AI-driven capital construction risk operation optimization management system

Abstract

The invention discloses an infrastructure risk operation optimization management system based on AI driving, and belongs to the field of computer data processing and commercial management, and the system comprises a multi-modal causal twinning construction module which integrates on-site multi-modal data streams to construct a dynamic space-time causal map; the risk evolution deduction module is used for performing anti-fact simulation based on a causal atlas to construct a prospective risk model; the collaborative configuration optimization module is used for solving an optimal collaborative defense strategy according to the risk model; the instruction analysis and digital prescription generation module is used for analyzing the defense strategy into a job digital prescription for a specific risk scene; and the intervention efficiency attribution and evolution correction module performs attribution analysis according to the execution effect of the digital prescription and adaptively updates the causal atlas. According to the method, a comprehensive method of constructing a dynamic causal map for risk deduction, coupling resource constraints for collaborative optimization and performing closed-loop feedback on a correction model is adopted, and active prediction, accurate intervention and continuous learning optimization of capital construction risks can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer data processing and business management, and particularly to an infrastructure risk operation optimization management system based on AI driving. BACKGROUND

[0002] Large infrastructure construction, as an important pillar of the national economy, involves complex planning, scheduling, operation and risk control in project management. Under the wave of digital transformation, project management is increasingly relying on data-driven decision support systems, aiming to optimize resource allocation, improve operational efficiency, and ensure personnel and asset safety through analysis of massive data. It usually integrates business intelligence, data mining and operation management technologies to achieve fine management of the entire project life cycle.

[0003] In the prior art, risk management of infrastructure projects usually relies on isolated information systems. For example, visual management using building information models, collecting equipment status data through Internet of Things sensors, or issuing weather warnings using environmental monitoring systems. In terms of data analysis, machine learning models based on historical data statistics are used to predict the probability of a single risk event, or fixed threshold rules are set for alarm. In terms of decision support, risk warning and resource scheduling, operation arrangement are usually separate processes, relying on the experience of management personnel for manual coordination and instruction.

[0004] However, the above prior art solutions have obvious defects. First, the information islands formed between various data systems cannot effectively reveal the dynamic and mutually coupled dependencies between different physical entities, resulting in blind spots in risk awareness. Second, statistical-based prediction models cannot explain their prediction results and cannot deduce rare but significant chain risk events. Finally, the separation of risk warning and operational decision-making leads to response delays, frequent resource allocation conflicts, and lack of evaluation and feedback on intervention measures, making the management system unable to continuously improve from practice and difficult to adapt to the complex and changing construction site environment. SUMMARY

[0005] To solve the above problems, the present application provides an infrastructure risk operation optimization management system based on AI driving, which adopts a comprehensive method of constructing a dynamic causal map for risk deduction, coupling resource constraints for collaborative optimization, and closed-loop feedback correction of the model, which can realize active foresight, precise intervention and continuous learning optimization of infrastructure risks.

[0006] The above objectives can be achieved by the following solutions: The AI-driven infrastructure risk operation optimization management system comprises a multi-modal causal twin construction module for real-time fusion of multi-modal data streams of an infrastructure site, the multi-modal data streams comprising building information model data, Internet of Things sensor array data, environmental monitoring data and personnel physiological sign data, deduction of nonlinear causal dependency relationships between physical entities by causal network, and construction of a dynamic spatiotemporal causal graph; a risk evolution deduction module for deduction of risk evolution paths and cascading effects triggered by a single point risk based on the dynamic spatiotemporal causal graph, construction of a forward-looking risk model by applying virtual disturbance to key nodes for counterfactual simulation; a collaborative configuration optimization module for construction of a dynamic programming model by taking the forward-looking risk model as a constraint condition and coupling resource dependency relationships of the dynamic spatiotemporal causal graph, solution of an optimal collaborative defense strategy under different resource constraints, and output of a time-varying risk intervention strategy matrix; an instruction analysis and digital prescription generation module for analysis of the time-varying risk intervention strategy matrix into an instruction paradigm and generation of an operation digital prescription for a specific dynamic risk scenario in combination with a real-time state snapshot of the dynamic spatiotemporal causal graph; and an intervention efficiency attribution and evolution correction module for continuous monitoring of a site state evolution after application of the operation digital prescription, comparison of a real evolution result with a prediction result of the forward-looking risk model, calculation of an efficiency deviation of an intervention measure and attribution analysis, and adaptive update of causal link weight in the dynamic spatiotemporal causal graph according to an attribution confidence.

[0007] Optionally, the multi-modal causal twin construction module comprises a data spatiotemporal alignment unit for mapping and synchronizing time stamps and spatial coordinates in the Internet of Things sensor array data, the environmental monitoring data and the personnel physiological sign data with the building information model data as a three-dimensional space reference, and output of a multi-modal data set; a cross-modal hidden feature extraction unit for learning and extracting internal states and interaction behaviors of physical entities based on the multi-modal data set, and obtaining a hidden feature vector; a causal structure learning unit for time series analysis of the hidden feature vector, quantization of nonlinear causal dependency relationships, search and determination of directed causal links and nonlinear function relationships of each hidden feature under the condition of meeting non-circulation constraints; and a graph instantiation and weighting unit for instantiation of a dynamic spatiotemporal causal graph with physical entities of the building information model data as nodes, the directed causal links as edges, and the influence strength of the nonlinear function relationship as weight.

[0008] Optionally, the risk evolution deduction module includes: a source point positioning and disturbance set generation unit, which is used to analyze the topological structure of the dynamic space-time causal graph, calculate the causal out-degree and influence weight of each node, locate the risk source point based on the calculation results, and generate a parameterized virtual disturbance set around the risk source point; a causal chain deduction and cascade effect simulation unit, which is used to apply the parameters in the parameterized virtual disturbance set to the risk source point one by one, perform discrete time iterative deduction in the dynamic space-time causal graph, simulate multiple risk evolution paths triggered by single-point risks and their cross-influences, and obtain a simulated evolution path; a forward-looking risk field construction unit, which is used to statistically aggregate the simulated evolution paths, construct a risk propagation probability field in a four-dimensional space-time coordinate system, quantify the probability that any future space-time point will be affected by the risk of a specific source point, and serve as a forward-looking risk model.

[0009] Optionally, the parameterized virtual disturbance set includes: a state offset vector set and a conduction failure operator set, wherein: the state offset vector set is used to simulate the extreme deterioration scenario of the physical entity's own attribute parameters in a statistical sense; the conduction failure operator set is used to simulate the local interruption or nonlinear distortion of the causal relationship transmission path between entities in the dynamic spatiotemporal causal graph.

[0010] Optionally, the system also includes: matching the latent feature vector with the simulation evolution path, calculating the distance between the initial state and the key feature points, evaluating the degree of convergence of the current infrastructure site state to each risk evolution path, and generating a path activation confidence vector.

[0011] Optionally, the collaborative configuration optimization module includes: a multi-dimensional constraint space construction unit, which is used to convert the risk propagation probability field of the forward-looking risk model into the state transfer cost of the dynamic programming model, and extract the logical dependency relationship between resource scheduling and task execution from the dynamic spatiotemporal causal graph as a state constraint, and jointly define the solution space; a strategy iteration solution unit, which is used to recursively solve the state-action sequence that can accumulate risk costs under the condition of satisfying the state constraints within the solution space to obtain the optimal collaborative defense strategy; a strategy matrix generation unit, which is used to map and discretize the optimal collaborative defense strategy in the time dimension to generate a time-varying risk intervention strategy matrix.

[0012] Optionally, the instruction parsing and digital prescription generation module includes: a policy matrix parsing unit, used to decode the time-varying risk intervention policy matrix into a structured intervention rule primitive; a dynamic risk scenario matching unit, used to identify dynamic risk scenarios with the path activation confidence vector, and match the optimal intervention rule from the structured intervention rule primitive; a digital prescription generation unit, used to call the real-time status snapshot of the dynamic spatiotemporal causal graph to instantiate the parameters in the optimal intervention rule, and encapsulate it into a standardized data package as a digital prescription for the operation and send it to the on-site execution end.

[0013] Optionally, the standardized data package includes: an operation instruction set, risk context data and performance verification parameters, wherein: the operation instruction set is used to define the intervention actions, parameter thresholds and execution timing for the target physical entity; the risk context data is used to encapsulate the activated risk evolution path, the real-time status of the risk source node, and the expected loss assessment after the intervention fails; the performance verification parameters are used to specify the monitoring indicators and state migration targets that need to be met for the intervention to be successful.

[0014] Optionally, the intervention efficacy attribution and evolution correction module includes: an effect monitoring and deviation calculation unit, which is used to continuously track the on-site data flow after the digital prescription for the operation is issued, and compare the actual monitoring indicators with the state migration targets predicted by the forward-looking risk model to quantify the efficacy deviation vector; a causal attribution and confidence assessment unit, which is used to perform contribution traceability analysis on the dynamic spatiotemporal causal graph based on the efficacy deviation vector, identify and calculate causal links, and generate attribution confidence for each identified causal link; a graph weight updating unit, which is used to update the parameters of the causal link with the attribution confidence as the weight, incorporate new observation evidence while retaining historical knowledge, and realize the adaptive evolution of the dynamic spatiotemporal causal graph.

[0015] Based on the same inventive concept, the application also provides an AI-driven infrastructure risk operation optimization management method, which comprises the following steps: fusing multi-modal data streams of an infrastructure site in real time, wherein the multi-modal data streams comprise building information model data, Internet of Things sensor array data, environmental monitoring data and personnel physiological sign data, deducing nonlinear causal dependence relationships between physical entities through causal network deduction, and constructing a dynamic spatiotemporal causal graph; based on the dynamic spatiotemporal causal graph, applying virtual disturbance to key nodes to perform counterfactual simulation, deducing risk evolution paths and cascading effects caused by single-point risks, and constructing a forward-looking risk model; taking the forward-looking risk model as a constraint condition, coupling resource dependence relationships of the dynamic spatiotemporal causal graph, constructing a dynamic programming model, solving optimal collaborative defense strategies under different resource constraints, and outputting a time-varying risk intervention strategy matrix; analyzing the time-varying risk intervention strategy matrix into an instruction paradigm, combining a real-time state snapshot of the dynamic spatiotemporal causal graph, and generating an operation digital prescription for a specific dynamic risk scenario; continuously monitoring the site state evolution after the operation digital prescription is applied, comparing the real evolution result with the prediction result of the forward-looking risk model, calculating the effectiveness deviation of the intervention measures and performing attribution analysis, and adaptively updating the causal link weight in the dynamic spatiotemporal causal graph according to the attribution confidence.

[0016] Compared with the prior art, the application has the following advantages: 1. The application realizes the fundamental understanding of infrastructure risks by deeply fusing multi-source heterogeneous data and constructing a dynamic spatiotemporal causal graph, and changes the risk management from the traditional passive response mode based on surface phenomenon correlation to the active prediction mode based on deep causal mechanism. This way can identify complex chain reactions caused by small events which are difficult to be found by traditional statistical methods, so that accurate early warning can be performed at the risk germination stage, and the predictability and accuracy of risk prevention and control are improved from the source; 2. The application integrates risk evolution deduction and resource collaborative allocation optimization, takes the forward-looking risk model as a constraint condition of dynamic programming, ensures that the generated intervention strategy is not only optimal in theory, but also considers the limitations of limited manpower, material resources and other resources in practice. This makes the decision scheme no longer an isolated risk alarm, but a directly executable and globally collaborative operation instruction, thereby solving the problem of disconnection between risk management and operation scheduling, and greatly improving the emergency response speed and resource utilization efficiency; 3、The application constructs a complete feedback loop from decision execution to model correction. By analyzing the deviation between the actual effect of the intervention measures and the prediction model, and using the causal attribution technology to accurately locate the shortcomings in the model, the adaptive learning and evolution of the core causal graph are realized. This enables continuous learning from practice, continuously improving its cognitive depth and decision quality in specific project environments, overcoming the fundamental defect of traditional models that are fixed and cannot adapt to dynamic changes, ensuring long-term effectiveness and reliability.

[0017] Other features and advantages of the present application will be set forth in the following description, and in part will be apparent from the description, or can be learned by practice of the application. The objects and other advantages of the application will be realized and attained by the structure particularly pointed out in the written description and claims thereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0019] Figure 1 is a framework diagram of the AI-driven infrastructure risk operation optimization management system according to an embodiment of the present application.

[0020] Figure 2 is a structural schematic diagram of the infrastructure risk operation optimization management system according to an embodiment of the present application.

[0021] Figure 3 is a dynamic spatio-temporal causal graph structure schematic diagram according to an embodiment of the present application.

[0022] Figure 4 is a time sequence probability distribution cloud rain map of risk evolution deduction according to an embodiment of the present application.

[0023] Figure 5 is a collaborative defense strategy optimization configuration schematic diagram according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0025] Referring Figure 1 , one embodiment of the present application proposes a capital construction risk operation optimization management system, which adopts a comprehensive method of constructing a dynamic causal graph for risk deduction, coupling resource constraints for collaborative optimization, and closed-loop feedback correction of the model, which can realize active prediction, accurate intervention and continuous learning optimization of capital construction risks.

[0026] As Figure 2 shown, the system of the embodiment specifically includes: A multi-modal causal twin construction module is used to fuse multi-modal data streams in real time on the capital construction site, the multi-modal data streams include building information model data, Internet of Things sensor array data, environmental monitoring data and personnel physiological sign data, to deduce the nonlinear causal dependence relationship between physical entities, and to construct a dynamic spatiotemporal causal graph; A risk evolution deduction module is used to apply virtual disturbance to key nodes based on the dynamic spatiotemporal causal graph to perform counterfactual simulation, deduce the risk evolution path and cascading effect caused by a single point risk, and construct a forward-looking risk model; A collaborative configuration optimization module is used to take the forward-looking risk model as a constraint condition, and couple the resource dependence relationship of the dynamic spatiotemporal causal graph to construct a dynamic programming model, solve the optimal collaborative defense strategy under different resource constraints, and output a time-varying risk intervention strategy matrix; An instruction analysis and digital prescription generation module is used to analyze the time-varying risk intervention strategy matrix into an instruction paradigm, and generate a job digital prescription for a specific dynamic risk scenario in combination with a real-time state snapshot of the dynamic spatiotemporal causal graph; An intervention efficiency attribution and evolution correction module is used to continuously monitor the site state evolution after applying the job digital prescription, compare the real evolution result with the prediction result of the forward-looking risk model, calculate the efficiency deviation of the intervention measures and perform attribution analysis, and adaptively update the causal link weight in the dynamic spatiotemporal causal graph according to the attribution confidence.

[0027] The comprehensive method of constructing a dynamic causal graph for risk deduction, coupling resource constraints for collaborative optimization, and closed-loop feedback correction of the model can realize active prediction, accurate intervention and continuous learning optimization of capital construction risks Optionally, the multi-modal causal twin construction module includes: A data spatiotemporal alignment unit is used to map and synchronize the timestamps and spatial coordinates in the Internet of Things sensor array data, the environmental monitoring data and the personnel physiological sign data with the building information model data as a three-dimensional space reference, and output a multi-modal data set; Specifically, the data space-time alignment unit first loads the building information model (BIM) data. The BIM model here is a comprehensive three-dimensional model containing building geometry, spatial relationship, geographic information and component attributes, serving as a unified spatial reference for all subsequent data. For each piece of data collected from the Internet of Things (IoT) sensor array, environmental monitoring station and personnel wearable device, the unit calls a coordinate transformation function to map its spatial coordinates to the global coordinate system of the BIM model. At the same time, the timestamps of all data are synchronized with the central server through the Network Time Protocol (NTP) to ensure the consistency of the time reference. After mapping and synchronization, the data streams from different sources are integrated into a structured multi-modal data set under a unified space-time reference, providing high-quality input for subsequent feature extraction.

[0028] The cross-modal hidden feature extraction unit is configured to learn and extract the internal state and interaction behavior of the physical entity based on the multi-modal data set, to obtain a hidden feature vector. Specifically, the cross-modal hidden feature extraction unit uses a variational auto-encoder (VAE) to process the multi-modal data set. The VAE model maps high-dimensional input data to a probability distribution in a low-dimensional hidden space through an encoder network, and then reconstructs the original data from the hidden space through a decoder network. The training objective of the model is to maximize the evidence lower bound (ELBO), and the optimization objective function is: , In the formula, is the input multi-modal data sample; is the hidden feature vector; is the prior distribution of the hidden feature vector, which is usually a standard normal distribution; is the posterior distribution learned by the encoder network with parameters ; is the conditional distribution learned by the decoder network with parameters ; is the reconstruction loss term, which is used to measure the similarity between the generated data and the original data; is the Kullback-Leibler divergence, which is used to measure the difference between the posterior distribution and the prior distribution. By optimizing the objective function, the hidden feature vector $z$ learned by the unit can capture the core and decoupled features representing the internal state and interaction behavior of the physical entity.

[0029] a causal structure learning unit configured to perform temporal analysis on the hidden feature vectors, quantify nonlinear causal dependencies, and search and determine directed causal links and nonlinear function relationships among the hidden features under the constraint of acyclic dependence; Specifically, the causal structure learning unit receives the sequence of time-ordered hidden feature vectors output by the previous unit. To identify the causal structure among these temporal features, the unit applies a structure discovery algorithm based on continuous optimization. The algorithm parameterizes the adjacency matrix of the causal graph and constructs a loss function that includes the goodness of data fitting and the sparsity of the graph structure as a regularization term. A key constraint is the acyclic constraint, which ensures that there is no directed loop in the learned causal graph. For example, this constraint can be implemented by a penalty function that sharply increases in value when the trace of the power of the adjacency matrix is not zero. During the training process, the unit iteratively solves by optimization methods such as gradient descent, and finally obtains an adjacency matrix that describes the directed causal links and corresponding nonlinear function relationships among the hidden features.

[0030] a graph instantiation and weighting unit configured to instantiate a dynamic spatiotemporal causal graph with the physical entities of the building information modeling data as nodes, the directed causal links as edges, and the influence strength of the nonlinear function relationships as weights.

[0031] Specifically, the graph instantiation and weighting unit first extracts explicit physical entities from the BIM data, such as a specific tower crane, a section of load-bearing wall, or a high-risk work area, and takes them as nodes of the graph. Then, the unit establishes directed edges between the corresponding nodes according to the adjacency matrix output by the causal structure learning unit, representing the directed causal links between them. The weight of the edge, i.e., the influence strength of the nonlinear function relationship, is determined by calculating the expected partial derivative of the dependent variable's hidden feature vector with respect to the independent variable's hidden feature vector. This influence strength value quantifies the expected impact of a small change in one node on another node. Through the above steps, the unit finally instantiates a weighted, directed, and dynamic spatiotemporal causal graph that dynamically reflects the complex causal relationships among the elements of the construction site, such as Figure 3 as shown in FIG. 6, which is a schematic diagram of the chordal visualization of the dynamic spatiotemporal causal graph of the present application. The outer arc segments represent physical entity nodes, and the thickness of the internal chords represents the causal influence strength between nodes.

[0032] Optionally, the risk evolution deduction module comprises: A source positioning and disturbance set generating unit is configured to analyze the topology of the dynamic spatiotemporal causal graph, calculate the causal out-degree and influence weight of each node, locate the risk source based on the calculation results, and generate a parameterized virtual disturbance set around the risk source. Specifically, the source positioning and disturbance set generating unit first receives the dynamic spatiotemporal causal graph generated by the multi-modal causal twin construction module. To locate the risk source, the unit analyzes the topology of the graph, calculates two core indicators for each node: one is the causal out-degree, i.e., the number of directed edges pointing from the node to other nodes; the other is the influence weight, i.e., the cumulative sum of all out-edge weight values of the node. Then, the unit uses a weighted model to calculate the keyness score of each node, which sums the causal out-degree and influence weight of the node by weighting coefficients. The weighting coefficients can be configured according to historical data analysis or expert knowledge base to balance the influence range and depth of the node. After the calculation, the unit compares the keyness score of each node with the keyness threshold, and any node with a score exceeding the threshold is located as a risk source. Finally, the unit generates a set of parameterized virtual disturbance sets around the located risk source to simulate potential extreme working conditions based on the attributes of the physical entity represented by the risk source and historical abnormal data.

[0033] A causal chain deduction and cascade effect simulation unit is configured to apply the parameters in the parameterized virtual disturbance set to the risk source one by one, perform discrete-time iterative deduction in the dynamic spatiotemporal causal graph, simulate multiple risk evolution paths triggered by a single risk and their cross-influences, and obtain simulated evolution paths. Specifically, the causal chain deduction and cascade effect simulation unit starts the counterfactual simulation process. The unit extracts disturbance parameters from the parameterized virtual disturbance set generated by the previous unit and applies them to the corresponding risk source one by one as the initial conditions for simulation. Under the driving of discrete time steps, the unit iteratively calculates and updates the future state of downstream nodes affected by the disturbed nodes based on the nonlinear function relationships defined in the dynamic spatiotemporal causal graph. For example, at time t+1, the state of a node is a function of its state at time t and the states of all upstream nodes pointing to it. This deduction process propagates along the causal links of the graph until the risk effect dissipates in the network or reaches the simulation cutoff time. The process record of a single complete deduction constitutes a simulated evolution path. By iterating the simulation of all parameters in the parameterized virtual disturbance set, the unit can generate a large number of simulated evolution paths, which comprehensively reveal the various risk evolution processes and complex cascade effects that can be triggered by a single risk.

[0034] A forward-looking risk field construction unit is configured to statistically aggregate the simulation evolution paths, construct a risk propagation probability field in a four-dimensional space-time coordinate system, and quantify the probability of a specific strategy source point affecting any space-time point in the future, and serve as a forward-looking risk model.

[0035] Specifically, the forward-looking risk field construction unit statistically aggregates all simulation evolution paths generated by the previous unit. In a four-dimensional space-time coordinate system including three-dimensional spatial coordinates and one-dimensional time coordinates, it statistically aggregates the total frequency of risk state transmission and occurrence in each four-dimensional space-time cell in all simulation times. The risk probability of any space-time point is determined by dividing the total number of simulation paths of the point by the total number of simulations. By calculating the entire coordinate system, the unit finally constructs a dynamic and quantitative risk propagation probability field. The risk propagation probability field intuitively shows the distribution density and diffusion trend of the risk triggered by a specific risk strategy source point in the future time and space dimensions, and is finally encapsulated as a forward-looking risk model for subsequent collaborative configuration optimization module calls, such as Figure 4 As shown in the time sequence probability distribution cloud rain map of risk evolution deduction in the application, the dynamic evolution process of the probability distribution of the key risk indicators over time is shown.

[0036] Optionally, the parameterized virtual disturbance set includes a state offset vector set and a conduction failure operator set, wherein: The state offset vector set is used to simulate the extreme deterioration scenario of the physical entity's own attribute parameters in a statistical sense; Specifically, for the state offset vector set, the generation process first identifies the key state parameters of the physical entity corresponding to a certain risk strategy source point, such as the operating temperature of the device, the strain value of the structure, or the physiological fatigue index of the personnel, and forms a baseline state vector according to the normal operation data. Subsequently, based on industry safety specifications or historical extreme failure data, one or more target state vectors under deterioration scenarios are defined. The state offset vector is the difference between the target state vector and the baseline state vector, and the calculation formula is: , In the formula, is the state offset vector; is the target state vector under the deterioration scenario; is the baseline state vector. By generating multiple state offset vectors representing different deterioration directions and degrees, the state offset vector set is formed. During simulation deduction, the current state of a certain risk strategy source point is superimposed on a state offset vector, which simulates the instantaneous deterioration of the entity's own attributes.

[0037] The conduction failure operator set is used to simulate local interruption or nonlinear distortion of the causal relationship transmission path between entities in the dynamic spatiotemporal causal graph.

[0038] Specifically, for the conduction failure operator set, attention is paid to the "edges" in the dynamic spatiotemporal causal graph, that is, the causal transmission relationship between physical entities. A conduction failure operator is defined as a mathematical transformation acting on the causal link transmission function. For example, one operator is used to simulate local interruption of the causal path, which temporarily sets the influence strength weight corresponding to the path to zero in the simulation iteration. Another operator is used to simulate nonlinear distortion, which acts on the original causal transmission function through a distortion function to simulate unexpected amplification, attenuation or delay of signals in the transmission process and other complex situations. These operators are selected from an operator library according to common failure modes, and together form the conduction failure operator set.

[0039] Optionally, the system further comprises: The hidden feature vector is matched with the simulation evolution path, the distance on the initial state and the key feature points is calculated, the convergence degree of the current infrastructure site state to each risk evolution path is evaluated, and a path activation confidence vector is generated.

[0040] Specifically, to dynamically evaluate which risk path the real state of the current infrastructure site is tending to, the processing flow first obtains the hidden feature vector generated by the cross-modal hidden feature extraction unit, which can represent the real state of the current site. At the same time, the flow calls a set of simulation evolution paths generated by the causal chain deduction and cascade effect simulation unit. For each simulation evolution path, the flow extracts the hidden feature vector of the initial state and the hidden feature vector of several key turning points on the path. Subsequently, the flow calculates the Euclidean distance between the real-time hidden feature vector of the current state and the initial state vector of each simulation path and the hidden feature vector of the nearest key turning point. The two distance values are weighted and summed by a weight factor to obtain a comprehensive path convergence degree index. The smaller the index, the closer the current state is to the simulation evolution path. To convert the convergence degrees of all paths into standardized probability values, the flow finally adopts the Softmax function for normalization processing, thereby generating a path activation confidence vector, and the calculation formula is: , In the formula, is the activation confidence of the i-th simulation evolution path; is the path convergence degree of the current real-time state and the i-th simulation evolution path; is the path convergence degree of the current real-time state and all candidate simulation evolution paths; is the path convergence degree of the current real-time state and the i-th simulation evolution path; is the path convergence degree of the current real-time state and all candidate simulation evolution paths; indicates summation over all paths. temperature hyper-parameter, which is derived from cross-validation to adjust the smoothness of the confidence probability distribution, The smaller the value, the sharper the probability distribution. In the final generated path activation confidence vector, the value of each element represents the credibility of the current infrastructure site state developing along the corresponding risk evolution path.

[0041] Optionally, the collaborative configuration optimization module comprises: A multi-dimensional constraint space construction unit is configured to convert the risk propagation probability field of the forward-looking risk model into a state transition cost of the dynamic programming model, and extract a logical dependency relationship between resource scheduling and task execution from the dynamic spatiotemporal causal graph as a state constraint, to jointly define a solution space. Specifically, the multi-dimensional constraint space construction unit is responsible for defining an accurate mathematical environment for subsequent optimization solving. This unit performs two key operations: First, it receives the risk propagation probability field output by the forward-looking risk model and converts the probability values in this probability field into state transition costs in the dynamic programming model. Specifically, in a given state, when taking a certain action to transition to the next state, if the spatiotemporal location of the next state corresponds to a higher risk probability, then the cost of this state transition will increase accordingly. Second, this unit analyzes the dynamic spatiotemporal causal graph and extracts the resource mutual exclusion relationship between physical entities and the logical sequence of task execution, such as a specific device being able to perform only one task at the same time, or a task having to be started after another task is completed. These logical dependency relationships are formalized as a set of state constraints, which collectively constitute the boundary conditions of the solution space.

[0042] A policy iteration solving unit is configured to recursively solve a state-action sequence that can accumulate risk costs under the condition of satisfying the state constraints in the solution space, to obtain an optimal collaborative defense strategy. Specifically, the policy iteration solving unit uses a dynamic programming algorithm based on value iteration or policy iteration to solve in the multi-dimensional constraint space defined by the previous unit. The goal of this unit is to find an optimal policy, which is a complete mapping that can guide the selection of actions in any possible state. The solving process is a recursive process that starts from the final target state and calculates the optimal value function of each intermediate state in reverse. In each iteration, the unit evaluates the expected future cumulative risk cost of taking all possible actions in a certain state and selects the action that minimizes this cost. This process is repeated iteratively until the value functions of all states converge, and the resulting converged strategy is the optimal collaborative defense strategy.

[0043] A policy matrix generation unit is configured to map and discretize the optimal collaborative defense strategy in the time dimension to generate a time-varying risk intervention strategy matrix.

[0044] Specifically, the strategy matrix generating unit is responsible for converting the state-based optimal coordinated defense strategy output by the previous unit into a directly applicable format associated with the project schedule. This unit discretizes the total project duration into a series of time steps. At each time step, the unit predicts the most likely state according to the construction plan and queries the optimal intervention action corresponding to that state from the optimal coordinated defense strategy. By arranging and combining the optimal actions corresponding to each time step, the unit finally generates a two-dimensional or higher-dimensional matrix. The dimensions of the matrix can include time, risk type, intervention object, etc., and the elements in the matrix are specific intervention instructions or parameters, thus forming a time-varying risk intervention strategy matrix, as shown in Figure 5 The left triangular matrix shows the synergy or conflict relationship between different intervention measures, and the right bipartite graph shows the defense effectiveness of each intervention measure on different risk evolution paths.

[0045] Optionally, the instruction analysis and digital prescription generating module comprises: A strategy matrix analysis unit for decoding the time-varying risk intervention strategy matrix into structured intervention rule primitives; Specifically, the strategy matrix analysis unit receives the time-varying risk intervention strategy matrix generated by the coordinated configuration optimization module. The matrix is a high-dimensional data structure, and the internal elements encode the optimal intervention action under different times and different risk scenarios. The unit uses a decoder to analyze the matrix and translate the strategy information contained therein into a set of logical and machine-readable structured intervention rule primitives. Each primitive defines an atomic-level intervention action framework, for example, a primitive can be defined as: "when [risk scenario type] is activated, perform [operation type] on [target entity type], adjust [key parameter] to [target threshold]".

[0046] A dynamic risk scenario matching unit for identifying a dynamic risk scenario with the path activation confidence vector and matching the optimal intervention rule from the structured intervention rule primitives; Specifically, the dynamic risk scenario matching unit obtains the generated path activation confidence vector in real time. The unit identifies the simulated evolution path corresponding to the element with the highest confidence value in the vector as the most likely to occur dynamic risk scenario. Subsequently, the unit uses the identified dynamic risk scenario type as an index to search and match in the set of structured intervention rule primitives generated by the previous unit. According to a priority rule base, the unit selects the optimal intervention rule from all matched rules to deal with the most urgent risk.

[0047] A digital prescription generating unit is configured to call the real-time state snapshot of the dynamic spatio-temporal causal graph to instantiate the parameters in the optimal intervention rule, and encapsulate it as a standardized data package as a job digital prescription and issue it to the on-site execution end.

[0048] Specifically, the digital prescription generating unit receives the optimal intervention rule matched by the previous unit. At this time, the rule is still parameterized. In order to make it fully executable, the unit calls the dynamic spatio-temporal causal graph to obtain the real-time state snapshot of the physical entity related to the placeholders defined in the rule, such as "[target entity type]" and "[key parameter]". For example, it queries the ID and location of all personnel in a specific area, or the current running power of a specific device. The unit uses these real-time data to fill and instantiate the parameters in the optimal intervention rule. Finally, the unit encapsulates this fully instantiated intervention rule into a standardized data package, which is the job digital prescription, and issues it to the designated on-site execution end, such as the augmented reality glasses of the on-site manager or the automatic control of the specific device, through the communication interface.

[0049] Optionally, the standardized data package includes a job instruction set, risk context data, and performance verification parameters, wherein: The job instruction set is used to define the intervention action, parameter threshold, and execution timing for the target physical entity. Specifically, for the job instruction set, as the core execution part of the standardized data package, it defines the intervention task that needs to be executed in a structured and machine-readable format. The content of the instruction set includes the unique identifier of the target physical entity, the specific intervention action type, such as "reduce device running power", "add temporary support to the structure", or "evacuate personnel in a specified area"; at the same time, the instruction set also contains quantitative parameter thresholds, such as "reduce power to eighty percent of the rated value or lower", and strict execution timing, including the start time of the intervention action, the time node that must be completed, or the allowed response time window.

[0050] The risk context data is used to encapsulate the activated risk evolution path, the real-time state of the risk source node, and the expected loss assessment after intervention failure. Specifically, for the risk context data, it aims to provide necessary background information for the on-site execution end to enhance the understanding of the instructions. The data encapsulates the key information of the risk evolution path that triggers this intervention and is activated, including the source node and transmission path of the risk. At the same time, the data also contains the real-time state snapshot of the risk source node at the moment of instruction generation, to reveal the severity and urgency of the problem. In addition, the data also provides an expected loss assessment after several interventions based on the forward-looking risk model, to emphasize the importance of intervention actions.

[0051] The performance verification parameter is used to specify the monitoring indicators and state transition targets that need to be met for the intervention to be successful.

[0052] Specifically, for the performance verification parameter, it sets a clear and quantifiable objective standard for the subsequent evaluation of the effectiveness of the intervention. The parameter specifies one or more indicators that need to be closely monitored after the implementation of the intervention measure, and gives the state transition targets that these indicators need to reach after the intervention is successful. For example, a performance verification parameter can be defined as: within five minutes after executing the "reduce device running power" instruction, the latent feature vector of the corresponding node of the device must be migrated to the safe state space. These parameters provide benchmark data for the subsequent intervention performance attribution and evolution correction module.

[0053] Optionally, the intervention performance attribution and evolution correction module comprises: An effect monitoring and deviation calculation unit for continuously tracking the field data stream after the digital work prescription is issued, and comparing the real monitoring indicators with the state transition targets predicted by the forward-looking risk model to quantify the performance deviation vector. Specifically, after the digital work prescription is issued to the field execution end, the intervention performance monitoring and deviation calculation unit is activated. The unit continuously tracks the field multi-modal data stream corresponding to the monitoring indicators specified in the performance verification parameter encapsulated in the digital work prescription, thereby obtaining the real state evolution result after the implementation of the intervention measure. Subsequently, the unit compares the time series of the real monitoring indicators with the state transition targets in the performance verification parameter, calculates the difference between them item by item, and finally generates a performance deviation vector that can quantify the gap between the actual effect and the expected effect of the intervention measure.

[0054] A causal attribution and confidence evaluation unit for performing contribution traceability analysis on the dynamic spatiotemporal causal graph based on the performance deviation vector, identifying and calculating causal links, and generating an attribution confidence for each identified causal link. Specifically, the causal attribution and confidence evaluation unit takes the performance deviation vector as the starting point of analysis, performs contribution traceability analysis on the dynamic spatiotemporal causal graph to explore the root cause of the deviation. This analysis process can use methods such as Shapley Value allocation or gradient integration, aiming to identify one or more causal links that contribute most to the final prediction deviation. The unit traces all causal paths from the intervention node to the performance monitoring node, and calculates the contribution of each causal link on the path to the total deviation. After identifying the causal links to be corrected, the unit also evaluates and generates an attribution confidence for each identified causal link according to the size of the deviation, the frequency of occurrence, and the score output by the attribution algorithm itself.

[0055] a graph weight updating unit configured to update parameters of the causal links with the attribution confidence as a weight, to incorporate new observation evidence on the basis of preserving historical knowledge, and to realize adaptive evolution of the dynamic spatiotemporal causal graph.

[0056] Specifically, the graph weight updating unit updates the dynamic spatiotemporal causal graph according to the causal link to be corrected and the corresponding attribution confidence output by the preceding unit. The unit updates the parameters of the identified causal link in a Bayesian manner with the attribution confidence as a learning rate or adjustment weight. This updating process is an incremental learning manner, which regards the existing parameters of the causal link as prior knowledge and regards the observed deviation this time as new evidence, to calculate the posterior parameters fused with the new evidence. In this way, the unit incorporates new observation evidence on the basis of preserving historical knowledge, realizes adaptive evolution of the dynamic spatiotemporal causal graph, and enables the dynamic spatiotemporal causal graph to more accurately reflect the physical laws of the real world.

[0057] Based on the same inventive concept, the application further provides an AI-driven infrastructure risk operation optimization management method, which comprises the following steps: fusing a multi-modal data stream of an infrastructure site in real time, the multi-modal data stream comprising building information model data, Internet of Things sensor array data, environmental monitoring data, and personnel physiological sign data, performing causal network deduction on a non-linear causal dependency relationship between physical entities, and constructing a dynamic spatiotemporal causal graph; based on the dynamic spatiotemporal causal graph, performing counterfactual simulation by applying a virtual disturbance to a key node, deducing a risk evolution path and cascading effect triggered by a single-point risk, and constructing a forward-looking risk model; taking the forward-looking risk model as a constraint condition, coupling a resource dependency relationship of the dynamic spatiotemporal causal graph, constructing a dynamic programming model, solving an optimal cooperative defense strategy under different resource constraints, and outputting a time-varying risk intervention strategy matrix; analyzing the time-varying risk intervention strategy matrix into an instruction paradigm, combining a real-time state snapshot of the dynamic spatiotemporal causal graph, and generating an operation digital prescription for a specific dynamic risk scenario; continuously monitoring a site state evolution after the operation digital prescription is applied, comparing a real evolution result with a prediction result of the forward-looking risk model, calculating an effectiveness deviation of the intervention measure and performing attribution analysis, and adaptively updating a weight of a causal link in the dynamic spatiotemporal causal graph according to an attribution confidence.

[0058] To verify the feasibility of the application, the application is applied to high-altitude operation risk management of a P1 main tower of a sea-crossing bridge.

[0059] The application firstly constructs a dynamic spatio-temporal causal graph containing nonlinear causal dependence relationships between entities by spatiotemporal alignment and feature extraction of multi-modal data streams such as BIM, Internet of Things and personnel physiology in the field through a multi-modal causal twin modeling module.

[0060] Subsequently, the risk evolution deduction module identifies that the "P1 main tower concrete curing state" is a risk source point based on the graph, and applies a parameterized virtual disturbance set simulating abnormal cooling to perform counterfactual simulation, thereby constructing a forward-looking risk model that predicts the risk of structural stress exceeding the limit within the next 24 hours. Based on this model, the collaborative configuration optimization module solves through dynamic programming to generate a time-varying risk intervention strategy matrix.

[0061] During project operation, it is monitored that the real-time state matches a certain pre-play risk path with a high degree of matching, generating an 88% path activation confidence vector. The instruction analysis and digital prescription generation module immediately responds, matches and instantiates the rules in the strategy matrix according to the activated scene, generates and issues a job digital prescription containing a job instruction set, risk context data and performance verification parameters, and the core instruction is: "immediately start the P1 main tower heat tracing and insulation system".

[0062] After intervention, the intervention performance attribution and evolution correction module monitors that there is a deviation between the actual strain value and the performance verification parameter, and through contribution degree tracing analysis, it locates that the model has insufficient weight estimation for the "high-altitude wind speed affecting heat dissipation". The module immediately updates the weight of the causal link according to the attribution confidence, completing the learning closed loop.

[0063] Table 1 Dynamic spatio-temporal causal graph construction and key node identification data representation example Table 2 Risk evolution deduction and digital prescription generation data representation example Table 3 Intervention performance attribution and model evolution correction data representation example As can be seen from the data in Tables 1 to 3 above, the method of the application can effectively support the intelligent management of infrastructure risks. Table 1 clearly shows that the "core tube concrete" is identified as the core risk source point through quantitative analysis. Table 2 records how to identify dynamic risks based on high-confidence path activation and quickly generate precise intervention instructions. Table 3 specifically embodies the closed-loop learning capability of the application, which can locate the knowledge defects in the model and perform adaptive correction through attribution analysis of the performance deviation after intervention, ensuring the continuous evolution and long-term effectiveness of the model.

[0064] It should be noted that the above formulae can be converted into unitless standard values or parameters of the same dimension that can be superimposed by using the dimensional consistency principle and mathematical standardization methods (for example, normalization processing, dimensionless parameter conversion, or unit system unification), so as to eliminate the interference of different dimensions on the operation logic, make the formula have mathematical operation rationality and objective law adaptability while preserving the original data distribution characteristics. It is a conventional technical means, and will not be described here. The electrical connection between the above-mentioned units does not necessarily mean direct connection, and indirect connection can also be used as long as the purpose of the application is achieved. The above-mentioned is only an exemplary embodiment of the application, and cannot limit the scope of the application.

[0065] That is, any equivalent changes and modifications made according to the teachings of the present application are still within the scope of the present application. Other embodiments of the present application will be readily apparent to those skilled in the art upon considering the description and practice of the principles disclosed herein. The present application is intended to cover any variations, uses, or adaptive changes of the present application that follow the general principles of the present application and include common knowledge or conventional technical means in the art that are not described in the present application.

Claims

1. AI-driven infrastructure risk operation optimization management system, characterized by: The system comprises: A multimodal causal twin construction module is used to integrate multimodal data streams from infrastructure sites in real time. The multimodal data streams include building information model data, IoT sensor array data, environmental monitoring data, and personnel physiological sign data. This module uses causal networks to deduce nonlinear causal dependencies between physical entities and construct a dynamic spatiotemporal causal graph. A risk evolution deduction module is used to apply virtual disturbances to key nodes to perform counterfactual simulations based on the dynamic spatiotemporal causal graph, deduce the risk evolution path and cascading effects caused by single-point risks, and build a forward-looking risk model; A collaborative configuration optimization module is used to use the forward-looking risk model as a constraint condition and couple the resource dependencies of the dynamic spatiotemporal causal graph to construct a dynamic programming model, solve the optimal collaborative defense strategy under different resource constraints, and output a time-varying risk intervention strategy matrix; An instruction parsing and digital prescription generation module is used to parse the time-varying risk intervention strategy matrix into an instruction paradigm, and generate a digital prescription for a specific dynamic risk scenario in combination with a real-time state snapshot of the dynamic spatiotemporal causal graph; The intervention effectiveness attribution and evolution correction module is used to continuously monitor the evolution of the on-site status after the application of the digital prescription for the operation, compare the actual evolution results with the predicted results of the forward-looking risk model, calculate the effectiveness deviation of the intervention measures and perform attribution analysis, and adaptively update the causal link weights in the dynamic spatiotemporal causal graph based on the attribution confidence.

2. The AI-driven infrastructure risk operation optimization management system according to claim 1 is characterized in that: The multimodal causal twin building module includes: a data spatiotemporal alignment unit, configured to map and synchronize the timestamps and spatial coordinates of the IoT sensor array data, the environmental monitoring data, and the personnel physiological sign data using the building information model data as a three-dimensional spatial reference, and output a multimodal data set; A cross-modal latent feature extraction unit, configured to learn and extract the intrinsic state and interaction behavior of the physical entity based on the multimodal dataset to obtain a latent feature vector; A causal structure learning unit is used to perform time series analysis on the latent feature vectors, quantify nonlinear causal dependencies, and search and determine directed causal links and nonlinear functional relationships of each latent feature under the condition of satisfying acyclic constraints; The graph instantiation and weighting unit is used to instantiate and generate a dynamic spatiotemporal causal graph using the physical entities of the building information model data as nodes, the directed causal links as edges, and the influence intensity of the nonlinear functional relationship as weight.

3. The AI-driven infrastructure risk operation optimization management system according to claim 2 is characterized in that: The risk evolution deduction module includes: A risk source location and disturbance set generation unit, configured to analyze the topological structure of the dynamic spatiotemporal causal graph, calculate the causal out-degree and influence weight of each node, locate the risk source point based on the calculation results, and generate a parameterized virtual disturbance set around the risk source point; A causal chain deduction and cascade effect simulation unit is used to apply the parameters of the parameterized virtual disturbance set to the risk source point one by one, perform discrete time iterative deduction in the dynamic spatiotemporal causal graph, simulate multiple risk evolution paths triggered by single-point risks and their cross-influences, and obtain a simulated evolution path; The forward-looking risk field construction unit is used to statistically aggregate the simulation evolution path, construct a risk propagation probability field in a four-dimensional space-time coordinate system, quantify the probability that any future space-time point will be affected by the risk of a specific source point, and serve as a forward-looking risk model.

4. The AI-driven infrastructure risk operation optimization management system according to claim 3 is characterized in that: The parameterized virtual disturbance set includes: a state offset vector set and a conduction failure operator set, wherein: The state offset vector set is used to simulate the extreme deterioration scenario of the physical entity's own attribute parameters in a statistical sense; The conduction failure operator set is used to simulate the local interruption or nonlinear distortion of the causal relationship transmission path between entities in the dynamic spatiotemporal causal graph.

5. The AI-driven infrastructure risk operation optimization management system according to claim 3 is characterized in that: The system further comprises: The latent feature vector is matched with the simulation evolution path, the distance between the initial state and the key feature points is calculated, the degree of convergence of the current infrastructure site state to each risk evolution path is evaluated, and a path activation confidence vector is generated.

6. The AI-driven infrastructure risk operation optimization management system according to claim 1 is characterized in that: The collaborative configuration optimization module includes: a multidimensional constraint space construction unit for converting the risk propagation probability field of the forward-looking risk model into the state transition cost of the dynamic programming model, and extracting the logical dependency between resource scheduling and task execution from the dynamic spatiotemporal causal graph as state constraints to jointly define a solution space; A strategy iteration solving unit is used to recursively solve, within the solution space, a state-action sequence that can accumulate risk costs while satisfying the state constraints, to obtain an optimal collaborative defense strategy; The strategy matrix generating unit is used to map and discretize the optimal collaborative defense strategy in the time dimension to generate a time-varying risk intervention strategy matrix.

7. The AI-driven infrastructure risk operation optimization management system according to claim 5 is characterized in that: The instruction parsing and digital prescription generation module includes: a strategy matrix parsing unit, configured to decode the time-varying risk intervention strategy matrix into structured intervention rule primitives; a dynamic risk scenario matching unit, configured to identify dynamic risk scenarios using the path activation confidence vector and match optimal intervention rules from the structured intervention rule primitives; The digital prescription generation unit is used to call the real-time status snapshot of the dynamic spatiotemporal causal graph to instantiate the parameters in the optimal intervention rule, and encapsulate it into a standardized data package as a digital prescription for the operation and send it to the on-site execution end.

8. The AI-driven infrastructure risk operation optimization management system according to claim 7 is characterized in that: The standardized data package includes: an operation instruction set, risk context data, and performance verification parameters, wherein: The job instruction set is used to define intervention actions, parameter thresholds, and execution timings for a target physical entity; The risk context data is used to encapsulate the activated risk evolution path, the real-time status of the risk source node, and the expected loss assessment after the intervention fails; The efficacy verification parameters are used to specify the monitoring indicators and state transition targets that need to be met for the intervention to be successful.

9. The AI-driven infrastructure risk operation optimization management system according to claim 1 is characterized in that: The intervention effectiveness attribution and evolution correction module includes: An effect monitoring and deviation calculation unit, configured to continuously track the on-site data flow after the digital prescription for the operation is issued, and compare the actual monitoring indicators with the state transition targets predicted by the forward-looking risk model to quantify the effectiveness deviation vector; a causal attribution and confidence evaluation unit, configured to perform contribution traceability analysis on the dynamic spatiotemporal causal graph based on the effectiveness deviation vector, identify and calculate causal links, and generate an attribution confidence for each identified causal link; The graph weight updating unit is used to update the parameters of the causal link with the attribution confidence as the weight, incorporate new observation evidence on the basis of retaining historical knowledge, and realize the adaptive evolution of the dynamic spatiotemporal causal graph.

10. An AI-driven infrastructure risk operation optimization management method, applied to an AI-driven infrastructure risk operation optimization management system according to any one of claims 1 to 9, characterized in that: The method comprises: Real-time integration of multimodal data streams from infrastructure construction sites, including building information model data, IoT sensor array data, environmental monitoring data, and personnel physiological sign data, to perform causal network deduction on the nonlinear causal dependencies between physical entities and construct a dynamic spatiotemporal causal graph. Based on the dynamic spatiotemporal causal graph, virtual disturbances are imposed on key nodes to conduct counterfactual simulations, deduce the risk evolution path and cascading effects caused by single-point risks, and build a forward-looking risk model; Taking the forward-looking risk model as a constraint and coupling it with the resource dependency of the dynamic spatiotemporal causal graph, a dynamic programming model is constructed to solve the optimal collaborative defense strategy under different resource constraints and output a time-varying risk intervention strategy matrix. Parsing the time-varying risk intervention strategy matrix into an instruction paradigm, and combining it with a real-time state snapshot of the dynamic spatiotemporal causal graph to generate a digital prescription for a specific dynamic risk scenario; Continuously monitor the evolution of on-site status after applying the digital prescription for the operation, compare the actual evolution results with the predicted results of the forward-looking risk model, calculate the effectiveness deviation of the intervention measures and perform attribution analysis, and adaptively update the causal link weights in the dynamic spatiotemporal causal graph based on the attribution confidence.

Images