding time-level mask images to train the spatiotemporal graph convolutional network; receiving real-time multi-source monitoring datasets and dynamically adjusting spatiotemporal weights through an attention mechanism to output the stress field prediction results for the future window.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves using advanced borehole radar detection data to correct stress prediction results and output dynamically updated stress concentration coefficients and risk warning levels. Specifically, this includes: calculating a stress concentration coefficient correction factor for the mask region based on the mask image and updating the stress concentration coefficient; and classifying the risk level based on the updated stress concentration coefficient.

[0010] In conjunction with the first aspect mentioned above, one possible implementation involves dynamically adjusting the rockburst discrimination threshold based on the stress concentration factor, and calculating a comprehensive risk index by combining the microseismic energy accumulation index and the surrounding rock deformation rate to determine the probability of a rockburst. Specifically, this process includes: acquiring the energy accumulation index from the microseismic monitoring system, inverting the source location using a three-dimensional velocity model, and calculating the accumulated elastic strain energy per unit volume of rock mass; acquiring the radial deformation rate of the surrounding rock through a distributed fiber optic strain monitoring system and calculating the average deformation rate; constructing a three-level risk assessment matrix, and then weighting and fusing the stress concentration factor, elastic strain energy, and average deformation rate after normalization to calculate the comprehensive risk index; and dynamically determining the probability of a rockburst based on the comprehensive risk index.

[0011] In conjunction with the first aspect mentioned above, one possible implementation involves inputting the risk warning level and rockburst probability into a multi-objective optimization model. This model, combined with decompression cost, support cost, and project schedule, generates the optimal control scheme. Specifically, this includes: constructing a multi-objective optimization model based on the NSGA-II algorithm. The model's optimization objectives are decompression cost, support cost, and project schedule delay rate, constrained by safety thresholds, resource consumption limits, and construction disturbance intensity. Real-time data streams of risk warning level, rockburst probability, stress concentration coefficient dynamic correction factor, and surrounding rock deformation rate are input into the multi-objective optimization model. An initial Pareto solution set is generated using a multi-objective genetic algorithm, and the scheme weights are adjusted. The control combination with the highest overall utility is selected as the optimal control scheme, and this is distributed in real-time to the intelligent decompression robot and shape memory alloy support system via edge computing nodes for execution. The optimal control scheme includes decompression parameters and support parameters.

[0012] Secondly, an active prevention system for rockburst sections in high-stress tunnels is provided, comprising: a multi-source data acquisition module, a feature-level fusion module, a physical information embedded neural network module, a spatiotemporal graph convolutional network module, a dynamic risk discrimination module, and a multi-objective optimization decision-making module. The multi-source data acquisition module collects multi-source monitoring datasets and performs synchronous preprocessing to establish a four-dimensional spatiotemporal coordinate system, obtaining a registered multimodal monitoring dataset. The feature-level fusion module fuses the multimodal monitoring datasets through a multimodal data fusion mechanism to generate a spatiotemporal correlation feature matrix. The physical information embedded neural network module performs calculations and inversions based on the spatiotemporal correlation feature matrix to obtain the initial geostress field inversion result. The spatiotemporal graph convolutional network module calls the advanced borehole radar mask image and combines it with the initial geostress field inversion result to predict the future geostress field and obtain the predicted result and correction factor. The dynamic risk discrimination module calculates a comprehensive risk index based on the corrected stress concentration coefficient, elastic strain energy, and average deformation rate, outputting the probability of rockburst occurrence and risk level. The multi-objective optimization decision-making module generates the optimal control scheme based on risk level, rock burst probability, correction factor, and surrounding rock deformation rate, and then transmits it to the execution layer for execution.

[0013] In conjunction with the second aspect mentioned above, in one possible implementation, the feature-level fusion module specifically includes a multimodal data synchronization and registration unit, a cross-modal feature extraction unit, and a multi-head attention fusion unit: the multimodal data synchronization and registration unit performs time synchronization on the multimodal monitoring dataset, followed by coordinate transformation and preprocessing to obtain a four-dimensional spatiotemporal registration dataset. The cross-modal feature extraction unit extracts features from the four-dimensional spatiotemporal registration dataset, generating a multi-dimensional feature matrix. The multi-head attention fusion unit calls the multi-dimensional feature matrix to perform multi-head attention calculation, outputting a spatiotemporal correlation feature vector.

[0014] In conjunction with the second aspect mentioned above, in one possible implementation, the physical information embedded neural network module specifically includes a physical information encoding layer, a physical equation solving layer, and a residual optimization layer. The physical information encoding layer extracts features based on spatiotemporal correlation feature vectors and separates subspaces using a feature decoupler to obtain multi-physics coupling features. The physical equation solving layer calculates the stress balance equation residuals from the multi-physics coupling features and, combined with the Hawke-Brown criterion and energy conservation equations, obtains the initial geostress field that satisfies physical constraints. The residual optimization layer calculates the composite loss function based on the stress balance equation residuals and data fitting loss, and iterates to generate the inversion result.

[0015] This application provides an active prevention method and system for rockburst sections in high-stress tunnels. By constructing a multi-source monitoring system and combining it with infrared thermal imaging, it accurately identifies rock mass temperature anomaly zones, improving the efficiency of locating high-stress risk points and solving the problems of insufficient coverage and high false negative rates in traditional single-monitoring and manual screening. Simultaneously, it extracts spatiotemporal correlation features through feature-level fusion and uses a PINN network embedded with engineering physics equations to invert the initial geostress field, balancing model physical interpretability and data coupling, reducing geostress inversion errors and directional errors, thus overcoming the bottlenecks of traditional AI inversion's black box nature and insufficient accuracy. Furthermore, it utilizes ST-GCN to mine historical stress evolution patterns and integrates advanced borehole radar data to correct prediction results, improving stress prediction accuracy and effectively addressing the problems of traditional prediction lag and lack of deep data support, allowing sufficient time for prevention and control during construction. At the same time, it adjusts the rockburst discrimination threshold based on the dynamic stress concentration coefficient and combines it with a multi-index calculation structure to achieve precise discrimination based on the specific rock mass, reducing the risk of false positives and false negatives in rockburst detection. Finally, by balancing cost and schedule through a multi-objective optimization model, the optimal control scheme is generated. This avoids the drawbacks of traditional experience-based control, improves construction efficiency, reduces the rock burst rate, resolves the contradiction between safety control and cost and schedule, and forms a closed-loop intelligent control system for the entire process, providing an efficient and safe intelligent solution for high ground stress tunnel construction.

[0016] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0017] Figure 1 A system architecture diagram of an active prevention method for rockburst sections in high-stress tunnels provided in this application embodiment; Figure 2 A flowchart illustrating an active prevention method for rockburst sections in high-stress tunnels, provided as an embodiment of this application; Figure 3 A flowchart illustrating an active prevention method for rockburst sections in high-stress tunnels, provided as an embodiment of this application; Figure 4 A flowchart illustrating an active prevention method for rockburst sections in high-stress tunnels, provided as an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an active prevention system for rockburst sections in high-stress tunnels, provided in an embodiment of this application. Detailed Implementation

[0018] In the description of this application, unless otherwise stated, "" means "or," for example, A / B can mean A or B. "And / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The words "first," "second," etc., do not limit the quantity or order of execution, and are not necessarily implied as unsuitable.

[0019] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0020] To address the shortcomings of existing technologies, such as reliance on a single type of equipment, limited data coverage, weak resistance to electromagnetic interference, microseismic positioning accuracy of only ±10m, significant lag in surrounding rock deformation monitoring, low efficiency and high likelihood of missed detections in manual risk point screening, and difficulty in obtaining comprehensive and accurate multi-source basic data.

[0021] However, the lack of effective multi-source data feature-level fusion methods in the data processing stage means that geostress inversion either relies on traditional hydraulic fracturing methods, stress relief methods, and other techniques, resulting in large errors (inversion errors of more than 8% in magnitude and 20°-30° in direction), or it uses AI models that do not incorporate engineering physical constraints such as rock mass constitutive relations and energy conservation, which have black box characteristics and the inversion process is lagging and cannot meet the needs of real-time dynamic analysis.

[0022] Meanwhile, stress prediction failed to fully explore historical stress evolution patterns and was not corrected by incorporating deep exploration data. The predicted results deviated significantly from the actual stress distribution, lacking sufficient foresight and accuracy, and making it difficult to allow sufficient time for prevention and control. Risk assessment, on the other hand, relied on fixed empirical criteria, failing to consider the brittle differences between different lithologies such as granite and marble, resulting in poor universality and a high rate of misjudgment and missed judgment.

[0023] Finally, the formulation of control schemes relies on manual experience, and the decompression and support parameters lack a dynamic adaptive adjustment mechanism. The failure to comprehensively balance decompression costs, support costs, and project progress easily leads to cost waste or inadequate safety control. Furthermore, the response is sluggish, failing to achieve rapid and efficient adaptive control, severely restricting construction safety and project efficiency. This application provides an active prevention method for rockburst sections in high-stress tunnels. This method constructs a multi-source monitoring system and combines it with infrared thermal imaging to accurately identify rock mass temperature anomaly zones, improving the efficiency of locating high-stress risk points and solving the problems of insufficient coverage and high miss rates in traditional single monitoring and manual screening. Simultaneously, by extracting spatiotemporal correlation features through feature-level fusion and using a PINN network embedded with engineering physics equations to invert the initial geostress field, it balances model physical interpretability and data coupling, reducing geostress inversion errors and directional errors, thus overcoming the bottlenecks of traditional AI inversion's black box nature and insufficient accuracy. By leveraging ST-GCN to uncover historical stress evolution patterns and integrating advanced borehole radar data to refine prediction results, the accuracy of stress prediction is improved. This effectively addresses the issues of traditional prediction lag and lack of deep data support, allowing sufficient time for prevention and control during construction. Simultaneously, adjusting the rockburst detection threshold based on the dynamic stress concentration coefficient and combining it with a multi-indicator comprehensive risk index structure enables precise identification of rockbursts on a case-by-case basis, reducing the risk of misjudgment or missed detection. Finally, a multi-objective optimization model balances cost and schedule to generate the optimal control scheme. This avoids the drawbacks of traditional experience-based control while improving construction efficiency and reducing the rockburst occurrence rate, resolving the conflict between safety control and cost / schedule. This forms a closed-loop intelligent control system throughout the entire process, providing an efficient and safe intelligent solution for high-stress tunnel construction.

[0024] like Figure 1 As shown in the embodiment of this application, an active prevention method for rockburst sections in high-stress tunnels includes: Step 101: Collect multi-source monitoring datasets and identify abnormal temperature areas on the rock surface based on thermal imaging data, and mark high stress risk points. The multi-source monitoring datasets include acoustic signals, strain data, image features and thermal imaging data.

[0025] Among these technologies, Distributed Fiber Optic Acoustic Sensing (DAS) detects acoustic signals generated by micro-fractures in surrounding rock using laser pulses, giving it a fully distributed and electromagnetic interference-resistant characteristic. Brillouin Optical Time Domain Reflectometry (BOTDR) measures the Brillouin frequency shift caused by fiber strain / temperature to invert surrounding rock deformation and stress distribution. A drilling robot equipped with a high-definition camera and ultrasonic probe can automatically drill detection holes along a preset path and transmit back images of borehole wall fractures and rock mass wave velocity data. Unmanned aerial vehicle (UAV)-borne infrared thermal imaging scans the temperature distribution on the tunnel wall surface to identify temperature anomaly zones formed by heat release from micro-fractures in the rock mass.

[0026] During tunnel face construction, a distributed fiber optic sensor network is first deployed along the tunnel axis to collect acoustic signals generated by micro-fractures in the surrounding rock in real time. Simultaneously, BOTDR technology is used to monitor strain and stress changes along the fiber optic line, obtaining data on deep deformation and stress distribution in the surrounding rock. This allows a drilling robot to automatically drill into the surrounding rock along a path planned in the BIM model. In-hole imaging and wave velocity testing are then used to obtain information on the degree of fracture development in the borehole wall and the integrity of the rock mass. Simultaneously, a drone equipped with an infrared thermal imager scans the tunnel cross-section, identifying localized stress concentration areas through changes in the rock surface temperature gradient, generating a distribution map of high-stress risk points.

[0027] Taking a deep-buried granite tunnel as an example: During excavation, DAS monitoring detected a sudden high-frequency micro-fracture signal (frequency 8kHz) 50m behind the tunnel face, and BOTDR simultaneously showed that the strain change in the surrounding rock in that area reached 0.5. The drilling robot then drilled a probe hole at the location and discovered tension cracks in the hole wall and an abnormally low core wave velocity. Infrared scanning by a drone showed that the temperature of the borehole wall in this area was 2.3°C higher than the surrounding area, leading to a comprehensive assessment that it was a high-stress concentration zone. Based on this, the system triggered an Intelligent Pressure Relief Robot (IPRR) to perform directional blasting pressure relief in the area, ultimately reducing the rockburst risk level from Level IV to Level II.

[0028] During practical use, DAS and BOTDR monitor the surrounding rock condition from acoustic and mechanical perspectives, respectively; while the drilling robot compensates for the low efficiency and high risk of manual drilling. Meanwhile, infrared thermal imaging provides a non-contact, rapid risk screening method. This enables multi-source sensors to achieve spatiotemporal coordination and data complementarity, solving the problems of insufficient coverage and poor anti-interference capabilities of single sensors in traditional monitoring.

[0029] Step 102: Perform feature-level fusion on the multi-source monitoring dataset, extract spatiotemporal correlation features and input physical information into the neural network, combine the rock mass constitutive relation and energy conservation equation to generate the initial geostress field inversion result.

[0030] Among them, Physical Information Embedded Neural Network (PINN) is a deep learning model that embeds rock mechanics governing equations (such as equilibrium equations and constitutive relations) as prior constraints into a neural network architecture. Rock mass constitutive relations are mathematical models describing the stress and strain behavior of surrounding rock (such as the Hawke-Brown criterion). The energy conservation equation is a physical law characterizing the energy transfer and dissipation of surrounding rock. .

[0031] In some implementations, a multi-source monitoring dataset is invoked, and a spatiotemporal registration algorithm is used to synchronize and align the DAS acoustic signals, BOTDR strain data, borehole robot borehole wall images, and infrared thermal imaging temperature data in time and space to eliminate sensor deployment bias. Then, a multi-scale feature extraction network (such as 3D-CNN for time-series acoustic signals, LSTM for strain time-series data, and CNN for image texture features) is used to extract features of each modality in parallel. The feature weights can then be dynamically weighted and fused using a cross-modal attention mechanism, allowing the fused high-dimensional features to be embedded into a neural network (PINN) by inputting physical information. At this point, the network has an built-in automatic differentiation module for the rock mass constitutive equation and energy conservation equation, calculating the physical equation residuals while inverting the geostress field. This is achieved by minimizing the weighted sum of data fitting loss and physical constraint loss. This generates an initial geostress field distribution that satisfies the laws of mechanics.

[0032] It should be noted that the PINN network calculates the residuals of the physical equations of the stress field in real time during the inversion process (e.g., stress balance error <5%), and optimizes the network parameters through backpropagation, so that the inversion results are consistent with both the monitoring data and the basic laws of rock mechanics. The multimodal fusion network dynamically adjusts the weights of different modal data through an attention mechanism (e.g., increasing the weight of infrared thermal imaging in brittle rock masses), effectively suppressing noise interference (e.g., DAS noise caused by blasting vibration is eliminated by LSTM time-series filtering).

[0033] Taking a deep-buried marble tunnel as an example: if DAS detects a sudden high-frequency micro-fracture signal (frequency 5kHz) 100m behind the tunnel face, BOTDR shows that the strain change in the surrounding rock in that area reaches 0.8. Images of the borehole wall captured by the drilling robot showed a 30% increase in fracture density, and infrared thermal imaging revealed an abnormal temperature increase of 2.8°C on the corresponding borehole wall. A multimodal fusion network extracted transient energy features from the acoustic signal, abrupt changes in strain data, fracture geometry from the image, and temperature gradient features from the thermal imaging. Through a cross-modal attention mechanism, higher weights were assigned to the thermal imaging features (weighting coefficient 0.45). The final PINN inversion results showed that the direction of the maximum principal stress in the area deviated from the initial prediction by 12°, and the stress concentration factor was corrected from 0.52 to 0.61, triggering a high-risk rockburst warning and initiating a directional water injection operation using an intelligent decompression robot.

[0034] Step 103: Based on the initial geostress field inversion results, analyze the historical stress evolution law using a spatiotemporal graph convolutional network, predict the geostress field distribution in the future window, and combine it with advanced borehole radar detection data to correct the stress prediction results and output dynamically updated stress concentration coefficients and risk warning levels.

[0035] In some implementations, a spatiotemporal graph convolutional network (ST-GCN) is constructed based on the initial results of the geostress field. Historical stress monitoring data (DAS / BOTDR) is mapped into high-dimensional feature vectors through node embedding layers. The spatiotemporal graph convolutional layers capture the correlation features between stress spatial distribution and temporal series. This is then combined with data on anomalies in surrounding rock wave velocity and fracture development detected by advanced borehole radar to dynamically adjust the spatiotemporal weight parameters of the ST-GCN. Finally, the predicted geostress field for the next hour is output, and the result is expressed using a formula. Calculate the dynamically updated stress concentration factor (where For strain sensitivity coefficient, For deformation rate), based on the threshold table (brittle rock mass) (≥0.55 triggers a high-risk warning) Generates a real-time risk warning level.

[0036] During the construction of a deep-buried sandstone tunnel, ST-GCN analysis of DAS monitoring data over the past 3 hours revealed an exponential increase in high-frequency micro-fracture energy in the right-side region of the tunnel face. Combined with a low-velocity anomaly zone detected by advanced radar 20m ahead (presumably a fault fracture zone), the spatiotemporal map convolution kernel parameters could be dynamically adjusted. Calculations showed that the maximum principal stress in this area would increase from 28MPa to 35MPa in the next hour, with the stress concentration factor jumping from 0.58 to 0.63. This triggered a Level IV high-risk warning and automatically activated an intelligent decompression robot to implement directional blasting, ultimately controlling the actual monitored stress concentration factor below 0.52.

[0037] The process of combining advanced borehole radar detection data to correct stress prediction results and output dynamically updated stress concentration coefficients and risk warning levels specifically includes: calculating the stress concentration coefficient correction factor for the mask region based on the mask image and updating the stress concentration coefficient; and classifying the risk level according to the updated stress concentration coefficient.

[0038] During its use, the drilling robot drills probe holes within a 50m range in front of the tunnel face according to the BIM-planned path. It uses drilling radar to scan the surrounding rock wave velocity distribution in real time, generating a mask map to mark abnormal areas. Then, the area ratio of the red region in the mask map (MaskArea / TotalArea) is calculated and substituted into the correction factor formula. Calculate the correction factor Therefore, the initial inversion results of the geostress field can be directly input into the ST-GCN model, and the spatiotemporal weight parameter wkr can be dynamically adjusted by the correction factor η to output the stress field prediction result σpred for the next hour. Finally, based on the updated stress concentration factor formula... (Where α is the attenuation coefficient and t is the distance from the working face) Recalculate the risk level. Therefore, when When <0.45, it is determined to be a Level I safety state, and routine monitoring should be maintained; when 0.45 ≤ A Level II warning is triggered when the value is <0.55, and the monitoring frequency is increased; when 0.55 ≤ When <0.65, initiate a Level III response and implement pre-depressurization; when If the value is ≥0.65, it is classified as Level IV, extremely high risk. Excavation is suspended and comprehensive prevention and control measures are initiated.

[0039] Step 104: Dynamically adjust the rockburst discrimination threshold based on the stress concentration coefficient, and calculate the comprehensive risk index by combining the microseismic energy accumulation index and the surrounding rock deformation rate to determine the probability of rockburst occurrence.

[0040] Among them, the rockburst detection threshold is the critical value for determining the risk of rockburst occurrence. The microseismic energy accumulation index is a quantitative indicator reflecting the degree of energy accumulation within the rock mass.

[0041] In some implementations, the discrimination threshold is dynamically adjusted based on the stress concentration coefficient and lithology (identified through core sampling by a borehole robot). This allows for the fusion of the energy accumulation index (EAI) from microseismic monitoring and the surrounding rock deformation rate from BOTDR monitoring. ), through weighted formula Calculate the comprehensive risk index when A value greater than 0.8 indicates an extremely high risk of rockburst.

[0042] The process of adjusting the discrimination threshold includes: lowering the threshold for brittle rock masses (such as granite) by 5%-10% (e.g., adjusting the original threshold from 0.55 to 0.50), and raising the threshold for ductile rock masses (such as marble) by 5%-10% (e.g., adjusting the original threshold from 0.55 to 0.60).

[0043] In a deep-buried quartz vein tunnel, when the stress concentration factor... When the RI = 0.58, due to the brittle nature of the quartz vein, the threshold can be lowered from 0.55 to 0.53, triggering an extremely high rockburst risk (RI = 0.82). At this point, the intelligent depressurization robot should be immediately activated to inject high-pressure water (15 MPa) to... The energy density was reduced to 0.48, and the microseismic energy accumulation index decreased by 60%, ultimately preventing a rockburst accident.

[0044] Step 105: Input the risk warning level and the probability of rock burst occurrence into the multi-objective optimization model, and generate the optimal control scheme by combining the pressure relief cost, support cost and project progress.

[0045] In some implementations, the stress concentration factor and comprehensive risk index are input into a multi-objective optimization model (NSGA-II algorithm). In this case, the multi-objective optimization model uses the decompression cost (yuan / cubic meter), support cost (yuan / linear meter), and project schedule delay rate (%) as optimization objectives, with a safety threshold (…). With a resource consumption limit of <0.55 (daily blasting charge ≤50kg) as a constraint, multiple feasible schemes are generated through Pareto front analysis. A reinforcement learning algorithm (PPO policy network) is then used, combined with historical decision-making effects (e.g., energy decay rate >80% after decompression is considered effective), to dynamically adjust the scheme weights and ultimately select the control combination with the highest overall utility (e.g., deep-hole blasting + high-strength support + 2-hour suspension of excavation).

[0046] The process of inputting risk warning level and rockburst probability into a multi-objective optimization model, and combining pressure relief cost, support cost, and project schedule to generate the optimal control scheme includes: constructing a multi-objective optimization model based on the NSGA-II algorithm, with pressure relief cost, support cost, and project schedule delay rate as optimization objectives, and safety threshold, resource consumption limit, and construction disturbance intensity as constraints; inputting real-time data streams of risk warning level, rockburst probability, stress concentration coefficient dynamic correction factor, and surrounding rock deformation rate into the multi-objective optimization model, generating an initial Pareto solution set through a multi-objective genetic algorithm, and adjusting the scheme weights; The optimal control scheme is selected based on the highest overall effectiveness and is then sent to the intelligent decompression robot and shape memory alloy support system in real time via edge computing nodes. The optimal control scheme includes decompression parameters and support parameters.

[0047] The NSGA-II algorithm is a multi-objective genetic algorithm based on non-dominated sorting that maintains population diversity through crowding distance calculation. It is used to generate multi-objective optimization solution sets for decompression cost, support cost, and project schedule delay rate. The PPO policy network is a policy network constructed using the Proximal Policy Optimization reinforcement learning algorithm, which optimizes the selection policy of the control scheme through the approximate policy gradient method.

[0048] In practical engineering applications, when the risk warning level reaches Level IV (high risk) and the probability of rockburst exceeds 80%, the system automatically triggers a multi-objective optimization process. First, based on the stress field data predicted by ST-GCN for a future window, the stress concentration coefficient correction factor η for each potential stress relief zone is calculated. Then, combined with the surrounding rock wave velocity data transmitted in real-time by the drilling robot (where the longitudinal wave velocity vp < 3500 m / s indicates a fractured zone), the stress relief hole depth (L = 1.8 × tunnel diameter) and spacing (D = 1.2 × tunnel diameter) are dynamically adjusted. Finally, the PPO strategy network is used to calculate the combined effectiveness of blasting stress relief (charge Q = 0.4 kg / m) and high-pressure water injection (pressure P = 15 MPa) in the current scenario, based on a historical scheme database (containing 200 control cases with different lithological combinations). Finally, the solution with the highest overall utility value (such as "drilling Φ50mm@1.2m quincunx-shaped boreholes 5m in front of the tunnel face, alternately injecting 15MPa high-pressure water and 0.3kg bentonite gel") is transmitted to the IPRR execution terminal via a 5G private network through edge computing nodes. At the same time, the self-reset mode of the SMA support system is activated (stiffness switching time <200ms). The entire process, from risk identification to solution execution, takes no more than 3 minutes, which is 15 times more efficient than traditional manual decision-making.

[0049] It is important to note that in high-stress soft rock sections (such as silty mudstone), the inclination angle of the pressure relief hole should be adjusted to a 15° elevation angle to prevent hole collapse. When the deformation rate of the surrounding rock exceeds 0.3% / min, the secondary response plan should be automatically activated (the activation threshold of the SMA anchor bolt pressure relief mode should be lowered to 3mm). Additionally, during high-temperature periods in summer (air temperature > 35℃), the high-pressure injection water temperature should be controlled at 18±2℃ to prevent thermal damage to the rock mass.

[0050] Taking the construction of a deep-buried granite tunnel (980m depth, σc=180MPa) as an example, when the tunnel face advanced to K12+456, the system monitored a jump in stress concentration coefficient to 0.62 and a microseismic energy accumulation index of 95J. After 127 iterations, the multi-objective optimization model generated the optimal solution: implementing "three-hole simultaneous unloading" (hole depth 24m, spacing 1.5m, explosive dosage 0.8kg / hole) 8m ahead of the tunnel face, combined with dynamic adjustment of SMA anchor stiffness (initial stiffness 65GPa → 42GPa after pressure relief). Actual construction showed that this solution reduced the stress concentration coefficient to 0.47 and controlled the surrounding rock deformation rate to within 0.12% / min, saving construction time and reducing overall costs compared to traditional pressure relief methods.

[0051] Based on the above technical solutions, by constructing a multi-source monitoring system and combining it with infrared thermal imaging to accurately identify rock mass temperature anomaly zones, the efficiency of locating high-stress risk points is improved, while also solving the problems of insufficient coverage of traditional single monitoring and high false negative rates in manual screening. Simultaneously, by extracting spatiotemporal correlation features through feature-level fusion and using a PINN network embedded with engineering physics equations to invert the initial geostress field, the system balances model physical interpretability and data coupling, reducing geostress inversion errors and directional errors, thus overcoming the bottlenecks of traditional AI inversion's black box nature and insufficient accuracy. Furthermore, by utilizing ST-GCN to mine historical stress evolution patterns and integrating advanced borehole radar data to correct prediction results, the accuracy of stress prediction is improved, effectively solving the problems of traditional prediction lag and lack of deep data support, allowing sufficient time for prevention and control during construction. At the same time, by adjusting the rockburst discrimination threshold based on the dynamic stress concentration coefficient and combining it with a multi-indicator structure to calculate a comprehensive risk index, precise discrimination based on the specific rock mass is achieved, reducing the risk of false or missed rockburst detections. Finally, by balancing cost and schedule through a multi-objective optimization model, the optimal control scheme is generated. This avoids the drawbacks of traditional experience-based control, improves construction efficiency, reduces the rock burst rate, resolves the contradiction between safety control and cost and schedule, and forms a closed-loop intelligent control system for the entire process, providing an efficient and safe intelligent solution for high ground stress tunnel construction.

[0052] In one possible implementation of the embodiments of this application, combined with Figure 1-2 As shown, the process of performing feature-level fusion on multi-source monitoring datasets and extracting spatiotemporal correlation features can be achieved through the following steps 201 to 206, which are explained in detail below: Step 201: Analyze the acoustic signals, strain data, image features and thermal imaging data in the multi-source monitoring data set and establish a four-dimensional spatiotemporal coordinate system. Perform synchronous registration through timestamp alignment and spatial coordinate transformation.

[0053] The four-dimensional spatiotemporal coordinate system refers to a unified data benchmark framework composed of a time axis (microseconds to hours) and three-dimensional spatial coordinates (X / Y / Z axes, millimeters to hundreds of meters), used to synchronize the time reference and spatial position relationship of multi-source sensor data. Spatial coordinate transformation maps the heterogeneous coordinate systems of multi-source data, such as distributed fiber optic sensor networks (DAS / BOTDR), drilling robots (polar coordinate system), and UAV-borne infrared thermal imaging (geodetic coordinate system), to the global coordinate system of the tunnel construction BIM model.

[0054] In some implementations, a four-dimensional spatiotemporal coordinate system is established with the tunnel axis as the Z-axis, the horizontal direction as the X / Y plane, and the time axis as the T-axis. A global positioning reference is achieved through the GNSS / BeiDou system and the inertial navigation module (INS). This allows the DAS and BOTDR to have built-in high-precision clock modules, aligning with the monitoring center's master clock via the IEEE 1588 Precision Time Protocol (PTP). Furthermore, the atomic clocks on the drilling robot and the UAV achieve sub-millisecond time anchoring via a wireless synchronization protocol. This allows for direct use of LiDAR point cloud registration technology to transform the 3D point cloud (polar coordinate system) of the borehole wall collected by the drilling robot to the tunnel BIM model (Cartesian coordinate system) using the ICP algorithm. The UAV's infrared thermal imaging data is then used to generate a DSM (Digital Surface Model) through the SFM (Structure of Motion Reconstruction) algorithm, which is then geometrically registered with the BIM model.

[0055] Therefore, wavelet transform is used to denoise the DAS acoustic signal, retaining the 0.5-2kHz rock micro-fracture frequency band. BOTDR strain data is filtered by Savitzky-Golay to eliminate temperature interference, and the surrounding rock stress tensor is inverted using the elasticity equation. The borehole robot images are used to extract the fracture network topology using an edge detection algorithm, and the rock integrity coefficient is calculated in conjunction with wave velocity test data. Finally, infrared thermal imaging data is used to mark temperature anomaly areas through threshold segmentation (ΔT > 2℃). Thus, spatiotemporal interpolation algorithms (such as Kriging interpolation) are used to convert discrete sensor data into a continuous four-dimensional spatiotemporal field.

[0056] Taking a deep-buried granite tunnel as an example, when construction reached kilometer marker K12+345 at the tunnel face: DAS detected a sudden high-frequency acoustic signal (dominant frequency 8kHz, peak energy 0.3V) 8m behind the working face. 2 (Hz), BOTDR simultaneously showed that the strain abrupt change in this region reached 0.6. ; The drilling robot drilled a probe hole according to the path planned by BIM (30° off the centerline of the working face, 15° inclination). When the hole was 6m deep, the ultrasonic probe detected an abnormal decrease in wave velocity (the longitudinal wave velocity dropped from 4500m / s to 3800m / s). The drone, equipped with an infrared thermal imager, scanned the tunnel cross-section and found that the temperature of the corresponding tunnel wall was 2.7°C higher than the surrounding area, marking it as a high-risk area; The four-dimensional spatiotemporal coordinate system unifies the above data to the kilometer point K12+345.087 (X=2345678.123m, Y=5432109.876m, Z=1234.567m, T=2025-12-3010:45:23), and generates a 5m×5m×1m gridded stress field model after spatiotemporal interpolation; Ultimately, the Intelligent Pressure Relief Robot (IPRR) was triggered to carry out directional blasting in the area (hole depth 18m, spacing 1.2m), reducing the stress concentration factor from 0.61 to 0.49 and avoiding rock bursts.

[0057] Step 202: Extract the time-frequency domain features from the registered acoustic signal to generate an acoustic energy distribution feature map.

[0058] In some implementations, an improved wavelet thresholding denoising algorithm (soft thresholding function combined with Bayes contraction rule) is used to filter out blasting vibration noise (energy attenuation >30dB for components with frequencies >20kHz) from the registered DAS time-series signal, leaving only the 0.5-10kHz effective frequency band of rock mass micro-fractures. The time-domain signal is then converted into a time-frequency spectrum using a Short-Time Fourier Transform (STFT), and a Gaussian window function is set (window length 512 points, overlap rate 75%) to generate a time-frequency matrix with a time resolution of 125ms and a frequency resolution of 48Hz.

[0059] A 3D-CNN model can be used to extract deep time-frequency features: the input layer is a 128×128×10 time-frequency cube (time span 10s), which passes through four sets of 3D convolutional layers (kernel size 3×3×3, stride 1, number of channels 64→128→256→512 respectively) to extract multi-scale spatiotemporal features, while the pooling layer uses spatial pyramid pooling (SPP) to enhance scale invariance. Finally, a softmax classifier is used to generate an acoustic energy distribution feature map, where high-energy regions (pixel value > 0.8) correspond to the rapid development zone of rock mass damage, and low-energy regions (pixel value < 0.3) correspond to stable rock mass. This feature map can then be fused with strain time-series features and thermal imaging temperature gradient features through a cross-modal attention mechanism to form a comprehensive spatiotemporal correlation feature characterizing the state of the rock mass.

[0060] During the construction of a deep sandstone tunnel (850m depth, surrounding rock grade IV): DAS detected a sudden high-frequency acoustic signal (main frequency 1.2kHz, duration 80ms) 35m behind the working face. After wavelet denoising, the effective frequency band of 0.8-2.5kHz is extracted, and the STFT generates a time-frequency matrix showing that the energy peak is concentrated around 1.5kHz. The 3D-CNN model extracts the spatiotemporal features of the time-frequency cube, and the output energy distribution feature map shows that the energy concentration area extends 12m along the tunnel axis and has a lateral diffusion width of 3.5m. Combined with a strain mutation amount of 0.4 The temperature difference between the infrared thermal imaging and the temperature is 2.1℃, and the area is identified as a medium rockburst risk zone through spatiotemporal correlation analysis. The intelligent decision-making system triggered pre-drilling for pressure relief (hole depth 4.2m, spacing 1.5m), and the measured stress concentration factor decreased from 0.58 to 0.45, thus preventing a disaster.

[0061] Step 203: Extract strain abrupt change features from the registered strain data to obtain the stress concentration area location signal.

[0062] In some implementations, after applying a Savitzky-Golay filter (window length 51 points, polynomial order 3) to the registered BOTDR time-series strain data to eliminate temperature drift noise (signals below 0.001℃ / min are considered valid), the strain change rate is calculated using the sliding window difference method. A threshold α = 0.1% / min is set to identify abrupt change points, thus constructing an LSTM model to extract deep temporal features: the input layer is a 128-dimensional strain time series (window length 120min), which passes through two bidirectional LSTM layers (128 hidden units) to capture long-term dependencies. The fully connected layer outputs a mutation intensity score (0-1). Then, the DBSCAN clustering algorithm (neighborhood radius 0.5m, minimum sample size 5) generates spatially continuous strain mutation regions. Combined with the elasticity constitutive equation, the signal is reconstructed, ultimately generating a stress concentration region location signal (location error < 0.5m, response time ≤ 10s). At this point, this signal, along with the acoustic energy distribution feature map and the thermal imaging temperature gradient feature, is analyzed through spatiotemporal correlation to form a multi-dimensional risk criterion.

[0063] During the construction of a deep-buried marble tunnel (720m depth, Class III surrounding rock): BOTDR monitoring showed a sudden increase in strain rate to 0.18% / min at a depth of 42m behind the working face; The effective signal was extracted by Savitzky-Golay filtering, and the LSTM model output a mutation intensity score of 0.89. DBSCAN clustering identified three consecutive strain abrupt change regions (X=2345678.345m, Y=5432109.678m, Z=1234.567m-1237.890m). Based on the high-energy area (energy > 0.7) shown in the acoustic energy map and the temperature difference of 2.3℃ in the infrared thermal imaging, it was determined to be a Class IV rockburst risk zone; The intelligent decision-making system initiates high-pressure water injection for pressure relief (injection pressure 18MPa, injection volume 0.3m³). 3 ( / min), the measured stress concentration factor decreased from 0.59 to 0.47, thus preventing disasters.

[0064] Step 204: Extract the geometric features of the hole wall cracks from the registered image features to obtain crack density, opening degree and azimuth data.

[0065] In some implementations, after the drilling robot drills a probe hole according to the BIM-planned path, its onboard industrial camera captures a panoramic image of the hole wall at a resolution of 0.1 mm / pixel. Simultaneously, a longitudinal wave velocity data (sampling rate 10 MHz) is acquired via an ultrasonic probe. This activates the image preprocessing step, which uses the CLAHE algorithm to enhance local contrast and combines it with OTSU adaptive threshold segmentation to extract the fracture region. The fracture contour is then extracted using the Canny edge detection algorithm, and the fracture azimuth distribution is statistically analyzed using the Hough transform.

[0066] Semantic segmentation of fracture images was performed using the U-Net model to output a fracture mask map (IoU > 0.85). The fracture opening was calculated using a fracture width measurement algorithm (error ± 0.1 mm). Finally, the number of fractures per unit area was counted using a fracture tracking algorithm, the fracture density was calculated, and the data was fused with wave velocity test data to generate a fracture development grade map (Grade I: intact rock mass; Grade V: fractured zone).

[0067] During the construction of a deep-buried quartz vein tunnel (980m depth, surrounding rock grade V): The drilling robot drilled a 50mm diameter probe hole to a depth of 8.2m; An industrial camera captured images of the borehole wall, revealing three sets of oblique fractures (azimuth angles of 35°, 145°, and 230°, respectively). Ultrasonic testing revealed an abnormal wave velocity in the intermediate fracture zone (vp decreased from 5200 m / s to 2800 m / s). The U-Net model segmented a fracture network with a total length of 12.7m, and the fracture density was calculated to be 8.3 fractures / m. 2 The maximum opening is 3.2mm; Based on the acoustic energy map (peak energy 0.6) and the infrared thermal imaging temperature difference of 2.9℃, it was determined to be a Class V rockburst risk zone; The intelligent decision-making system initiated directional blasting for pressure relief (hole depth 12m, explosive charge 0.6kg), and the measured stress concentration factor decreased from 0.63 to 0.44.

[0068] Step 205: Extract the temperature gradient features from the registered thermal imaging data to obtain the outline of the temperature anomaly zone on the rock surface.

[0069] In some implementations, an unmanned aerial vehicle (UAV)-borne infrared thermal imager performs a spiral scan along the tunnel cross-section to acquire a sequence of thermal infrared images with a resolution of 640×480, while simultaneously recording lidar point cloud data (point density ≥ 1000 points / m). 2 The Savitzky-Golay filter (window length 9, polynomial order 2) can be used to eliminate flight jitter noise, thereby converting pixel grayscale values ​​to absolute temperature through radiometric calibration.

[0070] Temperature anomaly regions were extracted using an improved threshold segmentation algorithm (Otsu + morphological opening operation). An initial threshold ΔT = 2℃ was set, and connected component analysis was performed on suspected regions. Three-dimensional coordinate mapping was then performed using UAV attitude angle data (pitch angle error < 0.5°) to ultimately generate the contour of the temperature anomaly zone on the rock surface (positioning error < 0.3m). This contour, along with the acoustic energy distribution feature map and strain abrupt change signal, was then analyzed in a spatiotemporal correlation to form a multi-dimensional risk criterion.

[0071] During the construction of a deep sandstone tunnel (650m depth, surrounding rock grade IV): Drone scanning revealed a persistent high-temperature zone in the upper right area of ​​the working face (temperature 38.7℃, surrounding area 25℃). An elliptical anomaly region with a diameter of 3.2m was extracted by threshold segmentation, with three-dimensional coordinates (X=2345678.456m, Y=5432109.789m, Z=1234.567m). The high-energy region (energy > 0.7) and strain abrupt change of 0.3 were observed in the acoustic energy map. It was determined to be a Level III rockburst risk zone; The intelligent decision-making system initiated advanced drilling for pressure relief (hole depth 3.5m, spacing 1.2m), which reduced the area of ​​the abnormal temperature zone by 62% and prevented the occurrence of disaster.

[0072] Step 206: Construct a multi-head attention mechanism to perform weighted fusion of acoustic energy distribution feature map, stress concentration area location signal, fracture density, opening degree and azimuth data and rock surface temperature anomaly area contour to generate spatiotemporal correlation features.

[0073] In some implementations, the acoustic energy distribution feature map is first normalized to the [0,1] interval, the stress concentration signal is normalized using z-score, the crack parameters (density, opening, azimuth) are normalized using min-max, and the temperature anomaly region contour is transformed to polar coordinates. This constructs a cross-modal attention computation framework comprising a feature embedding layer, a multi-head attention computation layer, a feature fusion layer, and a physical constraint verification layer.

[0074] That is, the feature embedding layer maps the four modal features to a 64-dimensional embedding space, generating a query matrix Q∈R. N×64 Key matrix K∈R N×64 Value matrix V∈R N×64 The multi-head attention computation layer computes eight attention heads in parallel, with each head processed through a different linear projection matrix. Focusing on local correlation features, among which , and ∈R 64×64The learningable projection matrix is ​​then used. In the feature fusion layer, the results of the eight attention heads are concatenated and compressed to the original dimension through a fully connected layer. The weighted fused features are then output and input into the physical constraint verification layer. The fused features are then input into the PINN network, and the residuals of the stress balance equation (error <5%) and the energy conservation equation (error <3%) are calculated using automatic differentiation. Backpropagation is then used to optimize the attention weight parameters.

[0075] During the construction of a deep-buried granite tunnel (920m depth, surrounding rock grade V): The acoustic energy map shows an energy peak of 0.82 at 15m behind the working face; The stress concentration signal was located in the region with X=2345678.567m and Y=5432109.890m; The crack parameters detected an opening of 3.5 mm and an azimuth angle of 125°. The outline of the temperature anomaly area shows an elliptical high-temperature zone (major axis 5.2m, minor axis 3.8m). Multi-head attention mechanism weighting: acoustic feature weight 0.42, stress feature weight 0.31, crack feature weight 0.18, temperature feature weight 0.09; The spatiotemporal correlation characteristics after fusion show that the risk zone expands asymmetrically along the tunnel axis, triggering the Intelligent Pressure Relief Robot (IPRR) to carry out directional blasting (hole depth 18m, explosive charge 0.7kg), and the measured stress concentration factor decreased from 0.67 to 0.43.

[0076] Construct a neural network model, which includes a feature encoding layer, a physical equation solving layer, and a residual optimization layer; The spatiotemporal correlation features are input into the feature encoding layer of the neural network model to obtain high-dimensional feature vectors, which are then fed back to the physical equation solving layer. The physical equation solution layer uses the rock mechanics governing equations as physical constraints and calculates the predicted stress and strain fields output by the network through automatic differentiation technology. The residual optimization layer calculates the fitting loss based on the predicted stress field and strain field values, and generates the initial geostress field inversion result by iterating the network parameters of the neural network model through the backpropagation algorithm.

[0077] During its use, the spatiotemporal correlation features can be invoked first to construct a neural network model (PINN model) architecture containing a feature encoding layer (such as a multilayer perceptron), a physical equation solving layer (with a built-in automatic differentiation module), and a residual optimization layer. The spatiotemporal correlation features can then be input into the feature encoding layer, which, through multiple nonlinear transformations, fuses and maps the multimodal acoustic energy, strain abrupt changes, crack geometry, and temperature gradient features into a high-dimensional feature vector H. The physical equation solving layer then receives the feature vector H and decodes it into a preliminary stress field. and strain field Predicted values. Simultaneously, the built-in rock mechanics governing equations (including the calibrated Hawke-Brown constitutive relation and the aforementioned energy conservation equations) are invoked, and automatic differentiation techniques are used to calculate... and The physical residuals are generated after substituting these equations. The residual optimization layer is then activated to calculate the physical residual loss. (i.e., the norm of the residuals of the aforementioned physical equations), and obtain the data fitting loss. , then and Based on the preset weights α=0.6 and β=0.4, through Calculate the data fitting loss (Where α and β are preset weighting coefficients, and α + β = 1), that is, comparison , Strain data directly derived from spatiotemporal correlation features The differences between them are then considered. Finally, the Adam optimizer is used to iteratively update all trainable parameters (weights and biases) of the PINN model through backpropagation, with the goal of making the PINN model more efficient. Minimize the loss function, and then iterate until the loss function converges to a stable threshold. Stop training at this point. This refers to a high-precision initial geostress field inversion result that satisfies physical laws and fits the monitoring data.

[0078] Taking the construction of a deep-buried granite tunnel (σc=150MPa) as an example, the DAS high-frequency signal and BOTDR strain mutation of 0.5 were collected in a certain area behind the tunnel face. The borehole images show dense micro-fractures, and the infrared thermal imaging shows a temperature rise of 2.5℃. After obtaining spatiotemporal correlation features through feature fusion, these features are input into the PINN model. PINN first fuses these features through a feature encoding layer, and then the physical equation solving layer solves the energy conservation equation based on the constitutive parameters of the granite (mb=28, s=0.005) obtained from AI inversion, as well as v, q, and r derived from the data. During training, the model continuously adjusts the parameters to ensure that the predicted stress field closely approximates the measured stress trend (data loss L_data) while also satisfying mechanical equilibrium and energy conservation (physical loss). Finally, the model outputs the initial geostress field of the region, showing that the maximum principal stress σ1 is 42 MPa, with a direction of N35°E. The error between the model and the subsequent verification results using the hydraulic fracturing method is less than 3%, and the directional error is less than 2°.

[0079] Based on the above technical solutions, a four-dimensional spatiotemporal coordinate system is constructed, and multi-source data are synchronously registered through timestamp alignment and spatial coordinate transformation. This breaks down the barriers of heterogeneous data fragmentation and solves the problems of spatiotemporal asynchrony and data incompatibility in traditional monitoring, laying a solid foundation for feature extraction. Simultaneously, for the registered data, acoustic time-frequency domain features are extracted to generate energy distribution maps, accurately locating rock mass damage areas and filtering noise interference; strain abrupt change features are extracted to achieve high-precision and rapid positioning of stress concentration areas, avoiding the risk of delayed or missed detections; quantitative parameters of borehole wall fractures are extracted to replace manual detection, improving efficiency and accuracy; and thermal imaging temperature gradient features are extracted to achieve non-contact, large-scale risk screening, solving the problems of narrow coverage and high risk in traditional temperature measurement. Finally, a multi-head attention mechanism is used to dynamically weight and fuse multimodal features, strengthening effective information, suppressing redundancy, and generating highly representative spatiotemporal correlation features. This provides accurate support for subsequent geostress inversion and rockburst identification, optimizing the monitoring and feature processing efficiency throughout the entire process.

[0080] Simultaneously, by constructing a dedicated PINN structure encompassing feature encoding, physical solution, and residual optimization, and training the network using dynamically constrained physical parameters derived from multi-source data, a balance between physical interpretability and data-driven accuracy in the geostress inversion process was achieved. This effectively solves the "black box" problem of traditional AI geostress inversion models, the weak anti-interference capability due to dependence on a single data source, and the technical problems of poor universality and large inversion errors (traditionally exceeding 8%) caused by the use of fixed empirical constitutive parameters. This reduces the inversion error of geostress magnitude and direction, laying a high-precision foundation for subsequent accurate prediction and proactive prevention.

[0081] In one possible implementation of the embodiments of this application, combined with Figure 1-3 As shown, the process of analyzing historical stress evolution patterns and predicting future window stress field distribution based on the initial inversion results using a spatiotemporal graph convolutional network can be achieved through the following steps 301 to 303, which are explained in detail below: Step 301: Construct a spatiotemporal graph convolutional network to map the initial geostress field inversion results into high-dimensional feature vectors, and convert the advanced borehole radar data into a binary mask image.

[0082] Spatiotemporal Graph Convolutional Network (ST-GCN) is a deep learning model that captures complex relationships between data by constructing a graph structure, capable of processing both spatial and temporal dimensions of data simultaneously. Advanced borehole radar is a device used for tunnel geological exploration. It transmits high-frequency electromagnetic waves to penetrate rock masses and generates radar images reflecting the internal structure of the surrounding rock based on the echo signals. A binarized mask image is a black-and-white image obtained by thresholding radar detection data. White areas represent anomalous geological bodies (such as faults and fracture zones), while black areas represent normal rock masses.

[0083] In some implementations, a spatiotemporal graph convolutional network is constructed. Its core is to map the initial geostress field inversion results into a high-dimensional feature vector through node embedding layers, where each node represents the stress state at a specific location in the tunnel. The network employs a multi-layer spatiotemporal convolutional structure. The spatial convolutional layers use Chebyshev multinomial approximation graph convolution operations, while the temporal convolutional layers use causal convolution to capture the temporal characteristics of stress evolution. Spatiotemporal feature fusion is achieved through gated linear units (GLUs).

[0084] During training, a Softmax attention mechanism is introduced to dynamically adjust the spatiotemporal weight parameter kr. This parameter is optimized by comparing historical stress window sequences with the surrounding rock wave velocity anomaly mask maps detected by advanced borehole radar at the corresponding times. When real-time multi-source monitoring data is received, the network adaptively adjusts the weights through the attention mechanism, outputs the geostress field prediction results for the future window, and corrects prediction errors by combining the binarized mask map of the radar data. For example, when the radar detects a low wave velocity anomaly area ahead, the mask map will mark the area as a high-risk target, and ST-GCN will automatically enhance the spatiotemporal weight of this area, making the prediction results more accurately reflect the actual geological conditions.

[0085] For example, acoustic and strain monitoring data near the tunnel face are first collected using DAS and BOTDR, and the initial geostress field is obtained by inversion using a PINN network. The inversion results are then input into the ST-GCN model, and the stress field data is mapped into a 128-dimensional feature vector through a node embedding layer. This allows the model to call pre-trained spatiotemporal convolutional kernels (e.g., 3 spatiotemporal convolutional layers, each with a kernel size of 3×3×3) to extract spatiotemporal correlation features. Simultaneously, a drilling robot drills probe holes along a preset path, and radar scans the surrounding rock in front of the tunnel face in real time, generating a wave velocity distribution image.

[0086] The system then uses the Otsu threshold segmentation algorithm to mark areas with wave velocities below 3500 m / s as red masks, while white areas represent normal rock masses. ST-GCN then dynamically adjusts the spatiotemporal attention weights based on the spatial distribution of anomalous areas in the mask image, reducing the prediction error of the stress concentration factor near the fault to within 8%. Finally, it outputs the geostress field prediction results for the next hour and the corrected risk warning level.

[0087] Step 302: Obtain the historical stress window sequence and the corresponding mask map, and train the spatiotemporal graph convolutional network.

[0088] In some implementations, a historical stress window sequence database is first established, where each window contains N consecutive hours of DAS microseismic signals, BOTDR strain data, and corresponding time-series advanced borehole radar detection results. Stress evolution segments from different time periods can then be extracted using a sliding window mechanism, forming a dataset containing T consecutive windows. During the training phase, real-time multi-source monitoring data is input into a spatiotemporal graph convolutional network. The initial geostress field inversion results are mapped into a 64-dimensional feature vector at the node embedding layer. The spatial convolutional layer uses Chebyshev multinomial approximation graph convolution operation, and the temporal convolutional layer uses causal convolution to capture the temporal features of stress evolution. Spatiotemporal feature fusion is achieved through gated linear units (GLUs).

[0089] During training, a Softmax attention mechanism is introduced to dynamically adjust the spatiotemporal weight parameter kr. This parameter is optimized by comparing the historical stress window sequence with the surrounding rock wave velocity anomaly mask map detected by the advanced borehole radar at the corresponding time. When the radar detects a low wave velocity anomaly area ahead, the mask map will mark the area as a high-risk target, and ST-GCN will automatically enhance the spatiotemporal weight of the area, so that the prediction results can more accurately reflect the actual geological conditions.

[0090] For example, when a deep-buried sandstone tunnel reached kilometer marker K12+345 at the tunnel face, the system automatically extracted a 3-hour historical stress window sequence (including DAS high-frequency micro-fracture signals, BOTDR strain abrupt changes, and radar detection data at the corresponding time). The ST-GCN model mapped the initial geostress field (σ1=42MPa, σ3=28MPa) into a 128-dimensional feature vector through a node embedding layer. The spatial convolutional layer used three Chebyshev convolutional kernels (size 5×5×5) to extract the spatial correlation features of the rock mass wave velocity anomaly area, and the temporal convolutional layer used a causal convolutional kernel (size 3×3) to capture the temporal pattern of stress growth. When a radar detection mask image was received (showing an anomaly area with wave velocity below 3500m / s in the upper right corner of the tunnel face), the Softmax attention mechanism increased the weight coefficient of the corresponding area to 0.6. After five iterations of training, the model's output stress concentration coefficient prediction error decreased from the initial 12% to less than 8%, successfully providing an early warning of the impending rockburst risk.

[0091] Step 303: Receive the real-time multi-source monitoring dataset, dynamically adjust the spatiotemporal weights through the attention mechanism, and output the stress field prediction results for the future window.

[0092] In some implementations, a multi-source monitoring dataset containing acoustic energy distribution, strain abrupt change location, fracture geometry features, and temperature anomaly profiles is received in real time and input into a pre-trained spatiotemporal graph convolutional network (ST-GCN). The network node embedding layer maps the initial geostress field inversion results into a 64-dimensional feature vector. The spatial convolutional layer uses Chebyshev multinomial approximation graph convolution operation, and the temporal convolutional layer uses causal convolution to capture the temporal features of stress evolution. This allows the Softmax attention mechanism to dynamically adjust the spatiotemporal weights ωkr according to the temporal change rate of each monitoring parameter: when acoustic energy suddenly increases or the strain abrupt change rate exceeds a threshold, the temporal dimension weight is increased; when thermal imaging shows changes in temperature gradient or an increase in fracture density, the spatial dimension weight is strengthened. For example, during the construction of a deep-buried granite tunnel, when the DAS detected a sudden increase of 30% in the energy of the high-frequency micro-fracture signal and infrared thermal imaging showed a local temperature increase of 2.8℃, the Softmax mechanism adjusted the corresponding spatiotemporal weight from the initial value of 0.3 to 0.6, making the stress concentration coefficient prediction value output by ST-GCN more accurately reflect the precursor characteristics of rockburst.

[0093] Based on the above technical solution, by employing a Spatiotemporal Graph Convolutional Network (ST-GCN) to map the initial geostress field inversion results into high-dimensional feature vectors and simultaneously converting advanced borehole radar data into binary mask maps, the problem of heterogeneous and difficult-to-coordinate stress data and geological exploration data formats in traditional models can be solved. This ensures the integrity and uniformity of the input data and provides a high-quality foundation for prediction. Furthermore, by training the ST-GCN with historical stress window sequences combined with corresponding radar mask maps, a dual-driven training mode of pattern discovery and geological constraints is formed. This overcomes the problems of weak generalization ability and poor adaptability of traditional models, enabling the generated model to adapt to different geological conditions and maintain stable prediction performance. Finally, by dynamically adjusting the spatiotemporal weights ωkr through an attention mechanism, the dynamic changes in geostress and risk areas are captured in real time. This solves the problems of lag and insufficient accuracy in traditional models, allowing for rapid output of future stress field prediction results during actual construction, reserving sufficient time for rockburst prevention and control, and ensuring safe and efficient construction progress.

[0094] In one possible implementation of the embodiments of this application, combined with Figure 1-4 As shown, the process of dynamically adjusting the rockburst discrimination threshold based on the stress concentration factor, combining the microseismic energy accumulation index and the surrounding rock deformation rate, calculating the comprehensive risk index, and determining the probability of rockburst occurrence can be achieved through the following steps 401 to 404, which are explained in detail below: Step 401: Collect the energy accumulation index of the microseismic monitoring system, obtain the source location through inversion using the three-dimensional velocity model, and calculate the accumulated elastic strain energy per unit volume of rock mass.

[0095] The microseismic monitoring system uses an array of sensors deployed in the surrounding rock of the tunnel to capture microseismic signals generated by rock fracturing in real time. The Energy Accumulation Index (EAI) is a quantitative indicator of the total energy released during a microseismic event. The three-dimensional velocity model is a geological structure model constructed based on the propagation speed of seismic waves in the rock mass.

[0096] In some implementations, waveform data acquired by a microseismic monitoring system can be used to first separate effective microseismic signals using time-frequency domain analysis. Travel time equations are then constructed using P-wave and S-wave velocity parameters from a three-dimensional velocity model, and the source coordinates are iteratively solved using a double-difference localization algorithm. Simultaneously, based on the principle of energy conservation, the amplitude integrals of each microseismic event are converted into energy values. Combining these with rock mass density and shear modulus parameters, the Green's function integral method is used to calculate the accumulated elastic strain energy per unit volume of rock. Finally, the source coordinates and the energy accumulation intensity are spatially visualized and overlaid for analysis.

[0097] Taking the construction of a deep-buried granite tunnel as an example, if the microseismic monitoring system continuously captures a cluster of microseismic events with a peak energy of 120 J for 72 hours, and the source is determined to be located 15 m in front of the tunnel face through three-dimensional velocity model inversion, the cumulative elastic strain energy per unit volume of rock mass in this area is calculated to be 8.2 MJ / m. 3 Based on the monitoring data of the surrounding rock deformation rate, the system automatically determined that the rockburst risk level had risen to level IV, triggering the intelligent decompression robot to carry out directional blasting, and successfully released 83% of the strain energy reserve in actual construction.

[0098] Step 402: Obtain the radial deformation rate of the surrounding rock through a distributed optical fiber strain monitoring system and calculate the average deformation rate.

[0099] In some implementations, Brillouin optical time-domain reflectometry (BOTDR) technology is used to acquire the surrounding rock strain data stream in real time through an optical fiber strain sensing network. After eliminating temperature drift noise using a Savitzky-Golay filter, the strain change rate is calculated using the sliding window difference method. Combined with the elasticity constitutive equation σ=E·ε (where σ is stress, E is elastic modulus, and ε is strain), a strain-deformation transformation model is established. The average deformation rate within a specified time window can then be directly calculated using time series analysis algorithms (such as moving average or exponential smoothing). Finally, the radial deformation rate data is mapped to spatial coordinates to form a visualized deformation cloud map.

[0100] Taking the construction of a deep-buried sandstone tunnel as an example, distributed fiber optic strain monitoring continuously collected strain data in a 30m area behind the tunnel face. The sliding window difference method revealed a sudden increase in the strain rate from 0.05% / min to 0.3% / min. Subsequently, wavelet threshold denoising was automatically applied, and the calculated average deformation rate was 0.18% / min. Combined with the microseismic energy accumulation index (EAI=65J), a yellow alert was triggered, and the intelligent decision-making system initiated advanced grouting reinforcement (grouting pressure 2MPa, grouting volume 0.8m). 3 The radial deformation rate of the surrounding rock is controlled within the safe threshold (0.2% / min).

[0101] Step 403: Construct a three-level risk assessment matrix. Normalize the stress concentration factor, elastic strain energy and average deformation rate, and then weight and fuse them to calculate the comprehensive risk index.

[0102] Among them, the stress concentration factor is the ratio of the maximum principal stress to the compressive strength of the rock, which reflects the degree of local stress anomaly. Elastic strain energy is the potential destructive energy stored in the rock mass under stress. The average deformation rate is the average velocity of radial deformation of the surrounding rock.

[0103] In some implementations, the min-max normalization method is used to adjust the stress concentration factor. (0-1) Elastic strain energy (0-10MJ / m 3 ), average deformation rate (0-1% / min) are mapped to the interval [0,1], and then weighted by a 4:3:3 ratio based on engineering experience, which can be used to perform a weighted summation formula. Calculate the comprehensive risk index When the stress concentration factor A red alert is triggered when the stress concentration factor is ≥0.55, and ≤0.45 is the stress concentration factor. A yellow alert is issued when the RI value is less than 0.55, and a three-level risk matrix is ​​established: Level I (RI < 0.4), Level II (0.4 ≤ RI < 0.6), and Level III (RI ≥ 0.6). Correspondingly, graded response strategies are formulated, including enhanced monitoring, partial pressure relief, and work stoppage for risk avoidance.

[0104] It should be noted that, for different lithologies (such as granite with high brittleness), the following measures should be taken: The weighting is increased to 50%, and reduced to 30% for sandstone with high plasticity. The parameters are automatically optimized during the construction phase (focusing on deformation rate in the early stage of excavation and stress concentration after support is completed) to ensure that the evaluation results are accurately matched with the actual project.

[0105] Taking the construction of a deep-buried marble tunnel as an example, monitoring data shows that the stress concentration coefficient at 20m behind the tunnel face is... c=0.52, elastic strain energy =5.8MJ / m 3 Average deformation rate When the deformation rate is 0.15% / min, the standardized values ​​are 0.78, 0.58, and 0.75, respectively, and the weighted average RI is 0.63. Since the area belongs to Class V surrounding rock and is adjacent to a fault zone, the system automatically increases the deformation rate weight to 40%, correcting the RI to 0.68, triggering a Level III risk warning. The intelligent decision-making system immediately initiates high-pressure water injection for pressure relief (injection volume 0.6 m³). 3 The overall risk index was reduced to 0.42 by 0.0000 min, successfully avoiding the occurrence of a rock eruption.

[0106] Step 404: Dynamically determine the probability of a rock eruption based on the comprehensive risk index.

[0107] In some implementation methods, a comprehensive risk index is obtained. ,when A value >0.8 indicates an extremely high risk of rockburst (probability >80%), while a value ≤0.6 indicates a risk of 0.8 or less. A value ≤0.8 indicates high risk (probability 40%–80%), while 0.4≤ <0.6 indicates medium risk (probability 10%-40%). A value <0.4 indicates low risk (probability <10%). The system incorporates a Bayesian probability model, continuously optimizing weight allocation and threshold settings based on historical rockburst case data.

[0108] Based on the above technical solutions, by employing microseismic monitoring combined with a three-dimensional velocity model to invert the seismic source location, the elastic strain energy per unit volume of rock mass is accurately quantified. This solves the problems of fuzzy positioning and inaccurate energy quantification in traditional microseismic monitoring, which lead to delayed precursor identification, allowing for earlier identification of high-risk areas. Simultaneously, distributed fiber optic strain monitoring enables continuous acquisition of the radial deformation rate of the surrounding rock and dynamic calculation of the average deformation rate, overcoming the limitations of limited coverage, data lag, and low manual efficiency of traditional equipment, providing real-time and reliable data support for risk assessment. Furthermore, in conjunction with a three-level risk assessment matrix, multi-dimensional indicators are normalized and weighted to generate a comprehensive risk index. This breaks through the limitations of traditional single indicators or empirical criteria, solving the problems of poor universality and high rates of misjudgment and missed judgment, allowing for precise risk level classification tailored to different lithologies and construction stages. Finally, the probability of rockburst occurrence is dynamically determined based on the comprehensive risk index, achieving real-time updates of risk quantification. This addresses the problems of fixed judgments and insufficient targeted decision-making in traditional methods, improving the timeliness and accuracy of prevention and control decisions, effectively reducing the rockburst accident rate, ensuring construction safety and progress, and balancing safety control with engineering benefits.

[0109] The above primarily describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, for example, an active prevention system for rockburst sections in high-stress tunnels, includes at least one of the hardware structures and software modules corresponding to the execution of each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in a hardware-driven or software-driven manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0110] This application embodiment can divide an active prevention system for rockburst sections in high-stress tunnels into functional units based on the above method example. For example, each function can be divided into separate functional units, or two or more functions can be integrated into the same processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents a logical functional division; other division methods may be used in actual implementation.

[0111] When using integrated units, Figure 5 This diagram illustrates a possible structural schematic of an active prevention system for high-stress tunnel rockburst sections, as described in the above embodiments. The active prevention system for high-stress tunnel rockburst sections includes: a multi-source data acquisition module, a feature-level fusion module, a physical information embedded neural network module, a spatiotemporal graph convolutional network module, a dynamic risk discrimination module, and a multi-objective optimization decision-making module. The multi-source data acquisition module acquires multi-source monitoring datasets and performs synchronous preprocessing to establish a four-dimensional spatiotemporal coordinate system, obtaining a registered multimodal monitoring dataset. The feature-level fusion module fuses the multimodal monitoring datasets through a multimodal data fusion mechanism to generate a spatiotemporal correlation feature matrix. The physical information embedded neural network module performs calculations and inversions based on the spatiotemporal correlation feature matrix to obtain the initial geostress field inversion result. The spatiotemporal graph convolutional network module calls the advanced borehole radar mask image and combines it with the initial geostress field inversion result to make predictions, obtaining the future geostress field prediction result and correction factor. The dynamic risk assessment module calculates a comprehensive risk index based on the corrected stress concentration factor, elastic strain energy, and average deformation rate, and outputs the probability of rockburst occurrence and the risk level. The multi-objective optimization decision-making module generates the optimal control scheme based on the risk level, rockburst occurrence probability, correction factor, and surrounding rock deformation rate, and transmits it to the execution layer for execution.

[0112] In conjunction with the second aspect mentioned above, in one possible implementation, the feature-level fusion module specifically includes a multimodal data synchronization and registration unit, a cross-modal feature extraction unit, and a multi-head attention fusion unit: the multimodal data synchronization and registration unit performs time synchronization on the multimodal monitoring dataset, followed by coordinate transformation and preprocessing to obtain a four-dimensional spatiotemporal registration dataset. The cross-modal feature extraction unit extracts features from the four-dimensional spatiotemporal registration dataset, generating a multi-dimensional feature matrix. The multi-head attention fusion unit calls the multi-dimensional feature matrix to perform multi-head attention calculation, outputting a spatiotemporal correlation feature vector.

[0113] In conjunction with the second aspect mentioned above, in one possible implementation, the physical information embedded neural network module specifically includes a physical information encoding layer, a physical equation solving layer, and a residual optimization layer. The physical information encoding layer extracts features based on spatiotemporal correlation feature vectors and separates subspaces using a feature decoupler to obtain multi-physics coupling features. The physical equation solving layer calculates the stress balance equation residuals from the multi-physics coupling features and, combined with the Hooke-Brown criterion and energy conservation equations, obtains the initial geostress field that satisfies physical constraints. The residual optimization layer calculates the composite loss function based on the stress balance equation residuals and data fitting loss, and iterates to generate the inversion result. In implementation, each step of the method provided in this embodiment can be completed by integrated logic circuits in the processor hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly manifested as being executed by a hardware processor, or being executed by a combination of hardware and software modules in the processor.

[0114] The processor in this application may include, but is not limited to, at least one of the following: a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), a microcontroller unit (MCU), or an artificial intelligence processor, etc., and other computing devices that run software. Each computing device may include one or more cores for executing software instructions to perform calculations or processing. The processor may be a standalone semiconductor chip or integrated with other circuits into a single semiconductor chip. For example, it may form a System-on-a-Chip (SoC) with other circuits (such as encoding / decoding circuits, hardware acceleration circuits, or various bus and interface circuits), or it may be integrated as a built-in processor within an ASIC. The ASIC of this integrated processor may be packaged separately or together with other circuits. In addition to the cores for executing software instructions to perform calculations or processing, the processor may further include necessary hardware accelerators, such as field-programmable gate arrays (FPGAs), programmable logic devices (PLDs), or logic circuits that implement dedicated logic operations.

[0115] The memory in the embodiments of this application may include at least one of the following types: read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions; random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions; or electrically erasable programmable-only memory (EEPROM). In some scenarios, the memory may also be compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media, or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but is not limited thereto.

[0116] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0117] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, the disclosure, and the appended claims, can understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0118] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and variations.

Claims

1. A method for active prevention of rockburst in high-stress tunnel sections, characterized in that, include: Collect multi-source monitoring datasets and identify rock mass surface temperature anomaly zones based on thermal imaging data, and mark high stress risk points. The multi-source monitoring datasets include acoustic signals, strain data, image features, and thermal imaging data. Feature-level fusion is performed on the multi-source monitoring dataset, spatiotemporal correlation features are extracted and input into physical information to embed into a neural network, and the initial geostress field inversion results are generated by combining the rock mass constitutive relation and energy conservation equation. Based on the initial geostress field inversion results, the historical stress evolution law is analyzed using a spatiotemporal graph convolutional network to predict the geostress field distribution in the future window. Combined with advanced borehole radar detection data, the stress prediction results are corrected, and dynamically updated stress concentration coefficients and risk warning levels are output. Based on the stress concentration coefficient, the rockburst discrimination threshold is dynamically adjusted, and the comprehensive risk index is calculated by combining the microseismic energy accumulation index and the surrounding rock deformation rate to determine the probability of rockburst occurrence. The risk warning level and the probability of rock burst occurrence are input into a multi-objective optimization model, and the optimal control scheme is generated by combining the pressure relief cost, support cost and project progress.

2. The active prevention method for rockburst sections in high-stress tunnels according to claim 1, characterized in that, The process of performing feature-level fusion on the multi-source monitoring dataset to extract spatiotemporal correlation features specifically includes: The acoustic signal, strain data, image features and thermal imaging data in the multi-source monitoring dataset are analyzed and a four-dimensional spatiotemporal coordinate system is established. Synchronous registration is performed through timestamp alignment and spatial coordinate transformation. Extract the time-frequency domain features from the registered acoustic signal to generate an acoustic energy distribution feature map; Extract strain abrupt change features from the registered strain data to obtain the stress concentration region location signal; Extract the geometric features of the hole wall cracks from the registered image features to obtain crack density, opening degree and azimuth data; The temperature gradient features in the registered thermal imaging data are extracted to obtain the outline of the temperature anomaly zone on the rock mass surface. A multi-head attention mechanism is constructed to weight and fuse the acoustic energy distribution feature map, stress concentration area location signal, fracture density, opening degree and azimuth data and the contour of the rock mass surface temperature anomaly area to generate spatiotemporal correlation features.

3. The active prevention method for rockburst sections in high-stress tunnels according to claim 2, characterized in that, The process of generating the initial geostress field inversion results by combining the constitutive relation of the rock mass with the energy conservation equation specifically includes: A neural network model is constructed, which includes a feature encoding layer, a physical equation solving layer, and a residual optimization layer. The spatiotemporal correlation features are input into the feature encoding layer of the neural network model to obtain a high-dimensional feature vector, which is then fed back to the physical equation solving layer. The physical equation solution layer uses the rock mechanics control equations as physical constraints and calculates the predicted stress field and strain field values ​​output by the network through automatic differentiation technology. The residual optimization layer calculates the fitting loss based on the predicted stress field and strain field values, and generates the initial geostress field inversion result by iterating the network parameters of the neural network model through the backpropagation algorithm.

4. The active prevention method for rockburst sections in high-stress tunnels according to claim 3, characterized in that, Based on the initial inversion results of the geostress field, the process of analyzing historical stress evolution patterns using a spatiotemporal graph convolutional network and predicting the distribution of the geostress field in future windows specifically includes: A spatiotemporal graph convolutional network is constructed to map the initial geostress field inversion results into high-dimensional feature vectors, and the advanced borehole radar data is converted into a binary mask image; Obtain the historical stress window sequence and the corresponding mask image at the time, and train the spatiotemporal graph convolutional network; It receives real-time multi-source monitoring datasets, dynamically adjusts spatiotemporal weights through an attention mechanism, and outputs stress field prediction results for future windows.

5. The active prevention method for rockburst sections in high-stress tunnels according to claim 4, characterized in that, The process of combining advanced borehole radar detection data to correct stress prediction results and output dynamically updated stress concentration coefficients and risk warning levels specifically includes: Calculate the stress concentration factor correction factor for the mask region based on the mask diagram and update the stress concentration factor; The risk level is classified according to the updated stress concentration factor.

6. The active prevention method for rockburst sections in high-stress tunnels according to claim 5, characterized in that, The process of dynamically adjusting the rockburst discrimination threshold based on the stress concentration coefficient, combining the microseismic energy accumulation index and the surrounding rock deformation rate, and calculating the comprehensive risk index to determine the probability of rockburst occurrence specifically includes: The energy accumulation index of the microseismic monitoring system is collected and the source location is obtained by inversion using a three-dimensional velocity model. The elastic strain energy accumulated per unit volume of rock mass is then calculated. The radial deformation rate of the surrounding rock was obtained through a distributed optical fiber strain monitoring system, and the average deformation rate was calculated. A three-level risk assessment matrix is ​​constructed. The stress concentration factor, elastic strain energy and average deformation rate are normalized and then weighted and fused to calculate the comprehensive risk index. The probability of a rock eruption is dynamically determined based on the comprehensive risk index.

7. The active prevention method for rockburst sections in high-stress tunnels according to claim 6, characterized in that, The process of inputting the risk warning level and rockburst probability into a multi-objective optimization model, and combining the decompression cost, support cost, and project schedule to generate the optimal control scheme, specifically includes: A multi-objective optimization model based on the NSGA-II algorithm is constructed. The multi-objective optimization model has the decompression cost, support cost and project schedule delay rate as optimization objectives, and safety threshold, resource consumption limit and construction disturbance intensity as constraints. The risk warning level, rock burst probability, stress concentration coefficient dynamic correction factor, and surrounding rock deformation rate real-time data stream are input into the multi-objective optimization model. An initial Pareto solution set is generated through a multi-objective genetic algorithm, and the scheme weights are adjusted. The optimal control scheme is selected based on the highest overall effectiveness and is then sent to the intelligent decompression robot and shape memory alloy support system in real time via edge computing nodes. The optimal control scheme includes decompression parameters and support parameters.

8. An active prevention system for rockburst sections in high-stress tunnels, characterized in that, An active prevention method for rockburst sections in high-stress tunnels, applicable to any one of claims 1-7, comprises a multi-source data acquisition module, a feature-level fusion module, a physical information embedded neural network module, a spatiotemporal graph convolutional network module, a dynamic risk discrimination module, and a multi-objective optimization decision-making module. The multi-source data acquisition module acquires multi-source monitoring datasets and performs synchronous preprocessing to establish a four-dimensional spatiotemporal coordinate system, thereby obtaining a registered multimodal monitoring dataset. The feature-level fusion module fuses the multimodal monitoring dataset through a multimodal data fusion mechanism to generate a spatiotemporal correlation feature matrix; The physical information embedded neural network module performs calculation inversion based on the spatiotemporal correlation feature matrix to obtain the initial geostress field inversion result; The spatiotemporal graph convolutional network module calls the advanced borehole radar mask map and combines it with the initial geostress field inversion results to make predictions, thereby obtaining the future geostress field prediction results and correction factors. The dynamic risk assessment module calculates a comprehensive risk index based on the corrected stress concentration factor, elastic strain energy, and average deformation rate, and outputs the probability of rock burst occurrence and risk level. The multi-objective optimization decision-making module generates the optimal control scheme based on the risk level, rock burst probability, correction factor, and surrounding rock deformation rate, and transmits it to the execution layer for execution.

9. The active prevention system for rockburst sections in high-stress tunnels according to claim 8, characterized in that, The feature-level fusion module specifically includes a multimodal data synchronization and registration unit, a cross-modal feature extraction unit, and a multi-head attention fusion unit: The multimodal data synchronization and registration unit performs time synchronization, coordinate transformation, and preprocessing on the multimodal monitoring dataset to obtain a four-dimensional spatiotemporal registration dataset. The cross-modal feature extraction unit extracts features from the four-dimensional spatiotemporal registration dataset to generate a multi-dimensional feature matrix. The multi-head attention fusion unit calls the multi-dimensional feature matrix to perform multi-head attention calculation and outputs a spatiotemporal correlation feature vector.

10. The active prevention system for rockburst sections in high-stress tunnels according to claim 9, characterized in that, The physical information embedded neural network module specifically includes a physical information encoding layer, a physical equation solving layer, and a residual optimization layer; The physical information encoding layer extracts features based on the spatiotemporal correlation feature vector and separates subspaces through a feature decoupler to obtain multi-physics coupling features; The physical equation solving layer calculates the stress balance equation residuals for the multi-physics coupling characteristics, and combines the Hawke-Brown criterion and the energy conservation equation to obtain the initial geostress field that satisfies the physical constraints. The residual optimization layer calculates the composite loss function and iterates based on the residual of the stress balance equation and the data fitting loss to generate the inversion result.

Images