Deep space multi-disaster superposition instability evaluation and advance control system and method

Abstract

The application relates to a deep space multi-disaster superimposed instability evaluation and advanced control system and method, and belongs to the technical field of underground engineering safety and intelligent operation and maintenance. The system comprises a multi-source synchronous monitoring module, an edge computing and data fusion module, a disaster evolution dynamic evaluation and prediction module and an advanced control execution module which are sequentially connected. The multi-source synchronous monitoring module comprises a vibration sensor array, a seepage pressure sensor array, a strain sensor array and a corrosion sensor array. The method comprises the following steps: installing a multi-source synchronous monitoring module; collecting and fusing multi-source heterogeneous data; constructing a dynamic Bayesian network model, recursively calculating the prior and posterior probabilities of disaster nodes; calculating the path trigger probability and evaluating the consequence severity, screening out the key path with high trigger probability and serious consequences; locking the intervention target node according to the key path identification result, establishing a multi-objective optimization model, and quantitatively generating a targeted quantitative intervention strategy; implementing precise intervention measures and real-time monitoring of response data. The application can greatly improve the initiative and reliability of deep engineering safety operation and maintenance.

Description

Technical Field

[0001] This invention belongs to the field of underground engineering safety and intelligent operation and maintenance technology, specifically a deep space multi-hazard superposition instability assessment and advanced control system and method. Background Technology

[0002] Deep underground spaces face multiple coupled challenges from high ground stress, high seepage pressure, strong disturbances, and corrosive environments. The surrounding rock and supporting structures not only endure instantaneous dynamic loads such as water pressure seepage, mechanical vibration, and blasting impacts, but also undergo progressive degradation through creep and chemical corrosion over decades of service. Crucially, these hazards do not simply coexist, but rather follow specific physicochemical mechanisms in time and space, forming a "hazard evolution sequence" with a clear causal order and transmission law. For example, dynamic disturbances induce fissures, opening channels for groundwater seepage and corrosive media intrusion, while the presence of fluids softens the rock mass and accelerates corrosion, ultimately inducing structural instability under continuous ground stress and creep. This chain-like, progressive evolution pattern constitutes a fundamental threat to the long-term safety of deep engineering projects. Current technologies generally detect these hazards through multi-parameter monitoring networks, but how to identify the inherent transmission logic from the complex appearance of multiple hazards coexisting and achieve proactive prevention and control remains a core scientific challenge that urgently needs to be overcome.

[0003] Current assessment methods for the risk of multiple coupled disasters can be broadly categorized into "parallel weighted comprehensive evaluation." This type of method, which uses a comprehensive risk index to reflect the superposition of multiple disasters, essentially normalizes the intensity of each disaster and then performs linear or nonlinear superposition based on empirical or statistical weights. While this approach can quantify the overall risk level to some extent, it suffers from a fundamental flaw: it treats the interactions between disasters as static or quasi-static mutual influences, failing to explicitly characterize and deduce their dynamic causal triggering relationships. Specifically, traditional models excel at answering "how dangerous it is now," but cannot answer how the danger will develop; that is, they cannot identify, under specific geological and engineering conditions, which initial disaster (e.g., vibration) is most likely to trigger which secondary disaster sequence (e.g., fracture expansion and increased seepage). Due to the lack of predictive ability for disaster evolution paths, corresponding prevention and control strategies are inevitably lagging and diffuse. Therefore, measures are typically only taken to reinforce various disaster phenomena after the comprehensive risk index exceeds a threshold. This approach fails to provide precise intervention in the early stages of the disaster chain and is also unable to block the most critical transmission links, resulting in low prevention and control efficiency and high economic costs. Therefore, there is an urgent need to provide a system and method for assessing and controlling instability caused by multiple disasters in deep space, so as to achieve a fundamental shift in the prevention and control paradigm from post-disaster response to pre-disaster prevention. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a multi-hazard superimposed instability assessment and advanced control system and method for deep space engineering. This system enables accurate perception, efficient processing, dynamic prediction, and targeted prevention and control of multiple hazard risks, significantly improving the initiative and reliability of safe operation and maintenance in deep engineering projects. This method achieves a fundamental shift in the prevention and control paradigm from post-disaster response to pre-disaster prevention, greatly improving the effectiveness and efficiency of prevention and control while effectively reducing costs, and ensuring the long-term safe operation of deep engineering projects.

[0005] To achieve the above objectives, the present invention provides a deep space multi-hazard superimposed instability assessment and advanced control system, including a multi-source synchronous monitoring module, an edge computing and data fusion module, a disaster evolution dynamic assessment and prediction module, and an advanced control execution module;

[0006] The multi-source synchronous monitoring module includes a vibration sensor array, a seepage pressure sensor array, a strain sensor array, and a corrosion sensor array deployed at key disaster nodes. It is used to synchronously and in real time collect the original time-series signals of vibration acceleration, seepage pressure, creep strain, and corrosion current density to form a multi-source heterogeneous monitoring dataset.

[0007] The edge computing and data fusion module is connected to the multi-source synchronous monitoring module and is used to perform spatiotemporal alignment, filtering, feature extraction and feature fusion on multi-source heterogeneous monitoring data to generate and output a spatiotemporally synchronized fused feature dataset.

[0008] The disaster evolution dynamic assessment and prediction module incorporates a dynamic disaster chain model based on a Bayesian inference network, a multi-disaster trigger probability calculation unit, and an advanced intervention strategy generator. The dynamic disaster chain model based on a Bayesian inference network is used to dynamically update the activation probability of disaster nodes based on monitoring data. The multi-disaster trigger probability calculation unit is used to calculate the trigger probability between disasters in real time based on the posterior probability of the dynamic Bayesian network and identify the most dangerous disaster evolution path. The advanced intervention strategy generator is used to automatically generate accurate preventive intervention strategies based on the most dangerous disaster evolution path.

[0009] The advanced control execution module is used to generate corresponding intervention instructions based on the preventive intervention strategy, and drive the corresponding prevention and control equipment or guide manpower to carry out advanced intervention in disasters based on the intervention instructions.

[0010] As a preferred embodiment, the vibration sensor array uses a wideband accelerometer with a monitoring range of 0.1 to 1000 Hz; the pressure sensor array is arranged along the potential seepage path and structural interface.

[0011] In this invention, the multi-source synchronous monitoring module facilitates the simultaneous acquisition of multi-dimensional data from key monitoring nodes, avoiding the limitations of single-parameter monitoring. Simultaneously, it allows for the simultaneous acquisition of heterogeneous multi-source data to comprehensively depict the original state of multiple hazard coupling in deep engineering, providing comprehensive and accurate data support for subsequent assessment and prediction. The edge computing and data fusion module enables partial data processing at the edge, allowing only a small amount of processed data to be sent, eliminating the need to transmit massive amounts of raw data to a remote platform, thus reducing data transmission pressure and latency. The dynamic disaster evolution assessment and prediction module, relying on a Bayesian inference network dynamic disaster chain model, enables dynamic updates of disaster node activation probabilities, overcoming the limitations of traditional static assessments. Furthermore, combined with a multi-hazard trigger probability calculation unit, it accurately identifies the most dangerous disaster evolution paths, preemptively locking in core risks and transforming post-event risk discovery into pre-event risk prediction, reserving sufficient window time for prevention and control, and avoiding delays in prevention and control. Meanwhile, the proactive intervention strategy generator can automatically match precise preventative strategies based on the most dangerous disaster evolution path. Combined with the proactive control execution module, this enables rapid strategy implementation without cumbersome manual decision-making. It avoids the crude, all-encompassing approach of traditional defense, reducing prevention costs and engineering disruption. Furthermore, targeted intervention directly addresses the core risks, improving disaster prevention success rates and ensuring the operational safety of deep engineering projects. The proactive control execution module facilitates targeted implementation of prevention and control strategies, ensuring timely intervention.

[0012] The system has a simple structure and a high degree of intelligence. It can achieve accurate perception, efficient processing, dynamic prediction and targeted prevention and control of multiple disaster risks, which can greatly improve the initiative and reliability of safety operation and maintenance of deep engineering projects.

[0013] This invention also provides a method for assessing and proactively controlling multi-hazard instability in deep space, employing a multi-hazard instability assessment and proactive control system for deep space, comprising the following steps:

[0014] S1: Installation of the multi-source synchronous monitoring module;

[0015] Multi-source synchronous monitoring modules are deployed at key disaster nodes in the project to ensure that the time synchronization accuracy and spatial deployment density of each sensor meet the monitoring requirements.

[0016] S2: Acquisition and fusion processing of multi-source heterogeneous data;

[0017] Vibration acceleration data are collected synchronously and in real time using a multi-source synchronous monitoring module. osmotic pressure data Creep strain data Corrosion current density data This process generates a multi-source heterogeneous original monitoring dataset. After preprocessing and spatiotemporal alignment, a feature layer fusion algorithm is used to extract sensitive features from various data types, generating... Time-space synchronization fusion feature dataset , serving as real-time input data for dynamic Bayesian networks;

[0018] S3: Dynamic Quantification of Disaster Status and Bayesian Network Update;

[0019] A dynamic Bayesian network model is constructed, defining the disaster state vector and the corresponding relationship between nodes. Prior probabilities are determined based on historical disaster data and expert experience. The likelihood function is trained through machine learning, and the disaster state transition probability is calibrated. Combining real-time monitoring data, the prior and posterior probabilities of disaster nodes are recursively calculated using Bayes' theorem and temporal state transition rules, thereby realizing the dynamic quantification of disaster state and network update.

[0020] S4: Calculation of multiple disaster triggering probability and identification of key evolution paths;

[0021] Based on the posterior probability of dynamic Bayesian networks, a time-varying impact matrix is ​​constructed to quantify the triggering intensity between disasters, calculate the path triggering probability and assess the severity of consequences, and finally screen out the critical paths with high triggering probability and severe consequences.

[0022] S5: Quantitative generation of proactive intervention strategies based on multi-objective decision-making;

[0023] Based on the critical path identification results, the target node for intervention was identified. Establish a multi-objective optimization model with risk-return, cost, and disturbance as the core, and generate targeted quantitative intervention strategies based on the built-in standardized strategy mapping library;

[0024] S6: Strategy Execution and Model Closed-Loop Optimization;

[0025] The proactive control execution module automatically implements precise intervention measures based on quantitative intervention strategies and monitors response data in real time. Based on feedback information, it uses optimization algorithms to dynamically adjust Bayesian network parameters, forming a closed-loop iterative mechanism of monitoring, evaluation, intervention, and optimization, continuously improving the accuracy of disaster prediction and the effectiveness of proactive prevention and control.

[0026] As a preferred approach, in S2, preprocessing includes outlier removal, missing value completion, data standardization, and environmental noise filtering of the original data; in spatiotemporal alignment processing, based on the spatial location coordinates of the sensors and a unified timestamp, accurate spatial coordinate matching and strict time series synchronization of multi-source heterogeneous data are achieved.

[0027] As a preferred option, the process of dynamic quantification of disaster status and Bayesian network update in S3 is as follows:

[0028] S31: Model initialization; Construct a dynamic Bayesian network model and define... Disaster state vector at any time , ,in, Indicates the first The activation status of disaster-like states. , Indicates the first Disaster-like state activated. Indicates the first Disaster-like status not activated;

[0029] Definition of the first Disaster-like nodes ,make ,but for Time of the first Post-hoc probability of disaster activation;

[0030] S32: Determination of Prior Probabilities and Likelihood Functions; Based on historical disaster data and expert experience, determine the prior probabilities at the initial moment of each disaster node. , The likelihood function is obtained through training with a machine learning algorithm. A conditional probability table characterizes the state of a disaster. Evidence observed in the state The probability of;

[0031] Based on disaster evolution patterns and historical time-series data, the temporal state transition probability of each disaster node is determined. , , characterization Disaster status at all times The probabilistic law of state evolution at any given moment;

[0032] S33: Dynamic posterior probability update; combined with Timing fusion feature dataset Based on Bayes' theorem and temporal state transition probabilities, the posterior probability is updated independently for each type of disaster node. The temporal prior probability is first recursively derived using formula (1). Then, by using formula (2), we can obtain... Time of the first Posterior probability of disaster-like nodes ;

[0033] (1);

[0034] (2).

[0035] As a preferred option, in S33 of S3, the conditional probability table uses the maximum likelihood estimation method within a time sliding window, each Using monitoring data within a window period to perform likelihood function A re-estimation is performed to dynamically reflect changes in the coupling relationships between disasters.

[0036] As a preferred option, the process of calculating the multi-hazard triggering probability and identifying the critical evolution path in S4 is as follows:

[0037] S41: Construction of Time-Varying Influence Matrix; Based on the updated posterior probability, construct the time-varying state transition influence matrix of causal relationships between disasters. Its elements Characterizing disasters disaster The intensity of the triggering effect;

[0038] S42: Path trigger probability calculation; for all possible disaster evolution paths. , Calculate the current active node set according to formula (3). Chain trigger probability of departure ;

[0039] (3);

[0040] In the formula, ;

[0041] S43: Quantification of Consequence Severity; using the risk matrix method or analytic hierarchy process, quantify the severity of the consequences S for each path. ;

[0042] S44: Critical path selection; setting early warning thresholds Filter out the trigger probability And the severity of the consequences S The highest critical disaster evolution path .

[0043] As a preferred approach, in S5, the process of quantifying and generating proactive intervention strategies based on multi-objective decision-making is as follows:

[0044] S51: Target node determined; locked. The next critical disaster node that will be triggered ;

[0045] S52: Strategy Mapping Matching; The proactive intervention strategy generator has a pre-stored strategy mapping library, based on locked key disaster nodes. The type is matched with the corresponding specific control measures combination and parameter quantification rules from the mapping library; the types of measures include, but are not limited to: local active vibration damper power adjustment for nodes with increased vibration; preventive precision grouting for nodes with crack propagation; and drainage hole opening control for nodes with a surge in seepage pressure.

[0046] S53: Multi-objective optimization modeling; Construct a decision-making model with the objectives of maximizing returns while minimizing intervention costs and minimizing engineering operation disturbances, and determine the quantitative formulas for key control parameters;

[0047] S54: Intervention strategy generation; when At that time, a prevention strategy is generated; when blocking the crack propagation node in the evolution process from vibration to crack propagation and then to sudden increase in seepage, the preventive grouting pressure is obtained according to formula (4). ;

[0048] (4);

[0049] In the formula, The background grouting pressure is expressed in Pa. This is the pressure-probability conversion coefficient; The baseline severity threshold is used for the consequences. This represents the severity of the consequences of the critical path.

[0050] As a preferred approach, the policy execution and model closed-loop optimization process in S6 is as follows:

[0051] S61: Implementation of intervention strategies; the proactive control execution module executes intervention measures according to the intervention strategy and simultaneously monitors disaster response data after the intervention; among which, the intelligent grouting system follows the preventive grouting pressure... Grouting intervention was carried out, and the crack propagation rate and seepage pressure changes were monitored simultaneously after the intervention.

[0052] S62: Model parameter optimization; with the goal of minimizing the log-likelihood loss function, the expectation-maximization algorithm or gradient descent algorithm is used to adjust the prior probability, likelihood function and state transition probability of the dynamic Bayesian network based on the intervention feedback data.

[0053] S63: Closed-loop iteration; repeat S3 to S6 to achieve a closed-loop process of disaster monitoring, assessment, intervention and optimization, and continuously improve the model's prediction accuracy and intervention effect.

[0054] As a preferred option, in S4 of S4, the warning threshold is... Based on the potential severity S of the identified path Dynamic adjustments are made, and the adjusted warning threshold is obtained according to formula (5). ;

[0055] (5);

[0056] In the formula, The baseline warning threshold; This is the threshold adjustment coefficient. .

[0057] This invention aims to address the technical pain point in multi-hazard risk assessment of deep engineering projects, where existing methods rely solely on weighted superposition of multiple parameters for static status assessment. This fails to reveal the causal transmission relationships and chain evolution patterns between disasters, ultimately leading to delayed prevention and control measures and a passive post-disaster response mode. First, an integrated monitoring network and stable transmission channels are established at key disaster nodes in the engineering project, simultaneously and in real-time collecting multi-source heterogeneous data such as vibration acceleration, seepage pressure, creep strain, and corrosion current density, providing real-time data support for disaster evolution prediction. Second, a core topology is constructed based on geomechanical mechanisms and historical disaster case data. A disaster state vector containing binary activation states and corresponding disaster nodes are defined, constructing a dynamic disaster chain model based on a Bayesian inference network. This effectively clarifies the correlation between the posterior probability of disaster node activation and its state, thereby facilitating the dynamic quantification of disaster states. Next, based on the updated posterior probabilities of disaster nodes, a time-varying causal influence matrix between disasters is constructed. This effectively identifies all disaster evolution paths and accurately calculates the chain trigger probability. The severity of path consequences is quantified, and key disaster evolution paths with high trigger probabilities and severe consequences are selected. This allows for precise and efficient identification of critical disaster nodes on the path that are about to be triggered, providing core targets for subsequent targeted interventions. Subsequently, the advanced intervention strategy generator incorporates a standardized strategy mapping library, which helps to accurately match specific prevention and control measures based on the identified key nodes. Finally, based on the quantified intervention strategies, targeted advanced interventions are initiated within the early warning window of disaster chain evolution, precisely cutting off the disaster transmission chain. Simultaneously, the optimization of Bayesian network parameters based on feedback information enables automatic model iteration, contributing to continuous improvement in prevention and control capabilities.

[0058] This invention achieves innovative breakthroughs in the following aspects:

[0059] First, a multi-hazard chain dynamic prediction method was established: breaking through the traditional static assessment model of multi-parameter parallel weighting, it is the first to incorporate the causal transmission and chain evolution relationship between disasters into the core of dynamic risk assessment. Relying on the Bayesian inference network dynamic disaster chain model, it has achieved a cognitive leap from single-point current situation risk assessment to full-link evolution path prediction. It can dynamically deduce the condition trigger probability between disasters, providing core theoretical and model basis for proactive prevention and control.

[0060] Second, it enables advanced identification and early warning of key nodes in disaster transmission: Based on real-time data, the probability of disaster chain triggering is dynamically calculated. Combined with the dynamic binding mechanism of consequence severity and early warning threshold, a lower probability threshold is adopted for high-consequence disaster chains, which significantly advances the risk warning window to before the actual occurrence of secondary disasters. It accurately identifies the most likely and most severe disaster evolution sequence and key blocking nodes, which significantly improves the foresight, timeliness and pertinence of the early warning.

[0061] Third, a new paradigm of precise and proactive intervention has been formed: abandoning the traditional extensive prevention and control model of comprehensive defense, relying on the strategy mapping library to achieve precise matching of key nodes and prevention and control measures, implementing targeted intervention in the early window period of disaster chain evolution, cutting off the disaster transmission path with minimal cost and minimal engineering disturbance, greatly improving the accuracy of prevention and control measures and the efficiency of resource utilization, and building a proactive prevention and control model of targeted policy implementation and pre-disaster interruption.

[0062] Fourth, a fully closed-loop intelligent proactive prevention and control system has been constructed: multi-source real-time monitoring, Bayesian chain evolution and inference, targeted proactive intervention, and online model optimization are deeply integrated to form an intelligent decision-making system that integrates perception, assessment, intervention, and optimization, and is self-iterative and self-optimizing. This system breaks through the limitations of fragmented application of existing technologies and provides a complete and feasible proactive safety operation and maintenance solution for deep engineering to cope with the coupled threats of multiple disasters in extreme and complex environments.

[0063] This method is simple to implement and highly intelligent, achieving a fundamental shift in the prevention and control paradigm from post-disaster response to pre-disaster prevention. It can significantly improve the effectiveness and efficiency of prevention and control, while effectively reducing prevention and control costs and ensuring the long-term safe operation of deep engineering projects. Attached Figure Description

[0064] Figure 1 This is a principle block diagram of the system part of this invention;

[0065] Figure 2 This is a flowchart of the method section of this invention. Detailed Implementation

[0066] The invention will now be further described with reference to the accompanying drawings.

[0067] like Figure 1 As shown, the present invention provides a deep space multi-hazard superimposed instability assessment and advanced control system, including a multi-source synchronous monitoring module, an edge computing and data fusion module, a disaster evolution dynamic assessment and prediction module, and an advanced control execution module;

[0068] The multi-source synchronous monitoring module includes a vibration sensor array, a seepage pressure sensor array, a strain sensor array, and a corrosion sensor array deployed at key disaster nodes. It is used to synchronously and in real time collect the original time-series signals of vibration acceleration, seepage pressure, creep strain, and corrosion current density to form a multi-source heterogeneous monitoring dataset.

[0069] As a preferred embodiment, the vibration sensor array employs a wideband accelerometer with a monitoring frequency range of 0.1–1000 Hz to capture dynamic disturbance signals such as blasting and mechanical vibration.

[0070] As a preferred embodiment, the pressure sensor array is arranged along the potential seepage path and structural interface, and has the ability to respond quickly to dynamic pressure, so as to monitor the spatiotemporal evolution of seepage pressure.

[0071] The edge computing and data fusion module is connected to the multi-source synchronous monitoring module and is used to perform spatiotemporal alignment, filtering, feature extraction and feature fusion on multi-source heterogeneous monitoring data to generate and output a spatiotemporally synchronized fused feature dataset.

[0072] The disaster evolution dynamic assessment and prediction module incorporates a dynamic disaster chain model based on a Bayesian inference network, a multi-disaster trigger probability calculation unit, and an advanced intervention strategy generator. The dynamic disaster chain model based on a Bayesian inference network is used to dynamically update the activation probability of disaster nodes based on monitoring data. The multi-disaster trigger probability calculation unit is used to calculate the trigger probability between disasters in real time based on the posterior probability of the dynamic Bayesian network and identify the most dangerous disaster evolution path. The advanced intervention strategy generator is used to automatically generate accurate preventive intervention strategies based on the most dangerous disaster evolution path.

[0073] As a preferred option, the dynamic disaster chain model based on Bayesian inference networks constructs a core topology based on geomechanical mechanisms and historical disaster case data. The node conditional probability table is dynamically updated through online learning algorithms combined with real-time monitoring data, accurately reflecting the dynamic changes in the evolution of disasters and the spatiotemporal coupling relationship.

[0074] The advanced control execution module is used to generate corresponding intervention instructions based on the preventive intervention strategy, and drive the corresponding prevention and control equipment or guide the human to carry out advanced intervention in disasters according to the intervention instructions, so as to realize precise intervention operations for key links in the evolution of disasters.

[0075] In one embodiment of the present invention, the vibration sensor array preferably adopts a triaxial MEMS accelerometer (such as ADXL354), with a sampling frequency of 2kHz and a range of ±50g. It is installed on the key cross-section of the surrounding rock of the roadway and the key nodes of the support structure. The time synchronization of all sensors is achieved through a synchronous triggering circuit, and the synchronization accuracy is better than 1ms.

[0076] Meanwhile, the pressure sensor array uses fiber optic grating piezometers (range 0–10 MPa, accuracy 0.1% FS), deployed every 5 m along the potential seepage path, and also arranged on the working face floor and sides, with a sampling frequency of 10 Hz. The strain sensor array uses vibrating wire strain gauges (range ±3000 με, accuracy 0.1% FS), with one cross-section every 10 m along the tunnel axis, and four gauges arranged in each cross-section (roof, floor, and sides), with a sampling frequency of 1 Hz. The corrosion sensor array uses linearly polarized resistance probes, embedded at the ends of anchor bolts and the grout interface, to monitor corrosion current density, with a sampling frequency of 0.01 Hz. All sensors are connected to the edge computing and data fusion module (edge ​​computing node) via RS485 bus and are synchronized with a unified time.

[0077] In this invention, the multi-source synchronous monitoring module facilitates the simultaneous acquisition of data from all dimensions at key monitoring nodes, avoiding the limitations of single-parameter monitoring. Simultaneously, it allows for the synchronous acquisition of heterogeneous multi-source data to comprehensively depict the original state of multiple hazard coupling in deep engineering, providing comprehensive and accurate data support for subsequent assessment and prediction, and reducing risk misjudgments caused by missing or asynchronous data. The edge computing and data fusion modules enable localized data spatiotemporal alignment, filtering, and feature fusion, eliminating the need to transmit massive amounts of raw data to remote platforms, significantly reducing data transmission pressure and latency, and ensuring the real-time performance of the fused feature dataset. Furthermore, filtering and noise reduction, along with feature extraction, optimize data quality and eliminate redundant information, allowing subsequent assessment modules to focus on core, effective data and improving assessment efficiency. By setting up a dynamic disaster evolution assessment and prediction module, and relying on a Bayesian inference network dynamic disaster chain model, the activation probability of disaster nodes can be dynamically updated, overcoming the limitations of traditional static assessments. Simultaneously, by combining a multi-disaster trigger probability calculation unit, it can accurately identify the most dangerous disaster evolution path, locking in the core risk in advance, transforming post-event risk discovery into pre-event risk prediction, reserving sufficient window time for prevention and control, and avoiding delays in prevention and control. Furthermore, the advanced intervention strategy generator can automatically match precise preventative strategies based on the most dangerous disaster evolution path, and, in conjunction with the advanced control execution module, enable rapid strategy implementation without cumbersome manual decision-making. This avoids the crude prevention and control model of comprehensive defense, reducing prevention and control costs and engineering disturbances, while targeted intervention directly addresses the core risk, improving the success rate of disaster prevention and control, and ensuring the operational safety of deep engineering projects. The advanced control execution module facilitates targeted implementation of prevention and control strategies, ensuring the timeliness of prevention and control.

[0078] The system has a simple structure and a high degree of intelligence. It can achieve accurate perception, efficient processing, dynamic prediction and targeted prevention and control of multiple disaster risks, which can greatly improve the initiative and reliability of safety operation and maintenance of deep engineering projects.

[0079] like Figure 2As shown, the present invention also provides a method for assessing and proactively controlling multi-hazard instability in deep space, employing a multi-hazard instability assessment and proactive control system for deep space, comprising the following steps:

[0080] S1: Installation of the multi-source synchronous monitoring module;

[0081] Multi-source synchronous monitoring modules are deployed at key disaster nodes in the project to ensure that the time synchronization accuracy and spatial deployment density of each sensor meet the monitoring requirements; at the same time, an integrated monitoring network and a stable data transmission channel are built to ensure the real-time, continuous and spatially correlated nature of the monitoring data.

[0082] S2: Acquisition and fusion processing of multi-source heterogeneous data;

[0083] Vibration acceleration data are collected synchronously and in real time using a multi-source synchronous monitoring module. osmotic pressure data Creep strain data Corrosion current density data This process generates a multi-source heterogeneous original monitoring dataset. After preprocessing and spatiotemporal alignment in the edge computing and data fusion module, a feature layer fusion algorithm is used to extract sensitive features from various data types, remove redundant features, and generate... Time-space synchronization fusion feature dataset , serving as real-time input data for dynamic Bayesian networks;

[0084] As a preferred approach, preprocessing includes operations such as outlier removal, missing value completion, data standardization, and environmental noise filtering of the original data. In spatiotemporal alignment processing, based on the spatial location coordinates of the sensors and a unified timestamp, precise matching of spatial coordinates and strict synchronization of time series of multi-source heterogeneous data are achieved to eliminate spatiotemporal heterogeneity bias.

[0085] As a preferred approach, the feature layer fusion algorithm can employ a feature fusion algorithm based on a multimodal attention mechanism. The fusion process is as follows: First, the vibration acceleration signal is bandpass filtered (0.1-1000Hz) and Hilbert transformed to extract the envelope energy; the pressure data is denoised using moving average; the strain data is decomposed into trends (using LOESS local weighted regression); and the corrosion current density is estimated using Kalman filtering. Then, the feature vectors of each mode are mapped to the same dimension (e.g., 64-dimensional) through a fully connected network, and the cross-modal weights are calculated using multi-head attention, finally obtaining the fused feature vector E(t).

[0086] S3: Dynamic Quantification of Disaster Status and Bayesian Network Update;

[0087] A dynamic Bayesian network model is constructed to obtain a dynamic disaster chain model based on a Bayesian inference network. The disaster state vector and the corresponding relationship between nodes are defined. The prior probability is determined based on historical disaster data and expert experience. The likelihood function is trained through machine learning, and the disaster state transition probability is calibrated. Combining real-time monitoring data, the prior probability and posterior probability of disaster nodes are recursively calculated using Bayes' theorem and the temporal state transition law, so as to realize the dynamic quantification of disaster state and network update.

[0088] As a preferred approach, the process of dynamic quantification of disaster status and Bayesian network update is as follows:

[0089] S31: Model initialization; constructing a dynamic Bayesian network model (DBN) and defining... Disaster state vector at any time , ,in, Indicates the first The activation status of disaster-like states. , Indicates the first Disaster-like state activated. Indicates the first Disaster-like status not activated;

[0090] Definition of the first Disaster-like nodes ,make ,but for Time of the first Post-hoc probability of disaster activation;

[0091] In one embodiment of the present invention, the process of defining a disaster node is as follows: "Vibration energy exceeds threshold" (≥10) -4 m 2 / s 3 ), This is defined as "crack propagation" (acoustic emission event rate > 50 times / h). This is defined as a "sudden increase in osmotic pressure" (pressure rise rate > 0.5 MPa / h). The corrosion current density is increased (>10 μA / cm²). "Accelerated creep deformation" (strain rate > 10) -5 / h). The causal relationship between nodes is: → → → ,as well as → (Corrosion weakens the fracture surface), forming a directed acyclic graph.

[0092] S32: Determination of Prior Probabilities and Likelihood Functions; Based on historical disaster data and expert experience, determine the prior probabilities at the initial moment of each disaster node. , ( ), The likelihood function is obtained through training with a machine learning algorithm. A conditional probability table characterizes the state of a disaster. Evidence observed in the state The probability of;

[0093] Based on disaster evolution patterns and historical time-series data, the temporal state transition probability of each disaster node is determined. , , characterization Disaster status at all times The probabilistic law of state evolution at any given moment;

[0094] S33: Dynamic posterior probability update; combined with Timing fusion feature dataset Based on Bayes' theorem and temporal state transition probabilities, the posterior probability is updated independently for each type of disaster node. The temporal prior probability is first recursively derived using formula (1). Then, by using formula (2), we can obtain... Time of the first Posterior probability of disaster-like nodes ;

[0095] (1);

[0096] (2);

[0097] In the formula, This is the time-series prior probability obtained recursively based on the monitoring data from the previous time step.

[0098] As a preferred option, in S33 of S3, the conditional probability table uses the maximum likelihood estimation method within a time sliding window, each Hour( Using monitoring data within the window period to analyze the likelihood function Re-estimation is performed to dynamically reflect changes in the coupling relationships between disasters and to continuously optimize the accuracy of probability mapping.

[0099] S4: Calculation of multiple disaster triggering probability and identification of key evolution paths;

[0100] The multi-hazard triggering probability calculation unit is based on the posterior probability of a dynamic Bayesian network. It constructs a time-varying impact matrix to quantify the triggering intensity between hazards, calculates the path triggering probability and assesses the severity of consequences, and finally selects the critical paths with high triggering probability and severe consequences.

[0101] As a preferred option, the process of calculating the multi-hazard triggering probability and identifying the critical evolution path is as follows:

[0102] S41: Construction of Time-Varying Influence Matrix; Based on the updated posterior probability, construct the time-varying state transition influence matrix of causal relationships between disasters. Its elements Characterizing disasters disaster The intensity of the triggering effect;

[0103] S42: Path trigger probability calculation; for all possible disaster evolution paths. , Calculate the current active node set according to formula (3). Chain trigger probability of departure ;

[0104] (3);

[0105] In the formula, ;

[0106] S43: Quantification of Consequence Severity; using the risk matrix method or analytic hierarchy process, quantify the severity of the consequences S for each path. ;

[0107] S44: Critical path selection; setting early warning thresholds (Based on engineering safety standards), the trigger probability is selected. And the severity of the consequences S The highest critical disaster evolution path .

[0108] As a preferred option, the warning threshold Based on the potential severity S of the identified path Dynamic adjustments are made, and the adjusted warning threshold is obtained according to formula (5). ;

[0109] (5);

[0110] In the formula, The baseline warning threshold; This is the threshold adjustment coefficient. Lower probability thresholds are used for high-consequence chains to trigger earlier warnings.

[0111] In one embodiment of the present invention, a baseline warning threshold is set. =0.6, threshold adjustment coefficient =0.3. Regarding the severity of the consequences. The path with value 30, after adjustment =0.6×exp(-0.3×(30 / 10-1))=0.6×exp(-0.6)=0.6×0.5488≈0.329. This threshold is lower than the baseline value, making high-consequence paths more likely to trigger early warnings and enabling earlier intervention.

[0112] S5: Quantitative generation of proactive intervention strategies based on multi-objective decision-making;

[0113] The proactive intervention strategy generator identifies target nodes for intervention based on critical path identification results. Establish a multi-objective optimization model with risk-return, cost, and disturbance as the core, and generate targeted quantitative intervention strategies based on the built-in standardized strategy mapping library;

[0114] As a preferred approach, the process of quantifying and generating proactive intervention strategies based on multi-objective decision-making is as follows:

[0115] S51: Target node determined; locked. The next critical disaster node that will be triggered ;

[0116] S52: Strategy Mapping Matching; The proactive intervention strategy generator has a pre-stored strategy mapping library, based on locked key disaster nodes. The type is matched with the corresponding specific control measures combination and parameter quantification rules from the mapping library; the types of measures include, but are not limited to: local active vibration damper power adjustment for nodes with increased vibration; preventive precision grouting for nodes with crack propagation; and drainage hole opening control for nodes with a surge in seepage pressure.

[0117] S53: Multi-objective optimization modeling; Construct a decision-making model with the objectives of maximizing returns while minimizing intervention costs and minimizing engineering operation disturbances, and determine the quantitative formulas for key control parameters;

[0118] S54: Intervention strategy generation; when At that time, a prevention strategy is generated; when blocking the crack propagation node in the evolution process from vibration to crack propagation and then to sudden increase in seepage, the preventive grouting pressure is obtained according to formula (4). ;

[0119] (4);

[0120] In the formula, The background grouting pressure is expressed in Pa. This is the pressure-probability conversion coefficient; The baseline severity threshold is used for the consequences.

[0121] In one embodiment of the present invention, the background grouting pressure Take 2.0 MPa as the pressure-probability conversion factor. Take 0.5 MPa as the baseline severity threshold. Take 10. When =0.85, =0.6, When =15, the calculation is as follows =2.0+0.5×(0.85-0.6)×ln(1+15 / 10)=2.0+0.125×ln(2.5)=2.0+0.125×0.916≈2.1145MPa. The grout used is a cement-water glass two-component grout with a water-cement ratio of 0.8:1 and a grouting rate of 20L / min.

[0122] S6: Strategy Execution and Model Closed-Loop Optimization;

[0123] The proactive control execution module automatically implements precise intervention measures based on quantitative intervention strategies and monitors response data in real time. Based on feedback information, it uses optimization algorithms to dynamically adjust Bayesian network parameters, forming a closed-loop iterative mechanism of monitoring, evaluation, intervention, and optimization, continuously improving the accuracy of disaster prediction and the effectiveness of proactive prevention and control.

[0124] As a preferred approach, the process of strategy execution and model closed-loop optimization is as follows:

[0125] S61: Implementation of intervention strategies; the proactive control execution module executes intervention measures according to the intervention strategy and simultaneously monitors disaster response data after the intervention; among which, the intelligent grouting system follows the preventive grouting pressure... Grouting intervention was carried out, and the crack propagation rate and seepage pressure changes were monitored simultaneously after the intervention.

[0126] S62: Model parameter optimization; with the goal of minimizing the log-likelihood loss function, the expectation-maximization algorithm or gradient descent algorithm is used to adjust the prior probability, likelihood function and state transition probability of the dynamic Bayesian network based on the intervention feedback data.

[0127] S63: Closed-loop iteration; repeat S3 to S6 to achieve a closed-loop process of disaster monitoring, assessment, intervention and optimization, and continuously improve the model's prediction accuracy and intervention effect.

[0128] This invention aims to address the technical pain point in multi-hazard risk assessment of deep engineering projects. Existing methods rely solely on static status assessment using multi-parameter weighted superposition, failing to reveal the causal transmission relationships and chain evolution patterns between hazards. This ultimately leads to delayed prevention and control measures and a passive "post-disaster response" mode. Targeting the complex risk scenarios of interaction and chain evolution of hazards such as water pressure seepage, mechanical vibration, surrounding rock creep, and structural corrosion in deep underground engineering projects, this invention constructs a proactive prevention and control system encompassing perception, assessment, intervention, and optimization. This achieves a fundamental shift in the prevention and control paradigm from "status assessment" to "process prediction," and from "post-disaster response" to "pre-disaster prevention," significantly improving the foresight and accuracy of assessment and control of chain risks in deep engineering projects. First, multi-source synchronous monitoring modules are deployed at key hazard nodes in the project, establishing an integrated monitoring network and stable transmission channels to synchronously and in real-time collect multi-source heterogeneous data such as vibration acceleration, seepage pressure, creep strain, and corrosion current density. After preprocessing and spatiotemporal alignment, sensitive features are extracted through feature layer fusion to generate a spatiotemporally synchronized fused feature dataset, providing real-time data support for hazard evolution prediction. Secondly, a dynamic disaster chain model based on a Bayesian inference network was constructed. This model relies on geomechanical mechanisms and historical disaster case data to build its core topology, defining disaster state vectors with binary activation states and corresponding disaster nodes, clarifying the correlation between the posterior probability of disaster node activation and its state. Initial prior probabilities were calibrated based on historical disaster data and expert experience. A likelihood function conditional probability table was obtained through machine learning training, and time-series state transition probabilities were calibrated in conjunction with disaster evolution patterns. Then, based on real-time fused data, the posterior probability of disaster node activation was recursively updated using Bayes' theorem, dynamically quantifying the disaster state. Next, based on the updated posterior probabilities of disaster nodes, a time-varying causal influence matrix between disasters was constructed, outlining all disaster evolution paths and calculating chain trigger probabilities. The risk matrix method was used to quantify the severity of path consequences, and the warning threshold was dynamically adjusted according to the potential severity of path consequences. Key disaster evolution paths with high trigger probabilities and severe consequences were selected, identifying key disaster nodes on the paths that are about to be triggered, providing core targets for targeted intervention. Subsequently, the advanced intervention strategy generator has a built-in standardized strategy mapping library to establish a one-to-one correspondence between key disaster node types, prevention and control measure combinations, and parameter quantification rules. Based on the locked key nodes, exclusive prevention and control measures are accurately matched. Finally, based on the matched prevention and control measures, a multi-objective optimization decision-making model is constructed to "maximize risk reduction benefits, minimize intervention costs, and minimize engineering operation disturbances." Intervention control parameters (such as preventive grouting pressure) are quantified and generated. Within the early warning window period of disaster chain evolution, targeted advanced intervention is initiated to accurately cut off the disaster transmission chain.Simultaneously, disaster response data is monitored after intervention. With the goal of minimizing the log-likelihood loss function, the model's prior probability and state transition probability are optimized through expectation maximization and gradient descent algorithms. Combined with the dynamic updating of the conditional probability table through online learning, the evaluation, identification, intervention, and optimization processes are repeatedly executed to achieve a closed-loop operation of model self-iteration and continuous improvement of prevention and control capabilities.

[0129] This method is simple to implement and highly intelligent, achieving a fundamental shift in the prevention and control paradigm from post-disaster response to pre-disaster prevention. It can significantly improve the effectiveness and efficiency of prevention and control, while effectively reducing prevention and control costs and ensuring the long-term safe operation of deep engineering projects.

Claims

1. A method for assessing and proactively controlling multi-hazard instability in deep space, characterized in that, Includes the following steps: S1: Deploy multi-source synchronous monitoring modules at key disaster nodes in the project; S2: Utilize a multi-source synchronous monitoring module to synchronously and in real-time acquire vibration acceleration data. osmotic pressure data Creep strain data Corrosion current density data This forms a multi-source heterogeneous original monitoring dataset; After preprocessing and spatiotemporal alignment, a feature layer fusion algorithm is used to extract sensitive features from various types of data to generate... Time-space synchronization fusion feature dataset , serving as real-time input data for dynamic Bayesian networks; S3: Construct a dynamic Bayesian network model, define the disaster state vector and the corresponding relationship between nodes, determine the prior probability based on historical disaster data and expert experience, train the likelihood function through machine learning, and calibrate the disaster state transition probability; combine real-time monitoring data, utilize Bayes' theorem and time-series state transition rules, recursively calculate the prior probability and posterior probability of disaster nodes, and realize the dynamic quantification of disaster state and network update. The process of achieving dynamic quantification and network updates of disaster status is as follows: S31: Construct a dynamic Bayesian network model and define... Disaster state vector at any time , ,in, Indicates the first The activation status of disaster-like states. , Indicates the first Disaster-like state activated. Indicates the first Disaster-like status not activated; Definition of the first Disaster-like nodes ,make ,but for Time of the first Post-hoc probability of disaster activation; S32: Based on historical disaster data and expert experience, determine the prior probability of each disaster node at its initial moment. , The likelihood function is obtained through training with a machine learning algorithm. A conditional probability table characterizes the state of a disaster. Evidence observed in the state The probability of; Based on disaster evolution patterns and historical time-series data, the temporal state transition probability of each disaster node is determined. , , characterization Disaster status at all times The probabilistic law of state evolution at any given moment; S33: Combination Timing fusion feature dataset Based on Bayes' theorem and temporal state transition probabilities, the posterior probability is updated independently for each type of disaster node. The temporal prior probability is first recursively derived using formula (1). Then, by using formula (2), we can obtain... Time of the first Posterior probability of disaster-like nodes ; （1）； （2）； S4: Based on the posterior probability of dynamic Bayesian networks, a time-varying impact matrix is ​​constructed to quantify the triggering intensity between disasters, calculate the path triggering probability and assess the severity of consequences, and finally screen out the critical paths with high triggering probability and severe consequences. S5: Identify the target node for intervention based on the critical path identification results. A multi-objective optimization model is established, and targeted quantitative intervention strategies are generated based on the built-in standardized strategy mapping library. S6: The proactive control execution module automatically implements precise intervention measures based on the quantitative intervention strategy and monitors response data in real time; based on feedback information, it uses optimization algorithms to dynamically adjust the Bayesian network parameters.

2. The method for assessing and proactively controlling multi-hazard instability in deep space according to claim 1, characterized in that, In S2, preprocessing includes outlier removal, missing value completion, data standardization, and environmental noise filtering of the raw data. In spatiotemporal alignment processing, based on the spatial location coordinates of the sensors and a unified timestamp, precise spatial coordinate matching and strict time series synchronization of multi-source heterogeneous data are achieved.

3. The method for assessing and proactively controlling multi-hazard instability in deep space according to claim 1, characterized in that, In S4, the process of selecting critical paths with high trigger probability and severe consequences is as follows: S41: Construct a time-varying state transition influence matrix of causal relationships between disasters based on the updated posterior probabilities. Its elements Characterizing disasters disaster The intensity of the triggering effect; S42: For all possible disaster evolution paths , Calculate the current active node set according to formula (3). Chain trigger probability of departure ; （3）； In the formula, ; S43: Use the risk matrix method or the analytic hierarchy process (AHP) to quantify the severity of the consequences S for each path. ; S44: Set warning threshold Filter out the trigger probability And the severity of the consequences S The highest critical disaster evolution path .

4. The method for assessing and proactively controlling multi-hazard instability in deep space according to claim 1, characterized in that, In S5, the process of generating targeted quantitative intervention strategies is as follows: S51: Lock The next critical disaster node that will be triggered ; S52: The proactive intervention strategy generator has a pre-stored strategy mapping library, based on locked key disaster nodes. The type is matched with the corresponding specific control measures combination and parameter quantification rules from the mapping library; the types of measures include, but are not limited to: local active vibration damper power adjustment for nodes with increased vibration; preventive precision grouting for nodes with crack propagation; and drainage hole opening control for nodes with a surge in seepage pressure. S53: Construct a decision-making model with the objectives of minimizing risk and maximizing returns, minimizing intervention costs, and minimizing engineering operation disturbances, and determine the quantitative formulas for key control parameters; S54: When At that time, a prevention strategy is generated; when blocking the crack propagation node in the evolution process from vibration to crack propagation and then to sudden increase in seepage, the preventive grouting pressure is obtained according to formula (4). ; （4）； In the formula, The background grouting pressure is expressed in Pa. This is the pressure-probability conversion coefficient; The baseline severity threshold is used for the consequences.

5. The method for assessing and proactively controlling multi-hazard instability in deep space according to claim 1, characterized in that, In S44 of S4, the warning threshold Based on the potential severity S of the identified path Dynamic adjustments are made according to formula (5); （5）； In the formula, The baseline warning threshold; This is the threshold adjustment coefficient. .

6. A deep space multi-hazard superimposed instability assessment and advance control system, used to implement the deep space multi-hazard superimposed instability assessment and advance control method as described in any one of claims 1 to 5, characterized in that, It includes a multi-source synchronous monitoring module, an edge computing and data fusion module, a disaster evolution dynamic assessment and prediction module, and an advanced control execution module; The multi-source synchronous monitoring module includes a vibration sensor array, a seepage pressure sensor array, a strain sensor array, and a corrosion sensor array deployed at key disaster nodes, used to collect multi-source heterogeneous monitoring datasets; The edge computing and data fusion module is connected to the multi-source synchronous monitoring module and is used to process multi-source heterogeneous monitoring data to generate and output a spatiotemporally synchronized fusion feature dataset. The disaster evolution dynamic assessment and prediction module incorporates a dynamic disaster chain model based on a Bayesian inference network, a multi-disaster trigger probability calculation unit, and an advanced intervention strategy generator. The dynamic disaster chain model based on a Bayesian inference network is used to dynamically update the activation probability of disaster nodes based on monitoring data. The multi-disaster trigger probability calculation unit is used to calculate the trigger probability between disasters in real time based on the posterior probability of the dynamic Bayesian network and identify the most dangerous disaster evolution path. The advanced intervention strategy generator is used to automatically generate accurate preventive intervention strategies based on the most dangerous disaster evolution path. The advanced control execution module is used to generate corresponding intervention instructions based on the preventive intervention strategy, and drive the corresponding prevention and control equipment or guide manpower to carry out advanced intervention in disasters based on the intervention instructions.

7. The deep space multi-hazard superimposed instability assessment and advanced control system according to claim 6, characterized in that, The vibration sensor array uses a wideband accelerometer; the pressure sensor array is arranged along the potential seepage path and structural interface.

Images