Construction risk analysis and emergency method

Abstract

The invention discloses a construction risk analysis and emergency method, and relates to the field of hydraulic engineering construction. The method comprises the following steps: initializing stratum loss transfer function parameters and a time sequence rule base on the basis of geological exploration reports and indoor tests by constructing digital twin bodies; collecting real-time monitoring data, inverting formation loss distribution, and correcting model parameters through an ensemble Kalman filtering data assimilation technology; calculating a dynamic threshold based on the updated model, matching a time sequence association rule, and predicting an engineering state in a rolling manner; calling a physical evolution model to reversely derive an optimal disposal parameter and issuing an instruction; according to the method, a dynamic closed loop of'perception, analysis, decision and learning 'is formed by constructing digital twins, performing data assimilation correction on parameters, performing dynamic threshold calculation and state prediction, calling a physical evolution model to generate a disposal instruction, and performing self-learning optimization on the parameters and the rule base; and accurate risk prevention and control are realized.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering construction, specifically to a construction risk analysis and emergency response method. Background Technology

[0002] In shield tunnel projects such as urban rail transit and integrated utility tunnels, the situation of tunneling under sensitive structures such as high-speed railways, dense building complexes, and important municipal pipelines is becoming increasingly common. The core technical challenge of such projects lies in the fact that traditional experience-based static risk management models are difficult to effectively cope with the dynamic risk evolution process characterized by "highly variable geological conditions, complex propagation paths of construction disturbances, and delayed and cascading environmental responses."

[0003] The existing technical solutions mainly suffer from the following systemic technical bottlenecks:

[0004] The static and empirical nature of risk assessment criteria: Risk warnings generally rely on preset, single, and fixed thresholds (such as settlement rate > 2 mm / d). However, the transmission of ground losses and pore water pressure changes caused by tunnel boring machine (TBM) disturbances to the surface is a complex spatiotemporal process. The same surface settlement rate can result in drastically different actual risk levels for underground structures (such as pile foundations) and surface facilities (such as high-speed rail tracks) under different geological combinations (such as soft top and hard bottom, rich in confined water) or different burial depths. Existing technologies completely ignore the spatiotemporal lag effect and attenuation characteristics of ground loss transmission, leading to a severe disconnect between warnings and the actual risk status. This results in either frequent false alarms interfering with construction or missed warnings causing accidents.

[0005] The isolation and one-sidedness of risk perception are highlighted: most existing monitoring systems are simply a collection of sensors, lacking in-depth correlation analysis based on physical models between data. For example, if only surface subsidence is monitored while changes in deep soil displacement and pore water pressure are ignored, it is impossible to accurately identify whether the subsidence is caused by improper shield tunneling posture or by erosion induced by the instability of the karst cave ahead. This "seeing the trees but not the forest" approach to perception fails to capture the precursors of the chain evolution of risks.

[0006] The lag and fragmentation of decision-making in emergency response are prominent issues: most emergency plans are merely textual clauses, severely disconnected from real-time, multi-source risk status information. When a hazard occurs, manual comprehensive judgment, hierarchical approval, and telephone dispatch are still required, resulting in a long timeframe from hazard awareness to effective response, often missing the optimal intervention window. Furthermore, the lack of effective coordination among various response actions (such as grouting, shield tunneling parameter adjustment, and traffic control) based on the same risk model can lead to mutual obstruction and negatively impact response effectiveness.

[0007] The system lacks self-adaptation and evolution capabilities: once the existing system parameters (such as early warning thresholds) are set, they become fixed and cannot be self-adjusted according to the differences in specific engineering geological conditions and construction technology. Furthermore, it cannot learn from past dangerous cases, resulting in poor applicability of the system at new work sites and a low level of intelligence.

[0008] Therefore, there is an urgent need in this field for an integrated intelligent prevention and control technology that can deeply integrate geomechanical mechanisms, construction disturbance theory and real-time monitoring data to achieve risk "advanced perception, precise quantification, collaborative intervention and autonomous learning". Summary of the Invention

[0009] To address the problems of existing technologies, this invention provides a construction risk analysis and emergency response method.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A construction risk analysis and emergency response method includes the following steps:

[0012] S1. Construct a digital twin integrating a 3D geological model, tunnel model, environmental model, and sensor network; initialize the formation loss transfer function based on the geological survey report and laboratory tests. Initial version of parameters and time-series rule base;

[0013] S2, Collection Inverting the actual stratigraphic loss distribution Model parameters are corrected by using ensemble Kalman filter data assimilation techniques;

[0014] S3, based on the updated Calculate dynamic thresholds and match temporal association rules to predict the engineering status of the next time step in a rolling manner.

[0015] S4. Call the physical evolution model corresponding to the risk type, reverse deduce the optimal handling parameters, and send the instructions to the equipment and external units simultaneously;

[0016] S5. Collect data from the entire process and use the self-learning unit to optimize the formation loss transfer function parameters and time series rule base, thereby completing the learning cycle.

[0017] The sensor network of the digital twin mentioned in step 1 includes sensor arrays at different depths of the strata.

[0018] The implementation process of the ensemble Kalman filter data assimilation technology in step 2 includes: inverting the ground loss distribution based on the real-time data of the tunnel boring machine PLC, and correcting the model parameters through 50-100 iterations.

[0019] In step S3, the dynamic early warning threshold is calculated based on the formation loss state at the current time t, predicting the future preset time period. The theoretical risk index value of the post-monitoring point is used as the predicted value for the future time t+. The dynamic early warning threshold.

[0020] In step S4, the self-learning unit adopts a reinforcement learning framework to adjust the safety factor of the formation loss transfer function and the time window of the time series rule.

[0021] Compared with existing technologies, the beneficial effects of this invention are: the construction risk analysis and emergency response method forms a dynamic closed loop of "perception, analysis, decision-making, and learning" by constructing a digital twin, assimilating and correcting parameters through data assimilation, calculating dynamic thresholds and predicting states, calling physical evolution models to generate disposal instructions, and self-learning to optimize parameters and rule bases, thereby achieving precise risk prevention and control. Detailed Implementation

[0022] The present invention will be further described in detail below through embodiments. These embodiments are only used to illustrate the present invention and do not limit the scope of the present invention.

[0023] A construction risk analysis and emergency response method, which is a dynamic closed-loop process driven by a physical model, includes the following steps:

[0024] S1. Digital Twin Initialization and Mechanism Model Embedding Features: Constructing a digital twin integrating a 3D geological model, tunnel model, environmental model, and sensor network; initializing the formation loss transfer function based on geological survey reports and laboratory tests. The parameters and the initial version of the time-series rule base provide a physical benchmark and initial model for system operation, ensuring a reliable foundation for subsequent data assimilation and risk assessment.

[0025] S2. Real-time data assimilation and model update features: data collection Inverting the actual stratigraphic loss distribution By employing ensemble Kalman filtering (EnKF) data assimilation techniques to correct model parameters, the digital twin is synchronized with the physical world. This compensates for uncertainties in geological data, calibrates the model in real time, and improves prediction accuracy.

[0026] S3. Dynamic Risk Forward-Looking Assessment Characteristics: Based on Updated... Calculate dynamic thresholds and match them with time-series association rules to predict the engineering status at a future time step (e.g., 30 minutes). This enables proactive risk identification, early detection of evolving trends, and allows for a time window for intervention.

[0027] S4. Model-Driven Precision and Collaborative Response: This approach invokes a physical evolution model tailored to the corresponding risk type, aiming to curb risk evolution. It then reverse-engineers the optimal response parameters and simultaneously sends instructions to equipment and external units. This avoids the blindness of experience-based responses, achieving "targeted intervention" and improving both efficiency and scientific rigor.

[0028] S5. Closed-Loop Learning and Model Evolution Features: Data from the entire process is collected, and the formation loss transfer function parameters and time-series rule base are optimized using self-learning units to complete the learning loop. To achieve continuous iteration of system capabilities, it needs to be adapted to different construction stages and geological sections, thereby improving its long-term applicability.

[0029] In step 1, the sensor network of the digital twin covers sensor arrays at different depths of the strata, specifically including hydrostatic level, deep settlement gauge, pore water pressure gauge and earth pressure cell. These devices are used to collect deformation, pressure and hydrological data of the strata.

[0030] In step 1, the initialization data of the formation loss transfer function parameters are derived from the soil physical and mechanical parameters in the geological survey report, such as elastic modulus, Poisson's ratio, permeability coefficient, and soil disturbance response curves obtained from laboratory tests.

[0031] The implementation process of the ensemble Kalman filter (EnKF) data assimilation technology described in step 2 includes: inverting the ground loss distribution based on real-time data (thrust, torque, excavated soil, grouting pressure and volume) from the tunnel boring machine's PLC, and correcting the model parameters through 50-100 iterations to control the synchronization error between the digital twin and the physical world to within 5%.

[0032] In step S3, the dynamic early warning threshold is calculated based on the formation loss state at the current time t, predicting the future preset time period. The theoretical risk index value of the post-monitoring point is used as the predicted value for the future time t+. The dynamic early warning threshold is calculated based on the theory of random media. Specifically, it uses the formation loss transfer function to predict the theoretical maximum deformation value of the monitoring point within a preset time period (30 minutes), and then multiplies it by a safety factor related to the sensitivity coefficient of the object being traversed (1.2 for high-speed rail tracks and 0.8 for general road surfaces) to obtain the dynamic early warning threshold. The time-series association rules include the time interval and spatial location association conditions of abnormal signals. For example, when the pore water pressure in front of the tunnel face decreases by more than ΔP threshold 1 (0.2 MPa) within Δt ≤ 5 minutes, and the settlement rate of the corresponding surface point exceeds the dynamic early warning threshold within the time window of [t water pressure drop + 5 min, t water pressure drop + 15 min], a karst collapse risk warning is triggered.

[0033] In step S4, when generating the disposal instruction, the goal is to curb or reverse the evolution of the risk. Key quantitative parameters for the disposal action are calculated using a physical evolution model to ensure precise disposal. The physical evolution model includes a fluid diffusion model. For karst collapse risk, the grouting volume is calculated using the formula Q=1.2×k×A×h / (μ×γ) (where k is the formation permeability coefficient, A is the affected area, h is the grouting depth, μ is the grout viscosity, and γ is the grout unit weight), and a target grouting pressure value (not less than 1.5 times the initial formation pressure) is simultaneously output.

[0034] In step S4, the self-learning unit adopts a reinforcement learning framework. By adjusting the safety factor (range 0.9-1.3) of the formation loss transfer function and the time window of the time series rule (adjustment step size ±2 minutes), the optimization objective is to make the false alarm rate ≤0.5% and the false alarm rate ≤5%.

[0035] This construction risk analysis and emergency response method is based on a dynamic analysis and emergency handling system, covering the following:

[0036] To address the issues of isolated and uncorrelated data in existing technologies, this solution incorporates a multi-source data sensing and fusion module. This module acquires and integrates multi-dimensional monitoring data related to tunnel boring machine (TBM) construction in real time, constructing a spatiotemporally synchronized engineering status field data. Specifically, a unified spatiotemporal reference framework for data is constructed, deploying a three-dimensional monitoring network covering the entire chain from "disturbance source to propagation path to affected object." Data is fused using a spatiotemporal registration algorithm based on a world coordinate system and unified timestamps. This includes: sensor arrays (static level, deep settlement gauge, pore water pressure gauge, earth pressure cell) at different depths in the strata (surface, tunnel crown, arch waist, and in front of the tunnel face); real-time data from the tunnel boring machine's PLC (thrust, torque, excavated soil, grouting pressure and volume); and monitoring data from external facilities (track geometry). This achieves comprehensive, high-frequency data acquisition and spatiotemporal alignment, generating three-dimensional, spatiotemporally synchronized engineering state field data. This provides a precise and unified data foundation for subsequent risk assessment.

[0037] The dynamic risk intelligent assessment module specifically includes the following components:

[0038] To avoid the shortcomings of the static lookup table method, a dynamic risk threshold generation unit is established, which considers the spatiotemporal transmission effect of formation loss. The warning threshold is an adaptive value that changes with time. The core formula is as follows:

[0039]

[0040] The parameters in the formula are defined as follows:

[0041] : The soil parameter vector at monitoring point i (including the elastic modulus of each soil layer) Poisson's ratio Permeability coefficient Its function is to reflect the influence of the inherent mechanical properties of the strata on the disturbance response and to ensure that the threshold is adapted to different geological conditions.

[0042] The sensitivity coefficient of the object to deformation at point i is used. For example, high-speed rail tracks have extremely high sensitivity, while general road surfaces have low sensitivity. Its function is to adjust the safety boundary according to the importance of the protected object, so as to avoid false alarms or missed alarms caused by a uniform threshold.

[0043] The cumulative ground loss volume and spatial distribution, derived from shield tunneling parameters (over-excavation and grouting filling rate), is used to quantify the core source of construction disturbance in real time and provide a disturbance basis for threshold prediction.

[0044] The real-time tunneling speed of the tunnel boring machine serves to reflect the rate characteristics of construction disturbances, thereby adapting to the risk evolution patterns under different tunneling rhythms.

[0045] The time lag constant from the occurrence of the disturbance to the generation of a significant response at the monitoring point (as determined by previous experiments or numerical simulations).

[0046] (To be determined), its function is to compensate for the spatiotemporal lag effect of formation loss transmission and to achieve forward-looking threshold setting.

[0047] function Predicting future time periods based on stochastic media theory / empirical formulas. The theoretical maximum deformation value of internal monitoring point i, multiplied by The relevant safety factor is obtained This allows the threshold to adapt to changes in construction progress, geological conditions, and construction parameters, thus solving the problem of the disconnect between traditional static thresholds and actual risks.

[0048] Specifically, it is configured as follows: for each monitoring point, based on the formation parameter vector at that point... Sensitivity coefficient for crossing objects The spatial distribution and volume of cumulative ground loss obtained from the inversion of real-time tunneling parameters of the shield machine. Real-time tunneling speed of shield tunneling and disturbance response time lag constant Through the built-in formation loss transfer function The dynamic early warning threshold for at least one risk indicator at the monitoring point is calculated in real time, and the dynamic early warning threshold is a function that changes with time.

[0049] In the specific dynamic risk threshold generation unit, the dynamic early warning threshold of monitoring point i at time t with respect to risk index I is... It is calculated using the following function model:

[0050]

[0051] in, This is a formation response calculation function based on random medium theory or empirical formulas, used to predict the theoretical maximum deformation value at monitoring points; To determine the safety factor corresponding to the risk level, based on the sensitivity of the object being traversed. Adjustment.

[0052] The multi-indicator time-series coupling analysis unit stores multiple time-series correlation rules defined based on the physical processes of different engineering risk evolution. Each rule clarifies the correlation conditions of multiple risk indicator abnormal signals in terms of occurrence order, time interval, and spatial location. This unit is configured to match real-time monitoring data with these time-series correlation rules. Once the monitoring data meets all the correlation conditions defined by a certain rule, a corresponding risk warning will be triggered. Specifically, a multi-indicator abnormal time-series correlation rule library based on the physical processes of risk evolution is constructed. The rules emphasize the consistency between the sequence, time interval, spatial correlation of abnormal signals and the geological disaster mechanism, rather than a simple logical "AND". Taking the karst collapse precursor rule as an example, the formula is as follows:

[0053]

[0054] Key parameters in the formula

[0055] The pressure range of the karst cave area in front of the face ...

[0056] Δt≤Δt1 represents the time threshold for a sudden drop in pore water pressure, where Δt is the actual monitored duration of the pressure drop, and Δt1 is the preset maximum allowable time. This threshold distinguishes between a "sudden drop" and normal fluctuations, ensuring that warnings are triggered only for rapid pressure changes within a short period, thereby eliminating interference from slow geological processes.

[0057] ΔP ≥ ΔP threshold 1 represents the threshold for the magnitude of pore water pressure drop, where ΔP is the actual pressure drop and ΔP threshold 1 is the minimum pressure drop for risk assessment. In the risk assessment of karst water inrush disasters in railway tunnels, by quantifying the severity of pressure anomalies, only when the pressure drop exceeds a specific threshold is it considered a potential precursor to karst water inrush.

[0058] Vs对应地表点 ≥λs(t): Vs is the settlement rate of the surface monitoring point, and λs(t) is the dynamic early warning threshold of the point at time t (which changes adaptively with construction progress and geological conditions); by comparing the settlement rate with the dynamic threshold, it is determined whether the surface deformation exceeds the safety boundary, reflecting the stability of the strata.

[0059] t_accelerated settlement ∈ [t_water pressure drop + Tmin, t_water pressure drop + Tmax]: This represents the time window in which settlement acceleration occurs. t_water pressure drop is the moment when the pore water pressure begins to decrease, and Tmin and Tmax are time intervals estimated based on the physical process of burrowing. This temporal logic of linking "water pressure drop (cause) - settlement acceleration (effect)" aligns with the geological evolution mechanism of karst collapse, avoiding misjudgment of unrelated settlement events.

[0060] The matching degree of the characteristic spectrum of F / M shield tunneling machine's "cutterhead contact cavity" is ≥85%, where F / M shield tunneling machine refers to the real-time monitoring data of the shield machine's thrust (F) or torque (M). By matching the real-time monitoring data with the typical characteristic spectrum (such as fluctuation frequency and amplitude) of the "cutterhead contact cavity," the matching degree must reach above 85%. From a mechanical response perspective, this can corroborate the existence of a cavity (such as a sinkhole), improving the specificity of risk identification and reducing the possibility of false alarms due to a single indicator.

[0061] The time window estimated based on the physical process of erosion is used to link the temporal logic of "water pressure drop (cause) - accelerated subsidence (effect)," which is consistent with the evolution mechanism of geological disasters.

[0062] Feature spectrum matching degree is used to quantify the consistency between the response and risk mode of tunnel boring machines, which can provide evidence for the source of disturbance and thus improve the specificity of the judgment.

[0063] Risk Level = Level I (High Risk of Karst Water Inrush and Collapse): When all the above conditions are met simultaneously, the system determines it to be at the highest risk level (Level I), i.e., an emergency state of particularly serious geological disaster that may cause more than 30 deaths or direct economic losses of more than 10 million yuan. The most stringent response procedures are triggered, such as immediately stopping tunneling and initiating grouting and sealing, to ensure construction safety.

[0064] The aforementioned parameters, through a chain relationship of "pressure anomaly, formation instability, and mechanical response," verify karst risks from multiple dimensions. The spatiotemporal coupling of water pressure and settlement must ensure a causal relationship between sudden pressure drops and accelerated settlement, rather than accidental synchronization. Dynamic thresholds and mechanical characteristics, combined with formation conditions and tunnel boring machine status, enable accurate risk identification and classification. By correlating multi-source indicators in space and time according to physical mechanisms, different risk causes under similar appearances can be distinguished, improving the accuracy of risk identification and the lead time for early warning, thus overcoming the perception limitations of existing technologies that "see the trees but not the forest."

[0065] Specifically, the time-series correlation rules stored in the multi-index time-series coupling analysis unit include karst collapse precursor rules, which are defined as follows: A pore water pressure within a preset risk area ahead of the tunnel face is within a certain range. The decrease value exceeds the first threshold within a certain time period. And within the time window Within this area, the subsidence rate at the corresponding surface point exceeds its dynamic early warning threshold. Furthermore, when the thrust or torque of the tunnel boring machine exhibits a fluctuation pattern matching the contact void characteristics of the cutting tools (matching degree ≥ 82%), it is judged as a high-risk karst collapse; among which, It is a time interval estimated based on the physical process of soil instability caused by erosion.

[0066] The model-driven collaborative response decision-making module is configured to generate an executable set of response instructions containing specific quantitative parameters based on the risk type and prediction information output by the dynamic risk intelligent assessment module, combined with the risk evolution physical model. The response plan generation is directly coupled with the risk evolution model, issuing instructions through standard industrial protocols (for example, in a case study of a subway shield tunnel passing under a high-speed railway station and a water-rich karst area, when the system triggers a Level I risk warning, the grouting instruction is automatically sent to the grouting trolley via the Modbus TCP protocol to initiate precise grouting, while simultaneously pushing a structured report to the railway department, promoting a shift in collaborative response from "passive alarm reception" to "proactive prevention"). It also pushes structured quantitative prediction reports to external units, rather than simply issuing alarms. Taking karst collapse risk response as an example, the grouting parameter calculation logic is as follows:

[0067]

[0068] In the formula: This refers to the grouting volume. The formation permeability coefficient, To affect the area, The viscosity of the slurry. The slurry density is [not specified]. The value represents the grouting depth, and a coefficient of 1.2 represents a safety redundancy. By generating executable instructions with quantifiable parameters, the system achieves precise and coordinated handling actions, shortening response time. Simultaneously, it provides external units with predictive data, promoting a shift in collaborative response from "passive alarm reception" to "proactive prevention," and addressing the issues of delayed and fragmented handling.

[0069] The model-driven collaborative decision-making module is configured to: for karst collapse risk, based on the pore water pressure drop rate, spatial gradient and formation permeability coefficient, use the fluid diffusion model to reverse calculate the key location, grout volume and target pressure value of grouting and sealing, and then generate grouting control commands.

[0070] The self-learning optimization unit employs a reinforcement learning framework that integrates mechanistic models and data-driven approaches. It defines a state space containing real-time monitoring data and theoretical safety boundaries, and an action space containing model parameters and rule-based adjustment actions. Both are coupled to the reward function through a closed-loop "action-state-reward" relationship. The reward function is...

[0071] ;

[0072] ω1: False negative rate weighting coefficient (highest priority), reflecting the "safety first" principle, with the largest value. It reduces the probability of unidentified risks through positive incentives, thus preventing safety incidents.

[0073] 1 - Missed Detection Rate: A metric for the accuracy of risk identification. The missed detection rate refers to the probability that "a risk actually exists but is not warned about." This represents the percentage of risks successfully identified. The lower the false negative rate, the higher this value, resulting in a larger reward function value and driving the system to prioritize reducing false negatives.

[0074] False alarm rate weighting coefficient (second highest priority), with a weight less than ω1. Through the penalty term ( This will reduce instances of false warnings despite no risk, and prevent waste of resources and disruption to construction.

[0075] False alarm rate: The probability of an incorrect warning, referring to the proportion of times when there is no risk, the system mistakenly judges it as risky. The higher this proportion, the larger the penalty and the lower the reward function value, prompting the system to improve the accuracy of warnings.

[0076] : Warning timeliness weighting coefficient, through penalty item ( The incentive system provides early warnings; the shorter the warning time (the smaller the denominator, the larger the value of this term), the heavier the penalty, thus pushing the system to extend the warning window.

[0077] Disposal cost weighting coefficient (lowest priority). Through penalty items ( To balance safety and economy, the frequency of using high-cost disposal solutions should be reduced. In general... Weighting coefficients ( Its function is to highlight the principle of safety first and to balance the relationship between underreporting, false reporting, early warning timeliness and handling costs.

[0078] Comprehensive cost of handling a single incident: This includes the resources consumed in risk management, such as direct and indirect costs like grouting materials, equipment wear and tear, and project delays. Higher costs result in greater penalties and lower reward function values, driving the system to choose a low-cost, high-efficiency handling strategy.

[0079] state space Includes monitoring data The theoretical safety boundary (the result of dynamic threshold calculation) serves to comprehensively reflect the deviation between the actual engineering state and the model prediction. The state space (S) forms the basis for calculating the false negative rate.

[0080] The specific state space includes real-time monitoring data (such as formation pressure and settlement rate) and theoretical safety boundaries (dynamic thresholds). These data directly determine the system's ability to perceive risks. If the state space data is insufficient or distorted (such as sensor malfunction leading to missing pressure data), the system may be unable to detect early signs of risk, resulting in an increased false negative rate.

[0081] If the theoretical safety boundary (dynamic threshold) is set unreasonably (e.g., the threshold is too high), the actual risk indicators will not be identified, which will also lead to underreporting.

[0082] Action space : Including fine-tuning functions Parameters (safety factor, time lag) The adjustment of the time window for time-series rules is crucial for achieving precise iteration of model parameters. Specifically, the action space involves adjusting parameters of the formation loss transfer function (such as the safety factor and time lag constant) and the time window for time-series rules, directly optimizing the sensitivity of risk identification. Decreasing the safety factor (making early warning more sensitive) or shortening the time window for time-series rules (speeding up response) can reduce false negatives caused by overly strict thresholds or delayed responses. Conversely, improper parameter adjustments (such as an excessively high safety factor) may lead to excessively high warning thresholds, causing risk signals to be ignored and increasing the false negative rate. Through interaction with real engineering environments, the mechanistic model parameters and rule base details are continuously optimized, adapting the system to specific geological conditions at work sites, achieving increasing accuracy with use, and addressing the problems of poor applicability and lack of evolutionary capability in existing systems. The state space provides the basic calculation data for the reward function, while the action space achieves optimization iteration through reward feedback, ultimately completing the adaptive adjustment of model parameters and time-series rules.

[0083] In summary, the state space provides the "raw data" for calculating the false negative rate, the action space directly affects the false negative rate through parameter adjustments, and the false negative rate drives system optimization through feedback from the reward function. These three elements form a dynamic cycle of "state awareness → action adjustment → false negative rate change → reward feedback → further optimization," ultimately minimizing the false negative rate.

[0084] Specifically, the state space of the self-learning optimization unit includes real-time monitoring data. The theoretical safety boundary calculated by the dynamic risk threshold generation unit, the action space includes adjusting the parameters in the formation loss transfer function and / or the time window parameters in the temporal association rules, and the reward function. These are the weighting coefficients, and This reflects the principle of prioritizing safety.

[0085] Project Case: A subway shield tunnel project passing under a high-speed railway station and a water-rich karst area

[0086] Project Overview and System Deployment

[0087] The tunnel needs to pass under the railway station building (sensitivity) It also includes a karst development area about 200m long with abundant water (cavity rate >50%), where the karst water is connected to the surface water system, posing an extremely high risk level.

[0088] Sensing network deployment: In addition to conventional monitoring, three rings of pore water pressure gauges are arranged in a ring at 5m, 10m and 15m outside the tunnel outline in the karst section; inclinometers and strain gauges are installed around the pile foundation of the station building; and earth pressure sensors are installed on the cutterhead and tail of the shield machine to form a three-dimensional monitoring network.

[0089] Digital Twin and Model Initialization: A 3D geological model is built based on detailed exploration data, identifying known karst caves. The formation loss transfer function is initialized. Formation parameters Taken from the geological survey report (silty clay layer) , limestone , Time lag constant Initially set to 120 minutes; initialize the time-series rule base, including karst collapse rules. With the window set to [30, 90] minutes, the threshold for sudden drop in pore water pressure is... .

[0090] System Operation Example: Early Warning and Response to Karst Water Inrush Risk

[0091] Time T0: The tunnel boring machine cutterhead approaches an unexplored small karst cave (digital twin shows it as a geological anomaly area). Sensors show that the pore water pressure 2 meters in front of the cutterhead (point P1) drops from 0.25 MPa to 0.18 MPa within 10 minutes. The dynamic threshold generation module generates thresholds based on current tunneling parameters (grouting filling rate 92%, tunneling speed 30 mm / min) and geological formation. The calculation showed that the current allowable water pressure drop threshold at this point was 0.05 MPa (the threshold for tightening in high-risk areas). The actual drop exceeded the threshold, triggering initial attention.

[0092] Time T0+40 minutes: The water pressure at point P1 continues to drop to 0.12 MPa, and the time-series coupling analysis module is activated. Testing shows: 1) The sudden drop in water pressure meets the condition ( , ); 2) Within the lag time window (35 minutes), the settlement rate at monitoring point S1 directly above the karst cave accelerated from 0.1 mm / d to 0.5 mm / d, exceeding the dynamic threshold. 3) The shield tunneling torque experienced short-term and drastic fluctuations, with a match rate of 85% with the characteristic spectrum of "cutterhead contact void". These three pieces of evidence are coupled with each other in time and space, triggering a Level I risk warning.

[0093] Time T0+41 minutes: The system initiates model-driven decision-making based on the water pressure drop rate (0.00325 MPa / min), the area of ​​influence (radius 8m), and the formation permeability coefficient. According to fluid simulation calculations, double-liquid grouting is required in the 5th and 7th rings behind the shield tail, with a target pressure of 0.75 MPa and an estimated grouting volume of 6 m³. At the same time, it is predicted that the cumulative settlement of point S1 will reach 3.2-3.8 mm in the next 2 hours, which is close to the alarm value of the station building pile foundation (4 mm).

[0094] Between T0 + 42 minutes and T0 + 60 minutes, the grouting command is automatically sent to the grouting trolley via the Modbus TCP protocol, and precise grouting is then initiated. The system pushes a structured report to the railway department, clearly and quantitatively presenting the prediction results and recommendations. Based on this, the railway department can conduct special inspections and deployments in advance.

[0095] Results: 25 minutes after grouting, the water pressure at point P1 stabilized at 0.22 MPa; the settlement rate at point S1 slowed to 0.08 mm / d, with a final cumulative settlement of 3.1 mm, consistent with the predicted value. This risk was detected 40 minutes in advance, and the response was timely and effective, without impacting the safe operation of the high-speed railway station.

[0096] Self-learning optimization example

[0097] Post-hoc self-learning optimization unit analysis of the event sequence revealed that the actual water pressure conduction and settlement response were faster than predicted by the initial model. Through reinforcement learning training, the time lag constant for this karst section was automatically determined. The timeframe was adjusted from 120 minutes to 70 minutes, and the threshold for matching the torque characteristic spectrum of "contact void" was fine-tuned to 82%. When the tunnel boring machine reached the next geologically similar section, the system's early warning lead time was increased to 55 minutes, the false alarm rate was further reduced, and the adaptability was significantly improved.

Claims

1. A construction risk analysis and emergency response method, characterized in that, Includes the following steps: S1. Construct a digital twin integrating a 3D geological model, tunnel model, environmental model, and sensor network; initialize the formation loss transfer function based on the geological survey report and laboratory tests. Initial version of parameters and time-series rule base; S2, Collection Inverting the actual stratigraphic loss distribution Model parameters are corrected by using ensemble Kalman filter data assimilation techniques; S3, based on the updated Calculate dynamic thresholds and match temporal association rules to predict the engineering status of the next time step in a rolling manner. S4. Invoke the physical evolution model that matches the corresponding risk type, reverse deduce the optimal treatment parameters, and synchronously send the instructions to the equipment and external units; S5. Collect data from the entire process, optimize the formation loss transfer function parameters and time series rule base through self-learning units, and complete the learning cycle.

2. The construction risk analysis and emergency response method according to claim 1, characterized in that, The sensor network of the digital twin mentioned in step 1 includes sensor arrays at different depths of the strata.

3. The construction risk analysis and emergency response method according to claim 1, characterized in that, The implementation process of the ensemble Kalman filter data assimilation technology in step 2 includes: inverting the ground loss distribution based on the real-time data of the tunnel boring machine PLC, and correcting the model parameters through 50-100 iterations.

4. The construction risk analysis and emergency response method according to claim 3, characterized in that, In step S3, the dynamic early warning threshold is calculated based on the formation loss state at the current time t, predicting the future preset time period. The theoretical risk index value of the post-monitoring point is used as the predicted value for the future time t+. The dynamic early warning threshold.

5. The construction risk analysis and emergency response method according to claim 4, characterized in that, In step S4, the self-learning unit adopts a reinforcement learning framework to adjust the safety factor of the formation loss transfer function and the time window of the time series rule.

Images